init

.gitattributes

@@ -0,0 +1,3 @@
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.ttc filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,18 @@
+*.pt
+*.zip
+*.7z
+*.rar
+*.ckpt
+*.onnx
+*.tar
+*.jpg
+*.ttc
+
+wandb
+data/backup
+data/dataset
+data/fonts
+data/font_statics*
+data/examples/annotations/*
+__pycache__
+.vscode

LICENSE

@@ -0,0 +1,674 @@
+                    GNU GENERAL PUBLIC LICENSE
+                       Version 3, 29 June 2007
+
+ Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
+ Everyone is permitted to copy and distribute verbatim copies
+ of this license document, but changing it is not allowed.
+
+                            Preamble
+
+  The GNU General Public License is a free, copyleft license for
+software and other kinds of works.
+
+  The licenses for most software and other practical works are designed
+to take away your freedom to share and change the works.  By contrast,
+the GNU General Public License is intended to guarantee your freedom to
+share and change all versions of a program--to make sure it remains free
+software for all its users.  We, the Free Software Foundation, use the
+GNU General Public License for most of our software; it applies also to
+any other work released this way by its authors.  You can apply it to
+your programs, too.
+
+  When we speak of free software, we are referring to freedom, not
+price.  Our General Public Licenses are designed to make sure that you
+have the freedom to distribute copies of free software (and charge for
+them if you wish), that you receive source code or can get it if you
+want it, that you can change the software or use pieces of it in new
+free programs, and that you know you can do these things.
+
+  To protect your rights, we need to prevent others from denying you
+these rights or asking you to surrender the rights.  Therefore, you have
+certain responsibilities if you distribute copies of the software, or if
+you modify it: responsibilities to respect the freedom of others.
+
+  For example, if you distribute copies of such a program, whether
+gratis or for a fee, you must pass on to the recipients the same
+freedoms that you received.  You must make sure that they, too, receive
+or can get the source code.  And you must show them these terms so they
+know their rights.
+
+  Developers that use the GNU GPL protect your rights with two steps:
+(1) assert copyright on the software, and (2) offer you this License
+giving you legal permission to copy, distribute and/or modify it.
+
+  For the developers' and authors' protection, the GPL clearly explains
+that there is no warranty for this free software.  For both users' and
+authors' sake, the GPL requires that modified versions be marked as
+changed, so that their problems will not be attributed erroneously to
+authors of previous versions.
+
+  Some devices are designed to deny users access to install or run
+modified versions of the software inside them, although the manufacturer
+can do so.  This is fundamentally incompatible with the aim of
+protecting users' freedom to change the software.  The systematic
+pattern of such abuse occurs in the area of products for individuals to
+use, which is precisely where it is most unacceptable.  Therefore, we
+have designed this version of the GPL to prohibit the practice for those
+products.  If such problems arise substantially in other domains, we
+stand ready to extend this provision to those domains in future versions
+of the GPL, as needed to protect the freedom of users.
+
+  Finally, every program is threatened constantly by software patents.
+States should not allow patents to restrict development and use of
+software on general-purpose computers, but in those that do, we wish to
+avoid the special danger that patents applied to a free program could
+make it effectively proprietary.  To prevent this, the GPL assures that
+patents cannot be used to render the program non-free.
+
+  The precise terms and conditions for copying, distribution and
+modification follow.
+
+                       TERMS AND CONDITIONS
+
+  0. Definitions.
+
+  "This License" refers to version 3 of the GNU General Public License.
+
+  "Copyright" also means copyright-like laws that apply to other kinds of
+works, such as semiconductor masks.
+
+  "The Program" refers to any copyrightable work licensed under this
+License.  Each licensee is addressed as "you".  "Licensees" and
+"recipients" may be individuals or organizations.
+
+  To "modify" a work means to copy from or adapt all or part of the work
+in a fashion requiring copyright permission, other than the making of an
+exact copy.  The resulting work is called a "modified version" of the
+earlier work or a work "based on" the earlier work.
+
+  A "covered work" means either the unmodified Program or a work based
+on the Program.
+
+  To "propagate" a work means to do anything with it that, without
+permission, would make you directly or secondarily liable for
+infringement under applicable copyright law, except executing it on a
+computer or modifying a private copy.  Propagation includes copying,
+distribution (with or without modification), making available to the
+public, and in some countries other activities as well.
+
+  To "convey" a work means any kind of propagation that enables other
+parties to make or receive copies.  Mere interaction with a user through
+a computer network, with no transfer of a copy, is not conveying.
+
+  An interactive user interface displays "Appropriate Legal Notices"
+to the extent that it includes a convenient and prominently visible
+feature that (1) displays an appropriate copyright notice, and (2)
+tells the user that there is no warranty for the work (except to the
+extent that warranties are provided), that licensees may convey the
+work under this License, and how to view a copy of this License.  If
+the interface presents a list of user commands or options, such as a
+menu, a prominent item in the list meets this criterion.
+
+  1. Source Code.
+
+  The "source code" for a work means the preferred form of the work
+for making modifications to it.  "Object code" means any non-source
+form of a work.
+
+  A "Standard Interface" means an interface that either is an official
+standard defined by a recognized standards body, or, in the case of
+interfaces specified for a particular programming language, one that
+is widely used among developers working in that language.
+
+  The "System Libraries" of an executable work include anything, other
+than the work as a whole, that (a) is included in the normal form of
+packaging a Major Component, but which is not part of that Major
+Component, and (b) serves only to enable use of the work with that
+Major Component, or to implement a Standard Interface for which an
+implementation is available to the public in source code form.  A
+"Major Component", in this context, means a major essential component
+(kernel, window system, and so on) of the specific operating system
+(if any) on which the executable work runs, or a compiler used to
+produce the work, or an object code interpreter used to run it.
+
+  The "Corresponding Source" for a work in object code form means all
+the source code needed to generate, install, and (for an executable
+work) run the object code and to modify the work, including scripts to
+control those activities.  However, it does not include the work's
+System Libraries, or general-purpose tools or generally available free
+programs which are used unmodified in performing those activities but
+which are not part of the work.  For example, Corresponding Source
+includes interface definition files associated with source files for
+the work, and the source code for shared libraries and dynamically
+linked subprograms that the work is specifically designed to require,
+such as by intimate data communication or control flow between those
+subprograms and other parts of the work.
+
+  The Corresponding Source need not include anything that users
+can regenerate automatically from other parts of the Corresponding
+Source.
+
+  The Corresponding Source for a work in source code form is that
+same work.
+
+  2. Basic Permissions.
+
+  All rights granted under this License are granted for the term of
+copyright on the Program, and are irrevocable provided the stated
+conditions are met.  This License explicitly affirms your unlimited
+permission to run the unmodified Program.  The output from running a
+covered work is covered by this License only if the output, given its
+content, constitutes a covered work.  This License acknowledges your
+rights of fair use or other equivalent, as provided by copyright law.
+
+  You may make, run and propagate covered works that you do not
+convey, without conditions so long as your license otherwise remains
+in force.  You may convey covered works to others for the sole purpose
+of having them make modifications exclusively for you, or provide you
+with facilities for running those works, provided that you comply with
+the terms of this License in conveying all material for which you do
+not control copyright.  Those thus making or running the covered works
+for you must do so exclusively on your behalf, under your direction
+and control, on terms that prohibit them from making any copies of
+your copyrighted material outside their relationship with you.
+
+  Conveying under any other circumstances is permitted solely under
+the conditions stated below.  Sublicensing is not allowed; section 10
+makes it unnecessary.
+
+  3. Protecting Users' Legal Rights From Anti-Circumvention Law.
+
+  No covered work shall be deemed part of an effective technological
+measure under any applicable law fulfilling obligations under article
+11 of the WIPO copyright treaty adopted on 20 December 1996, or
+similar laws prohibiting or restricting circumvention of such
+measures.
+
+  When you convey a covered work, you waive any legal power to forbid
+circumvention of technological measures to the extent such circumvention
+is effected by exercising rights under this License with respect to
+the covered work, and you disclaim any intention to limit operation or
+modification of the work as a means of enforcing, against the work's
+users, your or third parties' legal rights to forbid circumvention of
+technological measures.
+
+  4. Conveying Verbatim Copies.
+
+  You may convey verbatim copies of the Program's source code as you
+receive it, in any medium, provided that you conspicuously and
+appropriately publish on each copy an appropriate copyright notice;
+keep intact all notices stating that this License and any
+non-permissive terms added in accord with section 7 apply to the code;
+keep intact all notices of the absence of any warranty; and give all
+recipients a copy of this License along with the Program.
+
+  You may charge any price or no price for each copy that you convey,
+and you may offer support or warranty protection for a fee.
+
+  5. Conveying Modified Source Versions.
+
+  You may convey a work based on the Program, or the modifications to
+produce it from the Program, in the form of source code under the
+terms of section 4, provided that you also meet all of these conditions:
+
+    a) The work must carry prominent notices stating that you modified
+    it, and giving a relevant date.
+
+    b) The work must carry prominent notices stating that it is
+    released under this License and any conditions added under section
+    7.  This requirement modifies the requirement in section 4 to
+    "keep intact all notices".
+
+    c) You must license the entire work, as a whole, under this
+    License to anyone who comes into possession of a copy.  This
+    License will therefore apply, along with any applicable section 7
+    additional terms, to the whole of the work, and all its parts,
+    regardless of how they are packaged.  This License gives no
+    permission to license the work in any other way, but it does not
+    invalidate such permission if you have separately received it.
+
+    d) If the work has interactive user interfaces, each must display
+    Appropriate Legal Notices; however, if the Program has interactive
+    interfaces that do not display Appropriate Legal Notices, your
+    work need not make them do so.
+
+  A compilation of a covered work with other separate and independent
+works, which are not by their nature extensions of the covered work,
+and which are not combined with it such as to form a larger program,
+in or on a volume of a storage or distribution medium, is called an
+"aggregate" if the compilation and its resulting copyright are not
+used to limit the access or legal rights of the compilation's users
+beyond what the individual works permit.  Inclusion of a covered work
+in an aggregate does not cause this License to apply to the other
+parts of the aggregate.
+
+  6. Conveying Non-Source Forms.
+
+  You may convey a covered work in object code form under the terms
+of sections 4 and 5, provided that you also convey the
+machine-readable Corresponding Source under the terms of this License,
+in one of these ways:
+
+    a) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by the
+    Corresponding Source fixed on a durable physical medium
+    customarily used for software interchange.
+
+    b) Convey the object code in, or embodied in, a physical product
+    (including a physical distribution medium), accompanied by a
+    written offer, valid for at least three years and valid for as
+    long as you offer spare parts or customer support for that product
+    model, to give anyone who possesses the object code either (1) a
+    copy of the Corresponding Source for all the software in the
+    product that is covered by this License, on a durable physical
+    medium customarily used for software interchange, for a price no
+    more than your reasonable cost of physically performing this
+    conveying of source, or (2) access to copy the
+    Corresponding Source from a network server at no charge.
+
+    c) Convey individual copies of the object code with a copy of the
+    written offer to provide the Corresponding Source.  This
+    alternative is allowed only occasionally and noncommercially, and
+    only if you received the object code with such an offer, in accord
+    with subsection 6b.
+
+    d) Convey the object code by offering access from a designated
+    place (gratis or for a charge), and offer equivalent access to the
+    Corresponding Source in the same way through the same place at no
+    further charge.  You need not require recipients to copy the
+    Corresponding Source along with the object code.  If the place to
+    copy the object code is a network server, the Corresponding Source
+    may be on a different server (operated by you or a third party)
+    that supports equivalent copying facilities, provided you maintain
+    clear directions next to the object code saying where to find the
+    Corresponding Source.  Regardless of what server hosts the
+    Corresponding Source, you remain obligated to ensure that it is
+    available for as long as needed to satisfy these requirements.
+
+    e) Convey the object code using peer-to-peer transmission, provided
+    you inform other peers where the object code and Corresponding
+    Source of the work are being offered to the general public at no
+    charge under subsection 6d.
+
+  A separable portion of the object code, whose source code is excluded
+from the Corresponding Source as a System Library, need not be
+included in conveying the object code work.
+
+  A "User Product" is either (1) a "consumer product", which means any
+tangible personal property which is normally used for personal, family,
+or household purposes, or (2) anything designed or sold for incorporation
+into a dwelling.  In determining whether a product is a consumer product,
+doubtful cases shall be resolved in favor of coverage.  For a particular
+product received by a particular user, "normally used" refers to a
+typical or common use of that class of product, regardless of the status
+of the particular user or of the way in which the particular user
+actually uses, or expects or is expected to use, the product.  A product
+is a consumer product regardless of whether the product has substantial
+commercial, industrial or non-consumer uses, unless such uses represent
+the only significant mode of use of the product.
+
+  "Installation Information" for a User Product means any methods,
+procedures, authorization keys, or other information required to install
+and execute modified versions of a covered work in that User Product from
+a modified version of its Corresponding Source.  The information must
+suffice to ensure that the continued functioning of the modified object
+code is in no case prevented or interfered with solely because
+modification has been made.
+
+  If you convey an object code work under this section in, or with, or
+specifically for use in, a User Product, and the conveying occurs as
+part of a transaction in which the right of possession and use of the
+User Product is transferred to the recipient in perpetuity or for a
+fixed term (regardless of how the transaction is characterized), the
+Corresponding Source conveyed under this section must be accompanied
+by the Installation Information.  But this requirement does not apply
+if neither you nor any third party retains the ability to install
+modified object code on the User Product (for example, the work has
+been installed in ROM).
+
+  The requirement to provide Installation Information does not include a
+requirement to continue to provide support service, warranty, or updates
+for a work that has been modified or installed by the recipient, or for
+the User Product in which it has been modified or installed.  Access to a
+network may be denied when the modification itself materially and
+adversely affects the operation of the network or violates the rules and
+protocols for communication across the network.
+
+  Corresponding Source conveyed, and Installation Information provided,
+in accord with this section must be in a format that is publicly
+documented (and with an implementation available to the public in
+source code form), and must require no special password or key for
+unpacking, reading or copying.
+
+  7. Additional Terms.
+
+  "Additional permissions" are terms that supplement the terms of this
+License by making exceptions from one or more of its conditions.
+Additional permissions that are applicable to the entire Program shall
+be treated as though they were included in this License, to the extent
+that they are valid under applicable law.  If additional permissions
+apply only to part of the Program, that part may be used separately
+under those permissions, but the entire Program remains governed by
+this License without regard to the additional permissions.
+
+  When you convey a copy of a covered work, you may at your option
+remove any additional permissions from that copy, or from any part of
+it.  (Additional permissions may be written to require their own
+removal in certain cases when you modify the work.)  You may place
+additional permissions on material, added by you to a covered work,
+for which you have or can give appropriate copyright permission.
+
+  Notwithstanding any other provision of this License, for material you
+add to a covered work, you may (if authorized by the copyright holders of
+that material) supplement the terms of this License with terms:
+
+    a) Disclaiming warranty or limiting liability differently from the
+    terms of sections 15 and 16 of this License; or
+
+    b) Requiring preservation of specified reasonable legal notices or
+    author attributions in that material or in the Appropriate Legal
+    Notices displayed by works containing it; or
+
+    c) Prohibiting misrepresentation of the origin of that material, or
+    requiring that modified versions of such material be marked in
+    reasonable ways as different from the original version; or
+
+    d) Limiting the use for publicity purposes of names of licensors or
+    authors of the material; or
+
+    e) Declining to grant rights under trademark law for use of some
+    trade names, trademarks, or service marks; or
+
+    f) Requiring indemnification of licensors and authors of that
+    material by anyone who conveys the material (or modified versions of
+    it) with contractual assumptions of liability to the recipient, for
+    any liability that these contractual assumptions directly impose on
+    those licensors and authors.
+
+  All other non-permissive additional terms are considered "further
+restrictions" within the meaning of section 10.  If the Program as you
+received it, or any part of it, contains a notice stating that it is
+governed by this License along with a term that is a further
+restriction, you may remove that term.  If a license document contains
+a further restriction but permits relicensing or conveying under this
+License, you may add to a covered work material governed by the terms
+of that license document, provided that the further restriction does
+not survive such relicensing or conveying.
+
+  If you add terms to a covered work in accord with this section, you
+must place, in the relevant source files, a statement of the
+additional terms that apply to those files, or a notice indicating
+where to find the applicable terms.
+
+  Additional terms, permissive or non-permissive, may be stated in the
+form of a separately written license, or stated as exceptions;
+the above requirements apply either way.
+
+  8. Termination.
+
+  You may not propagate or modify a covered work except as expressly
+provided under this License.  Any attempt otherwise to propagate or
+modify it is void, and will automatically terminate your rights under
+this License (including any patent licenses granted under the third
+paragraph of section 11).
+
+  However, if you cease all violation of this License, then your
+license from a particular copyright holder is reinstated (a)
+provisionally, unless and until the copyright holder explicitly and
+finally terminates your license, and (b) permanently, if the copyright
+holder fails to notify you of the violation by some reasonable means
+prior to 60 days after the cessation.
+
+  Moreover, your license from a particular copyright holder is
+reinstated permanently if the copyright holder notifies you of the
+violation by some reasonable means, this is the first time you have
+received notice of violation of this License (for any work) from that
+copyright holder, and you cure the violation prior to 30 days after
+your receipt of the notice.
+
+  Termination of your rights under this section does not terminate the
+licenses of parties who have received copies or rights from you under
+this License.  If your rights have been terminated and not permanently
+reinstated, you do not qualify to receive new licenses for the same
+material under section 10.
+
+  9. Acceptance Not Required for Having Copies.
+
+  You are not required to accept this License in order to receive or
+run a copy of the Program.  Ancillary propagation of a covered work
+occurring solely as a consequence of using peer-to-peer transmission
+to receive a copy likewise does not require acceptance.  However,
+nothing other than this License grants you permission to propagate or
+modify any covered work.  These actions infringe copyright if you do
+not accept this License.  Therefore, by modifying or propagating a
+covered work, you indicate your acceptance of this License to do so.
+
+  10. Automatic Licensing of Downstream Recipients.
+
+  Each time you convey a covered work, the recipient automatically
+receives a license from the original licensors, to run, modify and
+propagate that work, subject to this License.  You are not responsible
+for enforcing compliance by third parties with this License.
+
+  An "entity transaction" is a transaction transferring control of an
+organization, or substantially all assets of one, or subdividing an
+organization, or merging organizations.  If propagation of a covered
+work results from an entity transaction, each party to that
+transaction who receives a copy of the work also receives whatever
+licenses to the work the party's predecessor in interest had or could
+give under the previous paragraph, plus a right to possession of the
+Corresponding Source of the work from the predecessor in interest, if
+the predecessor has it or can get it with reasonable efforts.
+
+  You may not impose any further restrictions on the exercise of the
+rights granted or affirmed under this License.  For example, you may
+not impose a license fee, royalty, or other charge for exercise of
+rights granted under this License, and you may not initiate litigation
+(including a cross-claim or counterclaim in a lawsuit) alleging that
+any patent claim is infringed by making, using, selling, offering for
+sale, or importing the Program or any portion of it.
+
+  11. Patents.
+
+  A "contributor" is a copyright holder who authorizes use under this
+License of the Program or a work on which the Program is based.  The
+work thus licensed is called the contributor's "contributor version".
+
+  A contributor's "essential patent claims" are all patent claims
+owned or controlled by the contributor, whether already acquired or
+hereafter acquired, that would be infringed by some manner, permitted
+by this License, of making, using, or selling its contributor version,
+but do not include claims that would be infringed only as a
+consequence of further modification of the contributor version.  For
+purposes of this definition, "control" includes the right to grant
+patent sublicenses in a manner consistent with the requirements of
+this License.
+
+  Each contributor grants you a non-exclusive, worldwide, royalty-free
+patent license under the contributor's essential patent claims, to
+make, use, sell, offer for sale, import and otherwise run, modify and
+propagate the contents of its contributor version.
+
+  In the following three paragraphs, a "patent license" is any express
+agreement or commitment, however denominated, not to enforce a patent
+(such as an express permission to practice a patent or covenant not to
+sue for patent infringement).  To "grant" such a patent license to a
+party means to make such an agreement or commitment not to enforce a
+patent against the party.
+
+  If you convey a covered work, knowingly relying on a patent license,
+and the Corresponding Source of the work is not available for anyone
+to copy, free of charge and under the terms of this License, through a
+publicly available network server or other readily accessible means,
+then you must either (1) cause the Corresponding Source to be so
+available, or (2) arrange to deprive yourself of the benefit of the
+patent license for this particular work, or (3) arrange, in a manner
+consistent with the requirements of this License, to extend the patent
+license to downstream recipients.  "Knowingly relying" means you have
+actual knowledge that, but for the patent license, your conveying the
+covered work in a country, or your recipient's use of the covered work
+in a country, would infringe one or more identifiable patents in that
+country that you have reason to believe are valid.
+
+  If, pursuant to or in connection with a single transaction or
+arrangement, you convey, or propagate by procuring conveyance of, a
+covered work, and grant a patent license to some of the parties
+receiving the covered work authorizing them to use, propagate, modify
+or convey a specific copy of the covered work, then the patent license
+you grant is automatically extended to all recipients of the covered
+work and works based on it.
+
+  A patent license is "discriminatory" if it does not include within
+the scope of its coverage, prohibits the exercise of, or is
+conditioned on the non-exercise of one or more of the rights that are
+specifically granted under this License.  You may not convey a covered
+work if you are a party to an arrangement with a third party that is
+in the business of distributing software, under which you make payment
+to the third party based on the extent of your activity of conveying
+the work, and under which the third party grants, to any of the
+parties who would receive the covered work from you, a discriminatory
+patent license (a) in connection with copies of the covered work
+conveyed by you (or copies made from those copies), or (b) primarily
+for and in connection with specific products or compilations that
+contain the covered work, unless you entered into that arrangement,
+or that patent license was granted, prior to 28 March 2007.
+
+  Nothing in this License shall be construed as excluding or limiting
+any implied license or other defenses to infringement that may
+otherwise be available to you under applicable patent law.
+
+  12. No Surrender of Others' Freedom.
+
+  If conditions are imposed on you (whether by court order, agreement or
+otherwise) that contradict the conditions of this License, they do not
+excuse you from the conditions of this License.  If you cannot convey a
+covered work so as to satisfy simultaneously your obligations under this
+License and any other pertinent obligations, then as a consequence you may
+not convey it at all.  For example, if you agree to terms that obligate you
+to collect a royalty for further conveying from those to whom you convey
+the Program, the only way you could satisfy both those terms and this
+License would be to refrain entirely from conveying the Program.
+
+  13. Use with the GNU Affero General Public License.
+
+  Notwithstanding any other provision of this License, you have
+permission to link or combine any covered work with a work licensed
+under version 3 of the GNU Affero General Public License into a single
+combined work, and to convey the resulting work.  The terms of this
+License will continue to apply to the part which is the covered work,
+but the special requirements of the GNU Affero General Public License,
+section 13, concerning interaction through a network will apply to the
+combination as such.
+
+  14. Revised Versions of this License.
+
+  The Free Software Foundation may publish revised and/or new versions of
+the GNU General Public License from time to time.  Such new versions will
+be similar in spirit to the present version, but may differ in detail to
+address new problems or concerns.
+
+  Each version is given a distinguishing version number.  If the
+Program specifies that a certain numbered version of the GNU General
+Public License "or any later version" applies to it, you have the
+option of following the terms and conditions either of that numbered
+version or of any later version published by the Free Software
+Foundation.  If the Program does not specify a version number of the
+GNU General Public License, you may choose any version ever published
+by the Free Software Foundation.
+
+  If the Program specifies that a proxy can decide which future
+versions of the GNU General Public License can be used, that proxy's
+public statement of acceptance of a version permanently authorizes you
+to choose that version for the Program.
+
+  Later license versions may give you additional or different
+permissions.  However, no additional obligations are imposed on any
+author or copyright holder as a result of your choosing to follow a
+later version.
+
+  15. Disclaimer of Warranty.
+
+  THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
+APPLICABLE LAW.  EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
+HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
+OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
+THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+PURPOSE.  THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
+IS WITH YOU.  SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
+ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
+
+  16. Limitation of Liability.
+
+  IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
+WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
+THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
+GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
+USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
+DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
+PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
+EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
+SUCH DAMAGES.
+
+  17. Interpretation of Sections 15 and 16.
+
+  If the disclaimer of warranty and limitation of liability provided
+above cannot be given local legal effect according to their terms,
+reviewing courts shall apply local law that most closely approximates
+an absolute waiver of all civil liability in connection with the
+Program, unless a warranty or assumption of liability accompanies a
+copy of the Program in return for a fee.
+
+                     END OF TERMS AND CONDITIONS
+
+            How to Apply These Terms to Your New Programs
+
+  If you develop a new program, and you want it to be of the greatest
+possible use to the public, the best way to achieve this is to make it
+free software which everyone can redistribute and change under these terms.
+
+  To do so, attach the following notices to the program.  It is safest
+to attach them to the start of each source file to most effectively
+state the exclusion of warranty; and each file should have at least
+the "copyright" line and a pointer to where the full notice is found.
+
+    <one line to give the program's name and a brief idea of what it does.>
+    Copyright (C) <year>  <name of author>
+
+    This program is free software: you can redistribute it and/or modify
+    it under the terms of the GNU General Public License as published by
+    the Free Software Foundation, either version 3 of the License, or
+    (at your option) any later version.
+
+    This program is distributed in the hope that it will be useful,
+    but WITHOUT ANY WARRANTY; without even the implied warranty of
+    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+    GNU General Public License for more details.
+
+    You should have received a copy of the GNU General Public License
+    along with this program.  If not, see <https://www.gnu.org/licenses/>.
+
+Also add information on how to contact you by electronic and paper mail.
+
+  If the program does terminal interaction, make it output a short
+notice like this when it starts in an interactive mode:
+
+    <program>  Copyright (C) <year>  <name of author>
+    This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
+    This is free software, and you are welcome to redistribute it
+    under certain conditions; type `show c' for details.
+
+The hypothetical commands `show w' and `show c' should show the appropriate
+parts of the General Public License.  Of course, your program's commands
+might be different; for a GUI interface, you would use an "about box".
+
+  You should also get your employer (if you work as a programmer) or school,
+if any, to sign a "copyright disclaimer" for the program, if necessary.
+For more information on this, and how to apply and follow the GNU GPL, see
+<https://www.gnu.org/licenses/>.
+
+  The GNU General Public License does not permit incorporating your program
+into proprietary programs.  If your program is a subroutine library, you
+may consider it more useful to permit linking proprietary applications with
+the library.  If this is what you want to do, use the GNU Lesser General
+Public License instead of this License.  But first, please read
+<https://www.gnu.org/licenses/why-not-lgpl.html>.

README.md

@@ -0,0 +1,33 @@
+This repository contains training scripts to train a text detector based on [manga-image-translator](https://github.com/zyddnys/manga-image-translator) which can extract bounding-boxes, text lines and segmentation of text from manga or comics to help further comics translation procedures such as text-removal, recognition, lettering, etc.  
+
+There are some awesome projects such as manga-image-translator, [manga_ocr](https://github.com/kha-white/manga_ocr), [SickZil-Machine](https://github.com/KUR-creative/SickZil-Machine) offer DL models to automize the remaining work, <s>we are working on a computer-aided comic/manga translation software which would (hopefully) put them together.</s>  see [BallonsTranslator](https://github.com/dmMaze/BallonsTranslator)[WIP]
+
+Download the text detection model from https://github.com/zyddnys/manga-image-translator/releases/tag/beta-0.2.1 or [Google Drive](https://drive.google.com/drive/folders/1cTsXP5NYTCjhPVxwScdhxqJleHuIOyXG?usp=sharing). 
+
+# Examples
+
+![AisazuNihaIrarenai-003](data/doc/AisazuNihaIrarenai-003.jpg)
+<sup>(source: [manga109](http://www.manga109.org/en/), © Yoshi Masako)</sup>
+
+![AisazuNihaIrarenai-003-mask](data/doc/AisazuNihaIrarenai-003-mask.png)
+
+![AisazuNihaIrarenai-003-bboxes](data/doc/AisazuNihaIrarenai-003-bboxes.jpg)
+
+# Training Details
+
+Our current model can be summarized as below.  
+
+<img src='data/doc/model.jpg'>  
+
+All models were trained on around 13 thousand anime & comic style images, 1/3 from Manga109-s, 1/3 from [DCM](https://digitalcomicmuseum.com/), and 1/3 are synthetic data in a weak supervision manner due to the lack of available high-quality annotations.   
+
+We used text detection model of manga-image-translator to generate text lines annotations for manga, and [Manga-Text-Segmentation](https://github.com/juvian/Manga-Text-Segmentation) with some post-processing to generate masks for both manga and comics. Synthetic data were generated using around 4k text-free anime-girls pictures from https://t.me/SugarPic, text-rendering, Unet and DBNet training scripts can be found in this repo.  Text block detector was trained using [yolov5 official repository](https://github.com/ultralytics/yolov5)  
+
+We would not (don't have the right) share training sets or fonts publicly, 2/3 of the training set is not so clean anyway, so the training is reproducible only if you have enough images and fonts, you can use the models this repo provided to generate labels for comics/manga, and the comic style text rendering script to generate synthetic data, please refer to [examples.ipynb](examples.ipynb) for more details. 
+
+## Acknowledgements
+
+* [https://github.com/zyddnys/manga-image-translator](https://github.com/zyddnys/manga-image-translator)
+* [https://github.com/juvian/Manga-Text-Segmentation](https://github.com/juvian/Manga-Text-Segmentation)
+* [https://github.com/ultralytics/yolov5](https://github.com/ultralytics/yolov5)
+* [https://github.com/WenmuZhou/DBNet.pytorch](https://github.com/WenmuZhou/DBNet.pytorch)

app.py

@@ -0,0 +1,107 @@
+from inference import model2annotations,traverse_by_dict, init_model
+import os
+import shutil
+import zipfile
+import datetime
+import gradio as gr
+import PIL.Image
+
+DESCRIPTION = "# [comic-text-detector](https://github.com/dmMaze/comic-text-detector)"
+INPUT_DIR = "./input"
+OUTPUT_DIR = "./output"
+TEMP_DIR = "./temp"
+
+os.makedirs(INPUT_DIR, exist_ok=True)
+os.makedirs(OUTPUT_DIR, exist_ok=True)
+os.makedirs(TEMP_DIR, exist_ok=True)
+
+def inference(model):
+    #model_path = f'{current_directory}/data/comictextdetector.pt'
+    
+    img_dir = './input'
+    save_dir = './output'
+    model2annotations(img_dir, save_dir, save_json=True, model=model)
+    return traverse_by_dict(img_dir, save_dir)
+current_directory = os.path.dirname(os.path.abspath(__file__))
+model_path = './data/comictextdetector.pt.onnx'
+model = init_model(model_path, device = 'cpu')
+
+def process_image_and_generate_zip(image_file: PIL.Image.Image) -> str:
+    if image_file is None:
+        return None, "请上传一张图片！"
+
+    # 1. 清空 ./input 文件夹
+    print(f"清空 {INPUT_DIR} 文件夹...")
+    for filename in os.listdir(INPUT_DIR):
+        file_path = os.path.join(INPUT_DIR, filename)
+        try:
+            if os.path.isfile(file_path) or os.path.islink(file_path):
+                os.unlink(file_path)
+            elif os.path.isdir(file_path):
+                shutil.rmtree(file_path)
+        except Exception as e:
+            print(f"无法删除 {file_path}. 原因: {e}")
+    print(f"{INPUT_DIR} 文件夹清空完成。")
+
+    # 2. 清空 ./output 文件夹 (通常在每次运行时也清空输出)
+    print(f"清空 {OUTPUT_DIR} 文件夹...")
+    for filename in os.listdir(OUTPUT_DIR):
+        file_path = os.path.join(OUTPUT_DIR, filename)
+        try:
+            if os.path.isfile(file_path) or os.path.islink(file_path):
+                os.unlink(file_path)
+            elif os.path.isdir(file_path):
+                shutil.rmtree(file_path)
+        except Exception as e:
+            print(f"无法删除 {file_path}. 原因: {e}")
+    print(f"{OUTPUT_DIR} 文件夹清空完成。")
+
+
+    # 3. 保存输入图片到 ./input 文件夹
+    input_image_path = os.path.join(INPUT_DIR, os.path.basename(image_file.name))
+    shutil.copyfile(image_file.name, input_image_path)
+    print(f"输入图片已保存到: {input_image_path}")
+
+    # 4. 调用模型
+    print("调用模型...")
+    inference(model)
+
+    # 5. 将 ./output 文件夹打包为zip文件
+    print("正在打包输出文件...")
+    timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
+    zip_filename = f"output_results_{timestamp}.zip"
+    zip_filepath = os.path.join(TEMP_DIR, zip_filename)
+
+    with zipfile.ZipFile(zip_filepath, 'w', zipfile.ZIP_DEFLATED) as zipf:
+        for root, _, files in os.walk(OUTPUT_DIR):
+            for file in files:
+                file_path = os.path.join(root, file)
+                # 计算在zip文件中的相对路径
+                arcname = os.path.relpath(file_path, OUTPUT_DIR)
+                zipf.write(file_path, arcname)
+    print(f"输出文件已打包到: {zip_filepath}")
+
+    return zip_filepath, "处理完成！请下载结果。"
+
+with gr.Blocks() as demo:
+    gr.Markdown("# 图像处理与结果下载")
+    gr.Markdown("上传一张图片，模型将对其进行处理，并将结果打包为ZIP文件供下载。")
+
+    with gr.Row():
+        image_input = gr.Image(type="filepath", label="上传图片")
+        # 直接显示处理后的图片，可选
+        # processed_image_output = gr.Image(label="处理后的图片")
+
+    run_button = gr.Button("运行模型并打包")
+    zip_output = gr.File(label="下载处理结果ZIP文件")
+    message_output = gr.Textbox(label="状态信息", interactive=False)
+
+    run_button.click(
+        fn=process_image_and_generate_zip,
+        inputs=image_input,
+        outputs=[zip_output, message_output] # 如果你想显示处理后的图片，这里可以添加 processed_image_output
+    )
+
+# 启动 Gradio 应用
+if __name__ == "__main__":
+    demo.launch()

basemodel.py

@@ -0,0 +1,273 @@
+
+from utils.general import CUDA, DEVICE
+from models.yolov5.yolo import Model
+import torch
+import cv2
+import numpy as np
+from models.yolov5.yolo import load_yolov5_ckpt
+from utils.yolov5_utils import fuse_conv_and_bn
+import glob
+import torch.nn as nn
+from utils.weight_init import init_weights
+from models.yolov5.common import C3, Conv
+from torchsummary import summary
+import torch.nn.functional as F
+import copy
+
+TEXTDET_MASK = 0
+TEXTDET_DET = 1
+TEXTDET_INFERENCE = 2
+
+class double_conv_up_c3(nn.Module):
+    def __init__(self, in_ch, mid_ch, out_ch, act=True):
+        super(double_conv_up_c3, self).__init__()
+        self.conv = nn.Sequential(
+        C3(in_ch+mid_ch, mid_ch, act=act),
+        nn.ConvTranspose2d(mid_ch, out_ch, kernel_size=4, stride = 2, padding=1, bias=False),
+        nn.BatchNorm2d(out_ch),
+        nn.ReLU(inplace=True),
+        )
+
+    def forward(self, x):
+        return self.conv(x)
+
+class double_conv_c3(nn.Module):
+    def __init__(self, in_ch, out_ch, stride=1, act=True):
+        super(double_conv_c3, self).__init__()
+        if stride > 1 :
+            self.down = nn.AvgPool2d(2,stride=2) if stride > 1 else None
+        self.conv = C3(in_ch, out_ch, act=act)
+
+    def forward(self, x):
+        if self.down is not None :
+            x = self.down(x)
+        x = self.conv(x)
+        return x
+
+class UnetHead(nn.Module):
+    def __init__(self, act=True) -> None:
+
+        super(UnetHead, self).__init__()
+        self.down_conv1 = double_conv_c3(512, 512, 2, act=act)
+        self.upconv0 = double_conv_up_c3(0, 512, 256, act=act)
+        self.upconv2 = double_conv_up_c3(256, 512, 256, act=act)
+        self.upconv3 = double_conv_up_c3(0, 512, 256, act=act)
+        self.upconv4 = double_conv_up_c3(128, 256, 128, act=act)
+        self.upconv5 = double_conv_up_c3(64, 128, 64, act=act)
+        self.upconv6 = nn.Sequential(
+            nn.ConvTranspose2d(64, 1, kernel_size=4, stride = 2, padding=1, bias=False),
+            nn.Sigmoid()
+        )
+
+    def forward(self, f160, f80, f40, f20, f3, forward_mode=TEXTDET_MASK):
+        # input: 640@3
+        d10 = self.down_conv1(f3) # 512@10
+        u20 = self.upconv0(d10)  # 256@10
+        u40 = self.upconv2(torch.cat([f20, u20], dim = 1)) # 256@40
+
+        if forward_mode == TEXTDET_DET:
+            return f80, f40, u40
+        else:
+            u80 = self.upconv3(torch.cat([f40, u40], dim = 1)) # 256@80
+            u160 = self.upconv4(torch.cat([f80, u80], dim = 1)) # 128@160
+            u320 = self.upconv5(torch.cat([f160, u160], dim = 1)) # 64@320
+            mask = self.upconv6(u320)
+            if forward_mode == TEXTDET_MASK:
+                return mask
+            else:
+                return mask, [f80, f40, u40]
+            
+    def init_weight(self, init_func):
+        self.apply(init_func)
+
+class DBHead(nn.Module):
+    def __init__(self, in_channels, k = 50, shrink_with_sigmoid=True, act=True):
+        super().__init__()
+        self.k = k
+        self.shrink_with_sigmoid = shrink_with_sigmoid
+        self.upconv3 = double_conv_up_c3(0, 512, 256, act=act)
+        self.upconv4 = double_conv_up_c3(128, 256, 128, act=act)
+        self.conv = nn.Sequential(
+            nn.Conv2d(128, in_channels, 1),
+            nn.BatchNorm2d(in_channels),
+            nn.ReLU(inplace=True)
+        )
+        self.binarize = nn.Sequential(
+            nn.Conv2d(in_channels, in_channels // 4, 3, padding=1),
+            nn.BatchNorm2d(in_channels // 4),
+            nn.ReLU(inplace=True),
+            nn.ConvTranspose2d(in_channels // 4, in_channels // 4, 2, 2),
+            nn.BatchNorm2d(in_channels // 4),
+            nn.ReLU(inplace=True),
+            nn.ConvTranspose2d(in_channels // 4, 1, 2, 2)
+            )
+        self.thresh = self._init_thresh(in_channels)
+
+    def forward(self, f80, f40, u40, shrink_with_sigmoid=True, step_eval=False):
+        shrink_with_sigmoid = self.shrink_with_sigmoid
+        u80 = self.upconv3(torch.cat([f40, u40], dim = 1)) # 256@80
+        x = self.upconv4(torch.cat([f80, u80], dim = 1)) # 128@160
+        x = self.conv(x)
+        threshold_maps = self.thresh(x)
+        x = self.binarize(x)
+        shrink_maps = torch.sigmoid(x)
+        
+        if self.training:
+            binary_maps = self.step_function(shrink_maps, threshold_maps)
+            if shrink_with_sigmoid:
+                return torch.cat((shrink_maps, threshold_maps, binary_maps), dim=1)
+            else:
+                return torch.cat((shrink_maps, threshold_maps, binary_maps, x), dim=1)
+        else:
+            if step_eval:
+                return self.step_function(shrink_maps, threshold_maps)
+            else:
+                return torch.cat((shrink_maps, threshold_maps), dim=1)
+
+    def init_weight(self, init_func):
+        self.apply(init_func)
+
+    def _init_thresh(self, inner_channels, serial=False, smooth=False, bias=False):
+        in_channels = inner_channels
+        if serial:
+            in_channels += 1
+        self.thresh = nn.Sequential(
+            nn.Conv2d(in_channels, inner_channels // 4, 3, padding=1, bias=bias),
+            nn.BatchNorm2d(inner_channels // 4),
+            nn.ReLU(inplace=True),
+            self._init_upsample(inner_channels // 4, inner_channels // 4, smooth=smooth, bias=bias),
+            nn.BatchNorm2d(inner_channels // 4),
+            nn.ReLU(inplace=True),
+            self._init_upsample(inner_channels // 4, 1, smooth=smooth, bias=bias),
+            nn.Sigmoid())
+        return self.thresh
+
+    def _init_upsample(self, in_channels, out_channels, smooth=False, bias=False):
+        if smooth:
+            inter_out_channels = out_channels
+            if out_channels == 1:
+                inter_out_channels = in_channels
+            module_list = [
+                nn.Upsample(scale_factor=2, mode='nearest'),
+                nn.Conv2d(in_channels, inter_out_channels, 3, 1, 1, bias=bias)]
+            if out_channels == 1:
+                module_list.append(nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, padding=1, bias=True))
+            return nn.Sequential(module_list)
+        else:
+            return nn.ConvTranspose2d(in_channels, out_channels, 2, 2)
+
+    def step_function(self, x, y):
+        return torch.reciprocal(1 + torch.exp(-self.k * (x - y)))
+
+class TextDetector(nn.Module):
+    def __init__(self, weights, map_location='cpu', forward_mode=TEXTDET_MASK, act=True):
+        super(TextDetector, self).__init__()
+
+        yolov5s_backbone = load_yolov5_ckpt(weights=weights, map_location=map_location)
+        yolov5s_backbone.eval()
+        out_indices = [1, 3, 5, 7, 9]
+        yolov5s_backbone.out_indices = out_indices
+        yolov5s_backbone.model = yolov5s_backbone.model[:max(out_indices)+1]
+        self.act = act
+        self.seg_net = UnetHead(act=act)
+        self.backbone = yolov5s_backbone
+        self.dbnet = None
+        self.forward_mode = forward_mode
+
+    def train_mask(self):
+        self.forward_mode = TEXTDET_MASK
+        self.backbone.eval()
+        self.seg_net.train()
+
+    def initialize_db(self, unet_weights):
+        self.dbnet = DBHead(64, act=self.act)
+        self.seg_net.load_state_dict(torch.load(unet_weights, map_location='cpu')['weights'])
+        self.dbnet.init_weight(init_weights)
+        self.dbnet.upconv3 = copy.deepcopy(self.seg_net.upconv3)
+        self.dbnet.upconv4 = copy.deepcopy(self.seg_net.upconv4)
+        del self.seg_net.upconv3
+        del self.seg_net.upconv4
+        del self.seg_net.upconv5
+        del self.seg_net.upconv6
+        # del self.seg_net.conv_mask
+    
+    def train_db(self):
+        self.forward_mode = TEXTDET_DET
+        self.backbone.eval()
+        self.seg_net.eval()
+        self.dbnet.train()
+
+    def forward(self, x):
+        forward_mode = self.forward_mode
+        with torch.no_grad():
+            outs = self.backbone(x)
+        if forward_mode == TEXTDET_MASK:
+            return self.seg_net(*outs, forward_mode=forward_mode)
+        elif forward_mode == TEXTDET_DET:
+            with torch.no_grad():
+                outs = self.seg_net(*outs, forward_mode=forward_mode)
+            return self.dbnet(*outs)
+
+def get_base_det_models(model_path, device='cpu', half=False, act='leaky'):
+    textdetector_dict = torch.load(model_path, map_location=device)
+    blk_det = load_yolov5_ckpt(textdetector_dict['blk_det'], map_location=device)
+    text_seg = UnetHead(act=act)
+    text_seg.load_state_dict(textdetector_dict['text_seg'])
+    text_det = DBHead(64, act=act)
+    text_det.load_state_dict(textdetector_dict['text_det'])
+    if half:
+        return blk_det.eval().half(), text_seg.eval().half(), text_det.eval().half()
+    return blk_det.eval().to(device), text_seg.eval().to(device), text_det.eval().to(device)
+
+class TextDetBase(nn.Module):
+    def __init__(self, model_path, device='cpu', half=False, fuse=False, act='leaky'):
+        super(TextDetBase, self).__init__()
+        self.blk_det, self.text_seg, self.text_det = get_base_det_models(model_path, device, half, act=act)
+        if fuse:
+            self.fuse()
+
+    def fuse(self):
+        def _fuse(model):
+            for m in model.modules():
+                if isinstance(m, (Conv)) and hasattr(m, 'bn'):
+                    m.conv = fuse_conv_and_bn(m.conv, m.bn)  # update conv
+                    delattr(m, 'bn')  # remove batchnorm
+                    m.forward = m.forward_fuse  # update forward
+            return model
+        self.text_seg = _fuse(self.text_seg)
+        self.text_det = _fuse(self.text_det)
+
+    def forward(self, features):
+        blks, features = self.blk_det(features, detect=True)
+        mask, features = self.text_seg(*features, forward_mode=TEXTDET_INFERENCE)
+        lines = self.text_det(*features, step_eval=False)
+        return blks[0], mask, lines
+
+class TextDetBaseDNN:
+    def __init__(self, input_size, model_path):
+        self.input_size = input_size
+        self.model = cv2.dnn.readNetFromONNX(model_path)
+        self.uoln = self.model.getUnconnectedOutLayersNames()
+    
+    def __call__(self, im_in):
+        blob = cv2.dnn.blobFromImage(im_in, scalefactor=1 / 255.0, size=(self.input_size, self.input_size))
+        self.model.setInput(blob)
+        blks, mask, lines_map  = self.model.forward(self.uoln)
+        return blks, mask, lines_map
+
+if __name__ == '__main__':
+    device = 'cuda'
+    weights = r'data/yolov5sblk.ckpt'
+
+    # yolov5s_backbone = load_yolov5_ckpt(weights=weights, map_location='cpu')
+
+    model = TextDetector(weights, map_location=DEVICE)
+    model.to(DEVICE)
+    model.train_mask()
+    summary(model, (3, 640, 640), device=DEVICE)
+
+    # model.initialize_db(unet_weights='data/unet_head.pt')
+    # model.train_db()
+    # summary(model, (3, 640, 640), device=DEVICE)
+
+

data/train_db_hyp.yaml

@@ -0,0 +1,51 @@
+data:
+  train_img_dir: 'dataset/train'
+  train_mask_dir: ''
+  val_img_dir: 'dataset/val'
+  val_mask_dir: ''
+  imgsz: 1024
+  augment: True
+  num_workers: 8
+  cache: True
+  aug_param:
+    hsv: 0.3
+    mini_mosaic: 0.7
+    flip_lr: 0.5
+    neg: 0.3
+    size_range: [0.85, 1.1]
+    rotate: 0.33
+    rotate_range: [-70, 70]
+  save_dir: 'results'
+
+train:
+  epochs: 160
+  linear_lr: False
+  optimizer: 'adam'
+  batch_size: 4
+  lr0: 0.01
+  lrf: 0.002
+  warm_up: True
+  momentum: 0.937
+  weight_decay: 0.00002
+  warmup_epochs: 3.0  # warmup epochs (fractions ok)
+  warmup_momentum: 0.8  # warmup initial momentum
+  warmup_bias_lr: 0.1  # warmup initial bias lr
+  eval_interval: 1
+  loss: 'bce'
+  accumulation_steps: 4
+
+model:
+  weights: 'data/yolov5sblk.ckpt'
+  unet_weights: 'data/unet_best.ckpt'
+  db_weights: ''
+  act: 'leaky'
+
+logger:
+  type: 'wandb'
+  run_id: ''
+  project: 'TextDetectDB'
+
+resume:
+  resume_training: False
+  ckpt: ''
+  

data/train_hyp.yaml

@@ -0,0 +1,45 @@
+data:
+  train_img_dir: 'dataset/train'
+  train_mask_dir: ''
+  val_img_dir: 'dataset/val'
+  val_mask_dir: ''
+  imgsz: 1024
+  augment: True
+  cache: True
+  aug_param:
+    hsv: 0.3
+    mini_mosaic: 0.5
+    flip_lr: 0.5
+    neg: 0.3
+    size_range: [0.7, 1]
+
+train:
+  epochs: 15
+  linear_lr: False
+  optimizer: 'adam'
+  batch_size: 4
+  lr0: 0.01
+  lrf: 0.005
+  momentum: 0.937
+  weight_decay: 0.0005
+  warmup_epochs: 3.0  # warmup epochs (fractions ok)
+  warmup_momentum: 0.8  # warmup initial momentum
+  warmup_bias_lr: 0.1  # warmup initial bias lr
+  eval_interval: 1
+  loss: 'dice'
+  accumulation_steps: 1
+
+model:
+  weights: 'data/yolov5sblk.ckpt'
+  act: 'leaky'
+
+logger:
+  type: 'wandb'
+  run_id: ''
+  project: ''
+  
+
+resume:
+  resume_training: False
+  ckpt: ''
+  

data/training_hyp.yaml

@@ -0,0 +1,44 @@
+data:
+  aug_param:
+    flip_lr: 0.5
+    hsv: 0.3
+    mini_mosaic: 0.5
+    neg: 0.3
+    size_range:
+    - 0.85
+    - 1.1
+  augment: true
+  cache: false
+  imgsz: 1024
+  train_img_dir:
+  - ../datasets/codat_manga_v3/images/train
+  - ../datasets/ComicErased/processed
+  train_mask_dir: ../datasets/ComicSegV2
+  val_img_dir:
+  - ../datasets/codat_manga_v3/images/val
+  val_mask_dir: ../datasets/ComicSegV2
+logger:
+  project: ''
+  run_id: ''
+  type: wandb
+model:
+  act: leaky
+  weights: data/yolov5sblk.ckpt
+resume:
+  ckpt: ''
+  resume_training: false
+train:
+  accumulation_steps: 4
+  batch_size: 4
+  epochs: 120
+  eval_interval: 1
+  linear_lr: false
+  loss: dice
+  lr0: 0.004
+  lrf: 0.005
+  momentum: 0.937
+  optimizer: adam
+  warmup_bias_lr: 0.1
+  warmup_epochs: 3.0
+  warmup_momentum: 0.8
+  weight_decay: 2.0e-05

db_dataset.py

@@ -0,0 +1,285 @@
+import numpy as np
+import yaml
+import torch
+import glob
+import os
+import os.path as osp
+import random
+from itertools import repeat
+from multiprocessing.pool import Pool, ThreadPool
+from pathlib import Path
+from threading import Thread
+import cv2
+from torch.utils.data import Dataset
+from tqdm import tqdm
+from pathlib import Path
+from torchvision import transforms
+from torch.utils.data import DataLoader, Dataset, dataloader
+from utils.general import LOGGER, Loggers, CUDA, DEVICE
+from utils.db_utils import MakeBorderMap, MakeShrinkMap
+from seg_dataset import augment_hsv
+from utils.imgproc_utils import rotate_polygons, letterbox, resize_keepasp
+from PIL import Image
+
+WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))  # DPP
+NUM_THREADS = min(8, max(1, os.cpu_count() - 1))  # number of multiprocessing threads
+IMG_EXT = ['.bmp', '.jpg', '.png', '.jpeg']
+
+def db_val_collate_fn(batchs):
+    cat_list = ['text_polys', 'ignore_tags']
+    ret_batchs = {}
+    for key in batchs[0].keys():
+        ret_batchs[key] = []
+        for batch in batchs:
+            if isinstance(batch[key], np.ndarray):
+                batch[key] = torch.from_numpy(batch[key])
+            ret_batchs[key].append(batch[key])
+        if key in cat_list:
+            pass
+        else:
+            ret_batchs[key] = torch.stack(ret_batchs[key], 0)
+    return ret_batchs
+
+class LoadImageAndAnnotations(Dataset):
+    def __init__(self, img_dir, ann_dir=None, img_size=640, augment=False, aug_param=None, cache=False, stride=128, cache_ann_only=True, with_ann=False):
+        if isinstance(img_dir, str):
+            self.img_dir = [img_dir]
+        elif isinstance(img_dir, list):
+            self.img_dir = img_dir
+        else:
+            raise Exception('unknown img_dir format')
+        
+        if ann_dir is None or ann_dir == '':
+            self.ann_dir = self.img_dir
+        else:
+            if isinstance(ann_dir, str):
+                self.ann_dir = [ann_dir]
+            elif isinstance(ann_dir, list):
+                self.ann_dir = ann_dir
+        self.with_ann = with_ann
+        self.make_border_map = MakeBorderMap(shrink_ratio=0.4)
+        self.make_shrink_map = MakeShrinkMap(shrink_ratio=0.4)
+        self.img_ann_list = []
+        self.img_size = (img_size, img_size)
+        self.stride = stride
+        self._augment = augment
+        if self._augment:
+            self._mini_mosaic = aug_param['mini_mosaic']
+            self._augment_hsv = aug_param['hsv']
+            self._flip_lr = aug_param['flip_lr']
+            self._neg = aug_param['neg']
+            self._rotate = aug_param['rotate']
+            self.rotate_range = aug_param['rotate_range']
+            size_range = aug_param['size_range'] 
+            if isinstance(size_range, list) and size_range[0] > 0:
+                min_size = round(img_size * size_range[0] / stride ) * stride
+                max_size = round(img_size * size_range[1] / stride ) * stride
+                self.valid_size = np.arange(min_size, max_size+1, stride)
+                self.multi_size = True
+            else:
+                self.valid_size = None
+                self.multi_size = False
+        for img_dir in self.img_dir:
+            for filep in glob.glob(osp.join(img_dir, "*")):
+                filename = osp.basename(filep)
+                file_suffix = Path(filename).suffix
+                if file_suffix not in IMG_EXT:
+                    continue
+                annname = 'line-' + filename.replace(file_suffix, '.txt')
+                for ann_dir in self.ann_dir:
+                    annp = osp.join(ann_dir, annname)
+                    if osp.exists(annp):
+                        self.img_ann_list.append((filep, annp))
+        self._img_transform = transforms.Compose([transforms.ToTensor()])
+
+        n = len(self.img_ann_list)
+        self.imgs, self.anns = [None] * n, [None] * n
+        gb = 0
+        if cache:
+            results = ThreadPool(NUM_THREADS).imap(lambda x: load_image_annotations(*x, max_size=img_size), zip(repeat(self), range(n)))
+            pbar = tqdm(enumerate(results), total=n)
+            for i, x in pbar:
+                im, self.anns[i] = x  # im, hw_orig, hw_resized = load_image_ann(self, i)
+                if not cache_ann_only:
+                    self.imgs[i] = im
+                    gb += self.imgs[i].nbytes
+                gb += self.anns[i].nbytes
+                if gb / 1E9 > 7:
+                    break
+                pbar.desc = f'Caching images ({gb / 1E9:.1f}GB )'
+            pbar.close()
+        
+    def initialize(self):
+        if self.augment:
+            if self.multi_size:
+                self.img_size = random.choice(self.valid_size)
+    
+    def transform(self, img):
+        cv2.cvtColor(img, cv2.COLOR_BGR2RGB, img)
+        img = img.astype(np.float32) / 255
+        img = self._img_transform(img)
+        return img
+
+    def mini_mosaic(self, img, ann):
+        im_h, im_w = img.shape[:2]
+        idx = random.randint(0, len(self)-1)
+        img2, ann2 = load_image_annotations(self, idx, self.img_size)
+        img2_h, img2_w = img2.shape[:2]
+        
+        if img2_h > img2_w:
+            imm_h = max(im_h, img2_h)
+            imm_w = im_w + img2_w
+            im_tmp = np.zeros((imm_h, imm_w, 3), np.uint8)
+            im_tmp[:im_h, :im_w] = img
+            im_tmp[:img2_h, im_w:] = img2
+            ann[:, :, 0] = ann[:, :, 0] * im_w / imm_w
+            ann[:, :, 1] = ann[:, :, 1] * im_h / imm_h
+            if ann2.shape[1] > 0:
+                ann2[:, :, 0] = ann2[:, :, 0] * img2_w / imm_w + im_w / imm_w
+                ann2[:, :, 1] = ann2[:, :, 1] * img2_h / imm_h
+                ann = np.concatenate((ann, ann2))
+            img = im_tmp
+            return img, ann
+                
+        else:
+            return img, ann
+
+    def augment(self, img, ann):
+        im_h, im_w = img.shape[0], img.shape[1]
+        if im_h > im_w and random.random() < self._mini_mosaic:
+            # imp2, annp2 = random.choice(self.img_ann_list)
+            img, ann = self.mini_mosaic(img, ann)
+
+        if random.random() < self._augment_hsv:
+            augment_hsv(img)
+        if random.random() < self._flip_lr:
+            cv2.flip(img, 1, img)
+            ann[:, :, 0] = 1 - ann[:, :, 0]
+        if random.random() < self._neg:
+            img = 255 - img
+        if random.random() < self._rotate:
+            degrees = random.uniform(self.rotate_range[0], self.rotate_range[1])
+            if abs(degrees) > 15:
+                img = Image.fromarray(img)
+                center = (img.width/2, img.height/2)
+                ann[:, :, 0] *= img.width
+                ann[:, :, 1] *= img.height
+                ann = ann.reshape(len(ann), -1)
+                img = img.rotate(degrees, resample=Image.BILINEAR, expand=1)
+                new_center = (img.width/2, img.height/2)
+                ann = rotate_polygons(center, ann, degrees, new_center, to_int=False)
+                ann = ann.reshape(len(ann), -1, 2)
+                ann[:, :, 0] /= img.width
+                ann[:, :, 1] /= img.height
+                img = np.asarray(img)
+        return img, ann
+
+    def inverse_transform(self, img: torch.Tensor, scale=255, to_uint8=True):
+        img = img.permute(1, 2, 0)
+        img = img * scale
+        img = img.cpu().numpy()
+        if to_uint8:
+            img = np.ascontiguousarray(img, np.uint8)
+        return img
+
+    def __len__(self):
+        return len(self.img_ann_list)
+
+    def __getitem__(self, idx):
+        img, ann = load_image_annotations(self, idx, self.img_size)
+        in_h, in_w = img.shape[:2]
+
+        if self._augment:
+            img, ann = self.augment(img, ann)
+        ignore_tags = [False] * ann.shape[0]
+
+        img, ratio, (dw, dh) = letterbox(img, new_shape=self.img_size, auto=False)
+        im_h, im_w = img.shape[:2]
+        if ann is not None:
+            ann[:, :, 0] *= (im_w - dw)
+            ann[:, :, 1] *= (im_h - dh)
+            ann = ann.astype(np.int64)
+        data_dict = {'imgs': img, 'text_polys': ann, 'ignore_tags': ignore_tags}
+
+        shrink_map = self.make_shrink_map(data_dict)
+        thresh_map = self.make_border_map(data_dict)
+        tp = thresh_map.pop('text_polys')
+        it = thresh_map.pop('ignore_tags')
+        if self.with_ann:
+            thresh_map['text_polys'] = torch.from_numpy(np.array(tp))
+            thresh_map['ignore_tags'] = torch.from_numpy(np.array(it))
+
+        thresh_map['imgs'] = self.transform(thresh_map['imgs'])
+        return thresh_map
+
+
+def load_image_annotations(self, i, max_size=None, ann_abs2rel=True):
+    # loads 1 image from dataset index 'i', returns im, original hw, resized hw
+    img, ann = self.imgs[i], self.anns[i]
+    imp, ann_path = self.img_ann_list[i]
+    if img is None:
+        img = cv2.imread(imp)
+    im_h, im_w = img.shape[:2]
+    if ann is None:
+        ann = np.loadtxt(ann_path)
+        if len(ann.shape) == 1:
+            ann = np.array([ann])
+        if ann_abs2rel:
+            ann[:, ::2] /= im_w
+            ann[:, 1::2] /= im_h
+        ann = ann.reshape(len(ann), -1, 2)
+    else:
+        ann = np.copy(ann)
+    if max_size is not None:
+        if isinstance(max_size, tuple):
+            max_size = max_size[0]
+        img = resize_keepasp(img, max_size)
+    return img, ann
+
+def create_dataloader(img_dir, ann_dir, imgsz, batch_size, augment=False, aug_param=None, cache=False, workers=8, shuffle=False, with_ann=False):
+    dataset = LoadImageAndAnnotations(img_dir, ann_dir, imgsz, augment, aug_param, cache, with_ann=with_ann)
+    batch_size = min(batch_size, len(dataset))
+    nw = min([os.cpu_count() // WORLD_SIZE, batch_size if batch_size > 1 else 0, workers])  # number of workers
+    if with_ann:
+        collate_fn = db_val_collate_fn
+    else:
+        collate_fn = None
+    loader = DataLoader(dataset, batch_size=batch_size, shuffle=shuffle, pin_memory=True, num_workers=nw, collate_fn=collate_fn)
+    return dataset, loader
+
+if __name__ == '__main__':
+    img_dir = 'data/dataset/db_sub'
+    hyp_p = r'data/train_db_hyp.yaml'
+    with open(hyp_p, 'r', encoding='utf8') as f:
+        hyp = yaml.safe_load(f.read())
+    hyp['data']['train_img_dir'] = img_dir
+    hyp['data']['cache'] = False
+    hyp_train, hyp_data, hyp_model, hyp_logger, hyp_resume = hyp['train'], hyp['data'], hyp['model'], hyp['logger'], hyp['resume']
+    batch_size = hyp_train['batch_size']
+    batch_size = 1
+    num_workers = 0
+    train_img_dir, train_mask_dir, imgsz, augment, aug_param = hyp_data['train_img_dir'], hyp_data['train_mask_dir'], hyp_data['imgsz'], hyp_data['augment'], hyp_data['aug_param']
+
+    train_dataset, train_loader = create_dataloader(train_img_dir, train_mask_dir, imgsz, batch_size, augment, aug_param, shuffle=True, workers=num_workers, cache=hyp_data['cache'], with_ann=True)
+
+    for ii in range(10):
+        
+        for batchs in train_loader:
+            train_dataset.initialize()
+            print(train_dataset.img_size)
+            img = batchs['imgs'][0]
+            
+            img = train_dataset.inverse_transform(img)
+            threshold_map = batchs['threshold_map'][0]
+            threshold_mask = batchs['threshold_mask'][0]
+            shrink_map = batchs['shrink_map'][0]
+            shrink_mask = batchs['shrink_mask'][0]
+            polys = batchs['text_polys'][0].numpy().astype(np.int32)
+            for p in polys:
+                cv2.polylines(img,[p],True,(255, 0, 0), thickness=2)
+            cv2.imshow('imgs', img)
+            cv2.imshow('threshold_map', threshold_map.numpy())
+            cv2.imshow('threshold_mask', threshold_mask.numpy())
+            cv2.imshow('shrink_map', shrink_map.numpy())
+            cv2.imshow('shrink_mask', shrink_mask.numpy())
+            cv2.waitKey(0) 

examples.ipynb

@@ -0,0 +1,185 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Annotate Comics/Manga\n",
+    "Download comictextdetector.pt and put it into data directory.\n",
+    "Run next block to generate following annotations for data\\examples\\AisazuNihaIrarenai-003.jpg:\n",
+    "- AisazuNihaIrarenai-003.txt: yolo format bounding boxes of english&japanese text block bounding boxes. 0 is eng.\n",
+    "- mask-AisazuNihaIrarenai-003.png\n",
+    "- line-AisazuNihaIrarenai-003.txt: icdar format bboxes of text lines."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "100%|██████████| 1/1 [00:04<00:00,  4.78s/it]\n"
+     ]
+    }
+   ],
+   "source": [
+    "from inference import model2annotations\n",
+    "\n",
+    "img_dir = r'data/examples'\n",
+    "model_path = r'data/comictextdetector.pt'\n",
+    "img_dir = r'data/examples'                              # can be dir list\n",
+    "save_dir = r'data/examples/annotations'\n",
+    "model2annotations(model_path, img_dir, save_dir, save_json=False)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Generate synthetic data\n",
+    "- current rendering script won't handle characters missing from fonts.\n",
+    "- Please use no-text images."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "100%|██████████| 10/10 [00:12<00:00,  1.23s/it]\n"
+     ]
+    }
+   ],
+   "source": [
+    "from text_rendering import ComicTextSampler, render_comictext, ALIGN_LEFT, ALIGN_CENTER\n",
+    "import copy\n",
+    "\n",
+    "ja_sampler_dict = {\n",
+    "                'num_txtblk': 20,\n",
+    "                'font': {\n",
+    "                        'font_dir': 'data/examples/fonts',   # font file directory\n",
+    "                        'font_statics': 'data/font_statics_en.csv',     # Just a font list file, please create your own list and ignore the last two cols.\n",
+    "                        'num': 1200,     # first 500 of the fontlist will be used \n",
+    "\n",
+    "                        # params to mimic comic/manga text style\n",
+    "                        'size': {'value': [0.02, 0.03, 0.15],\n",
+    "                                'prob': [1, 0.4, 0.15]},\n",
+    "                        'stroke_width': {'value': [0, 0.1, 0.15],\n",
+    "                                        'prob': [1, 0.5, 0.2]},\n",
+    "                        'color': {'value': ['black', 'white', 'random'],\n",
+    "                                'prob': [1, 1, 0.4]},\n",
+    "                },\n",
+    "                'text': {\n",
+    "                        'lang': 'ja',   # render japanese, 'en' for english\n",
+    "                        'orientation': {'value': [1, 0],    # 1 is vertical text.\n",
+    "                                                'prob': [1, 0.3]},\n",
+    "                        'rotation': {'value': [0, 30, 60],\n",
+    "                                                'prob': [1, 0.3, 0.1]},\n",
+    "                        'num_lines': {'value': [0.15],\n",
+    "                                'prob': [1]}, \n",
+    "                        'length': {'value': [0.3],\n",
+    "                                'prob': [1]},\n",
+    "                        'min_num_lines': 1,\n",
+    "                        'min_length': 3,\n",
+    "                        'alignment': {'value': [ALIGN_LEFT, ALIGN_CENTER],\n",
+    "                                'prob': [0.3, 1]}\n",
+    "                }\n",
+    "        }\n",
+    "\n",
+    "jp_cts = ComicTextSampler((845, 1280), ja_sampler_dict, seed=0)\n",
+    "eng_dict = copy.deepcopy(ja_sampler_dict)\n",
+    "eng_dict['text']['lang'] = 'en'\n",
+    "eng_dict['text']['orientation'] = {'value': [1, 0],\n",
+    "                                'prob': [0, 1]}\n",
+    "eng_cts = ComicTextSampler((845, 1280), eng_dict, seed=0)\n",
+    "\n",
+    "img_dir = r'data/examples'\n",
+    "save_dir = r'data/examples/annotations'\n",
+    " \n",
+    "render_comictext([eng_cts, jp_cts], img_dir, save_dir=save_dir, save_prefix=None, render_num=10, label_dir=None, show=False)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Training\n",
+    "### Train Text Block Detector\n",
+    "Train yolov5s using official repo of yolov5, assume the trained model is 'yolov5sblk.pt', go to the root directory of yolov5 and run following code.\n",
+    "\n",
+    "``` python\n",
+    "import torch\n",
+    "m = torch.load('yolov5sblk.pt')['model']\n",
+    "save_dict = {\n",
+    "    'cfg': m.yaml,\n",
+    "    'weights': m.state_dict()\n",
+    "}\n",
+    "torch.save(save_dict, 'yolov5sblk.ckpt')\n",
+    "```\n",
+    "### Train Text Segmentation Head\n",
+    "1. Put yolov5sblk.ckpt into data.   \n",
+    "2. Refer to train_seg.py for further details.  \n",
+    "\n",
+    "### Train DBHead\n",
+    "Please refer to train_db.py.\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Concat weights & export as onnx"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from utils.export import *\n",
+    "concate_models('data/yolov5sblk.ckpt', 'data/unet_best.ckpt', 'data/db_best.ckpt', 'data/textdetector.pt')\n",
+    "\n",
+    "batch_size, imgsz = 1, 1024\n",
+    "cuda = torch.cuda.is_available()\n",
+    "device = 'cpu'\n",
+    "im = torch.zeros(batch_size, 3, imgsz, imgsz).to(device)\n",
+    "model_path = r'data/textdetector.pt'\n",
+    "model = TextDetBase(model_path, device=device).to(device)\n",
+    "export_onnx(model, im, model_path, 11)"
+   ]
+  }
+ ],
+ "metadata": {
+  "interpreter": {
+   "hash": "545b34d9a5e72e2b90b819a16ec22002dd3dc9d66aaf1029c3177c6408a5603b"
+  },
+  "kernelspec": {
+   "display_name": "Python 3.9.7 64-bit",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.9.7"
+  },
+  "orig_nbformat": 4
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}

inference.py

@@ -0,0 +1,221 @@
+import json
+from basemodel import TextDetBase, TextDetBaseDNN
+import os.path as osp
+from tqdm import tqdm
+import numpy as np
+import cv2
+import torch
+from pathlib import Path
+import torch
+from utils.yolov5_utils import non_max_suppression
+from utils.db_utils import SegDetectorRepresenter
+from utils.io_utils import imread, imwrite, find_all_imgs, NumpyEncoder
+from utils.imgproc_utils import letterbox, xyxy2yolo, get_yololabel_strings
+from utils.textblock import TextBlock, group_output, visualize_textblocks
+from utils.textmask import refine_mask, refine_undetected_mask, REFINEMASK_INPAINT, REFINEMASK_ANNOTATION
+from pathlib import Path
+from typing import Union
+
+def init_model(model_path, device):
+    cuda = torch.cuda.is_available()
+    device = 'cuda' if cuda else 'cpu'
+    model = TextDetector(model_path=model_path, input_size=1024, device=device, act='leaky')  
+    return model
+
+def model2annotations(img_dir_list, save_dir, save_json=False, model=None):
+    if isinstance(img_dir_list, str):
+        img_dir_list = [img_dir_list]
+    # cuda = torch.cuda.is_available()
+    # device = 'cuda' if cuda else 'cpu'
+    # model = TextDetector(model_path=model_path, input_size=1024, device=device, act='leaky')  
+    imglist = []
+    for img_dir in img_dir_list:
+        imglist += find_all_imgs(img_dir, abs_path=True)
+    for img_path in tqdm(imglist):
+        imgname = osp.basename(img_path)
+        img = imread(img_path)
+        im_h, im_w = img.shape[:2]
+        imname = imgname.replace(Path(imgname).suffix, '')
+        maskname = 'mask-'+imname+'.png'
+        poly_save_path = osp.join(save_dir, 'line-' + imname + '.txt')
+        mask, mask_refined, blk_list = model(img, refine_mode=REFINEMASK_ANNOTATION, keep_undetected_mask=True)
+        polys = []
+        blk_xyxy = []
+        blk_dict_list = []
+        for blk in blk_list:
+            polys += blk.lines
+            blk_xyxy.append(blk.xyxy)
+            blk_dict_list.append(blk.to_dict())
+        blk_xyxy = xyxy2yolo(blk_xyxy, im_w, im_h)
+        if blk_xyxy is not None:
+            cls_list = [1] * len(blk_xyxy)
+            yolo_label = get_yololabel_strings(cls_list, blk_xyxy)
+        else:
+            yolo_label = ''
+        with open(osp.join(save_dir, imname+'.txt'), 'w', encoding='utf8') as f:
+            f.write(yolo_label)
+
+        # num_labels, labels, stats, centroids = cv2.connectedComponentsWithStats(mask)
+        # _, mask = cv2.threshold(mask, 50, 255, cv2.THRESH_BINARY)
+        # draw_connected_labels(num_labels, labels, stats, centroids)
+        # visualize_textblocks(img, blk_list)
+        # cv2.imshow('rst', img)
+        # cv2.imshow('mask', mask)
+        # cv2.imshow('mask_refined', mask_refined)
+        # cv2.waitKey(0)
+
+        if len(polys) != 0:
+            if isinstance(polys, list):
+                polys = np.array(polys)
+            polys = polys.reshape(-1, 8)
+            np.savetxt(poly_save_path, polys, fmt='%d')
+        if save_json:
+            with open(osp.join(save_dir, imname+'.json'), 'w', encoding='utf8') as f:
+                f.write(json.dumps(blk_dict_list, ensure_ascii=False, cls=NumpyEncoder))
+        imwrite(osp.join(save_dir, imgname), img)
+        imwrite(osp.join(save_dir, maskname), mask_refined)
+
+def preprocess_img(img, input_size=(1024, 1024), device='cpu', bgr2rgb=True, half=False, to_tensor=True):
+    if bgr2rgb:
+        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+    img_in, ratio, (dw, dh) = letterbox(img, new_shape=input_size, auto=False, stride=64)
+    if to_tensor:
+        img_in = img_in.transpose((2, 0, 1))[::-1]  # HWC to CHW, BGR to RGB
+        img_in = np.array([np.ascontiguousarray(img_in)]).astype(np.float32) / 255
+        if to_tensor:
+            img_in = torch.from_numpy(img_in).to(device)
+            if half:
+                img_in = img_in.half()
+    return img_in, ratio, int(dw), int(dh)
+
+def postprocess_mask(img: Union[torch.Tensor, np.ndarray], thresh=None):
+    # img = img.permute(1, 2, 0)
+    if isinstance(img, torch.Tensor):
+        img = img.squeeze_()
+        if img.device != 'cpu':
+            img = img.detach_().cpu()
+        img = img.numpy()
+    else:
+        img = img.squeeze()
+    if thresh is not None:
+        img = img > thresh
+    img = img * 255
+    # if isinstance(img, torch.Tensor):
+
+    return img.astype(np.uint8)
+
+def postprocess_yolo(det, conf_thresh, nms_thresh, resize_ratio, sort_func=None):
+    det = non_max_suppression(det, conf_thresh, nms_thresh)[0]
+    # bbox = det[..., 0:4]
+    if det.device != 'cpu':
+        det = det.detach_().cpu().numpy()
+    det[..., [0, 2]] = det[..., [0, 2]] * resize_ratio[0]
+    det[..., [1, 3]] = det[..., [1, 3]] * resize_ratio[1]
+    if sort_func is not None:
+        det = sort_func(det)
+
+    blines = det[..., 0:4].astype(np.int32)
+    confs = np.round(det[..., 4], 3)
+    cls = det[..., 5].astype(np.int32)
+    return blines, cls, confs
+
+class TextDetector:
+    lang_list = ['eng', 'ja', 'unknown']
+    langcls2idx = {'eng': 0, 'ja': 1, 'unknown': 2}
+
+    def __init__(self, model_path, input_size=1024, device='cpu', half=False, nms_thresh=0.35, conf_thresh=0.4, mask_thresh=0.3, act='leaky'):
+        super(TextDetector, self).__init__()
+        cuda = device == 'cuda'
+
+        if Path(model_path).suffix == '.onnx':
+            self.model = cv2.dnn.readNetFromONNX(model_path)
+            self.net = TextDetBaseDNN(input_size, model_path)
+            self.backend = 'opencv'
+        else:
+            self.net = TextDetBase(model_path, device=device, act=act)
+            self.backend = 'torch'
+        
+        if isinstance(input_size, int):
+            input_size = (input_size, input_size)
+        self.input_size = input_size
+        self.device = device
+        self.half = half
+        self.conf_thresh = conf_thresh
+        self.nms_thresh = nms_thresh
+        self.seg_rep = SegDetectorRepresenter(thresh=0.3)
+
+    @torch.no_grad()
+    def __call__(self, img, refine_mode=REFINEMASK_INPAINT, keep_undetected_mask=False):
+        img_in, ratio, dw, dh = preprocess_img(img, input_size=self.input_size, device=self.device, half=self.half, to_tensor=self.backend=='torch')
+        im_h, im_w = img.shape[:2]
+
+        blks, mask, lines_map = self.net(img_in)
+
+        resize_ratio = (im_w / (self.input_size[0] - dw), im_h / (self.input_size[1] - dh))
+        blks = postprocess_yolo(blks, self.conf_thresh, self.nms_thresh, resize_ratio)
+
+        if self.backend == 'opencv':
+            if mask.shape[1] == 2:     # some version of opencv spit out reversed result
+                tmp = mask
+                mask = lines_map
+                lines_map = tmp
+        mask = postprocess_mask(mask)
+
+        lines, scores = self.seg_rep(self.input_size, lines_map)
+        box_thresh = 0.6
+        idx = np.where(scores[0] > box_thresh)
+        lines, scores = lines[0][idx], scores[0][idx]
+        
+        # map output to input img
+        mask = mask[: mask.shape[0]-dh, : mask.shape[1]-dw]
+        mask = cv2.resize(mask, (im_w, im_h), interpolation=cv2.INTER_LINEAR)
+        if lines.size == 0 :
+            lines = []
+        else :
+            lines = lines.astype(np.float64)
+            lines[..., 0] *= resize_ratio[0]
+            lines[..., 1] *= resize_ratio[1]
+            lines = lines.astype(np.int32)
+        blk_list = group_output(blks, lines, im_w, im_h, mask)
+        mask_refined = refine_mask(img, mask, blk_list, refine_mode=refine_mode)
+        if keep_undetected_mask:
+            mask_refined = refine_undetected_mask(img, mask, mask_refined, blk_list, refine_mode=refine_mode)
+    
+        return mask, mask_refined, blk_list
+
+def traverse_by_dict(img_dir_list, dict_dir):
+    if isinstance(img_dir_list, str):
+        img_dir_list = [img_dir_list]
+    imglist = []
+    for img_dir in img_dir_list:
+        imglist += find_all_imgs(img_dir, abs_path=True)
+    for img_path in tqdm(imglist):
+        imgname = osp.basename(img_path)
+        imname = imgname.replace(Path(imgname).suffix, '')
+        mask_path = osp.join(dict_dir, 'mask-'+imname+'.png')
+        with open(osp.join(dict_dir, imname+'.json'), 'r', encoding='utf8') as f:
+            blk_dict_list = json.loads(f.read())
+            blk_list = [TextBlock(**blk_dict) for blk_dict in blk_dict_list]
+        img = cv2.imread(img_path)
+        mask = cv2.imread(mask_path, cv2.IMREAD_GRAYSCALE)
+        mask = refine_mask(img, mask, blk_list)
+        
+
+        visualize_textblocks(img, blk_list, path=dict_dir)
+        #cv2.imshow('im', img)
+        #cv2.imshow('mask', mask)
+        cv2.imwrite(f'{dict_dir}/labeled.png', img)
+        #cv2.imwrite('mask.png', mask)
+        #cv2.waitKey(0)
+        return len(blk_list)
+
+if __name__ == '__main__':
+    device = 'cpu'
+    
+    #model_path = 'data/comictextdetector.pt'
+    model_path = 'data/comictextdetector.pt.onnx'
+    
+    img_dir = r'../input'
+    save_dir = r'../output'
+    model2annotations(model_path, img_dir, save_dir, save_json=True)
+    traverse_by_dict(img_dir, save_dir)

models/yolov5/common.py

@@ -0,0 +1,289 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Common modules
+"""
+
+import json
+import math
+import platform
+import warnings
+from collections import OrderedDict, namedtuple
+from copy import copy
+from pathlib import Path
+
+import cv2
+import numpy as np
+import requests
+import torch
+import torch.nn as nn
+from PIL import Image
+from torch.cuda import amp
+
+from utils.yolov5_utils import make_divisible, initialize_weights, check_anchor_order, check_version, fuse_conv_and_bn
+
+def autopad(k, p=None):  # kernel, padding
+    # Pad to 'same'
+    if p is None:
+        p = k // 2 if isinstance(k, int) else [x // 2 for x in k]  # auto-pad
+    return p
+
+class Conv(nn.Module):
+    # Standard convolution
+    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True):  # ch_in, ch_out, kernel, stride, padding, groups
+        super().__init__()
+        self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g, bias=False)
+        self.bn = nn.BatchNorm2d(c2)
+        if isinstance(act, bool):
+            self.act = nn.SiLU() if act is True else (act if isinstance(act, nn.Module) else nn.Identity())
+        elif isinstance(act, str):
+            if act == 'leaky':
+                self.act = nn.LeakyReLU(0.1, inplace=True)
+            elif act == 'relu':
+                self.act = nn.ReLU(inplace=True)
+            else:
+                self.act = None
+    def forward(self, x):
+        return self.act(self.bn(self.conv(x)))
+
+    def forward_fuse(self, x):
+        return self.act(self.conv(x))
+
+
+class DWConv(Conv):
+    # Depth-wise convolution class
+    def __init__(self, c1, c2, k=1, s=1, act=True):  # ch_in, ch_out, kernel, stride, padding, groups
+        super().__init__(c1, c2, k, s, g=math.gcd(c1, c2), act=act)
+
+
+class TransformerLayer(nn.Module):
+    # Transformer layer https://arxiv.org/abs/2010.11929 (LayerNorm layers removed for better performance)
+    def __init__(self, c, num_heads):
+        super().__init__()
+        self.q = nn.Linear(c, c, bias=False)
+        self.k = nn.Linear(c, c, bias=False)
+        self.v = nn.Linear(c, c, bias=False)
+        self.ma = nn.MultiheadAttention(embed_dim=c, num_heads=num_heads)
+        self.fc1 = nn.Linear(c, c, bias=False)
+        self.fc2 = nn.Linear(c, c, bias=False)
+
+    def forward(self, x):
+        x = self.ma(self.q(x), self.k(x), self.v(x))[0] + x
+        x = self.fc2(self.fc1(x)) + x
+        return x
+
+
+class TransformerBlock(nn.Module):
+    # Vision Transformer https://arxiv.org/abs/2010.11929
+    def __init__(self, c1, c2, num_heads, num_layers):
+        super().__init__()
+        self.conv = None
+        if c1 != c2:
+            self.conv = Conv(c1, c2)
+        self.linear = nn.Linear(c2, c2)  # learnable position embedding
+        self.tr = nn.Sequential(*(TransformerLayer(c2, num_heads) for _ in range(num_layers)))
+        self.c2 = c2
+
+    def forward(self, x):
+        if self.conv is not None:
+            x = self.conv(x)
+        b, _, w, h = x.shape
+        p = x.flatten(2).permute(2, 0, 1)
+        return self.tr(p + self.linear(p)).permute(1, 2, 0).reshape(b, self.c2, w, h)
+
+
+class Bottleneck(nn.Module):
+    # Standard bottleneck
+    def __init__(self, c1, c2, shortcut=True, g=1, e=0.5, act=True):  # ch_in, ch_out, shortcut, groups, expansion
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1, act=act)
+        self.cv2 = Conv(c_, c2, 3, 1, g=g, act=act)
+        self.add = shortcut and c1 == c2
+
+    def forward(self, x):
+        return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
+
+
+class BottleneckCSP(nn.Module):
+    # CSP Bottleneck https://github.com/WongKinYiu/CrossStagePartialNetworks
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = nn.Conv2d(c1, c_, 1, 1, bias=False)
+        self.cv3 = nn.Conv2d(c_, c_, 1, 1, bias=False)
+        self.cv4 = Conv(2 * c_, c2, 1, 1)
+        self.bn = nn.BatchNorm2d(2 * c_)  # applied to cat(cv2, cv3)
+        self.act = nn.SiLU()
+        self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0) for _ in range(n)))
+
+    def forward(self, x):
+        y1 = self.cv3(self.m(self.cv1(x)))
+        y2 = self.cv2(x)
+        return self.cv4(self.act(self.bn(torch.cat((y1, y2), dim=1))))
+
+
+class C3(nn.Module):
+    # CSP Bottleneck with 3 convolutions
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5, act=True):  # ch_in, ch_out, number, shortcut, groups, expansion
+        super().__init__()
+        c_ = int(c2 * e)  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1, act=act)
+        self.cv2 = Conv(c1, c_, 1, 1, act=act)
+        self.cv3 = Conv(2 * c_, c2, 1, act=act)  # act=FReLU(c2)
+        self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, e=1.0, act=act) for _ in range(n)))
+        # self.m = nn.Sequential(*[CrossConv(c_, c_, 3, 1, g, 1.0, shortcut) for _ in range(n)])
+
+    def forward(self, x):
+        return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), dim=1))
+
+
+class C3TR(C3):
+    # C3 module with TransformerBlock()
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)
+        self.m = TransformerBlock(c_, c_, 4, n)
+
+
+class C3SPP(C3):
+    # C3 module with SPP()
+    def __init__(self, c1, c2, k=(5, 9, 13), n=1, shortcut=True, g=1, e=0.5):
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)
+        self.m = SPP(c_, c_, k)
+
+
+class C3Ghost(C3):
+    # C3 module with GhostBottleneck()
+    def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):
+        super().__init__(c1, c2, n, shortcut, g, e)
+        c_ = int(c2 * e)  # hidden channels
+        self.m = nn.Sequential(*(GhostBottleneck(c_, c_) for _ in range(n)))
+
+
+class SPP(nn.Module):
+    # Spatial Pyramid Pooling (SPP) layer https://arxiv.org/abs/1406.4729
+    def __init__(self, c1, c2, k=(5, 9, 13)):
+        super().__init__()
+        c_ = c1 // 2  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c_ * (len(k) + 1), c2, 1, 1)
+        self.m = nn.ModuleList([nn.MaxPool2d(kernel_size=x, stride=1, padding=x // 2) for x in k])
+
+    def forward(self, x):
+        x = self.cv1(x)
+        with warnings.catch_warnings():
+            warnings.simplefilter('ignore')  # suppress torch 1.9.0 max_pool2d() warning
+            return self.cv2(torch.cat([x] + [m(x) for m in self.m], 1))
+
+
+class SPPF(nn.Module):
+    # Spatial Pyramid Pooling - Fast (SPPF) layer for YOLOv5 by Glenn Jocher
+    def __init__(self, c1, c2, k=5):  # equivalent to SPP(k=(5, 9, 13))
+        super().__init__()
+        c_ = c1 // 2  # hidden channels
+        self.cv1 = Conv(c1, c_, 1, 1)
+        self.cv2 = Conv(c_ * 4, c2, 1, 1)
+        self.m = nn.MaxPool2d(kernel_size=k, stride=1, padding=k // 2)
+
+    def forward(self, x):
+        x = self.cv1(x)
+        with warnings.catch_warnings():
+            warnings.simplefilter('ignore')  # suppress torch 1.9.0 max_pool2d() warning
+            y1 = self.m(x)
+            y2 = self.m(y1)
+            return self.cv2(torch.cat([x, y1, y2, self.m(y2)], 1))
+
+
+class Focus(nn.Module):
+    # Focus wh information into c-space
+    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, act=True):  # ch_in, ch_out, kernel, stride, padding, groups
+        super().__init__()
+        self.conv = Conv(c1 * 4, c2, k, s, p, g, act)
+        # self.contract = Contract(gain=2)
+
+    def forward(self, x):  # x(b,c,w,h) -> y(b,4c,w/2,h/2)
+        return self.conv(torch.cat([x[..., ::2, ::2], x[..., 1::2, ::2], x[..., ::2, 1::2], x[..., 1::2, 1::2]], 1))
+        # return self.conv(self.contract(x))
+
+
+class GhostConv(nn.Module):
+    # Ghost Convolution https://github.com/huawei-noah/ghostnet
+    def __init__(self, c1, c2, k=1, s=1, g=1, act=True):  # ch_in, ch_out, kernel, stride, groups
+        super().__init__()
+        c_ = c2 // 2  # hidden channels
+        self.cv1 = Conv(c1, c_, k, s, None, g, act)
+        self.cv2 = Conv(c_, c_, 5, 1, None, c_, act)
+
+    def forward(self, x):
+        y = self.cv1(x)
+        return torch.cat([y, self.cv2(y)], 1)
+
+
+class GhostBottleneck(nn.Module):
+    # Ghost Bottleneck https://github.com/huawei-noah/ghostnet
+    def __init__(self, c1, c2, k=3, s=1):  # ch_in, ch_out, kernel, stride
+        super().__init__()
+        c_ = c2 // 2
+        self.conv = nn.Sequential(GhostConv(c1, c_, 1, 1),  # pw
+                                  DWConv(c_, c_, k, s, act=False) if s == 2 else nn.Identity(),  # dw
+                                  GhostConv(c_, c2, 1, 1, act=False))  # pw-linear
+        self.shortcut = nn.Sequential(DWConv(c1, c1, k, s, act=False),
+                                      Conv(c1, c2, 1, 1, act=False)) if s == 2 else nn.Identity()
+
+    def forward(self, x):
+        return self.conv(x) + self.shortcut(x)
+
+
+class Contract(nn.Module):
+    # Contract width-height into channels, i.e. x(1,64,80,80) to x(1,256,40,40)
+    def __init__(self, gain=2):
+        super().__init__()
+        self.gain = gain
+
+    def forward(self, x):
+        b, c, h, w = x.size()  # assert (h / s == 0) and (W / s == 0), 'Indivisible gain'
+        s = self.gain
+        x = x.view(b, c, h // s, s, w // s, s)  # x(1,64,40,2,40,2)
+        x = x.permute(0, 3, 5, 1, 2, 4).contiguous()  # x(1,2,2,64,40,40)
+        return x.view(b, c * s * s, h // s, w // s)  # x(1,256,40,40)
+
+
+class Expand(nn.Module):
+    # Expand channels into width-height, i.e. x(1,64,80,80) to x(1,16,160,160)
+    def __init__(self, gain=2):
+        super().__init__()
+        self.gain = gain
+
+    def forward(self, x):
+        b, c, h, w = x.size()  # assert C / s ** 2 == 0, 'Indivisible gain'
+        s = self.gain
+        x = x.view(b, s, s, c // s ** 2, h, w)  # x(1,2,2,16,80,80)
+        x = x.permute(0, 3, 4, 1, 5, 2).contiguous()  # x(1,16,80,2,80,2)
+        return x.view(b, c // s ** 2, h * s, w * s)  # x(1,16,160,160)
+
+
+class Concat(nn.Module):
+    # Concatenate a list of tensors along dimension
+    def __init__(self, dimension=1):
+        super().__init__()
+        self.d = dimension
+
+    def forward(self, x):
+        return torch.cat(x, self.d)
+
+
+class Classify(nn.Module):
+    # Classification head, i.e. x(b,c1,20,20) to x(b,c2)
+    def __init__(self, c1, c2, k=1, s=1, p=None, g=1):  # ch_in, ch_out, kernel, stride, padding, groups
+        super().__init__()
+        self.aap = nn.AdaptiveAvgPool2d(1)  # to x(b,c1,1,1)
+        self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p), groups=g)  # to x(b,c2,1,1)
+        self.flat = nn.Flatten()
+
+    def forward(self, x):
+        z = torch.cat([self.aap(y) for y in (x if isinstance(x, list) else [x])], 1)  # cat if list
+        return self.flat(self.conv(z))  # flatten to x(b,c2)
+
+

models/yolov5/yolo.py

@@ -0,0 +1,311 @@
+from operator import mod
+from cv2 import imshow
+from utils.yolov5_utils import scale_img
+from copy import deepcopy
+from .common import *
+
+class Detect(nn.Module):
+    stride = None  # strides computed during build
+    onnx_dynamic = False  # ONNX export parameter
+
+    def __init__(self, nc=80, anchors=(), ch=(), inplace=True):  # detection layer
+        super().__init__()
+        self.nc = nc  # number of classes
+        self.no = nc + 5  # number of outputs per anchor
+        self.nl = len(anchors)  # number of detection layers
+        self.na = len(anchors[0]) // 2  # number of anchors
+        self.grid = [torch.zeros(1)] * self.nl  # init grid
+        self.anchor_grid = [torch.zeros(1)] * self.nl  # init anchor grid
+        self.register_buffer('anchors', torch.tensor(anchors).float().view(self.nl, -1, 2))  # shape(nl,na,2)
+        self.m = nn.ModuleList(nn.Conv2d(x, self.no * self.na, 1) for x in ch)  # output conv
+        self.inplace = inplace  # use in-place ops (e.g. slice assignment)
+
+    def forward(self, x):
+        z = []  # inference output
+        for i in range(self.nl):
+            x[i] = self.m[i](x[i])  # conv
+            bs, _, ny, nx = x[i].shape  # x(bs,255,20,20) to x(bs,3,20,20,85)
+            x[i] = x[i].view(bs, self.na, self.no, ny, nx).permute(0, 1, 3, 4, 2).contiguous()
+
+            if not self.training:  # inference
+                if self.onnx_dynamic or self.grid[i].shape[2:4] != x[i].shape[2:4]:
+                    self.grid[i], self.anchor_grid[i] = self._make_grid(nx, ny, i)
+
+                y = x[i].sigmoid()
+                if self.inplace:
+                    y[..., 0:2] = (y[..., 0:2] * 2 - 0.5 + self.grid[i]) * self.stride[i]  # xy
+                    y[..., 2:4] = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh
+                else:  # for YOLOv5 on AWS Inferentia https://github.com/ultralytics/yolov5/pull/2953
+                    xy = (y[..., 0:2] * 2 - 0.5 + self.grid[i]) * self.stride[i]  # xy
+                    wh = (y[..., 2:4] * 2) ** 2 * self.anchor_grid[i]  # wh
+                    y = torch.cat((xy, wh, y[..., 4:]), -1)
+                z.append(y.view(bs, -1, self.no))
+
+        return x if self.training else (torch.cat(z, 1), x)
+
+    def _make_grid(self, nx=20, ny=20, i=0):
+        d = self.anchors[i].device
+        if check_version(torch.__version__, '1.10.0'):  # torch>=1.10.0 meshgrid workaround for torch>=0.7 compatibility
+            yv, xv = torch.meshgrid([torch.arange(ny, device=d), torch.arange(nx, device=d)], indexing='ij')
+        else:
+            yv, xv = torch.meshgrid([torch.arange(ny, device=d), torch.arange(nx, device=d)])
+        grid = torch.stack((xv, yv), 2).expand((1, self.na, ny, nx, 2)).float()
+        anchor_grid = (self.anchors[i].clone() * self.stride[i]) \
+            .view((1, self.na, 1, 1, 2)).expand((1, self.na, ny, nx, 2)).float()
+        return grid, anchor_grid
+
+class Model(nn.Module):
+    def __init__(self, cfg='yolov5s.yaml', ch=3, nc=None, anchors=None):  # model, input channels, number of classes
+        super().__init__()
+        self.out_indices = None
+        if isinstance(cfg, dict):
+            self.yaml = cfg  # model dict
+        else:  # is *.yaml
+            import yaml  # for torch hub
+            self.yaml_file = Path(cfg).name
+            with open(cfg, encoding='ascii', errors='ignore') as f:
+                self.yaml = yaml.safe_load(f)  # model dict
+
+        # Define model
+        ch = self.yaml['ch'] = self.yaml.get('ch', ch)  # input channels
+        if nc and nc != self.yaml['nc']:
+            # LOGGER.info(f"Overriding model.yaml nc={self.yaml['nc']} with nc={nc}")
+            self.yaml['nc'] = nc  # override yaml value
+        if anchors:
+            # LOGGER.info(f'Overriding model.yaml anchors with anchors={anchors}')
+            self.yaml['anchors'] = round(anchors)  # override yaml value
+        self.model, self.save = parse_model(deepcopy(self.yaml), ch=[ch])  # model, savelist
+        self.names = [str(i) for i in range(self.yaml['nc'])]  # default names
+        self.inplace = self.yaml.get('inplace', True)
+
+        # Build strides, anchors
+        m = self.model[-1]  # Detect()
+        # with torch.no_grad():
+        if isinstance(m, Detect):
+            s = 256  # 2x min stride
+            m.inplace = self.inplace
+            m.stride = torch.tensor([s / x.shape[-2] for x in self.forward(torch.zeros(1, ch, s, s))])  # forward
+            m.anchors /= m.stride.view(-1, 1, 1)
+            check_anchor_order(m)
+            self.stride = m.stride
+            self._initialize_biases()  # only run once
+
+        # Init weights, biases
+        initialize_weights(self)
+
+    def forward(self, x, augment=False, profile=False, visualize=False, detect=False):
+        if augment:
+            return self._forward_augment(x)  # augmented inference, None
+        return self._forward_once(x, profile, visualize, detect=detect)  # single-scale inference, train
+
+    def _forward_augment(self, x):
+        img_size = x.shape[-2:]  # height, width
+        s = [1, 0.83, 0.67]  # scales
+        f = [None, 3, None]  # flips (2-ud, 3-lr)
+        y = []  # outputs
+        for si, fi in zip(s, f):
+            xi = scale_img(x.flip(fi) if fi else x, si, gs=int(self.stride.max()))
+            yi = self._forward_once(xi)[0]  # forward
+            # cv2.imwrite(f'img_{si}.jpg', 255 * xi[0].cpu().numpy().transpose((1, 2, 0))[:, :, ::-1])  # save
+            yi = self._descale_pred(yi, fi, si, img_size)
+            y.append(yi)
+        y = self._clip_augmented(y)  # clip augmented tails
+        return torch.cat(y, 1), None  # augmented inference, train
+
+    def _forward_once(self, x, profile=False, visualize=False, detect=False):
+        y, dt = [], []  # outputs
+        z = []
+        for ii, m in enumerate(self.model):
+            if m.f != -1:  # if not from previous layer
+                x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f]  # from earlier layers
+            if profile:
+                self._profile_one_layer(m, x, dt)
+            x = m(x)  # run
+            y.append(x if m.i in self.save else None)  # save output
+            if self.out_indices is not None:
+                if m.i in self.out_indices:
+                    z.append(x)
+        if self.out_indices is not None:
+            if detect:
+                return x, z
+            else:
+                return z
+        else:
+            return x
+
+    def _descale_pred(self, p, flips, scale, img_size):
+        # de-scale predictions following augmented inference (inverse operation)
+        if self.inplace:
+            p[..., :4] /= scale  # de-scale
+            if flips == 2:
+                p[..., 1] = img_size[0] - p[..., 1]  # de-flip ud
+            elif flips == 3:
+                p[..., 0] = img_size[1] - p[..., 0]  # de-flip lr
+        else:
+            x, y, wh = p[..., 0:1] / scale, p[..., 1:2] / scale, p[..., 2:4] / scale  # de-scale
+            if flips == 2:
+                y = img_size[0] - y  # de-flip ud
+            elif flips == 3:
+                x = img_size[1] - x  # de-flip lr
+            p = torch.cat((x, y, wh, p[..., 4:]), -1)
+        return p
+
+    def _clip_augmented(self, y):
+        # Clip YOLOv5 augmented inference tails
+        nl = self.model[-1].nl  # number of detection layers (P3-P5)
+        g = sum(4 ** x for x in range(nl))  # grid points
+        e = 1  # exclude layer count
+        i = (y[0].shape[1] // g) * sum(4 ** x for x in range(e))  # indices
+        y[0] = y[0][:, :-i]  # large
+        i = (y[-1].shape[1] // g) * sum(4 ** (nl - 1 - x) for x in range(e))  # indices
+        y[-1] = y[-1][:, i:]  # small
+        return y
+
+    def _profile_one_layer(self, m, x, dt):
+        c = isinstance(m, Detect)  # is final layer, copy input as inplace fix
+        for _ in range(10):
+            m(x.copy() if c else x)
+
+
+    def _initialize_biases(self, cf=None):  # initialize biases into Detect(), cf is class frequency
+        # https://arxiv.org/abs/1708.02002 section 3.3
+        # cf = torch.bincount(torch.tensor(np.concatenate(dataset.labels, 0)[:, 0]).long(), minlength=nc) + 1.
+        m = self.model[-1]  # Detect() module
+        for mi, s in zip(m.m, m.stride):  # from
+            b = mi.bias.view(m.na, -1)  # conv.bias(255) to (3,85)
+            b.data[:, 4] += math.log(8 / (640 / s) ** 2)  # obj (8 objects per 640 image)
+            b.data[:, 5:] += math.log(0.6 / (m.nc - 0.999999)) if cf is None else torch.log(cf / cf.sum())  # cls
+            mi.bias = torch.nn.Parameter(b.view(-1), requires_grad=True)
+
+    def _print_biases(self):
+        m = self.model[-1]  # Detect() module
+        for mi in m.m:  # from
+            b = mi.bias.detach().view(m.na, -1).T  # conv.bias(255) to (3,85)
+
+    def fuse(self):  # fuse model Conv2d() + BatchNorm2d() layers
+        for m in self.model.modules():
+            if isinstance(m, (Conv, DWConv)) and hasattr(m, 'bn'):
+                m.conv = fuse_conv_and_bn(m.conv, m.bn)  # update conv
+                delattr(m, 'bn')  # remove batchnorm
+                m.forward = m.forward_fuse  # update forward
+        # self.info()
+        return self
+
+    # def info(self, verbose=False, img_size=640):  # print model information
+    #     model_info(self, verbose, img_size)
+
+    def _apply(self, fn):
+        # Apply to(), cpu(), cuda(), half() to model tensors that are not parameters or registered buffers
+        self = super()._apply(fn)
+        m = self.model[-1]  # Detect()
+        if isinstance(m, Detect):
+            m.stride = fn(m.stride)
+            m.grid = list(map(fn, m.grid))
+            if isinstance(m.anchor_grid, list):
+                m.anchor_grid = list(map(fn, m.anchor_grid))
+        return self
+
+def parse_model(d, ch):  # model_dict, input_channels(3)
+    # LOGGER.info(f"\n{'':>3}{'from':>18}{'n':>3}{'params':>10}  {'module':<40}{'arguments':<30}")
+    anchors, nc, gd, gw = d['anchors'], d['nc'], d['depth_multiple'], d['width_multiple']
+    na = (len(anchors[0]) // 2) if isinstance(anchors, list) else anchors  # number of anchors
+    no = na * (nc + 5)  # number of outputs = anchors * (classes + 5)
+
+    layers, save, c2 = [], [], ch[-1]  # layers, savelist, ch out
+    for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']):  # from, number, module, args
+        m = eval(m) if isinstance(m, str) else m  # eval strings
+        for j, a in enumerate(args):
+            try:
+                args[j] = eval(a) if isinstance(a, str) else a  # eval strings
+            except NameError:
+                pass
+
+        n = n_ = max(round(n * gd), 1) if n > 1 else n  # depth gain
+        if m in [Conv, GhostConv, Bottleneck, GhostBottleneck, SPP, SPPF, DWConv, Focus,
+                 BottleneckCSP, C3, C3TR, C3SPP, C3Ghost]:
+            c1, c2 = ch[f], args[0]
+            if c2 != no:  # if not output
+                c2 = make_divisible(c2 * gw, 8)
+
+            args = [c1, c2, *args[1:]]
+            if m in [BottleneckCSP, C3, C3TR, C3Ghost]:
+                args.insert(2, n)  # number of repeats
+                n = 1
+        elif m is nn.BatchNorm2d:
+            args = [ch[f]]
+        elif m is Concat:
+            c2 = sum(ch[x] for x in f)
+        elif m is Detect:
+            args.append([ch[x] for x in f])
+            if isinstance(args[1], int):  # number of anchors
+                args[1] = [list(range(args[1] * 2))] * len(f)
+        elif m is Contract:
+            c2 = ch[f] * args[0] ** 2
+        elif m is Expand:
+            c2 = ch[f] // args[0] ** 2
+        else:
+            c2 = ch[f]
+
+        m_ = nn.Sequential(*(m(*args) for _ in range(n))) if n > 1 else m(*args)  # module
+        t = str(m)[8:-2].replace('__main__.', '')  # module type
+        np = sum(x.numel() for x in m_.parameters())  # number params
+        m_.i, m_.f, m_.type, m_.np = i, f, t, np  # attach index, 'from' index, type, number params
+        # LOGGER.info(f'{i:>3}{str(f):>18}{n_:>3}{np:10.0f}  {t:<40}{str(args):<30}')  # print
+        save.extend(x % i for x in ([f] if isinstance(f, int) else f) if x != -1)  # append to savelist
+        layers.append(m_)
+        if i == 0:
+            ch = []
+        ch.append(c2)
+    return nn.Sequential(*layers), sorted(save)
+
+def load_yolov5(weights, map_location='cuda', fuse=True, inplace=True, out_indices=[1, 3, 5, 7, 9]):
+    if isinstance(weights, str):
+        ckpt = torch.load(weights, map_location=map_location)  # load
+    else:
+        ckpt = weights
+    
+    if fuse:
+        model = ckpt['model'].float().fuse().eval()  # FP32 model
+    else:
+        model = ckpt['model'].float().eval()  # without layer fuse
+
+    # Compatibility updates
+    for m in model.modules():
+        if type(m) in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU, Detect, Model]:
+            m.inplace = inplace  # pytorch 1.7.0 compatibility
+            if type(m) is Detect:
+                if not isinstance(m.anchor_grid, list):  # new Detect Layer compatibility
+                    delattr(m, 'anchor_grid')
+                    setattr(m, 'anchor_grid', [torch.zeros(1)] * m.nl)
+        elif type(m) is Conv:
+            m._non_persistent_buffers_set = set()  # pytorch 1.6.0 compatibility
+    model.out_indices = out_indices
+    return model
+
+@torch.no_grad()
+def load_yolov5_ckpt(weights, map_location='cpu', fuse=True, inplace=True, out_indices=[1, 3, 5, 7, 9]):
+    if isinstance(weights, str):
+        ckpt = torch.load(weights, map_location=map_location)  # load
+    else:
+        ckpt = weights
+    
+    model = Model(ckpt['cfg'])
+    model.load_state_dict(ckpt['weights'], strict=True)
+    
+    if fuse:
+        model = model.float().fuse().eval()  # FP32 model
+    else:
+        model = model.float().eval()  # without layer fuse
+
+    # Compatibility updates
+    for m in model.modules():
+        if type(m) in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU, Detect, Model]:
+            m.inplace = inplace  # pytorch 1.7.0 compatibility
+            if type(m) is Detect:
+                if not isinstance(m.anchor_grid, list):  # new Detect Layer compatibility
+                    delattr(m, 'anchor_grid')
+                    setattr(m, 'anchor_grid', [torch.zeros(1)] * m.nl)
+        elif type(m) is Conv:
+            m._non_persistent_buffers_set = set()  # pytorch 1.6.0 compatibility
+    model.out_indices = out_indices
+    return model

requirements.txt

@@ -0,0 +1,14 @@
+onnx>=1.9.0
+onnx-simplifier>=0.3.6
+opencv-python>=4.1.2
+Pillow>=7.1.2
+torch>=1.7.0
+torchvision>=0.8.1
+tqdm>=4.41.0
+torchsummary
+
+numpy
+wandb
+trdg
+gradio
+zipfile

seg_dataset.py

@@ -0,0 +1,253 @@
+import glob
+import os
+import os.path as osp
+import random
+from itertools import repeat
+from multiprocessing.pool import Pool, ThreadPool
+from pathlib import Path
+from threading import Thread
+from zipfile import ZipFile
+
+import cv2
+import numpy as np
+from numpy.lib.npyio import load
+from numpy.random import rand
+import torch
+import torch.nn.functional as F
+from torch.utils import data
+from torchvision.transforms.transforms import Compose
+
+from torch.utils.data import Dataset
+from tqdm import tqdm
+from pathlib import Path
+
+from tqdm import tqdm
+
+from torchvision import transforms
+import random
+from torch.utils.data import DataLoader, Dataset
+from utils.general import LOGGER, Loggers, CUDA, DEVICE
+from utils.imgproc_utils import resize_keepasp, letterbox
+from utils.io_utils import imread, imwrite
+
+WORLD_SIZE = int(os.getenv('WORLD_SIZE', 1))  # DPP
+NUM_THREADS = min(8, max(1, os.cpu_count() - 1))  # number of multiprocessing threads
+IMG_EXT = ['.bmp', '.jpg', '.png', '.jpeg']
+
+def augment_hsv(im, hgain=0.5, sgain=0.5, vgain=0.5):
+    # HSV color-space augmentation
+    if hgain or sgain or vgain:
+        r = np.random.uniform(-1, 1, 3) * [hgain, sgain, vgain] + 1  # random gains
+        hue, sat, val = cv2.split(cv2.cvtColor(im, cv2.COLOR_BGR2HSV))
+        dtype = im.dtype  # uint8
+
+        x = np.arange(0, 256, dtype=r.dtype)
+        lut_hue = ((x * r[0]) % 180).astype(dtype)
+        lut_sat = np.clip(x * r[1], 0, 255).astype(dtype)
+        lut_val = np.clip(x * r[2], 0, 255).astype(dtype)
+
+        im_hsv = cv2.merge((cv2.LUT(hue, lut_hue), cv2.LUT(sat, lut_sat), cv2.LUT(val, lut_val)))
+        cv2.cvtColor(im_hsv, cv2.COLOR_HSV2BGR, dst=im)  # no return needed
+
+def load_image_mask(self, i, max_size=None):
+    # loads 1 image from dataset index 'i', returns im, original hw, resized hw
+    img, mask = self.imgs[i], self.masks[i]
+    imp, maskp = self.img_mask_list[i]
+    if img is None:
+        img = cv2.imread(imp)
+    if mask is None:
+        mask = cv2.imread(maskp, cv2.IMREAD_GRAYSCALE)
+    if max_size is not None:
+        if isinstance(max_size, tuple):
+            max_size = max_size[0]
+        try:
+            img = resize_keepasp(img, max_size)
+            mask = resize_keepasp(mask, max_size, interpolation=cv2.INTER_AREA)
+        except:
+            pass
+    return img, mask
+
+def mini_mosaic(self, img, mask):
+    im_h, im_w = img.shape[0], img.shape[1]
+    idx = random.randint(0, len(self)-1)
+    img2, mask2 = load_image_mask(self, idx, self.img_size)
+    img2_h, img2_w = img2.shape[0], img2.shape[1]
+    ratio = img2_h / im_h
+    if img2_h > img2_w and ratio > 0.4 and ratio < 1.6:
+        im_h = max(im_h, img2_h)
+        im_w = im_w + img2_w
+        im_tmp = np.zeros((im_h, im_w, 3), np.uint8)
+        im_tmp[:img.shape[0], :img.shape[1]] = img
+        im_tmp[:img2_h, img.shape[1]:] = img2
+        mask_tmp = np.zeros((im_h, im_w), np.uint8)
+        mask_tmp[:img.shape[0], :img.shape[1]] = mask
+        mask_tmp[:img2_h, img.shape[1]:] = mask2
+
+        img = np.ascontiguousarray(im_tmp)
+        mask = np.ascontiguousarray(mask_tmp)
+    return img, mask
+
+class LoadImageAndMask(Dataset):
+    def __init__(self, img_dir, mask_dir=None, img_size=640, augment=False, aug_param=None, cache=False, stride=128, cache_mask_only=True):
+        if isinstance(img_dir, str):
+            self.img_dir = [img_dir]
+        elif isinstance(img_dir, list):
+            self.img_dir = img_dir
+        else:
+            raise Exception('unknown img_dir format')
+        
+        if mask_dir is None or mask_dir == '':
+            self.mask_dir = self.img_dir
+        else:
+            if isinstance(mask_dir, str):
+                self.mask_dir = [mask_dir]
+            elif isinstance(mask_dir, list):
+                self.mask_dir = mask_dir
+        
+        self.img_mask_list = []
+        self.img_size = (img_size, img_size)
+        self.stride = stride
+        self._augment = augment
+        if self._augment:
+            self._mini_mosaic = aug_param['mini_mosaic']
+            self._augment_hsv = aug_param['hsv']
+            self._flip_lr = aug_param['flip_lr']
+            self._neg = aug_param['neg']
+            size_range = aug_param['size_range'] 
+            if size_range[0] != -1:
+                min_size = round(img_size * size_range[0] / stride ) * stride
+                max_size = round(img_size * size_range[1] / stride ) * stride
+                self.valid_size = np.arange(min_size, max_size+1, stride)
+                self.multi_size = True
+            else:
+                self.valid_size = None
+                self.multi_size = False
+        for img_dir in self.img_dir:
+            for filep in glob.glob(osp.join(img_dir, "*")):
+                filename = osp.basename(filep)
+                file_suffix = Path(filename).suffix
+                if file_suffix.lower() not in IMG_EXT:
+                    continue
+                maskname = 'mask-' + filename.replace(file_suffix, '.png')
+                for mask_dir in self.mask_dir:
+                    maskp = osp.join(mask_dir, maskname)
+                    if osp.exists(maskp):
+                        self.img_mask_list.append((filep, maskp))
+        self._img_transform = transforms.Compose([transforms.ToTensor()])
+        self._mask_transform = transforms.Compose([transforms.ToTensor()])
+
+        n = len(self.img_mask_list)
+        self.imgs, self.masks = [None] * n, [None] * n
+        gb = 0
+        if cache:
+            results = ThreadPool(NUM_THREADS).imap(lambda x: load_image_mask(*x, max_size=img_size), zip(repeat(self), range(n)))
+            pbar = tqdm(enumerate(results), total=n)
+            for i, x in pbar:
+                im, self.masks[i] = x  # im, hw_orig, hw_resized = load_image_mask(self, i)
+                if not cache_mask_only:
+                    self.imgs[i] = im
+                    gb += self.imgs[i].nbytes
+                gb += self.masks[i].nbytes
+                if gb / 1E9 > 7:
+                    break
+                pbar.desc = f'Caching images ({gb / 1E9:.1f}GB )'
+            pbar.close()
+        
+    def initialize(self):
+        if self.augment:
+            if self.multi_size:
+                self.img_size = random.choice(self.valid_size)
+    
+    def transform(self, img, mask):
+        cv2.cvtColor(img, cv2.COLOR_BGR2RGB, img)
+        img = img.astype(np.float32) / 255
+        mask = (mask > 30).astype(np.float32)
+        # mask = mask / 255
+        img = self._img_transform(img)
+        mask = self._mask_transform(mask)
+        return img, mask
+
+    def augment(self, img, mask):
+        im_h, im_w = img.shape[0], img.shape[1]
+        if im_h > im_w and random.random() < self._mini_mosaic:
+            # imp2, maskp2 = random.choice(self.img_mask_list)
+            img, mask = mini_mosaic(self, img, mask)
+
+        img, ratio, (dw, dh) = letterbox(img, new_shape=self.img_size, auto=False)
+        mask, ratio, (dw, dh) = letterbox(mask, new_shape=self.img_size, auto=False)
+
+        if random.random() < self._augment_hsv:
+            augment_hsv(img)
+        if random.random() < self._flip_lr:
+            cv2.flip(img, 1, img)
+            cv2.flip(mask, 1, mask)
+        if random.random() < self._neg:
+            img = 255 - img
+        return img, mask
+
+    def inverse_transform(self, img: torch.Tensor):
+        img = img.permute(1, 2, 0)
+        img = img * 255
+        img = img.cpu().numpy().astype(np.uint8)
+        return img
+
+    def __len__(self):
+        return len(self.img_mask_list)
+
+    def __getitem__(self, idx):
+        img, mask = load_image_mask(self, idx, self.img_size)
+        if self._augment:
+            img, mask = self.augment(img, mask)
+        else:
+            img, ratio, (dw, dh) = letterbox(img, new_shape=self.img_size, auto=False)
+            mask, ratio, (dw, dh) = letterbox(mask, new_shape=self.img_size, auto=False)
+        return self.transform(img, mask)
+
+def create_dataloader(img_dir, mask_dir, imgsz, batch_size, augment=False, aug_param=None, cache=False, workers=8, shuffle=False):
+    dataset = LoadImageAndMask(img_dir, mask_dir, imgsz, augment, aug_param, cache)
+    batch_size = min(batch_size, len(dataset))
+    nw = min([os.cpu_count() // WORLD_SIZE, batch_size if batch_size > 1 else 0, workers])  # number of workers
+    loader = DataLoader(dataset, batch_size=batch_size, shuffle=shuffle, pin_memory=True, num_workers=nw)
+    return dataset, loader
+
+if __name__ == '__main__':
+    random.seed(42)
+    torch.random.manual_seed(42)
+    np.random.seed(42)
+    import yaml
+    hyp_p = r'data/train_hyp.yaml'
+    with open(hyp_p, 'r', encoding='utf8') as f:
+        hyp = yaml.safe_load(f.read())
+    hyp['data']['train_img_dir'] = [r'D:/neonbub/datasets/codat_manga_v3/images/train', r'D:/neonbub/datasets/ComicErased/processed']
+    hyp['data']['val_img_dir'] = [r'D:/neonbub/datasets/codat_manga_v3/images/val']
+    hyp['data']['train_mask_dir'] = r'D:/neonbub/datasets/ComicSegV2'
+    hyp['data']['val_mask_dir'] = r'D:/neonbub/datasets/ComicSegV2'
+    hyp['data']['cache'] = False
+
+    hyp_train, hyp_data, hyp_model, hyp_logger, hyp_resume = hyp['train'], hyp['data'], hyp['model'], hyp['logger'], hyp['resume']
+    
+    batch_size = hyp_train['batch_size']
+    batch_size = 4
+    num_workers = 2
+
+    train_img_dir, train_mask_dir, imgsz, augment, aug_param = hyp_data['train_img_dir'], hyp_data['train_mask_dir'], hyp_data['imgsz'], hyp_data['augment'], hyp_data['aug_param']
+    val_img_dir, val_mask_dir = hyp_data['val_img_dir'], hyp_data['val_mask_dir']
+    train_dataset, train_loader = create_dataloader(train_img_dir, train_mask_dir, imgsz, batch_size, augment, aug_param, shuffle=True, workers=num_workers, cache=hyp_data['cache'])
+    val_dataset, val_loader = create_dataloader(val_img_dir, val_mask_dir, imgsz, batch_size, augment=False, shuffle=False, workers=num_workers, cache=hyp_data['cache'])
+    LOGGER.info(f'num training imgs: {len(train_dataset)}, num val imgs: {len(val_dataset)}')
+    
+    for epoch in range(0, 4):  # epoch ------------------------------------------------------------------
+        train_dataset.initialize()
+        pbar = enumerate(train_loader)
+        pbar = tqdm(pbar, total=len(train_loader), bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}')  # progress bar
+        pbar.set_description(f' training size: {train_dataset.img_size}')
+        for i, (imgs, masks) in pbar:
+            img, mask = imgs[0], masks[0]
+            imgs = imgs
+            masks = masks
+            img = train_dataset.inverse_transform(img)
+            mask = train_dataset.inverse_transform(mask)
+            cv2.imshow('img', img)
+            cv2.imshow('mask', mask)
+            cv2.waitKey(0)
+        pbar.close()

text_rendering.py

@@ -0,0 +1,545 @@
+from copy import copy
+from http.client import IM_USED
+import pathlib
+import shutil
+import PIL
+import cv2
+
+import numpy as np
+import os.path as osp
+import os
+from PIL import Image, ImageColor, ImageFont, ImageDraw, ImageFilter, ImageOps
+import random
+
+from numpy.random import rand
+from trdg.utils import load_dict, load_fonts
+from tqdm import tqdm
+import pandas as pd
+import sys
+sys.path.append(os.getcwd())
+from utils.io_utils import find_all_imgs, imread, imwrite
+from utils.imgproc_utils import *
+import copy
+
+ALIGN_LEFT = 0
+ALIGN_CENTER = 1
+ALIGN_RIGHT = 2
+
+ORIENTATION_HOR = 0
+ORIENTATION_VER = 1
+
+def get_textlines_from_langdict(lang_dict, num_line, line_len, sampler=None):
+    textlines = []
+    dict_len = len(lang_dict)
+    for ii in range(num_line):
+        line = ''
+        for jj in range(line_len):
+            line += lang_dict[random.randrange(dict_len)] + ' '
+        textlines.append(line[:line_len])
+    if sampler is None:
+        return textlines
+    return textlines
+
+def draw_text_polygons(img, text_polygons, color=None):
+    if isinstance(img, PIL.Image.Image):
+        img = np.array(img)
+    img = np.copy(img)
+    for poly in text_polygons:
+        if color is None:
+            randcolor = (random.randint(0,255), random.randint(0,255), random.randint(0,255))
+        else:
+            randcolor = color
+        cv2.polylines(img,[poly.reshape((-1, 1, 2))],True,randcolor, thickness=2)
+    return img
+
+def draw_textblk(textlines, font, 
+                     fill='black',
+                     stroke_width=0,
+                     stroke_fill='grey',
+                     spacing=0,
+                     rotation=0,
+                     orientation=ORIENTATION_HOR,
+                     alignment=ALIGN_LEFT):
+
+    text_size = np.array([font.getsize(line) for line in textlines])
+    if orientation == ORIENTATION_HOR:
+        line_widths, line_heights = text_size[:, 0], text_size[:, 1]
+        textblk_w = max(text_size[:, 0]) + 3*stroke_width 
+        textblk_h = (len(textlines) - 1) * spacing + text_size[:, 1].sum() + 3*stroke_width
+    else:
+        line_widths, line_heights = text_size[:, 1], text_size[:, 0]
+        textblk_w = line_widths.sum() + 3*stroke_width 
+        textblk_h = max(line_heights) + 3*stroke_width
+    if orientation == ORIENTATION_VER:
+        textblk_h += font.size * 3  # some fonts are not correctly aligned
+    
+    txtblk_img = Image.new("RGBA", (textblk_w, textblk_h), (255, 255, 255, 255))
+    txtblk_draw = ImageDraw.Draw(txtblk_img)
+    txtblk_draw.fontmode = '1'      # disable anti-aliasing
+    txtblk_mask = Image.new("L", (textblk_w, textblk_h), (0))
+    tmp_msk = txtblk_mask.copy()
+    tmp_msk_draw = ImageDraw.Draw(tmp_msk)
+    tmp_msk_draw.fontmode = '1'
+
+    textpolygons = []
+    if orientation == ORIENTATION_VER:     
+        for ii, line in enumerate(textlines):
+            x_offset = sum(line_widths[:ii]) + stroke_width
+            for jj, char in enumerate(line):
+                txtblk_draw.text((x_offset, jj*font.size), char, font=font, fill=fill, stroke_width=stroke_width, stroke_fill=stroke_fill)
+                tmp_msk_draw.text((x_offset, jj*font.size), char, font=font, fill='white', stroke_width=stroke_width, stroke_fill='white')
+            valid_bbox = tmp_msk.getbbox()
+            if valid_bbox is None:
+                continue
+            txtblk_mask.paste(tmp_msk, mask=tmp_msk)
+            tmp_msk.paste('black', [0, 0, tmp_msk.size[0],tmp_msk.size[1]])
+            textpolygons.append([valid_bbox[0], valid_bbox[1], valid_bbox[2]-valid_bbox[0], valid_bbox[3]-valid_bbox[1]])
+    else:
+        for ii, line in enumerate(textlines):
+            x_offset = stroke_width
+            y_offset = sum(line_heights[0:ii]) + stroke_width
+            if alignment == ALIGN_CENTER:
+                x_offset += (textblk_w - line_widths[ii]) / 2
+            txtblk_draw.text((x_offset, y_offset), line, font=font, fill=fill, stroke_width=stroke_width, stroke_fill=stroke_fill)
+            tmp_msk_draw.text((x_offset, y_offset), line, font=font, fill='white', stroke_width=stroke_width, stroke_fill='white')
+            valid_bbox = tmp_msk.getbbox()
+            if valid_bbox is None:
+                continue
+            txtblk_mask.paste(tmp_msk, mask=tmp_msk)
+            tmp_msk.paste('black', [0, 0, tmp_msk.size[0],tmp_msk.size[1]])
+            textpolygons.append([valid_bbox[0], valid_bbox[1], valid_bbox[2]-valid_bbox[0], valid_bbox[3]-valid_bbox[1]])
+    bbox = txtblk_mask.getbbox()
+    if bbox is None:
+        return None, None, None
+    textpolygons = np.array(textpolygons)
+    textpolygons = xywh2xyxypoly(textpolygons)
+    txtblk_img, txtblk_mask = txtblk_img.crop(bbox), txtblk_mask.crop(bbox)
+    textpolygons[:, ::2] = np.clip(textpolygons[:, ::2] - bbox[0], 0, txtblk_mask.width-1)
+    textpolygons[:, 1::2] = np.clip(textpolygons[:, 1::2] - bbox[1], 0, txtblk_mask.height-1)
+    if rotation != 0:
+        center = (txtblk_img.width/2, txtblk_img.height/2)
+        txtblk_img = txtblk_img.rotate(rotation, Image.BICUBIC, expand=1)
+        txtblk_mask = txtblk_mask.rotate(rotation, Image.BICUBIC, expand=1)
+        new_center = (txtblk_img.width / 2, txtblk_img.height / 2)
+        textpolygons = rotate_polygons(center, textpolygons, rotation, new_center)
+    # txtblk_img, txtblk_mask = txtblk_img.crop(bbox), txtblk_mask.crop(bbox)
+    # textpolygons[:, ::2] = np.clip(textpolygons[:, ::2] - bbox[0], 0, txtblk_mask.width-1)
+    # textpolygons[:, 1::2] = np.clip(textpolygons[:, 1::2] - bbox[1], 0, txtblk_mask.height-1)
+    return txtblk_img, txtblk_mask, textpolygons
+
+def create_random_sampler(value, prob):
+    if isinstance(prob, list):
+        prob = np.array(prob).astype(np.float32)
+    prob /= prob.sum()
+    sampler = lambda : np.random.choice(value, replace=False, p=prob)
+    return sampler
+
+class ScaledSampler:
+    def __init__(self, func_args, func='default'):
+        if func == 'default':
+            self.sampler_func = create_random_sampler(**func_args)
+        else:
+            raise NotImplementedError()
+        pass
+    def __call__(self, scaler=None, to_int=True):
+        value = self.sampler_func()
+        if scaler is not None:
+            value = scaler * value
+            if to_int:
+                value = int(round(value))
+        return value
+        pass
+
+class RandColorSampler:
+    def __init__(self, func_args, func='default'):
+        if func == 'default':
+            self.sampler_func = create_random_sampler(**func_args)
+        else:
+            raise NotImplementedError()
+        pass
+    def __call__(self, scaler=None):
+        value = self.sampler_func()
+        if value == 'random':
+            return (random.randint(0,255), random.randint(0,255), random.randint(0,255), 255)
+        return value
+
+class TextLinesSampler:
+    def __init__(self, page_size, sampler_dict):
+        self.page_w, self.page_h = page_size
+        self.lang = sampler_dict['lang']
+        self.lang_dict = load_dict(lang=self.lang)
+        self.orientation_sampler = ScaledSampler(sampler_dict['orientation'])
+        self.numlines_sampler = ScaledSampler(sampler_dict['num_lines'])
+        self.length_sampler = ScaledSampler(sampler_dict['length'])
+        self.min_num_lines = sampler_dict['min_num_lines']
+        self.min_length = sampler_dict['min_length']
+        self.alignment_sampler = create_random_sampler(**sampler_dict['alignment'])
+        self.rotation_sampler = create_random_sampler(**sampler_dict['rotation'])
+        
+    def __call__(self, page_w=None, page_h=None, font_size=1):
+        if page_w == None:
+            page_w = self.page_w
+        if page_h == None:
+            page_h = self.page_h
+        orientation = self.orientation_sampler()
+        rotation = self.rotation_sampler()
+        if rotation != 0:
+            rotation = random.randint(-rotation, rotation)
+        num_lines = max(self.numlines_sampler(page_h/font_size), self.min_num_lines)
+        num_lines = random.randint(self.min_num_lines, num_lines)
+        max_length = max(self.length_sampler(page_h/font_size), self.min_length)
+
+        textlines = []
+        dict_len = len(self.lang_dict)
+        for ii in range(num_lines):
+            line = ''
+            length = random.randint(self.min_length, max_length)
+            for jj in range(length):
+                line += self.lang_dict[random.randrange(dict_len)] + ' '
+            textlines.append(line[:length])
+        return textlines, orientation, self.alignment_sampler(), rotation
+
+class FontSampler:
+    def __init__(self, font_dict, page_size) -> None:
+        font_statics = font_dict['font_statics']
+        font_dir = font_dict['font_dir']
+        self.page_size = page_size
+
+        self.size_sampler = ScaledSampler(font_dict['size'])
+        self.color_sampler = RandColorSampler(font_dict['color'])
+        self.sw_sampler = ScaledSampler(font_dict['stroke_width'])
+        
+        self.font_dir = font_dir
+        self.sampler_range = font_dict['num']
+        self.font_idx = 0
+
+        font_statics = pd.read_csv(font_statics)
+        self.font_list = list()
+        for fontname in font_statics['font']:
+            if osp.exists(osp.join(self.font_dir, fontname)):
+                self.font_list.append(fontname)
+                if len(self.font_list) >= self.sampler_range:
+                    break
+        assert len(self.font_list) > 0
+
+    def __call__(self, page_size = None):
+        if page_size is None:
+            page_size = self.page_size
+        page_w, page_h = page_size
+        fontsize = self.size_sampler(page_h)
+        stroke_width = self.sw_sampler(fontsize)
+        color = self.color_sampler()
+        if color == 'black':
+            sw_color = (255, 255, 255, 255)
+        elif color == 'white':
+            sw_color = (0, 0, 0, 255)
+        else:
+            sw_color = self.color_sampler()
+        # while (True):
+        #     self.font_idx = random.randrange(0, self.sampler_range)
+        #     fontname = self.font_statics.iloc[self.font_idx]['font']
+        #     font_path = osp.join(self.font_dir, fontname)
+        #     if osp.exists(font_path):
+        #         break
+        self.font_idx = random.randrange(0, self.sampler_range) % len(self.font_list)
+        font_path = osp.join(self.font_dir, self.font_list[self.font_idx])
+        font = ImageFont.truetype(font_path, fontsize)
+        
+        return font, color, stroke_width, sw_color
+
+
+class TextBlkSampler:
+    def __init__(self, page_size, max_tries, bboxlist=[]):
+        self.page_w, self.page_h = page_size
+        self.bboxlist = bboxlist
+        self.max_tries = max_tries
+        self.max_padding = int(round(0.05 * self.page_h))
+
+    def __call__(self, bbox_w, bbox_h, padding=0, page_size=None):
+        padding = int(round(padding))
+        if page_size is not None:
+            page_w, page_h = page_size
+        else:
+            page_w, page_h = self.page_w, self.page_h
+        padding = min(self.max_padding, padding)
+        bbox_w += 2*padding
+        bbox_h += 2*padding
+        x_range = page_w-bbox_w-1
+        y_range = page_h-bbox_h-1
+        if x_range < 0 or y_range < 0:
+            return None
+        for ii in range(self.max_tries):
+            x, y = random.randint(0, x_range), random.randint(0, y_range)
+            bbox_padded = [x, y, x + bbox_w, y + bbox_h]
+            collide = False
+            for bbox_exist in self.bboxlist:
+                if union_area(bbox_exist, bbox_padded) > 0:
+                    collide = True
+                    break
+            if not collide:
+                break
+        if not collide:
+            bbox = [bbox_padded[0]+padding, bbox_padded[1]+padding, bbox_padded[2]-padding, bbox_padded[3]-padding]
+            # bbox = [int(bb) for bb in bbox]
+            self.bboxlist.append(bbox)
+            return bbox
+        return None
+
+    def initialize(self, page_w, page_h, bboxlist=None, to_xywh=False):
+        if bboxlist is None:
+            self.bboxlist = []
+        else:
+            if to_xywh:
+                self.bboxlist = yolo_xywh2xyxy(bboxlist, page_w, page_h)
+                if self.bboxlist is not None:
+                    self.bboxlist = self.bboxlist.tolist()
+                else:
+                    self.bboxlist = []
+                
+
+LANG_DICT = {'en': 0, 'ja': 1}
+def lang2cls(lang: str) -> int:
+    return LANG_DICT[lang]
+def cls2lang(cls: int) -> str:
+    return list(LANG_DICT.keys())[cls]
+
+def get_max_var_color(mean_bgcolor):
+    color_candidate = np.clip(np.array([mean_bgcolor-127, mean_bgcolor+127]), 0, 255).astype(np.int64)
+    max_var_color = [c[0] if abs(c[0]-mean_bgcolor[ii]) > abs(c[1]-mean_bgcolor[ii]) else c[1] for ii, c in enumerate(zip(color_candidate[0], color_candidate[1]))]
+    max_var_color = (max_var_color[0], max_var_color[1], max_var_color[2])
+    return max_var_color
+
+
+class ComicTextSampler:
+    def __init__(self, page_size, sampler_dict, seed=None):
+        if seed is not None:
+            random.seed(seed)
+            np.random.seed(seed)
+        self.page_size = page_size
+        self.num_txtblk = sampler_dict['num_txtblk']
+        self.font_dict = sampler_dict['font']
+        self.text_dict = sampler_dict['text']
+        
+        self.textlines_sampler = TextLinesSampler(page_size, sampler_dict['text'])
+        self.font_sampler = FontSampler(self.font_dict, self.page_size)
+        self.textblk_sampler = TextBlkSampler(page_size, max_tries=20)
+
+        self.lang = sampler_dict['text']['lang']
+
+    def drawtext_one_page(self, page_size=None, bboxlist=None, im_in=None, adaptive_color=False):
+        if page_size is not None:
+            page_w, page_h = page_size
+        else:
+            page_w, page_h = self.page_size
+        if im_in is None:
+            canvas = Image.new("RGBA", (page_w, page_h), 'white')
+        else:
+            canvas = Image.fromarray(cv2.cvtColor(im_in, cv2.COLOR_BGR2RGB))
+            page_w, page_h = canvas.width, canvas.height
+        canvas_msk = Image.new("L", (page_w, page_h), 'black')
+        canvas_draw = ImageDraw.Draw(canvas)
+        block_dicts = {}
+        yolo_labels = []
+        textpolylines = []
+        self.textblk_sampler.initialize(page_w, page_h, bboxlist, True)
+        for ii in range(self.num_txtblk):
+            font, color, stroke_width, sw_color = self.font_sampler(page_size=self.page_size)
+            textlines, orientation, alignment, rotation = self.textlines_sampler(font_size=font.size)
+            txtblk_img, txtblk_mask, textpolygons = draw_textblk(textlines, font, fill=color, stroke_width=stroke_width, stroke_fill=sw_color, orientation=orientation, alignment=alignment, rotation=rotation)
+            if txtblk_mask is None:
+                continue
+            bbox = self.textblk_sampler(txtblk_img.width, txtblk_img.height, font.size*1.2, page_size=(page_w, page_h))
+            if bbox is not None:
+                x1, y1, x2, y2 = bbox[0], bbox[1], bbox[0] + txtblk_mask.width, bbox[1] + txtblk_mask.height
+                re_draw = False
+                if im_in is not None:
+                    mean_bgcolor = np.mean(im_in[y1: y2, x1: x2], axis=(0, 1))
+                    max_var_color = get_max_var_color(mean_bgcolor)
+                    # color_candidate = np.clip(np.array([mean_bgcolor-127, mean_bgcolor+127]), 0, 255).astype(np.int64)
+                    # max_var_color = [c[0] if abs(c[0]-mean_bgcolor[ii]) > abs(c[1]-mean_bgcolor[ii]) else c[1] for ii, c in enumerate(zip(color_candidate[0], color_candidate[1]))]
+                    # max_var_color = (max_var_color[0], max_var_color[1], max_var_color[2])
+                    if color == 'black':
+                        color_rep = np.array([0, 0, 0])
+                    elif color == 'white':
+                        color_rep = np.array([255, 255, 255])
+                    else:
+                        color_rep = np.array(color[:3])
+                    color_var = np.sum(np.abs(mean_bgcolor - color_rep))
+                    if not adaptive_color:
+                        if color_var < 127:
+                            color = max_var_color
+
+                            sw_color = get_max_var_color(np.array(color))
+                            re_draw = True
+                    else:
+                        color = max_var_color
+                        sw_color = get_max_var_color(np.array(color))
+                        re_draw = True
+                if stroke_width != 0 and im_in is not None:
+                    # sw_color = get_max_var_color(color)
+                    re_draw = True
+                if re_draw:
+                    txtblk_img, txtblk_mask, textpolygons = draw_textblk(textlines, font, fill=color, stroke_width=stroke_width, stroke_fill=sw_color, orientation=orientation, alignment=alignment, rotation=rotation)
+                blk_dict = {
+                    'lang': self.lang, 
+                    'lang_cls': lang2cls(self.lang),
+                    'xyxy': [x1, y1, x2, y2],
+                    'polylines': textpolygons
+                }
+                block_dicts[str(ii)+'-'+self.lang] = blk_dict
+                textpolygons[:, ::2] += x1
+                textpolygons[:, 1::2] += y1
+                textpolylines += textpolygons.astype(np.int64).tolist()
+                yolo_labels += [[x1, y1, x2, y2]]
+                canvas.paste(txtblk_img, (bbox[0], bbox[1]), mask=txtblk_mask)
+                canvas_msk.paste(txtblk_mask, (bbox[0], bbox[1]), mask=txtblk_mask)
+
+        rst = cv2.cvtColor(np.array(canvas), cv2.COLOR_RGB2BGR)
+        rst_msk = np.array(canvas_msk)
+        yolo_labels = xyxy2yolo(np.array(yolo_labels), page_w, page_h)
+        if yolo_labels is not None:
+            cls = np.ones((yolo_labels.shape[0], 1)) * lang2cls(self.lang)
+            yolo_labels = np.concatenate((cls, yolo_labels), axis=1)
+        return rst, rst_msk, block_dicts, yolo_labels, np.array(textpolylines)
+
+def render_comictext(comic_sampler_list, img_dir, label_dir=None, render_num=700, save_dir=None, save_prefix=None, show=False):
+    if osp.exists(osp.join(img_dir, 'statistics.csv')):
+        statistics = pd.read_csv(osp.join(img_dir, 'statistics.csv'))
+    else:
+        statistics = None
+        imglist = find_all_imgs(img_dir)
+        # render_num = min(render_num, len(imglist))
+    num_im = len(imglist)
+    for ii in tqdm(range(render_num)):
+        im_idx = ii % num_im
+        if statistics is not None:
+            imgname = statistics.loc[im_idx]['name']
+        else:
+            imgname = imglist[im_idx]
+        img = imread(osp.join(img_dir, imgname))
+        cs_idx = ii % len(comic_sampler_list)
+        bboxlist = []
+        labels = None
+        if label_dir is not None:
+            labelname = imgname.replace(pathlib.Path(imgname).suffix, '.txt')
+            label_path = osp.join(label_dir, labelname)
+            labels = np.loadtxt(label_path)
+            if len(labels) != 0:
+                if len(labels.shape) == 1:
+                    labels = np.array([labels])
+                clslist, bboxlist = labels[:, 0], np.copy(labels[:, 1:])
+            else:
+                labels = None
+                bboxlist = []
+        rst, rst_msk, block_dicts, yolo_labels, textpolylines = comic_sampler_list[cs_idx].drawtext_one_page(im_in=img, bboxlist=bboxlist, adaptive_color=True)
+        if save_dir is not None:
+            if save_prefix is not None:
+                save_name = save_prefix + '{0:09d}'.format(ii) + '.jpg'
+            else:
+                save_name = 'syn-' + imgname
+            yolo_save_path = osp.join(save_dir, save_name.replace(pathlib.Path(save_name).suffix, '.txt'))
+            content = ''
+            if yolo_labels is not None:
+                if labels is None:
+                    content = get_yololabel_strings(yolo_labels[:, 0], yolo_labels[:, 1:])
+                else:
+                    yolo_labels = np.concatenate((labels, yolo_labels))
+                    content = get_yololabel_strings(yolo_labels[:, 0], yolo_labels[:, 1:])
+            if content == '' and label_dir is not None:
+                shutil.copy(label_path, yolo_save_path)
+            else:
+                with open(yolo_save_path, 'w', encoding='utf8') as f:
+                    f.write(content)
+                
+            linepoly_save_path = osp.join(save_dir, 'line-'+osp.basename(yolo_save_path))
+            np.savetxt(linepoly_save_path, textpolylines, fmt='%d')
+            imwrite(osp.join(save_dir, save_name), rst, ext='.jpg')
+            imwrite(osp.join(save_dir, 'mask-'+save_name), rst_msk)
+
+        if show:
+            for pts in textpolylines:
+                rst = cv2.polylines(rst, [np.array(pts).reshape((-1, 1, 2))], color=(255, 0, 0), isClosed=True, thickness=2)
+            cv2.imshow('rst', rst)
+            cv2.waitKey(0)
+
+
+if __name__ == '__main__':
+
+    eng_sampler_dict = {
+                    'num_txtblk': 20,
+                    'font': {
+                            'font_dir': 'data/fonts',
+                            'font_statics': 'data/font_statics_en.csv',
+                            'num': 500,
+                            'size': {'value': [0.02, 0.03, 0.15],
+                                    'prob': [1, 0.4, 0.15]},
+                            'stroke_width': {'value': [0, 0.1, 0.15],
+                                            'prob': [1, 0.2, 0.2]},
+                            'color': {'value': ['black', 'random'],
+                                        'prob': [1, 0.4]},
+                    },
+                    'text': {
+                        'lang': 'en',
+                        'orientation': {'value': [1, 0],
+                                            'prob': [0, 1]},
+                        'rotation': {'value': [0, 30, 60],
+                                            'prob': [1, 0.3, 0.1]},
+                        'num_lines': {'value': [0.15],
+                                'prob': [1]}, 
+                        'length': {'value': [1],
+                                'prob': [1]},
+                        'min_num_lines': 1,
+                        'min_length': 3,
+                        'alignment': {'value': [ALIGN_LEFT, ALIGN_CENTER],
+                                'prob': [0.3, 1]}
+                    }
+                }
+
+    ja_sampler_dict = {
+                    'num_txtblk': 20,
+                    'font': {
+                            'font_dir': 'data/fonts',   # font file directory
+                            'font_statics': 'data/font_statics_jp.csv',     # Just a font list to use, please create your own list and ignore the last two cols.
+                            'num': 500,     # first 500 of the fontlist will be used 
+                            # params to 
+                            'size': {'value': [0.02, 0.03, 0.15],
+                                    'prob': [1, 0.4, 0.15]},
+                            'stroke_width': {'value': [0, 0.1, 0.15],
+                                            'prob': [1, 0.5, 0.2]},
+                            'color': {'value': ['black', 'white', 'random'],
+                                        'prob': [1, 1, 0.4]},
+                    },
+                    'text': {
+                        'lang': 'ja',   # render japanese, 'en' for english
+                        'orientation': {'value': [1, 0],    # 1 is vertical text.
+                                            'prob': [1, 0.3]},
+                        'rotation': {'value': [0, 30, 60],
+                                            'prob': [1, 0.3, 0.1]},
+                        'num_lines': {'value': [0.15],
+                                'prob': [1]}, 
+                        'length': {'value': [0.3],
+                                'prob': [1]},
+                        'min_num_lines': 1,
+                        'min_length': 3,
+                        'alignment': {'value': [ALIGN_LEFT, ALIGN_CENTER],
+                                'prob': [0.3, 1]}
+                    }
+                }
+
+
+
+    # random.seed(0)
+    # cts = ComicTextSampler((845, 1280), sampler_dict, seed=0)
+    # jp_cts = ComicTextSampler((845, 1280), ja_sampler_dict, seed=0)
+    
+    # img_dir = r'../../datasets/pixanimegirls'
+    # save_dir = r'../../datasets/pixanimegirls/processed'
+    # os.makedirs(save_dir, exist_ok=True)
+
+    # img_dir = r'../../datasets/ComicErased'
+    # label_dir = img_dir
+    # save_dir = r'../../datasets/ComicErased/processed'
+    # os.makedirs(save_dir, exist_ok=True)
+    # render_comictext([jp_cts, cts], img_dir, save_dir=save_dir, save_prefix=None, render_num=4000, label_dir=None)
+
+

train_db.py

@@ -0,0 +1,226 @@
+from torch.autograd.grad_mode import F
+from torch.nn.functional import sigmoid
+from torch.nn.modules.loss import CrossEntropyLoss
+from torch.optim import SGD, Adam, lr_scheduler
+from tqdm import tqdm
+import math
+from torch.cuda import amp
+import torch
+from utils.loss import DBLoss
+import torch.nn as nn
+import yaml
+from basemodel import TextDetector
+from utils.db_utils import SegDetectorRepresenter, QuadMetric
+import numpy as np
+from datetime import datetime
+from torchsummary import summary
+import numexpr
+import os
+import shutil
+os.environ['NUMEXPR_MAX_THREADS'] = str(numexpr.detect_number_of_cores())
+
+from db_dataset import create_dataloader
+from utils.general import LOGGER, Loggers, CUDA, DEVICE
+import time
+import random
+
+torch.random.manual_seed(0)
+random.seed(0)
+np.random.seed(0)
+
+def one_cycle(y1=0.0, y2=1.0, steps=100):
+    return lambda x: ((1 - math.cos(x * math.pi / steps)) / 2) * (y2 - y1) + y1
+
+def eval_model(model: nn.Module, val_loader, post_process, metric_cls):
+    # global DEVICE
+    raw_metrics = []
+    total_frame = 0.0
+    total_time = 0.0
+    model.eval()
+    for i, batch in tqdm(enumerate(val_loader), total=len(val_loader), desc='test model'):
+        with torch.no_grad():
+            # 数据进行转换和丢到gpu
+            for key, value in batch.items():
+                if value is not None:
+                    if isinstance(value, torch.Tensor):
+                        batch[key] = value.to(DEVICE)
+            start = time.time()
+            with amp.autocast():
+                preds = model(batch['imgs'])
+            boxes, scores = post_process(batch, preds,is_output_polygon=False)
+            total_frame += batch['imgs'].size()[0]
+            total_time += time.time() - start
+            raw_metric = metric_cls.validate_measure(batch, (boxes, scores))
+            raw_metrics.append(raw_metric)
+    metrics = metric_cls.gather_measure(raw_metrics)
+    LOGGER.info('FPS:{}'.format(total_frame / total_time))
+    return metrics['recall'].avg, metrics['precision'].avg, metrics['fmeasure'].avg
+
+def train(hyp):
+    start_epoch = 0
+    hyp_train, hyp_data, hyp_model, hyp_logger, hyp_resume = hyp['train'], hyp['data'], hyp['model'], hyp['logger'], hyp['resume']
+    epochs = hyp_train['epochs']
+    batch_size = hyp_train['batch_size']
+
+    scaler = amp.GradScaler(enabled=CUDA)
+    criterion = DBLoss()
+    use_bce = False
+    if hyp_train['loss'] == 'bce':
+        use_bce = True
+    shrink_with_sigmoid = not use_bce
+
+    model = TextDetector(hyp_model['weights'], map_location='cpu', act=hyp_model['act'])
+    model.initialize_db(hyp_model['unet_weights'])
+    model.dbnet.shrink_with_sigmoid = shrink_with_sigmoid
+    model.train_db()
+    model.to(DEVICE)
+
+    if hyp_model['db_weights'] != '':
+        model.dbnet.load_state_dict(torch.load(hyp_model['db_weights'])['weights'])
+    if hyp_train['optimizer'] == 'adam': 
+        optimizer = Adam(model.dbnet.parameters(), lr=hyp_train['lr0'], betas=(0.937, 0.999), weight_decay=0.00002)  # adjust beta1 to momentum
+    else:
+        optimizer = SGD(model.dbnet.parameters(), lr=hyp_train['lr0'], momentum=hyp_train['momentum'], nesterov=True, weight_decay=hyp_train['weight_decay'])
+    
+    if hyp_train['linear_lr']:
+        lf = lambda x: (1 - x / (epochs - 1)) * (1.0 - hyp_train['lrf']) + hyp_train['lrf']  # linear
+    else:
+        lf = one_cycle(1, hyp_train['lrf'], epochs)  # cosine 1->hyp['lrf']
+
+    if hyp_train['linear_lr']:
+        lf = lambda x: (1 - x / (epochs - 1)) * (1.0 - hyp_train['lrf']) + hyp_train['lrf']  # linear
+    else:
+        lf = one_cycle(1, hyp_train['lrf'], epochs)  # cosine 1->hyp['lrf']
+    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)  # plot_lr_scheduler(optimizer, scheduler, epochs)
+    
+    logger = None
+    if hyp_resume['resume_training']:
+        LOGGER.info(f'resume traning ... ')
+        ckpt = torch.load(hyp_resume['ckpt'], map_location=DEVICE)
+        model.dbnet.load_state_dict(ckpt['weights'])
+        optimizer.load_state_dict(ckpt['optimizer'])
+        scheduler.load_state_dict(ckpt['scheduler'])
+        scheduler.step()
+        start_epoch = ckpt['epoch'] + 1
+        hyp_logger['run_id'] = ckpt['run_id']
+        logger = Loggers(hyp)
+
+    else:
+        # if hyp_logger['type'] == 'wandb':
+        logger = Loggers(hyp)
+
+    train_img_dir, train_mask_dir, imgsz, augment, aug_param = hyp_data['train_img_dir'], hyp_data['train_mask_dir'], hyp_data['imgsz'], hyp_data['augment'], hyp_data['aug_param']
+    val_img_dir, val_mask_dir = hyp_data['val_img_dir'], hyp_data['val_mask_dir']
+    train_dataset, train_loader = create_dataloader(train_img_dir, train_mask_dir, imgsz, batch_size, augment, aug_param, shuffle=True, workers=hyp_data['num_workers'], cache=hyp_data['cache'])
+    val_dataset, val_loader = create_dataloader(val_img_dir, val_mask_dir, imgsz, batch_size, augment=False, shuffle=False, workers=hyp_data['num_workers'], cache=hyp_data['cache'], with_ann=True)
+    nb = len(train_loader)
+    nw = max(round(3 * nb), 700)
+
+    LOGGER.info(f'num training imgs: {len(train_dataset)}, num val imgs: {len(val_dataset)}')
+
+    eval_interval = hyp_train['eval_interval']
+    best_f1 = best_epoch = -1
+    best_val_loss = np.inf
+
+    accumulation_steps = hyp_train['accumulation_steps']
+    summary(model, (3, 640, 640), device=DEVICE)
+    metric_cls = QuadMetric()
+    post_process = SegDetectorRepresenter(thresh=0.5)
+    best_f1 = -1
+    for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
+        model.train_db()
+        pbar = enumerate(train_loader)
+        pbar = tqdm(pbar, total=nb, bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}')  # progress bar
+        m_loss = 0
+        m_loss_s = 0
+        m_loss_t = 0
+        m_loss_b = 0
+        for i, batchs in pbar:
+            if (i+2) % 256 == 0:
+                train_dataset.initialize()
+                pbar.set_description(f' training size: {train_dataset.img_size}')
+            # warm up
+            if hyp_train['warm_up']:
+                ni = i + nb * epoch
+                if ni <= nw:
+                    xi = [0, nw]  # x interp
+                    for j, x in enumerate(optimizer.param_groups):
+                        x['lr'] = np.interp(ni, xi, [hyp_train['warmup_bias_lr'] if j == 2 else 0.0, x['initial_lr'] * lf(epoch)])
+                        if 'momentum' in x:
+                            x['momentum'] = np.interp(ni, xi, [hyp_train['warmup_momentum'], hyp_train['momentum']])
+
+            with amp.autocast():
+                for key in batchs.keys():
+                    batchs[key] = batchs[key].cuda()
+                preds = model(batchs['imgs'])
+                metric = criterion(preds, batchs, use_bce)
+            loss = metric['loss'] / accumulation_steps
+            scaler.scale(loss).backward()
+            if (i+1) % accumulation_steps == 0:
+                scaler.step(optimizer)
+                scaler.update()
+                optimizer.zero_grad()
+            m_loss = (m_loss * i + metric['loss'].detach()) / (i + 1)
+            m_loss_s = (m_loss_s * i + metric['loss_shrink_maps'].detach()) / (i + 1)
+            m_loss_t = (m_loss_t * i + metric['loss_threshold_maps'].detach()) / (i + 1)
+            m_loss_b = (m_loss_b * i + metric['loss_binary_maps'].detach()) / (i + 1)
+
+        if i % eval_interval == 0:
+            recall, precision, fmeasure = eval_model(model,  val_loader, post_process, metric_cls)
+            log_dict = {}
+            log_dict['train/lr'] = optimizer.param_groups[0]['lr']
+            log_dict['train/loss'] = m_loss
+            log_dict['train/loss_shrink'] = m_loss_s
+            log_dict['train/loss_threshold'] = m_loss_t
+            log_dict['train/loss_binary_maps'] = m_loss_b
+            log_dict['eval/recall'] = recall
+            log_dict['eval/precision'] = precision
+            log_dict['eval/f1'] = fmeasure
+            
+            save_best = best_f1 < fmeasure
+            if save_best:
+                best_f1 = fmeasure
+            last_ckpt = {'epoch': epoch,
+                        'best_f1': best_f1,
+                        'weights': model.dbnet.state_dict(),
+                        'best_val_loss': best_val_loss,
+                        'optimizer': optimizer.state_dict(),
+                        'scheduler': scheduler.state_dict(),
+                        'run_id': logger.wandb.id if logger.wandb is not None else None,
+                        'date': datetime.now().isoformat(),
+                        'hyp': hyp}
+            torch.save(last_ckpt, 'data/db_last.ckpt')
+            if save_best:
+                shutil.copy('data/db_last.ckpt', 'data/db_best.ckpt')
+            if logger is not None:
+                logger.on_train_epoch_end(epoch, log_dict)
+        scheduler.step()
+        pbar.close()
+
+if __name__ == '__main__':
+    hyp_p = r'data/train_db_hyp.yaml'
+    with open(hyp_p, 'r', encoding='utf8') as f:
+        hyp = yaml.safe_load(f.read())
+
+    # hyp['data']['train_img_dir'] = r'../datasets/pixanimegirls/processed'
+    hyp['data']['train_img_dir'] = [r'../datasets/codat_manga_v3/images/train', r'../datasets/codat_manga_v3/images/val', r'../datasets/pixanimegirls/processed']
+    hyp['data']['train_mask_dir'] = r'../datasets/TextLines'
+    # hyp['data']['train_img_dir'] = r'data/dataset/db_sub'
+    hyp['data']['val_img_dir'] = r'data/dataset/db_sub'
+    hyp['data']['cache'] = False
+    # hyp['data']['aug_param']['size_range'] = [-1]
+
+    hyp['train']['lr0'] = 0.01
+    hyp['train']['lrf'] = 0.002
+    hyp['train']['weight_decay'] = 0.00002
+    hyp['train']['batch_size'] = 4
+    hyp['train']['epochs'] = 160
+    # hyp['train']['optimizer'] = 'sgd'
+
+    hyp['train']['loss'] = 'bce'
+    hyp['logger']['type'] =  'wandb'
+
+    # hyp['resume']['resume_training'] = True
+    # hyp['resume']['ckpt'] = 'data/db_last_bk.ckpt'
+    # hyp['model']['db_weights'] = r'data/db_last.ckpt'
+    train(hyp)

train_seg.py

@@ -0,0 +1,210 @@
+import torch
+from torch.optim import SGD, Adam, lr_scheduler
+from tqdm import tqdm
+import math
+from torch.cuda import amp
+import torch
+from utils.loss import BinaryDiceLoss
+import torch.nn as nn
+import yaml
+from basemodel import TextDetector
+import numpy as np
+from datetime import datetime
+from torchsummary import summary
+import numexpr
+import os
+import shutil
+os.environ['NUMEXPR_MAX_THREADS'] = str(numexpr.detect_number_of_cores())
+
+from seg_dataset import create_dataloader
+from utils.general import LOGGER, Loggers, CUDA, DEVICE
+import random
+
+torch.random.manual_seed(0)
+random.seed(0)
+np.random.seed(0)
+
+
+def one_cycle(y1=0.0, y2=1.0, steps=100):
+    return lambda x: ((1 - math.cos(x * math.pi / steps)) / 2) * (y2 - y1) + y1
+
+def eval_model(model: nn.Module, val_loader):
+    global DEVICE
+    loss_func = BinaryDiceLoss()
+    pbar = enumerate(val_loader)
+    nb = len(val_loader)
+    model.eval()
+    pbar = tqdm(pbar, total=nb, bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}')  # progress bar
+    pr = tp = gt = m_loss = 0
+    with torch.no_grad():
+        for i, (imgs, masks) in pbar:
+            imgs = imgs.to(DEVICE)
+            masks = masks.to(DEVICE)
+            pred = model(imgs)
+            imgs.detach_()
+            del imgs
+            tp += torch.mul(pred, masks).sum().detach_()
+            gt += masks.sum().detach_()
+            pr += pred.sum().detach_()
+            loss = loss_func(pred, masks)
+            m_loss = (m_loss * i + loss.detach()) / (i + 1)
+            masks.detach_()
+            del masks
+    recall = tp / gt
+    precision = tp / pr
+    return recall, precision, m_loss
+
+def train(hyp):
+    with open(r'data/training_hyp.yaml', 'w', encoding='utf8') as f:
+        yaml.safe_dump(hyp, f)
+    start_epoch = 0
+    hyp_train, hyp_data, hyp_model, hyp_logger, hyp_resume = hyp['train'], hyp['data'], hyp['model'], hyp['logger'], hyp['resume']
+    epochs = hyp_train['epochs']
+    batch_size = hyp_train['batch_size']
+    model = TextDetector(**hyp_model)
+    if CUDA:
+        model.cuda()
+    params = model.seg_net.parameters()
+    
+    if hyp_train['optimizer'] == 'adam':
+        optimizer = Adam(params, lr=hyp_train['lr0'], betas=(hyp_train['momentum'], 0.999), weight_decay=hyp_train['weight_decay'])  # adjust beta1 to momentum
+    else:
+        optimizer = SGD(params, lr=hyp_train['lr0'], momentum=hyp_train['momentum'], nesterov=True, weight_decay=hyp_train['weight_decay'])
+    
+    if hyp_train['linear_lr']:
+        lf = lambda x: (1 - x / (epochs - 1)) * (1.0 - hyp_train['lrf']) + hyp_train['lrf']  # linear
+    else:
+        lf = one_cycle(1, hyp_train['lrf'], epochs)  # cosine 1->hyp['lrf']
+
+    scaler = amp.GradScaler(enabled=CUDA)
+    loss_func = BinaryDiceLoss()
+
+    # Scheduler
+    if hyp_train['linear_lr']:
+        lf = lambda x: (1 - x / (epochs - 1)) * (1.0 - hyp_train['lrf']) + hyp_train['lrf']  # linear
+    else:
+        lf = one_cycle(1, hyp_train['lrf'], epochs)  # cosine 1->hyp['lrf']
+    scheduler = lr_scheduler.LambdaLR(optimizer, lr_lambda=lf)  # plot_lr_scheduler(optimizer, scheduler, epochs)
+    
+    logger = None
+    if hyp_resume['resume_training']:
+        LOGGER.info(f'resume traning ... ')
+        ckpt = torch.load(hyp_resume['ckpt'], map_location=DEVICE)
+        model.seg_net.load_state_dict(ckpt['weights'])
+        optimizer.load_state_dict(ckpt['optimizer'])
+        scheduler.load_state_dict(ckpt['scheduler'])
+        scheduler.step()
+        start_epoch = ckpt['epoch'] + 1
+        hyp_logger['run_id'] = ckpt['run_id']
+        logger = Loggers(hyp)
+
+    else:
+        if hyp_logger['type'] == 'wandb':
+            logger = Loggers(hyp)
+
+    num_workers = 8
+    train_img_dir, train_mask_dir, imgsz, augment, aug_param = hyp_data['train_img_dir'], hyp_data['train_mask_dir'], hyp_data['imgsz'], hyp_data['augment'], hyp_data['aug_param']
+    val_img_dir, val_mask_dir = hyp_data['val_img_dir'], hyp_data['val_mask_dir']
+    train_dataset, train_loader = create_dataloader(train_img_dir, train_mask_dir, imgsz, batch_size, augment, aug_param, shuffle=True, workers=num_workers, cache=hyp_data['cache'])
+    val_dataset, val_loader = create_dataloader(val_img_dir, val_mask_dir, imgsz, 4, augment=False, shuffle=False, workers=num_workers, cache=hyp_data['cache'])
+    nb = len(train_loader)
+    nw = max(round(3 * nb), 700)
+
+    LOGGER.info(f'num training imgs: {len(train_dataset)}, num val imgs: {len(val_dataset)}')
+
+    eval_interval = hyp_train['eval_interval']
+    best_f1 = -1
+    best_val_loss = np.inf
+    accumulation_steps = hyp_train['accumulation_steps']
+    summary(model, (3, 640, 640), device=DEVICE)
+    for epoch in range(start_epoch, epochs):  # epoch ------------------------------------------------------------------
+        
+        model.train_mask()
+        train_dataset.initialize() 
+        pbar = enumerate(train_loader)
+        pbar = tqdm(pbar, total=nb, bar_format='{l_bar}{bar:10}{r_bar}{bar:-10b}')  # progress bar
+        
+        m_loss = 0
+        for i, (imgs, masks) in pbar:
+            
+            pbar.set_description(f' training size: {train_dataset.img_size}')
+            # warm up
+            ni = i + nb * epoch
+            if ni <= nw:
+                xi = [0, nw]  # x interp
+                for j, x in enumerate(optimizer.param_groups):
+                    x['lr'] = np.interp(ni, xi, [hyp_train['warmup_bias_lr'] if j == 2 else 0.0, x['initial_lr'] * lf(epoch)])
+                    if 'momentum' in x:
+                        x['momentum'] = np.interp(ni, xi, [hyp_train['warmup_momentum'], hyp_train['momentum']])
+
+            imgs, masks = imgs.to(DEVICE), masks.to(DEVICE)
+            with amp.autocast():
+                preds = model(imgs)
+                imgs.detach_()
+                del imgs
+                loss = loss_func(preds, masks)
+                masks.detach_()
+                del masks
+            scaler.scale(loss).backward()
+            if i % accumulation_steps == 0:
+                scaler.step(optimizer)
+                scaler.update()
+                optimizer.zero_grad()
+            m_loss = (m_loss * i + loss.detach()) / (i + 1)
+        
+        if (epoch + 1) % eval_interval == 0:
+            recall, precision, eval_m_loss = eval_model(model, val_loader)
+            f1 = 2 * recall * precision / (recall + precision)
+            last_ckpt = {'epoch': epoch,
+                        'best_f1': best_f1,
+                        'weights': model.seg_net.state_dict(),
+                        'best_val_loss': best_val_loss,
+                        'optimizer': optimizer.state_dict(),
+                        'scheduler': scheduler.state_dict(),
+                        'run_id': logger.wandb.id if logger is not None else None,
+                        'date': datetime.now().isoformat(),
+                        'hyp': hyp}
+            torch.save(last_ckpt, 'data/unet_last.ckpt')
+            if best_f1 < f1:
+                best_f1 = f1
+                LOGGER.info(f'saving model at epoch {epoch}, best val f1: {best_f1}')
+                shutil.copy2('data/unet_last.ckpt', 'data/unet_best.ckpt')
+            LOGGER.info(f'epoch {epoch}/{epochs-1} loss: {m_loss} precision: {precision} recall: {recall}')
+            if logger is not None:
+                log_dict = {}
+                log_dict['train/lr'] = optimizer.param_groups[0]['lr']
+                log_dict['train/loss'] = m_loss
+                log_dict['eval/recall'] = recall
+                log_dict['eval/precision'] = precision
+                log_dict['eval/f1'] = f1
+                log_dict['eval/eval_m_loss'] = eval_m_loss
+                logger.on_train_epoch_end(epoch, log_dict)
+        scheduler.step()
+        pbar.close()
+
+if __name__ == '__main__':
+    hyp_p = r'data/train_hyp.yaml'
+    with open(hyp_p, 'r', encoding='utf8') as f:
+        hyp = yaml.safe_load(f.read())
+
+    hyp['data']['train_img_dir'] = [r'../datasets/codat_manga_v3/images/train', r'../datasets/ComicErased/processed']
+    # hyp['data']['train_img_dir'] = [r'../datasets/codat_manga_v3/images/val']
+    hyp['data']['val_img_dir'] = [r'../datasets/codat_manga_v3/images/val']
+    hyp['data']['train_mask_dir'] = r'../datasets/ComicSegV2'
+    hyp['data']['val_mask_dir'] = r'../datasets/ComicSegV2'
+    hyp['data']['imgsz'] = 1024
+    hyp['data']['cache'] = False
+    hyp['data']['aug_param']['neg'] = 0.3
+    hyp['data']['aug_param']['size_range'] = [0.85, 1.1]
+
+    hyp['train']['lr0'] = 0.004
+    hyp['train']['lrf'] = 0.005
+    hyp['train']['weight_decay'] = 0.00002
+    hyp['train']['epochs'] = 120
+    hyp['train']['accumulation_steps'] = 4
+    hyp['train']['batch_size'] = 4
+    hyp['logger']['type'] = 'wandb'
+
+    # hyp['resume']['resume_training'] = True
+    # hyp['resume']['ckpt'] = 'data/unet_last.ckpt'
+    train(hyp)

utils/db_utils.py

@@ -0,0 +1,701 @@
+import cv2
+import numpy as np
+import pyclipper
+from shapely.geometry import Polygon
+from collections import namedtuple
+import torch
+import warnings
+warnings.filterwarnings('ignore')
+
+
+def iou_rotate(box_a, box_b, method='union'):
+    rect_a = cv2.minAreaRect(box_a)
+    rect_b = cv2.minAreaRect(box_b)
+    r1 = cv2.rotatedRectangleIntersection(rect_a, rect_b)
+    if r1[0] == 0:
+        return 0
+    else:
+        inter_area = cv2.contourArea(r1[1])
+        area_a = cv2.contourArea(box_a)
+        area_b = cv2.contourArea(box_b)
+        union_area = area_a + area_b - inter_area
+        if union_area == 0 or inter_area == 0:
+            return 0
+        if method == 'union':
+            iou = inter_area / union_area
+        elif method == 'intersection':
+            iou = inter_area / min(area_a, area_b)
+        else:
+            raise NotImplementedError
+        return iou
+
+class SegDetectorRepresenter():
+    def __init__(self, thresh=0.3, box_thresh=0.7, max_candidates=1000, unclip_ratio=1.5):
+        self.min_size = 3
+        self.thresh = thresh
+        self.box_thresh = box_thresh
+        self.max_candidates = max_candidates
+        self.unclip_ratio = unclip_ratio
+
+    def __call__(self, batch, pred, is_output_polygon=False):
+        '''
+        batch: (image, polygons, ignore_tags
+        batch: a dict produced by dataloaders.
+            image: tensor of shape (N, C, H, W).
+            polygons: tensor of shape (N, K, 4, 2), the polygons of objective regions.
+            ignore_tags: tensor of shape (N, K), indicates whether a region is ignorable or not.
+            shape: the original shape of images.
+            filename: the original filenames of images.
+        pred:
+            binary: text region segmentation map, with shape (N, H, W)
+            thresh: [if exists] thresh hold prediction with shape (N, H, W)
+            thresh_binary: [if exists] binarized with threshold, (N, H, W)
+        '''
+        pred = pred[:, 0, :, :]
+        segmentation = self.binarize(pred)
+        boxes_batch = []
+        scores_batch = []
+        # print(pred.size())
+        batch_size = pred.size(0) if isinstance(pred, torch.Tensor) else pred.shape[0]
+        for batch_index in range(batch_size):
+            # height, width = batch['shape'][batch_index]
+            height, width = pred.shape[1], pred.shape[2]
+            if is_output_polygon:
+                boxes, scores = self.polygons_from_bitmap(pred[batch_index], segmentation[batch_index], width, height)
+            else:
+                boxes, scores = self.boxes_from_bitmap(pred[batch_index], segmentation[batch_index], width, height)
+            boxes_batch.append(boxes)
+            scores_batch.append(scores)
+        return boxes_batch, scores_batch
+
+    def binarize(self, pred):
+        return pred > self.thresh
+
+    def polygons_from_bitmap(self, pred, _bitmap, dest_width, dest_height):
+        '''
+        _bitmap: single map with shape (H, W),
+            whose values are binarized as {0, 1}
+        '''
+
+        assert len(_bitmap.shape) == 2
+        bitmap = _bitmap.cpu().numpy()  # The first channel
+        pred = pred.cpu().detach().numpy()
+        height, width = bitmap.shape
+        boxes = []
+        scores = []
+
+        contours, _ = cv2.findContours((bitmap * 255).astype(np.uint8), cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
+
+        for contour in contours[:self.max_candidates]:
+            epsilon = 0.005 * cv2.arcLength(contour, True)
+            approx = cv2.approxPolyDP(contour, epsilon, True)
+            points = approx.reshape((-1, 2))
+            if points.shape[0] < 4:
+                continue
+            # _, sside = self.get_mini_boxes(contour)
+            # if sside < self.min_size:
+            #     continue
+            score = self.box_score_fast(pred, contour.squeeze(1))
+            if self.box_thresh > score:
+                continue
+
+            if points.shape[0] > 2:
+                box = self.unclip(points, unclip_ratio=self.unclip_ratio)
+                if len(box) > 1:
+                    continue
+            else:
+                continue
+            box = box.reshape(-1, 2)
+            _, sside = self.get_mini_boxes(box.reshape((-1, 1, 2)))
+            if sside < self.min_size + 2:
+                continue
+
+            if not isinstance(dest_width, int):
+                dest_width = dest_width.item()
+                dest_height = dest_height.item()
+
+            box[:, 0] = np.clip(np.round(box[:, 0] / width * dest_width), 0, dest_width)
+            box[:, 1] = np.clip(np.round(box[:, 1] / height * dest_height), 0, dest_height)
+            boxes.append(box)
+            scores.append(score)
+        return boxes, scores
+
+    def boxes_from_bitmap(self, pred, _bitmap, dest_width, dest_height):
+        '''
+        _bitmap: single map with shape (H, W),
+            whose values are binarized as {0, 1}
+        '''
+
+        assert len(_bitmap.shape) == 2
+        if isinstance(pred, torch.Tensor):
+            bitmap = _bitmap.cpu().numpy()  # The first channel
+            pred = pred.cpu().detach().numpy()
+        else:
+            bitmap = _bitmap
+        height, width = bitmap.shape
+        contours, _ = cv2.findContours((bitmap * 255).astype(np.uint8), cv2.RETR_LIST, cv2.CHAIN_APPROX_SIMPLE)
+        num_contours = min(len(contours), self.max_candidates)
+        boxes = np.zeros((num_contours, 4, 2), dtype=np.int16)
+        scores = np.zeros((num_contours,), dtype=np.float32)
+
+        for index in range(num_contours):
+            contour = contours[index].squeeze(1)
+            points, sside = self.get_mini_boxes(contour)
+            # if sside < self.min_size:
+            #     continue
+            if sside < 2:
+                continue
+            points = np.array(points)
+            score = self.box_score_fast(pred, contour)
+            # if self.box_thresh > score:
+            #     continue
+
+            box = self.unclip(points, unclip_ratio=self.unclip_ratio).reshape(-1, 1, 2)
+            box, sside = self.get_mini_boxes(box)
+            # if sside < 5:
+            #     continue
+            box = np.array(box)
+            if not isinstance(dest_width, int):
+                dest_width = dest_width.item()
+                dest_height = dest_height.item()
+
+            box[:, 0] = np.clip(np.round(box[:, 0] / width * dest_width), 0, dest_width)
+            box[:, 1] = np.clip(np.round(box[:, 1] / height * dest_height), 0, dest_height)
+            boxes[index, :, :] = box.astype(np.int16)
+            scores[index] = score
+        return boxes, scores
+
+    def unclip(self, box, unclip_ratio=1.5):
+        poly = Polygon(box)
+        distance = poly.area * unclip_ratio / poly.length
+        offset = pyclipper.PyclipperOffset()
+        offset.AddPath(box, pyclipper.JT_ROUND, pyclipper.ET_CLOSEDPOLYGON)
+        expanded = np.array(offset.Execute(distance))
+        return expanded
+
+    def get_mini_boxes(self, contour):
+        bounding_box = cv2.minAreaRect(contour)
+        points = sorted(list(cv2.boxPoints(bounding_box)), key=lambda x: x[0])
+
+        index_1, index_2, index_3, index_4 = 0, 1, 2, 3
+        if points[1][1] > points[0][1]:
+            index_1 = 0
+            index_4 = 1
+        else:
+            index_1 = 1
+            index_4 = 0
+        if points[3][1] > points[2][1]:
+            index_2 = 2
+            index_3 = 3
+        else:
+            index_2 = 3
+            index_3 = 2
+
+        box = [points[index_1], points[index_2], points[index_3], points[index_4]]
+        return box, min(bounding_box[1])
+
+    def box_score_fast(self, bitmap, _box):
+        h, w = bitmap.shape[:2]
+        box = _box.copy()
+        xmin = np.clip(np.floor(box[:, 0].min()).astype(np.int64), 0, w - 1)
+        xmax = np.clip(np.ceil(box[:, 0].max()).astype(np.int64), 0, w - 1)
+        ymin = np.clip(np.floor(box[:, 1].min()).astype(np.int64), 0, h - 1)
+        ymax = np.clip(np.ceil(box[:, 1].max()).astype(np.int64), 0, h - 1)
+
+        mask = np.zeros((ymax - ymin + 1, xmax - xmin + 1), dtype=np.uint8)
+        box[:, 0] = box[:, 0] - xmin
+        box[:, 1] = box[:, 1] - ymin
+        cv2.fillPoly(mask, box.reshape(1, -1, 2).astype(np.int32), 1)
+        if bitmap.dtype == np.float16:
+            bitmap = bitmap.astype(np.float32)
+        return cv2.mean(bitmap[ymin:ymax + 1, xmin:xmax + 1], mask)[0]
+
+class AverageMeter(object):
+    """Computes and stores the average and current value"""
+
+    def __init__(self):
+        self.reset()
+
+    def reset(self):
+        self.val = 0
+        self.avg = 0
+        self.sum = 0
+        self.count = 0
+
+    def update(self, val, n=1):
+        self.val = val
+        self.sum += val * n
+        self.count += n
+        self.avg = self.sum / self.count
+        return self
+
+
+class DetectionIoUEvaluator(object):
+    def __init__(self, is_output_polygon=False, iou_constraint=0.5, area_precision_constraint=0.5):
+        self.is_output_polygon = is_output_polygon
+        self.iou_constraint = iou_constraint
+        self.area_precision_constraint = area_precision_constraint
+
+    def evaluate_image(self, gt, pred):
+
+        def get_union(pD, pG):
+            return Polygon(pD).union(Polygon(pG)).area
+
+        def get_intersection_over_union(pD, pG):
+            return get_intersection(pD, pG) / get_union(pD, pG)
+
+        def get_intersection(pD, pG):
+            return Polygon(pD).intersection(Polygon(pG)).area
+
+        def compute_ap(confList, matchList, numGtCare):
+            correct = 0
+            AP = 0
+            if len(confList) > 0:
+                confList = np.array(confList)
+                matchList = np.array(matchList)
+                sorted_ind = np.argsort(-confList)
+                confList = confList[sorted_ind]
+                matchList = matchList[sorted_ind]
+                for n in range(len(confList)):
+                    match = matchList[n]
+                    if match:
+                        correct += 1
+                        AP += float(correct) / (n + 1)
+
+                if numGtCare > 0:
+                    AP /= numGtCare
+
+            return AP
+
+        perSampleMetrics = {}
+
+        matchedSum = 0
+
+        Rectangle = namedtuple('Rectangle', 'xmin ymin xmax ymax')
+
+        numGlobalCareGt = 0
+        numGlobalCareDet = 0
+
+        arrGlobalConfidences = []
+        arrGlobalMatches = []
+
+        recall = 0
+        precision = 0
+        hmean = 0
+
+        detMatched = 0
+
+        iouMat = np.empty([1, 1])
+
+        gtPols = []
+        detPols = []
+
+        gtPolPoints = []
+        detPolPoints = []
+
+        # Array of Ground Truth Polygons' keys marked as don't Care
+        gtDontCarePolsNum = []
+        # Array of Detected Polygons' matched with a don't Care GT
+        detDontCarePolsNum = []
+
+        pairs = []
+        detMatchedNums = []
+
+        arrSampleConfidences = []
+        arrSampleMatch = []
+
+        evaluationLog = ""
+
+        for n in range(len(gt)):
+            points = gt[n]['points']
+            # transcription = gt[n]['text']
+            dontCare = gt[n]['ignore']
+
+            if not Polygon(points).is_valid or not Polygon(points).is_simple:
+                continue
+
+            gtPol = points
+            gtPols.append(gtPol)
+            gtPolPoints.append(points)
+            if dontCare:
+                gtDontCarePolsNum.append(len(gtPols) - 1)
+
+        evaluationLog += "GT polygons: " + str(len(gtPols)) + (" (" + str(len(
+            gtDontCarePolsNum)) + " don't care)\n" if len(gtDontCarePolsNum) > 0 else "\n")
+
+        for n in range(len(pred)):
+            points = pred[n]['points']
+            if not Polygon(points).is_valid or not Polygon(points).is_simple:
+                continue
+
+            detPol = points
+            detPols.append(detPol)
+            detPolPoints.append(points)
+            if len(gtDontCarePolsNum) > 0:
+                for dontCarePol in gtDontCarePolsNum:
+                    dontCarePol = gtPols[dontCarePol]
+                    intersected_area = get_intersection(dontCarePol, detPol)
+                    pdDimensions = Polygon(detPol).area
+                    precision = 0 if pdDimensions == 0 else intersected_area / pdDimensions
+                    if (precision > self.area_precision_constraint):
+                        detDontCarePolsNum.append(len(detPols) - 1)
+                        break
+
+        evaluationLog += "DET polygons: " + str(len(detPols)) + (" (" + str(len(
+            detDontCarePolsNum)) + " don't care)\n" if len(detDontCarePolsNum) > 0 else "\n")
+
+        if len(gtPols) > 0 and len(detPols) > 0:
+            # Calculate IoU and precision matrixs
+            outputShape = [len(gtPols), len(detPols)]
+            iouMat = np.empty(outputShape)
+            gtRectMat = np.zeros(len(gtPols), np.int8)
+            detRectMat = np.zeros(len(detPols), np.int8)
+            if self.is_output_polygon:
+                for gtNum in range(len(gtPols)):
+                    for detNum in range(len(detPols)):
+                        pG = gtPols[gtNum]
+                        pD = detPols[detNum]
+                        iouMat[gtNum, detNum] = get_intersection_over_union(pD, pG)
+            else:
+                # gtPols = np.float32(gtPols)
+                # detPols = np.float32(detPols)
+                for gtNum in range(len(gtPols)):
+                    for detNum in range(len(detPols)):
+                        pG = np.float32(gtPols[gtNum])
+                        pD = np.float32(detPols[detNum])
+                        iouMat[gtNum, detNum] = iou_rotate(pD, pG)
+            for gtNum in range(len(gtPols)):
+                for detNum in range(len(detPols)):
+                    if gtRectMat[gtNum] == 0 and detRectMat[
+                        detNum] == 0 and gtNum not in gtDontCarePolsNum and detNum not in detDontCarePolsNum:
+                        if iouMat[gtNum, detNum] > self.iou_constraint:
+                            gtRectMat[gtNum] = 1
+                            detRectMat[detNum] = 1
+                            detMatched += 1
+                            pairs.append({'gt': gtNum, 'det': detNum})
+                            detMatchedNums.append(detNum)
+                            evaluationLog += "Match GT #" + \
+                                             str(gtNum) + " with Det #" + str(detNum) + "\n"
+
+        numGtCare = (len(gtPols) - len(gtDontCarePolsNum))
+        numDetCare = (len(detPols) - len(detDontCarePolsNum))
+        if numGtCare == 0:
+            recall = float(1)
+            precision = float(0) if numDetCare > 0 else float(1)
+        else:
+            recall = float(detMatched) / numGtCare
+            precision = 0 if numDetCare == 0 else float(
+                detMatched) / numDetCare
+
+        hmean = 0 if (precision + recall) == 0 else 2.0 * \
+                                                    precision * recall / (precision + recall)
+
+        matchedSum += detMatched
+        numGlobalCareGt += numGtCare
+        numGlobalCareDet += numDetCare
+
+        perSampleMetrics = {
+            'precision': precision,
+            'recall': recall,
+            'hmean': hmean,
+            'pairs': pairs,
+            'iouMat': [] if len(detPols) > 100 else iouMat.tolist(),
+            'gtPolPoints': gtPolPoints,
+            'detPolPoints': detPolPoints,
+            'gtCare': numGtCare,
+            'detCare': numDetCare,
+            'gtDontCare': gtDontCarePolsNum,
+            'detDontCare': detDontCarePolsNum,
+            'detMatched': detMatched,
+            'evaluationLog': evaluationLog
+        }
+
+        return perSampleMetrics
+
+    def combine_results(self, results):
+        numGlobalCareGt = 0
+        numGlobalCareDet = 0
+        matchedSum = 0
+        for result in results:
+            numGlobalCareGt += result['gtCare']
+            numGlobalCareDet += result['detCare']
+            matchedSum += result['detMatched']
+
+        methodRecall = 0 if numGlobalCareGt == 0 else float(
+            matchedSum) / numGlobalCareGt
+        methodPrecision = 0 if numGlobalCareDet == 0 else float(
+            matchedSum) / numGlobalCareDet
+        methodHmean = 0 if methodRecall + methodPrecision == 0 else 2 * \
+                                                                    methodRecall * methodPrecision / (
+                                                                            methodRecall + methodPrecision)
+
+        methodMetrics = {'precision': methodPrecision,
+                         'recall': methodRecall, 'hmean': methodHmean}
+
+        return methodMetrics
+
+class QuadMetric():
+    def __init__(self, is_output_polygon=False):
+        self.is_output_polygon = is_output_polygon
+        self.evaluator = DetectionIoUEvaluator(is_output_polygon=is_output_polygon)
+
+    def measure(self, batch, output, box_thresh=0.6):
+        '''
+        batch: (image, polygons, ignore_tags
+        batch: a dict produced by dataloaders.
+            image: tensor of shape (N, C, H, W).
+            polygons: tensor of shape (N, K, 4, 2), the polygons of objective regions.
+            ignore_tags: tensor of shape (N, K), indicates whether a region is ignorable or not.
+            shape: the original shape of images.
+            filename: the original filenames of images.
+        output: (polygons, ...)
+        '''
+        results = []
+        gt_polyons_batch = batch['text_polys']
+        ignore_tags_batch = batch['ignore_tags']
+        pred_polygons_batch = np.array(output[0])
+        pred_scores_batch = np.array(output[1])
+        for polygons, pred_polygons, pred_scores, ignore_tags in zip(gt_polyons_batch, pred_polygons_batch, pred_scores_batch, ignore_tags_batch):
+            gt = [dict(points=np.int64(polygons[i]), ignore=ignore_tags[i]) for i in range(len(polygons))]
+            if self.is_output_polygon:
+                pred = [dict(points=pred_polygons[i]) for i in range(len(pred_polygons))]
+            else:
+                pred = []
+                # print(pred_polygons.shape)
+                for i in range(pred_polygons.shape[0]):
+                    if pred_scores[i] >= box_thresh:
+                        # print(pred_polygons[i,:,:].tolist())
+                        pred.append(dict(points=pred_polygons[i, :, :].astype(np.int64)))
+                # pred = [dict(points=pred_polygons[i,:,:].tolist()) if pred_scores[i] >= box_thresh for i in range(pred_polygons.shape[0])]
+            results.append(self.evaluator.evaluate_image(gt, pred))
+        return results
+
+    def validate_measure(self, batch, output, box_thresh=0.6):
+        return self.measure(batch, output, box_thresh)
+
+    def evaluate_measure(self, batch, output):
+        return self.measure(batch, output), np.linspace(0, batch['image'].shape[0]).tolist()
+
+    def gather_measure(self, raw_metrics):
+        raw_metrics = [image_metrics
+                       for batch_metrics in raw_metrics
+                       for image_metrics in batch_metrics]
+
+        result = self.evaluator.combine_results(raw_metrics)
+
+        precision = AverageMeter()
+        recall = AverageMeter()
+        fmeasure = AverageMeter()
+
+        precision.update(result['precision'], n=len(raw_metrics))
+        recall.update(result['recall'], n=len(raw_metrics))
+        fmeasure_score = 2 * precision.val * recall.val / (precision.val + recall.val + 1e-8)
+        fmeasure.update(fmeasure_score)
+
+        return {
+            'precision': precision,
+            'recall': recall,
+            'fmeasure': fmeasure
+        }
+
+def shrink_polygon_py(polygon, shrink_ratio):
+    """
+    对框进行缩放，返回去的比例为1/shrink_ratio 即可
+    """
+    cx = polygon[:, 0].mean()
+    cy = polygon[:, 1].mean()
+    polygon[:, 0] = cx + (polygon[:, 0] - cx) * shrink_ratio
+    polygon[:, 1] = cy + (polygon[:, 1] - cy) * shrink_ratio
+    return polygon
+
+
+def shrink_polygon_pyclipper(polygon, shrink_ratio):
+    from shapely.geometry import Polygon
+    import pyclipper
+    polygon_shape = Polygon(polygon)
+    distance = polygon_shape.area * (1 - np.power(shrink_ratio, 2)) / polygon_shape.length
+    subject = [tuple(l) for l in polygon]
+    padding = pyclipper.PyclipperOffset()
+    padding.AddPath(subject, pyclipper.JT_ROUND, pyclipper.ET_CLOSEDPOLYGON)
+    shrunk = padding.Execute(-distance)
+    if shrunk == []:
+        shrunk = np.array(shrunk)
+    else:
+        shrunk = np.array(shrunk[0]).reshape(-1, 2)
+    return shrunk
+
+class MakeShrinkMap():
+    r'''
+    Making binary mask from detection data with ICDAR format.
+    Typically following the process of class `MakeICDARData`.
+    '''
+
+    def __init__(self, min_text_size=4, shrink_ratio=0.4, shrink_type='pyclipper'):
+        shrink_func_dict = {'py': shrink_polygon_py, 'pyclipper': shrink_polygon_pyclipper}
+        self.shrink_func = shrink_func_dict[shrink_type]
+        self.min_text_size = min_text_size
+        self.shrink_ratio = shrink_ratio
+
+    def __call__(self, data: dict) -> dict:
+        """
+        从scales中随机选择一个尺度，对图片和文本框进行缩放
+        :param data: {'imgs':,'text_polys':,'texts':,'ignore_tags':}
+        :return:
+        """
+        image = data['imgs']
+        text_polys = data['text_polys']
+        ignore_tags = data['ignore_tags']
+
+        h, w = image.shape[:2]
+        text_polys, ignore_tags = self.validate_polygons(text_polys, ignore_tags, h, w)
+        gt = np.zeros((h, w), dtype=np.float32)
+        mask = np.ones((h, w), dtype=np.float32)
+        for i in range(len(text_polys)):
+            polygon = text_polys[i]
+            height = max(polygon[:, 1]) - min(polygon[:, 1])
+            width = max(polygon[:, 0]) - min(polygon[:, 0])
+            if ignore_tags[i] or min(height, width) < self.min_text_size:
+                cv2.fillPoly(mask, polygon.astype(np.int32)[np.newaxis, :, :], 0)
+                ignore_tags[i] = True
+            else:
+                shrunk = self.shrink_func(polygon, self.shrink_ratio)
+                if shrunk.size == 0:
+                    cv2.fillPoly(mask, polygon.astype(np.int32)[np.newaxis, :, :], 0)
+                    ignore_tags[i] = True
+                    continue
+                cv2.fillPoly(gt, [shrunk.astype(np.int32)], 1)
+
+        data['shrink_map'] = gt
+        data['shrink_mask'] = mask
+        return data
+
+    def validate_polygons(self, polygons, ignore_tags, h, w):
+        '''
+        polygons (numpy.array, required): of shape (num_instances, num_points, 2)
+        '''
+        if len(polygons) == 0:
+            return polygons, ignore_tags
+        assert len(polygons) == len(ignore_tags)
+        for polygon in polygons:
+            polygon[:, 0] = np.clip(polygon[:, 0], 0, w - 1)
+            polygon[:, 1] = np.clip(polygon[:, 1], 0, h - 1)
+
+        for i in range(len(polygons)):
+            area = self.polygon_area(polygons[i])
+            if abs(area) < 1:
+                ignore_tags[i] = True
+            if area > 0:
+                polygons[i] = polygons[i][::-1, :]
+        return polygons, ignore_tags
+
+    def polygon_area(self, polygon):
+        return cv2.contourArea(polygon)
+
+
+class MakeBorderMap():
+    def __init__(self, shrink_ratio=0.4, thresh_min=0.3, thresh_max=0.7):
+        self.shrink_ratio = shrink_ratio
+        self.thresh_min = thresh_min
+        self.thresh_max = thresh_max
+
+    def __call__(self, data: dict) -> dict:
+        """
+        从scales中随机选择一个尺度，对图片和文本框进行缩放
+        :param data: {'imgs':,'text_polys':,'texts':,'ignore_tags':}
+        :return:
+        """
+        im = data['imgs']
+        text_polys = data['text_polys']
+        ignore_tags = data['ignore_tags']
+
+        canvas = np.zeros(im.shape[:2], dtype=np.float32)
+        mask = np.zeros(im.shape[:2], dtype=np.float32)
+
+        for i in range(len(text_polys)):
+            if ignore_tags[i]:
+                continue
+            self.draw_border_map(text_polys[i], canvas, mask=mask)
+        canvas = canvas * (self.thresh_max - self.thresh_min) + self.thresh_min
+
+        data['threshold_map'] = canvas
+        data['threshold_mask'] = mask
+        return data
+
+    def draw_border_map(self, polygon, canvas, mask):
+        polygon = np.array(polygon)
+        assert polygon.ndim == 2
+        assert polygon.shape[1] == 2
+
+        polygon_shape = Polygon(polygon)
+        if polygon_shape.area <= 0:
+            return
+        distance = polygon_shape.area * (1 - np.power(self.shrink_ratio, 2)) / polygon_shape.length
+        subject = [tuple(l) for l in polygon]
+        padding = pyclipper.PyclipperOffset()
+        padding.AddPath(subject, pyclipper.JT_ROUND,
+                        pyclipper.ET_CLOSEDPOLYGON)
+
+        padded_polygon = np.array(padding.Execute(distance)[0])
+        cv2.fillPoly(mask, [padded_polygon.astype(np.int32)], 1.0)
+
+        xmin = padded_polygon[:, 0].min()
+        xmax = padded_polygon[:, 0].max()
+        ymin = padded_polygon[:, 1].min()
+        ymax = padded_polygon[:, 1].max()
+        width = xmax - xmin + 1
+        height = ymax - ymin + 1
+
+        polygon[:, 0] = polygon[:, 0] - xmin
+        polygon[:, 1] = polygon[:, 1] - ymin
+
+        xs = np.broadcast_to(
+            np.linspace(0, width - 1, num=width).reshape(1, width), (height, width))
+        ys = np.broadcast_to(
+            np.linspace(0, height - 1, num=height).reshape(height, 1), (height, width))
+
+        distance_map = np.zeros(
+            (polygon.shape[0], height, width), dtype=np.float32)
+        for i in range(polygon.shape[0]):
+            j = (i + 1) % polygon.shape[0]
+            absolute_distance = self.distance(xs, ys, polygon[i], polygon[j])
+            distance_map[i] = np.clip(absolute_distance / distance, 0, 1)
+        distance_map = distance_map.min(axis=0)
+
+        xmin_valid = min(max(0, xmin), canvas.shape[1] - 1)
+        xmax_valid = min(max(0, xmax), canvas.shape[1] - 1)
+        ymin_valid = min(max(0, ymin), canvas.shape[0] - 1)
+        ymax_valid = min(max(0, ymax), canvas.shape[0] - 1)
+        canvas[ymin_valid:ymax_valid + 1, xmin_valid:xmax_valid + 1] = np.fmax(
+            1 - distance_map[
+                ymin_valid - ymin:ymax_valid - ymax + height,
+                xmin_valid - xmin:xmax_valid - xmax + width],
+            canvas[ymin_valid:ymax_valid + 1, xmin_valid:xmax_valid + 1])
+
+    def distance(self, xs, ys, point_1, point_2):
+        '''
+        compute the distance from point to a line
+        ys: coordinates in the first axis
+        xs: coordinates in the second axis
+        point_1, point_2: (x, y), the end of the line
+        '''
+        height, width = xs.shape[:2]
+        square_distance_1 = np.square(xs - point_1[0]) + np.square(ys - point_1[1])
+        square_distance_2 = np.square(xs - point_2[0]) + np.square(ys - point_2[1])
+        square_distance = np.square(point_1[0] - point_2[0]) + np.square(point_1[1] - point_2[1])
+
+        cosin = (square_distance - square_distance_1 - square_distance_2) / (2 * np.sqrt(square_distance_1 * square_distance_2))
+        square_sin = 1 - np.square(cosin)
+        square_sin = np.nan_to_num(square_sin)
+
+        result = np.sqrt(square_distance_1 * square_distance_2 * square_sin / square_distance)
+        result[cosin < 0] = np.sqrt(np.fmin(square_distance_1, square_distance_2))[cosin < 0]
+        return result
+
+    def extend_line(self, point_1, point_2, result):
+        ex_point_1 = (int(round(point_1[0] + (point_1[0] - point_2[0]) * (1 + self.shrink_ratio))),
+                      int(round(point_1[1] + (point_1[1] - point_2[1]) * (1 + self.shrink_ratio))))
+        cv2.line(result, tuple(ex_point_1), tuple(point_1), 4096.0, 1, lineType=cv2.LINE_AA, shift=0)
+        ex_point_2 = (int(round(point_2[0] + (point_2[0] - point_1[0]) * (1 + self.shrink_ratio))),
+                      int(round(point_2[1] + (point_2[1] - point_1[1]) * (1 + self.shrink_ratio))))
+        cv2.line(result, tuple(ex_point_2), tuple(point_2), 4096.0, 1, lineType=cv2.LINE_AA, shift=0)
+        return ex_point_1, ex_point_2

utils/export.py

@@ -0,0 +1,58 @@
+
+from cv2 import imshow
+from matplotlib import lines
+import numpy as np
+import onnxruntime
+import cv2
+import torch
+import onnx
+from basemodel import TextDetBase
+import onnxsim
+from models.yolov5.common import Conv
+from models.yolov5.yolo import Detect
+import torch.nn as nn
+import time
+from seg_dataset import letterbox
+from utils.yolov5_utils import fuse_conv_and_bn
+
+class SiLU(nn.Module):  # export-friendly version of nn.SiLU()
+    @staticmethod
+    def forward(x):
+        return x * torch.sigmoid(x)
+
+def concate_models(blk_weights, seg_weights, det_weights, save_path):
+    textdetector_dict = dict()
+    textdetector_dict['blk_det'] = torch.load(blk_weights, map_location='cpu')
+    textdetector_dict['text_seg'] = torch.load(seg_weights, map_location='cpu')['weights']
+    textdetector_dict['text_det'] = torch.load(det_weights, map_location='cpu')['weights']
+    torch.save(textdetector_dict, save_path)
+
+def export_onnx(model, im, file, opset, train=False, simplify=True, dynamic=False, inplace=False):
+    # YOLOv5 ONNX export
+    f = file + '.onnx'
+    for k, m in model.named_modules():
+        if isinstance(m, Conv):  # assign export-friendly activations
+            if isinstance(m.act, nn.SiLU):
+                m.act = SiLU()
+        elif isinstance(m, Detect):
+            m.inplace = inplace
+            m.onnx_dynamic = False
+    torch.onnx.export(model, im, f, verbose=False, opset_version=opset,
+                        training=torch.onnx.TrainingMode.TRAINING if train else torch.onnx.TrainingMode.EVAL,
+                        do_constant_folding=not train,
+                        input_names=['images'],
+                        output_names=['blk', 'seg', 'det'],
+                        dynamic_axes={'images': {0: 'batch', 2: 'height', 3: 'width'},  # shape(1,3,640,640)
+                                    'output': {0: 'batch', 1: 'anchors'}  # shape(1,25200,85)
+                                    } if dynamic else None)
+
+    # Checks
+    model_onnx = onnx.load(f)  # load onnx model
+    onnx.checker.check_model(model_onnx)  # check onnx model
+
+    model_onnx, check = onnxsim.simplify(
+        model_onnx,
+        dynamic_input_shape=dynamic,
+        input_shapes={'images': list(im.shape)} if dynamic else None)
+    assert check, 'assert check failed'
+    onnx.save(model_onnx, f)

utils/general.py

@@ -0,0 +1,65 @@
+
+import os
+import logging
+import wandb
+import torch
+
+def set_logging(name=None, verbose=True):
+    for handler in logging.root.handlers[:]:
+        logging.root.removeHandler(handler)
+    # Sets level and returns logger
+    rank = int(os.getenv('RANK', -1))  # rank in world for Multi-GPU trainings
+    logging.basicConfig(format="%(message)s", level=logging.INFO if (verbose and rank in (-1, 0)) else logging.WARNING)
+    return logging.getLogger(name)
+
+LOGGER = set_logging(__name__)  # define globally (used in train.py, val.py, detect.py, etc.)
+
+LOGGERS = ('csv', 'tb', 'wandb')
+
+CUDA = True if torch.cuda.is_available() else False
+DEVICE = 'cuda' if CUDA else 'cpu'
+
+LOGGER_WANDB = 'wandb'
+LOGGER_TENSORBOARD = 'tb'
+
+
+class Loggers():
+    def __init__(self, hyp):
+        self.type = hyp['logger']['type']
+        self.epochs = hyp['train']['epochs']
+        self.wandb = None
+        self.writer = None
+        if self.type == LOGGER_WANDB:
+            if hyp['logger']['project'] == '':
+                project = 'ComicTextDetector'
+            else:
+                project = hyp['logger']['project']
+            if hyp['logger']['run_id'] == '':
+                self.wandb = wandb.init(project=project, config=hyp, resume='allow')
+            else:
+                self.wandb = wandb.init(project=project, config=hyp, resume='must', id=hyp['logger']['run_id'])
+        elif self.type == LOGGER_TENSORBOARD:
+            from torch.utils.tensorboard import SummaryWriter
+            self.writer = SummaryWriter(hyp['data']['save_dir'])
+
+    def on_train_batch_end(self, metrics):
+        # Callback runs on train batch end
+        if self.wandb:
+            self.wandb.log(metrics)
+        pass
+
+    def on_train_epoch_end(self, epoch, metrics):
+        LOGGER.info(f'fin epoch {epoch}/{self.epochs}, metrics: {metrics}')
+        if self.type == LOGGER_WANDB:
+            self.wandb.log(metrics)
+        elif self.type == LOGGER_TENSORBOARD:
+            for key in metrics.keys():
+                self.writer.add_scalar(key, metrics[key], epoch)
+
+    def on_model_save(self, last, epoch, final_epoch, best_fitness, fi):
+        # Callback runs on model save event
+        if self.wandb:
+            if ((epoch + 1) % self.opt.save_period == 0 and not final_epoch) and self.opt.save_period != -1:
+                self.wandb.log_model(last.parent, self.opt, epoch, fi, best_model=best_fitness == fi)
+
+

utils/imgproc_utils.py

@@ -0,0 +1,192 @@
+import numpy as np
+import cv2
+import random
+
+def hex2bgr(hex):
+    gmask = 254 << 8
+    rmask = 254
+    b = hex >> 16
+    g = (hex & gmask) >> 8
+    r = hex & rmask
+    return np.stack([b, g, r]).transpose()
+
+def union_area(bboxa, bboxb):
+    x1 = max(bboxa[0], bboxb[0])
+    y1 = max(bboxa[1], bboxb[1])
+    x2 = min(bboxa[2], bboxb[2])
+    y2 = min(bboxa[3], bboxb[3])
+    if y2 < y1 or x2 < x1:
+        return -1
+    return (y2 - y1) * (x2 - x1)
+
+def get_yololabel_strings(clslist, labellist):
+    content = ''
+    for cls, xywh in zip(clslist, labellist):
+        content += str(int(cls)) + ' ' + ' '.join([str(e) for e in xywh]) + '\n'
+    if len(content) != 0:
+        content = content[:-1]
+    return content
+
+# 4 points bbox to 8 points polygon
+def xywh2xyxypoly(xywh, to_int=True):
+    xyxypoly = np.tile(xywh[:, [0, 1]], 4)
+    xyxypoly[:, [2, 4]] += xywh[:, [2]]
+    xyxypoly[:, [5, 7]] += xywh[:, [3]]
+    if to_int:
+        xyxypoly = xyxypoly.astype(np.int64)
+    return xyxypoly
+
+def xyxy2yolo(xyxy, w: int, h: int):
+    if xyxy == [] or len(xyxy) == 0:
+        return None
+    if isinstance(xyxy, list):
+        xyxy = np.array(xyxy)
+    if len(xyxy.shape) == 1:
+        xyxy = np.array([xyxy])
+    yolo = np.copy(xyxy).astype(np.float64)
+    yolo[:, [0, 2]] =  yolo[:, [0, 2]] / w
+    yolo[:, [1, 3]] = yolo[:, [1, 3]] / h
+    yolo[:, [2, 3]] -= yolo[:, [0, 1]]
+    yolo[:, [0, 1]] += yolo[:, [2, 3]] / 2
+    return yolo
+
+def yolo_xywh2xyxy(xywh: np.array, w: int, h:  int, to_int=True):
+    if xywh is None:
+        return None
+    if len(xywh) == 0:
+        return None
+    if len(xywh.shape) == 1:
+        xywh = np.array([xywh])
+    xywh[:, [0, 2]] *= w
+    xywh[:, [1, 3]] *= h
+    xywh[:, [0, 1]] -= xywh[:, [2, 3]] / 2
+    xywh[:, [2, 3]] += xywh[:, [0, 1]]
+    if to_int:
+        xywh = xywh.astype(np.int64)
+    return xywh
+
+def rotate_polygons(center, polygons, rotation, new_center=None, to_int=True):
+    if new_center is None:
+        new_center = center
+    rotation = np.deg2rad(rotation)
+    s, c = np.sin(rotation), np.cos(rotation)
+    polygons = polygons.astype(np.float32)
+    
+    polygons[:, 1::2] -= center[1]
+    polygons[:, ::2] -= center[0]
+    rotated = np.copy(polygons)
+    rotated[:, 1::2] = polygons[:, 1::2] * c - polygons[:, ::2] * s
+    rotated[:, ::2] = polygons[:, 1::2] * s + polygons[:, ::2] * c
+    rotated[:, 1::2] += new_center[1]
+    rotated[:, ::2] += new_center[0]
+    if to_int:
+        return rotated.astype(np.int64)
+    return rotated
+
+def letterbox(im, new_shape=(640, 640), color=(0, 0, 0), auto=False, scaleFill=False, scaleup=True, stride=128):
+    # Resize and pad image while meeting stride-multiple constraints
+    shape = im.shape[:2]  # current shape [height, width]
+    if not isinstance(new_shape, tuple):
+        new_shape = (new_shape, new_shape)
+
+    # Scale ratio (new / old)
+    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
+    if not scaleup:  # only scale down, do not scale up (for better val mAP)
+        r = min(r, 1.0)
+
+    # Compute padding
+    ratio = r, r  # width, height ratios
+    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
+    dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1]  # wh padding
+    if auto:  # minimum rectangle
+        dw, dh = np.mod(dw, stride), np.mod(dh, stride)  # wh padding
+    elif scaleFill:  # stretch
+        dw, dh = 0.0, 0.0
+        new_unpad = (new_shape[1], new_shape[0])
+        ratio = new_shape[1] / shape[1], new_shape[0] / shape[0]  # width, height ratios
+
+    # dw /= 2  # divide padding into 2 sides
+    # dh /= 2
+    dh, dw = int(dh), int(dw)
+
+    if shape[::-1] != new_unpad:  # resize
+        im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)
+    top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))
+    left, right = int(round(dw - 0.1)), int(round(dw + 0.1))
+    im = cv2.copyMakeBorder(im, 0, dh, 0, dw, cv2.BORDER_CONSTANT, value=color)  # add border
+    return im, ratio, (dw, dh)
+
+def resize_keepasp(im, new_shape=640, scaleup=True, interpolation=cv2.INTER_LINEAR, stride=None):
+    shape = im.shape[:2]  # current shape [height, width]
+
+    if new_shape is not None:
+        if not isinstance(new_shape, tuple):
+            new_shape = (new_shape, new_shape)
+    else:
+        new_shape = shape
+
+    # Scale ratio (new / old)
+    r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])
+    if not scaleup:  # only scale down, do not scale up (for better val mAP)
+        r = min(r, 1.0)
+
+    new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))
+
+    if stride is not None:
+        h, w = new_unpad
+        if new_shape[0] % stride != 0 :
+            new_h = (stride - (new_shape[0] % stride)) + h
+        else :
+            new_h = h
+        if w % stride != 0 :
+            new_w = (stride - (w % stride)) + w
+        else :
+            new_w = w
+        new_unpad = (new_h, new_w)
+
+    if shape[::-1] != new_unpad:  # resize
+        im = cv2.resize(im, new_unpad, interpolation=interpolation)
+    return im
+
+def expand_textwindow(img_size, xyxy, expand_r=8, shrink=False):
+    im_h, im_w = img_size[:2]
+    x1, y1 , x2, y2 = xyxy
+    w = x2 - x1
+    h = y2 - y1
+    paddings = int(round((max(h, w) * 0.25 + min(h, w) * 0.75) / expand_r))
+    if shrink:
+        paddings *= -1
+    x1, y1 = max(0, x1 - paddings), max(0, y1 - paddings)
+    x2, y2 = min(im_w-1, x2+paddings), min(im_h-1, y2+paddings)
+    return [x1, y1, x2, y2]
+
+def draw_connected_labels(num_labels, labels, stats, centroids, names="draw_connected_labels", skip_background=True):
+    labdraw = np.zeros((labels.shape[0], labels.shape[1], 3), dtype=np.uint8)
+    max_ind = 0
+    if isinstance(num_labels, int):
+        num_labels = range(num_labels)
+    
+    # for ind, lab in enumerate((range(num_labels))):
+    for lab in num_labels:
+        if skip_background and lab == 0:
+            continue
+        randcolor = (random.randint(0,255), random.randint(0,255), random.randint(0,255))
+        labdraw[np.where(labels==lab)] = randcolor
+        maxr, minr = 0.5, 0.001
+        maxw, maxh = stats[max_ind][2] * maxr, stats[max_ind][3] * maxr
+        minarea = labdraw.shape[0] * labdraw.shape[1] * minr
+
+        stat = stats[lab]
+        bboxarea = stat[2] * stat[3]
+        if stat[2] < maxw and stat[3] < maxh and bboxarea > minarea:
+            pix = np.zeros((labels.shape[0], labels.shape[1]), dtype=np.uint8)
+            pix[np.where(labels==lab)] = 255
+
+            rect = cv2.minAreaRect(cv2.findNonZero(pix))
+            box = np.int0(cv2.boxPoints(rect))
+            labdraw = cv2.drawContours(labdraw, [box], 0, randcolor, 2)
+            labdraw = cv2.circle(labdraw, (int(centroids[lab][0]),int(centroids[lab][1])), radius=5, color=(random.randint(0,255), random.randint(0,255), random.randint(0,255)), thickness=-1)                
+
+    cv2.imshow(names, labdraw)
+    return labdraw
+

utils/io_utils.py

@@ -0,0 +1,53 @@
+import os
+import os.path as osp
+import glob
+from pathlib import Path
+import cv2
+import numpy as np
+import json
+
+IMG_EXT = ['.bmp', '.jpg', '.png', '.jpeg']
+
+NP_BOOL_TYPES = (np.bool_, np.bool8)
+NP_FLOAT_TYPES = (np.float_, np.float16, np.float32, np.float64)
+NP_INT_TYPES = (np.int_, np.int8, np.int16, np.int32, np.int64, np.uint, np.uint8, np.uint16, np.uint32, np.uint64)
+
+# https://stackoverflow.com/questions/26646362/numpy-array-is-not-json-serializable
+class NumpyEncoder(json.JSONEncoder):
+    def default(self, obj):
+        if isinstance(obj, np.ndarray):
+            return obj.tolist()
+        elif isinstance(obj, np.ScalarType):
+            if isinstance(obj, NP_BOOL_TYPES):
+                return bool(obj)
+            elif isinstance(obj, NP_FLOAT_TYPES):
+                return float(obj)
+            elif isinstance(obj, NP_INT_TYPES):
+                return int(obj)
+        return json.JSONEncoder.default(self, obj)
+
+def find_all_imgs(img_dir, abs_path=False):
+    imglist = list()
+    for filep in glob.glob(osp.join(img_dir, "*")):
+        filename = osp.basename(filep)
+        file_suffix = Path(filename).suffix
+        if file_suffix.lower() not in IMG_EXT:
+            continue
+        if abs_path:
+            imglist.append(filep)
+        else:
+            imglist.append(filename)
+    return imglist
+
+imread = lambda imgpath, read_type=cv2.IMREAD_COLOR: cv2.imdecode(np.fromfile(imgpath, dtype=np.uint8), read_type)
+# def imread(imgpath, read_type=cv2.IMREAD_COLOR):
+#     img = cv2.imdecode(np.fromfile(imgpath, dtype=np.uint8), read_type)
+#     return img
+
+def imwrite(img_path, img, ext='.png'):
+    suffix = Path(img_path).suffix
+    if suffix != '':
+        img_path = img_path.replace(suffix, ext)
+    else:
+        img_path += ext
+    cv2.imencode(ext, img)[1].tofile(img_path)

utils/loss.py

@@ -0,0 +1,188 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+"""Modified from https://github.com/LikeLy-Journey/SegmenTron/blob/master/
+segmentron/solver/loss.py (Apache-2.0 License)"""
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.nn.modules.loss import BCEWithLogitsLoss
+
+
+class BinaryDiceLoss(nn.Module):
+    """Dice loss of binary class
+    Args:
+        smooth: A float number to smooth loss, and avoid NaN error, default: 1
+        p: Denominator value: \sum{x^p} + \sum{y^p}, default: 2
+        predict: A tensor of shape [N, *]
+        target: A tensor of shape same with predict
+        reduction: Reduction method to apply, return mean over batch if 'mean',
+            return sum if 'sum', return a tensor of shape [N,] if 'none'
+    Returns:
+        Loss tensor according to arg reduction
+    Raise:
+        Exception if unexpected reduction
+    """
+    def __init__(self, smooth=1, p=2, reduction='mean'):
+        super(BinaryDiceLoss, self).__init__()
+        self.smooth = smooth
+        self.p = p
+        self.reduction = reduction
+
+    def forward(self, predict, target):
+        assert predict.shape[0] == target.shape[0], "predict & target batch size don't match"
+        predict = predict.contiguous().view(predict.shape[0], -1)
+        target = target.contiguous().view(target.shape[0], -1)
+
+        num = torch.sum(torch.mul(predict, target), dim=1) + self.smooth
+        den = torch.sum(predict.pow(self.p) + target.pow(self.p), dim=1) + self.smooth
+
+        loss = 1 - num / den
+
+        if self.reduction == 'mean':
+            return loss.mean()
+        elif self.reduction == 'sum':
+            return loss.sum()
+        elif self.reduction == 'none':
+            return loss
+        else:
+            raise Exception('Unexpected reduction {}'.format(self.reduction))
+
+
+class BalanceCrossEntropyLoss(nn.Module):
+    '''
+    Balanced cross entropy loss.
+    Shape:
+        - Input: :math:`(N, 1, H, W)`
+        - GT: :math:`(N, 1, H, W)`, same shape as the input
+        - Mask: :math:`(N, H, W)`, same spatial shape as the input
+        - Output: scalar.
+
+    Examples::
+
+        >>> m = nn.Sigmoid()
+        >>> loss = nn.BCELoss()
+        >>> input = torch.randn(3, requires_grad=True)
+        >>> target = torch.empty(3).random_(2)
+        >>> output = loss(m(input), target)
+        >>> output.backward()
+    '''
+
+    def __init__(self, negative_ratio=3.0, eps=1e-6):
+        super(BalanceCrossEntropyLoss, self).__init__()
+        self.negative_ratio = negative_ratio
+        self.eps = eps
+
+    def forward(self,
+                pred: torch.Tensor,
+                gt: torch.Tensor,
+                mask: torch.Tensor,
+                return_origin=False):
+        '''
+        Args:
+            pred: shape :math:`(N, 1, H, W)`, the prediction of network
+            gt: shape :math:`(N, 1, H, W)`, the target
+            mask: shape :math:`(N, H, W)`, the mask indicates positive regions
+        '''
+        positive = (gt * mask).byte()
+        negative = ((1 - gt) * mask).byte()
+        positive_count = int(positive.float().sum())
+        negative_count = min(int(negative.float().sum()), int(positive_count * self.negative_ratio))
+        # loss = nn.functional.binary_cross_entropy(pred, gt, reduction='none')
+        loss = nn.functional.binary_cross_entropy_with_logits(pred, gt, reduction='none')
+        positive_loss = loss * positive.float()
+        negative_loss = loss * negative.float()
+        # negative_loss, _ = torch.topk(negative_loss.view(-1).contiguous(), negative_count)
+        negative_loss, _ = negative_loss.view(-1).topk(negative_count)
+
+        balance_loss = (positive_loss.sum() + negative_loss.sum()) / (positive_count + negative_count + self.eps)
+
+        if return_origin:
+            return balance_loss, loss
+        return balance_loss
+
+
+class DiceLoss(nn.Module):
+    '''
+    Loss function from https://arxiv.org/abs/1707.03237,
+    where iou computation is introduced heatmap manner to measure the
+    diversity between tow heatmaps.
+    '''
+
+    def __init__(self, eps=1e-6):
+        super(DiceLoss, self).__init__()
+        self.eps = eps
+
+    def forward(self, pred: torch.Tensor, gt, mask, weights=None):
+        '''
+        pred: one or two heatmaps of shape (N, 1, H, W),
+            the losses of tow heatmaps are added together.
+        gt: (N, 1, H, W)
+        mask: (N, H, W)
+        '''
+        return self._compute(pred, gt, mask, weights)
+
+    def _compute(self, pred, gt, mask, weights):
+        if pred.dim() == 4:
+            pred = pred[:, 0, :, :]
+            gt = gt[:, 0, :, :]
+        assert pred.shape == gt.shape
+        assert pred.shape == mask.shape
+        if weights is not None:
+            assert weights.shape == mask.shape
+            mask = weights * mask
+        intersection = (pred * gt * mask).sum()
+
+        union = (pred * mask).sum() + (gt * mask).sum() + self.eps
+        loss = 1 - 2.0 * intersection / union
+        assert loss <= 1
+        return loss
+
+
+class MaskL1Loss(nn.Module):
+    def __init__(self, eps=1e-6):
+        super(MaskL1Loss, self).__init__()
+        self.eps = eps
+
+    def forward(self, pred: torch.Tensor, gt, mask):
+        loss = (torch.abs(pred - gt) * mask).sum() / (mask.sum() + self.eps)
+        return loss
+
+class DBLoss(nn.Module):
+    def __init__(self, alpha=3.0, beta=1.0, ohem_ratio=3, reduction='mean', eps=1e-6):
+        """
+        Implement PSE Loss.
+        :param alpha: binary_map loss 前面的系数
+        :param beta: threshold_map loss 前面的系数
+        :param ohem_ratio: OHEM的比例
+        :param reduction: 'mean' or 'sum'对 batch里的loss 算均值或求和
+        """
+        super().__init__()
+        assert reduction in ['mean', 'sum'], " reduction must in ['mean','sum']"
+        self.alpha = alpha
+        self.beta = beta
+        self.bce_loss = BalanceCrossEntropyLoss(negative_ratio=ohem_ratio)
+        self.dice_loss = DiceLoss(eps=eps)
+        self.l1_loss = MaskL1Loss(eps=eps)
+        self.ohem_ratio = ohem_ratio
+        self.reduction = reduction
+
+    def forward(self, pred, batch, use_bce=True):
+        shrink_maps = pred[:, 0, :, :]
+        threshold_maps = pred[:, 1, :, :]
+        binary_maps = pred[:, 2, :, :]
+
+        if use_bce:
+            loss_shrink_maps = self.bce_loss(pred[:, 3, :, :], batch['shrink_map'], batch['shrink_mask']) + self.dice_loss(shrink_maps, batch['shrink_map'], batch['shrink_mask'])
+        else:
+            loss_shrink_maps = self.dice_loss(shrink_maps, batch['shrink_map'], batch['shrink_mask'])
+            
+        loss_threshold_maps = self.l1_loss(threshold_maps, batch['threshold_map'], batch['threshold_mask'])
+        metrics = dict(loss_shrink_maps=loss_shrink_maps, loss_threshold_maps=loss_threshold_maps)
+        if pred.size()[1] > 2:
+            loss_binary_maps = self.dice_loss(binary_maps, batch['shrink_map'], batch['shrink_mask']) + self.bce_loss(binary_maps, batch['shrink_map'], batch['shrink_mask'])
+            metrics['loss_binary_maps'] = loss_binary_maps
+            loss_all = self.alpha * loss_shrink_maps + self.beta * loss_threshold_maps + loss_binary_maps
+            metrics['loss'] = loss_all
+        else:
+            metrics['loss'] = loss_shrink_maps
+        return metrics
+

utils/textblock.py

@@ -0,0 +1,526 @@
+from typing import List
+import numpy as np
+from shapely.geometry import Polygon
+import math
+import copy
+from utils.imgproc_utils import union_area, xywh2xyxypoly, rotate_polygons
+import cv2
+
+LANG_LIST = ['eng', 'ja', 'unknown']
+LANGCLS2IDX = {'eng': 0, 'ja': 1, 'unknown': 2}
+
+class TextBlock(object):
+    def __init__(self, xyxy: List, 
+                       lines: List = None, 
+                       language: str = 'unknown',
+                       vertical: bool = False, 
+                       font_size: float = -1,
+                       distance: List = None,
+                       angle: int = 0,
+                       vec: List = None,
+                       norm: float = -1,
+                       merged: bool = False,
+                       weight: float = -1,
+                       text: List = None,
+                       translation: str = "",
+                       fg_r = 0,
+                       fg_g = 0,
+                       fg_b = 0,
+                       bg_r = 0,
+                       bg_g = 0,
+                       bg_b = 0,                
+                       line_spacing = 1.,
+                       font_family: str = "",
+                       bold: bool = False,
+                       underline: bool = False,
+                       italic: bool = False,
+                       alignment: int = -1,
+                       alpha: float = 255,
+                       rich_text: str = "",
+                       _bounding_rect: List = None,
+                       accumulate_color = True,
+                       default_stroke_width = 0.2,
+                       target_lang: str = "",
+                       **kwargs) -> None:
+        self.xyxy = [int(num) for num in xyxy]                    # boundingbox of textblock
+        self.lines = [] if lines is None else lines     # polygons of textlines
+        self.vertical = vertical            # orientation of textlines
+        self.language = language
+        self.font_size = font_size          # font pixel size
+        self.distance = None if distance is None else np.array(distance, np.float64)   # distance between textlines and "origin"          
+        self.angle = angle                  # rotation angle of textlines
+
+        self.vec = None if vec is None else np.array(vec, np.float64) # primary vector of textblock
+        self.norm = norm                    # primary norm of textblock
+        self.merged = merged
+        self.weight = weight
+
+        self.text = text if text is not None else []
+        self.prob = 1
+
+        self.translation = translation
+
+        # note they're accumulative rgb values of textlines
+        self.fg_r = fg_r                       
+        self.fg_g = fg_g
+        self.fg_b = fg_b
+        self.bg_r = bg_r
+        self.bg_g = bg_g
+        self.bg_b = bg_b
+
+        # self.stroke_width = stroke_width
+        self.font_family: str = font_family
+        self.bold: bool = bold
+        self.underline: bool = underline
+        self.italic: bool = italic
+        self.alpha = alpha
+        self.rich_text = rich_text
+        self.line_spacing = line_spacing
+        # self.alignment = alignment
+        self._alignment = alignment
+        self._target_lang = target_lang
+
+        self._bounding_rect = _bounding_rect
+        self.default_stroke_width = default_stroke_width
+        self.accumulate_color = accumulate_color
+
+    def adjust_bbox(self, with_bbox=False):
+        lines = self.lines_array().astype(np.int32)
+        if with_bbox:
+            self.xyxy[0] = min(lines[..., 0].min(), self.xyxy[0])
+            self.xyxy[1] = min(lines[..., 1].min(), self.xyxy[1])
+            self.xyxy[2] = max(lines[..., 0].max(), self.xyxy[2])
+            self.xyxy[3] = max(lines[..., 1].max(), self.xyxy[3])
+        else:
+            self.xyxy[0] = lines[..., 0].min()
+            self.xyxy[1] = lines[..., 1].min()
+            self.xyxy[2] = lines[..., 0].max()
+            self.xyxy[3] = lines[..., 1].max()
+
+    def sort_lines(self):
+        if self.distance is not None:
+            idx = np.argsort(self.distance)
+            self.distance = self.distance[idx]
+            lines = np.array(self.lines, dtype=np.int32)
+            self.lines = lines[idx].tolist()
+
+    def lines_array(self, dtype=np.float64):
+        return np.array(self.lines, dtype=dtype)
+
+    def aspect_ratio(self) -> float:
+        min_rect = self.min_rect()
+        middle_pnts = (min_rect[:, [1, 2, 3, 0]] + min_rect) / 2
+        norm_v = np.linalg.norm(middle_pnts[:, 2] - middle_pnts[:, 0])
+        norm_h = np.linalg.norm(middle_pnts[:, 1] - middle_pnts[:, 3])
+        return norm_v / norm_h
+
+    def center(self):
+        xyxy = np.array(self.xyxy)
+        return (xyxy[:2] + xyxy[2:]) / 2
+    
+    def min_rect(self, rotate_back=True):
+        angled = self.angle != 0
+        center = self.center()
+        polygons = self.lines_array().reshape(-1, 8)
+        if angled:
+            polygons = rotate_polygons(center, polygons, self.angle)
+        min_x = polygons[:, ::2].min()
+        min_y = polygons[:, 1::2].min()
+        max_x = polygons[:, ::2].max()
+        max_y = polygons[:, 1::2].max()
+        min_bbox = np.array([[min_x, min_y, max_x, min_y, max_x, max_y, min_x, max_y]])
+        if angled and rotate_back:
+            min_bbox = rotate_polygons(center, min_bbox, -self.angle)
+        return min_bbox.reshape(-1, 4, 2).astype(np.int64)
+
+    # equivalent to qt's boundingRect, ignore angle
+    def bounding_rect(self):
+        if self._bounding_rect is None:
+        # if True:
+            min_bbox = self.min_rect(rotate_back=False)[0]
+            x, y = min_bbox[0]
+            w, h = min_bbox[2] - min_bbox[0]
+            return [x, y, w, h]
+        return self._bounding_rect
+
+    def __getattribute__(self, name: str):
+        if name == 'pts':
+            return self.lines_array()
+        # else:
+        return object.__getattribute__(self, name)
+
+    def __len__(self):
+        return len(self.lines)
+
+    def __getitem__(self, idx):
+        return self.lines[idx]
+
+    def to_dict(self):
+        blk_dict = copy.deepcopy(vars(self))
+        return blk_dict
+
+    def get_transformed_region(self, img, idx, textheight) -> np.ndarray :
+        im_h, im_w = img.shape[:2]
+        direction = 'v' if self.vertical else 'h'
+        src_pts = np.array(self.lines[idx], dtype=np.float64)
+
+        if self.language == 'eng' or (self.language == 'unknown' and not self.vertical):
+            e_size = self.font_size / 3
+            src_pts[..., 0] += np.array([-e_size, e_size, e_size, -e_size])
+            src_pts[..., 1] += np.array([-e_size, -e_size, e_size, e_size])
+            src_pts[..., 0] = np.clip(src_pts[..., 0], 0, im_w)
+            src_pts[..., 1] = np.clip(src_pts[..., 1], 0, im_h)
+
+        middle_pnt = (src_pts[[1, 2, 3, 0]] + src_pts) / 2
+        vec_v = middle_pnt[2] - middle_pnt[0]   # vertical vectors of textlines
+        vec_h = middle_pnt[1] - middle_pnt[3]   # horizontal vectors of textlines
+        ratio = np.linalg.norm(vec_v) / np.linalg.norm(vec_h)
+
+        if direction == 'h' :
+            h = int(textheight)
+            w = int(round(textheight / ratio))
+            dst_pts = np.array([[0, 0], [w - 1, 0], [w - 1, h - 1], [0, h - 1]]).astype(np.float32)
+            M, _ = cv2.findHomography(src_pts, dst_pts, cv2.RANSAC, 5.0)
+            region = cv2.warpPerspective(img, M, (w, h))
+        elif direction == 'v' :
+            w = int(textheight)
+            h = int(round(textheight * ratio))
+            dst_pts = np.array([[0, 0], [w - 1, 0], [w - 1, h - 1], [0, h - 1]]).astype(np.float32)
+            M, _ = cv2.findHomography(src_pts, dst_pts, cv2.RANSAC, 5.0)
+            region = cv2.warpPerspective(img, M, (w, h))
+            region = cv2.rotate(region, cv2.ROTATE_90_COUNTERCLOCKWISE)
+        # cv2.imshow('region'+str(idx), region)
+        # cv2.waitKey(0)
+        return region
+
+    def get_text(self):
+        if isinstance(self.text, str):
+            return self.text
+        return ' '.join(self.text).strip()
+
+    def set_font_colors(self, frgb, srgb, accumulate=True):
+        self.accumulate_color = accumulate
+        num_lines = len(self.lines) if accumulate and len(self.lines) > 0 else 1
+        # set font color
+        frgb = np.array(frgb) * num_lines
+        self.fg_r, self.fg_g, self.fg_b = frgb
+        # set stroke color  
+        srgb = np.array(srgb) * num_lines
+        self.bg_r, self.bg_g, self.bg_b = srgb
+
+    def get_font_colors(self, bgr=False):
+        num_lines = len(self.lines)
+        frgb = np.array([self.fg_r, self.fg_g, self.fg_b])
+        brgb = np.array([self.bg_r, self.bg_g, self.bg_b])
+        if self.accumulate_color:
+            if num_lines > 0:
+                frgb = (frgb / num_lines).astype(np.int32)
+                brgb = (brgb / num_lines).astype(np.int32)
+                if bgr:
+                    return frgb[::-1], brgb[::-1]
+                else:
+                    return frgb, brgb
+            else:
+                return [0, 0, 0], [0, 0, 0]
+        else:
+            return frgb, brgb
+
+    def xywh(self):
+        x, y, w, h = self.xyxy
+        return [x, y, w-x, h-y]
+
+    # alignleft: 0, center: 1, right: 2 
+    def alignment(self):
+        if self._alignment >= 0:
+            return self._alignment
+        elif self.vertical:
+            return 0
+        lines = self.lines_array()
+        if len(lines) == 1:
+            return 0
+        angled = self.angle != 0
+        polygons = lines.reshape(-1, 8)
+        if angled:
+            polygons = rotate_polygons((0, 0), polygons, self.angle)
+        polygons = polygons.reshape(-1, 4, 2)
+        
+        left_std = np.std(polygons[:, 0, 0])
+        # right_std = np.std(polygons[:, 1, 0])
+        center_std = np.std((polygons[:, 0, 0] + polygons[:, 1, 0]) / 2)
+        if left_std < center_std:
+            return 0
+        else:
+            return 1
+
+    def target_lang(self):
+        return self.target_lang
+
+    @property
+    def stroke_width(self):
+        var = np.array([self.fg_r, self.fg_g, self.fg_b]) \
+            - np.array([self.bg_r, self.bg_g, self.bg_b])
+        var = np.abs(var).sum()
+        if var > 40:
+            return self.default_stroke_width
+        return 0
+
+def sort_textblk_list(blk_list: List[TextBlock], im_w: int, im_h: int) -> List[TextBlock]:
+    if len(blk_list) == 0:
+        return blk_list
+    num_ja = 0
+    xyxy = []
+    for blk in blk_list:
+        if blk.language == 'ja':
+            num_ja += 1
+        xyxy.append(blk.xyxy)
+    xyxy = np.array(xyxy)
+    flip_lr = num_ja > len(blk_list) / 2
+    im_oriw = im_w
+    if im_w > im_h:
+        im_w /= 2
+    num_gridy, num_gridx = 4, 3
+    img_area = im_h * im_w
+    center_x = (xyxy[:, 0] + xyxy[:, 2]) / 2
+    if flip_lr:
+        if im_w != im_oriw:
+            center_x = im_oriw - center_x
+        else:
+            center_x = im_w - center_x
+    grid_x = (center_x / im_w * num_gridx).astype(np.int32)
+    center_y = (xyxy[:, 1] + xyxy[:, 3]) / 2
+    grid_y = (center_y / im_h * num_gridy).astype(np.int32)
+    grid_indices = grid_y * num_gridx + grid_x
+    grid_weights = grid_indices * img_area + 1.2 * (center_x - grid_x * im_w / num_gridx) + (center_y - grid_y * im_h / num_gridy)
+    if im_w != im_oriw:
+        grid_weights[np.where(grid_x >= num_gridx)] += img_area * num_gridy * num_gridx
+    
+    for blk, weight in zip(blk_list, grid_weights):
+        blk.weight = weight
+    blk_list.sort(key=lambda blk: blk.weight)
+    return blk_list
+
+def examine_textblk(blk: TextBlock, im_w: int, im_h: int, sort: bool = False) -> None:
+    lines = blk.lines_array()
+    middle_pnts = (lines[:, [1, 2, 3, 0]] + lines) / 2
+    vec_v = middle_pnts[:, 2] - middle_pnts[:, 0]   # vertical vectors of textlines
+    vec_h = middle_pnts[:, 1] - middle_pnts[:, 3]   # horizontal vectors of textlines
+    # if sum of vertical vectors is longer, then text orientation is vertical, and vice versa.
+    center_pnts = (lines[:, 0] + lines[:, 2]) / 2
+    v = np.sum(vec_v, axis=0)
+    h = np.sum(vec_h, axis=0)
+    norm_v, norm_h = np.linalg.norm(v), np.linalg.norm(h)
+    if blk.language == 'ja':
+        vertical = norm_v > norm_h
+    else:
+        vertical = norm_v > norm_h * 2
+    # calculate distance between textlines and origin 
+    if vertical:
+        primary_vec, primary_norm = v, norm_v
+        distance_vectors = center_pnts - np.array([[im_w, 0]], dtype=np.float64)   # vertical manga text is read from right to left, so origin is (imw, 0)
+        font_size = int(round(norm_h / len(lines)))
+    else:
+        primary_vec, primary_norm = h, norm_h
+        distance_vectors = center_pnts - np.array([[0, 0]], dtype=np.float64)
+        font_size = int(round(norm_v / len(lines)))
+    
+    rotation_angle = int(math.atan2(primary_vec[1], primary_vec[0]) / math.pi * 180)     # rotation angle of textlines
+    distance = np.linalg.norm(distance_vectors, axis=1)     # distance between textlinecenters and origin
+    rad_matrix = np.arccos(np.einsum('ij, j->i', distance_vectors, primary_vec) / (distance * primary_norm))
+    distance = np.abs(np.sin(rad_matrix) * distance)
+    blk.lines = lines.astype(np.int32).tolist()
+    blk.distance = distance
+    blk.angle = rotation_angle
+    if vertical:
+        blk.angle -= 90
+    if abs(blk.angle) < 3:
+        blk.angle = 0
+    blk.font_size = font_size
+    blk.vertical = vertical
+    blk.vec = primary_vec
+    blk.norm = primary_norm
+    if sort:
+        blk.sort_lines()
+
+def try_merge_textline(blk: TextBlock, blk2: TextBlock, fntsize_tol=1.3, distance_tol=2) -> bool:
+    if blk2.merged:
+        return False
+    fntsize_div = blk.font_size / blk2.font_size
+    num_l1, num_l2 = len(blk), len(blk2)
+    fntsz_avg = (blk.font_size * num_l1 + blk2.font_size * num_l2) / (num_l1 + num_l2)
+    vec_prod = blk.vec @ blk2.vec
+    vec_sum = blk.vec + blk2.vec
+    cos_vec = vec_prod / blk.norm / blk2.norm
+    distance = blk2.distance[-1] - blk.distance[-1]
+    distance_p1 = np.linalg.norm(np.array(blk2.lines[-1][0]) - np.array(blk.lines[-1][0]))
+    l1, l2 = Polygon(blk.lines[-1]), Polygon(blk2.lines[-1])
+    if not l1.intersects(l2):
+        if fntsize_div > fntsize_tol or 1 / fntsize_div > fntsize_tol:
+            return False
+        if abs(cos_vec) < 0.866:   # cos30
+            return False
+        if distance > distance_tol * fntsz_avg or distance_p1 > fntsz_avg * 2.5:
+            return False
+    # merge
+    blk.lines.append(blk2.lines[0])
+    blk.vec = vec_sum
+    blk.angle = int(round(np.rad2deg(math.atan2(vec_sum[1], vec_sum[0]))))
+    if blk.vertical:
+        blk.angle -= 90
+    blk.norm = np.linalg.norm(vec_sum)
+    blk.distance = np.append(blk.distance, blk2.distance[-1])
+    blk.font_size = fntsz_avg
+    blk2.merged = True
+    return True
+
+def merge_textlines(blk_list: List[TextBlock]) -> List[TextBlock]:
+    if len(blk_list) < 2:
+        return blk_list
+    blk_list.sort(key=lambda blk: blk.distance[0])
+    merged_list = []
+    for ii, current_blk in enumerate(blk_list):
+        if current_blk.merged:
+            continue
+        for jj, blk in enumerate(blk_list[ii+1:]):
+            try_merge_textline(current_blk, blk)
+        merged_list.append(current_blk)
+    for blk in merged_list:
+        blk.adjust_bbox(with_bbox=False)
+    return merged_list
+
+def split_textblk(blk: TextBlock):
+    font_size, distance, lines = blk.font_size, blk.distance, blk.lines
+    l0 = np.array(blk.lines[0])
+    lines.sort(key=lambda line: np.linalg.norm(np.array(line[0]) - l0[0]))
+    distance_tol = font_size * 2
+    current_blk = copy.deepcopy(blk)
+    current_blk.lines = [l0]
+    sub_blk_list = [current_blk]
+    textblock_splitted = False
+    for jj, line in enumerate(lines[1:]):
+        l1, l2 = Polygon(lines[jj]), Polygon(line)
+        split = False
+        if not l1.intersects(l2):
+            line_disance = abs(distance[jj+1] - distance[jj])
+            if line_disance > distance_tol:
+                split = True
+            elif blk.vertical and abs(blk.angle) < 15:
+                if len(current_blk.lines) > 1 or line_disance > font_size:
+                    split = abs(lines[jj][0][1] - line[0][1]) > font_size
+        if split:
+            current_blk = copy.deepcopy(current_blk)
+            current_blk.lines = [line]
+            sub_blk_list.append(current_blk)
+        else:
+            current_blk.lines.append(line)
+    if len(sub_blk_list) > 1:
+        textblock_splitted = True
+        for current_blk in sub_blk_list:
+            current_blk.adjust_bbox(with_bbox=False)
+    return textblock_splitted, sub_blk_list
+
+def group_output(blks, lines, im_w, im_h, mask=None, sort_blklist=True) -> List[TextBlock]:
+    blk_list: List[TextBlock] = []
+    scattered_lines = {'ver': [], 'hor': []}
+    for bbox, cls, conf in zip(*blks):
+        # cls could give wrong result
+        blk_list.append(TextBlock(bbox, language=LANG_LIST[cls]))
+
+    # step1: filter & assign lines to textblocks
+    bbox_score_thresh = 0.4
+    mask_score_thresh = 0.1
+    for ii, line in enumerate(lines):
+        bx1, bx2 = line[:, 0].min(), line[:, 0].max()
+        by1, by2 = line[:, 1].min(), line[:, 1].max()
+        bbox_score, bbox_idx = -1, -1
+        line_area = (by2-by1) * (bx2-bx1)
+        for jj, blk in enumerate(blk_list):
+            score = union_area(blk.xyxy, [bx1, by1, bx2, by2]) / line_area
+            if bbox_score < score:
+                bbox_score = score
+                bbox_idx = jj
+        if bbox_score > bbox_score_thresh:
+            blk_list[bbox_idx].lines.append(line)
+        else:   # if no textblock was assigned, check whether there is "enough" textmask
+            if mask is not None:
+                mask_score = mask[by1: by2, bx1: bx2].mean() / 255
+                if mask_score < mask_score_thresh:
+                    continue
+            blk = TextBlock([bx1, by1, bx2, by2], [line])
+            examine_textblk(blk, im_w, im_h, sort=False)
+            if blk.vertical:
+                scattered_lines['ver'].append(blk)
+            else:
+                scattered_lines['hor'].append(blk)
+
+    # step2: filter textblocks, sort & split textlines
+    final_blk_list = []
+    for blk in blk_list:
+        # filter textblocks 
+        if len(blk.lines) == 0:
+            bx1, by1, bx2, by2 = blk.xyxy
+            if mask is not None:
+                mask_score = mask[by1: by2, bx1: bx2].mean() / 255
+                if mask_score < mask_score_thresh:
+                    continue
+            xywh = np.array([[bx1, by1, bx2-bx1, by2-by1]])
+            blk.lines = xywh2xyxypoly(xywh).reshape(-1, 4, 2).tolist()
+        examine_textblk(blk, im_w, im_h, sort=True)
+        
+        # split manga text if there is a distance gap
+        textblock_splitted = False
+        if len(blk.lines) > 1:
+            if blk.language == 'ja':
+                textblock_splitted = True
+            elif blk.vertical:
+                textblock_splitted = True
+        if textblock_splitted:
+            textblock_splitted, sub_blk_list = split_textblk(blk)
+        else:
+            sub_blk_list = [blk]
+        # modify textblock to fit its textlines
+        if not textblock_splitted:
+            for blk in sub_blk_list:
+                blk.adjust_bbox(with_bbox=True)
+        final_blk_list += sub_blk_list
+
+    # step3: merge scattered lines, sort textblocks by "grid"
+    final_blk_list += merge_textlines(scattered_lines['hor'])
+    final_blk_list += merge_textlines(scattered_lines['ver'])
+    if sort_blklist:
+        final_blk_list = sort_textblk_list(final_blk_list, im_w, im_h)
+
+    for blk in final_blk_list:
+        if blk.language == 'eng' and not blk.vertical:
+            num_lines = len(blk.lines)
+            if num_lines == 0:
+                continue
+            # blk.line_spacing = blk.bounding_rect()[3] / num_lines / blk.font_size
+            expand_size = max(int(blk.font_size * 0.1), 2)
+            rad = np.deg2rad(blk.angle)
+            shifted_vec = np.array([[[-1, -1],[1, -1],[1, 1],[-1, 1]]])
+            shifted_vec = shifted_vec * np.array([[[np.sin(rad), np.cos(rad)]]]) * expand_size
+            lines = blk.lines_array() + shifted_vec
+            lines[..., 0] = np.clip(lines[..., 0], 0, im_w-1)
+            lines[..., 1] = np.clip(lines[..., 1], 0, im_h-1)
+            blk.lines = lines.astype(np.int64).tolist()
+            blk.font_size += expand_size
+            
+    return final_blk_list
+
+def visualize_textblocks(canvas, blk_list:  List[TextBlock], path = '../output/'):
+    lw = max(round(sum(canvas.shape) / 2 * 0.003), 2)  # line width
+    for ii, blk in enumerate(blk_list):
+        bx1, by1, bx2, by2 = blk.xyxy
+        cv2.rectangle(canvas, (bx1, by1), (bx2, by2), (127, 255, 127), lw)
+        cut_img = canvas[by1:by2, bx1:bx2]
+        cv2.imwrite(path + f'/cut_image_{ii}.png', cut_img)
+        lines = blk.lines_array(dtype=np.int32)
+        for jj, line in enumerate(lines):
+            cv2.putText(canvas, str(jj), line[0], cv2.FONT_HERSHEY_SIMPLEX, 0.7, (255,127,0), 1)
+            cv2.polylines(canvas, [line], True, (0,127,255), 2)
+        cv2.polylines(canvas, [blk.min_rect()], True, (127,127,0), 2)
+        center = [int((bx1 + bx2)/2), int((by1 + by2)/2)]
+        cv2.putText(canvas, str(blk.angle), center, cv2.FONT_HERSHEY_SIMPLEX, 1, (127,127,255), 2)
+        cv2.putText(canvas, str(ii), (bx1, by1 + lw + 2), 0, lw / 3, (255,127,127), max(lw-1, 1), cv2.LINE_AA)
+    return canvas
+

utils/textmask.py

@@ -0,0 +1,267 @@
+from os import stat
+from typing import List
+import cv2
+import numpy as np
+from .textblock import TextBlock
+from .imgproc_utils import draw_connected_labels, expand_textwindow, union_area
+
+WHITE = (255, 255, 255)
+BLACK = (0, 0, 0)
+LANG_ENG = 0
+LANG_JPN = 1
+
+REFINEMASK_INPAINT = 0
+REFINEMASK_ANNOTATION = 1
+
+def get_topk_color(color_list, bins, k=3, color_var=10, bin_tol=0.001):
+    idx = np.argsort(bins * -1)
+    color_list, bins = color_list[idx], bins[idx]
+    top_colors = [color_list[0]]
+    bin_tol = np.sum(bins) * bin_tol
+    if len(color_list) > 1:
+        for color, bin in zip(color_list[1:], bins[1:]):
+            if np.abs(np.array(top_colors) - color).min() > color_var:
+                top_colors.append(color)
+            if len(top_colors) >= k or bin < bin_tol:
+                break
+    return top_colors
+
+def minxor_thresh(threshed, mask, dilate=False):
+    neg_threshed = 255 - threshed
+    e_size = 1
+    if dilate:
+        element = cv2.getStructuringElement(cv2.MORPH_RECT, (2 * e_size + 1, 2 * e_size + 1),(e_size, e_size))
+        neg_threshed = cv2.dilate(neg_threshed, element, iterations=1)
+        threshed = cv2.dilate(threshed, element, iterations=1)
+    neg_xor_sum = cv2.bitwise_xor(neg_threshed, mask).sum()
+    xor_sum = cv2.bitwise_xor(threshed, mask).sum()
+    if neg_xor_sum < xor_sum:
+        return neg_threshed, neg_xor_sum
+    else:
+        return threshed, xor_sum
+
+def get_otsuthresh_masklist(img, pred_mask, per_channel=False) -> List[np.ndarray]:
+    channels = [img[..., 0], img[..., 1], img[..., 2]]
+    mask_list = []
+    for c in channels:
+        _, threshed = cv2.threshold(c, 1, 255, cv2.THRESH_OTSU+cv2.THRESH_BINARY)
+        threshed, xor_sum = minxor_thresh(threshed, pred_mask, dilate=False)
+        mask_list.append([threshed, xor_sum])
+    mask_list.sort(key=lambda x: x[1])
+    if per_channel:
+        return mask_list
+    else:
+        return [mask_list[0]]
+
+def get_topk_masklist(im_grey, pred_mask):
+    if len(im_grey.shape) == 3 and im_grey.shape[-1] == 3:
+        im_grey = cv2.cvtColor(im_grey, cv2.COLOR_BGR2GRAY)
+    msk = np.ascontiguousarray(pred_mask)
+    candidate_grey_px = im_grey[np.where(cv2.erode(msk, np.ones((3,3), np.uint8), iterations=1) > 127)]
+    bin, his = np.histogram(candidate_grey_px, bins=255)
+    topk_color = get_topk_color(his, bin, color_var=10, k=3)
+    color_range = 30
+    mask_list = list()
+    for ii, color in enumerate(topk_color):
+        c_top = min(color+color_range, 255)
+        c_bottom = c_top - 2 * color_range
+        threshed = cv2.inRange(im_grey, c_bottom, c_top)
+        threshed, xor_sum = minxor_thresh(threshed, msk)
+        mask_list.append([threshed, xor_sum])
+    return mask_list
+
+def merge_mask_list(mask_list, pred_mask, blk: TextBlock = None, pred_thresh=30, text_window=None, filter_with_lines=False, refine_mode=REFINEMASK_INPAINT):
+    mask_list.sort(key=lambda x: x[1])
+    linemask = None
+    if blk is not None and filter_with_lines:
+        linemask = np.zeros_like(pred_mask)
+        lines = blk.lines_array(dtype=np.int64)
+        for line in lines:
+            line[..., 0] -= text_window[0]
+            line[..., 1] -= text_window[1]
+            cv2.fillPoly(linemask, [line], 255)
+        linemask = cv2.dilate(linemask, np.ones((3, 3), np.uint8), iterations=3)
+    
+    if pred_thresh > 0:
+        e_size = 1
+        element = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (2 * e_size + 1, 2 * e_size + 1),(e_size, e_size))      
+        pred_mask = cv2.erode(pred_mask, element, iterations=1)
+        _, pred_mask = cv2.threshold(pred_mask, 60, 255, cv2.THRESH_BINARY)
+    connectivity = 8
+    mask_merged = np.zeros_like(pred_mask)
+    for ii, (candidate_mask, xor_sum) in enumerate(mask_list):
+        num_labels, labels, stats, centroids = cv2.connectedComponentsWithStats(candidate_mask, connectivity, cv2.CV_16U)
+        for label_index, stat, centroid in zip(range(num_labels), stats, centroids):
+            if label_index != 0: # skip background label
+                x, y, w, h, area = stat
+                if w * h < 3:
+                    continue
+                x1, y1, x2, y2 = x, y, x+w, y+h
+                label_local = labels[y1: y2, x1: x2]
+                label_coordinates = np.where(label_local==label_index)
+                tmp_merged = np.zeros_like(label_local, np.uint8)
+                tmp_merged[label_coordinates] = 255
+                tmp_merged = cv2.bitwise_or(mask_merged[y1: y2, x1: x2], tmp_merged)
+                xor_merged = cv2.bitwise_xor(tmp_merged, pred_mask[y1: y2, x1: x2]).sum()
+                xor_origin = cv2.bitwise_xor(mask_merged[y1: y2, x1: x2], pred_mask[y1: y2, x1: x2]).sum()
+                if xor_merged < xor_origin:
+                    mask_merged[y1: y2, x1: x2] = tmp_merged
+
+    if refine_mode == REFINEMASK_INPAINT:
+        mask_merged = cv2.dilate(mask_merged, np.ones((3, 3), np.uint8), iterations=1)
+    # fill holes
+    num_labels, labels, stats, centroids = cv2.connectedComponentsWithStats(255-mask_merged, connectivity, cv2.CV_16U)
+    sorted_area = np.sort(stats[:, -1])
+    if len(sorted_area) > 1:
+        area_thresh = sorted_area[-2]
+    else:
+        area_thresh = sorted_area[-1]
+    for label_index, stat, centroid in zip(range(num_labels), stats, centroids):
+        x, y, w, h, area = stat
+        if area < area_thresh:
+            x1, y1, x2, y2 = x, y, x+w, y+h
+            label_local = labels[y1: y2, x1: x2]
+            label_coordinates = np.where(label_local==label_index)
+            tmp_merged = np.zeros_like(label_local, np.uint8)
+            tmp_merged[label_coordinates] = 255
+            tmp_merged = cv2.bitwise_or(mask_merged[y1: y2, x1: x2], tmp_merged)
+            xor_merged = cv2.bitwise_xor(tmp_merged, pred_mask[y1: y2, x1: x2]).sum()
+            xor_origin = cv2.bitwise_xor(mask_merged[y1: y2, x1: x2], pred_mask[y1: y2, x1: x2]).sum()
+            if xor_merged < xor_origin:
+                mask_merged[y1: y2, x1: x2] = tmp_merged
+    return mask_merged
+
+
+def refine_undetected_mask(img: np.ndarray, mask_pred: np.ndarray, mask_refined: np.ndarray, blk_list: List[TextBlock], refine_mode=REFINEMASK_INPAINT):
+    mask_pred[np.where(mask_refined > 30)] = 0
+    _, pred_mask_t = cv2.threshold(mask_pred, 30, 255, cv2.THRESH_BINARY)
+    num_labels, labels, stats, centroids = cv2.connectedComponentsWithStats(pred_mask_t, 4, cv2.CV_16U)
+    valid_labels = np.where(stats[:, -1] > 50)[0]
+    seg_blk_list = []
+    if len(valid_labels) > 0:
+        for lab_index in valid_labels[1:]:
+            x, y, w, h, area = stats[lab_index]
+            bx1, by1 = x, y
+            bx2, by2 = x+w, y+h
+            bbox = [bx1, by1, bx2, by2]
+            bbox_score = -1
+            for blk in blk_list:
+                bbox_s = union_area(blk.xyxy, bbox)
+                if bbox_s > bbox_score:
+                    bbox_score = bbox_s
+            if bbox_score / w / h < 0.5:
+                seg_blk_list.append(TextBlock(bbox))
+    if len(seg_blk_list) > 0:
+        mask_refined = cv2.bitwise_or(mask_refined, refine_mask(img, mask_pred, seg_blk_list, refine_mode=refine_mode))
+    return mask_refined
+
+
+def refine_mask(img: np.ndarray, pred_mask: np.ndarray, blk_list: List[TextBlock], refine_mode: int = REFINEMASK_INPAINT) -> np.ndarray:
+    mask_refined = np.zeros_like(pred_mask)
+    for blk in blk_list:
+        bx1, by1, bx2, by2 = expand_textwindow(img.shape, blk.xyxy, expand_r=16)
+        im = np.ascontiguousarray(img[by1: by2, bx1: bx2])
+        msk = np.ascontiguousarray(pred_mask[by1: by2, bx1: bx2])
+        mask_list = get_topk_masklist(im, msk)
+        mask_list += get_otsuthresh_masklist(im, msk, per_channel=False)
+        mask_merged = merge_mask_list(mask_list, msk, blk=blk, text_window=[bx1, by1, bx2, by2], refine_mode=refine_mode)
+        mask_refined[by1: by2, bx1: bx2] = cv2.bitwise_or(mask_refined[by1: by2, bx1: bx2], mask_merged)
+    return mask_refined
+
+# def extract_textballoon(img, pred_textmsk=None, global_mask=None):
+#     if len(img.shape) > 2 and img.shape[2] == 3:
+#         img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
+#     im_h, im_w = img.shape[0], img.shape[1]
+#     hyp_textmsk = np.zeros((im_h, im_w), np.uint8)
+#     thresh_val, threshed = cv2.threshold(img, 1, 255, cv2.THRESH_OTSU+cv2.THRESH_BINARY)
+#     xormap_sum = cv2.bitwise_xor(threshed, pred_textmsk).sum()
+#     neg_threshed = 255 - threshed
+#     neg_xormap_sum = cv2.bitwise_xor(neg_threshed, pred_textmsk).sum()
+#     neg_thresh = neg_xormap_sum < xormap_sum
+#     if neg_thresh:
+#         threshed = neg_threshed
+#     thresh_info = {'thresh_val': thresh_val,'neg_thresh': neg_thresh}
+#     connectivity = 8
+#     num_labels, labels, stats, centroids = cv2.connectedComponentsWithStats(threshed, connectivity, cv2.CV_16U)
+#     label_unchanged = np.copy(labels)
+#     if global_mask is not None:
+#         labels[np.where(global_mask==0)] = 0
+#     text_labels = []
+#     if pred_textmsk is not None:
+#         text_score_thresh = 0.5
+#         textbbox_map = np.zeros_like(pred_textmsk)
+#         for label_index, stat, centroid in zip(range(num_labels), stats, centroids):
+#             if label_index != 0: # skip background label
+#                 x, y, w, h, area = stat
+#                 area *= 255
+#                 x1, y1, x2, y2 = x, y, x+w, y+h
+#                 label_local = labels[y1: y2, x1: x2]
+#                 label_coordinates = np.where(label_local==label_index)
+#                 tmp_merged = np.zeros((h, w), np.uint8)
+#                 tmp_merged[label_coordinates] = 255
+#                 andmap = cv2.bitwise_and(tmp_merged, pred_textmsk[y1: y2, x1: x2])
+#                 text_score = andmap.sum() / area
+#                 if text_score > text_score_thresh:
+#                     text_labels.append(label_index)
+#                     hyp_textmsk[y1: y2, x1: x2][label_coordinates] = 255
+#     labels = label_unchanged
+#     bubble_msk = np.zeros((img.shape[0], img.shape[1]), np.uint8)
+#     bubble_msk[np.where(labels==0)] = 255
+#     # if lang == LANG_JPN:
+#     bubble_msk = cv2.erode(bubble_msk, (3, 3), iterations=1)
+#     line_thickness = 2
+#     cv2.rectangle(bubble_msk, (0, 0), (im_w, im_h), BLACK, line_thickness, cv2.LINE_8)
+#     contours, hiers = cv2.findContours(bubble_msk, cv2.RETR_CCOMP, cv2.CHAIN_APPROX_NONE)
+
+#     brect_area_thresh = im_h * im_w * 0.4
+#     min_brect_area = np.inf
+#     ballon_index = -1
+#     maximum_pixsum = -1
+#     for ii, contour in enumerate(contours):
+#         brect = cv2.boundingRect(contours[ii])
+#         brect_area = brect[2] * brect[3]
+#         if brect_area > brect_area_thresh and brect_area < min_brect_area:
+#             tmp_ballonmsk = np.zeros_like(bubble_msk)
+#             tmp_ballonmsk = cv2.drawContours(tmp_ballonmsk, contours, ii, WHITE, cv2.FILLED)
+#             andmap_sum = cv2.bitwise_and(tmp_ballonmsk, hyp_textmsk).sum()
+#             if andmap_sum > maximum_pixsum:
+#                 maximum_pixsum = andmap_sum
+#                 min_brect_area = brect_area
+#                 ballon_index = ii
+#     if ballon_index != -1:
+#         bubble_msk = np.zeros_like(bubble_msk)
+#         bubble_msk = cv2.drawContours(bubble_msk, contours, ballon_index, WHITE, cv2.FILLED)
+#     hyp_textmsk = cv2.bitwise_and(hyp_textmsk, bubble_msk)
+#     return hyp_textmsk, bubble_msk, thresh_info, (num_labels, label_unchanged, stats, centroids, text_labels)
+
+# def extract_textballoon_channelwise(img, pred_textmsk, test_grey=True, global_mask=None):
+#     c_list = [img[:, :, i] for i in range(3)]
+#     if test_grey:
+#         c_list.append(cv2.cvtColor(img, cv2.COLOR_BGR2GRAY))
+#     best_xorpix_sum = np.inf
+#     best_cindex = best_hyptextmsk = best_bubblemsk = best_thresh_info = best_component_stats = None
+#     for c_index, channel in enumerate(c_list):
+#         hyp_textmsk, bubble_msk, thresh_info, component_stats = extract_textballoon(channel, pred_textmsk, global_mask=global_mask)
+#         pixor_sum = cv2.bitwise_xor(hyp_textmsk, pred_textmsk).sum()
+#         if pixor_sum < best_xorpix_sum:
+#             best_xorpix_sum = pixor_sum
+#             best_cindex = c_index
+#             best_hyptextmsk, best_bubblemsk, best_thresh_info, best_component_stats = hyp_textmsk, bubble_msk, thresh_info, component_stats
+#     return best_hyptextmsk, best_bubblemsk, best_component_stats
+
+# def refine_textmask(img, pred_mask, channel_wise=True, find_leaveouts=True, global_mask=None):
+#     hyp_textmsk, bubble_msk, component_stats = extract_textballoon_channelwise(img, pred_mask, global_mask=global_mask)
+#     num_labels, labels, stats, centroids, text_labels = component_stats
+#     stats = np.array(stats)
+#     text_stats = stats[text_labels]
+#     if find_leaveouts and len(text_stats) > 0:
+#         median_h = np.median(text_stats[:, 3])
+#         for label, label_h in zip(range(num_labels), stats[:, 3]):
+#             if label == 0 or label in text_labels:
+#                 continue
+#             if label_h > 0.5 * median_h and label_h < 1.5 * median_h:
+#                 hyp_textmsk[np.where(labels==label)] = 255
+#         hyp_textmsk = cv2.bitwise_and(hyp_textmsk, bubble_msk)
+#         if global_mask is not None:
+#             hyp_textmsk = cv2.bitwise_and(hyp_textmsk, global_mask)
+#     return hyp_textmsk, bubble_msk

utils/weight_init.py

@@ -0,0 +1,103 @@
+import torch.nn as nn
+import torch
+
+def constant_init(module, val, bias=0):
+    nn.init.constant_(module.weight, val)
+    if hasattr(module, 'bias') and module.bias is not None:
+        nn.init.constant_(module.bias, bias)
+
+def xavier_init(module, gain=1, bias=0, distribution='normal'):
+    assert distribution in ['uniform', 'normal']
+    if distribution == 'uniform':
+        nn.init.xavier_uniform_(module.weight, gain=gain)
+    else:
+        nn.init.xavier_normal_(module.weight, gain=gain)
+    if hasattr(module, 'bias') and module.bias is not None:
+        nn.init.constant_(module.bias, bias)
+
+
+def normal_init(module, mean=0, std=1, bias=0):
+    nn.init.normal_(module.weight, mean, std)
+    if hasattr(module, 'bias') and module.bias is not None:
+        nn.init.constant_(module.bias, bias)
+
+
+def uniform_init(module, a=0, b=1, bias=0):
+    nn.init.uniform_(module.weight, a, b)
+    if hasattr(module, 'bias') and module.bias is not None:
+        nn.init.constant_(module.bias, bias)
+
+
+def kaiming_init(module,
+                 a=0,
+                 is_rnn=False,
+                 mode='fan_in',
+                 nonlinearity='leaky_relu',
+                 bias=0,
+                 distribution='normal'):
+    assert distribution in ['uniform', 'normal']
+    if distribution == 'uniform':
+        if is_rnn:
+            for name, param in module.named_parameters():
+                if 'bias' in name:
+                    nn.init.constant_(param, bias)
+                elif 'weight' in name:
+                    nn.init.kaiming_uniform_(param,
+                                             a=a,
+                                             mode=mode,
+                                             nonlinearity=nonlinearity)
+        else:
+            nn.init.kaiming_uniform_(module.weight,
+                                     a=a,
+                                     mode=mode,
+                                     nonlinearity=nonlinearity)
+
+    else:
+        if is_rnn:
+            for name, param in module.named_parameters():
+                if 'bias' in name:
+                    nn.init.constant_(param, bias)
+                elif 'weight' in name:
+                    nn.init.kaiming_normal_(param,
+                                            a=a,
+                                            mode=mode,
+                                            nonlinearity=nonlinearity)
+        else:
+            nn.init.kaiming_normal_(module.weight,
+                                    a=a,
+                                    mode=mode,
+                                    nonlinearity=nonlinearity)
+
+    if not is_rnn and hasattr(module, 'bias') and module.bias is not None:
+        nn.init.constant_(module.bias, bias)
+
+
+def bilinear_kernel(in_channels, out_channels, kernel_size):
+    factor = (kernel_size + 1) // 2
+    if kernel_size % 2 == 1:
+        center = factor - 1
+    else:
+        center = factor - 0.5
+    og = (torch.arange(kernel_size).reshape(-1, 1),
+          torch.arange(kernel_size).reshape(1, -1))
+    filt = (1 - torch.abs(og[0] - center) / factor) * \
+           (1 - torch.abs(og[1] - center) / factor)
+    weight = torch.zeros((in_channels, out_channels,
+                          kernel_size, kernel_size))
+    weight[range(in_channels), range(out_channels), :, :] = filt
+    return weight
+
+
+def init_weights(m):
+    # for m in modules:
+
+    if isinstance(m, nn.Conv2d):
+        kaiming_init(m)
+    elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
+        constant_init(m, 1)
+    elif isinstance(m, nn.Linear):
+        xavier_init(m)
+    elif isinstance(m, (nn.LSTM, nn.LSTMCell)):
+        kaiming_init(m, is_rnn=True)
+    # elif isinstance(m, nn.ConvTranspose2d):
+    #     m.weight.data.copy_(bilinear_kernel(m.in_channels, m.out_channels, 4));

utils/yolov5_utils.py

@@ -0,0 +1,243 @@
+import math
+import torch
+import torch.nn as nn
+import pkg_resources as pkg
+import torch.nn.functional as F
+import cv2
+import numpy as np
+import time
+import torchvision
+
+def scale_img(img, ratio=1.0, same_shape=False, gs=32):  # img(16,3,256,416)
+    # scales img(bs,3,y,x) by ratio constrained to gs-multiple
+    if ratio == 1.0:
+        return img
+    else:
+        h, w = img.shape[2:]
+        s = (int(h * ratio), int(w * ratio))  # new size
+        img = F.interpolate(img, size=s, mode='bilinear', align_corners=False)  # resize
+        if not same_shape:  # pad/crop img
+            h, w = (math.ceil(x * ratio / gs) * gs for x in (h, w))
+        return F.pad(img, [0, w - s[1], 0, h - s[0]], value=0.447)  # value = imagenet mean
+
+def fuse_conv_and_bn(conv, bn):
+    # Fuse convolution and batchnorm layers https://tehnokv.com/posts/fusing-batchnorm-and-conv/
+    fusedconv = nn.Conv2d(conv.in_channels,
+                          conv.out_channels,
+                          kernel_size=conv.kernel_size,
+                          stride=conv.stride,
+                          padding=conv.padding,
+                          groups=conv.groups,
+                          bias=True).requires_grad_(False).to(conv.weight.device)
+
+    # prepare filters
+    w_conv = conv.weight.clone().view(conv.out_channels, -1)
+    w_bn = torch.diag(bn.weight.div(torch.sqrt(bn.eps + bn.running_var)))
+    fusedconv.weight.copy_(torch.mm(w_bn, w_conv).view(fusedconv.weight.shape))
+
+    # prepare spatial bias
+    b_conv = torch.zeros(conv.weight.size(0), device=conv.weight.device) if conv.bias is None else conv.bias
+    b_bn = bn.bias - bn.weight.mul(bn.running_mean).div(torch.sqrt(bn.running_var + bn.eps))
+    fusedconv.bias.copy_(torch.mm(w_bn, b_conv.reshape(-1, 1)).reshape(-1) + b_bn)
+
+    return fusedconv
+
+def check_anchor_order(m):
+    # Check anchor order against stride order for YOLOv5 Detect() module m, and correct if necessary
+    a = m.anchors.prod(-1).view(-1)  # anchor area
+    da = a[-1] - a[0]  # delta a
+    ds = m.stride[-1] - m.stride[0]  # delta s
+    if da.sign() != ds.sign():  # same order
+        m.anchors[:] = m.anchors.flip(0)
+
+def initialize_weights(model):
+    for m in model.modules():
+        t = type(m)
+        if t is nn.Conv2d:
+            pass  # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
+        elif t is nn.BatchNorm2d:
+            m.eps = 1e-3
+            m.momentum = 0.03
+        elif t in [nn.Hardswish, nn.LeakyReLU, nn.ReLU, nn.ReLU6, nn.SiLU]:
+            m.inplace = True
+
+def make_divisible(x, divisor):
+    # Returns nearest x divisible by divisor
+    if isinstance(divisor, torch.Tensor):
+        divisor = int(divisor.max())  # to int
+    return math.ceil(x / divisor) * divisor
+
+def intersect_dicts(da, db, exclude=()):
+    # Dictionary intersection of matching keys and shapes, omitting 'exclude' keys, using da values
+    return {k: v for k, v in da.items() if k in db and not any(x in k for x in exclude) and v.shape == db[k].shape}
+
+def check_version(current='0.0.0', minimum='0.0.0', name='version ', pinned=False, hard=False):
+    # Check version vs. required version
+    current, minimum = (pkg.parse_version(x) for x in (current, minimum))
+    result = (current == minimum) if pinned else (current >= minimum)  # bool
+    if hard:  # assert min requirements met
+        assert result, f'{name}{minimum} required by YOLOv5, but {name}{current} is currently installed'
+    else:
+        return result
+
+class Colors:
+    # Ultralytics color palette https://ultralytics.com/
+    def __init__(self):
+        # hex = matplotlib.colors.TABLEAU_COLORS.values()
+        hex = ('FF3838', 'FF9D97', 'FF701F', 'FFB21D', 'CFD231', '48F90A', '92CC17', '3DDB86', '1A9334', '00D4BB',
+               '2C99A8', '00C2FF', '344593', '6473FF', '0018EC', '8438FF', '520085', 'CB38FF', 'FF95C8', 'FF37C7')
+        self.palette = [self.hex2rgb('#' + c) for c in hex]
+        self.n = len(self.palette)
+
+    def __call__(self, i, bgr=False):
+        c = self.palette[int(i) % self.n]
+        return (c[2], c[1], c[0]) if bgr else c
+
+    @staticmethod
+    def hex2rgb(h):  # rgb order (PIL)
+        return tuple(int(h[1 + i:1 + i + 2], 16) for i in (0, 2, 4))
+
+def box_iou(box1, box2):
+    # https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py
+    """
+    Return intersection-over-union (Jaccard index) of boxes.
+    Both sets of boxes are expected to be in (x1, y1, x2, y2) format.
+    Arguments:
+        box1 (Tensor[N, 4])
+        box2 (Tensor[M, 4])
+    Returns:
+        iou (Tensor[N, M]): the NxM matrix containing the pairwise
+            IoU values for every element in boxes1 and boxes2
+    """
+
+    def box_area(box):
+        # box = 4xn
+        return (box[2] - box[0]) * (box[3] - box[1])
+
+    area1 = box_area(box1.T)
+    area2 = box_area(box2.T)
+
+    # inter(N,M) = (rb(N,M,2) - lt(N,M,2)).clamp(0).prod(2)
+    inter = (torch.min(box1[:, None, 2:], box2[:, 2:]) - torch.max(box1[:, None, :2], box2[:, :2])).clamp(0).prod(2)
+    return inter / (area1[:, None] + area2 - inter)  # iou = inter / (area1 + area2 - inter)
+
+def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=False,
+                        labels=(), max_det=300):
+    """Runs Non-Maximum Suppression (NMS) on inference results
+
+    Returns:
+         list of detections, on (n,6) tensor per image [xyxy, conf, cls]
+    """
+
+    if isinstance(prediction, np.ndarray):
+        prediction = torch.from_numpy(prediction)
+
+    nc = prediction.shape[2] - 5  # number of classes
+    xc = prediction[..., 4] > conf_thres  # candidates
+
+    # Checks
+    assert 0 <= conf_thres <= 1, f'Invalid Confidence threshold {conf_thres}, valid values are between 0.0 and 1.0'
+    assert 0 <= iou_thres <= 1, f'Invalid IoU {iou_thres}, valid values are between 0.0 and 1.0'
+
+    # Settings
+    min_wh, max_wh = 2, 4096  # (pixels) minimum and maximum box width and height
+    max_nms = 30000  # maximum number of boxes into torchvision.ops.nms()
+    time_limit = 10.0  # seconds to quit after
+    redundant = True  # require redundant detections
+    multi_label &= nc > 1  # multiple labels per box (adds 0.5ms/img)
+    merge = False  # use merge-NMS
+
+    t = time.time()
+    output = [torch.zeros((0, 6), device=prediction.device)] * prediction.shape[0]
+    for xi, x in enumerate(prediction):  # image index, image inference
+        # Apply constraints
+        # x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0  # width-height
+        x = x[xc[xi]]  # confidence
+
+        # Cat apriori labels if autolabelling
+        if labels and len(labels[xi]):
+            l = labels[xi]
+            v = torch.zeros((len(l), nc + 5), device=x.device)
+            v[:, :4] = l[:, 1:5]  # box
+            v[:, 4] = 1.0  # conf
+            v[range(len(l)), l[:, 0].long() + 5] = 1.0  # cls
+            x = torch.cat((x, v), 0)
+
+        # If none remain process next image
+        if not x.shape[0]:
+            continue
+
+        # Compute conf
+        x[:, 5:] *= x[:, 4:5]  # conf = obj_conf * cls_conf
+
+        # Box (center x, center y, width, height) to (x1, y1, x2, y2)
+        box = xywh2xyxy(x[:, :4])
+
+        # Detections matrix nx6 (xyxy, conf, cls)
+        if multi_label:
+            i, j = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T
+            x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1)
+        else:  # best class only
+            conf, j = x[:, 5:].max(1, keepdim=True)
+            x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres]
+
+        # Filter by class
+        if classes is not None:
+            x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)]
+
+        # Apply finite constraint
+        # if not torch.isfinite(x).all():
+        #     x = x[torch.isfinite(x).all(1)]
+
+        # Check shape
+        n = x.shape[0]  # number of boxes
+        if not n:  # no boxes
+            continue
+        elif n > max_nms:  # excess boxes
+            x = x[x[:, 4].argsort(descending=True)[:max_nms]]  # sort by confidence
+
+        # Batched NMS
+        c = x[:, 5:6] * (0 if agnostic else max_wh)  # classes
+        boxes, scores = x[:, :4] + c, x[:, 4]  # boxes (offset by class), scores
+        i = torchvision.ops.nms(boxes, scores, iou_thres)  # NMS
+        if i.shape[0] > max_det:  # limit detections
+            i = i[:max_det]
+        if merge and (1 < n < 3E3):  # Merge NMS (boxes merged using weighted mean)
+            # update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)
+            iou = box_iou(boxes[i], boxes) > iou_thres  # iou matrix
+            weights = iou * scores[None]  # box weights
+            x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True)  # merged boxes
+            if redundant:
+                i = i[iou.sum(1) > 1]  # require redundancy
+
+        output[xi] = x[i]
+        if (time.time() - t) > time_limit:
+            print(f'WARNING: NMS time limit {time_limit}s exceeded')
+            break  # time limit exceeded
+
+    return output
+
+def xywh2xyxy(x):
+    # Convert nx4 boxes from [x, y, w, h] to [x1, y1, x2, y2] where xy1=top-left, xy2=bottom-right
+    y = x.clone() if isinstance(x, torch.Tensor) else np.copy(x)
+    y[:, 0] = x[:, 0] - x[:, 2] / 2  # top left x
+    y[:, 1] = x[:, 1] - x[:, 3] / 2  # top left y
+    y[:, 2] = x[:, 0] + x[:, 2] / 2  # bottom right x
+    y[:, 3] = x[:, 1] + x[:, 3] / 2  # bottom right y
+    return y
+
+DEFAULT_LANG_LIST = ['eng', 'ja']
+def draw_bbox(pred, img, lang_list=None):
+    if lang_list is None:
+        lang_list = DEFAULT_LANG_LIST
+    lw = max(round(sum(img.shape) / 2 * 0.003), 2)  # line width
+    pred = pred.astype(np.int32)
+    colors = Colors()
+    img = np.copy(img)
+    for ii, obj in enumerate(pred):
+        p1, p2 = (obj[0], obj[1]), (obj[2], obj[3])
+        label = lang_list[obj[-1]] + str(ii+1)
+        cv2.rectangle(img, p1, p2, colors(obj[-1], bgr=True), lw, lineType=cv2.LINE_AA)
+        t_w, t_h = cv2.getTextSize(label, 0, fontScale=lw / 3, thickness=lw)[0]
+        cv2.putText(img, label, (p1[0], p1[1] + t_h + 2), 0, lw / 3, colors(obj[-1], bgr=True), max(lw-1, 1), cv2.LINE_AA)
+    return img