MohamedDhouib1 commited on
Commit
be4c88d
·
verified ·
1 Parent(s): ce77644

Upload README.md, doc_protocol/README.md, tdoc_protocol/README.md

Browse files
Files changed (3) hide show
  1. README.md +75 -0
  2. doc_protocol/README.md +3 -0
  3. tdoc_protocol/README.md +3 -0
README.md ADDED
@@ -0,0 +1,75 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-nc-4.0
3
+ library_name: pytorch
4
+ pipeline_tag: image-segmentation
5
+ tags:
6
+ - document-forensics
7
+ - tampering-localization
8
+ - image-forensics
9
+ ---
10
+
11
+ # DiffNet
12
+
13
+ DiffNet localises tampered regions in document images. Given a document image, it predicts a per-pixel tampering
14
+ probability map and an image-level tampering probability.
15
+
16
+ Paper: [arXiv:2606.22285](https://arxiv.org/abs/2606.22285) · Code: <https://github.com/Mohamed-Dhouib/DiffNet>
17
+
18
+ ## Weights
19
+
20
+ One folder per training protocol of the paper. Each folder contains the weights (`.safetensors`, network
21
+ state_dict exported with `training/export_weights.py`) and the training config (`diffnet.yaml`).
22
+
23
+ | Folder | Protocol | Inference input size | Status |
24
+ |---|---|---|---|
25
+ | `doc_protocol/` | Doc Protocol: trained on the DocTamper training set (512x512); tested on the DocTamper test sets and on T-SROIE, OSTF, Tampered-IC13, RealTextManipulation | 512 | coming soon |
26
+ | `tdoc_protocol/` | Syn2Real-TDoc protocol: trained on TDoc-2.8M (1024x1024) from the `doc_protocol/` weights; tested zero-shot on RTM, FindItAgain, FindIt | 1024 | coming soon |
27
+
28
+ ## Download
29
+
30
+ ```python
31
+ from huggingface_hub import hf_hub_download
32
+
33
+ repo_id = "MohamedDhouib1/DiffNet"
34
+ weights = hf_hub_download(repo_id, "doc_protocol/diffnet_doc_protocol.safetensors")
35
+ config = hf_hub_download(repo_id, "doc_protocol/diffnet.yaml")
36
+ ```
37
+
38
+ ## Inference
39
+
40
+ The model code is in the GitHub repository (installation details in its README):
41
+
42
+ ```bash
43
+ git clone https://github.com/Mohamed-Dhouib/DiffNet DiffNet
44
+ cd DiffNet
45
+ pip install -r requirements.txt # after torch and mmcv, see the README
46
+ cd training
47
+ python infer.py --weights <path>/diffnet_doc_protocol.safetensors --input <image-or-folder> \
48
+ --output-dir <output-folder> --input-size 512
49
+ ```
50
+
51
+ Use `--input-size 1024` with the `tdoc_protocol/` weights. Inference also runs on CPU (`--device cpu`). Larger
52
+ images are processed with non-overlapping windows of the input size, smaller ones are padded with white. For each
53
+ image, `infer.py` writes `<name>_prob.png` (probability x 255) and `<name>_mask.png` (probability > `--threshold`,
54
+ default 0.5, 0/255) and prints the image-level probability.
55
+
56
+ Training and evaluation instructions: `README.md` and `REPRODUCIBILITY.md` of the GitHub repository. Results are
57
+ reported in the paper.
58
+
59
+ ## Citation
60
+
61
+ ```bibtex
62
+ @inproceedings{dhouib2026efficient,
63
+ title={Efficient Document Tampering Localization with Multi-level Discrepancy Features and Unified DCT--Quantization Embedding},
64
+ author={Dhouib, Mohamed and Zhu, Ye and Vanier, Sonia and Shabou, Aymen},
65
+ booktitle={European Conference on Computer Vision},
66
+ pages={110--127},
67
+ year={2026},
68
+ organization={Springer}
69
+ }
70
+ ```
71
+
72
+ ## Licence
73
+
74
+ The weights are released under CC BY-NC 4.0 (they are initialised from the ConvNeXt-V2 ImageNet checkpoints,
75
+ which are CC BY-NC 4.0). The code on GitHub is MIT.
doc_protocol/README.md ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ # doc_protocol
2
+
3
+ Weights trained under the Doc Protocol will be added to this folder.
tdoc_protocol/README.md ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ # tdoc_protocol
2
+
3
+ Weights trained under the Syn2Real-TDoc protocol will be added to this folder.