comicocr / utils /textblock_mask.py
fasdfsa's picture
init
f6f8d06
import cv2
import numpy as np
from typing import Tuple
from .imgproc_utils import draw_connected_labels
from .stroke_width_calculator import strokewidth_check
opencv_inpaint = lambda img, mask: cv2.inpaint(img, mask, 3, cv2.INPAINT_NS)
def show_img_by_dict(imgdicts):
for keyname in imgdicts.keys():
cv2.imshow(keyname, imgdicts[keyname])
cv2.waitKey(0)
# 计算文本rgb均值
def letter_calculator(img, mask, bground_rgb, show_process=False):
gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
# rgb to grey
aver_bground_rgb = 0.299 * bground_rgb[0] + 0.587 * bground_rgb[1] + 0.114 * bground_rgb[2]
thresh_low = 127
retval, threshed = cv2.threshold(gray, 127, 255, cv2.THRESH_OTSU)
if aver_bground_rgb < thresh_low:
threshed = 255 - threshed
threshed = 255 - threshed
threshed = cv2.bitwise_and(threshed, mask)
le_region = np.where(threshed==255)
mat_region = img[le_region]
if mat_region.shape[0] == 0:
# retval, threshed = cv2.threshold(gray, 20, 255, cv2.THRESH_BINARY)
# cv2.imshow("xxx", threshed)
# cv2.imshow("2xxx", img)
# cv2.waitKey(0)
return [-1, -1, -1], threshed
letter_rgb = np.mean(mat_region, axis=0).astype(int).tolist()
if show_process:
cv2.imshow("thresh", threshed)
# ocr_protest(threshed)
imgcp = np.copy(img)
imgcp *= 0
imgcp += 127
imgcp[le_region] = letter_rgb
cv2.imshow("letter_img", imgcp)
# cv2.waitKey(0)
return letter_rgb, threshed
# 预处理让文本颜色提取准确点
def usm(src):
# Handle RGBA images by converting to RGB for processing
if len(src.shape) == 3 and src.shape[2] == 4:
src = cv2.cvtColor(src, cv2.COLOR_RGBA2RGB)
blur_img = cv2.GaussianBlur(src, (0, 0), 5)
usm = cv2.addWeighted(src, 1.5, blur_img, -0.5, 0)
h, w = src.shape[:2]
result = np.zeros([h, w*2, 3], dtype=src.dtype)
result[0:h,0:w,:] = src
result[0:h,w:2*w,:] = usm
return usm
# 计算文本rgb均值方法2,可能用中位数代替均值会好点
def textrgb_calculator(img, text_mask, show_process=False):
text_mask = cv2.erode(text_mask, (3, 3), iterations=1)
usm_img = usm(img)
overall_meanrgb = np.mean(usm_img[np.where(text_mask==255)], axis=0)
if show_process:
colored_text_board = np.zeros((img.shape[0], img.shape[1], 3), dtype=np.uint8) + 127
colored_text_board[np.where(text_mask==255)] = overall_meanrgb
cv2.imshow("usm", usm_img)
cv2.imshow("textcolor", colored_text_board)
return overall_meanrgb.astype(np.uint8)
# 计算背景rgb均值和标准差
def bground_calculator(buble_img, back_ground_mask, dilate=True):
kernel = np.ones((3,3),np.uint8)
if dilate:
back_ground_mask = cv2.dilate(back_ground_mask, kernel, iterations = 1)
bground_region = np.where(back_ground_mask==0)
sd = -1
if len(bground_region[0]) != 0:
pix_array = buble_img[bground_region]
bground_aver = np.mean(pix_array, axis=0).astype(int)
pix_array - bground_aver
gray = cv2.cvtColor(buble_img, cv2.COLOR_RGB2GRAY)
gray_pixarray = gray[bground_region]
gray_aver = np.mean(gray_pixarray)
gray_pixarray = gray_pixarray - gray_aver
gray_pixarray = np.power(gray_pixarray, 2)
# gray_pixarray = np.sqrt(gray_pixarray)
sd = np.mean(gray_pixarray)
else: bground_aver = np.array([-1, -1, -1])
return bground_aver, bground_region, sd
# 输入:文本块roi,分割出文本mask,根据mask计算文本bgr均值和标准差,决定纯色覆盖/inpaint修复
def canny_flood(img, show_process=False, inpaint_sdthresh=10, **kwargs):
# cv2.setNumThreads(4)
WHITE = (255, 255, 255)
BLACK = (0, 0, 0)
kernel = np.ones((3,3),np.uint8)
orih, oriw = img.shape[0], img.shape[1]
# Handle RGBA images by converting to RGB for processing
if len(img.shape) == 3 and img.shape[2] == 4:
# Convert RGBA to RGB for processing
img = cv2.cvtColor(img, cv2.COLOR_RGBA2RGB)
scaleR = 1
if orih > 300 and oriw > 300:
scaleR = 0.6
elif orih < 120 or oriw < 120:
scaleR = 1.4
if scaleR != 1:
h, w = img.shape[0], img.shape[1]
orimg = np.copy(img)
img = cv2.resize(img, (int(w*scaleR), int(h*scaleR)), interpolation=cv2.INTER_AREA)
h, w = img.shape[0], img.shape[1]
img_area = h * w
cpimg = cv2.GaussianBlur(img,(3,3),cv2.BORDER_DEFAULT)
detected_edges = cv2.Canny(cpimg, 70, 140, L2gradient=True, apertureSize=3)
cv2.rectangle(detected_edges, (0, 0), (w-1, h-1), WHITE, 1, cv2.LINE_8)
cons, hiers = cv2.findContours(detected_edges, cv2.RETR_CCOMP, cv2.CHAIN_APPROX_NONE)
cv2.rectangle(detected_edges, (0, 0), (w-1, h-1), BLACK, 1, cv2.LINE_8)
ballon_mask, outer_index = np.zeros((h, w), np.uint8), -1
min_retval = np.inf
mask = np.zeros((h, w), np.uint8)
difres = 10
seedpnt = (int(w/2), int(h/2))
for ii in range(len(cons)):
rect = cv2.boundingRect(cons[ii])
if rect[2]*rect[3] < img_area*0.4:
continue
mask = cv2.drawContours(mask, cons, ii, (255), 2)
cpmask = np.copy(mask)
cv2.rectangle(mask, (0, 0), (w-1, h-1), WHITE, 1, cv2.LINE_8)
retval, _, _, rect = cv2.floodFill(cpmask, mask=None, seedPoint=seedpnt, flags=4, newVal=(127), loDiff=(difres, difres, difres), upDiff=(difres, difres, difres))
if retval <= img_area * 0.3:
mask = cv2.drawContours(mask, cons, ii, (0), 2)
if retval < min_retval and retval > img_area * 0.3:
min_retval = retval
ballon_mask = cpmask
ballon_mask = 127 - ballon_mask
ballon_mask = cv2.dilate(ballon_mask, kernel,iterations = 1)
outer_area, _, _, rect = cv2.floodFill(ballon_mask, mask=None, seedPoint=seedpnt, flags=4, newVal=(30), loDiff=(difres, difres, difres), upDiff=(difres, difres, difres))
ballon_mask = 30 - ballon_mask
retval, ballon_mask = cv2.threshold(ballon_mask, 1, 255, cv2.THRESH_BINARY)
ballon_mask = cv2.bitwise_not(ballon_mask, ballon_mask)
detected_edges = cv2.dilate(detected_edges, kernel, iterations = 1)
for ii in range(2):
detected_edges = cv2.bitwise_and(detected_edges, ballon_mask)
mask = np.copy(detected_edges)
bgarea1, _, _, rect = cv2.floodFill(mask, mask=None, seedPoint=(0, 0), flags=4, newVal=(127), loDiff=(difres, difres, difres), upDiff=(difres, difres, difres))
bgarea2, _, _, rect = cv2.floodFill(mask, mask=None, seedPoint=(detected_edges.shape[1]-1, detected_edges.shape[0]-1), flags=4, newVal=(127), loDiff=(difres, difres, difres), upDiff=(difres, difres, difres))
txt_area = min(img_area - bgarea1, img_area - bgarea2)
ratio_ob = txt_area / outer_area
ballon_mask = cv2.erode(ballon_mask, kernel,iterations = 1)
if ratio_ob < 0.85:
break
mask = 127 - mask
retval, mask = cv2.threshold(mask, 1, 255, cv2.THRESH_BINARY)
if scaleR != 1:
img = orimg
ballon_mask = cv2.resize(ballon_mask, (oriw, orih))
mask = cv2.resize(mask, (oriw, orih))
bg_mask = cv2.bitwise_or(mask, 255-ballon_mask)
mask = cv2.bitwise_and(mask, ballon_mask)
bground_aver, bground_region, sd = bground_calculator(img, bg_mask)
inner_rect = None
threshed = np.zeros((img.shape[0], img.shape[1]), np.uint8)
if bground_aver[0] != -1:
letter_aver, threshed = letter_calculator(img, mask, bground_aver, show_process=show_process)
if letter_aver[0] != -1:
mask = cv2.dilate(threshed, kernel, iterations=1)
inner_rect = cv2.boundingRect(cv2.findNonZero(mask))
else: letter_aver = [0, 0, 0]
if sd != -1 and sd < inpaint_sdthresh:
need_inpaint = False
else:
need_inpaint = True
if show_process:
print(f"\nneed_inpaint: {need_inpaint}, sd: {sd}, {type(inner_rect)}")
show_img_by_dict({"outermask": ballon_mask, "detect": detected_edges, "mask": mask})
if isinstance(inner_rect, tuple):
inner_rect = [ii for ii in inner_rect]
if inner_rect is None:
inner_rect = [-1, -1, -1, -1]
else:
inner_rect.append(-1)
bground_aver = bground_aver.astype(np.uint8)
bub_dict = {"rgb": letter_aver,
"bground_rgb": bground_aver,
"inner_rect": inner_rect,
"need_inpaint": need_inpaint}
return mask, ballon_mask, bub_dict
# 输入:文本块roi,分割出文本mask,根据mask计算文本bgr均值和标准差,决定纯色覆盖/inpaint修复
def connected_canny_flood(img, show_process=False, inpaint_sdthresh=10, apply_strokewidth_check=0, **kwargs):
# Handle RGBA images by converting to RGB for processing
if len(img.shape) == 3 and img.shape[2] == 4:
# Convert RGBA to RGB for processing
img = cv2.cvtColor(img, cv2.COLOR_RGBA2RGB)
# 寻找最可能是气泡的外轮廓mask
def find_outermask(img):
connectivity = 4
num_labels, labels, stats, centroids = cv2.connectedComponentsWithStats(img, connectivity, cv2.CV_16U)
drawtext = np.zeros((img.shape[0], img.shape[1]), np.uint8)
max_ind = np.argmax(stats[:, 4])
maxbbox_area, sec_ind = -1, -1
for ind, stat in enumerate(stats):
if ind != max_ind:
bbarea = stat[2] * stat[3]
if bbarea > maxbbox_area:
maxbbox_area = bbarea
sec_ind = ind
drawtext[np.where(labels==max_ind)] = 255
cv2.rectangle(drawtext, (0, 0), (img.shape[1]-1, img.shape[0]-1), (0, 0, 0), 1, cv2.LINE_8)
cons, hiers = cv2.findContours(drawtext, cv2.RETR_CCOMP, cv2.CHAIN_APPROX_NONE)
img_area = img.shape[0] * img.shape[1]
rects = np.array([cv2.boundingRect(cnt) for cnt in cons])
rect_area = np.array([rect[2] * rect[3] for rect in rects])
quali_ind = np.where(rect_area > img_area * 0.3)[0]
ballon_mask = np.zeros((img.shape[0], img.shape[1]), np.uint8)
for ind in quali_ind:
ballon_mask = cv2.drawContours(ballon_mask, cons, ind, (255), 2)
seedpnt = (int(ballon_mask.shape[1]/2), int(ballon_mask.shape[0]/2))
difres = 10
retval, _, _, rect = cv2.floodFill(ballon_mask, mask=None, seedPoint=seedpnt, flags=4, newVal=(127), loDiff=(difres, difres, difres), upDiff=(difres, difres, difres))
ballon_mask = 255 - cv2.threshold(ballon_mask - 127, 1, 255, cv2.THRESH_BINARY)[1]
return num_labels, labels, stats, centroids, ballon_mask
# BGR直接转灰度图可能导致文本区域和背景难以区分,比如测试样例中的黑底红字
# 但是总有一个通道文本和背景容易区分
# 返回最容易区分的那个通道
def ccctest(img, crop_r=0.1):
# img = usm(img)
maxh = 100
if img.shape[0] > maxh:
scaleR = maxh / img.shape[0]
im = cv2.resize(img, (int(img.shape[1]*scaleR), int(img.shape[0]*scaleR)), interpolation=cv2.INTER_AREA)
else:
im = img
textlabel_counter = 0
reverse = False
c_ind = 0
num_labels, labels, stats, centroids, pseduo_outermask = find_outermask(cv2.threshold(cv2.cvtColor(im, cv2.COLOR_RGB2GRAY), 1, 255, cv2.THRESH_OTSU+cv2.THRESH_BINARY)[1])
grayim = np.expand_dims(np.array(cv2.cvtColor(im, cv2.COLOR_RGB2GRAY)), axis=2)
im = np.append(im, grayim, axis=2)
outer_cords = np.where(pseduo_outermask==255)
for bgr_ind in range(4):
channel = im[:, :, bgr_ind]
ret, thresh = cv2.threshold(channel, 1, 255, cv2.THRESH_OTSU+cv2.THRESH_BINARY)
tmp_reverse = False
if np.mean(thresh[outer_cords]) > 160:
thresh = 255 - thresh
tmp_reverse = True
num_labels, labels, stats, centroids = cv2.connectedComponentsWithStats(thresh, 4, cv2.CV_16U)
# draw_connected_labels(num_labels, labels, stats, centroids)
# cv2.waitKey(0)
max_ind = np.argmax(stats[:, 4])
maxr, minr = 0.5, 0.001
maxw, maxh = stats[max_ind][2] * maxr, stats[max_ind][3] * maxr
minarea = im.shape[0] * im.shape[1] * minr
tmp_counter = 0
for stat in stats:
bboxarea = stat[2] * stat[3]
if stat[2] < maxw and stat[3] < maxh and bboxarea > minarea:
tmp_counter += 1
if tmp_counter > textlabel_counter:
textlabel_counter = tmp_counter
c_ind = bgr_ind
reverse = tmp_reverse
return c_ind, reverse
channel_index, reverse = ccctest(img)
chanel = img[:, :, channel_index] if channel_index < 3 else cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
ret, thresh = cv2.threshold(chanel, 1, 255, cv2.THRESH_OTSU+cv2.THRESH_BINARY)
# reverse to get white text on black bg
if reverse:
thresh = 255 - thresh
num_labels, labels, stats, centroids, ballon_mask = find_outermask(thresh)
img_area = img.shape[0] * img.shape[1]
text_mask = np.zeros((img.shape[0], img.shape[1]), np.uint8)
max_ind = np.argmax(stats[:, 4])
for lab in (range(num_labels)):
stat = stats[lab]
if lab != max_ind and stat[4] < img_area * 0.4:
labcord = np.where(labels==lab)
text_mask[labcord] = 255
text_mask = cv2.bitwise_and(text_mask, ballon_mask)
if apply_strokewidth_check > 0:
text_mask = strokewidth_check(text_mask, labels, num_labels, stats, debug_type=show_process-1)
text_color = textrgb_calculator(img, text_mask, show_process=show_process)
inner_rect = cv2.boundingRect(cv2.findNonZero(cv2.dilate(text_mask, (3, 3), iterations=1)))
inner_rect = [ii for ii in inner_rect]
inner_rect.append(-1)
bg_mask = cv2.bitwise_or(text_mask, 255-ballon_mask)
bground_aver, bground_region, sd = bground_calculator(img, bg_mask)
mask = cv2.GaussianBlur(text_mask,(3,3),cv2.BORDER_DEFAULT)
_, mask = cv2.threshold(mask, 1, 255, cv2.THRESH_BINARY)
if sd != -1 and sd < inpaint_sdthresh:
need_inpaint = False
else:
need_inpaint = True
if show_process:
print(f"\nuse inpaint: {need_inpaint}, sd: {sd}, {type(inner_rect)}")
draw_connected_labels(num_labels, labels, stats, centroids)
show_img_by_dict({"thresh": thresh, "ori": img, "outer": ballon_mask, "text": text_mask, "bgmask": bg_mask})
bground_aver = bground_aver.astype(np.uint8)
bub_dict = {"rgb": text_color,
"bground_rgb": bground_aver,
"inner_rect": inner_rect,
"need_inpaint": need_inpaint}
return mask, ballon_mask, bub_dict
def existing_mask(img, mask: np.ndarray):
bub_dict = {"rgb": [0, 0, 0],"bground_rgb": [255, 255, 255],"need_inpaint": True}
return mask, mask, bub_dict
def extract_ballon_mask(img: np.ndarray, mask: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
'''
Given original img and text mask (cropped)
return ballon mask & non text mask
'''
# Handle RGBA images by converting to RGB for processing
if len(img.shape) == 3 and img.shape[2] == 4:
img = cv2.cvtColor(img, cv2.COLOR_RGBA2RGB)
img = cv2.GaussianBlur(img,(3,3),cv2.BORDER_DEFAULT)
h, w = img.shape[:2]
text_sum = np.sum(mask)
cannyed = cv2.Canny(img, 70, 140, L2gradient=True, apertureSize=3)
e_size = 1
element = cv2.getStructuringElement(cv2.MORPH_RECT, (2 * e_size + 1, 2 * e_size + 1),(e_size, e_size))
cannyed = cv2.dilate(cannyed, element, iterations=1)
br = cv2.boundingRect(cv2.findNonZero(mask))
br_xyxy = [br[0], br[1], br[0] + br[2], br[1] + br[3]]
# draw the bounding rect in case there is no closed ballon
cv2.rectangle(cannyed, (0, 0), (w-1, h-1), (255, 255, 255), 1, cv2.LINE_8)
cannyed = cv2.bitwise_and(cannyed, 255 - mask)
cons, _ = cv2.findContours(cannyed, cv2.RETR_CCOMP, cv2.CHAIN_APPROX_NONE)
min_ballon_area = w * h
ballon_mask = None
non_text_mask = None
# minimum contour which covers all text mask must be the ballon
for ii, con in enumerate(cons):
br_c = cv2.boundingRect(con)
br_c = [br_c[0], br_c[1], br_c[0] + br_c[2], br_c[1] + br_c[3]]
if br_c[0] > br_xyxy[0] or br_c[1] > br_xyxy[1] or br_c[2] < br_xyxy[2] or br_c[3] < br_xyxy[3]:
continue
tmp = np.zeros_like(cannyed)
cv2.drawContours(tmp, cons, ii, (255, 255, 255), -1, cv2.LINE_8)
if cv2.bitwise_and(tmp, mask).sum() >= text_sum:
con_area = cv2.contourArea(con)
if con_area < min_ballon_area:
min_ballon_area = con_area
ballon_mask = tmp
if ballon_mask is not None:
non_text_mask = cv2.bitwise_and(ballon_mask, 255 - mask)
# cv2.imshow('ballon', ballon_mask)
# cv2.imshow('non_text', non_text_mask)
# cv2.imshow('im', img)
# cv2.imshow('msk', mask)
# cv2.imshow('canny', cannyed)
# cv2.waitKey(0)
return ballon_mask, non_text_mask