sudacode/owocr
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Relicense, comic text detector code from mokuro in manga OCR engine, add split paragraph code/option to disable

Browse Source
This commit is contained in:
AuroraWright
2025-10-22 12:17:39 +02:00
parent 665837a764
commit d5b570a438
9 changed files with 882 additions and 221 deletions
Download Patch File Download Diff File Expand all files Collapse all files

168
owocr/ocr.py
View File

@@ -8,6 +8,7 @@ import logging
from math import sqrt, sin, cos, atan2
import json
import base64
import urllib
from urllib.parse import urlparse, parse_qs
from dataclasses import dataclass, field, asdict
from typing import List, Optional
@@ -20,6 +21,10 @@ import requests
try:
from manga_ocr import MangaOcr as MOCR
from comic_text_detector.inference import TextDetector
from scipy.signal.windows import gaussian
import torch
import cv2
except ImportError:
pass
@@ -366,30 +371,179 @@ class MangaOcr:
available = False
local = True
manual_language = False
coordinate_support = False
coordinate_support = True
threading_support = True
def __init__(self, config={}):
if 'manga_ocr' not in sys.modules:
logger.warning('manga-ocr not available, Manga OCR will not work!')
elif 'scipy' not in sys.modules:
logger.warning('scipy not available, Manga OCR will not work!')
else:
comic_text_detector_path = Path.home() / ".cache" / "manga-ocr"
comic_text_detector_file = comic_text_detector_path / "comictextdetector.pt"
if not comic_text_detector_file.exists():
comic_text_detector_path.mkdir(parents=True, exist_ok=True)
logger.info('Downloading comic text detector model ' + str(comic_text_detector_file))
try:
urllib.request.urlretrieve('https://github.com/zyddnys/manga-image-translator/releases/download/beta-0.3/comictextdetector.pt', str(comic_text_detector_file))
except:
logger.warning('Download failed. Manga OCR will not work!')
return
pretrained_model_name_or_path = config.get('pretrained_model_name_or_path', 'kha-white/manga-ocr-base')
force_cpu = config.get('force_cpu', False)
logger.disable('manga_ocr')
logging.getLogger('transformers').setLevel(logging.ERROR) # silence transformers >=4.46 warnings
from manga_ocr import ocr
ocr.post_process = empty_post_process
logger.info(f'Loading Manga OCR model')
self.model = MOCR(pretrained_model_name_or_path, force_cpu)
if not force_cpu and torch.cuda.is_available():
device = 'cuda'
elif not force_cpu and torch.backends.mps.is_available():
device = 'mps'
else:
device = 'cpu'
logger.info(f'Loading comic text detector model, using device {device}')
self.text_detector_model = TextDetector(model_path=comic_text_detector_file, input_size=1024, device=device, act='leaky')
self.available = True
logger.info('Manga OCR ready')
def _convert_line_bbox(self, rect, img_width, img_height):
x1, y1 = float(rect[0][0]), float(rect[0][1])
x2, y2 = float(rect[1][0]), float(rect[1][1])
x3, y3 = float(rect[2][0]), float(rect[2][1])
x4, y4 = float(rect[3][0]), float(rect[3][1])
return quad_to_bounding_box(x1, y1, x2, y2, x3, y3, x4, y4, img_width, img_height)
def _convert_box_bbox(self, rect, img_width, img_height):
x1, y1, x2, y2 = rect
width_px = x2 - x1
height_px = y2 - y1
center_x_px = (x1 + x2) / 2
center_y_px = (y1 + y2) / 2
width_norm = width_px / img_width
height_norm = height_px / img_height
center_x_norm = center_x_px / img_width
center_y_norm = center_y_px / img_height
return BoundingBox(
center_x=float(center_x_norm),
center_y=float(center_y_norm),
width=float(width_norm),
height=float(height_norm)
)
# from https://github.com/kha-white/mokuro/blob/master/mokuro/manga_page_ocr.py
def _split_into_chunks(self, img, mask_refined, blk, line_idx, textheight, max_ratio, anchor_window):
line_crop = blk.get_transformed_region(img, line_idx, textheight)
h, w, *_ = line_crop.shape
ratio = w / h
if ratio <= max_ratio:
return [line_crop], []
else:
k = gaussian(textheight * 2, textheight / 8)
line_mask = blk.get_transformed_region(mask_refined, line_idx, textheight)
num_chunks = int(np.ceil(ratio / max_ratio))
anchors = np.linspace(0, w, num_chunks + 1)[1:-1]
line_density = line_mask.sum(axis=0)
line_density = np.convolve(line_density, k, 'same')
line_density /= line_density.max()
anchor_window *= textheight
cut_points = []
for anchor in anchors:
anchor = int(anchor)
n0 = np.clip(anchor - anchor_window // 2, 0, w)
n1 = np.clip(anchor + anchor_window // 2, 0, w)
p = line_density[n0:n1].argmin()
p += n0
cut_points.append(p)
return np.split(line_crop, cut_points, axis=1), cut_points
# derived from https://github.com/kha-white/mokuro/blob/master/mokuro/manga_page_ocr.py
def _to_generic_result(self, mask_refined, blk_list, img_np, img_height, img_width):
paragraphs = []
for blk_idx, blk in enumerate(blk_list):
lines = []
for line_idx, line in enumerate(blk.lines_array()):
if blk.vertical:
max_ratio = 16
else:
max_ratio = 8
line_crops, cut_points = self._split_into_chunks(
img_np,
mask_refined,
blk,
line_idx,
textheight=64,
max_ratio=max_ratio,
anchor_window=2,
)
l_text = ''
for line_crop in line_crops:
if blk.vertical:
line_crop = cv2.rotate(line_crop, cv2.ROTATE_90_CLOCKWISE)
l_text += self.model(Image.fromarray(line_crop))
l_bbox = self._convert_line_bbox(line.tolist(), img_width, img_height)
word = Word(
text=l_text,
bounding_box=l_bbox
)
words = [word]
line = Line(
text=l_text,
bounding_box=l_bbox,
words=words
)
lines.append(line)
p_bbox = self._convert_box_bbox(list(blk.xyxy), img_width, img_height)
writing_direction = 'TOP_TO_BOTTOM' if blk.vertical else None
paragraph = Paragraph(bounding_box=p_bbox, lines=lines, writing_direction=writing_direction)
paragraphs.append(paragraph)
return OcrResult(
image_properties=ImageProperties(width=img_width, height=img_height),
paragraphs=paragraphs
)
def __call__(self, img):
img, is_path = input_to_pil_image(img)
if not img:
return (False, 'Invalid image provided')
x = (True, [self.model(img)])
img_np = pil_image_to_numpy_array(img)
img_width, img_height = img.size
_, mask_refined, blk_list = self.text_detector_model(img_np, refine_mode=1, keep_undetected_mask=True)
ocr_result = self._to_generic_result(mask_refined, blk_list, img_np, img_height, img_width)
x = (True, ocr_result)
if is_path:
img.close()
@@ -631,10 +785,10 @@ class GoogleLens:
filter.filter_type = LensOverlayFilterType.AUTO_FILTER
request.objects_request.request_context.client_context.client_filters.filter.append(filter)
image_data = self._preprocess(img)
request.objects_request.image_data.payload.image_bytes = image_data[0]
request.objects_request.image_data.image_metadata.width = image_data[1]
request.objects_request.image_data.image_metadata.height = image_data[2]
img_bytes, img_width, img_height = self._preprocess(img)
request.objects_request.image_data.payload.image_bytes = img_bytes
request.objects_request.image_data.image_metadata.width = img_width
request.objects_request.image_data.image_metadata.height = img_height
payload = request.SerializeToString()
@@ -678,7 +832,7 @@ class GoogleLens:
new_h = int(new_w / aspect_ratio)
img = img.resize((new_w, new_h), Image.Resampling.LANCZOS)
return (pil_image_to_bytes(img), img.width, img.height)
return pil_image_to_bytes(img), img.width, img.height
class Bing:
name = 'bing'