Spaces:
Running
Running
File size: 20,696 Bytes
836388f 42dc069 836388f c04ffe5 836388f c04ffe5 836388f 42dc069 836388f 42dc069 836388f 42dc069 836388f 42dc069 836388f 42dc069 836388f 42dc069 836388f 42dc069 836388f 42dc069 836388f 42dc069 836388f 42dc069 836388f 42dc069 836388f 42dc069 836388f |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 |
"""
Image segmentation utility for OCR preprocessing.
Separates text regions from image regions to improve OCR accuracy on mixed-content documents.
Based on Mistral AI cookbook examples.
"""
import cv2
import numpy as np
from PIL import Image
import io
import base64
import logging
from pathlib import Path
from typing import Tuple, List, Dict, Union, Optional
# Configure logging
logging.basicConfig(level=logging.INFO,
format='%(asctime)s - %(name)s - %(levelname)s - %(message)s')
logger = logging.getLogger(__name__)
def segment_image_for_ocr(image_path: Union[str, Path], vision_enabled: bool = True, preserve_content: bool = True) -> Dict[str, Union[Image.Image, str]]:
"""
Segment an image into text and image regions for improved OCR processing.
Args:
image_path: Path to the image file
vision_enabled: Whether the vision model is enabled
Returns:
Dict containing:
- 'text_regions': PIL Image with highlighted text regions
- 'image_regions': PIL Image with highlighted image regions
- 'text_mask_base64': Base64 string of text mask for visualization
- 'combined_result': PIL Image with combined processing approach
"""
# Convert to Path object if string
image_file = Path(image_path) if isinstance(image_path, str) else image_path
# Log start of processing
logger.info(f"Segmenting image for OCR: {image_file.name}")
try:
# Open original image with PIL for compatibility
with Image.open(image_file) as pil_img:
# --- 2 · Stop "text page detected as image" when vision model is off ---
if not vision_enabled:
# Import the entropy calculator from utils.image_utils
from utils.image_utils import calculate_image_entropy
# Calculate entropy to determine if this is line art or blank
ent = calculate_image_entropy(pil_img)
if ent < 3.5: # Heuristically low → line-art or blank page
logger.info(f"Low entropy image detected ({ent:.2f}), classifying as illustration")
# Return minimal result for illustration
return {
'text_regions': None,
'image_regions': pil_img,
'text_mask_base64': None,
'combined_result': None,
'text_regions_coordinates': []
}
# Convert to RGB if not already
if pil_img.mode != 'RGB':
pil_img = pil_img.convert('RGB')
# Convert PIL image to OpenCV format
img = np.array(pil_img)
img_rgb = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
# Create grayscale version for text detection
gray = cv2.cvtColor(img_rgb, cv2.COLOR_BGR2GRAY)
# Step 1: Apply adaptive thresholding to identify potential text areas
# This works well for printed text against contrasting backgrounds
binary = cv2.adaptiveThreshold(gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
cv2.THRESH_BINARY_INV, 11, 2)
# Step 2: Perform morphological operations to connect text components
# Use a combination of horizontal and vertical kernels for better text detection
# in historical documents with mixed content
horiz_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (15, 1))
vert_kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (1, 3))
# Apply horizontal dilation to connect characters in a line
horiz_dilation = cv2.dilate(binary, horiz_kernel, iterations=1)
# Apply vertical dilation to connect lines in a paragraph
vert_dilation = cv2.dilate(binary, vert_kernel, iterations=1)
# Combine both dilations for better region detection
dilation = cv2.bitwise_or(horiz_dilation, vert_dilation)
# Step 3: Find contours which will correspond to text blocks
contours, _ = cv2.findContours(dilation, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
# Prepare masks to separate text and image regions
text_mask = np.zeros_like(gray)
# Step 4: Filter contours based on size to identify text regions
min_area = 50 # Lower minimum area to catch smaller text blocks in historical documents
max_area = img.shape[0] * img.shape[1] * 0.4 # Reduced max to avoid capturing too much
text_regions = []
for contour in contours:
area = cv2.contourArea(contour)
# Filter by area to avoid noise
if min_area < area < max_area:
# Get the bounding rectangle
x, y, w, h = cv2.boundingRect(contour)
# Calculate aspect ratio - text regions typically have wider aspect ratio
aspect_ratio = w / h
# Calculate density of dark pixels in the region (text is typically dense)
roi = binary[y:y+h, x:x+w]
dark_pixel_density = np.sum(roi > 0) / (w * h)
# Special handling for historical documents
# Check for position - text is often at the bottom in historical prints
y_position_ratio = y / img.shape[0] # Normalized y position (0 at top, 1 at bottom)
# Bottom regions get preferential treatment as text
is_bottom_region = y_position_ratio > 0.7
# Check if part of a text block cluster (horizontal proximity)
is_text_cluster = False
# Check already identified text regions for proximity
for tx, ty, tw, th in text_regions:
# Check if horizontally aligned and close
if abs((ty + th/2) - (y + h/2)) < max(th, h) and \
abs((tx + tw) - x) < 20: # Near each other horizontally
is_text_cluster = True
break
# More inclusive classification for historical documents
# 1. Typical text characteristics OR
# 2. Bottom position (likely text in historical prints) OR
# 3. Part of a text cluster OR
# 4. Surrounded by other text
is_text_region = ((aspect_ratio > 1.05 or aspect_ratio < 0.9) and dark_pixel_density > 0.1) or \
(is_bottom_region and dark_pixel_density > 0.08) or \
is_text_cluster
if is_text_region:
# Add to text regions list
text_regions.append((x, y, w, h))
# Add to text mask
cv2.rectangle(text_mask, (x, y), (x+w, y+h), 255, -1)
# Step 5: Create visualization for debugging
text_regions_vis = img_rgb.copy()
for x, y, w, h in text_regions:
cv2.rectangle(text_regions_vis, (x, y), (x+w, y+h), (0, 255, 0), 2)
# ENHANCED APPROACH FOR HISTORICAL DOCUMENTS:
# We'll identify different regions including titles at the top of the document
# First, look for potential title text at the top of the document
image_height = img.shape[0]
image_width = img.shape[1]
# Examine the top 20% of the image for potential title text
title_section_height = int(image_height * 0.2)
title_mask = np.zeros_like(gray)
title_mask[:title_section_height, :] = 255
# Find potential title blocks in the top section
title_contours, _ = cv2.findContours(
cv2.bitwise_and(dilation, title_mask),
cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE
)
# Extract title regions with more permissive criteria
title_regions = []
for contour in title_contours:
area = cv2.contourArea(contour)
# Use more permissive criteria for title regions
if area > min_area * 0.8: # Smaller minimum area for titles
x, y, w, h = cv2.boundingRect(contour)
# Title regions typically have wider aspect ratio
aspect_ratio = w / h
# More permissive density check for titles that might be stylized
roi = binary[y:y+h, x:x+w]
dark_pixel_density = np.sum(roi > 0) / (w * h)
# Check if this might be a title
# Titles tend to be wider, in the center, and at the top
is_wide = aspect_ratio > 2.0
is_centered = abs((x + w/2) - (image_width/2)) < (image_width * 0.3)
is_at_top = y < title_section_height
# If it looks like a title or has good text characteristics
if (is_wide and is_centered and is_at_top) or \
(is_at_top and dark_pixel_density > 0.1):
title_regions.append((x, y, w, h))
# Now handle the main content with our standard approach
# Use fixed regions for the main content - typically below the title
# For primary content, assume most text is in the bottom 70%
text_section_start = int(image_height * 0.7) # Start main text section at 70% down
# Create text mask combining the title regions and main text area
text_mask = np.zeros_like(gray)
text_mask[text_section_start:, :] = 255
# Add title regions to the text mask
for x, y, w, h in title_regions:
# Add some padding around title regions
pad_x = max(5, int(w * 0.05))
pad_y = max(5, int(h * 0.05))
x_start = max(0, x - pad_x)
y_start = max(0, y - pad_y)
x_end = min(image_width, x + w + pad_x)
y_end = min(image_height, y + h + pad_y)
# Add title region to the text mask
text_mask[y_start:y_end, x_start:x_end] = 255
# Image mask is the inverse of text mask - for visualization only
image_mask = np.zeros_like(gray)
image_mask[text_mask == 0] = 255
# For main text regions, find blocks of text in the bottom part
# Create a temporary mask for the main text section
temp_mask = np.zeros_like(gray)
temp_mask[text_section_start:, :] = 255
# Find text regions for visualization purposes
text_regions = []
# Start with any title regions we found
text_regions.extend(title_regions)
# Then find text regions in the main content area
text_region_contours, _ = cv2.findContours(
cv2.bitwise_and(dilation, temp_mask),
cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE
)
# Add each detected region
for contour in text_region_contours:
x, y, w, h = cv2.boundingRect(contour)
if w > 10 and h > 5: # Minimum size to be considered text
text_regions.append((x, y, w, h))
# Add the entire bottom section as a fallback text region if none detected
if len(text_regions) == 0:
x, y = 0, text_section_start
w, h = img.shape[1], img.shape[0] - text_section_start
text_regions.append((x, y, w, h))
# Create image regions visualization
image_regions_vis = img_rgb.copy()
# Top section is image
cv2.rectangle(image_regions_vis, (0, 0), (img.shape[1], text_section_start), (0, 0, 255), 2)
# Bottom section has text - draw green boxes around detected text regions
text_regions_vis = img_rgb.copy()
for x, y, w, h in text_regions:
cv2.rectangle(text_regions_vis, (x, y), (x+w, y+h), (0, 255, 0), 2)
# For OCR: CRITICAL - Don't modify the image content
# Only create a non-destructive enhanced version
# For text detection visualization:
text_regions_vis = img_rgb.copy()
for x, y, w, h in text_regions:
cv2.rectangle(text_regions_vis, (x, y), (x+w, y+h), (0, 255, 0), 2)
# For image region visualization:
image_regions_vis = img_rgb.copy()
cv2.rectangle(image_regions_vis, (0, 0), (img.shape[1], text_section_start), (0, 0, 255), 2)
# Create a minimally enhanced version of the original image
# that preserves ALL content (both text and image)
combined_result = img_rgb.copy()
# Apply gentle contrast enhancement if requested
if not preserve_content:
# Use a subtle CLAHE enhancement to improve OCR without losing content
lab_img = cv2.cvtColor(img_rgb, cv2.COLOR_BGR2LAB)
l, a, b = cv2.split(lab_img)
# Very mild CLAHE settings to preserve text
clahe = cv2.createCLAHE(clipLimit=1.5, tileGridSize=(8, 8))
cl = clahe.apply(l)
# Merge channels back
enhanced_lab = cv2.merge((cl, a, b))
combined_result = cv2.cvtColor(enhanced_lab, cv2.COLOR_LAB2BGR)
# Extract individual region images for separate OCR processing
region_images = []
if text_regions:
for idx, (x, y, w, h) in enumerate(text_regions):
# Add padding around region (10% of width/height)
pad_x = max(5, int(w * 0.1))
pad_y = max(5, int(h * 0.1))
# Ensure coordinates stay within image bounds
x_start = max(0, x - pad_x)
y_start = max(0, y - pad_y)
x_end = min(img_rgb.shape[1], x + w + pad_x)
y_end = min(img_rgb.shape[0], y + h + pad_y)
# Extract region with padding
region = img_rgb[y_start:y_end, x_start:x_end].copy()
# Store region with its coordinates
region_info = {
'image': region,
'coordinates': (x, y, w, h),
'padded_coordinates': (x_start, y_start, x_end - x_start, y_end - y_start),
'order': idx
}
region_images.append(region_info)
# Convert visualization results back to PIL Images
text_regions_pil = Image.fromarray(cv2.cvtColor(text_regions_vis, cv2.COLOR_BGR2RGB))
image_regions_pil = Image.fromarray(cv2.cvtColor(image_regions_vis, cv2.COLOR_BGR2RGB))
combined_result_pil = Image.fromarray(cv2.cvtColor(combined_result, cv2.COLOR_BGR2RGB))
# Create base64 representation of text mask for visualization
_, buffer = cv2.imencode('.png', text_mask)
text_mask_base64 = base64.b64encode(buffer).decode('utf-8')
# Convert region images to PIL format
region_pil_images = []
for region_info in region_images:
region_pil = Image.fromarray(cv2.cvtColor(region_info['image'], cv2.COLOR_BGR2RGB))
region_info['pil_image'] = region_pil
region_pil_images.append(region_info)
# Return the segmentation results
return {
'text_regions': text_regions_pil,
'image_regions': image_regions_pil,
'text_mask_base64': f"data:image/png;base64,{text_mask_base64}",
'combined_result': combined_result_pil,
'text_regions_coordinates': text_regions,
'region_images': region_pil_images
}
except Exception as e:
logger.error(f"Error segmenting image {image_file.name}: {str(e)}")
# Return None values if processing fails
return {
'text_regions': None,
'image_regions': None,
'text_mask_base64': None,
'combined_result': None,
'text_regions_coordinates': []
}
def process_segmented_image(image_path: Union[str, Path], output_dir: Optional[Path] = None, preserve_content: bool = True) -> Dict:
"""
Process an image using segmentation for improved OCR, saving visualization outputs.
Args:
image_path: Path to the image file
output_dir: Optional directory to save visualization outputs
Returns:
Dictionary with processing results and paths to output files
"""
# Convert to Path object if string
image_file = Path(image_path) if isinstance(image_path, str) else image_path
# Create output directory if not provided
if output_dir is None:
output_dir = Path("output") / "segmentation"
output_dir.mkdir(parents=True, exist_ok=True)
# Process the image with segmentation
segmentation_results = segment_image_for_ocr(image_file)
# Prepare results dictionary
results = {
'original_image': str(image_file),
'output_files': {}
}
# Save visualization outputs if segmentation was successful
if segmentation_results['text_regions'] is not None:
# Save text regions visualization
text_regions_path = output_dir / f"{image_file.stem}_text_regions.jpg"
segmentation_results['text_regions'].save(text_regions_path)
results['output_files']['text_regions'] = str(text_regions_path)
# Save image regions visualization
image_regions_path = output_dir / f"{image_file.stem}_image_regions.jpg"
segmentation_results['image_regions'].save(image_regions_path)
results['output_files']['image_regions'] = str(image_regions_path)
# Save combined result
combined_path = output_dir / f"{image_file.stem}_combined.jpg"
segmentation_results['combined_result'].save(combined_path)
results['output_files']['combined_result'] = str(combined_path)
# Save text mask visualization
text_mask_path = output_dir / f"{image_file.stem}_text_mask.png"
# Save text mask from base64
if segmentation_results['text_mask_base64']:
base64_data = segmentation_results['text_mask_base64'].split(',')[1]
with open(text_mask_path, 'wb') as f:
f.write(base64.b64decode(base64_data))
results['output_files']['text_mask'] = str(text_mask_path)
# Add detected text regions count
results['text_regions_count'] = len(segmentation_results['text_regions_coordinates'])
results['text_regions_coordinates'] = segmentation_results['text_regions_coordinates']
return results
if __name__ == "__main__":
# Simple test - process a sample image if run directly
import sys
if len(sys.argv) > 1:
image_path = sys.argv[1]
else:
# Default to testing with the magician image
image_path = "input/magician-or-bottle-cungerer.jpg"
logger.info(f"Testing image segmentation on {image_path}")
results = process_segmented_image(image_path)
# Print results summary
logger.info(f"Segmentation complete. Found {results.get('text_regions_count', 0)} text regions.")
logger.info(f"Output files saved to: {[path for path in results.get('output_files', {}).values()]}")
|