367 lines
13 KiB
Python
367 lines
13 KiB
Python
"""Image preprocessing for optimal OCR results."""
|
|
|
|
from pathlib import Path
|
|
from typing import List
|
|
|
|
import numpy as np
|
|
import cv2
|
|
|
|
try:
|
|
import pdf2image
|
|
PDF_AVAILABLE = True
|
|
except ImportError:
|
|
PDF_AVAILABLE = False
|
|
|
|
|
|
class ImagePreprocessor:
|
|
"""Preprocess receipt images for OCR."""
|
|
|
|
def _add_safety_padding(self, image: np.ndarray, padding: int = 50) -> np.ndarray:
|
|
"""Add white padding around image to protect edge content during rotation.
|
|
|
|
This prevents left/right margin truncation in OCR by ensuring text near
|
|
edges isn't lost during deskew rotation.
|
|
"""
|
|
if len(image.shape) == 2:
|
|
# Grayscale
|
|
return cv2.copyMakeBorder(
|
|
image, padding, padding, padding, padding,
|
|
cv2.BORDER_CONSTANT, value=255
|
|
)
|
|
else:
|
|
# Color (BGR)
|
|
return cv2.copyMakeBorder(
|
|
image, padding, padding, padding, padding,
|
|
cv2.BORDER_CONSTANT, value=(255, 255, 255)
|
|
)
|
|
|
|
def load_image(self, path: Path) -> np.ndarray:
|
|
"""Load image from file."""
|
|
image = cv2.imread(str(path))
|
|
if image is None:
|
|
raise ValueError(f"Could not load image: {path}")
|
|
return image
|
|
|
|
def pdf_to_images(self, path: Path, dpi: int = 300) -> List[np.ndarray]:
|
|
"""
|
|
Convert PDF to images.
|
|
|
|
Args:
|
|
path: Path to PDF file
|
|
dpi: Resolution (300 = fast & good quality, 400 = better but slower)
|
|
"""
|
|
if not PDF_AVAILABLE:
|
|
raise RuntimeError("pdf2image not available. Install with: pip install pdf2image")
|
|
images = pdf2image.convert_from_path(str(path), dpi=dpi)
|
|
return [np.array(img) for img in images]
|
|
|
|
def preprocess(self, image: np.ndarray, high_quality: bool = True) -> np.ndarray:
|
|
"""
|
|
Apply LIGHT preprocessing - better for clear PDFs.
|
|
Heavy binarization can destroy text on clear images.
|
|
"""
|
|
return self.preprocess_light(image)
|
|
|
|
def preprocess_light(self, image: np.ndarray) -> np.ndarray:
|
|
"""
|
|
Light preprocessing for CLEAR images (PDFs, good scans).
|
|
Preserves original quality, only enhances contrast.
|
|
"""
|
|
# 0. Add safety padding to protect edge content during deskew rotation
|
|
image = self._add_safety_padding(image)
|
|
|
|
# 1. Grayscale
|
|
if len(image.shape) == 3:
|
|
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
|
|
else:
|
|
gray = image.copy()
|
|
|
|
# 2a. Scale DOWN if any side exceeds 4000px (PaddleOCR limit)
|
|
height, width = gray.shape
|
|
max_side = max(height, width)
|
|
if max_side > 4000:
|
|
scale = 4000 / max_side
|
|
gray = cv2.resize(gray, None, fx=scale, fy=scale, interpolation=cv2.INTER_AREA)
|
|
height, width = gray.shape
|
|
|
|
# 2b. Scale UP if too small
|
|
if width < 1500:
|
|
scale = 1500 / width
|
|
# Ensure we don't exceed 4000px after upscaling
|
|
new_width = int(width * scale)
|
|
new_height = int(height * scale)
|
|
if max(new_width, new_height) > 4000:
|
|
scale = 4000 / max(new_width, new_height)
|
|
gray = cv2.resize(gray, None, fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC)
|
|
|
|
# 3. Deskew
|
|
gray = self._deskew(gray)
|
|
|
|
# 4. Light contrast enhancement only
|
|
clahe = cv2.createCLAHE(clipLimit=1.5, tileGridSize=(8, 8))
|
|
enhanced = clahe.apply(gray)
|
|
|
|
# NO binarization, NO morphological ops - preserve original quality
|
|
return enhanced
|
|
|
|
def preprocess_medium(self, image: np.ndarray) -> np.ndarray:
|
|
"""
|
|
Medium preprocessing for MIXED-QUALITY images.
|
|
Balance between Light (too gentle) and Heavy (too aggressive).
|
|
|
|
Use cases:
|
|
- Moderately faded receipts
|
|
- Photos with uneven lighting
|
|
- Scans with slight blur
|
|
|
|
Preprocessing steps:
|
|
- Moderate contrast enhancement (CLAHE clipLimit=2.0)
|
|
- Light denoising (fastNlMeansDenoising h=6)
|
|
- Gentle sharpening
|
|
- NO binarization (preserves text boundaries)
|
|
- NO morphological operations (avoids digit concatenation)
|
|
|
|
This method was created to replace preprocess_heavy() which caused
|
|
digit concatenation errors on high-quality PDFs (85.99 → 859,762.16).
|
|
"""
|
|
# 0. Add safety padding to protect edge content during deskew rotation
|
|
image = self._add_safety_padding(image)
|
|
|
|
# 1. Grayscale
|
|
if len(image.shape) == 3:
|
|
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
|
|
else:
|
|
gray = image.copy()
|
|
|
|
# 2a. Scale DOWN if any side exceeds 4000px (PaddleOCR limit)
|
|
height, width = gray.shape
|
|
max_side = max(height, width)
|
|
if max_side > 4000:
|
|
scale = 4000 / max_side
|
|
gray = cv2.resize(gray, None, fx=scale, fy=scale, interpolation=cv2.INTER_AREA)
|
|
height, width = gray.shape
|
|
|
|
# 2b. Scale UP if too small
|
|
if width < 1500:
|
|
scale = 1500 / width
|
|
# Ensure we don't exceed 4000px after upscaling
|
|
new_width = int(width * scale)
|
|
new_height = int(height * scale)
|
|
if max(new_width, new_height) > 4000:
|
|
scale = 4000 / max(new_width, new_height)
|
|
gray = cv2.resize(gray, None, fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC)
|
|
|
|
# 3. Deskew
|
|
gray = self._deskew(gray)
|
|
|
|
# 4. Moderate contrast enhancement (CLAHE clipLimit=2.0)
|
|
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
|
|
enhanced = clahe.apply(gray)
|
|
|
|
# 5. Light denoising (less aggressive than Heavy)
|
|
denoised = cv2.fastNlMeansDenoising(enhanced, h=6, templateWindowSize=7, searchWindowSize=15)
|
|
|
|
# 6. Gentle sharpening
|
|
gaussian = cv2.GaussianBlur(denoised, (0, 0), 1.0)
|
|
sharpened = cv2.addWeighted(denoised, 1.3, gaussian, -0.3, 0)
|
|
|
|
# NO binarization, NO morphological operations
|
|
# This preserves text boundaries and avoids digit concatenation
|
|
return sharpened
|
|
|
|
def preprocess_heavy(self, image: np.ndarray) -> np.ndarray:
|
|
"""
|
|
Heavy preprocessing for FADED thermal receipts.
|
|
Aggressive binarization to recover faded text.
|
|
|
|
⚠️ DEPRECATED: Use preprocess_medium() instead.
|
|
Heavy preprocessing causes digit concatenation on clear PDFs
|
|
(e.g., 85.99 → 859,762.16 due to binarization + morphological operations).
|
|
Kept for backward compatibility only.
|
|
"""
|
|
# 0. Add safety padding to protect edge content during deskew rotation
|
|
image = self._add_safety_padding(image)
|
|
|
|
# 1. Grayscale
|
|
if len(image.shape) == 3:
|
|
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
|
|
else:
|
|
gray = image.copy()
|
|
|
|
# 2a. Scale DOWN if any side exceeds 4000px (PaddleOCR limit)
|
|
height, width = gray.shape
|
|
max_side = max(height, width)
|
|
if max_side > 4000:
|
|
scale = 4000 / max_side
|
|
gray = cv2.resize(gray, None, fx=scale, fy=scale, interpolation=cv2.INTER_AREA)
|
|
height, width = gray.shape
|
|
|
|
# 2b. Scale UP if too small (larger = better OCR)
|
|
if width < 1500:
|
|
scale = 1500 / width
|
|
# Ensure we don't exceed 4000px after upscaling
|
|
new_width = int(width * scale)
|
|
new_height = int(height * scale)
|
|
if max(new_width, new_height) > 4000:
|
|
scale = 4000 / max(new_width, new_height)
|
|
gray = cv2.resize(gray, None, fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC)
|
|
|
|
# 3. Deskew
|
|
gray = self._deskew(gray)
|
|
|
|
# 4. Contrast enhancement with CLAHE
|
|
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
|
|
enhanced = clahe.apply(gray)
|
|
|
|
# 5. Denoise
|
|
denoised = cv2.fastNlMeansDenoising(enhanced, h=8, templateWindowSize=7, searchWindowSize=21)
|
|
|
|
# 6. Sharpening
|
|
gaussian = cv2.GaussianBlur(denoised, (0, 0), 2.0)
|
|
sharpened = cv2.addWeighted(denoised, 1.5, gaussian, -0.5, 0)
|
|
|
|
# 7. Adaptive thresholding (binarization)
|
|
binary = cv2.adaptiveThreshold(
|
|
sharpened, 255,
|
|
cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
|
|
cv2.THRESH_BINARY,
|
|
blockSize=11, C=5
|
|
)
|
|
|
|
# 8. Morphological operations
|
|
kernel_close = cv2.getStructuringElement(cv2.MORPH_RECT, (2, 2))
|
|
result = cv2.morphologyEx(binary, cv2.MORPH_CLOSE, kernel_close)
|
|
|
|
return result
|
|
|
|
def preprocess_for_tesseract(self, image: np.ndarray, binarize: bool = False,
|
|
padding: int = 0, clahe_clip: float = 1.5) -> np.ndarray:
|
|
"""
|
|
Tesseract-optimized preprocessing (based on comprehensive benchmark).
|
|
|
|
BENCHMARK FINDINGS:
|
|
- DPI 200 is optimal (not 300!)
|
|
- Padding 40px fixes left margin truncation issues
|
|
- CLAHE 1.5 for most receipts, 2.0 for difficult ones
|
|
- NO deskew, NO denoising for clear PDFs
|
|
|
|
Recommended usage:
|
|
- Simple receipts: padding=0, clahe_clip=1.5
|
|
- Complex receipts: padding=40, clahe_clip=1.5
|
|
- Difficult/faded: padding=40, clahe_clip=2.0, binarize=True
|
|
|
|
Args:
|
|
image: Input image (RGB from pdf2image or BGR from OpenCV)
|
|
binarize: Apply Otsu binarization (for faded receipts)
|
|
padding: White padding in pixels (40px recommended for edge protection)
|
|
clahe_clip: CLAHE clip limit (1.5 normal, 2.0 for difficult)
|
|
|
|
Returns:
|
|
Preprocessed grayscale image
|
|
"""
|
|
# 1. Grayscale (handle both RGB and BGR)
|
|
if len(image.shape) == 3:
|
|
gray = cv2.cvtColor(image, cv2.COLOR_RGB2GRAY)
|
|
else:
|
|
gray = image.copy()
|
|
|
|
# 2. Add padding if specified (protects against left margin truncation)
|
|
if padding > 0:
|
|
gray = cv2.copyMakeBorder(
|
|
gray, padding, padding, padding, padding,
|
|
cv2.BORDER_CONSTANT, value=255
|
|
)
|
|
|
|
# 3. CLAHE contrast enhancement
|
|
clahe = cv2.createCLAHE(clipLimit=clahe_clip, tileGridSize=(8, 8))
|
|
enhanced = clahe.apply(gray)
|
|
|
|
# NO deskew, NO denoising - these DEGRADE quality on clear PDFs!
|
|
|
|
if not binarize:
|
|
return enhanced
|
|
|
|
# Binarization only for faded receipts
|
|
_, binary = cv2.threshold(enhanced, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
|
|
|
|
# Ensure correct polarity
|
|
if np.mean(binary) < 127:
|
|
binary = 255 - binary
|
|
|
|
return binary
|
|
|
|
def preprocess_for_tesseract_padded(self, image: np.ndarray) -> np.ndarray:
|
|
"""
|
|
Tesseract preprocessing with optimal padding (40px).
|
|
|
|
Best for complex receipts where left margin gets truncated.
|
|
"""
|
|
return self.preprocess_for_tesseract(image, padding=40)
|
|
|
|
def preprocess_for_tesseract_faded(self, image: np.ndarray) -> np.ndarray:
|
|
"""
|
|
Tesseract preprocessing for FADED thermal receipts.
|
|
Uses binarization to recover faded text.
|
|
"""
|
|
return self.preprocess_for_tesseract(image, binarize=True)
|
|
|
|
def get_all_variants(self, image: np.ndarray) -> List[np.ndarray]:
|
|
"""
|
|
Generate 2 preprocessing variants for OCR (fast mode).
|
|
Returns: [light_processed, heavy_processed]
|
|
"""
|
|
return [
|
|
self.preprocess_light(image),
|
|
self.preprocess_heavy(image),
|
|
]
|
|
|
|
def _deskew(self, image: np.ndarray) -> np.ndarray:
|
|
"""Correct image rotation/skew using Hough lines.
|
|
|
|
Uses expanded canvas to preserve all content during rotation,
|
|
preventing left/right margin truncation.
|
|
"""
|
|
edges = cv2.Canny(image, 50, 150, apertureSize=3)
|
|
lines = cv2.HoughLinesP(
|
|
edges, 1, np.pi / 180,
|
|
threshold=100, minLineLength=100, maxLineGap=10
|
|
)
|
|
|
|
if lines is None:
|
|
return image
|
|
|
|
angles = []
|
|
for line in lines:
|
|
x1, y1, x2, y2 = line[0]
|
|
angle = np.arctan2(y2 - y1, x2 - x1) * 180 / np.pi
|
|
if abs(angle) < 45:
|
|
angles.append(angle)
|
|
|
|
if not angles:
|
|
return image
|
|
|
|
median_angle = np.median(angles)
|
|
if abs(median_angle) < 0.5:
|
|
return image
|
|
|
|
h, w = image.shape[:2]
|
|
center = (w // 2, h // 2)
|
|
M = cv2.getRotationMatrix2D(center, median_angle, 1.0)
|
|
|
|
# Calculate new canvas size to fit entire rotated image (prevents edge truncation)
|
|
cos_angle = abs(np.cos(np.radians(median_angle)))
|
|
sin_angle = abs(np.sin(np.radians(median_angle)))
|
|
new_w = int(h * sin_angle + w * cos_angle)
|
|
new_h = int(h * cos_angle + w * sin_angle)
|
|
|
|
# Adjust rotation matrix for new canvas center
|
|
M[0, 2] += (new_w - w) / 2
|
|
M[1, 2] += (new_h - h) / 2
|
|
|
|
return cv2.warpAffine(
|
|
image, M, (new_w, new_h),
|
|
flags=cv2.INTER_CUBIC,
|
|
borderMode=cv2.BORDER_CONSTANT,
|
|
borderValue=255 # White background (grayscale)
|
|
)
|