romfast/roa2web-service-auto
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: Improve OCR adaptive pipeline with early exit and better pattern matching

Browse Source 
- Add adaptive 3-step OCR pipeline with early exit when all 5 fields found
- Add pattern for "C. I. F." with spaces (OCR artifact from PaddleOCR)
- Add pattern for YYYY. MM. DD date format with spaces (OMV/Petrom receipts)
- Add pattern for "OTAL TAXE" with T cut off and reversed amount position
- Make TVA rate pattern more flexible (code letter optional, handle "-21%")
- Replace logger.info with print(flush=True) for better debugging visibility
- Improve OCRPreview.vue to show extraction progress and raw OCR text

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
This commit is contained in:
Marius Mutu
2025-12-13 01:54:52 +02:00
parent 6c3dd89f6d
commit 9f06482681
9 changed files with 952 additions and 116 deletions
Download Patch File Download Diff File Expand all files Collapse all files

105
data-entry-app/backend/app/services/image_preprocessor.py
View File

@@ -23,37 +23,57 @@ class ImagePreprocessor:
raise ValueError(f"Could not load image: {path}")
return image
def pdf_to_images(self, path: Path, dpi: int = 400) -> List[np.ndarray]:
def pdf_to_images(self, path: Path, dpi: int = 300) -> List[np.ndarray]:
"""
Convert PDF to images with high DPI for better OCR.
Convert PDF to images.
Args:
path: Path to PDF file
dpi: Resolution (400 recommended for receipts, higher = better quality but slower)
dpi: Resolution (300 = fast & good quality, 400 = better but slower)
"""
if not PDF_AVAILABLE:
raise RuntimeError("pdf2image not available. Install with: pip install pdf2image")
# Use 400 DPI for better text recognition on thermal receipts
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 preprocessing pipeline for thermal receipt images.
Apply LIGHT preprocessing - better for clear PDFs.
Heavy binarization can destroy text on clear images.
"""
return self.preprocess_light(image)
Pipeline:
1. Convert to grayscale
2. Resize if too small (min 1500px width for high quality)
3. Deskew (straighten rotated text)
4. Contrast enhancement (CLAHE)
5. Denoise (Non-local means)
6. Sharpening (for clearer text edges)
7. Adaptive thresholding (binarization)
8. Morphological operations (connect broken chars)
def preprocess_light(self, image: np.ndarray) -> np.ndarray:
"""
Light preprocessing for CLEAR images (PDFs, good scans).
Preserves original quality, only enhances contrast.
"""
# 1. Grayscale
if len(image.shape) == 3:
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
else:
gray = image.copy()
Args:
image: Input image (BGR or grayscale)
high_quality: If True, apply more aggressive preprocessing
# 2. Resize if too small
height, width = gray.shape
if width < 1500:
scale = 1500 / width
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_heavy(self, image: np.ndarray) -> np.ndarray:
"""
Heavy preprocessing for FADED thermal receipts.
Aggressive binarization to recover faded text.
"""
# 1. Grayscale
if len(image.shape) == 3:
@@ -63,57 +83,48 @@ class ImagePreprocessor:
# 2. Resize if too small (larger = better OCR)
height, width = gray.shape
min_width = 1500 if high_quality else 1000
if width < min_width:
scale = min_width / width
gray = cv2.resize(
gray, None, fx=scale, fy=scale,
interpolation=cv2.INTER_CUBIC
)
if width < 1500:
scale = 1500 / width
gray = cv2.resize(gray, None, fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC)
# 3. Deskew
gray = self._deskew(gray)
# 4. Contrast enhancement with CLAHE (Contrast Limited Adaptive Histogram Equalization)
# 4. Contrast enhancement with CLAHE
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
enhanced = clahe.apply(gray)
# 5. Denoise (slightly less aggressive to preserve text details)
denoised = cv2.fastNlMeansDenoising(
enhanced, h=8, # Lower h = preserve more details
templateWindowSize=7,
searchWindowSize=21
)
# 5. Denoise
denoised = cv2.fastNlMeansDenoising(enhanced, h=8, templateWindowSize=7, searchWindowSize=21)
# 6. Sharpening to enhance text edges
if high_quality:
# Unsharp mask for better text clarity
gaussian = cv2.GaussianBlur(denoised, (0, 0), 2.0)
sharpened = cv2.addWeighted(denoised, 1.5, gaussian, -0.5, 0)
else:
sharpened = denoised
# 6. Sharpening
gaussian = cv2.GaussianBlur(denoised, (0, 0), 2.0)
sharpened = cv2.addWeighted(denoised, 1.5, gaussian, -0.5, 0)
# 7. Adaptive thresholding with optimized parameters
# 7. Adaptive thresholding (binarization)
binary = cv2.adaptiveThreshold(
sharpened, 255,
cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
cv2.THRESH_BINARY,
blockSize=11, # Smaller block = better for small text
C=5 # Lower C = darker result, better for faded receipts
blockSize=11, C=5
)
# 8. Morphological operations
# Close small gaps in characters
kernel_close = cv2.getStructuringElement(cv2.MORPH_RECT, (2, 2))
result = cv2.morphologyEx(binary, cv2.MORPH_CLOSE, kernel_close)
# Optional: Remove small noise spots
if high_quality:
kernel_open = cv2.getStructuringElement(cv2.MORPH_RECT, (1, 1))
result = cv2.morphologyEx(result, cv2.MORPH_OPEN, kernel_open)
return result
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."""
edges = cv2.Canny(image, 50, 150, apertureSize=3)