How OCR Works: The 5-Step Process

Production Preprocessing Pipeline

import cv2
import numpy as np
from scipy import ndimage
from skimage.filters import threshold_sauvola

def advanced_preprocess(image_path):
    """Production-grade preprocessing pipeline"""
    # Load image with error handling
    img = cv2.imread(image_path, cv2.IMREAD_GRAYSCALE)
    if img is None:
        raise ValueError(f"Could not load image: {image_path}")
    
    # Step 1: DPI normalization (300 DPI standard)
    height, width = img.shape
    if width > 4000:  # High-res document
        scale = 3000 / width
        new_width = int(width * scale)
        new_height = int(height * scale)
        img = cv2.resize(img, (new_width, new_height), 
                        interpolation=cv2.INTER_AREA)
    
    # Step 2: Advanced noise reduction
    # Bilateral filter preserves edges while reducing noise
    denoised = cv2.bilateralFilter(img, 9, 75, 75)
    
    # Step 3: Adaptive binarization (handles uneven lighting)
    # Sauvola method works better than Otsu for documents
    window_size = 25
    k = 0.2  # Sensitivity parameter
    binary = threshold_sauvola(denoised, window_size=window_size, k=k)
    binary = (denoised > binary).astype(np.uint8) * 255
    
    # Step 4: Skew correction using Hough transform
    edges = cv2.Canny(binary, 50, 150, apertureSize=3)
    lines = cv2.HoughLines(edges, 1, np.pi/180, threshold=100)
    
    if lines is not None:
        angles = []
        for rho, theta in lines[0]:
            angle = np.rad2deg(theta) - 90
            angles.append(angle)
        
        # Use median angle for robustness
        skew_angle = np.median(angles)
        
        if abs(skew_angle) > 0.5:  # Only correct significant skew
            (h, w) = binary.shape[:2]
            center = (w // 2, h // 2)
            M = cv2.getRotationMatrix2D(center, skew_angle, 1.0)
            binary = cv2.warpAffine(binary, M, (w, h),
                                  flags=cv2.INTER_CUBIC,
                                  borderMode=cv2.BORDER_REPLICATE)
    
    # Step 5: Morphological operations for cleanup
    # Remove small noise and fill gaps in characters
    kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (2, 2))
    binary = cv2.morphologyEx(binary, cv2.MORPH_CLOSE, kernel)
    
    return binary, {"skew_corrected": abs(skew_angle) > 0.5 if 'skew_angle' in locals() else False}