如何用pytesseract与OpenCV识别被遮挡及带横线的变形文本？

Let's tackle those two annoying recognition problems you're facing—missing text near horizontal lines and failed detection when there's large content above. Here are actionable tweaks and configurations to get pytesseract working reliably:

1. Fix Missing Text Near Horizontal Lines #

Horizontal lines can confuse tesseract's text detection by merging with text regions. Try these steps:

a. Preprocess the Image to Remove Lines #

Use OpenCV to detect and erase horizontal lines before passing to pytesseract. This cleans up the image so tesseract focuses on text:

import cv2
import numpy as np
from pytesseract import Output, image_to_data

def remove_horizontal_lines(img):
    # Convert to grayscale
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
    # Create a horizontal kernel to target lines
    kernel = np.ones((1, 15), np.uint8)
    # Detect horizontal lines using morphological operations
    detected_lines = cv2.morphologyEx(gray, cv2.MORPH_OPEN, kernel, iterations=2)
    # Subtract lines from original grayscale image
    cleaned_img = cv2.subtract(gray, detected_lines)
    # Apply threshold to get a high-contrast binary image
    _, thresh_img = cv2.threshold(cleaned_img, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)
    return thresh_img

# Usage example
processed_img = remove_horizontal_lines(your_input_img)
td = image_to_data(processed_img, output_type=Output.DICT)

b. Adjust Tesseract's Text Detection Thresholds #

Add config flags to make tesseract more sensitive to text near lines:

config = r'--oem 3 --psm 6 -c textord_min_linesize=1 textord_max_linesize=50'
td = image_to_data(img, output_type=Output.DICT, config=config)

• --oem 3: Uses the default hybrid engine (combines LSTM and legacy models for better compatibility)
• --psm 6: Treats the image as a single uniform text block (ideal for structured content like your examples)
• textord_min_linesize=1: Lowers the minimum line height threshold so short lines near horizontal rules aren't ignored
• textord_max_linesize=50: Sets a reasonable maximum line height to avoid merging text with large line elements

2. Fix Failed Detection of Lower Text with Large Content Above #

When there's a large image or text block at the top, tesseract might skip lower regions if it thinks they're not part of the main text. Fix this with better page segmentation:

a. Use a Flexible Page Segmentation Mode #

Instead of the default, use --psm 11 which tells tesseract to search for sparse text across the entire image, even with large gaps or blocks:

config = r'--oem 3 --psm 11'
td = image_to_data(img, output_type=Output.DICT, config=config)

If your content is mostly structured but has large gaps, --psm 3 (fully automatic page segmentation) combined with --psm 6 can also work, but psm 11 is more reliable for scattered layouts.

b. Disable Aggressive Text Region Filtering #

Add flags to prevent tesseract from discarding small or distant text regions:

config = r'--oem 3 --psm 11 -c textord_noise_rejection_threshold=0.1 textord_min_xheight=2'
td = image_to_data(img, output_type=Output.DICT, config=config)

• textord_noise_rejection_threshold=0.1: Reduces noise rejection so smaller text blocks aren't mistaken for noise
• textord_min_xheight=2: Lowers the minimum character height threshold to catch smaller text below large elements

3. Additional Debugging Tips #

• Visualize Detection Boxes: Print the bounding boxes from image_to_data to see which regions tesseract is picking up. This helps confirm if the issue is detection or recognition:

  for o in range(len(td['level'])):
      x, y, w, h = td['left'][o], td['top'][o], td['width'][o], td['height'][o]
      cv2.rectangle(img, (x, y), (x+w, y+h), (0, 255, 0), 2)
  cv2.imshow('Detection Boxes', img)
  cv2.waitKey(0)
  

• Try Adaptive Thresholding: If your image has uneven lighting, use adaptive thresholding instead of Otsu's to improve contrast:

  adaptive_thresh = cv2.adaptiveThreshold(gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY_INV, 11, 2)
  

内容的提问来源于stack exchange，提问作者Navpreet Devpuri