求助：基于Python的三类图像文本分割（Otsu算法失效）

Hey there! Let's work through your text binarization challenges. You noted that Otsu's global thresholding works perfectly for your noise-free third image type, but fails on the other two—let's apply targeted techniques to fix each failure case:

Problem 1: Light Gray Text on the Right Fails to Segment #

Your first image has text with lower grayscale values on the right, which gets lost with Otsu's global threshold (since it calculates one threshold for the entire image).

Solution: Adaptive Thresholding #

Adaptive thresholding calculates thresholds for small local regions, making it ideal for uneven lighting or varying text contrast. Here's how to implement it:

import cv2

# Read grayscale image
gray = cv2.imread("/home/shrouk/Pictures/f1.png", 0)

# Apply adaptive thresholding (adjust block size and C value based on your image)
# BLOCK_SIZE must be odd; C is subtracted from the local mean to adjust threshold sensitivity
thresholded = cv2.adaptiveThreshold(
    gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY_INV, 11, 2
)

cv2.imshow("Fixed Light Text Segmentation", thresholded)
cv2.waitKey(0)
cv2.destroyAllWindows()

If adaptive thresholding still leaves some noise, you can add a small Gaussian blur first to smooth the image:

blurred = cv2.GaussianBlur(gray, (3, 3), 0)
thresholded = cv2.adaptiveThreshold(blurred, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY_INV, 11, 2)

Problem 2: Dark Background Causes Text Expansion (Bloating) #

Otsu's thresholding on dark background images can incorrectly include dark background pixels as part of the text, leading to bloated characters.

Solution: Reverse Thresholding + Morphological Erosion #

First, invert the image (since text is likely lighter than the dark background), apply Otsu, then use erosion to shrink the bloated text back to its original shape:

import cv2
import numpy as np

# Read grayscale image
gray = cv2.imread("/home/shrouk/Pictures/f2.png", 0)

# Invert the image (light text on dark background becomes dark text on light)
inverted_gray = cv2.bitwise_not(gray)

# Apply Otsu's thresholding on the inverted image
_, thresholded = cv2.threshold(inverted_gray, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)

# Define a small kernel for morphological erosion (adjust size based on bloat severity)
kernel = np.ones((2, 2), np.uint8)
eroded = cv2.erode(thresholded, kernel, iterations=1)

cv2.imshow("Fixed Bloated Text", eroded)
cv2.waitKey(0)
cv2.destroyAllWindows()

Alternatively, you can use adaptive thresholding with inverse binary here too—often it avoids the need for post-processing erosion:

thresholded = cv2.adaptiveThreshold(
    gray, 255, cv2.ADAPTIVE_THRESH_MEAN_C, cv2.THRESH_BINARY, 15, 4
)

For Your Noise-Free Third Image #

Your existing Otsu code works great here, but you can keep it as-is or add a quick median blur to ensure no tiny noise creeps in:

gray = cv2.imread("/home/shrouk/Pictures/f3.png", 0)
blurred = cv2.medianBlur(gray, 3)
_, thresholded = cv2.threshold(blurred, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
cv2.imshow("Noise-Free Segmentation", thresholded)

Key Notes #

• Adjust block sizes, C values, and kernel sizes based on your specific image dimensions and text thickness.
• Always test with your actual images—tweaking these parameters is often needed for perfect results.

内容的提问来源于stack exchange，提问作者shrouk mansour