基于Raspberry Pi 4的Python CAPTCHA图像逆扭曲实现求助

Hey Michael, glad to hear manual Photoshop fixes work with Tesseract—let's replicate that automatedly using pure Python tools that play seamlessly with Raspbian, since you're having OpenCV deployment issues. Here's a step-by-step plan:

1. Install Required Libraries #

First, make sure you have these lightweight alternatives to OpenCV installed:

• Pillow (for image manipulation)
• numpy (for numerical calculations)
• pytesseract (for OCR)
• Tesseract OCR engine itself

Run these commands in your terminal:

sudo apt update && sudo apt install tesseract-ocr
pip install pillow numpy pytesseract

2. Understand the S-Curve Warp Model #

Your CAPTCHA uses a horizontal S-shaped warp, which is essentially a sine wave applied to the x-coordinate of each pixel. For a given pixel at (x, y) in the original flat image, the warped position (x', y) would look like:
x' = x + A * sin(2πy/H)
Where:

• A = amplitude of the warp (how much pixels shift horizontally)
• H = height of the CAPTCHA image

To reverse this, we calculate the original x-coordinate from the warped one:
x = x' - A * sin(2πy/H)

3. Implement the Unwarping Function #

Here's a Python script that applies this inverse warp, then passes the flattened image to Tesseract for recognition:

from PIL import Image
import numpy as np
import pytesseract

def unwarp_s_captcha(img_path, warp_amplitude=20):
    # Load image and convert to grayscale (improves Tesseract accuracy)
    img = Image.open(img_path).convert('L')
    img_np = np.array(img)
    height, width = img_np.shape

    # Initialize empty array for unwarped image
    unwarped_img_np = np.full_like(img_np, 255)  # Fill background with white

    # Iterate over each pixel row to apply inverse warp
    for y in range(height):
        # Calculate horizontal offset for this row based on sine curve
        offset = int(warp_amplitude * np.sin(2 * np.pi * y / height))
        for x in range(width):
            # Compute original x-coordinate before warp
            original_x = x - offset
            # Only copy pixels that fall within the original image bounds
            if 0 <= original_x < width:
                unwarped_img_np[y, x] = img_np[y, original_x]

    # Convert back to PIL Image
    unwarped_img = Image.fromarray(unwarped_img_np)
    return unwarped_img

# Example usage
if __name__ == "__main__":
    # Adjust warp_amplitude based on your CAPTCHA's distortion level
    unwarped_image = unwarp_s_captcha("your_captcha.png", warp_amplitude=18)
    unwarped_image.save("unwarped_captcha.png")

    # Configure Tesseract to focus on alphanumeric characters only
    tesseract_config = "--psm 10 --oem 3 -c tessedit_char_whitelist=ABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789"
    captcha_text = pytesseract.image_to_string(unwarped_image, config=tesseract_config)
    print(f"Recognized CAPTCHA: {captcha_text.strip()}")

4. Fine-Tune for Your CAPTCHA #

• Adjust warp_amplitude: Start with a value between 15-25, then tweak it based on how well the unwarped image looks (open unwarped_captcha.png to check). If characters are still curved, increase the amplitude; if they're shifted too far, decrease it.
• Add Preprocessing: For better Tesseract accuracy, add a binarization step after converting to grayscale:

  # Add this right after converting to grayscale
  img = img.point(lambda pixel: 0 if pixel < 127 else 255, '1')
  

  This turns the image into pure black-and-white, eliminating gray noise.

5. Test with Multiple CAPTCHA Samples #

Once you have your amplitude dialed in, test the script with your uploaded CAPTCHA examples to ensure consistency. If some samples have slightly different warp intensities, you could add a small adjustment loop to auto-tune the amplitude based on edge detection (using Pillow's edge filters), but the manual tweak should work for most uniform CAPTCHAs.

内容的提问来源于stack exchange，提问作者Michael