使用OpenCV2提取验证码字符效果不佳，寻求正确操作指导

Hey there! I’ve wrestled with plenty of stubborn CAPTCHAs using OpenCV, so let me share a tried-and-true workflow that usually turns bad extraction results around. More often than not, the issue is missing or misconfigured preprocessing steps—let’s break this down properly:

Captchas are designed to mess with basic image processing, so skipping these steps will almost guarantee bad results:

• Grayscale Conversion: Strip away color channels to simplify processing. Use:

  gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
  

• Adaptive Thresholding: Forget basic global thresholding—captchas almost always have uneven lighting. Adaptive thresholding adjusts to local pixel values, making characters pop:

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

  We use THRESH_BINARY_INV to invert colors (characters white, background black) which makes contour detection easier.
• Noise Reduction: Clean up tiny speckles with morphological operations. An "open" operation (erosion followed by dilation) works great:

  kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (2, 2))
  cleaned = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, kernel)
  

  For heavier noise, add a median blur before thresholding: gray = cv2.medianBlur(gray, 3)

Once your image is clean, you need to isolate individual characters:

• Find External Contours: We only care about outer contours (not inner gaps in letters like 'O' or 'A'):

  contours, _ = cv2.findContours(cleaned.copy(), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
  

• Filter Invalid Contours: Get rid of noise and background blobs by checking area and aspect ratio (tweak these values to match your captcha):

  valid_contours = []
  for cnt in contours:
      area = cv2.contourArea(cnt)
      x, y, w, h = cv2.boundingRect(cnt)
      # Adjust thresholds based on your captcha's character size
      if 80 < area < 1200 and 0.2 < w/h < 2:
          valid_contours.append(cnt)
  

• Sort Contours Left-to-Right: Captcha characters are ordered, so sort by the x-coordinate of each contour's bounding box:

  valid_contours = sorted(valid_contours, key=lambda x: cv2.boundingRect(x)[0])
  

• Extract & Standardize Characters: Crop each character and resize to a uniform size (great for feeding into OCR or a model):

  extracted_chars = []
  for cnt in valid_contours:
      x, y, w, h = cv2.boundingRect(cnt)
      char_img = cleaned[y:y+h, x:x+w]
      # Resize to a consistent size (28x28 is standard for digit/char models)
      char_img = cv2.resize(char_img, (28, 28))
      extracted_chars.append(char_img)
  

If your captcha has lines, skewed characters, or weird distortions:

• Remove Interference Lines: Target horizontal/vertical lines with specialized kernels:

  # Remove horizontal lines
  kernel_h = cv2.getStructuringElement(cv2.MORPH_RECT, (10, 1))
  horizontal_lines = cv2.morphologyEx(thresh, cv2.MORPH_OPEN, kernel_h)
  no_lines = cv2.subtract(thresh, horizontal_lines)
  

• Skew Correction: If characters are tilted, use minAreaRect to find the rotation angle and straighten them:

  rect = cv2.minAreaRect(cnt)
  angle = rect[2]
  # Adjust angle for proper rotation (depends on contour orientation)
  if angle > 45:
      angle = 90 - angle
  # Rotate the character
  (h, w) = char_img.shape[:2]
  center = (w // 2, h // 2)
  M = cv2.getRotationMatrix2D(center, angle, 1.0)
  rotated = cv2.warpAffine(char_img, M, (w, h), flags=cv2.INTER_CUBIC, borderMode=cv2.BORDER_REPLICATE)
  

After every step, use cv2.imshow("Step Name", img) to check what’s happening. If your contours are picking up noise, tweak the area/aspect ratio thresholds. If characters are merging, adjust the kernel size in morphological operations.

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