如何用OpenCV识别嵌套框并计数？解决框内文字干扰问题

Hey there! Let’s work through this nested box detection and counting problem—those overlapping white characters can be a real pain, but we’ve got some practical fixes to try out.

1. First: Eliminate White Text Interference (Preprocessing) #

The white text is messing with your edge detection and thresholding, so let’s clean it up first:

• Mask & Remove White Text: If your text is pure white (or very light), create a mask to target those bright regions and overwrite them with the box background color. This keeps the box edges intact while getting rid of text:

  # Assuming your input is a grayscale image named 'gray'
  _, text_mask = cv2.threshold(gray, 240, 255, cv2.THRESH_BINARY)
  # Replace text regions with the box background (adjust based on your image's background)
  gray_clean = cv2.copyTo(gray, cv2.bitwise_not(text_mask))
  

• Morphological Opening: Use an opening operation (erosion followed by dilation) to erase small white blobs (your text) without damaging larger box structures. Tweak the kernel size based on how big your text is:

  kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 3))
  gray_clean = cv2.morphologyEx(gray, cv2.MORPH_OPEN, kernel)
  

2. Optimize Thresholding & Edge Detection #

Global thresholding might fail with text interference—switch to more robust methods:

• Adaptive Thresholding: This adjusts thresholds locally, so it’s better at distinguishing boxes from background even when text is present:

  binary = cv2.adaptiveThreshold(
      gray_clean, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, 
      cv2.THRESH_BINARY_INV, 11, 2
  )
  

• Canny Edges + Closing: Get clean edges with Canny, then use a closing operation to reconnect edges broken by text:

  edges = cv2.Canny(gray_clean, 50, 150)
  kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (2, 2))
  edges_closed = cv2.morphologyEx(edges, cv2.MORPH_CLOSE, kernel)
  

3. Detect & Count Nested Boxes Using Contour Hierarchy #

Once you have a clean binary/edge image, use contour detection with hierarchy to handle nesting:

• Retrieve Contour Hierarchy: Use cv2.RETR_TREE to get parent-child relationships between contours—perfect for identifying nested boxes:

  contours, hierarchy = cv2.findContours(binary, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
  

• Filter Out Small Contours: Ignore tiny leftover blobs (text remnants) by setting a minimum area threshold:

  min_box_area = 100  # Adjust based on your box size
  valid_boxes = [cnt for cnt in contours if cv2.contourArea(cnt) > min_box_area]
  

• Count Nested Boxes: Use the hierarchy array to track parent-child relationships. If you just need a total count, len(valid_boxes) works. For nested levels, loop through the hierarchy to count how many contours have parent/child links.

4. Extra: Validate Boxes with Contour Approximation #

Make sure you’re only counting actual rectangular boxes by approximating contours to quadrilaterals:

final_boxes = []
for cnt in valid_boxes:
    epsilon = 0.02 * cv2.arcLength(cnt, True)
    approx = cv2.approxPolyDP(cnt, epsilon, True)
    if len(approx) == 4:  # Only keep quadrilateral shapes
        final_boxes.append(approx)

# Total box count:
print(f"Total nested boxes detected: {len(final_boxes)}")

Play around with the parameters (threshold values, kernel sizes, min area) to match your specific images—every dataset is a little different!

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