如何用OpenCV生成网格及药盒单元格的内部区域掩码？

Hey there! Let's work through your two mask-generation questions—they're actually pretty similar once you shift away from Hough Transform for the pill box task, which makes sense since Hough can be finicky with real-world camera noise and imperfect lines.

For general white-line grid images, the key is to isolate the grid lines, then extract the enclosed regions. Here's a step-by-step approach:

• Preprocess the image
  First, convert to grayscale and use adaptive thresholding to handle uneven lighting (way more reliable than fixed thresholds):

  import cv2
  import numpy as np
  
  img = cv2.imread("grid_image.jpg")
  gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
  # Adaptive threshold to turn white lines into white foreground, dark background
  binary = cv2.adaptiveThreshold(
      gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY, 11, 2
  )
  

• Clean up the grid lines
  Use morphological operations to fix broken lines and remove small noise:

  # Create a rectangular structuring element
  kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 3))
  # Close operation fills gaps in lines; open operation removes tiny noise
  closed = cv2.morphologyEx(binary, cv2.MORPH_CLOSE, kernel, iterations=2)
  opened = cv2.morphologyEx(closed, cv2.MORPH_OPEN, kernel)
  

• Extract internal region masks
  Find contours in the cleaned grid, then filter out the outer grid frame and fill the internal cells:

  contours, hierarchy = cv2.findContours(opened, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
  mask = np.zeros_like(gray)
  
  # Filter contours to keep only internal cells (adjust area ranges for your image)
  for cnt in contours:
      area = cv2.contourArea(cnt)
      # Skip the outer frame (too large) and tiny noise (too small)
      if 800 < area < (gray.shape[0] * gray.shape[1]) // 2:
          cv2.drawContours(mask, [cnt], -1, 255, thickness=cv2.FILLED)
  

  The mask variable now holds the filled internal regions of your grid.

Hough Transform struggles with real-world camera images (noise, distorted lines, uneven lighting from your Raspberry Pi cam). Let's swap to a contour-based approach that's more robust:

• Preprocess for camera images
  Add Gaussian blur to reduce camera noise before thresholding:

  # Capture image from Pi cam (or load from file)
  # img = pi_cam.capture()
  img = cv2.imread("pill_box.jpg")
  gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
  blur = cv2.GaussianBlur(gray, (5, 5), 0)
  
  # Adaptive threshold to handle shadows/lighting changes
  binary = cv2.adaptiveThreshold(
      blur, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY_INV, 15, 4
  )
  

• Isolate the pill box first
  Crop out the pill box from the background to eliminate distractions:

  # Find the largest contour (should be the pill box frame)
  contours, _ = cv2.findContours(binary, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
  pill_box_contour = max(contours, key=cv2.contourArea)
  
  # Create a mask for the pill box and crop the ROI
  box_mask = np.zeros_like(gray)
  cv2.drawContours(box_mask, [pill_box_contour], -1, 255, cv2.FILLED)
  pill_box_roi = cv2.bitwise_and(gray, gray, mask=box_mask)
  

• Extract individual cell masks
  Now find contours inside the pill box ROI, filtering for cells of similar size:

  # Re-threshold the ROI for cleaner cell lines
  _, roi_binary = cv2.threshold(pill_box_roi, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
  cell_contours, _ = cv2.findContours(roi_binary, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
  
  cell_masks = []
  # Adjust area range to match your pill cells' actual size
  for cnt in cell_contours:
      area = cv2.contourArea(cnt)
      if 600 < area < 2200:
          cell_mask = np.zeros_like(pill_box_roi)
          cv2.drawContours(cell_mask, [cnt], -1, 255, cv2.FILLED)
          cell_masks.append(cell_mask)
  

• Check for pills in each cell
  Use the masks to compare cell content against empty background (adjust the mean threshold based on your pill/box colors):

  for idx, cell_mask in enumerate(cell_masks):
      # Extract only the cell's pixels
      cell_pixels = cv2.bitwise_and(pill_box_roi, pill_box_roi, mask=cell_mask)
      # Calculate mean brightness of the cell
      mean_brightness = cv2.mean(cell_pixels, mask=cell_mask)[0]
      
      # Empty cells will be close to the box's background brightness; pills will differ
      if mean_brightness < 140:  # Tweak this value for your setup
          print(f"Cell {idx+1} has a pill!")
      else:
          print(f"Cell {idx+1} is empty.")
  

Pro Tip for Regular Pill Boxes #

If your pill box is a perfect grid (e.g., 2x7), use perspective transformation first to straighten the box into a rectangle. Then you can split the image into equal-sized cells directly (no contour hunting needed)—this is super reliable if you can detect the box's four corner points.

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