OpenCV Haar Cascade车牌检测误报过多及漏检问题求解

Hey Johan,

I’ve dealt with similar Haar Cascade false positive issues for European license plates before, especially with the generic eu.xml model targeting Swedish plates. Since you’ve already tuned the detectMultiScale parameters to your current best, let’s focus on preprocessing tweaks and post-filtering steps that’ll cut down on false positives (road surfaces, signs) without sacrificing detection rate. Here’s a breakdown of actionable solutions:

1. Preprocessing: Reduce Noise & Enhance Plate Contrast #

Haar cascades rely heavily on consistent contrast and clear features. Tweaking your input before detection can eliminate many irrelevant regions:

• Adaptive Histogram Equalization (CLAHE)：Instead of standard equalizeHist, CLAHE preserves local contrast, which makes plate characters pop while minimizing road texture noise. Try this:

  clahe = cv.createCLAHE(clipLimit=2.0, tileGridSize=(8,8))
  gray_enhanced = clahe.apply(gray)
  # Use this enhanced grayscale for detection
  plates = lp_cascade.detectMultiScale(gray_enhanced, 1.1, 2, cv.CASCADE_DO_CANNY_PRUNING, (30,6), (160,34))
  

• ROI Restriction：If your images are taken from a vehicle, license plates are almost always on lower portions of the frame (attached to cars). Define a region of interest (ROI) that excludes upper areas (where signs are) and distant road surfaces. For example:

  # Assume plates are in the bottom 60% of the image
  roi = gray[int(gray.shape[0]*0.4):, :]
  # Adjust detection coordinates back to original image later
  plates = lp_cascade.detectMultiScale(roi, 1.1, 2, cv.CASCADE_DO_CANNY_PRUNING, (30,6), (160,34))
  for (x,y,w,h) in plates:
      x_orig = x
      y_orig = y + int(gray.shape[0]*0.4)
      # Use (x_orig, y_orig, w, h) for drawing/processing
  

2. Post-Filtering: Validate Detected Candidates #

This is the most impactful step for cutting false positives. We’ll filter out candidates that don’t match Swedish plate characteristics:

• Aspect Ratio Check：Swedish plates have a fixed aspect ratio (~4.7:1, width to height). Reject any boxes that fall outside a reasonable range (e.g., 4.2–5.2):

  valid_plates = []
  for (x,y,w,h) in plates:
      aspect_ratio = w / float(h)
      if 4.2 <= aspect_ratio <= 5.2:
          valid_plates.append((x,y,w,h))
  

• Color Validation：Swedish plates are mostly white with a blue top bar, or yellow for temporary plates. Check the color of the detected ROI from the original BGR image:

  filtered_plates = []
  for (x,y,w,h) in valid_plates:
      roi = img[y:y+h, x:x+w]
      avg_bgr = cv.mean(roi)[:3]
      # Check for white (high BGR values)
      is_white = all(val > 200 for val in avg_bgr)
      # Check for yellow (low B, high G/R)
      is_yellow = avg_bgr[0] < 50 and avg_bgr[1] > 200 and avg_bgr[2] > 200
      # Check for blue top bar (at least 10% of ROI is blue)
      blue_mask = cv.inRange(roi, (200, 0, 0), (255, 100, 100))
      has_blue = cv.countNonZero(blue_mask) > (w*h)*0.1
      if is_white or is_yellow or has_blue:
          filtered_plates.append((x,y,w,h))
  

• Non-Maximum Suppression (NMS)：Merge overlapping boxes (common with Haar cascades) and filter out low-priority candidates. Use OpenCV’s built-in NMS:

  # Prepare boxes and dummy confidences (adjust weights based on validation)
  boxes = [[x, y, x+w, y+h] for (x,y,w,h) in filtered_plates]
  confidences = [0.9 if (4.5 <= w/h <= 4.9) else 0.7 for (x,y,w,h) in filtered_plates]
  
  # Apply NMS
  indices = cv.dnn.NMSBoxes(boxes, confidences, score_threshold=0.5, nms_threshold=0.3)
  final_plates = [filtered_plates[i] for i in indices]
  

• Character Contour Check：License plates have distinct character shapes. For each candidate ROI, binarize the image and count valid character-sized contours:

  def has_valid_characters(roi_gray, min_chars=3):
      _, thresh = cv.threshold(roi_gray, 127, 255, cv.THRESH_BINARY_INV)
      contours, _ = cv.findContours(thresh, cv.RETR_EXTERNAL, cv.CHAIN_APPROX_SIMPLE)
      # Count contours that match character size (adjust based on image resolution)
      char_count = 0
      for cnt in contours:
          area = cv.contourArea(cnt)
          if 20 < area < 200:  # Tweak these values for your image scale
              char_count +=1
      return char_count >= min_chars
  
  # Add this to your filtering loop
  final_plates = []
  for (x,y,w,h) in filtered_plates:
      roi_gray = gray[y:y+h, x:x+w]
      if has_valid_characters(roi_gray):
          final_plates.append((x,y,w,h))
  

3. Minor Parameter Tweaks (If You Haven’t Tried These) #

You mentioned your current parameters are optimal, but these small adjustments might help without hurting detection:

• Increase minNeighbors slightly：Try 3 instead of 2. This requires more overlapping detections to keep a box, which reduces false positives but may miss some plates—test with your dataset.
• Tighten minSize/maxSize to match Swedish plate dimensions：Based on real-world sizes (520x110mm), adjust to (30,6) and (160,34) (scaled to your image resolution) to filter out boxes that are too small/large.

Final Notes #

Start with the post-filtering steps (aspect ratio + color check) since they’re non-intrusive and won’t reduce detection rate. Add CLAHE preprocessing next, then ROI restriction if your images have consistent plate positions. These steps should drastically cut down on road/sign false positives while keeping your valid plate detections intact.

内容的提问来源于stack exchange，提问作者Johan