基于Go语言gocv的棋盘游戏全棋子鲁棒性检测技术咨询

Hey there! Let's work through this robust chess piece detection problem together—your current progress with contour detection is a solid foundation, so let's build on that to fix the white/dark piece issues and cut down on manual parameter tweaking.

1. Fix Preprocessing for All Piece Colors #

Your current grayscale + fixed threshold approach struggles with white and dark pieces because it's tied to color intensity directly. Instead, let's decouple color from brightness:

• Use HSV's Value Channel: Convert your frame to HSV and extract the V channel—this isolates brightness, so white (high V) and dark maroon (low V) pieces will both stand out against most backgrounds.
• Adaptive Thresholding: Ditch fixed thresholds for adaptive ones to handle uneven lighting on the board:

  // Convert to HSV and grab the Value channel
  hsvMat := gocv.NewMat()
  gocv.CvtColor(clone, &hsvMat, gocv.ColorRGBToHSV)
  channels := gocv.Split(hsvMat)
  vChannel := channels[2] // HSV order: H, S, V
  
  // Adaptive threshold to handle lighting variations
  gocv.AdaptiveThreshold(vChannel, &vChannel, 255, gocv.AdaptiveThresholdGaussianC, gocv.ThresholdBinaryInv, 11, 2)
  

• Better Morphological Operations: Use elliptical kernels (they match circular pieces better) and combine close + open operations to fill gaps in pieces and remove small noise:

  // Close operation: fill holes inside pieces
  closeKernel := gocv.GetStructuringElement(gocv.MorphEllipse, image.Pt(5,5))
  gocv.MorphologyEx(vChannel, &vChannel, gocv.MorphClose, closeKernel)
  
  // Open operation: remove tiny noise blobs
  openKernel := gocv.GetStructuringElement(gocv.MorphEllipse, image.Pt(3,3))
  gocv.MorphologyEx(vChannel, &vChannel, gocv.MorphOpen, openKernel)
  

2. Refine Contour Filtering Logic #

Your current checks are good, but adding these metrics will drastically reduce false positives and catch more valid pieces:

• Circularity: A perfect circle has a circularity of ~1.0—filter contours that fall within a reasonable range (0.7–0.9 works for most real-world circular pieces):

  area := gocv.ContourArea(cnt)
  perimeter := gocv.ArcLength(cnt, true) // Use true for closed contours
  if perimeter == 0 {
      continue
  }
  circularity := 4 * math.Pi * area / (perimeter * perimeter)
  if circularity < 0.7 {
      continue
  }
  

• Aspect Ratio: For elliptical fits, ensure the width/height ratio is close to 1 (0.8–1.2) to rule out oblong shapes:

  rect := gocv.FitEllipse(cnt)
  aspectRatio := float64(rect.Width) / float64(rect.Height)
  if aspectRatio < 0.8 || aspectRatio > 1.2 {
      continue
  }
  

• Area Bounds: Have the user click one piece once to calculate an average area, then filter contours within ±30% of that value. This eliminates tiny noise and large board artifacts.

3. Fix Hough Circles (If You Want to Use It) #

Your Hough Circle params were likely off for your specific board. Try these tweaks:

• dp = 1.5 (more precise for high-res boards)
• minDist set to the actual pixel distance between piece centers on your board
• param1 = 40 (lower Canny threshold to catch faint edges on dark pieces)
• param2 = 30 (balance between detection count and false positives)
• minRadius/maxRadius set to the actual radius range of your pieces

Combine it with contour detection: use contours to get candidate regions, then run Hough Circles only within those regions to reduce computation and noise.

4. Minimize Manual Interaction #

Your ideas for reducing slider tweaks are great—here's how to implement them simply:

• Single Piece Calibration: Let the user click/drag a box around one piece. Calculate its average area, circularity, and brightness range from the V channel. Use these values as automatic thresholds for all other pieces.
• Background Masking: Have the user click a blank spot on the board to capture the background's HSV values. Mask out the background before processing to eliminate non-piece noise:

  // After capturing background HSV range (lowBg, highBg)
  mask := gocv.NewMat()
  gocv.InRange(hsvMat, lowBg, highBg, &mask)
  gocv.BitwiseNot(mask, &mask) // Keep non-background pixels
  gocv.BitwiseAnd(clone, clone, &clone, mask)
  

• Board Grid Alignment: If your board is a regular grid, let the user click the four corners. Apply a perspective transform to straighten the board—this makes piece positions uniform and easier to detect (you can even check each grid cell for a piece).

Example Integrated Code Snippet #

Here's how to update your contour detection loop with the above fixes:

// Assume we've run calibration to get avgArea, minArea, maxArea
matchColor := color.RGBA{0, 255, 0, 0}
cnts := gocv.FindContours(vChannel, gocv.RetrievalExternal, gocv.ChainApproxSimple) // Grab only outer contours

for i := 0; i < cnts.Size(); i++ {
    cnt := cnts.At(i)
    area := gocv.ContourArea(cnt)

    // Filter by area range
    if area < minArea || area > maxArea {
        continue
    }

    // Check circularity
    perimeter := gocv.ArcLength(cnt, true)
    if perimeter == 0 {
        continue
    }
    circularity := 4 * math.Pi * area / (perimeter * perimeter)
    if circularity < 0.7 {
        continue
    }

    // Fit ellipse and check aspect ratio
    rect := gocv.FitEllipse(cnt)
    aspectRatio := float64(rect.Width) / float64(rect.Height)
    if aspectRatio < 0.8 || aspectRatio > 1.2 {
        continue
    }

    // Allow small tolerance for center alignment
    rectCenterX := (rect.BoundingRect.Max.X + rect.BoundingRect.Min.X) / 2
    rectCenterY := (rect.BoundingRect.Min.Y + rect.BoundingRect.Max.Y) / 2
    if abs(rect.Center.X - rectCenterX) > 5 || abs(rect.Center.Y - rectCenterY) > 5 {
        continue
    }

    // All checks passed—this is a valid piece
    gocv.Circle(&colorImage, image.Pt(rect.Center.X, rect.Center.Y), (rect.Height+rect.Width)/4, matchColor, 3)
    pieces = append(pieces, image.Pt(rect.Center.X, rect.Center.Y))
}

// Helper function for absolute value
func abs(x int) int {
    if x < 0 {
        return -x
    }
    return x
}

These changes should make your detection far more robust across different piece colors, and the calibration steps will cut down on manual slider tweaking significantly.

内容的提问来源于stack exchange，提问作者amlwwalker