如何用OpenCV去除棋盘工作区扫描中的相机噪声以获无畸变结果

Hey there! Let's fix that noise problem messing up your chessboard scan. Camera noise is such a nuisance—it throws off edge detection, leads to wonky contour picks, and ruins the final perspective transform. Here are some practical, code-ready tweaks to your workflow that should get you that crisp, undistorted chessboard output:

1. Boost Preprocessing for Targeted Noise Reduction #

Your current Gaussian blur helps, but camera images often have a mix of Gaussian noise and salt-and-pepper noise. Adding a median blur first will knock out those sharp, random noise specks while keeping edges intact:

gray = cv2.cvtColor(image,cv2.COLOR_BGR2GRAY)
# Handle salt-and-pepper noise with median blur (adjust kernel size if needed)
gray = cv2.medianBlur(gray, 5)
# Follow up with Gaussian blur to smooth high-frequency noise
gray = cv2.GaussianBlur(gray,(3,3),0)

If your image has more soft, blurry noise, swap the median blur for a bilateral filter—it preserves edges while blurring uniform noise:

gray = cv2.bilateralFilter(gray, 9, 75, 75)

2. Clean Up Edge Detection #

The Canny edge detector is sensitive to noise, so feeding it a cleaner input will make a huge difference. Try using adaptive thresholding first to get a binary image, then run Canny on that:

# Adaptive thresholding to isolate the chessboard from background noise
thresh = cv2.adaptiveThreshold(gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C, cv2.THRESH_BINARY_INV, 11, 2)
# Now run Canny on the thresholded image for cleaner edges
edged = cv2.Canny(thresh, 50, 150)

You can also tweak the Canny thresholds—lower the lower bound if edges are missing, or raise the upper bound if too much noise is showing up.

3. Morphological Operations to Polish the Edge Map #

Even after filtering, edges might have tiny gaps or noise blobs. Use morphological operations to clean them up:

# Create a small rectangular kernel
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3,3))
# Close small gaps in edges (dilate then erode)
edged = cv2.morphologyEx(edged, cv2.MORPH_CLOSE, kernel)
# Optional: Remove tiny noise blobs (erode then dilate)
edged = cv2.morphologyEx(edged, cv2.MORPH_OPEN, kernel)

4. Make Contour Selection More Robust #

Right now, you're only checking if a contour has 4 points—but noise can create fake quadrilateral contours. Add checks for aspect ratio (chessboards are square-ish) and minimum size to filter out junk:

for c in cnts:
    peri = cv2.arcLength(c,True)
    approx = cv2.approxPolyDP(c,0.02*peri, True)
    
    if len(approx) == 4:
        # Get bounding rectangle to check aspect ratio
        x, y, w_cnt, h_cnt = cv2.boundingRect(approx)
        aspect_ratio = float(w_cnt) / h_cnt
        # Allow a small range around 1 (adjust if your chessboard is rectangular)
        if 0.8 <= aspect_ratio <= 1.2:
            # Ensure the contour is large enough (covers at least 20% of the image)
            if cv2.contourArea(c) > (image.shape[0] * image.shape[1]) * 0.2:
                screenCnt = approx
                break

5. Fine-Tune Thresholding for the Warped Image #

The final thresholding step can also be thrown off by residual noise. Add a quick blur before applying local thresholding, or try Otsu's automatic thresholding:

warped = cv2.cvtColor(warped,cv2.COLOR_BGR2GRAY)
# Blur the warped image to reduce residual noise
warped = cv2.GaussianBlur(warped, (3,3), 0)
# Option 1: Use Otsu's thresholding for automatic optimal value
_, warped = cv2.threshold(warped, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
# Option 2: Adjust local threshold parameters (increase block size/offset)
# T = threshold_local(warped, 15, offset = 15, method = "gaussian")
# warped = (warped > T).astype("uint8")*255

If you still see distortion, sharing a bit more detail about the noise (e.g., is it tiny white/black specks, or blurry grain?) would help narrow things down further.

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