Blurring and smoothing (Gaussian, median, bilateral) in Computer Vision - Model Metrics & Evaluation

Metrics & Evaluation - Blurring and smoothing (Gaussian, median, bilateral)

Which metric matters for this concept and WHY

For blurring and smoothing in images, the key metric is Mean Squared Error (MSE) or Peak Signal-to-Noise Ratio (PSNR) when comparing the smoothed image to a clean reference. These metrics show how well noise is reduced without losing important details. Another important metric is Edge Preservation Index, which measures how well edges are kept sharp after smoothing. This matters because smoothing should reduce noise but keep edges clear for good image understanding.

Confusion matrix or equivalent visualization (ASCII)

Original Image:    Noisy Image:       Smoothed Image:
+---------+       +---------+       +---------+
|  Detail |       |  Noise  |       |  Smooth |
|  Edges  |       |  Added  |       |  Edges  |
+---------+       +---------+       +---------+

Metrics compare pixel differences:
MSE = sum((Original - Smoothed)^2) / total_pixels
PSNR = 10 * log10((MAX_pixel^2) / MSE)

Edge Preservation Index:
High value means edges kept sharp after smoothing.
    

Precision vs Recall (or equivalent tradeoff) with concrete examples

In smoothing, there is a tradeoff between noise removal and edge preservation:

• Gaussian blur removes noise well but can blur edges, losing details.
• Median blur removes salt-and-pepper noise and keeps edges better than Gaussian.
• Bilateral filter removes noise while preserving edges best but is slower.

Choosing a smoothing method depends on what matters more: clean image (noise removal) or sharp edges (details).

What "good" vs "bad" metric values look like for this use case

Good smoothing:

• Low MSE (close to zero) indicating noise is reduced.
• High PSNR (above 30 dB) showing good image quality.
• High Edge Preservation Index (close to 1) meaning edges are sharp.

Bad smoothing:

• High MSE showing noise or distortion remains.
• Low PSNR (below 20 dB) indicating poor quality.
• Low Edge Preservation Index (close to 0) meaning edges are blurred.

Metrics pitfalls (accuracy paradox, data leakage, overfitting indicators)

• Over-smoothing: Metrics like MSE may improve but image details are lost, hurting downstream tasks.
• Ignoring edge preservation: Only measuring noise reduction can miss that edges are blurred.
• Data leakage: Using the same noisy image as reference inflates quality metrics.
• Accuracy paradox: A very smooth but blurry image may score well on noise metrics but look worse to humans.

Self-check

Your smoothing model reduces noise well with a low MSE but the Edge Preservation Index is very low. Is it good for image tasks that need sharp edges? Why or why not?

Answer: No, because while noise is reduced, edges are blurred. This harms tasks like object detection that rely on clear edges. A balance is needed.

Key Result

Effective smoothing balances low noise (low MSE) with high edge preservation for best image quality.

Practice

1. Which filter is best for removing salt-and-pepper noise while preserving edges in an image?

easy

A. Bilateral filter

B. Median filter

C. Gaussian filter

D. Box filter

Solution

1. Step 1: Understand salt-and-pepper noise characteristics

   Salt-and-pepper noise appears as random black and white pixels, which median filters handle well by replacing each pixel with the median of neighbors.

2. Step 2: Compare filters for edge preservation

   Median filters remove noise without blurring edges, unlike Gaussian filters which blur edges, and bilateral filters which are more complex but less effective for salt-and-pepper noise.

3. Final Answer:

   Median filter -> Option B

4. Quick Check:

   Salt-and-pepper noise = Median filter [OK]

Hint: Median filter excels at salt-and-pepper noise removal [OK]

Common Mistakes:

• Choosing Gaussian filter which blurs edges
• Confusing bilateral filter with median filter
• Using box filter which blurs noise and edges

2. Which of the following is the correct OpenCV function call to apply a Gaussian blur with a 5x5 kernel on an image stored in img?

easy

A. cv2.GaussianBlur(img, (5,5), 0)

B. cv2.medianBlur(img, 5)

C. cv2.blur(img, (5,5))

D. cv2.bilateralFilter(img, 5, 75, 75)

Solution

1. Step 1: Identify Gaussian blur function syntax

   OpenCV's GaussianBlur requires the image, kernel size as a tuple, and sigma (0 means auto).

2. Step 2: Match options to syntax

   cv2.GaussianBlur(img, (5,5), 0) matches the correct function and parameters for Gaussian blur with 5x5 kernel.

3. Final Answer:

   cv2.GaussianBlur(img, (5,5), 0) -> Option A

4. Quick Check:

   Gaussian blur syntax = cv2.GaussianBlur(img, (5,5), 0) [OK]

Hint: GaussianBlur uses tuple kernel size and sigma parameter [OK]

Common Mistakes:

• Using medianBlur for Gaussian blur
• Passing kernel size as single integer instead of tuple
• Confusing bilateralFilter parameters

3. What will be the effect of applying a bilateral filter with parameters d=9, sigmaColor=75, sigmaSpace=75 on a noisy image?

medium

A. Sharpens edges and increases noise

B. Blurs the entire image uniformly

C. Removes only salt-and-pepper noise

D. Smooths noise while preserving edges

Solution

1. Step 1: Understand bilateral filter parameters

   d controls neighborhood size; sigmaColor controls color similarity; sigmaSpace controls spatial closeness. Together they smooth noise but keep edges sharp.

2. Step 2: Analyze filter effect on noisy image

   Bilateral filter smooths noise while preserving edges, unlike Gaussian which blurs edges or median which targets salt-and-pepper noise.

3. Final Answer:

   Smooths noise while preserving edges -> Option D

4. Quick Check:

   Bilateral filter = noise smoothing + edge preservation [OK]

Hint: Bilateral filter smooths noise but keeps edges sharp [OK]

Common Mistakes:

• Thinking bilateral filter blurs edges
• Confusing bilateral with median filter
• Assuming it only removes salt-and-pepper noise

4. You applied a median filter with kernel size 3 on an image but the noise is still visible. What is the likely mistake?

medium

A. Median filter does not remove noise

B. Kernel size must be even number

C. Kernel size is too small to remove noise effectively

D. Median filter blurs edges too much

Solution

1. Step 1: Understand median filter kernel size effect

   Small kernel sizes may not cover enough pixels to remove noise effectively.

2. Step 2: Identify correct kernel size usage

   Median filter kernels must be odd-sized and larger kernels remove more noise but may blur details.

3. Final Answer:

   Kernel size is too small to remove noise effectively -> Option C

4. Quick Check:

   Increase kernel size for better noise removal [OK]

Hint: Use larger odd kernel sizes for stronger noise removal [OK]

Common Mistakes:

• Using even kernel sizes (invalid)
• Expecting median filter to blur edges heavily
• Thinking median filter can't remove noise

5. You want to denoise a photo with both Gaussian noise and preserve sharp edges of objects. Which sequence of filters should you apply for best results?

hard

A. Apply Gaussian blur first, then bilateral filter

B. Apply median filter first, then Gaussian blur

C. Apply bilateral filter first, then median filter

D. Apply median filter only

Solution

1. Step 1: Understand noise types and filter strengths

   Gaussian noise is best reduced by Gaussian blur; bilateral filter preserves edges while smoothing remaining noise.

2. Step 2: Determine filter order for best effect

   Applying Gaussian blur first reduces Gaussian noise; then bilateral filter smooths while preserving edges, improving overall quality.

3. Final Answer:

   Apply Gaussian blur first, then bilateral filter -> Option A

4. Quick Check:

   Gaussian blur + bilateral filter = noise reduction + edge preservation [OK]

Hint: Use Gaussian blur then bilateral filter for noise + edges [OK]

Common Mistakes:

• Applying median filter first which doesn't target Gaussian noise well
• Using only one filter for mixed noise
• Applying bilateral filter before Gaussian blur