Convolution Kernel Laboratory
Hand-design 3×3 convolution kernels — blurs, edge detectors, sharpeners — and see exactly what a CNN's first layer discovers automatically.
- convolution
- CNN
- edge detection
- filters
Original image
After convolution
Kernel weights
What CNNs learn
A trained CNN's first layer typically rediscovers exactly these shapes: edge detectors resembling Sobel kernels, blob detectors resembling Gaussian blur, and corner detectors formed by combining edges in multiple directions. A kernel whose weights sum to zero (Laplacian, Sobel) is a high-pass filter — it is silent on flat regions and fires only where intensity changes. The difference is that a CNN discovers these weights from data instead of a human designing them by hand.