Computer Vision: Edge Detection Kernels

Test Image

Convolution Kernel

View

Edge threshold 35%

Scan speed 6 / s

Live Readouts

Scan window

(0, 0)

|∇I| at window

0.00

Edge pixels

Peak |∇I|

0.00

How it works

An image is a grid of intensities I(x,y). A convolution kernel K is a small matrix slid over every pixel; at each position the output is the weighted sum of the neighbourhood:

O(x,y) = ΣᵢΣⱼ K(i,j) · I(x+i, y+j)

The Sobel operator uses two 3×3 kernels to estimate the horizontal and vertical intensity gradient:

Gx = [-1 0 1; -2 0 2; -1 0 1]
Gy = [-1 -2 -1; 0 0 0; 1 2 1]
|∇I| = √(Gx² + Gy²)

Wherever brightness changes sharply — an object boundary, a shadow edge, a facial feature — |∇I| spikes. Thresholding that gradient turns it into a binary edge map. The Laplacian kernel (a single 3×3 mask) instead responds to curvature in intensity, highlighting fine detail and noise more aggressively.

Test image buttons — swap the synthetic scene the kernel scans.
Kernel buttons — choose which convolution mask computes the relief height and colour.
View toggle — switch between the raw grayscale image and the computed edge map.
Threshold — pixels whose gradient exceeds it are marked as edges (bright) versus suppressed (dim).
Scan window — an animated 3×3 neighbourhood sweeps the grid exactly as a real convolution pass would, one output pixel at a time.

Real-world relevance: Sobel/Laplacian edge maps were the standard hand-engineered feature extractor before deep learning, and they still run as a preprocessing stage in lightweight embedded vision — lane-boundary detection in driver-assist systems, document-edge alignment in scanning apps, and facial-landmark localisation pipelines. A trained CNN effectively learns kernels like these (and many more complex ones) automatically from data.

Low value High value / edge

⚠ Couldn't load the 3D engineThree.js failed to load from the CDN. Check your connection and reload.