Every classical edge detector starts the same way: slide a small kernel over the image and combine nearby pixel intensities into a single number at each position. The left grid is the source image, rendered as a height field where taller, brighter columns mean brighter pixels. A translucent kernel window sweeps across it exactly like a real convolution pass. The right grid is the resulting edge map — tall, accent-coloured columns mark pixels the chosen operator flags as a boundary.
Gx, Gy) to estimate the horizontal and vertical intensity gradient, then combines them into a gradient magnitude √(Gx²+Gy²). Large magnitude = steep brightness change = likely edge.The Canny detector, published by John F. Canny in 1986, is still the reference baseline computer-vision libraries compare new edge detectors against — its non-maximum suppression and hysteresis stages are what let it produce thin, continuous, low-noise contours instead of the thick, speckled response of a raw gradient magnitude.
A synthetic test image is rendered as a 3D pixel-height grid on the left; a sliding 3×3 kernel window scans it exactly as a real convolution pass would, and the resulting edge response — Sobel, Laplacian or full Canny — builds up as a second height grid on the right.
How gradient kernels turn local intensity differences into an edge signal, and why Canny's non-maximum suppression and hysteresis thresholding produce thinner, cleaner boundaries than a raw gradient-magnitude cutoff.
Pick a test pattern and an operator, then drag the threshold slider to see edges appear and disappear. Add noise to see which operator degrades first, and slow or speed up the kernel sweep to follow the convolution step by step.
The Sobel operator dates to 1968, the Laplacian-of-Gaussian approach to the 1980s, and Canny's multi-stage detector to 1986 — yet all three are still shipped in every major computer-vision library today.