Edge Detection Online

How the Edge Detector works

Edge detection online: getting a good result

Edge detection finds the places in a picture where brightness changes sharply: the border between a face and the background, the lines of a building, the rim of a cup. This tool uses the Canny algorithm, which is still the standard approach forty years after it was published. It first blurs the image slightly, then measures the brightness gradient at every pixel, thins the result down to one-pixel-wide ridges, and finally decides which ridges are real edges using two thresholds.

Low and high threshold work together in a step called hysteresis. Any pixel whose gradient is above the high threshold is definitely an edge. Pixels between the two thresholds are kept only if they connect to a definite edge, which is how the algorithm follows a faint contour that is attached to a strong one without also picking up isolated speckle. A good starting ratio is high at two to three times low. Lower both to pull in more detail; raise them to keep only the strongest outlines.

Blur before detecting sets the size of the Gaussian blur applied first. A small blur removes sensor noise and fine grain that would otherwise show up as thousands of tiny edges; a bigger blur also smooths away fabric weave, foliage and other texture so only the large shapes remain. Gradient chooses how the strength of an edge is measured: the fast L1 norm adds the horizontal and vertical changes, while the precise L2 norm takes the true magnitude, which gives slightly cleaner edges on diagonals at a small cost in speed. Output flips between white edges on black, the conventional way to view a detector's output, and black edges on white, which looks more like a line drawing and prints better.

Very large photos are scaled down to 2400 pixels on the long side before processing so the result appears quickly. The edge pixel count in the stats is a useful gauge: a few percent of the frame usually means a clean outline, while fifteen percent or more means texture is being picked up and the thresholds or blur should go up. The library that does the work is loaded on first use and the image never leaves your device.

Related tools