Canny Edge Detection

Non-maximum suppression thins the gradient magnitude map by suppressing pixels that are not local maxima along the gradient direction. This step is crucial because Gaussian derivatives produce thick gradient responses near edges. By keeping only the locally strongest responses (perpendicular to the edge direction), NMS produces single-pixel-wide edge contours. Without NMS, edge maps would be thick and diffuse.

Canny uses two thresholds: a high threshold (e.g., 0.3 × max gradient) and a low threshold (e.g., 0.1 × max gradient). Pixels above the high threshold are immediately classified as edges. Pixels below the low threshold are rejected. Pixels between the thresholds are kept only if they are connected to pixels above the high threshold. This hysteresis mechanism suppresses noise while preserving thin, weakly-defined edges that are part of true contours.

Gaussian blur reduces noise before computing gradients. Since gradient operators are sensitive to high-frequency noise, smoothing first ensures that detected edges correspond to true intensity changes rather than noise. The smoothing scale (Gaussian sigma) is a tuning parameter: larger sigma removes more noise but also blurs edges; smaller sigma preserves edge sharpness but may be sensitive to noise.

Threshold selection depends on image contrast and noise level. A common heuristic is to use high threshold = 0.2–0.3 × max gradient and low threshold = 0.05–0.1 × max gradient. Alternatively, Otsu's method or histogram analysis can suggest thresholds. For interactive applications, thresholds can be exposed as sliders. For batch processing, automated threshold selection methods (e.g., based on image statistics) are preferable.