Edge detection is one of the foundational operations in computer vision, based on the mathematical concept of image derivatives. An edge in an image corresponds to a rapid change in pixel intensity, which in calculus terms is a region of high gradient magnitude. The gradient of a 2D image at any point is a vector pointing in the direction of greatest intensity increase, with magnitude proportional to the rate of change. Computing this gradient is the basis of all edge detection algorithms.

The Sobel operator approximates the image gradient using two 3x3 convolution kernels. The horizontal kernel Gx = [[-1,0,1],[-2,0,2],[-1,0,1]] detects vertical edges by computing the intensity difference between left and right neighbors, with extra weight on the immediate neighbors. The vertical kernel Gy = [[-1,-2,-1],[0,0,0],[1,2,1]] detects horizontal edges similarly. The gradient magnitude at each pixel is computed as sqrt(Gx^2 + Gy^2), and the gradient direction as arctan(Gy/Gx). The Sobel operator produces relatively thick edge lines because it responds to gradients over a 3-pixel neighborhood, and it is somewhat sensitive to noise because it does not include explicit noise reduction.

The Canny edge detector, developed by John Canny in 1986, is considered the gold standard for edge detection and addresses the limitations of simpler operators through a multi-stage pipeline. First, Gaussian smoothing removes noise that could cause false edge responses. Second, gradient computation (typically using Sobel operators) finds edge strength and direction at every pixel. Third, non-maximum suppression thins the edges to single-pixel width by examining each pixel along the gradient direction and keeping only pixels that are local maxima. This elegant step converts the thick gradient responses into crisp, thin edge lines. Fourth, hysteresis thresholding uses two threshold values to classify edges: pixels above the high threshold are definite edges, pixels below the low threshold are definite non-edges, and pixels between the thresholds are edges only if connected to definite edge pixels. This dual-threshold approach preserves weak but valid edges that are part of continuous contours while eliminating isolated noise responses.

In modern computer vision, edge detection serves as a preprocessing step for higher-level tasks. Object recognition systems use edges to identify shape boundaries, autonomous vehicles use lane edge detection for navigation, medical imaging uses edge detection to identify organ boundaries and abnormalities, and industrial inspection systems use edges to detect defects in manufactured parts. The Hough transform, which detects geometric shapes like lines and circles, operates on edge maps to find structural patterns in images.