Why smooth an image before detecting edges?

Differentiation amplifies noise, so smoothing first prevents the detector from flagging noise as edges; the smoothing scale sets which size of detail is treated as an edge.

Why does the Canny detector have multiple stages?

Each stage handles a separate goal: smoothing controls noise, gradient computation finds candidates, non-maximum suppression thins them to single-pixel edges, and hysteresis thresholding keeps weak edges only when connected to strong ones.

Edge and Contour Detection

Edge and contour detection locate the boundaries in an image where intensity changes sharply, which often correspond to the outlines of objects and surface discontinuities.

Definition

An edge is a location of significant local intensity change, and edge detection is the identification of such locations, typically by analyzing the image gradient or the zero crossings of a smoothed second derivative.

Scope

This topic covers gradient-based edge operators, the role of smoothing before differentiation, the multi-stage Canny detector with non-maximum suppression and hysteresis thresholding, zero-crossing approaches, and the linking of edge points into continuous contours.

Core questions

Where does intensity change abruptly in an image?
How is differentiation made robust to noise?
How are thick gradient responses thinned to one-pixel edges?
How are isolated edge points joined into meaningful contours?

Key concepts

Image gradient
Gradient operators
Non-maximum suppression
Hysteresis thresholding
Laplacian of Gaussian and zero crossings
Contour linking

Key theories

Canny edge detection: Derived from criteria of good detection, good localization, and a single response per edge, the Canny detector smooths the image, computes gradients, suppresses non-maximal responses, and links edges by hysteresis thresholding, remaining a standard baseline.
Marr-Hildreth zero crossings: Edges are located at the zero crossings of the Laplacian of a Gaussian-smoothed image, tying edge detection to a computational theory of early vision and to multi-scale analysis.

Clinical relevance

Edge and contour detection feed segmentation, shape analysis, and object recognition, and are used in medical imaging, industrial inspection, and feature extraction pipelines across computer vision.

History

Marr and Hildreth's 1980 theory linked edges to zero crossings of a smoothed Laplacian, and Canny's 1986 optimal-detector formulation became the most widely used edge detector, later complemented by learned boundary detectors.

Key figures

John Canny
David Marr
Ellen Hildreth

Seminal works

canny1986
marr1980

Frequently asked questions

Why smooth an image before detecting edges?: Differentiation amplifies noise, so smoothing first prevents the detector from flagging noise as edges; the smoothing scale sets which size of detail is treated as an edge.
Why does the Canny detector have multiple stages?: Each stage handles a separate goal: smoothing controls noise, gradient computation finds candidates, non-maximum suppression thins them to single-pixel edges, and hysteresis thresholding keeps weak edges only when connected to strong ones.