Enhanced Canny Edge Detection Using Curvature Consistency

Edges are often considered as primary image artifacts for extraction by low-level processing techniques, and the starting point for many computer vision techniques. As a result, reliable edge detection has long been a research goal. This paper describes initial investigations into recovering reliable edges using curvature models. Essentially, we modify Canny’s edge detector, using a curvature consistency process to adjust the gradient direction estimates prior to finding the zero crossings in those directions. 1 Background Edge detection has long been an active research area in the image processing and computer vision community, and the literature is extensive, with comprehensive surveys available (e.g. [11]). The Canny edge detector [1] is arguably the de facto standard for edge detectors. Its underlying process can be viewed as finding the zero crossings of the second derivative calculated in the gradient direction. A further process, hysteresis thresholding, is sometimes used to dismiss weak edges whilst maintaining edge continuity. Before non-maximal suppression and hysteresis thresholding, the basic Canny edge detector can be defined as the solution of: @I @g = g ( @I @x @I @x@y @I @y@x @I @y )