AN OBJECTIVE AND SUBJECTIVE EVALUATION OF EDGE-DETECTION METHODS IN IMAGES

Edge detection is an important first step in the analysis of images. In this paper, several components of edge detectors are identified. For one component, derivative computation, it is shown that the use of derivatives of Gaussians as point-spread functions is equivalent to a Gaussian-weighted polynomial approximation of the neighbourhood of each pixel. This leads to a possible refinement of Canny's operator. The components are compared objectively using an adapted version of Haralick's test. This gives greater understanding than comparing edge detectors that differ in respect of more than one component. Several edge detectors obtained by mixing components are also compared subjectively using a real image. It will emerge that the use of a Gaussian-weighted approximation is to be preferred to an equally weighted approximation for derivative computation. Furthermore, it will emerge that the second derivative in the gradient direction offers better localization than the Laplacian. Despite this, Haralick's operator performs worse than the Laplacian of a Gaussian operator with improved edge-strength computation.