Scale detection based on statistical characteristics of edges in the scale space

A study of the behavior of edges in the scale space is presented in this paper. Three statistical parameters, intensity average, root-mean-square, and variance of pixel intensity in edge images are used as the basis for evaluating visual quality of edge images. A parametric family of edge images containing edge information at several scales is used to create the edge statistic curves. It is found that the number of local minima in edge statistic curves is related to the number of scales in an image. This leads to an automated method for detecting scales in an image and a criteria for selecting optimal parameters for a multiscale edge operator.