Multiresolution edge detection and classification: noise characterization

In this work we present the method we have developed in order to achieve edge detection and classification in gray level images for five different contour types: step, ramp, stair, pulse and noise. The edge detection method is based in a multiresolution analysis using Mallat and Zhong's wavelet, which is compared with the gaussian-based one. The edge classification has been made analyzing the first six wavelet coefficient evolution across scales at the edge position. We have implemented a decision algorithm based on a simple second order polynomial fitting, and a subsequent numerical analysis of the obtained polynomial in order to discriminate between the five contour types. At the end of the process we obtain the edge position and the belonging class for each contour pixel. The main advance of this work is the characterization of the edge class 'noise'. In this class are included the gaussian noise and the irrelevant low-level contours that appear in non-thresholded image. We can filter this new edge simply by eliminating the edges classified in this group without affecting low-level, but important, contours.