A fuzzy logic and neural network approach to boundary detection for noisy imagery

Abstract In this paper, an edge relaxation method utilizing fuzzy logic and neural networks is presented. The goal is to extract physically meaningful edge segments. The candidates for edge segments are first estimated using a local derivative operator with a window of small size. To enhance the performance of edge detection, in addition to the output of a derivative operator, the spatial relationships among the neighboring edge segments around an edge segment of interest are used as additional source of information. Fuzzy rules, each of which is associated with a neighborhood pattern defined by the spatial relationships among the neighboring edge segments, are used as a computational framework of collecting the evidence for the existence of an edge segment. These fuzzy rules are trained by a specially structured neural network which performs fuzzy reasoning operation. Through a series of experiments, the proposed edge relaxation scheme is found to exhibit reasonable performance of eliminating false edges, strengthening weak real edges and bridging small gaps between edge segments on object boundary.