A NOVEL EDGE DETECTION TECHNIQUE FOR 2-D IMAGES USING FUZZY-NEURAL HYBRID SYSTEM

Abstract

Edges are important features widely used in image processing, pattern recognition, and artificial vision systems. For numerous image processing tasks, these edges contain very useful structural information about objects boundaries and. image segments in a highly compact form. In this dissertation, a new edge detection technique has been proposed using fuzzy-neural hybrid system. Here edge detection process is considered as classification of input pattern space into edge and non-edge regions. The proposed system consists of three stages, namely adaptive fuzzification, pattern classification using Radial Basis Function network (RBF) and thinning process. The proposed RBF network has only six hidden neurons and six output neurons. To improve the performance on real images, a threshold factor, which is function of gradual variations in gray-level, contrast and non-uniform lighting effects, has been incorporated. This threshold factor is estimated using difference histogram. The proposed algorithm has less computational complexity compared to other neural network based edge detectors. , The proposed algorithm is tested for both synthetic and real images and comparisons are made with relatively new techniques. Qualitative and quantitative comparisons are presented by considering the parameters like edge orientation, edge localization, missing edges, false edges, noise effect and computational complexity. The proposed technique is implemented in MATLAB 6.5, has generated around 250 edge maps of synthetic and real images for analysis, In MATLAB 6.5, to process 256 x 256 pixel size images, the proposed technique has taken 14 minutes where as FNN with 1-1op1ield network [24] for edge enhancement has taken 145 minutes. Thus, overall comparisons, with respect to the above parameters clearly establish the superiority of the proposed technique over other techniques.