TRAINING A NEURAL NETWORK FOR MOMENT BASED IMAGE EDGE DETECTION

Edge detection is a crucial step to computer vision. Currently, there is not a single edge detector that has both efficiency and reliability. Traditional differential filter\|based algorithms have the advantage of theoretical strictness , but require excessive post\|processing. This paper introduces a neural network edge detector that takes advantage of moments features. It functions as a neural pattern classifier that directly estimates the posterior probability from the training data set. Two subsystems can be distinguished and different kinds of learning rules are used. For the end\|user, it works as a black box that directly transforms raw images into the edge maps so no complicated postprocessing is required. Tests on both simulated and real images showed the proposed neural network edge detector is superior to traditional operators.