Boundary detection based on neural networks model

Abstract

A new model of feedforward neural networks is proposed for solving the problem of robust boundary detection. Structurally, it is based on a circular mask which is characterized as a symmetrical neural network. By analyzing the weighted intermediate pattern, a dominant pattern is found to appear repeatedly for similar boundary orientation. Hence, a piecewise linearized edge could be detected by this approximation. Experimental results show that this new architecture can be applied to experience scaling effect by changing the mask size.<>