Texture identification using artificial neural networks and 2D-Autoregressive Model

Abstract

As an important aspect of image analysis, texture identification has been pursued by many researchers. Among techniques developed, the approach of modeling texture images through a 2-D Autoregressive (AR) Model is of special interest. The major problem with the modeling methods is the estimation of parameters due to the intensive amount of computation involved. From a parallel computing perspective, parameter estimation can be implemented by learning procedure of a neural network, and texture classification can be mapped into a neural computation. A multilayer network is proposed which consists of three subnets, namely the input subnet (ISN), the analysis subnet (ASN) and the classification subnet (CSN). The network obtains the classification capability through an adaptive learning proceedure. In the processing phase, images proceed through the network without the preprocessing and feature extraction required by many other techniques. An integrated texture segmentation technique is proposed to segment textured images. The technique is implemented by comparing local region properties, which are represented by a 2-D AR model, in a hierarchical manner. It is able to grow all regions in a textured image simultaneously starting from initially decided internal regions until smooth boundaries are formed between all adjacent regions. The performances of the classification and segmentation techniques are shown by experiments on natural textured images.