Textured image segmentation using autoregressive model and artificial neural network

In this paper, we use a two-dimensional (2-D) AR model for texture description. The coefficients of the AR model as the parameters can thus be used to identify textured images. These processes are ideally suited to implementation by neural networks which are well known for their parallel execution and adaptive learning abilities. The proposed network consists of three subnets, namely the input subnet (ISN), the analysis subnet (ASN) and the classification subnet (CSN), respectively. The neural network obtains parameters for a 2-D AR model on a given texture through an adaptive learning procedure, and segments an input image into regions with the learned textures. Furthermore, a textured image which has a certain degree of deformation with respect to one of the possible texture classes can be correctly classified by the network. The network is easy to extend because of its modular structure in which all channels work independently. A region growing technique for texture segmentation is implemented by comparing local region properties. 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 performance of the proposed network has been examined on real textured images. In the classification phase, images proceed through the network without the preprocessing and feature extraction required by many other techniques. Hence, overall computation time has been considerably reduced.