Texture segmentation method considering optimum number of segmentation areas by using neural networks

Abstract

The automatic decision of the optimum number of homogeneous areas which constitute an image is very difficult and important task in the image segmentation problem. We propose a new segmentation method of an image composed of some kinds of textures with randomness by using both unsupervised and supervised neural networks. After a texture image is divided into many small windows with the same size, the feature vectors in those windows are extracted by using two-dimensional autoregressive model and fractal dimension. The clustering of feature vectors is performed to some extent by applying Kohonen's self-organizing neural networks which are unsupervised neural networks and the maximum candidate number of the homogeneous areas in the image is obtained. Here we define the evaluation function which measures the segmentation quality and execute the further clustering of the feature vectors recursively to be maximum candidate number by applying decision-based neural networks which are supervised neural networks. Then the optimum number of clusters is decided according to the value of the evaluation function and the result of clustering feature vectors is mapped to the original image. In numerical examples the validity of this method is verified.