Considering optimum number of segmentation areas

Abstract

Image segmentation is an important preprocessing step before object recognition and is a process of partitioning an image into different regions with homogeneity in some image characteristics. For the image segmentation problem, deciding the optimum number of homogeneous texture areas which compose an image is very difficult, and there seems to be only a few applications of genetic algorithms (GAs). In this paper we introduce a new segmentation method of an image composed of textures with randomness by using GAs. This method can decide the optimum number of segmentation areas in an image automatically. After an image is divided into many small rectangular windows with the same size, a two-dimensional autoregressive model, fractal dimension, mean and variance extracted from the data in each small window are used as a feature vector of the small window. The clustering of feature vectors is performed to some extent by using Kohonen's self-organizing neural networks and we can get the rough clustering result and a candidate number of clusters. This number also means a candidate number of segmentation areas in an image. Here we define the evaluation function which measures the clustering quality and execute the further clustering of feature vectors recursively in and around the obtained candidate number by using GAs. Finally, the optimum number of clusters is decided according to the value of the evaluation function and the optimum clustering result of feature vectors is mapped to the original image. In numerical examples the validity of this method is verified.