A new approach to texture segmentation is presented. This approach consists of two procedures based on classifications. In the first step, an image is divided into n by n blocks and features are calculated for each block. Then, blocks are classified by means of cluster analysis. In the cluster analysis $g^2tr(S_w)$ is utilized as criterion measure in order to find the number of cluster in the feature space. And K-means algorithm is uses for partitioning samples into a given number of clusters. As its result, the image is coarsely segmentated. In the second step, boundary blocks are determined and the distribution of clusters in the feature space is estimated. And then, only pixels in the boundary blocks are classified. At this time, k by k window centered on a pixel is used to estimate features. But the window region is not homogeneous. Hence the proposed method solved this problem using eight homogeneity masks. As the second step results image is finely segmented. Conventional methods need a priori knowledge before segmentation, but the proposed method needs not such a knowledge. And by using two step procedures we can find more accurate boundaries between texture regions and reduce processing time. We obtained excellent experiment results on artificial images.