Abstract: A major problem in content based image retrieval （CBIR） is how to obtain the robust and efficient segmentation of images. To satisfy the need of CBIR, a new algorithm for segmentation color images by fusing color and texture and by cluster analysis is presented in this paper. It fuses color and texture features by linear weights, and the weights are auto determined according to the distribution of the image spectrum power. It uses a new proposed clustering algorithm that is based on the coding cost and the concept of self annealing. The clustering algorithm can determine the number of clusters automatically,and also is insensitive to the choice of initial clusters. The new algorithm is applied to segment a number of natural images, and its efficiency is confirmed by the experiments.