To determine the number of clusters in the traditional FLICM algorithm, this paper designs splitting-merging operations based on the characteristics of the FLICM algorithm which uses the mean of pixels to describe the cluster. Then, an acceptance probability is designed to promote the convergence of the algorithm, which can prevent it from being trapped into local extremes and realize the self-adaptiveness of the thresholds, simultaneously. Qualitative and quantitative analysis on the experimental results from a series of synthetic and gray texture image segmentation demonstrates the adaptability and accuracy of the proposed approach compared with the traditional iterative self organizing data analysis techniques algorithm(ISODATA).