Abstract: With the development of Content-Based Image Retrieval (CBIR),the solution of "semantic gap" which exists between the low-features and the high-level semantic features has become the key problems of the semantic image retrieval．To avoid the general method maps an image into a class of semantic image,and reflect the natural scenery image contains a wealth of high-level semantic information and multi-homing type,this paper presented a process of repeating use of the optimal threshold for a roughly extraction of the largest target area with the color image．This color target area is comparatively singleness in the natural scenery images．On the basis of the divided regions,this paper extracted the color and shape features of each region,at last,the fuzzy nerve network was used to map low-features into the high-level semantic features,so it finally realized the image attribute information transfer effectively and obtained the high-level semantic automatically．Experimental results show that the method of image segmentation of natural color image can effectively extract the target object．It also has a certain degree of robustness to the noise images．The accurate retrieval rate approaches 90% and the recall rate also achieves 75% in some class image of the nature image database．The ex-perimental result shows the effectiveness and advancement of this method in the natural image retrieval.

Key words: Content-based image retrieval,Semantic image retrieval,Image segmentation,Optimal threshold,Robustness,Fuzzy nerve network