The Research Of Multi-Feature Combination And Relevance Feedback In Image Retrieval

With the rapid development of Internet and Multimedia, image retrieval has become the research focus at home and abroad. How to set up the effective mechanism of image description and index has turned into a problem that must be solved at once. At present, image retrieval has widely been used in the fields of the RS, Medicine, Computer Vision, Military, and so on.Feature extraction is a critical stage of image retrieval.The performance of image retrieval is directly affected by the choice of feature extraction method.Through research and analysis for the traditional feature extraction methods in depth, a new image retrieval method that utilizes concentric circles to extract color features is proposed in the paper. The method splits image by concentric circles strategy, in which dynamic weight of image is acquired according to the characteristics of human visual physiology and mind,and obtains integrated weighted eigenvector of the image. Ten groups of images from database are tested.It selects 4 of each group to make up 40 inquiries.Experimental results show that the average accuracy comparing with global histogram and traditional method of block improves, and the method possesses geometric invariance.Since using a single feature to express image possesses content description unilateralism and shortcomings that lack distinction between information, an integrated multi-feature image retrieval method gets studying in depth, however, how to organize these features achieving the best results is always a studying difficulty in image retrieval.A feature weight assignment approach based on relevance feedback of utilizing genetic algorithm is proposed.The method embeds the user's query intent into feedback process by using a means of interaction and determines the fitness value of the image through the user's rank for the image,then using the genetic algorithm carries out genetic operations and adjusts the feature weight for the image that the user can accept.This method takes the user's subjective intent into account, which narrows the semantic gap between low-level feature and high-level semantics,obtaining better search results in a relatively short period of time.The experimental results show that the recall and precision of this proposed approach is better. This approach can get the best retrieval results for different integrating multiple features.