The Research And Realization On Region-Based Fuzzy Feature To Content-Based Image Retrieval

With the rapid development of technologies of computer and multimedia and the wide-spread use of Internet, there is more and more of image information. How to rapidly find out images needed is starving for resolve. Content-based Image Retrieval (CBIR) is aimed at searching out images which satisfy user's needs, so is widely researched. Fuzzy set theory advances development of CBIR, extends CBIR from precision of calculating and caters for human's fuzzy thoughts. Emphasizing on this character, the thesis discusses image retrieval techniques which are syncretized Fuzzy set, analyzes and compares the characters of Fuzzy c-Means clustering and artificial immune network aiNet clustering, finally presents an improved region-based fuzzy feature to content-based image retrieval better in speed and hit rate than the original method as the realized system proves. The main works are as follows. ¨¨To summarize key techniques of modules of CBIR system such as feature extraction, indexing scheme, similarity measure, query specification, relevance feedback and performance evaluation. ¨¨To discuss some fuzzy techniques of feature extraction including crude fuzzy histograms, fuzzy paradigm-based histograms, combined fuzzy histograms and fuzzy geometrical features. ¨¨To analyze and compare the characters of Fuzzy c-Means clustering and artificial immune network aiNet clustering, conclude that Fuzzy c-Means clustering does not work well if the underlying classes or clusters deviate strongly from hyperspherical structures, aiNet is capable of reducing redundancy, describing data structure, including the shape of clusters by experiments on relatively gathering data set, scattering data set, annular data set and corkscrew data set. ¨¨To present and realize an improved region-based fuzzy feature to content-based image retrieval. On the basis of original method, three improvements are presented, they are data reduction but no influence to results by quantization and aggregation to Fuzzy c-Means, describing regions by shape based moments and adding feature of spatial information by neighboring table. Experiment results show that the new is better in speed and hit rate.