The Research Of Content-based Image Retrieval System

With the popularity of research on Database and Computer Visualization, Content-based image retrieval (CBIR) technology has been a hit recently. In terms of this topic, we did a research and discussion on image retrieval based on color, texture and shape respectively, made a conclusion of variant algorithms on feature extraction and match. In addition, we also made a comparison between different algorithms and especially did a number of experiments on color-based feature extraction.Further, this essay raises a new combined algorism based on content-based image retrieval. It mainly makes use of histogram to express image-structure information and therefore improved color-based retrieval efficiency greatly. It scans database image to have a template match, at the same time segments images dynamically into several blocks and does the retrieving according to Euclidean distance between blocks and its main histogram information. Finally compound both factors as the final similarity criteria, in which we can adjust the block number and weight values so as to get a better retrieving.Also, a new similarity measures for comparing two color histograms are described for the purpose of a better improvement: the dissimilitude distance DS* and the similarity distance S. Taking both into consideration, we have a combination of them as E. E isincorporated into the exponentiation part of the Gibbs distribution while DS* is compared to four similarity measures: L₁, L₂ (Euclidean distance), the similarity measure E in addition toGibbs distributions integrating E. In order to overcome the limitations of some previous information retrieval measures in evaluating the efficiency of an image retrieval process, three variants of a new effectiveness measure are proposed and experimented on an image collection for various similarity measures. Experimental results show that retrieval effectiveness is the highest for E+Gibbs and the lowest for the Euclidean distance. They also illustrate the superiority of our approach towards similarity analysis and retrieval effectiveness computation. For the sake of a further improvement, two new technologies are raised in the end of this paper, that is, the extended database technology and the maximum variance between clusters method of image segmentation based on PSO.