Study Of Content-based Image Retrieval Based On Multi-feature Fusion Of Interest Points

CBIR (Content-based image retrieval) is currently a hot research topic. It is abetter solution to solve the problem that how to efficiently find the desired content inthe large image database. In CBIR, image features become the object of study, andimage retrieval are achieved by operating these features. Interest points are thesepoints those have rich content in an image, at present, most of the image retrieval isaccording to features of the image or features fusion of interest points. But in thesemethods produce redundant information easily, which makes the calculation morecomplex, also it makes the retrieval error by loss of part of image features.This thesis presents a content-based image retrieval method that based ondistance distribution histogram of interest points and multi-feature fusion, and verifiesits effectiveness. The main research results and innovations are as follows:(1) By analysis of the existing interest point extraction algorithm, thenintroduces the working principle, and the advantages and disadvantages of eachalgorithm. Finally, this thesis improves the Harris algorithm, and put it as thedetection algorithm. Then some representative points are selected by some rule,finally the region of interest is confirmed.(2) At present, many methods easily lost global features. This thesis uses thedistance distribution histogram of interest points as global features, and presents acontent-based image retrieval that based on distance distribution histogram of interestpoints and multi-feature fusion. It is able to describe the image information roundlyand exactly.(3) Because color histogram lack of spatial information, this thesis uses themethods of fan zoning division, to enhance the spatial information of the colorhistogram.(4) Considering the distance distribution histogram of interest points and imagefeatures such as color and texture, this thesis presents a method that based on distancedistribution histogram of interest points and multi-feature fusion. Experimentsuggested that this algorithm is effective.