SIFT Feature Based Similar Image Retrieval Technology Research

With the development of modern economy and human civilization, the growth rate of information is totally no longer as it has been like. And at the same time the simple needs to search the vast amounts of data for things we needed become complicated. As to the reasons that images are more intuitive and easy to understand, the demand of image retrieval is growing with the continuous improvement of the retrieval technique. Image retrieval has gone through the process of text-based image retrieval technology, content-based image retrieval technology and the combination of these two technologies. This paper focuses on the analysis, improvement of the SIFT feature and its organization and retrieval. Methods to improve the image retrieval results are proposed. And the final similar image retrieval system based on SIFT feature is developed. The main contributions in this paper are as follows:(1) Firstly, the basic theory of multi-scale technology used in image presentation is introduced. The algorithm and extracting method of SIFT is elaborated in detail. And SIFT’s features include translation invariant, scale invariant, rotation invariant and affine invariant are analyzed. Experiments on extracting SIFT feature is carried.(2) After describing the traditional nearest neighbor retrieval method, the principle of KD tree is elaborated and it is detailed analyzed that how the KD tree is used to SIFT Algorithm. The experiment is designed to show the matching results.(3) It is mainly described how the texture-based vocabulary tree structure is used to create index for the image features to achieve the target of improving the efficiency of the search. As to solve the problem of the non-stop iteration in the visual vocabulary tree creating, the final solution of using sliding window to change the cluster’s initial center.(4) Experimental verification of the SIFT’s basic characteristics are carried out. The image retrieval experiment based on the original SIFT features and experiment using visual vocabulary tree are both designed and the experimental data is given which is analyzed to make a conclusion. And the matching algorithm is to be improved so that the searching result is better. A proper matching method which is appropriate for SIFT is proposed.The experimental results show that the proposing algorithm and the designed can reach to a satisfactory result.