Research On Rerank Algorithms In Image Retrieval

With the rapid development of multimedia and network technology, the amount and the type of digital images are both very huge nowadays, and are increasing every day, how to rapidly and accurately find the image information that users need in these huge data, and present the retrieval results to the user with a certain rank way, are hot problems in research.Content-based image retrieval(CBIR) technology is a currently widely used way of retrieval.However, in the component part of CBIR, the rank algorithm directly determines the retrieval results which the user can see, so the rank algorithm plays a vital role in the performance of CBIR technology. With the improvement of retrieval performance of search engine, the current rank algorithms are mainly based on image similarity between different images, namely, the image with the most similar to query image will be in the front of the list of results. But this way of ranking doesn’t consider the diversity of image retrieval results, so the rank algorithms based on diversity are proposed. Visualrank is a very famous method in rank filed. However, considering the diversity of images results, people proposed a multiclass visualrank algorithm, namely the Multiclass Visualrank algorithm. Based on Visualrank algorithm and Multiclass Visualrank algorithm, this paper made the following work:1. This paper proposes a new image similarity rerank algorithm based on VisualRank algorithm.While extracting the SIFT features of an image, it needs to extract a large number of key points, which will cause a large amount of calculation and high computational complexity. In this paper, by extracting Space Pyramid Matching( SPM) features to improve the SIFT features of images, then, according to the idea of Pagerank algorithm to rank the images, and the results will be presented to the users at last. Experimentalresults show that the proposed algorithm can get better rank results, and the computational complexity is lower, which is more similar with the needs of users.2. Based on the Multiclass Visualrank algorithm, an improved image diversity rerank algorithm is proposed. Multiclass Visualrank algorithm based on the basis of Visualrank algorithm, firstly extracting the SIFT features of images, then using Ncuts clustering algorithm to cluster in image set, in each category,rerank the images within a class, the final results are presented to the user in the form of classification. But NCuts algorithm is based on each pixel of images for clustering, so the amount of calculation is large,which is not conducive to deal with real-time images. Based on these problems, this paper uses the combination of Mean Shift and Ncuts clustering algorithm for clustering, then according to the idea of Pagerank algorithm to rerank the images, the results will be presented to the users finally. In this paper, the algorithm is compared with other algorithms, experimental results show that the algorithm can achieve ideal results both in time complexity and relevance.