Research On Graph-based Reranking For Image Search

With the development of the search engine, users have been accustomed to search all kinds of information, including text, images and videos on the Internet with the help of various types of search engines. Traditional text-based image retrieval system, most of which rely on key words search. It is the text information contains a lot of noise, this will lead to the search results are not ideal. Existing mainstream Internet search engines, such as Google, Bing, Baidu etc.. Most of these search engines use the text information around the image to achieve the image search and rank. These search engines lack the intrinsic link between the image and the contents of the image itself, which leads to the unsatisfactory results of the text-based image search results. Now the researchers focus on how to improve the quality of image search results.Image reranking is based on the initial search results. We can extract the inner link of the image and the contents of the image itself. Image reranking is to rank the initial results, then the final reranking results will be the user needs. At present, according to the different framework, the image reranking methods can be classified into four categories:based on linear combination, based on clustering, based on classification and based on graph. The existing graph-based image reranking methods generally used pseudo relevance methods for initial ranking results.. These methods consider the front images to have high scores. But it is not a truth. As the initial ranking results are based on the text retrieval, the accuracy of the ranking results is low. It may cause the back of the image to be the user’s need. The contribution of this work is summarized as follows:1. To improve the effectiveness of reranking algorithm for image retrieval, this paper presents a multimodal graph-based reranking through random walk algorithm. Firstly, different from the existing reranking algorithms that initialize the relevance score list of the retrieved images according to the returned image sequence, the proposed method integrates multimodal to catch more information and employs a multimodal random walk algorithm to initialize the relevance score list of the retrieved images. Then the proposed method optimizes the objective function using a multimodal graph-based reranking algorithm, in which an iteration procedure is used to update the parameters and relevance score list. Finally, the retrieved images are reordered according to the relevance score list. Experimental results demonstrate that the proposed reranking algorithm performs better than some other state-of-art algorithms.2. Multimodal graph-based reranking through similarity integral algorithm is put forward to solve the problem of imperfect result of image research build on text. First, this algorithm will generate six image modals after receiving dataset and apply mode concept into the similarity integral algorithm. This solution can improve the accuracy of research during the process in the next Iteration of algorithm. Then take the image reranking score returned by the multimodal similarity integral algorithm as the input of the multimodal graph-based learning. After receiving the reranking result, the objective function, of which the reranking algorithm is generated by using normalized graph Laplacian regularizer, will be minimized. Finally, the six image modals will iterate continuously before it done. According to the tests, under certain conditions, the reranking algorithm referred in the text is better than the multimodal graph-based learning and multimodal similarity integral algorithm.