Content-based Concept Modeling And Image Search Reranking

With the rapid development of Internet, multimedia technologies, such as image, video and audio, are widely used on the Internet during recent decades. Digital images resources are also increasing rapidly. What’s more, the Social Network Service (SNS), such as Facebook, Twitter, Instagram, Whatsapp, Renren, Weibo, Wechat, makes the multimedia service popular recent years. Users play two different roles on the Internet, that is, the multimedia resources providers and the multimedia resources consumers.The users acquire multimedia resources via Internet. And, retrieval technologies make the processing more efficient. For the research of image retrieval, how to index the users’expected images fast and accurately becomes a hot topic. But, there are some junk images existing in the initial search ranking list. Based on this observation, we re-rank the initial ranking list so that the re-ranked list shows more query-relative images on the top of the list. The work proposes two different reranking algorithms. Applying these reranking methods on initial ranking list, we get more query-relative images on the top of the list.The contributions of this paper can be summarized as follows.First of all, we give the definition of image search reranking and list some basic concepts. What’s more, we make a summarization of state-of-art reranking methods. Considering the importance of image visual features for image search reranking, we also explore some visual features at this part. Secondly, we propose an adaptive method for image search reranking. We extract different image features, such as color, texture and local SIFT for reranking. The method can automatically choose visual features to re-rank the initial image search result. Also, this method can achieve better performance than image search result. Furthermore, we propose an adaptive query prototype modeling method for image search reranking. This method uses top N images of the initial search result for concept modeling. Then, it re-ranks the initial image search result using the trained concept model. Also, we use the local SIFT feature for image reranking. During the phase of training, an incremental learning method is introduced. Finally, we conclude our work and present further discussion on future work.