Research Of Neighborhood Reversibility In Media Retrieval

With the rapid development of information technology, vast and varied digital media data is produced and published on the internet. It is important for digital media retrieval to meet users’ information need by quickly and precisely finding out the required data from the huge dataset.In practical study of media retrieval, neighborhood non-reversibility relationship can generally be observed in the media retrieval process. In particular, if the media object i is in the retrieval result list of the media object j and the media object j is also in the retrieval result list of the media object i, we can say the two objects meet the property of neighborhood reversibility. According to the observations from lots of media retrieval processes, if two media objects meet neighborhood reversibility relationship, the two media objects are most probably relevant. This property can be used to improve the media retrieval quality. However, less effort has been paid on the study of neighborhood reversibility relationship in the multimedia community.In this thesis, we focus mainly on the reversibility of neighborhood relationships and the sensibility of neighborhood reversibility in media retrieval. The main contributions of this thesis include:(1) For the neighborhood reversibility in image retrieval, we propose two effective and efficient image re-ranking schemes, i.e. hard re-ranking and soft re-ranking schemes. By recording a K-NN distance for each database image in offline stage, we can void high computational cost and quickly reconstruct the reversibility relationship in online stage. Experimental results show that the proposed re-ranking schemes can remarkably improve the image retrieval quality in terms of effectiveness and efficiency.(2) To address the effect of neighborhood size, we propose an adaptive selection scheme of neighborhood size to handle the performance difference of neighborhood reversibility learning schemes on varied image datasets. Experimental results show the proposed scheme can greatly improve the robustness of reversibility verifying process.(3) For cross-media retrieval, the neighborhood reversibility is introduced in the multimodal feature space creation process to improve the precision of semantic correlations. In particular, we first analyze a kind of cross-media retrieval framework based on neighborhood relationship. And then, we apply the neighborhood reversibility to the calculation of similarity matrix in this framework. Experimental results show that the proposed scheme can remarkably improve the search quality of the cross-media retrieval framework.