Image Retrieval By Joint Nonnegative Matrix Factorization With Sparse Regularization

Recently, social tagging has become more and more popular. Now social tagging is one of the defining characteristics of Web 2.0. Users are free to upload, share and annotate different media in various social tagging systems, for example, users can upload images in Flickr and label them, and can also share videos in YouTube. Meanwhile, users are able to retrieve any web-accessible items of interest. As these increasing social tagging systems provide us with much convenience, they also propose challenges and chances to our research.Since the tags annotated by users are often noisy, ambiguous, and subjective, how to deal with these problems in order to improve the exactness of the tags and to raise the precision of tag based information retrieval? How to utilize the rich multiple data sources so as to achieve the transfer learning between different sources? All these problems have become the hot research areas recently.Inspired by the recent advances of sparse coding and shared subspace learning, in this paper we propose an approach, namely Multi-source Boosting by Sparse Nonnegative Matrix Factorization (MtBSNMF). The proposed algorithm, analyses the multiple data sources via sparse nonnegative matrix factorization simultaneously, learning a shared structure and the corresponding individual structure for each data source. In this way, we achieve the transfer learning across different sources.In this thesis, we mainly implement two types of applications on the proposed algorithm. The first is on two different data sources linked by tags, MtBSNMF is applied to image retrieval. The second is on image sources connected by visual words, and we implement sample query and tag prediction with the assist of MtBSNMF. The experimental results demonstrate the effectiveness and feasibility of the proposed approach.