Diversified textual features based image retrieval

Abstract

How to target the users’ demands faster and more accurately has become a hot issue in the domain of image retrieval. Most of the existing techniques focus on retrieving the most relevant images to the query, which will reduce the search efficiency and make the retrieval process boring for users. For these reasons, diversity-induced image retrieval is proposed to guarantee that the retrieval results are not only relevant to the query, but also cover the aspects of the query as many as possible. Most of the traditional diversity-induced methods seldom use the multi-modal information to compensate each other. Moreover, the traditional textual features come from the complete textual information which can not highlight the unique property of an image. In this paper, we propose a new retrieval framework that combines visual and textual information in an effective way. Firstly, Latent Dirichlet Allocation (LDA) is applied on visual features to construct topics. Then textual information is used to improve the coherence of the topics. To extract diversified textual features, we propose a new method focusing on the unique words. For topic improvement, we design a new strategy that removes outliers from the constructed graph. Finally, the representative images are selected as retrieval results. The experiments on the database of MediaEval 2013 prove the superiority of our method. Besides, we also illustrate the results of diversified textual feature extraction and similarity graph clustering, which indicates the effectiveness of the proposed method.