Creating Descriptive Visual Word Tree for Tag Ranking of Social Image

Abstract

Visual words image representation has been very popular in tag ranking of social image. However, the descriptive power of detected features is not evaluated in traditional visual words model, resulting in many non-descriptive local features are clustered as visual words. Moreover, the time consumption will be increased during quantification phase due to compare with all visual words. To overcome these problems, the descriptive visual word tree is created for tag ranking of social image in this paper. Firstly, the social image saliency region is detected with visual attention model. Then, the descriptive visual word tree can be created by clustering the descriptive Scale Invariant Feature Transform (SIFT) features which consist of refined SIFT features and corner SIFT features extracted from saliency region. Finally, the tags of weighted neighbor images, which are found by descriptive visual word tree, are applied to rank the tags. Experimental results indicate that the proposed descriptive visual word tree can effectively improve the accuracy of tag ranking of social image as well as reduce the time of tag ranking.