Fusing Visual and Textual Retrieval Techniques to Effectively Search Large Collections of Wikipedia Images

Abstract

This paper presents an experimental study that examines the performance of various combination techniques for content-based image retrieval using a fusion of visual and textual search results. The evaluation is comprehensively benchmarked using more than 160,000 samples from INEX-MM2006 images dataset and the corresponding XML documents. For visual search, we have successfully combined Hough transform, Object’s color histogram, and Texture (H.O.T). For comparison purposes, we used the provided UvA features. Based on the evaluation, our submissions show that Uva+Text combination performs most effectively, but it is closely followed by our H.O.T- (visual only) feature. Moreover, H.O.T+Text performance is still better than UvA (visual) only. These findings show that the combination of effective text and visual search results can improve the overall performance of CBIR in Wikipedia collections which contain a heterogeneous (i.e. wide) range of genres and topics.