Web image retrieval on ImagEVAL

Abstract

We present in this article an efficient visuo-textual Web Image Retrieval system (WIR), which is the second best system according to the official European ImagEVAL 2006 campaign evaluation. It uses very simple tfidf textual analysis, and subband entropy profile visual features. Our mean fusion model represents a simple but nearly state of the art WIR. We depict analyses of the fusion behavior of each query. We then demonstrate that "visualness" of images, and "textualness" of web page, relative to the discriminant power of each features, are concept dependant, and that fusion model could take advantage of their possible complementarity. We finally discuss on their automatic estimations that may enhance WIR.