String distance for automatic image classification

Abstract

The Bag-of-visual Words model has recently become the most popular representation to depict image content. It has proven to be quite effective for many multimedia and vision applications, especially for object recognition and scene classification or automatic image annotation. This model however ignores the spatial layout of features within images, which is yet discriminative for category classification. In this paper, we present a novel approach based on string matching to take into account geometric correspondences between images and facilitate category recognition. First, we propose to represent images as strings of histogram second, we introduce a new string distance in the context of image comparison. This distance automatically identifies local alignments between sub image regions and allows merging groups of similar sub-regions. Experiments on several dataset such as Scene-15, Caltech-101 and Pascal 2007 show that the proposed approach outperforms the classical BOW method and is competitive with state-of-the art techniques for image classification.