Visual homograph: A novel basic visual element

The bag-of-visual-words (BoW) model has been broadly applied to represent images in a variety of image retrieval applications. Ideal visual word vocabulary should obtain two important properties, completeness and distinctiveness (i.e., containing all the similar image patches and eliminating different ones accurately). However, the traditional definition of basic visual elements tend to lack in either generality or particularity, which will lead to the incoherence problem and further ruin the image retrieval performance. To address this issue, in this paper, we propose a novel basic visual element called Visual Homograph - groups of local patches which have indivisible descriptors but differ in origins. To generate Visual Homograph based vocabulary, we also propose a novel similarity metric named Integrated Perceptual Similarity which systematically combines center-based and boundary-based measures. Experimental results on standard datasets demonstrate the superiority of Visual Homograph based method over state-of-the-art visual vocabulary generation approaches, in terms of alleviating incoherence problem, and enhancing retrieval performance in practical applications.