Discriminative Image Representation for Classification

Abstract

The Bag-of-visual Words (BoW) image representation is a classical method applied for various problems in the fields of multimedia and computer vision. During the process of BoW image representation, one of the core problems is to generate discriminative and descriptive visual words. In this paper, in order to represent the image completely, we propose a visual word filtering algorithm, which filters the lower discriminative and descriptive visual words. Based on the traditional method of generating visual words, the filtering algorithm includes two steps: 1) calculate the probability distribution of the various visual words, and then, delete the words with gentle probability distribution; 2) delete the visual words with less instances. In this way, the generated visual features become more discriminative and descriptive, furthermore, multiple cues fusion, such as shape, color, texture, is also taken into account, we compare our approach with traditional Bag-of-visual Words method applied for image classification on three benchmark datasets, and the performances of the classification all get improvements to some extent.