Integrated patch model: A generative model for image categorization based on feature selection

Abstract

Image categorization could be treated as an effective solution to enable keyword-based image retrieval. In this paper, we propose a novel image categorization approach by learning semantic concepts of image categories. In order to choose representative features and meanwhile reduce noisy features, a three-step feature selection strategy is proposed. First, salient patches are detected. Then all the detected salient patches are clustered and the visual keyword vocabulary is constructed. Finally, the region of dominance and the salient entropy measure are calculated to reduce the similar and non-common noises of salient patches. Based on the selected visual keywords, the Integrated Patch (IP) model is proposed to describe and categorize images. As a generative model, the IP model represents the appearance of the combination of the visual keywords, considering the diversity of the object or the scene. The parameters are estimated by the EM algorithm. The experimental results on the Corel image dataset demonstrate that the proposed feature selection and the image description model are effective in image categorization.