Visual saliency coding for image categorization

Image categorization is a challenging task and image representation is a key problem in categorization. Many works have improved Bag-of-Words model to help image representation. However, they ignored the visual saliency information which is useful for image understanding. In this paper, we propose a novel visual saliency coding method based on Bag-of-Words model to represent images effectively. Our method combines visual saliency information with the local feature descriptors before they are clustered and quantized. Thus, after clustering, the quantized visual words represent the local image descriptors that are not only similar in their appearance, but also similar about their visual saliency. Furthermore, the visual words also contain some spatial segmentation and shape information which also help image understanding. We have evaluated our methods on Caltech 101 dataset, and demonstrated the effectiveness of our method.