Create visual word pairs dynamically based on sparse codes of SIFT features for image categorization

Image categorization is an important issue in computer vision. The bag-of-visual words(BOV) model which ignores spatial restriction of local features has gained state-of-the-art performance in recent years. The basic BOV model uses k-means to form codebook. As sparse codes can better represent local features, we use sparse codes of SIFT features instead of k-means to form codebook. Additional, as local features in most categories have spatial dependence in real world, this paper proposed to use visual word pairs to represent the spatial information between words. To reduce the complexity both in time and storage, we add word pairs dynamically. Our experiments show that our algorithm can improve the categorization performance.