Improving the Discriminability of Dictionary by Gist Information Detection

Image representations using code words from a visual dictionary are widely applied in object detection and categorization. Traditionally, there are two types of methods to construct a dictionary: k-means and optimization-based method. The former cannot achieve a good discriminability because it extracts too many background features. The latter needs to cooperate with coding methods and brings about high computational complexity. In this paper, we present an effective method based on Gist information detection to obtain a more discriminative dictionary with low computational cost. First, we partition the image into increasingly fine sub-regions, and calculate the Gist information of each region. Then extract more features from sub-regions with richer information and fewer features from ones with less information. Finally construct a dictionary using the non-uniform sampling features. Experiments on Caltech101 show that our method can achieve a better performance than traditionally k-means and the optimization-based method. Hence our dictionary has a better discrimination.