Large margin based discriminative sparse coding for image classification

Sparse has attractive performance in codebook generation. This paper proposes a Large Margin based discriminative method which utilize the discriminative information. Specifically, the codes corresponding to similar local features of images from different classes are pushed further apart, while the codes corresponding to similar local features of images from the same class are pulled closer together. We formulate the above two issues to generate sparse code of each feature point, and adopt multi-scale spatial max pooling approach to obtain feature representation of each image. Then, a linear SPM kernel is employed in the one-vs-all SVM training for image classification. Experiments on 15 Scenes and Caltech 256 dataset show that our method is effective.