Computer Science · Computer Vision and Pattern Recognition
Invariant hierarchical sparse coding for object recognition via bags of atoms
Xiaoxia Sun, Nasser M. Nasrabadi, Trac D. Tran
See on the mapAbstract
In this paper, we introduce a novel local feature-based hierarchical framework to produce invariant sparse codes for object recognition. In order to enforce the invariant property for each sample patch (local feature descriptor) in the image, its sparse code is recovered with a dedicated dictionary whose atoms are adaptively chosen from several bags of candidate atoms. The single-layer invariant sparse coding model is naturally extended to a multi-layer hierarchical architecture to further improve the invariance of the sparse codes. Experimental results on the popular image datasets, including MNIST, CIFAR-10 and STL-10, verify the efficacy and robustness of the proposed algorithm.