Ideal Code Constrained Supervised Sparse Coding

In this paper, we proposed a novel sparse coding algorithm by using the class labels to constrain the learning of codebook and sparse code. We not only use the class label to train the classifier, but also use it to construct class conditional codewords to make the sparse code as discriminative as possible. We first construct ideal sparse codes with regarding to the class conditional codewords, and then constrain the learned sparse codes to the ideal sparse codes. We proposed a novel loss function composed of parse reconstruction error, classification error, and the ideal sparse code constrain error. This problem can be optimized by using the transitional KSVD method. In this way, we may learn a discriminative classifier and a discriminative codebook simultaneously. Moreover, using this codebook, the learnt the sparse codes of the same class are similar to each other. Finally, exhaustive experimental results show that the proposed algorithm outperforms other sparse coding methods.