Ideal Code Constrained Supervised Sparse Coding

In this paper, we proposed a novel sparse codingalgorithm by using the class labels to constrain the learningof codebook and sparse code. We not only use the classlabel to train the classifier, but also use it to constructclass conditional codewords to make the sparse code asdiscriminative as possible. We first construct ideal sparsecodes with regarding to the class conditional codewords,and then constrain the learned sparse codes to the idealsparse codes. We proposed a novel loss function composed ofparse reconstruction error, classification error, and the idealsparse code constrain error. This problem can be optimizedby using the transitional KSVD method. In this way, wemay learn a discriminative classifier and a discriminativecodebook simultaneously. Moreover, using this codebook, thelearnt the sparse codes of the same class are similar to eachother. Finally, exhaustive experimental results show thatthe proposed algorithm outperforms other sparse codingmethods.