Weighted Discriminative Sparse Coding for Image Classification

In the discriminative sparse coding, the reconstruction residual over each class-specific subdictionary can provide great discriminative and label information. In this paper, we propose a weighted discriminative sparse coding method by using the residual as the weight. For a test sample, we first compute its sparse code over each learnt sub-dictionary, and then use the reconstruction residual over each sub-dictionary to weight the corresponding sub-dictionary, thereby forming a weighted sample-specific structure dictionary, over which we compute a new sparse code for the test sample. This code carries more discriminative information about interclass difference. Our method yields a unique structure dictionary for each test sample, so that samples with the same class labels have more similar distributions of dictionary atom contributions. Experimental results demonstrate that the proposed method outperforms some state-of-the-art methods under the same learning conditions.