A Dictionary Learning Method Based on Self-adaptive Locality-Sensitive Sparse Representation

Abstract

In traditional sparse dictionary methods, dictionary obtained from the dictionary update stage is used directly for sparse coding stage and the loss of data potential discriminative information during the learning process could be caused. To overcome this problem, we propose a dictionary learning algorithm based on self-adaptive locality-sensitive sparse representation. The algorithm considers the relationship of linear representation between dictionary atoms to build a self-adaptive local adaptor for sparse coding. Each dictionary atom will be linear represented and reconstructed by the rest dictionary atoms from the same category, and the adaptor will be constructed by balancing the differences between each reconstructed dictionary atom and each training sample. Under the premise of guaranteeing sparsity, the proposed algorithm guides sparse coding process gradually by using the designed self-adaptive local adaptor to adjust the dictionary adaptively, then gains the potential locality and similarity of the dictionary atoms efficiently and fastly, so that the optimized dictionary can represent training samples more faithfully and the sparse coefficients of samples can better express the intrinsic nature of the training samples. We execute several experiments on two commonly used video sets, and the experimental result shows that the dictionary obtained from the new dictionary learning model can be better adapted to video data, and improve the accuracy of video classification, which verifies the effectiveness of our proposed algorithm.