关键词: 字典学习, 综合型-解析型字典对, 局部保持, 分类, 识别

Abstract: Dictionary learning (DL) has been widely applied in various computer vision problems. Most existing DL methods aim to learn a synthesis dictionary to represent the input signal while enforcing the representation coefficients and/or representation residuals to be discriminative. However, the l0 or l1-norm sparsity constraint on the representation coefficients adopted in most DL methods makes the learning phase time consuming. Hence, a new DL framework, namely analysis dictionary learning, has been proposed to deal with this problem. Besides, jointly learning a synthesis dictionary and an analysis dictionary has also drawn much attention recently, which can not only reduce the time complexity in the learning phase, but also lead to very competitive accuracy in a variety of classification and recognition tasks. This paper is motivated by the recently proposed dictionary pair learning algorithms. Our proposed novel Locality Preserving Discriminative Synthesis-Analysis Dictionary Pair Learning Algorithm integrates the local structure of training data, which is totally ignored in most applications of dictionary pair learning. We evaluate the proposed algorithm on three databases, including two face databases (Extended YaleB and AR) and one image categorization database (Caltech101). Based on the experiment results, we conclude that our algorithm can largely reduce the computational burden and improve the accuracies.

Key words: Synthesis-Analysis dictionary pair learning, Locality preserving, Classification, Recognition