Double sparse local fisher discriminant analysis for facial expression recognition

In this paper, we propose a novel feature extraction method called double sparse local Fisher discriminant analysis (DSLFDA), which is an extension of the local Fisher discriminant analysis (LFDA) algorithm. The proposed method combines the idea of sparse representation to construct an adaptive graph to describe the structure information of the samples. Meanwhile, to obtain the sparse projection vectors, we first transform the original generalized eigenvalue problem to a regression-type problem with two variables. Then, l1 penalty was added to the objective function in the regression problem. One disadvantage of the sparse projection vectors is that which elements or regions of the pattern are important for each sparse projection vector. Experiments on the JAFEE and Cohn-Kande facial expression database show that the proposed DSLFDA is effective for recognition tasks and achieves competitive performance compared with other feature extraction methods.