Effective Micro-Expression Recognition Using Relaxed K-SVD Algorithm

Hao Zheng,Jie Zhu,Zhongxue Yang,Zhong Jin

DOI: https://doi.org/10.1007/s13042-017-0684-6

2017-01-01

International Journal of Machine Learning and Cybernetics

Abstract:Micro-expression recognition has been a challenging problem in computer vision due to its subtlety, which are often hard to be concealed. In the paper, a relaxed K-SVD algorithm (RK-SVD) to learn sparse dictionary for spontaneous micro-expression recognition is proposed. In RK-SVD, the reconstruction error and the classification error are considered, while the variance of sparse coefficients is minimized to address the similarity of same classes and the distinctiveness of different classes. The optimization is implemented by the K-SVD algorithm and stochastic gradient descent algorithm. Finally a single overcomplete dictionary and an optimal linear classifier are learned simultaneously. Experimental results on two spontaneous micro-expression databases, namely CASME and CASME II, show that the performance of the new proposed algorithm is superior to other state-of-the-art algorithms.