Expression recognition algorithm based on local and global feature reconstruction residuals

In this paper, an algorithm is proposed to improve expression recognition by combining reconstructed residuals of local and global information. Firstly, positional relation is introduced into local binary features encoding, and a non-negative matrix decomposition method with sparse correlation constraints is employed to perform reducing of dimension. Then, the features are encoded by bag-of-words model to form local features. Secondly, a discrete wavelet transform is used to emphasis on expression information and extract expression-enhanced global features. Thirdly, collaboration Represents is used to calculate local partial reconstruction residuals and is combined with global reconstruction residuals to construct the final residual to be classified at the decision level. Experiments on Jaffe, TEFID and CK+ data sets prove the effectiveness of the proposed algorithm.