Performance Evaluation of Stationary Wavelet Features with Least Squares SVM for Facial Expression Recognition

Abstract

This paper presents an improved method for recognition of facial expressions. The facial expression images are first decomposed using 3-level stationary wavelet transform(SWT). The feature vectors are then derived from the coefficients of high-frequency SWT sub-bands.Moreover,these high-frequency coefficients are useful to retain the edge information from expression images. SWT overcomes the issue of translation variant that traditional discrete wavelet transform (DWT) suffers. Also, SWT performs considerably well on translated images. To generate a set of compressed and discriminant features, linear discriminant analysis + principal component analysis (LDA+PCA) is applied. Finally, least squares SVM is used for the classification. Japanese female facial expression (JAFFE) and the Extended Cohn-Kanade (CK+) datasets are used to evaluate the system proposed. The proposed method achieves an accuracy of 98.72% and 98.63% on JAFFE and CK+ datasets, respectively. Experimental results based on 5-fold cross-validation test indicate the superiority of the proposed system over state-of-the-art schemes.Besides, the effectiveness of the SWT features is compared with the DWT features.