Abstract

This study presents a modified classification and regression tree (M-CRT) framework based on difference expression images, to address the facial expression recognition (FER) problem. The authors firstly obtain facial expressional details by calculating the difference between the images of basic expressions and images of neutral expression, which reflect the information irrelevant to identities. Local binary patterns and supervised descent method are, respectively, used to obtain the global and local features from difference expression images.

M-CRT model is developed for FER, which uses recursive segmentation to find the best classification decision according to the attributes of the global and local features, respectively. Compared with traditional methods, M-CRT can simultaneously maximise intra-class purity and distance between classes, which improves the discriminating power for classification. Experimental results on Japanese Female Facial Expression and CK+ database verify the effectiveness of their method.