Human vision inspired feature extraction for facial expression recognition

Facial expression is a powerful way for human emotional communications. According to various applications, automatic facial expression recognition becomes an interesting problem for researchers in different areas. An automatic facial expression recognition system recognizes the expressed emotion in input facial image using several processing stages. Feature extraction is a vital step in facial expression recognition. One of the most widely used techniques for feature extraction in machine vision is utilizing Gabor filter which is sensitive to lines at various orientations. The disadvantage of Gabor filters is their computational cost and large feature vector length. Inspiring the human vision system and stimulations of complex cells, in this paper, firstly, facial image is convolved with Gabor filters. Then, the achieved convolution matrices are properly coded based on the maximum and minimum responses. Finally, the feature vector is obtained by calculating the histogram of these codes. The length of achieved histogram for 16 and 8 Gabor filters are 240 and 56, respectively, which is considerably less than keeping all Gabor responses. The proposed method is (person-independently) evaluated on four facial expression recognition datasets including CK+, SFEW, MMI, and RAF-DB. The experimental results show that the proposed method outperforms existing image texture descriptors in facial expression recognition in both controlled and uncontrolled images.