Facial Expression Recognition Based On Monogenic Binary Coding

Facial expression recognition (FER) is a technology which detects and locates face images from the complex background firstly, extracts expression features from all the obtained face images and classifies the images to the corresponding expression category with computer technology. FER has potential applicable prospect in medical field, distance education, etc. And it promotes the development of human-machine interaction, emotional computing, machine vision, etc. Therefore, it’s significant meaningful to research on facial expression recognition. However, under the influence of the complexity of expression itself and external environment etc., up to now FER rate especially person-independent FER rate is generally not high, which seriously restricts its actual application in many fields. Therefore, to improve recognition rate and robustness of FER algorithms is a challenging topic.Monogenic signal representation and monogenic binary coding(MBC) are researched and applied in FER field in this paper. And on facial expression database experiments related are carried out. The work of research in this paper is as follows.(1) Monogenic binary coding (MBC) applied in facial expression feature extraction and Block-based Fisher Linear Discriminant(BFLD) algorithm used for feature dimension reduction are researched. It is firstly introduced that band-pass monogenic signal representation based on Riesz transform decomposes an image into multi-scale amplitude, phase and orientation three complementary components. Then monogenic binary coding (MBC) is used to encode each monogenic components (i.e., amplitude, orientation and phase), and histogram feature extraction process is researched. Considering SSS (Small Size Sample) problem, BFLD is adopted to reduce feature dimension. Cosine Similarity measure showing direction difference is adopted for classification.(2) FER based on MBC is researched, each monogenic component (i.e., amplitude, orientation and phase) can be used individually for FER, three components can also be fused for more effective FER. Experiments carried out on JAFFE and Cohn-Kanade(CK) facial expression database verified the effectiveness of MBC based methods, among which the fusion method is most effective. Compared with state-of-the-art facial expression recognition algorithms such as residual fusion of LBP+SRC, Local Phase Quantization (LPQ) plus SRC, the results show that the MBC based methods especially the fusion approach are more effective. In person-independent experiments, the FER rate of fusion approach is70.95%on JAFFE database and76.86%on CK database. In person-dependent experiments, the FER rate of fusion approach is as high as98.1%on JAFFE database and100%on CK database.(3) The robustness to occlusion of the MBC based methods is researched in person-independent experiments. Experiments are carried out on JAFFE and CK databases with occlusion. By occluding each test sample10%,20%,30%,40%,50%, the results show both the FER method based on monogenic orientation component and the fusion method for FER are more robust to occlusion than residual fusion of LBP+SRC, Local Phase Quantization (LPQ) plus SRC etc.