Multi-mode emotion recognition based on generalized discriminative canonical correlation analysis

In recent years, emotion recognition encounters difficulties: accuracy and robustness. In order to improve performance of the emotion recognition system, here we present a novel multimode emotion recognition system, including visual and audio information. Meanwhile, a feature fusion algorithm named Generalized discriminative canonical correlation analysis (GDCCA) is proposed and utilized in this system. First, we extract video image features through 2D Gabor wavelet and obtain the statistical and annotation features of audio. Then the above features are fused by GDCCA to be a fusion feature which is fed into SVM classifier to get the recognition results. Finally, the novel emotion recognition system is applied on the database of BUBE (Beihang University Biomodal Emotion Database) and the experiments provide extensive illustrations of the systems performance.