Facial Expression Recognition Using Deep Belief Network

Emotional understanding and expression is a fundamental basis for human-computer interaction, and how to read the human mind through facial expression recognition technology has become a hot issue. Large dimension of image data, sample calibration difficulties, and small size training sample set make the efficient facial expression recognition task difficult. DBN (Deep Belief Network) achieves a layered understanding of knowledge through abstracting multiple representation and layers of the characteristics (knowledge) of learning objects according to the graded information processing of human brain. On basis of DBNs, this paper proposes a deep facial expression recognition system. The system first trains the Restricted Boltzmann Machine (RBM) network using the training sample set to obtain its RBM network-based probability distribution model, that is, the mapping relationship between network input and output; then, the multi-layered RBM network forms layered learning structure, realizing feature learning changes from fine-grained to coarse-grained scale, to achieve multi-layered feature learning and hierarchical levels of abstraction. Finally, the recognition system realizes facial expression recognition by resorting to training Softmax networks, which automatically learn expression image. The expression database JAFFE based experimental results show that our proposed deep facial expression recognition systemcan achieve realization of characteristic abstraction and form a good probability classification structure, and highest recognition rate can reach 98.75%.