Facial expression recognition based on transfer learning from deep convolutional networks

It is well-known that deep models could extract robust and abstract features. We propose a efficient facial expression recognition model based on transfer features from deep convolutional networks (ConvNets). We train the deep ConvNets through the task of 1580-class face identification on the MSRA-CFW database and transfer high-level features from the trained deep model to recognize expression. To train and test the facial expression recognition model on a large scope, we built a facial expression database of seven basic emotion states and 2062 imbalanced samples depending on four facial expression databases (CK+, JAFFE, KDEF, Pain expressions form PICS). Compared with 50.65% recognition rate based on Gabor features with the seven-class SVM and 78.84% recognition rate based on distance features with the seven-class SVM, we achieve average 80.49% recognition rate with the seven-class SVM classifier on the self-built facial expression database. Considering occluded face in reality, we test our model in the occluded condition and demonstrate the model could keep its ability of classification in the small occlusion case. To increase the ability further, we improve the facial expression recognition model. The modified model merges high-level features transferred from two trained deep ConvNets of the same structure and the different training sets. The modified model obviously improves its ability of classification in the occluded condition and achieves average 81.50% accuracy on the self-built facial expression database.