Sparse Autoencoder for Facial Expression Recognition

Facial expression recognition has become one of the most interesting topics in human computer interaction. A lot of methods have been proposed and studied for facial expression recognition. Among some of these methods, feature extraction is very important. However, feature extraction in these methods involves human intervention more or less which may make the recognition task unstable and not suitable for practical application. To overcome these problems, a deep learning method for facial expression recognition is proposed. First, it divides the training images to 7 groups correspond to 7 expressions to train 7 sparse auto encoder networks. Then, it lets each training group gets thought the network that have been trained by this group, and each training image gets 150 outputs as the training features. In testing step, a testing image is input into the 7 trained networks and gets 150 outputs as the testing features of each network. Comparing the square of error between testing features with the training features in each network, and classifying the testing image into the class with the minimum error. Experiments on JAFFE database has proved the effectiveness of the proposed method.