Abstract: According to the problem that traditional facial expression recognition method could not act a robust performance, we propose an algorithm based on deep learning. First of all, we train two sparse auto encoder in two scales, and the parameter of the hidden layer should be a series of convolutional kernel, we use these kernels to extract firstlayer features. Then we get secondlayer features through maxpooling operators, it improves the invariance of the features. Finally we parallelize seven fourlayers neural networks to accomplish the recognition task. The experiment result shows this deep neural networks structure act a robust performance in facial expression recognition task in the case of the test samples’ ID information did not appear in the training samples.

Key words: facial expression recognition, deep learning, neural networks, auto sparse encoder