| Facial expression recognition is a leading subject of pattern recognition and computer vision field, due to the property of non-contact, concealment, easy to understand, low cost of devices, face recognition has been increasingly used in security monitoring, human-computer interaction, artificial intelligence and e-commerce security. In recent years, based on sparse representation and depth of learning methods in face recognition achieved good results, gaining more and more researchers’ attention, but the facial expression recognition research stalled,This paper summarize the basic theory lasted researches, build two deep learning network structures for facial expression recognition and extraction identity maintain features.The feature extracted by traditional method is not stabilize to complex changes like illumination, facial expressions, pose, the traditional features can not accurately describe the facial features, I proposed a deep learning structure for facial expressions recognition, which use big data to find robust features, first,I use unsupervised learning methods get a series of convolution kernels in two scales, this process is only need no labels data, second I use the convolution kernel extraction features, max-pooling enhance the robustness and reduce the dimension of features, The third, I use the parallel network to handle the features,in the last layer is a softmax classifier, we fine-tune the network to achieve a supervised network for classification, experimental results show that the algorithm has higher recognition rate and better robustness.For the nonlinear mapping ability of based on convolution neural network,this paper propose a method of extraction identity maintain feature by a convolution neural network, in this network I use some novel method to updates parameter, and use two loss function to regularize and train the network, use the JAFFE and C-K dataset to train and test this network, and compare the different face features extraction by the network. |
