Facial Expression Recognition Based On Deep Learning

With the development of science and technology,facial expression recognition plays an im-portant role in the field of computer vision.It has been widely used in psychotherapy,education,artificial intelligence interaction and so on.Methods of facial expression recognition can be roughly classified as detector based and deep learning based.The detector based methods of facial expres-sion recognition consist of four steps,namely face detection,face alignment,feature extraction and facial expression recognition.Face detection and face alignment are hot research fields.For facial expression recognition methods,researches are focused on the design of facial expression features and facial expression classification algorithm.In deep learning based methods,we merge facial extraction and facial expression classification to learn facial expression features automatically.Facial expression recognition can be applied to static-based images or video-based sequences.For the static image,this paper proposes a new deep learning network structure and training method for facial expression recognition,which is called detail-perception transform network.The net-work structure consists of parallel nested small filter structures,which can effectively extract the characteristics of different receptive fields and reduce the network parameters effectively by using convolution.Thus the network can extract discriminative features and improve the training speed.Meanwhile,residual terms are introduced into the network,which can effectively avoid the loss problem of deep network information.In addition,the network is learned under a transfer learning framework,which uses large data from a similar problem to train the network model.In the paper,we propose that the boost-based method can improve the network accuracy effectively.For the video sequence,this paper proposes a multi-task recurrent neural network model.The model is learned by utilizing the multi-task learning of face recognition and facial expression recognition.Both tasks are promoted and learned by each other.The recurrent neural network structure can record the video information effectively.The model consists of coding network,face recognition network,temporary recurrent neural network and facial expression recognition network structure.To verify the multi-task recurrent neural network model,we collect a 500-video dataset of facial expression.We verify the detail-perception transfer learning network on the datasets Cohn-kanada+ and Kaggle.The experiments demonstrates that the network structure and training method can effec-tively improve the accuracy.The multi-task recurrent neural network based on video sequences is conducted on the I-PFE dataset.The experiment argues that merging the face recognition and facial expression recognition can better classify temporal-related information and temporal-independent information,and improve the accuracy of facial expression recognition.Meanwhile,the proposed method achieves better performance than state-of-the-art methods.