Facial Expression Recognition Based On Deep Learning

With the development of science and technology,facial expression recognition has become more and more important in the field of computer vision,and has been widely used in psychotherapy,online education and artificial intelligence interaction.From the perspective of research and development of facial expression recognition,it can be roughly divided into two categories.One is a classifier-based approach,and the other is an end-to-end deep learning approach.The traditional face expression recognition methods can be divided into four steps:face detection,face key point location,feature extraction and face expression recognition.The first two tasks have become a necessary research area.The feature extraction,feature selection and facial expression classification algorithms of human face are the core of facial expression recognition algorithm.The method of combining the two steps of feature extraction and facial expression recognition and learning together can automatically learn to select features related to facial expressions to perform facial expression recognition tasks based on the features of the deep learning method.At present,facial expression recognition can be divided into two categories,based on image and video sequence-based recognition methods.For static images,this paper proposes a convolution neural network structure based on Island loss for the static image expression recognition.The innovation of this method lies in the fact that the convolution neural network is used to extract facial expression features,which is more reasonable and accurate than artificially designed ones.The convolutional neural network uses a shallow network structure in order to avoid overfitting.In addition,a new loss function,Island loss,was proposed.It is designed to solve the difficulty of extracting facial expression features and improve the discriminability of features.Greatly reduce the possibility of misjudgment.For video sequences,this paper proposes a local bidirectional recursive recurrent neural network(PHRNN)structure for extracting timing information between video sequences.This model performs two-way recursive recurrent network(BRNN)on various parts of the face.It can extract information on the change of each part in the time series,and then fuses in the high-level network,and finally obtains information on the change of the overall face shape in time series.In order to further improve the accuracy of the expression recognition of the video sequence,the method of model fusion is used to combine the spatial information and the temporal information to perform face expression prediction of the video sequence.For images,we performed experiments on FER-2013 and Cohn-Kanade(CK+)datasets.Experiments have shown that this method can improve the accuracy and robustness of recognition.For the video,we performed experiments on CK+,Oulu-CASIA,and MMI datasets.The experimental results show that the proposed model fusion method achieves better recognition results.Using the method of model fusion,a richer expression-related information can be used to perform the final expression prediction,and the recognition rate is greatly improved.