Facial Expression Recognition Based On Deep Learning

As an intelligent human-computer interaction technology,facial expression recognition is an important part of artificial intelligence in emotional computing,and has been widely concerned and studied by scholars both at home and abroad.At present,the research work of facial expression recognition is still in the experimental stage,the ability of characteristics represent and expression classification need to be further improved for traditional facial expression recognition methods.Deep learning has a significant advantage in the field of computer vision and artificial intelligence.Therefore,it is of great theoretical and practical value to study facial expression recognition through deep learning in this thesis.The main works are as follows:(1)Considering the increase of hidden layer in convolution neural network(CNN)will make the training more difficult,this thesis proposed a CNN intensive training method based on learning rate optimization by analysing the causes of blocking information flow in the network,which lays the foundation for the later training of facial expression model.There are two parts in this method: one is to construct the optimization function about the learning rate according to the current position and the gradient information in each iteration of the model,and obtain the current optimal learning rate,and provide the optimal step size of the model movement;the other part is the hierarchical adjustment of the learning rate in order to enhance the ability of the parameters in the shallow structure of the model to improve the ability to extract the effective features.The experimental results show that the method can accelerate the convergence of the model and help to jump out of the local minimum.(2)In order to solve the problem of insufficient representation of traditional facial features,this thesis studied the data preprocessing,facial feature extraction and expression classification of facial expression recognition by deep learning technique.Firstly,the expression method of expression sample data with rotation and translation invariance is designed and implemented,and the method of data balance of expression samples which can provide unbiased learning is proposed.Secondly,the facial feature extraction method based on CNN is adopted.By setting the type and number of layer in CNN,the size of the convolution kernel,and the size of each layer and the selection of the activation functions,which includes three convolution layers,three pool layers,and the feature extraction structure that joint RRe LU and Sigmoid activated function;Lastly,we fusion SVM to classify and match facial expressions.The experimental results show that the model has strong ability of expression recognition.Through the comparative experiment and analysis in the Cohn-Kanade expression database,it is proved that this method has obvious advantages in the classification performance compared with the traditional method,and has stronger generalization ability.