Research On Facial Expression Recognition Based On Deep Convolutional Neural Network

Facial expression is an important way for people to express their feelings.In recent years,with the development of computer realm,facial expression recognition has become a current research hot spot and made remarkable progress,which can be applied to human-computer interaction,emotional computing and other computer vision fields.The development of artificial intelligence and deep learning has better facilitated the study of facial expression recognition.The facial expression features extracted by the traditional facial expression recognition algorithm based on machine learning are interfered by human factors,so that the training model has insufficient generalization ability and low recognition accuracy.The excellent performance of convolution neural network in deep learning algorithm in facial expression recognition task has attracted the attention of many scholars.However,facial expression recognition based on deep learning in real scenes is still affected by several factors such as human posture,facial occlusion,background environment and light interference,thus the accuracy of recognition still needs further improvement.Therefore,this paper proposes an improved facial expression recognition model based on deep convolution neural network for improving the accuracy of facial expression recognition.The specific work and innovations are as follows:(1)Aiming at the low accuracy of traditional facial expression recognition models,two improved convolution neural network models are proposed in this paper for comparison and selection.The first model is the optimized VGG12 network which is applied to facial expression recognition task.The structure of VGG12 network is based on VGGNet,with its system structure changed and Dropout added to avoid overfitting.The second model is the deep convolution neural network(DCNN)based on depth separable convolution.The structure of DCNN refers to the superposition of convolution blocks in VGG12,but the core convolution layer is replaced by the depth separable convolution layer,together with the use of convolutional residual blocks.In the case of reducing the parameters,the network can extract multi-scale feature information and effectively retain the detailed features.Through experiments,it is found that DCNN not only has fewer parameters,but also has a slightly higher accuracy than VGG12.Therefore,DCNN is chosen as the basic feature extraction network.(2)The lack of diversity and easily distinguishable feature information in the input data of the convolution neural network may affect the performance of the network and lead to insufficient extraction of facial expression features.In order to solve the problems above,the attention mechanism can be introduced into the network to make it ignore irrelevant features in the images and focus on effective ones.Therefore,this thesis adds the convolution block attention module to the DCNN model to form the new DCNN-CBAM network.So that the network can calculate the attention map for the given input feature mapping along the channel and space dimensions,and then combine the input feature mapping with its attention map to complete the refinement of features so as to improve the feature expression ability of the network.The experimental results on different datasets show that the introduction of the attention mechanism into the DCNN network can effectively improve the performance of the network.(3)In view of the problem that the Re LU activation function will cause negative neuron necrosis in the process of network training,and the traditional Softmax loss function can not solve the situation that there are large differences between similar expressions and small differences between different types of expressions in the training data of facial expressions,Mish activation function and AMSoftmax loss function are introduced into DCNN-CBAM network.The Mish activation function can avoid neuronal necrosis,while the AM-Softmax loss function can maximize the difference between classes.The experiment proves that the introduction of Mish activation function and AM-Softmax loss function can improve the training of DCNN-CBAM and make it have better feature expression ability.