Research On Facial Expression Recognition Based On Convolutional Neural Network

Facial expression is the most important way of human emotion expression,and it is also a kind of non-verbal signal that is indispensable in the social communication between people.In recent years,with the rapid development of computer technology,exploring the emotional interaction between humans and machines has become a hot topic in current research.In order to accurately recognize facial expressions,many scholars have proposed a variety of practical methods.These methods can be roughly divided into two categories:expression recognition methods based on traditional machine learning and expression recognition methods based on deep learning.Among them,the deep learning algorithm can learn the essential characteristics of the data autonomously in an "end-to-end" manner,and has become the mainstream algorithm of current expression recognition research.This article starts from optimizing the network structure of deep learning algorithms,reducing the interference of redundant information,and improving the ability to distinguish features.The main research contents of the paper are as follows:(1)Summarize and analyze the structure and characteristics of the VGGNet network,and propose an improved VGGNet network.First,replace the last average pooling layer with a global pooling layer,while reducing the number of fully connected layers to reduce the amount of network parameters;then,a batch normalization layer and a DropBlock layer are added to avoid overfitting problems;finally,Island Loss and Softmax loss are jointly used as a new loss function to improve the discrimination of features.Experiments on CK+database show that the recognition rate of the algorithm is 95.41%,which is higher than 93.87%of VGGNet,thus proving the effectiveness of the algorithm.(2)In order to effectively use the information of the expression area,an expression recognition algorithm based on local feature fusion is proposed.The algorithm uses a multi-channel convolutional neural network consisting of a backbone part,a block part,a local feature extraction part,and a feature fusion and classification part.Among them,the improved VGGNet network is used as its backbone part to convert facial expression images into global feature maps;the custom block layer uses the information such as the center and block size of the block after the scale conversion to extract a Group local feature map;local feature extraction part is used to further extract the features of each local feature map;feature fusion and classification part is used to realize the classification of facial expressions.Through a series of experiments,it is proved that this algorithm has a higher recognition rate than the expression recognition algorithm based on the improved VGGNet network.(3)Since the features extracted by the convolutional layer are a combination of spatial and channel information,the information contained in different spaces and channel positions is somewhat different in importance,and it is proposed to introduce the attention mechanism to the multi-channel convolutional neural network.On the one hand,the channel attention module enhances the characteristics of key channels and suppresses the channel features containing redundant information;on the other hand,the regional attention module adaptively captures the importance of each local block,thereby reducing occlusion or non-expression Adverse effects of regional characteristics.The experimental results on the CK+database after occlusion simulation show that the algorithm is also robust to occlusion.