Study On Facial Expression Recognition Algorithm Based On Convolutional Neural Network

Facial expression is a part of body language,which plays an irreplaceable role in interpersonal communication and social life.With the rapid development of computer hardware and the arrival of the era of artificial intelligence,how to make the computer recognize facial expressions,further understand people's inner thoughts,fully grasp people's behavioral intentions,has become a hot research field.In recent years,the research results on facial expressions have been widely used in the fields of safe driving,online education,auxiliary medical care,and commercial marketing,etc.The traditional method of facial expression recognition is to extract expression features by designing feature extraction operator manually,and then using appropriate classification algorithms to classify facial expressions.The traditional recognition method has the characteristic of fast recognition speed,but its recognition accuracy is often low.Compared with the traditional method,the convolutional neural network can automatically extract the deep features of facial expressions by constructing multiple convolutional layers.On the one hand,the error caused by artificial feature extraction is avoided,but on the other hand,it has strong robustness and generalization ability.However,it requires a large amount of labeled training data for network training.In practice,it is often difficult to pay a high price to manually label a large number of expression data sets.Therefore,a lightweight domain-adaptive convolutional neural network is designed in this paper.The network is based on a lightweight structure.Not only is a dual-mode attention module designed in the structure to make the model pay more attention to subtle changes in expression,but also through the domain adaptive learning strategy to use large data sets with tags for auxiliary training,making the model more conducive to wide application in real scenarios.The main research contents of this paper include the following aspects:(1)Taking the classical AlexNet network structure as a reference,a lightweight convolutional neural network model is designed,and the feasibility of the proposed network is verified by comparison of performance on public data sets.The proposed network first uses cascaded small convolution kernels instead of large convolution kernels.The increased non-linear transformation operation further improves the generalization ability when the number of model parameters is reduced.And then batch normalization is used instead of random deactivation,which speeds up the convergence of the network and suppresses overfrtting to a some extent.Finally,the Squeeze-and-Excitation(SE)module is combined to make different feature channels have different contribution,and feature extraction is more efficient under a simplified network structure.(2)A dual-mode attention network is designed.This network is based on the lightweight convolutional neural network proposed in this paper.The dual-mode attention network takes the dual-mode attention module as the core.By differently selecting the appropriate attention mode and applying it to the designed lightweight convolutional neural network,the network has a good feature expression for the subtle expression areas.The dual-mode attention module creates a self-learning weighting matrix of the same size as the feature map.During the process of the network training,the weighting matrix can change its weight value,so that the important feature areas are enhanced after multiplying the feature map,which is not important.Eventually,the effectiveness of the proposed attention module is verified by comparison experiments on the RAF-DB and CK+datasets.(3)A convolutional neural network based on domain adaptation is designed.This network is based on the dual mode attention network proposed in this paper.It realizes cross-domain knowledge sharing through domain adaptive learning methods,and solves the problem of fewer labeled corpora in practical application scenarios.The domain adaptive network firstly extracts the high-dimensional expression features of the source domain expression data and the target domain expression data through the fully connected layer,then calculates the maximum mean distance of the extracted features,and finally reduces the maximum mean distance as the optimization goal of the network.The knowledge of the source domain is transferred to the target domain,so that the unlabeled expression data in the target domain can be correctly classified.