Research And Application Of Deep Learning Model For Face Emotion Recognition

In recent years,deep learning model has been developed rapidly,and deep convolution neural network as one of the methods has been widely used in computer vision.There are many factors affecting the performance of deep learning model.The selection of activation function and the structure of convolutional neural network have important effects on the performance of deep learning model.Aiming to improve the accuracy of facial emotion recognition model,this paper analyses the advantages and disadvantages of traditional activation functions such as ReLU,L-ReLU,P-ReLU,and compares them with new Swish activation function.Swish function is introduced into the deep learning model of facial emotion,and an optimized back propagation algorithm is proposed.A novel deep learning model Swish-FER-CNNs for facial emotion recognition is proposed by using multi-layer small size convolution block instead of large-scale convolution module to extract detailed features.The experimental results show that the recognition accuracy of Swish activation function is higher than that of ReLU,L-ReLU and P-ReLU activation functions.Combined with Swish activation function,the recognition accuracy of deep learning nerual network Swish-FER-CNNS is improved by 4.02%compared with the existing model.Swish-FER-CNNs nerual network improves the accuracy of facial emotion recognition,but it takes a long time to converge.The training time of deep model is another important index of network performance.In order to reduce the training time of model,this paper propose a brand new activation function S-ReLU based on the characteristics of easy calculation of positive half-axis derivative of ReLU activation function and strong robustness of negative half-axis of Swish activation function,and applies it to the deep learning model Swish-FER-CNNs,which is proposed earlier in this paper.Experiments show that Swish-FER-CNNs converge rapidly,the training time is greatly reduced while maintaining a high accuracy.At the same time,the TEST-ReLU activation function is constructed in reverse and tested on MNIST and FER-2013 datasets.The results show that the deep nerual network using TEST-ReLU activation function works well on MNIST,but it is difficult to converge on FER-2013,with a large fluctuation range.And the recognition accuracy decreases obviously,which proves the reliability and correctness of S-ReLU.The S-mobile-CNNs is constructed by simplifying Swish-FER-CNNs parameters with deep separable convolution.The experimental results show that the new network has small size,lesscomputation,fewer network parameters and good accuracy.Finally,Swish-FER-CNNs is applied to real industrial scene.A face emotion analysis system based on deep learning model is designed,which consists of two parts: face detection and face emotion recognition.HOG + SVM algorithm is used for face detection in video stream,and Swish-FER-CNNS is used for face emotion analysis.In addition,other deep learning model is used to design functional modules:face recognition and gender recognition.Based on the lightweight network S-mobile-CNNs,a face emotion recognition system at the mobile device is constructed,which has the characteristics of small size,small computation and easy operation.