Research On Facial Expression Recognition Algorithm Based On VGG19

Facial expression is one of the most important factors reflecting a person’s mental state,and it plays an indispensable role in interpersonal communication.With the rapid development of deep learning,the use of convolutional neural networks for facial expression recognition has become a mainstream method.Convolutional neural networks can automatically extract facial expression features,but how to optimize convolutional neural networks to extract more effective facial expression features and improve the recognition accuracy of the model remains the focus of facial expression recognition research.This thesis is based on VGG19 network and focuses on facial expression recognition.With the goals of improving the recognition accuracy of the network and reducing the number of parameters,the optimization of convolutional neural networks is studied.The main research work of this thesis is summarized as follows:(1)An improved network model MS-VGG19 based on VGG19 is proposed.Firstly,use Mish activation function to replace the Re LU activation function after the original network VGG19 convolution layer.Secondly,after feature extraction of facial expression images in the convolutional layer,Soft Pool is used to alleviate the problem of information loss caused by downsampling to improve classification accuracy.Finally,Batch Normalization(BN)is added to the network to accelerate the convergence of the model,and prevent the model from overfitting.The improved network model is validated on three datasets: FER-2013,CK+,and RAF-DB.Compared with the original VGG19 network,the recognition accuracy is improved by1.12%,2.27%,and 2.09%,respectively.The experimental results validate the effectiveness of the improved network model.(2)A lightweight network model DGE-VGG based on ECA attention mechanism is proposed.Firstly,introduce an efficient channel attention mechanism(ECA)into the network.Due to the fact that convolutional neural networks can only uniformly extract image information,but facial expression information is only distributed in local areas of the image.Adding attention mechanisms to the network can enable the network to filter out more effective feature information from facial expression images and improve the model’s ability to express important feature information.Secondly,in order to reduce the number of network parameters and make the network more lightweight,depthwise separable convolution is introduced,and global average pooling is used to replace the first two fully connected layers in the network.Finally,experiments are conducted on three publicly available datasets FER-2013,CK+,and RAF-DB.Compared to the original VGG19 network,the recognition accuracy on the datasets is further improved by 2.36%,5.75%,and 4.61%,respectively.The experimental results showed that the improved model has better recognition performance than the original model,proving the effectiveness of the improved method.