Expression Recognition Of Ward Monitoring Based On Lightweight Convolutional Neural Network

In recent years,the accelerating trend of population aging has made the global medical system face the serious problem of manpower shortage.The outbreak of COVID-19 in 2020 has aggravated the scarcity of efficient medical resources around the world,and the need to use artificial intelligence to relieve the shortage of medical resources has become more and more urgent.At present,there are few research schemes for ward monitoring that combine facial expression recognition and edge computing.Aiming at this situation and the problem of insufficient accuracy of facial expression recognition in lightweight convolutional neural network,the following in-depth research work is carried out in this paper:1.The traditional lightweight convolutional neural network Xception is refined into D-Xception to enhance the expression recognition rate.The improved method of this paper is adding the local dense connection and transition layer to the middle flow of the traditional Xception,which makes the middle flow into three improved dense block.The concrete measure of improved method is concatenating the input and output feature map of the four groups of depthwise separable convolution as the output of every improved dense block.And a 1×1 convolution in transition layer is used to adjust the channel number of dense block's output.In addition,the entry flow of traditional Xception is reduced to two convolution layers and two 1×1 residual blocks,while the exit flow structure remains unchanged.Finally,an improved network named D-Xception is obtained.The improved network can enhance the feature reuse and improve the expression recognition rate through local dense connection,while the number of parameters can be reduced by adjusting the number of channels.2.The improved network D-Xception is compared with the mainstream network in the application of facial expression recognition.The experiment is conducted on an open dataset FER2013 containing 35,887 emotion images,which is divided into a training set and a test set by a ratio of 8:2.The test results are compared with Dense Net,Mobile Net,Xception and other networks on accuracy and parameters.The experimental results show that the parameters of network D-Xception is 7.5×10~6,which is one third of the number of traditional Xception,and the accuracy of facial expression recognition can reach 70.97%,which is 2.7%higher than the 68.24%accuracy of traditional Xception.Through the analysis of confusion matrix,it is found that there is a problem of unbalanced expression recognition in the network,and the weighted cross entropy loss function is used to improve the network.Finally the expression recognition rate of Anger and Sad are increased by 4%and 3%respectively,making expression recognition more balanced.3.Based on the above work at the first and second points,combined with improved network D-Xception and edge computing,a ward monitoring system based on patient expression recognition is designed and implemented.The image preprocessing,face detection,face recognition and expression recognition modules of the system are all deployed on the hardware platform Raspberry Pi 3B and Intel neural computing stick,and the warning module runs on the upper computer.The system is verified from two aspects of functionality and robustness in this paper.Functional experiment results show that on the hardware platform,the system can directly recognize 7 kinds of basic expression of the patient in the camera view at 5 FPS,and analyze the patient's emotional status.When patients present abnormal status,the system can send out early warning to medical staff through TCP Socket connection.The robust experiment results show that the system can adapt to different lighting environment such as indoor night and intense sunlight,and can monitor the status of multiple patients in real time.In conclusion,this paper proposes an improved lightweight convolutional neural network D-Xception in theory and proves that the expression recognition accuracy of the D-Xception is increased by 2.7%compared with traditional Xception through experiments.A ward monitoring system based on D-Xception expression recognition is designed and implemented on hardware Raspberry Pi 3B and neural computation stick.Furthermore,the complete function of the system is verified,which proves that the network D-Xception can be deployed on embedded devices for expression recognition at 5 FPS,and the combination with edge computing can provide a practical and effective alternative for real-time monitoring of intelligent wards in the future.