Neonatal Pain Expression Recognition Based On 3D Convolutional Neural Network

Modern medicine shows that persistent pain stimuli can damage the newborn's physiology and nervous system.In view of the fact that newborns cannot express their pain in words,the pain assessment for newborns currently relies on professional medical personnel.This assessment method mainly includes the subjective assessment and the shortage of manpower.Therefore,it is of great significance and value to develop an automatic pain assessment system for newborns to take timely measures of analgesia for the medical staff to relieve the pain of the newborn.In recent years,deep learning as a data-driven feature learning method has greatly promoted the development of the face expression recognition field.This article deeply discusses and studies the application of 3D Convolutional Neural Network(3D CNN)in neonatal pain expression recognition.The main work is as follows:(1)Established a video library for newborn facial expressions.The video library is the foundation of the research.This article has established a facial expression video library for newborns through video capture,video editing,and normalization.(2)A method for recognizing the pain of newborns based on C3D(Convolutional 3D Neural Network)was studied.C3 D can simultaneously extract features in the spatial domain and temporal domain,and can accurately express the nonlinear characteristics of facial expressions of newborns.In order to explore the influence of network depth on the recognition rate,two C3 D structures were studied: C3 D with 5 convolutional layers(C3D-5)and C3 D with 8 convolutional layers(C3D-8).Through experimental verification,C3D-5 achieved a recognition rate of 50.53% on the facial expression video database of newborns,and C3D-8 achieved a recognition rate of 51.27%.(3)A method based on R3D(3D Residual Convolutional Neural Network)for recognizing the expression of pain in newborns was studied.Network depth is an important factor affecting the final recognition effect.R3 D uses the concept of residual learning to deepen the depth of the network.The features extracted from 3D convolutions are more abstract and advanced.Through experimental verification,the recognition rate of neonatal pain expression based on R3 D reached58.84%.(4)A P3D(Pseudo 3D Convolutional Neural Network)based neonatal pain expression recognition method was studied.P3 D solves the 2D spatial domain convolution and 1D temporal domain convolution by integrating the traditional 3D volume.The network parameters are greatlyreduced.At the same time,the concept of residual learning is applied to deepen the network depth again.Through experimental verification,the recognition rate of neonatal pain expression based on P3 D has achieved a recognition rate of 55.69%.Although P3D-based recognition rate of neonatal pain expression is lower than that of R3 D,P3D has fewer network parameters and lower computational cost,thus greatly reducing training time and reducing the network model.