Development And Implementation Of Neonatal Pain Expression Recognition Demonstration System

Studies have shown that neonates have the ability to perceive pain at birth,and that a large number of painful stimuli have a significant adverse effect on the growth and development of neonates.Therefore,timely and effective pain assessment and early intervention are of great significance.However,neonates lack the ability to express pain levels in language,at present,it mainly relies on manual pain levels assessment and it is time-consuming and laborious and subjective.Therefore,an efficient,accurate and convenient automated evaluation system is urgently needed to assist parents and medical staff to judge the pain levels.At present,deep learning-based neonatal pain expression recognition technology has made some progress,but because the commonly used deep learning neural network architecture is complex,requires a lot of computing resources and high hardware configuration,usually needs to be deployed and run on a large server,so the related techniques of neonatal pain expression recognition is further explored and a different approach is proposed in this thesis.In particular,the compression,transplantation and deployment of neural network on mobile terminal is studied,and the construction and design of neonatal pain levels automatic evaluation system on mobile terminal is completed in this thesis.The main research work includes:(1)To establish and expand the image and video database of neonatal pain expression.Both deep learning and traditional machine learning methods require image data as support.Therefore,the main job in the early days was to collect facial expression videos of neonates,and scientifically evaluate them by professional medical staff,and we establish a video database composed of more than 500 videos.Based on this,the video database was manually cut and calibrated,and finally a database of neonatal pain expression images containing about 12,000 images of quiet,crying,low-grade pain and high-grade pain was established by us.(2)To study the neonatal face detection algorithm.The study used a Single Shot MultiBox Detector(SSD)algorithm for facial detection of neonates.The advantage of the SSD algorithm is that the detection speed is fast,and the real-time requirement can be achieved under the better detection results.The disadvantage is that the detection results on the small target is not ideal,but usually the ratio of the whole photographed picture occupied by the neonatal face area in the clinical application is not too small,so this shortcoming can be ignored.The experimental results show that the SSD algorithm still has a good detection results in the case of slight occlusion anddeflection of the neonatal face.(3)To study the neonatal pain expression recognition algorithm.The basic principles of traditional machine learning methods(LBP+SVM),VGG network,and MobileNet network and their effects in neonatal pain expression recognition is studied and compared in this thesis.For the four types of neonatal expression(quiet,crying,low-grade pain and high-grade pain),the VGG network has the best recognition effect,and the recognition accuracy rate is 79.84%,but the VGG network has more parameters and more calculation than MobileNet,and can only be deployed on the server.MobileNet’s recognition accuracy rate is the second,about 76.25%,but its parameter is small,and the calculation is small,suitable for being transplanted to mobile phone terminals;LBP+SVM algorithm has the lowest recognition accuracy,about 65.7%.(4)To design the neonatal pain expression recognition system.The theory of compression transplantation of convolutional neural network and the compression effects of various compression algorithms on VGG network is studied and compared in this thesis.Finally,we complete the design and development of neonatal pain expression recognition demonstration System based on LBP+SVM,VGG(distributed on cloud server)and MobileNet network.