基于深度学习的老年人疼痛表情识别

【作者】 张敏；

【导师】 孙瑜；

【作者基本信息】 南京理工大学 ， 控制理论与控制工程， 2021， 硕士

【摘要】 随着我国城镇老龄人口“空巢化”现象日益严峻,越来越多的老年人处于独居状态。对于口齿不清、行动不变的老年残障人士,饮食和大小便需求可由护工定时完成,但在缺少看护、出现异常疼痛情况时,现有的语音、手势等交互方式往往不能进行实时检测并报警,使得老人无法获得及时救助。因此本文基于深度学习相关知识,对老人面部检测、疼痛表情识别技术进行研究,构建老人疼痛识别系统,实现对老人健康状况的实时监测。本文主要工作如下:(1)构建疼痛表情数据集。目前疼痛表情识别系统服务对象主要是新生儿,国内外针对疼痛表情识别构建的数据库主要来自于实验室下场景,因此某些疼痛表情并非自发产生,数据集自身数量及质量的不足限制着疼痛自动识别系统开发与应用泛化。针对此问题,本课题组成员实地走访江苏地区养老护理场所,通过相关调研,拍摄了大量的老人不同疼痛状态的图片,与专业的医护人员进行疼痛等级评分,构建了老人疼痛表情数据集,并且采用了裁剪、翻转、旋转和加入噪音点等方式有效扩充了样本数量,丰富了研究对象的样本。(2)分析并改进面部检测和对齐技术。由于护理床上的老人可能具有不同的面部姿态,直接使用现有的人脸检测器取得的效果并不理想。因此,本文基于多任务卷积神经网络构建老人面部检测器,针对该网络在检测速度和老人面部对齐上的不足,结合网络与老人自身面部特点对其进行改进,提高检测速度和面部对齐的精确度。(3)探究疼痛表情识别技术。经典的Alex Net网络训练易发生过拟合和疼痛表情的识别率低下问题,本文对经典的Alex Net网络在训练时做了一些改进:预训练-微调、多尺寸卷积核和批量归一化。针对老人表情图片对不同光照不具备鲁棒性问题,本文将局部二值模式特征映射和Alex Net网络结合进行老人疼痛表情识别。(4)进行老人疼痛表情识别系统的应用研究。结合老人面部检测与疼痛表情识别算法的研究,构建一个完整的老人疼痛表情识别分类系统,将其应用在实验室自主研发的护理床上进行测试研究。更多还原

【Abstract】 As the "empty nest" phenomenon of the elderly population in Chinese country’s urban areas has become increasingly severe,more and more elderly people are living alone.For elderly disabled people with slurred speech and limited mobility,the needs of diet and toileting can be completed by nursing staff regularly.However,when there is a lack of nursing care and abnormal pain,the existing interactive methods such as voice and gestures often cannot be detected and alarmed in real time,making it impossible for the elderly to obtain timely assistance.Therefore,based on the related knowledge of deep learning,this thesis studies the facial detection and pain recognition technology of the elderly,and builds a pain recognition system for the elderly to realize real-time monitoring of the health status of the elderly.The main work content is as follows:(1)Construct pain expression data set.At present,the service objects of the pain expression recognition system are mainly newborns.The databases built for pain expression recognition at home and abroad mainly come from the scenes in the laboratory,so some painful expressions are not spontaneous.The defects in the quantity and quality of the data set limit the algorithm development and application generalization of the automatic pain recognition system.(2)Analyze and improve face detection and alignment technology.Since the elderly on the nursing bed may have different facial postures,The effect of directly using the existing face detector is not ideal.Therefore,this thesis builds an elderly face detector based on a multi-task convolutional neural network.Aiming at the shortcomings of the network in detection speed and facial alignment of the elderly,it is improved by combining the characteristics of the network and the facial features of the elderly to improve the detection speed and accuracy of facial alignment.(3)Explore the pain facial expression recognition technology.The classic Alex Net network training is prone to overfitting,and the recognition rate of painful expressions is low.This article has made some improvements to the classic Alex Net network during training: Pre-training-finetuning,multi-size convolution kernel and batch normalization.Aiming at the problem that elderly facial expression pictures are not robust to different illuminations,this thesis combines local binary pattern feature mapping and Alex Net network to recognize elderly painful facial expressions.(4)Carry out the application research of the elderly pain facial expression recognition system.Combining the research of elderly face detection and pain expression recognition algorithm,a complete elderly pain expression recognition classification system is constructed,and it is applied to the nursing bed independently developed by the laboratory for testing and research.更多还原