基于CNN和表情识别技术的帕金森病诊断多任务学习研究  

A MULTI-TASK-LEARNING STUDY OF PARKINSON'S DISEASE DIAGNOSIS BASED ON CNN AND FACIAL EXPRESSION RECOGNITION TECHNIQUES

作　　者：毛政翔 章思远 肖世明 黄伟[1,2] Mao Zhengxiang;Zhang Siyuan;Xiao Shiming;Huang Wei(School of Information Engineering,Nanchang University,Nanchang 330031,Jiangxi,China;Information Office,Nanchang University,Nanchang 330031,Jiangxi,China)

机构地区：[1]南昌大学信息工程学院,江西南昌330031 [2]南昌大学信息化办公室,江西南昌330031

出　　处：《计算机应用与软件》2023年第11期163-169,265,共8页Computer Applications and Software

基　　金：国家自然科学基金项目(61862043);江西省自然科学基金重点项目(20181ACB20006)。

摘　　要：针对帕金森病患者筛查这一问题,提出一种基于卷积神经网络和面部表情识别技术的帕金森病患者识别的多任务学习网络模型(Densely connected multi-scale convolutional network,DMSNet),该模型主要采用密集连接网络和多尺度卷积核结构。通过获得表情预测结果后,将不同表情特征向量组合用以获得帕金森病患者预测结果。在OuluCASIA和PDface数据库上进行5折交叉实验后,该模型在帕金森病筛查上的有效性得到验证。此外,该模型采用多任务学习机制的优越性也通过与其只进行帕金森病预测的单任务学习模型对比后得到验证。Aimed at the problem of screening Parkinson's disease patients,a multi-task learning network model(DMSNet)for Parkinson's disease recognition based on convolutional neural networks and facial expression recognition technology is proposed.This model mainly used dense connection networks and multi-scale convolution kernel structure.After obtaining the expression prediction results,different expression feature vectors were combined to obtain the prediction results of Parkinson's disease patients.After performing a 5-fold crossover experiment on the OuluCASIA and PDface databases,the effectiveness of this model in screening for Parkinson's disease was verified.In addition,compared with the single-task learning model that only predicts Parkinson's disease,the superiority of the multi-task learning mechanism of this model was also verified.

关 键 词：卷积神经网络 图像识别 表情识别 帕金森病筛查 DMSNet 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]