摘要
针对新生儿疼痛表情识别任务中由于有类别标签样本数量不足而导致分类准确率不高的问题,提出了一种基于图的半监督深度学习(Graph-based Semi-supervised Deep Learning,GSDL)方法。首先,使用训练集中少量有类别标签的样本对深度神经网络模型进行初步训练,得到初始模型;然后,利用初始模型提取有类别标签样本与无类别标签样本的特征向量,并使用提取的特征向量构建一个邻接矩阵,进而构建一个图,在构建的图上通过标签传播算法推测出无类别标签样本的伪标签;最后,使用所有样本及其标签对深度神经网络模型进行微调,得到最终的新生儿疼痛表情识别分类模型。在新生儿疼痛表情数据集上的实验结果表明,在使用相同数量的有类别标签样本情况下,文中提出的GSDL模型的分类准确率优于传统的有监督深度学习模型,也高于现有的半监督深度学习模型(Mean-Teachers,MT),验证了GSDL方法在新生儿疼痛表情识别中的有效性。
The accuracy of recognizing neonatal facial expressions of pain is not high due to the insufficient labeled samples for training.In this regard,this paper proposes a graph-based semi-supervised deep learning(GSDL)method.First,a few labeled samples in the training set are used to train the deep neural network,and the initial model is obtained.Second,the initial model is used to extract the feature vectors of the labeled and the unlabeled samples,and the vectors are then used to construct an adjacency matrix and a graph.Third,the pseudo labels of the unlabeled samples are inferred from the constructed graph by using the label propagation algorithm.Finally,all the samples and their labels are used to fine turn the deep neural network and the final classification model for neonatal facial expressions of pain is obtained.The experimental results on the neonatal facial pain expression dataset show that the classification accuracy of the proposed GSDL model is superior to that of the traditional supervised deep learning model and the existing semi-supervised deep learning model MT(Mean-Teachers)when the same number of labeled samples are used.This demonstrates the effectiveness of the GSDL method in the recognition for neonatal facial expressions of pain.
作者
卢官明
宋统帅
楼亦墨
郑浩伟
闫静杰
LU Guanming;SONG Tongshuai;LOU Yimo;ZHENG Haowei;YAN Jingjie(College of Telecommunications&Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing 210003,China)
出处
《南京邮电大学学报（自然科学版）》
北大核心
2022年第1期53-61,共9页
Journal of Nanjing University of Posts and Telecommunications：Natural Science Edition
基金
国家自然科学基金(72074038,61971236)
江苏省重点研发计划(BE2016775)
江苏省研究生科研与实践创新计划(KYCX21_0742)资助项目。
关键词
半监督学习
基于图的半监督学习
半监督深度学习
新生儿疼痛
表情识别
semi-supervised learning
graph-based semi-supervised learning
semi-supervised deep leaning
neonatal pain
facial expression recognition
