半监督特征重构的无约束条件下人脸表情识别方法

Semi-supervised feature reconstruction for unconstrained face expression recognition

针对无约束条件下人脸表情数据的标签不足问题,本文提出了半监督特征重构的人脸表情识别方法。首先,构建了多头卷积神经网络,每个卷积头倾向于提取不同的特征,以丰富特征多样性;其次,使用五官关键点坐标构建结果用于提取局部特征的关键特征分支;再次,构建了自训练重采样模块,利用少数类的高精度预测补充标签数量;最后,使用一致性正则损失和伪标签方法学习无标签数据。分别在包含400个标签的RAF-DB和FERPlus数据集和包含250个标签的CIFAR-10数据集上进行测试,实验结果表明,所提方法的识别率分别达到68.88%、77.38%和95.75%,方法优于经典的半监督学习方法。

Aiming at the problem of insufficient labels for facial expression data under unconstrained conditions,this paper proposes a facial expression recognition method based on semi-supervised feature reconstruction learning.Firstly,a multi-head convolution neural network is constructed with each convolution head tended for extracting different features to enrich feature diversity.Secondly,the key feature branch for extracting local features is constructed using the landmarks of a face.Thirdly,a self-training resampling module is constructed to supplement the number of labels by using the high-precision prediction of categories with a small number of images.Finally,unlabeled data is learned using a consistent regularized loss and pseudo-labels.Experimental results show that the proposed method achieves 68.88%,77.38%and 95.75%recognition rates respectively on the RAF-DB and FERPlus datasets containing 400 labels and the CIFAR-10 dataset containing 250 labels which outperforming the classic semisupervised learning methods.