基于半监督对比学习的人脸表情识别

Facial expression recognition based on semi-supervised contrast learning

为解决面部表情识别(FER)大规模表情收集困难和现有表情无法满足实际细粒度需求的问题,在ResNet系列网络,以Resnet-18残差网络作为基础,首先引入图像预处理算法处理人脸表情图片,然后利用半监督学习方法将未标记数据与标记数据相结合,用以描述输入空间的数据分布。最后利用对比学习方法扩大类间距,减少类内差异。该方法在RAF-DB真实场景人脸表情识别数据集上进行了测试,其中2000个有标签的训练集测试准确率为81.37%,4000个有标签的训练集测试准确率为83.63%。

In this paper,a method of a facial expression recognition based on semi-supervised contrast learning is proposed.Based on Resnet-18 residual network,the image preprocessing module is added to process the input expression pictures.The advantage of semi-supervised learning is fully utilized,and the unlabeled data is combined with labeled data to better describe the synthesized data in the input space.The network is also used for comparative learning methods to expand the class spacing when clustering,while new data is automatically labeled in the embedded space by the class centroid distance.The method was tested on RAF-DB field facial expression recognition datasets,in which 2000 labeled training sets had a test accuracy of 81.37%and 4000 labeled training sets had a test accuracy of 83.63%.