基于深度学习的可靠表情数据识别

Recognition of Reliable Facial Expression Data Based on Deep Learning

当前深度学习已成为表情识别领域的重要研究方法,但此方法应用于真实环境或者复合表情数据库下时识别准确率非常低下。为此提出一种深度局部关联神经网络DLR-VGGNet(Deep Locality-Relevance VGGNet)的可靠表情数据识别方法,首先在VGGNet网络添加一个新的监督层,即局部关联损失(LR loss),提高深层特征的判别能力,之后在不同的人脸表情数据库中基于这种DLR-VGGNet网络进行训练并且进行网络参数微调和测试。最后,RAF-DB数据库中对7类基本表情和11类复合表情做基准实验以及在SFEW和CK+数据库中做对比实验,实验结果表明在真实环境基于DLR-VGGNet的方法优于传统的手工特征提取方法。

At present,the deep learning field has become one of the hotspots in various fields.The traditional facial expression recognition method is applied to the real environment or the compound expression database with very low accuracy.This paper presents a Deep Locality-Relevance VGGNet(DLR-VGGNet),first adds a new supervision layer in the VGGNet network,namely the local association(LR loss),improving the ability of identify deep feature,after the facial expression is different in the number of database in the DLR-VGGNet network based on the network parameters and training fine tuning and testing.Finally,7 basic expressions and 11 kinds of compound expressions in RAF-DB database are tested as a benchmark,and a comparative experiment is made in SFEW and CK+database.The experimental results show that the DLR-VGGNet method is superior to the traditional manual feature extraction method in real expression recognition environment.