基于过渡帧概念训练的微表情检测深度网络

Deep micro-expression spotting network training based onconcept of transition frame

为了更准确地从视频中检测面部微表情,针对微表情数据库样本规模较小的特点,采用迁移学习方法将深度卷积神经网络应用于微表情检测问题.选取预训练过的深度卷积神经网络模型,保留卷积层及预训练参数,添加全连接层和分类器,构造一个二分类的微表情检测深度网络(MesNet).为了去除微表情数据库中影响网络训练的噪声标签,提出过渡帧的概念和自适应识别过渡帧算法. MesNet在CASME Ⅱ、SMIC-E-HS与CAS(ME)2数据库上的曲线下面积(AUC)分别达到0.955 6、0.933 8与0.785 3,其中在CASME Ⅱ短视频数据库和CAS(ME)2长视频数据库上均取得最优结果,表明MesNet具有高精度和广适用范围的特点;过渡帧对比实验结果表明,构造训练集时从原始视频中去除过渡帧能够有效提高MesNet微表情检测性能.

A deep convolutional neural network was applied in view of the small sample size of micro-expression databases, in order to spot facial micro-expressions more accurately from videos through transfer learning. A pretrained deep convolutional neural network model was selected, and the convolutional layers and the pre-trained parameters were reserved. The full connected layer and the classifier were added after these layers to construct a deep binary classification micro-expression spotting network(MesNet). The concept of transition frame and an adaptive recognition algorithm of transition frames were proposed to remove the noisy labels from micro-expression databases that disturbed the network training. Experimental results show that the AUC values of MesNet on CASME Ⅱ, SMIC-E-HS and CAS(ME)2 reach 0.955 6, 0.933 8 and 0.785 3, respectively. Among three databases, MesNet achieves state-of-the-art results both on CASME Ⅱ which is a short video database and CAS(ME)2 which is a long video database. It shows that the proposed MesNet has the characteristics of high accuracy and wide application range. Comparison experiment results of the transition frame show that removing the transition frames from original videos when constructing the training set can effectively improve the micro-expression spotting performance of MesNet.