结合空洞卷积的CNN实时微表情识别算法

Real-time micro-expression recognition algorithm based on atrous convolutions for CNN

随着CNN等基于深度特征的人脸自发式微表情识别分类方法逐渐完善,相比于传统的特征提取方法更易满足应用实时性,针对微表情持续时间短、动作幅度细微,在多卷积层叠加会丢失图像中的细微信息的问题,为了完善细节信息,充分提取微表情细微特征,提出结合空洞卷积核及人脸自动校正算法,完善CNN特征提取过程,通过自动人脸矫正适应实际应用中的实时识别分类,在CASME及CASMEⅡ微表情公开数据集上完成模型训练及测试,通过损失函数方案对比提高模型鲁棒性,CASME中准确率为70.16%,CASMEⅡ中准确率为72.26%;实时识别帧率在60 fps。该方法能有效地提高微表情识别准确率,满足实时性要求,且具有较好的鲁棒性和泛化能力。

Based on the depth feature of facial micro-expression recognition,such as CNN,the classification method of facial micro-expression recognition is gradually improved.Compared with the traditional feature extraction method,it is easier to meet the real-time application.In order to perfect the details and extract the fine features of micro-expressions,this paper proposed a new algorithm combining the atrous convolutions kernel and the automatic correction of face to improve the feature extraction process of CNN network.It trained and tested the model on CASME and CASMEⅡmicro-expression public data sets through real-time recognition in real-time application of automatic face correction.It improved the robustness of the model by comparing the loss function schemes.The accuracy of the method in CASME is 70.16%and witch in CASMEⅡis 72.26%.Real-time recognition frame rate at 60 fps.This method can effectively improve the accuracy of micro-expression recognition,meet the real-time requirements,and has good robustness and generalization ability.