用于微表情识别的改进双流浅层卷积神经网络  

Enhanced Dual-stream Shallow Convolutional Neural Network for Facial Micro-expression Recognition

作　　者：李昆仑[1] 陈栋[1] 王珺 王怡辉 LI Kun-lun;CHEN Dong;WANG Jun;WANG Yi-hui(Electronic Information Engineering College,Hebei University,Baoding 071000,China)

机构地区：[1]河北大学电子信息工程学院,河北保定071000

出　　处：《小型微型计算机系统》2021年第6期1219-1226,共8页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(61672205)资助.

摘　　要：在微表情自动识别任务中,浅层卷积神经网络和深层网络相比更好地改善了网络训练过拟合的情况,但是多数浅层卷积神经网络存在输入特征单一和提取高维有效特征能力不足的问题.针对上述问题本文同时使用图像的灰度特征和运动特征表征原图像,并且提出了一种改进双流浅层卷积神经网络(Enhanced Dual-stream Shallow Convolutional Neural Network,EDSSNet)用于微表情的识别.本文首先使用欧拉视频放大算法和TV-L1光流法对视频关键帧处理,提取图像的灰度特征和运动特征,然后用空洞卷积和注意力模块改进双流浅层卷积网络模型,提高网络提取有效特征的能力,最后将两种特征输入网络训练后进行分类.理论分析及在CASMEⅡ、SMIC-HS和SAMM微表情数据库上的实验结果均表明了改进模型的有效性.In micro-expression recognition,comparing with deep neural networks,shallow networks show a better performance in solving the overfitting problems,but most shallow convolutional neural networks have the problems of single input features and insufficient ability to extract high-level effective features.In response to above problems,the algorithm in this paper uses both the grayscale feature and the motion feature of the image to represent the original image.This paper proposes an enhanced dual-stream shallow convolutional neural network(EDSSNet)to identify micro-expression.The algorithm in this paper first uses Euler video magnification algorithm and Total Variation-L1 optical flow method to obtain the gray-scale features and motion features of the image.Then,we use the atrous convolution and attention module to improve the dual-stream shallow convolution network model,which can enhance the ability of the network to extract effective features.Finally,two kinds of features are input into the network for classification.Theoretical analysis and experimental results on the CASMEⅡ,SMIC-HS and SAMM micro-expression databases prove the effectiveness of the improved model.

关 键 词：微表情识别 双流卷积神经网络 欧拉视频放大算法 TV-L1光流法 空洞卷积 注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]