一种基于CNN与LSTM结合的微表情识别算法  

A Micro-expression Recognition Algorithm Based on Convolutional Neural Network and Long Short-Term Memory

作　　者：吴进 闵育 马思敏 张伟华 WU Jin;MIN Yu;MA Simin;ZHANG Weihua(School of Electronic Engineering,Xi′an University of Posts and Telecommunications,Xi′an 710121,China)

机构地区：[1]西安邮电大学电子工程学院

出　　处：《电讯技术》2020年第1期1-7,共7页Telecommunication Engineering

基　　金：国家自然科学基金资助项目(61772417,61634004,61602377);陕西省重点研发计划(2017GY-060);陕西省自然科学基础研究计划项目(2018JM4018)

摘　　要：微表情表现强度微弱且非常短暂。针对微表情识别效果不理想的问题,以视觉几何组(Visual Geometry Group,VGG)网络为基础,提出卷积神经网络(Convolutionnal Neural Network,CNN)与长短期记忆网络(Long Short-Term Memory,LSTM)结合的识别算法。CNN提取数据集CASME II的空域特征,LSTM处理时域特征,实现空域与时域特征的结合。针对深度学习训练困难以及过拟合问题,加入批量归一化算法与丢弃法,提高网络训练速度,有效防止过拟合。针对数据集稀缺的问题,固定每次读取帧序列的长度,随机生成起始帧的位置,不断循环读取以遍历整个数据集并达到数据扩增。根据实验结果,五类微表情(高兴、惊讶、厌恶、抑郁、其他)识别率最高可达72.3%。Micro-expression is very weak and lasts for a short time.Based on visual geometry group(VGG),a recognition algorithm combining convolutional neural network(CNN)and long short-term memory(LSTM)is proposed for the problem of unsatisfactory recognition.The spatial domain features of the data set CASME II are extracted by CNN,temporal domain features are handled by LSTM,thus achieving the combination of spatial and temporal domain features.For the problems of the difficult of deep learning training and over-fitting,the Batch Normalization algorithm and Dropout are added to improve the network training speed and effectively prevent over-fitting.For the problem of data set scarcity,the length of each read frame sequence is fixed,the position of the start frame is randomly generated,and the loop is continuously read to traverse the entire data set and achieve data enhancement.According to experiment result,the recognition rate of five classes micro-expression(happy,surprise,disgust,repression,others)is up to 72.3%.

关 键 词：微表情识别 深度学习 卷积神经网络 长短期记忆网络 批量归一化算法 丢弃法 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]