| 摘要: |
| 微表情是一种持续时间很短暂的面部表情。针对其识别率低的问题，提出了一种基于三维卷积神经网络（3D Convolutionnal Neural Network，3D-CNN）的微表情识别算法。使用Keras作为网络框架，在3D-VGG-Block(3Dimension Visual Geometry Group Block,3D-VGG-Block)的基础上加入批量归一化算法以及丢弃法，提升网络深度与训练速度的同时有效地防止过拟合；针对数据集稀少的问题，采取随机设置起始帧的位置，提前设定每次读取帧序列的长度，循环操作，在将所有数据均遍历的同时，达到数据增广的目的。该算法在CASME II数据集上的识别率最高达68.85%，在识别率上有一定优势。 |
| 关键词: 微表情识别 深度学习 三维卷积神经网络 批量归一化算法 丢弃法 |
| DOI： |
|
| 基金项目:国家自然科学基金资助项目（61772417，61634004，61602377）；陕西省科技统筹创新工程项目（2016KTZDGY02-04-02）；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） |
|
| A micro-expression recognition algorithm based on 3D-CNN |
| WU Jin,MIN Yu,LI Cong,ZHANG Weihua |
| (School of Electronic and Engineering,Xi′an University of Posts and Telecommunications,Xi′an 710121,China) |
| Abstract: |
| Micro-expression is a facial expression that lasts for a short time.For the problem of low recognition rate,a micro-expression recognition algorithm based on 3D convolutional neural network(3D-CNN) is proposed.Specifically,Keras is used as a network framework,and the batch normalization algorithm and dropout are added on the basis of 3D visual geometry group block(3D-VGG-Block).It effectively prevents overfitting while improving network depth and training speed.For the problem of rare data sets,the position of the starting frame is randomly set.Meanwhile,the length of the sequence of frames which need to be read is preseted every time,and the loop operation is performed.It achieves the goal of data augmentation while traversing all data.The recognition rate of the algorithm on the CASME II dataset is up to 68.85%,which proves its advantage in recognition rate. |
| Key words: micro-expression recognition deep learning 3D convolutional neural network(3D-CNN) batch normalization algorithm dropout algorithm |
《电讯技术》编辑部 