摘要
由于微表情持续时间短、动作幅度小,因此微表情自动识别一直是一个具有挑战性的问题。针对上述问题,提出一种用于微表情识别的三维卷积神经网络进化(Three-Dimensional Convolutional Neural Network Evolution,C3DEvol)方法。该方法使用能有效提取动态信息的三维卷积神经网络(Three-Dimensional Convolutional Neural Network,C3D)来提取微表情在时域和空域上的特征;同时使用具有全局搜索和优化能力的遗传算法对C3D的网络结构进行优化,以获取最优的C3D网络结构和避免局部优化。利用CASME2数据集在带有两块NVIDIA Titan X GPU的工作站上开展了实验,结果表明C3DEvol微表情自动识别的准确率达到63.71%,优于现有的微表情自动识别方法。
Due to the short duration of micro-expressions and the small amplitude of motion,the automatic recognition of micro-expressions is still a challenging problem.Aiming at the problems,this paper proposes a Three-Dimensional Convolutional Neural Network Evolution(C3DEvol)method for micro-expression recognition.In the C3DEvol,three-dimensional Convolutional Neural Network(C3D)which can extract dynamic information effectively is used to extract micro-expression features in time domain and space domain.At the same time,the genetic algorithm with the capabilities of global search and optimization is used to optimize the network structure of C3D in order to obtain the optimal network structure and avoid local optimization.Experiments are performed on a workstation with two NVIDIA Titan X GPUs using the CASME2 dataset.Experiments show that the accuracy of C3DEvol micro-expression automatic recognition reaches 63.71%,which is better than the existing micro-expression automatic recognition method.
作者
梁正友
何景琳
孙宇
LIANG Zheng-you;HE Jing-lin;SUN Yu(School of Computer and Electronics Information,Guangxi University,Nanning 530004,China)
出处
《计算机科学》
CSCD
北大核心
2020年第8期227-232,共6页
Computer Science
基金
国家自然科学基金(61763002)。
关键词
微表情识别
遗传算法
三维卷积神经网络
特征提取
网络结构优化
Micro-expression recognition
Genetic algorithm
Three-dimensional convolutional neural network
Feature extraction
Network structure optimization
