基于ME-Xception卷积神经网络的微表情识别  

Micro-expression Recognition Based on ME-Xception Convolutional Neural Network

作　　者：陈汤慧 高美凤 CHEN Tanghui;GAO Meifeng(Key Laboratory of Advanced Process Control for Light Industry,Ministry of Education,Jiangnan University,School of Internet of Things Engineering,Wuxi,Jiangsu 214122,China)

机构地区：[1]江南大学轻工过程先进控制教育部重点实验室,物联网工程学院,江苏无锡214122

出　　处：《信号处理》2022年第5期992-1000,共9页Journal of Signal Processing

基　　金：国家自然科学基金(61373126)。

摘　　要：微表情(Micro Expression,ME)存在面部肌肉运动幅度小、数据集样本少的问题,这会导致神经网络在学习过程中难以捕获有效特征和提高识别精度,因此,本文提出了一种基于改进Mini-Xception卷积神经网络的微表情识别方法。首先,在预处理阶段根据余弦相似度计算得到放大倍数,对微表情进行自适应运动放大;接着改进Mini-Xception网络模型,具体操作为在输入层两侧添加投影层以重整输入特征,将通道注意力机制加入由深度可分离卷积层和批归一化层组成的循环模块中,以此来构建ME-Xception网络模型;最后,将ME-Xception网络模型用于微表情识别任务,在CASMEⅡ、SAMM和SMIC数据集上进行实验,结果表明该方法有效提高了识别精度,与其他主流算法相比可以获得较好的识别性能。Micro-expression has the problems of small facial muscle motion range and few dataset samples,which will make it difficult for neural network to capture effective features in the learning process and improve accuracy. Therefore,a micro-expression recognition method based on improved Mini-Xception convolutional neural network is proposed. Firstly,in the preprocessing stage,the magnification is calculated according to the cosine similarity,and the micro-expression is amplified adaptively. Then,the Mini-Xception model is improved. The specific operation is to add projection layer on both sides of the input layer to reorganize the input characteristics,and add the channel attention mechanism to the circular module composed of deep separable convolution layer and batch normalization layer to construct the ME-Xception model. Finally,the ME-Xception model is applied to the task of micro-expression recognition. Experiments are carried out on CASME Ⅱ,SAMM and SMIC datasets. The results show that this algorithm effectively improves the recognition accuracy and can obtain better recognition performance compared with other mainstream algorithms.

关 键 词：微表情识别 运动放大 卷积神经网络 迁移学习 光流 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]