首页|融合光流与多注意力机制的微表情识别

融合光流与多注意力机制的微表情识别

扫码查看
针对微表情发生在面部特定区域,且持续时间短、运动幅度小,导致识别准确率较低的问题,探究一种融合光流和多注意力机制的人脸微表情识别方法.首先,利用总变差和L1范数的光流算法(TVL1)提取顶点帧与起始帧的面部光流特征,计算光流应变作为补充特征;然后,使用三种光流特征作为输入,利用引入全局注意力机制(GAM)和双注意力机制(DA)的残差神经网络(Resblock)进行特征提取和微表情识别,以减少面部信息弥散,并对全局光流特征进行放大;最后,在CASME和CASME Ⅱ数据集上进行分类实验,验证所提方法对正面人脸微表情识别的可行性和有效性.
Facial Micro-expressions Recognition Based on Fusion Optical Flow and Multi-attention Mechanism
Aiming at the problem of low recognition accuracy due to the short duration and small motion range of micro-expressions occurring in specific areas of face,this paper explores a face micro-expression recognition method based on optical flow and multi-attention mechanism.Firstly,the facial optical flow features of vertex frames and initial frames are extracted by using Total Variation and L1 norm(TVL1)algorithm,and the optical flow strain is calculated as a supplementary feature;Then,three kinds of opti-cal flow features are used as inputs,and ResBlock,which introduces Global Attention Mechanism(GAM)and Dual Attention mechanism(DA),is applied to extract features and classify micro-expres-sions,so as to reduce facial information dispersion and enlarge global optical flow features.Finally,the classification experiments on CASME and CASME Ⅱ data sets are carried out to verify the feasibility and effectiveness of the proposed method for frontal face classification.

fusion optical flowmicro-expressions recognitionattention mechanismresidual neural net-workdeep learning

唐绍语、魏利胜

展开 >

安徽工程大学电气工程学院,安徽芜湖 241000

融合光流 微表情识别 注意力机制 残差神经网络 深度学习

2024

安徽工程大学学报
安徽工程大学

安徽工程大学学报

影响因子:0.289
ISSN:2095-0977
年,卷(期):2024.39(5)