基于改进YOLOv5s的微表情识别研究  

Research on Micro-Expression Recognition Based on Improved YOLOv5s

作　　者：宋明芮 唐海 徐洪胜[1] SONG Mingrui;TANG Hai;XU Hongsheng(School of Electrical and Information Engineering,Hubei University of Automotive Technology,Shiyan Hubei 442002,China)

机构地区：[1]湖北汽车工业学院电气与信息工程学院,湖北十堰442002

出　　处：《重庆科技学院学报（自然科学版）》2023年第6期65-71,共7页Journal of Chongqing University of Science and Technology：Natural Sciences Edition

基　　金：湖北省教育科学规划2020年度重点课题“大数据驱动下的个性化在线学习精准推送服务体系研究”(2020GA045)。

摘　　要：传统的表情识别方法虽然具有较高的识别精度,但数据处理的速度较慢。针对传统表情识别方法无法及时捕捉微表情变化的问题,提出了一种基于改进YOLOv5s的微表情识别方法。在YOLOv5s的主干网络结构中添加具有无参注意力机制的SimAM模块,在不增加原模型参数的基础上分配三维注意力权重,以提高模型的特征提取能力。实验结果表明,改进YOLOv5s模型的平均精度均值比原YOLOv5s模型更高,能够实时捕捉微表情的变化,满足实时任务的需求。Although the traditional expression recognition method has high recognition accuracy,the speed of data processing is slow.Aiming at the problem that traditional expression recognition methods cannot capture micro-expression changes in time,a micro-expression recognition method based on improved YOLOv5s is proposed.The SimAM module with parameter-free attention mechanism is added into the backbone network structure of YOLOv5s,and the three-dimensional attention weight is assigned without increasing the parameters of the original model to improve the feature extraction ability of the model.Experimental results show that the average accuracy of the improved YOLOv5s model is higher than that of the original YOLOv5s model,which can capture the changes of micro-expressions in real time to meet the needs of real-time tasks.

关 键 词：YOLOv5s 微表情识别 目标检测 深度学习 SimAM 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]