基于YOLOv5的微表情识别技术研究
Research on micro-expression recognition technology based on YOLOv5
张山山 1胡志慧1
作者信息
- 1. 湖北汽车工业学院电气与信息工程学院,442000,湖北省十堰市
- 折叠
摘要
针对微表情数据集较少、泛化能力较弱和微表情识别系统对设备要求较高等问题进行研究.对3个主流数据集进行融合筛选,重构了一个多样本表情数据集;采用在微表情数据集中融入部分宏表情的方法解决了模型泛化能力不强的问题;基于YOLOv5模型进行训练和测试,获得了相对便于移植的微表情识别系统.结果表明:经过改进数据集训练后的模型泛化能力有所提升;训练好的模型在自己构建的微表情数据集(Z-MES)上的识别准确率达到79.4%,优于改进前的微表情识别效果.
Abstract
To solve the problems of small data sets,weak generalization ability and higher requirements for equipment and GPU of micro-expression recognition system,three mainstream data sets are fused and screened,and a multi-source expression data set was reconstructed.The problem of poor generalization a-bility was solved by incorporating some macro-expressions into the micro-expression data set.By training and testing based on YOLOv5 model,a relatively portable micro-expression recognition system is ob-tained.The results show that the model generalization ability is improved after the improved data set train-ing.The recognition accuracy of the trained model on the Z-MES micro-expression data set is 79.4%,which is better than before the improvement.
关键词
微表情识别/YOLOv5/数据集Key words
micro-expression recognition/YOLOv5/dataset引用本文复制引用
基金项目
湖北省教育科学规划重点课题(2020GA045)
十堰市科技计划项目(19Y134)
出版年
2024
国家科技期刊平台
NSTL
万方数据