基于优化YOLOv7⁃Tiny的表情识别算法  

Expression Recognition Algorithm Based on Optimized YOLOv7-Tiny

作　　者：常星花 王建荣[2] CHANG Xinghua;WANG Jianrong(Department of Mathematics,Taiyuan University,Taiyuan 030032,China;School of Mathematics and Statistics,Shanxi University,Taiyuan 030006,China)

机构地区：[1]太原学院数学系,山西太原030032 [2]山西大学数学与统计学院,山西太原030006

出　　处：《测试技术学报》2025年第2期113-120,共8页Journal of Test and Measurement Technology

基　　金：中国博士后科学基金资助项目(2021M692400);山西省基础研究计划项目(202303021212360);山西省高等学校科技创新计划项目(2023L379);山西省高等学校科技创新平台项目(2022P016)。

摘　　要：表情识别不仅能够提升人机交互体验,推动情感计算的发展,还可以辅助心理健康评估和治疗,提升社会安全和监控效率。为了提高表情识别的检测平均精度,提出了一种基于优化YOLOv7-tiny的表情识别算法。首先,将YOLOv7-tiny中原有的激活函数替换为Mish函数,提高了模型的优化能力;在YOLOv7-tiny的主干网络上再增加CA注意力机制,提高了对目标感兴趣区域的注意,增加了检测的平均精度;最后,将Neck层的上采样部分替换为轻量级上采样算子CARAFE,提高了特征融合能力。实验结果表明,优化后算法的检测效果有了明显的提升,与原始YOLOv7-tiny相比,模型的mAP@0.5提高了1.6百分点,达到88.6%,mAP@0.5:0.95提高了1.3百分点,达到64%;图片检测速度达到每张图片5.0 ms,而且模型保持了轻量化。Expression recognition can not only improve human-computer interaction experience and promote the development of emotional computing,but also assist in mental health assessment and treatment,and improve social security and monitoring efficiency.To improve the detection average accuracy of expression recognition,this paper proposes an expression recognition algorithm based on improved YOLOv7-tiny.Firstly,the original activation function of YOLOv7-tiny is replaced with the Mish func‐tion,which improves the optimization ability of the model.Furthermore,the CA attention mechanism is added to the backbone network of YOLOv7-tiny to improve the attention to the target area of interest and increase the average accuracy of detection.Finally,the upsampling part of the Neck layer is replaced by the lightweight upsampling operator CARAFE to improve the feature fusion capability.Experimental results show that the detection effect of the improved detection algorithm is significantly improved.Compared with the original YOLOv7-tiny,the detection effect of the enhanced detection algorithm is increased by 1.6 percent point to 88.6%,and that of mAP0.5:0.95 is increased by 1.3 percent point to 64%.Image detection speed reaches 5.0 ms per image and the model remains lightweight.

关 键 词：目标检测 表情识别 YOLOv7-tiny 注意力机制 Mish函数 CARAFE算子 

分 类 号：R540.4[医药卫生—心血管疾病] TP183[医药卫生—内科学]