改进YOLOv8的轻量化猪脸关键点检测  

Improved Yolov8 for Lightweight Pen Face Keypoints Detection

作　　者：周玉蝶 张春燕[1,2] 乔印虎 陈泽伟[1] ZHOU Yudie;ZHANG Chunyan;QIAO Yinhu;CHEN Zewei(Anhui Science and Technology University,Anhui Chuzhou,233100,China;Anhui Yunlong Grain Machinery Co.,LTD.,Anhui Fuyang,236500,China)

机构地区：[1]安徽科技学院,安徽滁州233100 [2]安徽云龙粮机有限公司,安徽阜阳236500

出　　处：《杭州电子科技大学学报（自然科学版）》2024年第2期51-64,102,共15页Journal of Hangzhou Dianzi University：Natural Sciences

基　　金：2022年滁州市科技计划项目(2022ZN014,2022ZN016);2021年度教育部中国高校产学研创新基金(2021BCE01004);安徽省教育厅自然科学重大项目:2022AH040238。安徽省粮食机械乡村振兴协同技术服务中心项目(GXXT-2022-077);安徽省自然科学基金面上项目(2308085ME178)。

摘　　要：为了解决YOLOv8-pose关键点检测网络结构复杂、计算量多、参数多等问题,提出一种基于YOLOv8-pose改进的轻量化模型SGS-YOLOv8-pose。在原有的YOLOv8-pose网络上替换骨干网络为ShuffleNetv2,将GhostConv模块整合到Neck中,降低网络计算量,在网络层中加入SE注意力机制,提高模型计算精度;同时采用了WIOU损失函数,使模型的收敛效果更好;改进后的模型与原始模型相比,目标框检测精度提升0.30%,浮点计算量减少33.70%,参数量减少38.55%,模型大小降低了37.18%。试验结果表明:在不同轻量化网络与特征融合方式下,在不同的注意力机制下作对比下,改进后的SGS-YOLOv8-pose模型,能够在保持较高检测精度下使模型浮点计算量和参数量大幅降低,便于后期模型的部署。In order to solve the problems of complex structure,high computational complexity,and multiple parameters in the YOLOv8-pose keypoints detection network,an improved lightweight model SGS-YOLOv8-pose based on YOLOv8-pose is proposed.It replaces the backbone network with ShuffleNetv2 on the original YOLOv8-pose network,integrates the GhostConv module into the neck to reduce network computation,and addes SEAttention mechanism in the network layer to improve model computation accuracy,and simultaneously uses the WIOU loss function to make the model more effective.Compared with the original model,the improved model achieved a 0.30%improvement in target box detection accuracy,a 33.70%reduction in floating-point computation,and a 38.55%reduction in parameter size,and model size reduced by 37.18%.The experimental results show that under different lightweight networks and feature fusion methods,and under different attention mechanisms,the improved SGS-YOLOv8 pose model and the improved SGS-YOLOv8 pose model can significantly reduce the floating-point computation and parameter count of the model while maintaining high detection accuracy,making it easier to deploy the model in the later stage.

关 键 词：YOLOv8 ShuffleNetv2 注意力机制 关键点检测 

分 类 号：S8-1[农业科学—畜牧兽医] TP391[自动化与计算机技术—计算机应用技术]