面向复杂背景环境下垃圾检测的YOLOv8n轻量化改进  

Lightweight improvement of YOLOv8n for garbage detection in complex background environments

作　　者：孙世政[1] 何玲玲 郑帅 徐向阳[1] 陈仁祥 Sun Shizheng;He Lingling;Zheng Shuai;Xu Xiangyang;Chen Renxiang(School of Mechatronics and Vehicle Engineering,Chongqing Jiaotong University,Chongqing 400074,China;School of Shipping and Naval Architecture,Chongqing Jiaotong University,Chongqing 400074,China)

机构地区：[1]重庆交通大学机电与车辆工程学院,重庆400074 [2]重庆交通大学航运与船舶工程学院,重庆400074

出　　处：《电子测量与仪器学报》2025年第2期136-146,共11页Journal of Electronic Measurement and Instrumentation

基　　金：重庆市技术创新与应用发展专项重大项目(CSTB2023TIAD-STX0016)、重庆市自然科学基金创新发展联合基金(CSTB2023NSCQ-LZX0081)项目资助。

摘　　要：垃圾检测与分类对推动绿色经济和实现低碳循环具有重要意义,面向复杂背景环境的垃圾检测模型存在参数量大、计算成本高等问题,限制了模型在资源受限设备上的应用。为解决上述问题,提出一种轻量化的GCAW-YOLOv8n模型,旨在平衡模型轻量化与精度检测。首先,在YOLOv8n骨干网络中引入GhostNet网络中的C3Ghost和GhostConv模块,有效降低模型参数量;其次,添加上下文锚点注意力机制,增强特征提取能力,提升检测精度;然后,在特征融合阶段,构建渐近特征金字塔网络,提升多尺度目标检测能力;接着,采用WIoU v3边界损失函数优化网络边界框回归性能;最后,结合Taco数据集和人工采集数据集进行了模型验证实验。实验结果表明,相比原YOLOv8n模型,改进后的GCAW-YOLOv8n模型在模型参数量Params和计算量FLOPs分别降低了14.3%和33.3%,而精确度和召回率分别提高了4.4%和1.9%,同时mAP@0.5达到了81.3%,提升了0.7%。改进模型更好地平衡了模型轻量化和检测精度,对模型部署与应用至边缘端检测装备具有重要的工程意义。Garbage detection and classification are essential for promoting the green economy and achieving a low-carbon circular economy.However,current models face challenges such as large parameters and high computational costs,limiting their deployment on resource-constrained devices.To address these issues,a lightweight GCAW-YOLOv8n model is proposed that balances model size and detection accuracy.Firstly,the C3Ghost and GhostConv modules from GhostNet are integrated into the YOLOv8n backbone to reduce parameters.Secondly,the context anchor attention is introduced to enhance feature extraction and detection accuracy.Then,the asymptotic feature pyramid network is used to improve multi-scale detection,and the WIoU v3 loss function optimizes bounding box regression.Finally,the improved model is validated using the Taco dataset and a custom dataset.Experimental results show that,compared with the original YOLOv8n model,the GCAW-YOLOv8n model reduces parameters by 14.3%and floating-point operations by 33.3%,while precision and recall increase by 4.4%and 1.9%,respectively.The mAP@0.5 improves to 81.3%,a 0.7%gain.This model achieves a better balance between lightweight design and detection accuracy,making it suitable for deployment in edge devices for garbage detection.

关 键 词：垃圾检测 轻量化YOLOv8n GhostNet 上下文锚点注意力机制 渐近特征金字塔 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TN912[自动化与计算机技术—计算机科学与技术]