| 孙世政,何玲玲,郑帅,徐向阳,陈仁祥.面向复杂背景环境下垃圾检测的YOLOv8n 轻量化改进[J].电子测量与仪器学报,2025,39(2):136-146 |
| 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 detectionin complex background environments |
| |
| DOI： |
| 中文关键词: 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 |
| 英文关键词:garbage detection lightweight YOLOv8n GhostNet context anchor attention asymptotic feature pyramid network |
| 基金项目: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 |
|
|
| Author | Institution |
| Sun Shizheng | School of Mechatronics and Vehicle Engineering, Chongqing Jiaotong University, Chongqing 400074, China |
| He Lingling | School of Mechatronics and Vehicle Engineering, Chongqing Jiaotong University, Chongqing 400074, China |
| Zheng Shuai | School of Shipping and Naval Architecture, Chongqing Jiaotong University, Chongqing 400074, China |
| Xu Xiangyang | School of Mechatronics and Vehicle Engineering, Chongqing Jiaotong University, Chongqing 400074, China |
| Chen Renxiang | School of Mechatronics and Vehicle Engineering, Chongqing Jiaotong University, Chongqing 400074, China |
|
| 摘要点击次数: 19 |
| 全文下载次数: 58 |
| 中文摘要: |
| 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。 |
| 英文摘要: |
| 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. |
| 查看全文 查看/发表评论 下载PDF阅读器 |
|
|
|
