改进YOLOv8n的轻量化光伏电池EL缺陷检测  

Improved YOLOv8n for EL Defect Detection in Lightweight Photovoltaic Cells

作　　者：刘峪 朱文忠 何鑫 李杰 LIU Yu;ZHU Wenzhong;HE Xin;LI Jie(Artificial Intelligence Key Laboratory,School of Automation and Information Engineering,Sichuan University of Science&Engineering,Yibin 643000,China;School of Computer Science and Engineering,Sichuan University of Science&Engineering,Yibin 643000,China)

机构地区：[1]四川轻化工大学自动化与信息工程学院人工智能重点实验室,宜宾643000 [2]四川轻化工大学计算机科学与工程学院,宜宾643000

出　　处：《组合机床与自动化加工技术》2025年第3期87-93,共7页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：四川省科技研发重点项目(2023YFS0371);四川省科技创新(苗子工程)培育项目(2022049);企业信息化与物联网测控技术四川省高校重点实验室基金项目(2022WYY03)。

摘　　要：针对电致发光(electroluminescence,EL)成像背景下光伏电池片小目标缺陷识别难和检测模型体积大的问题,提出一种HPE-YOLOv8n的轻量化检测模型。首先在骨干层引入HGNetv2网络,提升模型对图像中EL缺陷的特征提取能力,减少参数量和计算量;其次在颈部层提出一种改进特征融合网络(P2 CSP efficient dual layer aggregation networks,P2-CEDN),有效地整合和聚合不同层级的特征信息,降低参数量,增强模型对小目标的检测能力;最后在输出层利用参数共享的思想改进解耦头EfficientHead,降低计算量;同时引入损失函数WIoU,提高检测模型的整体性能。实验结果表明,提出的HPE-YOLOv8n检测模型与原YOLOv8n模型相比,参数量和计算量分别降低了64.5%和16%,mAP@0.5和mAP@0.5:0.95指标分别提升3%和3.1%,每秒帧数可达96.15。为实际工业出厂前的光伏电池片缺陷检测提供技术参考。Aiming at the problems of difficult identification of small target defects and large size of the detection model for photovoltaic cell wafers under the background of electroluminescence(EL)imaging,a lightweight detection model of HPE-YOLOv8n is proposed.Firstly,an HGNetv2 network is introduced in the back-bone layer to improve the feature extraction ability of the model in complex scenes and reduce the number of parameters and computation.Secondly,an improved feature fusion network is proposed in the neck layer(P2 CSP efficient dual layer aggregation networks,P2-CEDN)is proposed in the neck layer,which effectively integrates and aggregates feature information from different layers,reduces the number of parameters,and enhances the model′s ability to detect small targets.Finally,the EfficientHead is improved in the output layer by utilizing the idea of parameter sharing to reduce the computational amount;meanwhile,the loss function WIoU is introduced to improve the overall performance of the detection model.The experimental results show that the proposed HPE-YOLOv8n detection model reduces the number of parameters and computation by 64.5%and 16%,respectively,compared with the original YOLOv8n model,meanwhile the mAP@0.5 and mAP@0.5:0.95 metrics are improved by 3%and 3.1%,respectively,and the number of frames per second can be up to 96.15.Provide technical reference for the defect detection of PV cells before leaving the factory in real industry.

关 键 词：YOLOv8n 轻量化 光伏电池片 小目标 缺陷检测 

分 类 号：TH165[机械工程—机械制造及自动化] TG659[金属学及工艺—金属切削加工及机床]