多尺度的轻量化光伏电池EL缺陷检测  

Multi-Scale Lightweight EL Defect Detection for Photovoltaic Cells

作　　者：孙铁强 赵成伟 宋超 肖鹏程 SUN Tieqiang;ZHAO Chengwei;SONG Chao;XIAO Pengcheng(College of Artificial Intelligence,North China University of Science and Technology,Tangshan 063210,China;Hebei Provincial Key Laboratory of Industrial Intelligent Perception,North China University of Science and Technology,Tangshan 063210,China;College of Metallurgy and Energy,North China University of Science and Technology,Tangshan 063210,China)

机构地区：[1]华北理工大学人工智能学院,唐山063210 [2]华北理工大学河北工业智能感知重点实验室,唐山063210 [3]华北理工大学冶金与能源学院,唐山063210

出　　处：《组合机床与自动化加工技术》2025年第3期177-181,186,共6页Modular Machine Tool & Automatic Manufacturing Technique

基　　金：河北省“三三三人才工程”资助项目(A202102002);河北省创新能力提升计划项目(23561007D);2023年唐山市重点研发项目(23140204A)。

摘　　要：针对光伏电池EL缺陷图像特点,提出了一种多尺度的轻量化光伏电池EL缺陷检测框架YOLOV8-GF。首先,采用轻量级特征提取网络HGNetV2作为骨干,并引入RepConv对其改进,增强多尺度特征融合能力;其次,提出自适应权重下采样(AFDS)方法,增强网络对形态多变目标的处理能力;同时,改进头部采用共享参数结构并引入新型多尺度特征提取卷积(MLSA),减少参数冗余并提高特征提取能力;最后,使用Inner-IOU思想改进Shape-IOU提出Inner-Shape-IOU作为损失函数,提高网络鲁棒性。实验结果表明:在光伏电池EL缺陷数据集得到模型的参数量与计算量为2.41 M和5.8 GFLOPs,仅为YOLOv8n的77.4%和70.7%,mAP达到91.62%,较YOLOv8n提升1.4%。Based on the characteristics of photovoltaic cell EL defect images,a multi-scale lightweight detection framework named YOLOV8-GF is proposed for photovoltaic cell EL defect detection.Firstly,a lightweight feature extraction network HGNetV2 is employed as the backbone and improved by introducing RepConv to enhance multi-scale feature fusion capabilities.Secondly,an adaptive feature downsampling(AFDS)method is proposed to enhance the network′s ability to handle targets with varying morphologies.Simultaneously,the improved head adopts a shared parameter structure and introduces a novel multi-scale feature extraction convolution(MLSA),reducing parameter redundancy and improving feature extraction capabilities.Finally,the Inner-IOU concept is utilized to improve Shape-IOU,resulting in Inner-Shape-IOU as the loss function,which enhances the network′s robustness.Experimental results demonstrate that the proposed model achieves a parameter count and computational cost of 2.41 M and 5.8 GFLOPs,respectively,which are only 77.4%and 70.7%of those of YOLOv8n.The mAP reaches 91.62%,representing a 1.4%improvement compared to YOLOv8n.

关 键 词：缺陷检测 YOLOv8 轻量化 光伏电池EL 多尺度 

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]