基于YOLOv5的太阳能电池片EL图像缺陷检测算法  

Solar Cells EL Image Defect Detection and Classification Based on Improved YOLOv5

作　　者：张德钰 ZHANG Deyu(College of Advanced Manufacturing,Fuzhou University,Quanzhou 362251,China)

机构地区：[1]福州大学先进制造学院,福建泉州362251

出　　处：《电视技术》2024年第1期42-47,55,共7页Video Engineering

摘　　要：EL图像可用于精准检测太阳电池及光伏组件的内在缺陷,太阳能电池片EL图像缺陷会受到复杂背景的干扰使其变得难以识别。为此,提出一种改进的YOLOv5深度学习模型,用于提高缺陷检测的可靠性和准确性。该模型采用CutMix数据增强对EL图像的处理,在Backbone中使用改进混合域注意力替换原有网络的内容安全策略(Content Security Policy,CSP)部分,提高模型的特征提取能力。同时,引入特征融合(Feature Fusion Module,FFM)模块有效融合不同维度的特征,达到背景抑制的效果。基于PVEL-AD公开数据集的实验结果表明,相较于原来的YOLOv5s模型,改进后的模型参数量从7.02×10^(6)下降到6.79×10^(6),且mAP50准确率从71.11%提升到87.74%。EL images can be used to accurately detect the inherent defects of solar cells and photovoltaic modules.The defects of EL images of solar cells can be interfered with by complex backgrounds,making them difficult to identify.Therefore,an improved YOLOv5 deep learning model is proposed to improve the reliability anda ccuracy of defect detection.The model uses CutMix data to enahnce EL image processing,and replaces the Content Security Policy(CSP)part of the original network with improved mixed domain atteniotn in Backbone to improve the feature extraction capability of the model.At the same time,the Feature Fusion Module(FFM)is introduced to effectively fuse features of different dimensions to achieve the effect of background suppression.The experimental results based on the public data set of PVEL-AD show that compared with the origina YlOLOv5s model,the number of parameters of the improved modeils reduced from 7.02×10^(6) to 6.79×10^(6),and the accuracy of mAP50 is increased from 71.11%to 87.74%.

关 键 词：太阳能电池片 图像缺陷 目标检测 深度学习 YOLOv5 

分 类 号：TM914.4[电气工程—电力电子与电力传动]