复杂背景下的多晶硅太阳电池缺陷检测  

DETECTION OF DEFECTS IN POLYSILICON SOLAR CELLS IN COMPLEX BACKGROUNDS

作　　者：廖力达[1] 罗晓 黄斌 刘亮[1] 冯飞[3] 陈为强 Liao Lida;Luo Xiao;Huang Bin;Liu Liang;Feng Fei;Chen Weiqiang(School of Energy and Power Engineering,Changsha University of Science and Technology,Changsha 410114,China;Engineering College,University of South Australia,Adelaide 5000,Australia;Tongdao Zimu Wind Energy Development Co.,Ltd.,Huaihua 412500,China)

机构地区：[1]长沙理工大学能源与动力工程学院,长沙410114 [2]南澳大学工学院,阿德莱德5000 [3]通道子舞风能开发有限公司,怀化418500

出　　处：《太阳能学报》2024年第9期295-303,共9页Acta Energiae Solaris Sinica

基　　金：湖南省自然科学基金(2021JJ30717)。

摘　　要：提出并设计名为OD-YOLO缺陷检测模型来改善多晶硅太阳电池电致发光成像中复杂背景干扰缺陷检测效果的问题。使用二次卷积模块(TwiceConv-OD)过滤掉复杂晶粒背景干扰,增强模型对缺陷本身的关注力;提出anchor-plus1分配策略来增加模型在面对复杂背景时获取更多的缺陷正样本数量,提升模型的准确率与召回率,减少漏检误检;使用K-means++算法初始化锚框尺寸,聚类后的锚框更能代表检测样本中所有缺陷的几何形貌,能更好适应多晶硅太阳电池的缺陷尺度差异。经公开缺陷检测数据集PVEL-AD-2021的实验验证:OD-YOLO模型的均值平均精度(mAP)达到89.4%,相比YOLOV5s缺陷检测模型提升3%,准确率提升4.8%,召回率提升1.9%,参数减少4.5%,检测速度为104帧/s。A defect detection model named OD-YOLO is proposed and designed to improve the problem of complex background interfering with the defect detection effect in electroluminescence imaging of polycrystalline silicon solar cells.We use the secondary convolution module(TwiceConv-OD)to filter out the complex grain background interference and enhance the model's focus on the defects themselves;we propose the anchor-plus1 allocation strategy to increase the number of defect positive samples obtained by the model in the face of the complex background,which improves the model's accuracy and recall and reduces the omission of misdetections;and we use the K-means++algorithm to initialize the anchor frames and cluster the anchor frames to be more representative of all the detected samples.The size of the anchor frame is initialized using the K-means++algorithm,and the clustered anchor frame is more representative of the geometrical shape of all defects in the detected samples,which is better adapted to the differences in defect scales of polysilicon solar cells.Experimentally verified by the publicly available defect detection dataset PVEL-AD-2021:the mean average precision(mAP)of the OD-YOLO model reaches 89.4%,which is improved by 3%compared to the YOLOV5s defect detection model,the accuracy is improved by 4.8%,the recall rate is improved by 1.9%,the parameters are reduced by 4.5%,and the speed is 104 frames per second.

关 键 词：太阳电池 计算机视觉 表面缺陷 卷积神经网络 

分 类 号：TN383.1[电子电信—物理电子学]