轻量化M-CNN的太阳电池表面缺陷识别  

LIGHTWEIGHT M-CNN SOLAR CELL SURFACE DEFECTIDENTIFICATION

作　　者：陶志勇[1] 易廷军 林森 杜福廷 Tao Zhiyong;Yi Tingjun;Lin Sen;Du Futing(School of Electronic and Information Engineering,Liaoning Technical University,Huludao 125105,China;School of Automation and Electrical Engineering,Shenyang Ligong University,Shenyang 110159,China)

机构地区：[1]辽宁工程技术大学电子与信息工程学院,葫芦岛125105 [2]沈阳理工大学自动化与电气工程学院,沈阳110159

出　　处：《太阳能学报》2024年第6期341-348,共8页Acta Energiae Solaris Sinica

基　　金：辽宁省教育厅研究生教改项目六位一体、产教融合的专业学位硕士研究生实践创新能力培养模式探索(LNYJG2023117)。

摘　　要：针对太阳电池EL图像,提出一种高识别率的轻量化M-CNN网络模型。首先,该网络模型将拼接的多特征图融合通道注意力机制;然后,引入Ghost卷积层降低模型参数;最后,利用普通卷积层取代最大池化层进行特征空间降维。实验结果表明:在自建裂痕、阴影、微小瑕疵、无缺陷图像数据库共15767片上,M-CNN对粗糙分类检测准确率和瑕疵分类检测准确率分别是99.83%和93.38%,模型参数量是1.29 MB。相较先进的MobileNetV3、DeepVit和MobileVit等网络,M-CNN有缺陷识别率高和模型参数量低的优势。A lightweight micro convolutional neural network(M-CNN)model with high recognition rate is proposed for EL images of solar cells.The network model incorporates a fusion channel attention mechanism to merge multiple feature maps.Introducing Ghost convolutional layers to reduce the model parameters,and using ordinary convolutional layers to replace maximum pooling layers for feature space dimensionality reduction.Experimental results show that on a self-built database of 15767 EL images of cracks,shadows,minor defects,and no defects,M-CNN achieves an accuracy of 99.83%and 93.38%for rough classification and flaw classification detection,respectively,with a model parameter count of 1.29 MB.Notably,compared to advanced networks such as MobileNetV3,DeepVit,and MobileVit,M-CNN has the advantages of superior defect recognition rate and lower model parameter count.

关 键 词：太阳电池 深度学习 卷积神经网络 图像处理 缺陷识别 

分 类 号：TM914.4[电气工程—电力电子与电力传动]