Abstract: There are many kinds of defects in the surface of solar cells,with large differences in size and shape,and high labeling cost,which brings difficulties to the task of defect segmentation.In order to improve the segmentation performance of solar cell surface defects,this paper proposes a novel weakly supervised deep learning network with multi-spectral multi-scale class activation mapping,called MMCAMNet.Firstly,a multi-spectral deep learning network structure is proposed,which realizes the multi-channel fusion of spectral information,and enhances the feature extraction capability of the MMCAM-Net.Secondly,a multi-scale network structure is designed to achieve high-level and low-level defect feature fusion and improve MMCAM-Net＇s defect global and local information extraction capabilities.Finally,the image-level label data set is used to train MMCAM-Net,which realizes weakly supervised segmentation of solar cell surface defects.Experimental results show that the IoU of the network model has increased by 15%-20%,and satisfactory results are gained.

Key words: class activation mapping, multi-scale, multispectral, defect detection, solar cells

摘要： 光伏电池外观缺陷种类多、大小和形状差异较大,标注成本高,这为缺陷分割任务带来困难。为了提高光伏电池外观缺陷的分割性能,文中提出一种基于多光谱多尺度类激活映射的弱监督深度学习网络模型(MMCAM-Net)。首先,提出了多光谱深度学习网络结构,实现了光谱信息的多通道融合,增强了MMCAM-Net网络的精细化特征提取能力;其次,设计了多尺度网络结构,实现缺陷信息的高级特征和低级特征融合,增强了MMCAM-Net的缺陷全局与局部信息提取能力;最后,使用图像级标签的数据集来训练MMCAM-Net,实现了光伏电池表面缺陷的弱监督分割。实验结果表明,该网络模型的缺陷分割平均IoU提高了15%-20%,取得了较好的效果。

关键词:  , 类激活映射, 多尺度, 多光谱, 缺陷分割, 光伏电池