传统光伏面板缺陷检测任务以人工目视方法为主，存在效率低、精度差、成本高等问题。提出基于深度学习的融合光伏面板可见光图像与红外图像的缺陷检测网络，即多源图像融合网络(Multi-source Fusion Network, MF-Net），实现光伏面板的缺陷检测。 MF-Net以 YOLOv3 tiny为主干结构，并针对光伏面板缺陷特征进行网络结构改进。其中包括：在特征提取模块中增加网络深度并融入密集块结构，使得 MF-Net能够融合更多高层语义信息的同时增强特征的选择；将双尺度检测增加为三尺度检测，以提高网络对不同尺寸缺陷的适用度。此外，提出自适应融合模块，使特征图融合过程中可以根据像素邻域信息自适应分配融合系数。实验结果表明，相比基于 YOLOv3 tiny的融合网络，改进后的融合检测网络 mAP提高 7.41%；自适应融合模块使 mAP进一步提升 2.14%，且自适应融合模块能够有效提高特征的显著性；在与单一图像（仅有可见光图像或红外图像）的检测网络及其他融合图像检测网络的对比实验中，所提出的网络 F1 score最高（F1＝0.86）。

This study proposed a multisource fusion network (MF-Net) that combines visible and infrared images for the inspection of a photovoltaic panel to achieve photovoltaic panel defect detection, defect clas-sification, and localization. The limitations of the traditional methods include low efficiency, low accuracy, and high cost. In this study, a defect detection network was designed based on the backbone of YOLOv3-tiny. Deep layers are added to the network, constituting a dense block structure to augment semantic information on fused feature maps. The detection scale of the network was extended to improve its applicability at different scales. In addition, an adaptive weight fusion strategy was proposed to achieve feature map fusion, where the fusion coefficients can be allocated according to the pixel neighborhood information. Compared with the backbone, the results show that the mAP of our network improved by 7.41%. The performance improves (by approximately 2.14% mAP) when the weighted fusion strategy is replaced with ours, and the significance of the features can be effectively improved. Relative to other networks, the proposed network that takes the fused images as the input has the highest performance in terms of the F1 score (F1=0.86).