Improved Yolov5 Photovoltaic Panel Defect Detection Algorithm

Aiming at the problem of low detection accuracy in photovoltaic panel defect detection, which exists in the defects with different shapes, complex and variable background texture, this paper proposes an improved yolov5s network model containing RepNCSPELAN4 module and CARAFE module. Firstly, the RepNCSPELAN4 module is used to replace the C3 layer in the trunk and neck to enrich the feature extraction capability of the network and improve the feature expression capability of the network. Secondly, the content-aware feature reorganisation module CARAFE is introduced to handle multi-scale information and contextual information more effectively and improve the accuracy of prediction. Compared with the original yolov5s network, the improved model shows a 2.1 % increase in mAP with a 19.2% reduction in the number of parameters. The results show that the improved yolov5s detection model can effectively locate and identify EL defect images of PV cells.