Detection Method of Photovoltaic Panel Defect Based on Improved Mask R-CNN

Abstract

To solve the low efficiency and precision of uncrewed inspection in photovoltaic power stations, a segmentation method of improving the defective photovoltaic panels based on improved Mask R-CNN is proposed. The atrous spatial pyramid pooling and spatial attention mechanism were introduced into the extraction network to improve detection accuracy. Uncrewed aerial vehicle infrared video of the panels is used to input the network model for defect detection. As a result, the automatic annotation of the defect position is achieved, significantly improving the efficiency and precision of uncrewed inspection. The precision, recall and FPS are used as performance metrics to evaluate U-Net, PSPNet, Mask R-CNN and the algorithm in this paper. Experiments show that the detection precision of the four models are 77.3%, 82.2%, 84.0% and 89.8% respectively, the recall are 79.4%, 79.0%, 81.6% and 84.4% respectively, and the FPS are 12.5, 9.6, 8.6 and 8.2 respectively. Although the FPS of this algorithm is slightly reduced, the precision and recall have been greatly improved, and can be applied to industry.