Improved Mask R-CNN Network Method for PV Panel Defect Detection

Abstract: With the increasing popularity of photovoltaic power generation, the demand for photovoltaic panel defect detection in the industry is also increasing. Deep learning can automatically extract individual photovoltaic panels from images or videos, and perform the defect detection task on it. Aiming at the problem of low detection accuracy of existing deep learning-based photovoltaic panel defect detection methods, an improved Mask R-CNN photovoltaic panel defect detection algorithm is proposed. To improve the training performance, the feature pyramid (FPN) structure is improved, and the cascade network based on attention guidance is adopted to fuse more features and prevent the loss of shallow semantic information to a certain extent. Secondly, Group Normalization (GN) is used to replace Batch Normalization (BN) in the traditional high-performance deep neural network models. The quality of the self-made dataset is improved by Mosaic data enhancement to prevent accuracy loss due to insufficient sample size in the dataset. The effectiveness of the algorithm is verified by the self-made dataset and the public COCO2017 dataset. The improved Mask R-CNN algorithm has a detection accuracy of more than 89% on the self-made photovoltaic panel dataset and 44.6% bounding box average precision (APbbox) and 41.5% mask average precision (APmask) on the COCO2017 dataset, which is 6.4% and 5.8% higher than the original Mask R-CNN algorithm respectively. Finally, to comprehensively analyze the detection performance of the improved algorithm in photovoltaic panel defect detection tasks, the common deep learning-based defect detection algorithms for photovoltaic panel defect detection are summarized. Based on this, a comparison and summary of the improved algorithm in this paper are conducted.