Impact of Class Grouping and Relabelling on Defect Detection in Solar Panels using YOLOv9

The use of solar panels is increasing rapidly globally, emphasizing the need of effective monitoring and maintenance strategies. This can be achieved by detecting and resolving the problems as early as possible to maintain high energy generation. Computer vision techniques have proven to be effective in this regard. Most of the conventional defect detection methods are inefficient to deal with multiple classes of defects, thereby delaying the response to address the problems. This study aims to fill in this gap using an efficient computer vision model, you only look once (YOLO) v9, to determine the impact of grouping defect classes on performance of solar panel defect detection models. The original six classes are grouped into three and two classes, respectively, using modified dataset and weighted average modes. Class grouping as well as the grouping methodology significantly impacts the performance of the model. The modified dataset method of class grouping performing better. The mean average precision at threshold 50% (mAP50) for 2-class grouping is 80.2%, which surpasses grouping into three classes by 3.2% and the original six-class setup by 5.6%.