UnetYOLO: A Novel Object Detection Network for Photovoltaics Defect Detection using Electroluminescence Images

The growing global focus on solar energy underscores the need for high efficiency in photovoltaics. Consequently, maintaining efficiency by effective and precise defect inspection in photovoltaic panels has become an important issue. Manual inspection is subject to high labor cost and error susceptibility. Therefore, machine learning (ML) and artificial intelligence (AI) solutions have been explored by researchers to develop vision inspection algorithms for this application. However, prevailing methods exhibit shortcomings in precision and are often restricted to specific defect types. This study introduces an object detection model designed for the precise detection and localization of defects in electroluminescent images. Our approach integrates the state-of-the-art object detection algorithm YOLOV8 with U-Net. The proposed method was tested using the PVEL-AD dataset which features 12 different photovoltaic defects. The new method achieved precision at 91.66%, recall at 90.27%, mAP50 at 93.42%, and mAP50-95 at 68.23%, which significantly improved over the best result previous reported on the same dataset (for example, precisions below 88%, mAP50 values below 87%, and mAP50-95 values below 50%). A significant enhancement over previous methods is the proposed model's accurate detection of defect areas concerning both bounding box size and location.