A novel cost-function for transformerbased YOLO algorithm to detect photovoltaic panel defects

Abstract

Solar panel defects can lead to substantial efficiency loss and increased maintenance expenses. Conventional defect detection methods are often slow and ineffective. Thisstudy revisits the You Only Look Once (YOLO) algorithm and its variations, assessing their efficacy in identifying defects in thermal images of solar panels. Subsequently, we introduce a novel YOLO algorithm, termed YOLOS-PV, built uponthe transformer-based YOLOS algorithm. The proposed algorithm introduces newloss function weights to prioritize localized objects and visualize the attention mapof each transformer head within the YOLOS algorithm. In the experiments, theYOLOS-PV achieves a [email protected]:0.95 score of 0.894, surpassing the efficiency ofother YOLO variants. Code implementation can be found here: tella26/YOLOS-PV (github.com).

Keywords

• solar panels
• defect detection
• deep learning
• yolo models
• cost function