Exploring Drone-Based Inspection for Detecting Anomalies in PV using Darknet and Image Colorization

This paper explores the use of deep learning (DL) techniques for drone-based anomaly detection in solar photovoltaic (PV) systems, utilizing the Darknet architecture. Addressing challenges like binary classification dominance and inadequate preprocessing, our work employs extensive dataset cleaning and uses colorized images captured by drones to enhance model robustness. By leveraging the fastai library, we streamline the training process for efficient model development. Our approach detects 10 distinct classes of anomalies, providing a more detailed and comprehensive defect detection system. Specifically, we attained a precision of 92.08% and a recall of 92.03%. Additionally, the F1 score, reaching 91.98%, reflected a balanced performance between precision and recall, indicating robustness in classification across various classes. The main contribution of this research lies in improving the accuracy and reliability of solar panel anomaly detection, offering a scalable and effective solution for real-world applications.