Dust Detection on Solar Panels Using DINOv2 with Multi-Head Self-Attention

Recently, detecting dust and damage on solar panels has become feasible with deep learning models. However, existing solutions often struggle with either model performance or practical implementation challenges. To address these limitations, this paper presents a novel lightweight and robust framework for dust detection on solar panels, leveraging the DINOv2 transformer model enhanced with a multi-head self-attention mechanism. The proposed approach integrates advanced data augmentation and fine-tuning strategies to optimize performance, achieving state-of-the-art results with an accuracy of 94.1% and an F1-score of 93.2%, significantly outperforming traditional CNN-based methodologies. Additionally, the trained model is lightweight, requiring fewer parameters, which reduces memory, computation, and resource demands, leading to lower inference time. The success of this approach highlights the potential of transformer architectures in computer vision tasks, particularly for capturing long-range dependencies and contextual relationships. By striking a balance between model efficiency and real-world application constraints, the proposed solution offers a practical and scalable approach for automated dust detection, enabling proactive maintenance and optimizing the performance of solar power installations.