The rapid expansion of photovoltaic (PV) installations necessitates efficient, autonomous inspection systems to identify thermal hotspots that account for 15-30% of energy losses and pose safety hazards. Traditional UAV-based infrared detection methods face critical challenges: CNN-based detectors lack sensitivity to subtle thermal gradients, while transformer-based approaches incur excessive computational costs unsuitable for rapid processing on portable ground stations or edge nodes. This paper presents SolarIR-DETR, a lightweight transformer framework specifically designed for near-real-time UAV-Assisted PV hotspot detection. Our approach integrates three key innovations: (1) a C2f-EFFM backbone with multi-scale dynamic convolutions for anisotropic thermal feature extraction, (2) a Hybrid-Scale Feature Modulation (HFM) head employing cross-level channel attention to enhance hotspot sensitivity, and (3) an AIFI-FSABM attention mechanism that separates low- and high-frequency components for efficient anomaly detection. Experimental results demonstrate that SolarIR-DETR achieves 82.82% mAP50 and 78.32% Recall-outperforming RT-DETR-R18 by + 2.13% and + 5.02%-while reducing parameters by 29.5% and computational cost by 21.3%. This compact footprint (44.8 GFLOPs) specifically targets deployment on capable portable ground stations, yielding a projected ~ 30 FPS processing of UAV video downlinks. Furthermore, cross-dataset validation across severe thermal modality shifts (e.g., pseudo-color to grayscale) demonstrates the architecture's consistent structural resilience against feature collapse. By relying on high-frequency spatial morphology rather than cross-channel color semantics, SolarIR-DETR establishes itself as a highly reliable and practical solution for large-scale autonomous PV infrastructure monitoring.
Li et al. (Tue,) studied this question.