GF-RTDETR: lightweight transformer-based method for surface defect detection in photovoltaic panels

Abstract Photovoltaic (PV) panels convert sunlight into direct current electricity, but their exposure to outdoor conditions makes them susceptible to defects like cracks and hotspots. Current defect detection methods face challenges including deployment difficulties and low detection accuracy. To address these limitations, we propose a lightweight algorithm called GF-RTDETR. Our method proposes the Cross-scale Enhanced Feature Pyramid Fusion Network (CEFPFN), which leverages multiple feature extraction branches to improve detection accuracy across scales. The Dual Convolution Block (DualConv_Block) is integrated into the backbone network, simultaneously optimizing feature representation and computational efficiency. We design AIFI-HiLo by combining High-Frequency and Low-Frequency (HiLo) attention with Attention-based Intra-scale Feature Interaction (AIFI), thereby enhancing detection of small and overlapping defects. Additionally, we introduce Inner Generalized Intersection over Union (Inner-GIoU), a novel loss function that incorporates adaptive scaling mechanisms to accelerate convergence. Experimental results demonstrate that GF-RTDETR significantly outperforms RT-DETR, achieving a 20.3% reduction in parameters, 14.4% decrease in GFLOPs, 20.8% reduction in model size, while improving mAP50 by 3.2% (reaching 97.4%), mAP50:95 by 2.6%, and FPS by 21%. Validation on the PVEL-AD dataset confirms its generalization capability, achieving an mAP50 of 82.5%, which establishes GF-RTDETR as a cost-effective and efficient solution for industrial defect inspection.