A lightweight network for photovoltaic cell defect detection with feature enhancement and attention mechanism

Detecting defects in photovoltaic (PV) modules is essential due to the challenges posed by multiscale variations, small targets, and complex backgrounds. Existing methods often fail to balance detection accuracy and efficiency. To address these challenges, we proposed Global Perception-YOLO (GP-YOLO), an efficient PV cell defect detector that combined high precision with a lightweight design. First, Reversible Column Networks (RevCol) were employed as the backbone of YOLOv8, enabling efficient propagation of multi-level features. This approach reduces parameter count and GFLOPs, while improving the richness and fidelity of feature representations. Second, the Global Attention Mechanism Transformation Module (GATM) was designed to analyse the interrelationships among input data, identifying and amplifying key features. Finally, the GATM module replaced the ConModule in the PAN path, forming a GP-PAN structure. This captured global context to reduce detail loss during convolutions and upsampling, while mitigating information degradation during feature fusion. On PVEL-AD, GP-YOLO improves mAP@0.5 and mAP@0.5:0.95 by 2.47% and 5.25%, while reducing parameters and GFLOPs by 22.52% and 20.77% compared with the baseline, indicating strong potential for PV cell defect detection. Evaluations on Enterprise PV and NEU further confirm superior cross-domain generalisation and robustness.