A novel lightweight detector for precise measurement of defects in photovoltaic panels in complex industrial scenarios

Defect detection is crucial for maintaining the functionality and productivity of photovoltaic (PV) systems. However, existing object detection algorithms face significant challenges in PV panel defect detection, including interference from complex backgrounds, limited small target detection capabilities, and difficulties in handling multi-scale defects. To address these issues, a lightweight PV panel defect detection model based on real-time detection transformer is proposed. Firstly, to tackle the problem of defect feature loss due to complex background noise, a lightweight backbone network, utilising the partial convolution (Pconv) operator, is designed to improve the model’s feature extraction capability in complex heterogeneous environments while keeping the computational cost low. Next, to effectively utilise the boundary texture information of small target defects, a local global collaborative attention mechanism is proposed by integrating both local and global feature information for identifying small defect targets. Finally, a multipath hybrid enhancement module is developed by aggregating features from multiple convolutional paths for recognising defects of various shapes. The results show that the improved model achieves an AP of 64.1% and an APS of 33.8% on the public PVEL-AD dataset, with improvements of 2.0% and 3.5%, respectively, compared to the baseline model.