Small-sample defect detection of twin-network photovoltaic panels based on multi-scale dynamic fusion

As the global energy transition accelerates, the photovoltaic industry has become a core pillar of renewable energy. However, photovoltaic panels are prone to defects such as microcracks, grid breaks, and hot spots during production and operation, leading to a 20%-50% reduction in power generation efficiency and, in severe cases, even causing safety incidents. Existing detection methods suffer from weak generalization capabilities and high false-negative rates in scenarios involving novel or rare defects with small sample sizes. This paper proposes a small-sample defect detection method based on a multi-scale dynamic fusion twin network, which addresses the issue of traditional twin neural networks inadequately capturing defect features across different scales by introducing a multi-scale feature dynamic weighting module. Experimental results show that the method achieves an mAP of 86.3% on the comprehensive dataset, an 8.7% improvement over the original twin network, with a recall rate of 82.1% and inference speed maintained at 28 FPS, meeting industrial real-time detection requirements. This research provides an efficient solution for defect detection in small-sample scenarios, significantly reducing maintenance costs and enhancing power plant safety.