MFPD-Net: Multiscale Feature Fusion-based Photovoltaic Panel Defect Detection Network

With the rapid development of photovoltaic technology, ensuring the quality of photovoltaic modules has become increasingly crucial. Defect detection in photovoltaic panels is essential within the photovoltaic industry, as it directly influences the performance and reliability of these modules. Nonetheless, challenges such as image noise and the small size of defect targets in surface defect detection frequently result in information loss during feature fusion. Firstly, ADP-GFPN (Adaptive Detail Preservation Generalized Feature Pyramid Network) module is proposed based on GFPN (Generalized Feature Pyramid Network), which further improves feature fusion efficiency and performance. This module boosts the ability of the model to detect small-sized defects while preserving more critical detail information. Secondly, Exponential Moving Average Slide (EMAS) Loss is proposed. The loss not only emphasizes hard samples during training but also balances the influence of historical and current observational data, resulting in a smoother and more stable loss value. This enhancement strengthens the model's robustness and generalization capability. Experimental results indicate that MFPD-Net outperforms the baseline, with a 5.4% increase in Precision and a 2.2% improvement in mAP50.