Marine Photovoltaic Module Salt Detection via Semantic-Driven Feature Optimization in Mask R-CNN

Offshore floating photovoltaic systems are highly susceptible to salt crystallization on the surfaces of photovoltaic modules, highlighting the need for intelligent inspection and cleaning technologies to improve operational efficiency and overcome the limitations of conventional manual maintenance methods. However, the presence of surface gridlines on the photovoltaic modules introduces significant visual interference, which complicates the accurate identification of salt deposition regions. To address this challenge, a semantic information-guided detection framework is proposed to enable precise segmentation of salt-affected areas. The key innovation lies in the effective classification of gridlines as background features by extracting semantic priors through low-level thresholding, which are then fused with the original red-green-blue image to construct a four-channel input. This fusion enhances the model’s ability to extract and discriminate features related to salt crystallization. Experimental results demonstrate that the proposed method achieves a 4.6% improvement in segmentation accuracy and a 3.7% increase in recognition accuracy compared to conventional models, based on evaluation metrics such as mean average precision and F1-score. The proposed framework offers a robust technical foundation for developing intelligent maintenance systems tailored to offshore floating photovoltaic applications.