BEACH-Net: boundary explicit–implicit aware cross-scale hierarchical network for maritime scene semantic segmentation

Abstract In recent years, semantic segmentation of maritime scenes has become increasingly important for tasks such as unmanned navigation, monitoring, and safety perception. However, maritime environments still present several inherent challenges, including highly dynamic water-surface textures, ambiguous and unstable semantic boundaries caused by reflections and specular highlights, and substantial variations in object scale. These factors make it difficult for conventional segmentation networks to maintain stable and accurate predictions in the vicinity of category boundaries. To address these issues, we propose BEACH-Net, a boundary-explicit–implicit-aware cross-scale hierarchical network tailored for maritime scenes. The framework incorporates explicit boundary modeling, structured implicit representation of key boundary points, and cross-scale implicit interaction between boundary and semantic features, effectively improving the model’s structural understanding of actual category boundaries. A comprehensive set of quantitative evaluations and ablation studies is conducted on multiple maritime-scene datasets. BEACH-Net achieves a μ edge of 15.5 px on the MODD2 dataset, outperforming existing methods and fully demonstrating the effectiveness and superiority of the proposed framework.