SEAD-Net:Complex underwater image segmentation via semantic-enhanced and detail-aware collaboration

Underwater image segmentation is crucial for marine resource exploration and ecological monitoring, yet remains challenging due to severe image degradation, target camouflage, and limited large-scale datasets. To tackle these issues, we propose SEAD-Net, a dual-stream feature-complementary framework that fuses fine-grained structural cues from a general segmentation foundation model with high-level semantics and global context from a self-supervised semantic representation foundation model, improving both boundary precision and target coherence. To enable stable fusion and mitigate spatial–semantic misalignment, we introduce an equal-scale cross-modulation fusion layer that performs lightweight cross-branch modulation and conflict suppression in a hierarchical manner (shallow layers for spatial consistency, deep layers for semantic consistency). Furthermore, to handle large appearance variations and strong background interference in underwater objects, we design a U-shaped fusion decoder with discrepancy-aware fusion and hybrid attention, combined with multi-scale deep supervision to progressively recover fine details and enhance structural integrity. Experiments demonstrate that SEAD-Net achieves 76.5% and 84.0% m I o U on RMAS and MAS3K, respectively, and reaches an S α of 93.62% on USOD10K, outperforming the best baseline by 1.47%, demonstrating strong effectiveness and generalization in complex underwater scenarios.