DSAFG-UNet: Dual-Branch Semantic Adaptive Fusion Gate UNet for Infrared Sea Surface Image Segmentation

Infrared sea surface image segmentation is a fundamental task in marine perception, but its performance degrades significantly due to faint wave textures, blurred sea–sky boundaries, and low contrast. To address these issues, this paper proposes a Dual-branch Semantic Adaptive Fusion Gate UNet (DSAFG-UNet). The method replaces the static concatenation in the skip connections of standard UNet with a Dual-branch Semantic Adaptive Fusion Gate (DSAF-Gate) module, which adaptively fuses global semantic and local edge information to dynamically filter skip-connection features. Meanwhile, a multi-scale edge loss and a multi-scale context module are introduced to complement the gating mechanism from the perspectives of boundary supervision and global perception, respectively. Experiments show that DSAFG-UNet achieves an mIoU of 91.96% and a sea-class F1-Score of 99.40%, outperforming gating-based models, multi-scale context models, and attention-based models on multiple accuracy metrics, with good stability and certain generalization capability on a limited external sample, although its computational efficiency still needs improvement. In summary, the proposed model achieves strong overall performance in infrared sea surface segmentation and provides an effective solution for sea-region segmentation.