GFANet: An Efficient and Accurate Water Segmentation Network

Accurate water body detection is essential for autonomous navigation and 1 operational planning of Unmanned Surface Vehicles (USVs). To address model adaptability 2 to ambiguous boundaries caused by diverse scenarios and climatic conditions, this study 3 proposes GFANet (Global-Local Feature Attention Network) for real-time water surface 4 semantic segmentation of camera-captured images. First, a Global-Local Feature (GLF) 5 Extraction module is proposed, integrating a self-attention-based local feature extractor 6 and a multi-scale global feature extractor for parallel feature learning, thereby enhancing 7 hierarchical feature representation. Second, a Gated Attention (GA) Module is designed 8 with a dual-branch gating mechanism to implement noise suppression and efficient low- 9 level feature utilization. The method was validated on three publicly available datasets in 10 relevant domains.Experimental results on the Riwa dataset show GFANet achieves state-of- 11 the-art segmentation performance (4.41M parameters, 7.15 GFLOPs) with mIoU 82.29% 12 and mPA 89.49%. Comparable performance metrics were obtained on the USVInland and 13 WaterSeg datasets.Additionally, GFANet achieves 154.98 FPS processing speed, meeting 14 real-time segmentation requirements. Experimental results verify that GFANet achieves an 15 optimal balance between high segmentation accuracy and real-time processing efficiency.