EdgeFlowNet: A Lightweight, Dynamic Edge-Aware Network for Precise Water Surface Boundary Detection

Abstract Water surface edge detection demands lightweight solutions for real-time marine navigation, maintaining high precision under diverse environmental conditions. This paper introduces EdgeFlowNet, a lightweight and environment-robust edge-aware network designed specifically for water surface boundary detection. EdgeFlowNet integrates a Dynamic Importance Weighting (DIW) module for adaptive fusion of edge and semantic features, a hybrid dual-branch architecture combining classical edge operators with deep learning, and binary classification optimization for extreme parameter efficiency. Evaluated on the WaterScenes dataset across 192 environmental conditions, EdgeFlowNet-Efficient (74K parameters) achieves 97.83% mIoU at 76.8 FPS, outperforming PiDiNet by 0.67% mIoU with 52.6% fewer parameters and 82% faster processing. Our approach demonstrates strong robustness and an improved accuracy-efficiency trade-off, facilitating deployment in resource-constrained autonomous marine systems. The source code of our EdgeFlowNet model is available at https://github.com/Wuuu10/EdgeFlowNet.