U-DualMAR: Dual Differentiation Network for Maritime Small Target Segmentation

Small obstacles are critical for target perception in Unmanned Surface Vehicles, but their segmentation accuracy directly impacts navigation safety. To address this issue, this paper proposes a collaborative optimization framework based on Spatial Differentiation and Class Differentiation. Spatial Differentiation enhances the feature representation capability of small targets in key regions through spatial weight encoding, while Class Differentiation dynamically adjusts inter-class interference constraints via a class-sensitive loss function, further improving the recognition accuracy of small targets. To validate the effectiveness of the proposed method, comparative experiments were conducted using a six-tier small target classification system and the authoritative maritime segmentation benchmark MODS (Maritime Object Detection and Segmentation). The results show that our method achieves a TPR (True Positive Rate) value of 71.2% in danger zones small obstacle detection, outperforming the state-of-the-art model WaSR by 7.1 percentage points, while reducing the false detection rate by 19.1%.