SURT-DETR: Sea-Sky Line Context-Aware and Uncertainty-Guided Small Object Detection for Unmanned Surface Vehicles

Abstract

Detecting small objects within complex maritime environments presents a persistent challenge for Unmanned Surface Vehicle (USV) perception. Extreme scale variations often render distant targets as indistinct visual features, while background interference from waves and sea-sky boundaries further obscures object delineation. Consequently, state-of-the-art detectors frequently suffer from inaccurate localization and limited generalization capabilities. To address these limitations, we propose SURT-DETR, a robust Transformer-based framework tailored for USV perception. Constructed upon the RT-DETR architecture, SURT-DETR integrates three innovations: (1) the Sea-Sky Line Context-Aware Feature Enhancement Module (SLCAM) to model geometric priors and suppress background noise; (2) the Uncertainty-Guided Small Object Decoder Branch (UGSD) to enhance tiny target detection; and (3) an Uncertainty-Regularized NWD Loss to mitigate scale sensitivity and localization uncertainty. Experimental results on the Pohang Canal Dataset verify the effectiveness of our method. The SURT-DETR achieves an overall mAP of 79.5% and smallobject mAP s of 69.2%. Compared with the RT-DETR baseline, our model obtains clear accuracy gains on overall and small-target detection metrics with a modest parameter increment of 1.2M. Ablation experiments further confirm the complementarity and efficacy of the proposed modules, making the method applicable for practical USV maritime perception tasks.