W-ESDETR: An Efficient Scale-Aware Transformer Detection Network for Inland Waterway Surveillance

This is the authors' abstract. We don't add key points for this paper.

Visual ship surveillance in inland waterways remains challenging because complex shoreline backgrounds and adverse weather make it difficult to balance real-time speed with precise localization. To address this problem, we propose W-ESDETR (Wavelet-Efficient Scale-aware DETR), an efficient Transformer-based detection network built on RT-DETRv2. An Enhanced Feature-correlation with Large Selective-kernel (EFLS) module is employed in the hybrid encoder to reduce background interference, while a Wavelet-enhanced Multi-scale Dense Feature Fusion Network (W-MDFFN) with an Affinity Parametric Wavelet Downsampling (APWD) operator preserves high-frequency texture details for distinguishing visually similar ship classes. In addition, the EB-Inner-MPDIoU loss, adapted from the YOLO-TSL formulation, is adopted to improve bounding-box localization under blurred vessel contours. Experimental results show that W-ESDETR achieves 99.18% mAP@50 on SeaShips and 95.2% mAP@50 on our self-constructed inland dataset at 173.4 FPS, with 8.8% fewer parameters than RT-DETRv2-R18. It also increases mAP@50:95 from 78.2% to 79.6% on the self-constructed dataset. These results indicate that W-ESDETR improves detection accuracy over the baseline while maintaining comparable real-time inference efficiency, providing a high-precision, low-latency solution for intelligent inland waterway surveillance.