UDCNet: A U-Net Guided Dual-Branch Cross-Attention Network for SAR Object Detection

Synthetic aperture radar (SAR) object detection often suffers from speckle noise and deformation of diverse target shapes, leading to an inability for the algorithm to effectively distinguish between foreground and background. To address these challenges, we propose a dual-branch framework with a se...

Full description

Saved in:
Bibliographic Details
Main Authors: Siyang Huang, Liushun Hu, Zhangjunjie Cheng, Shaojing Su, Junyu Wei, Xiaozhong Tong, Zongqing Zhao
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Subjects:
Online Access:https://ieeexplore.ieee.org/document/11071349/
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Synthetic aperture radar (SAR) object detection often suffers from speckle noise and deformation of diverse target shapes, leading to an inability for the algorithm to effectively distinguish between foreground and background. To address these challenges, we propose a dual-branch framework with a segmentation surrogate task that leverages foreground and background cues to guide robust multiscale feature extraction. Our method comprises three key components: 1) a U-Net guided dual-branch architecture with cross-attention module, which introduces a U-Net-based segmentation branch to generate “region-based hints” for cross-attention fusion with the detection branch; 2) Shi-Tomasi corner detection with convex polygon optimization for pseudo mask generation, enabling cost-effective segmentation masks from bounding label information, alleviating the need for exhaustive pixel-level annotations; and 3) a two-stage training strategy, which first jointly trains both branches, then freezes segmentation branch to finetune the detection backbone, mitigating gradient conflicts. Extensive experiments on remote sensing dataset, SARDet-100k, validate the detection superiority of our framework [+4.2% mean Average Precision (mAP)], especially in medium (+3.9% mAP) and small target detection (+5.2% mAP). Our findings highlight the importance of task-specific priors in SAR detection pipelines and confirm that combining pseudolabeled segmentation with a targeted training procedure within our proposed framework yields robust, scalable performance for SAR object detection.
ISSN:1939-1404
2151-1535