摘要：

基于卷积神经网络的目标检测算法已经在合成孔径雷达(SAR)图像的海上船舶检测中取得了显著进展, 但在近岸场景中, 海岸线、建筑物和其他背景干扰物的强散射信号仍然对船舶的检测精度造成了挑战。针对以上问题, 提出一种基于双分支网络的SAR图像船舶检测方法, 利用时序SAR图像提取的伪背景信息作为一个分支, 联合船舶目标图像的分支, 输入主干网络并行计算特征, 利用双分支的融合特征增强模型对船舶目标的检测能力。同时, 引入特征增强模块(FEM), 通过注意力机制和特征融合对齐模块(FFAM)对特征图进行增强和优化, 增强浅层特征中的语义信息并实现层级特征之间更精细化的融合。在此基础上, 双分支特征融合策略采用动态门控模块(DGM), 联合双分支特征生成动态门控权重, 自适应地调整双分支特征的融合比例, 从而加强对目标特征的关注。在时序SAR影像船舶数据集上的实验结果表明, 与YOLOv11、YOLOv8等主流的旋转框检测方法相比, 所提方法达到了最高的精确率和平均精确率(AP), 尤其在目标密集的近岸场景中达到较高召回率的同时保持较高的精确率, 能够更精确地检测到停靠在海岸线的船舶目标。

关键词: 船舶检测, 合成孔径雷达时序影像, 双分支深度网络, 特征对齐, 动态门控机制

Abstract:

Convolutional neural network-based object detection algorithms have achieved significant progress in ship detection using Synthetic Aperture Radar (SAR) images. However, in inshore scenes, strong scattering signals from coastlines, buildings, and other background clutter limit ship detection accuracy. To address this issue, this study proposes a dual-branch network-based SAR ship detection method. This method uses pseudo-background information extracted from time-series SAR images in one branch, combines it with the ship target image is processed in the other branch, and inputs them into a backbone network for parallel feature extraction. The ship detection capability is further enhanced through a feature fusion model that utilizes dual-branch features. In addition, a Feature Enhancement Module (FEM) is introduced that employs an attention mechanism and a Feature Fusion and Alignment Module (FFAM) to optimize and refine the feature maps, thereby enriching semantic information in shallow features and achieving finer fusion between hierarchical features. A Dynamic Gated Module (DGM) is applied in the dual-branch feature fusion strategy, generating dynamic gated weights to adaptively adjust the fusion ratio and enhance focus on the target features. Experimental results on a temporal SAR image ship dataset show that compared with mainstream bounding box detection methods, such as YOLOv11 and YOLOv8, the proposed method achieves the highest Average Precision (AP). In inshore scenes with densely populated objects, it achieves high recall and precision rates and is able to detect ships docked at the coastline more accurately.

Key words: ship detection, Synthetic Aperture Radar (SAR) time-series imagery, dual-branch deep network, feature alignment, dynamic gated mechanism