一种形态学引导的解耦式SAR舰船有向检测框架

汪泽宇; 王青松

doi:10.11999/JEIT250979

一种形态学引导的解耦式SAR舰船有向检测框架

doi: 10.11999/JEIT250979 cstr: 32379.14.JEIT250979

汪泽宇¹,
王青松^1, ,

中山大学电子与通信工程学院，深圳，518107

基金项目: 国家自然科学基金(62273365)，“小米青年学者”项目

详细信息

作者简介:
汪泽宇：男，研究生，研究方向为深度学习、目标检测、表征学习

王青松：男，教授，研究方向为遥感图像精化处理、协同探测感知与信息融合

通讯作者:
王青松　wangqs5@mail.sysu.edu.cn

中图分类号: TP75; TP391.41
计量
- 文章访问数: 36
- HTML全文浏览量: 14
- PDF下载量: 0
- 被引次数: 0
出版历程
- 收稿日期: 2025-09-24
- 修回日期: 2025-12-12
- 录用日期: 2025-12-12
- 网络出版日期: 2025-12-25

A Morphology-Guided Decoupled Framework for Oriented SAR Ship Detection

WANG Zeyu¹,
WANG Qingsong^{1
, ,}

School of Electronics and Communication Engineering, Sun Yat-sen University, Shenzhen 518107, China

Funds: The National Natural Science Foundation of China(62273365), Xiaomi Young Talents Program

摘要

摘要: 合成孔径雷达(SAR)以其全天时、全天候的观测能力，在遥感检测中得到了广泛应用。然而，受限于标注精度，目前主流的 SAR 目标检测方法多依赖水平框标注，难以实现精确的目标角度和尺度估计。同时，尽管弱监督学习在光学图像中的角度预测取得了进展，但其忽视了 SAR 特有的成像几何，难以有效泛化。为解决上述挑战，该文提出了一种融合 SAR 成像机理的有向舰船检测新框架，核心思想在于将检测任务解耦为定位与方向估计两个独立的子模块。其中，定位模块可以直接利用任意现有的、在水平框标注上训练的检测器；而方向估计模块则在一个专门构建的形态学合成二值数据集上进行全监督训练。该框架的优势在于无需修改原有检测器结构和重新训练的前提下，即插即用地赋予模型高精度的有向框预测能力。实验验证了所提方法在多个数据集上相较于现有仅依赖水平框监督的方法表现出更优的性能，部分场景中甚至超越全监督方法，体现出强大的有效性与工程应用价值
- 有向目标检测 /
- 形态学仿真 /
- 仿真驱动学习 /
- 解耦检测
Abstract: Objective Synthetic Aperture Radar (SAR) is an indispensable tool in remote sensing, yet the effective detection of ships in SAR imagery remains a significant challenge. Mainstream deep learning methods typically rely on Horizontal Bounding Boxes (HBBs) annotations, which are insufficient for estimating the precise orientation and scale of ships. While Oriented Bounding Boxes (OBBs) provide this data, obtaining accurate OBBs annotations for SAR is prohibitively expensive and often unreliable due to SAR-specific phenomena like speckle noise and geometric distortions. Weakly Supervised Object Detection (WSOD) presents a promising alternative, but existing techniques developed for optical images do not generalize well to the SAR domain. This study departs from the traditional WSOD paradigm to propose a novel, simulation-driven decoupled framework. The objective is to enable existing HBB detectors to predict precise OBBs without structural modification by training a specialized orientation module on a fully-supervised synthetic dataset that mimics core SAR target morphology. Methods The proposed method decouples the complex task of oriented object detection into two sequential sub-problems: coarse localization and fine-grained orientation estimation (Fig 1). First, an Axis-aligned Localization Module, consisting of a standard HBBs detection network (e.g., YOLOX), is trained on available HBB labels to identify candidate regions of interest. This stage leverages the high performance of existing detectors to achieve high-recall, coarse localization, outputting a set of image patches containing potential ship targets. Second, to learn orientation without OBBs labels, a large-scale Morphological Simulation Dataset of binary images is constructed. This synthetic dataset forms the core of the method. The process begins with generating simple binary rectangles with randomized aspect ratios and known ground-truth orientations. To simulate the appearance of binarized SAR ship targets, a series of morphological modulations are applied, including edge and regional erosion and dilation to create boundary ambiguity, and the addition of structured “strong scattering cross noise” to mimic SAR artifacts. This yields a large training set of binary images with precise orientation labels. Third, an Orientation Estimation Module, using a lightweight ResNet-18 network, is trained exclusively on this synthetic dataset. It learns to predict an object’s orientation and refine its aspect ratio based purely on shape and contour. During inference, candidate patches from the localization module are binarized and fed into this orientation module. The final Oriented Bounding Box is then generated by fusing the spatial coordinates from the initial HBBs with the predicted angle and refined dimensions. Results and Discussions The efficacy of the proposed method is extensively evaluated on two public SAR ship detection benchmarks: HRSID and SSDD. The model is trained using only HBBs annotations, while its performance is evaluated against ground-truth OBBs using the standard Average Precision (AP50) and Recall (R) metrics. The method demonstrates superior performance compared to existing weakly supervised techniques and is competitive with fully supervised ones (Table 1 and Table 2). On the challenging HRSID dataset, the proposed method achieves an AP50 of 84.3% and a recall of 91.9%. This result not only surpasses other weakly supervised methods like H2Rbox-v2 (56.2% AP50) and the method from Yue et al. (81.5% AP50), but also exceeds the performance of several fully supervised algorithms, including R-RetinaNet (72.7% AP50) and S2ANet (80.8% AP50). A similar performance advantage is observed on the SSDD dataset, where the method attains an AP50 of 89.4%, representing a significant improvement over the state-of-the-art weakly supervised result of 87.3%. Visual inspection of the detection results confirms these quantitative findings (Fig 1). The proposed method exhibits a markedly lower rate of missed detections, particularly for small and densely clustered ships, when compared to other weakly supervised approaches. This enhanced robustness is attributed to the high-recall nature of the first-stage localization network combined with the accurate orientation cues learned from the morphological dataset. To address key methodological questions, new experiments were conducted. First, the domain gap between the synthetic and real data was analyzed. A UMAP visualization of the network's high-dimensional features (Fig. 5) shows a high degree of overlap and a similar manifold structure between the two domains, confirming their morphological similarity. Second, a crucial ablation study on the morphological components (Fig 4) demonstrated that each simulation element progressively contributes to the final performance, justifying the high-fidelity simulation process. Conclusions This paper presents a novel and effective morphology-guided, decoupled framework for oriented ship detection in SAR images. By decoupling localization and orientation, the framework enables standard HBB detectors to perform high-precision oriented detection without retraining. The central innovation is the fully-supervised morphological simulation dataset, which allows a dedicated module to learn robust orientation features from structural contours, bypassing the challenges of real SAR data. Extensive experiments demonstrate that this approach significantly outperforms existing HBB-supervised methods and is competitive with fully supervised ones. The method's plug-and-play nature highlights its practical value.
- Oriented Objective Detection /
- Morphological Simulation /
- Simulation-Driven Learning /
- Decoupled Detection

HTML全文

图 1 模型总览图，(c)模块训练集来源于形态学仿真的数据，在推理阶段将(a)模块的二值化切片作为输入得到最终的有向检测结果。

下载: 全尺寸图片幻灯片

图 2 不同仿真条件单独作用时的效果图

下载: 全尺寸图片幻灯片

图 3 不同方法的检测可视化结果对比图，第一列中的绿色矩形代表数据集的矩形框标注，第二列到第三列中的绿色矩形代表网络正确预测的目标，红色矩形代表网络的虚警

下载: 全尺寸图片幻灯片

图 4 仿真数据集样本多样性结果展示，虚线代表使用了全部的仿真条件的性能指标。各实验结果根据对照组进行归一化显示处理

下载: 全尺寸图片幻灯片

图 5 不同源的数据分布降维可视化，仿真样本数据集与HRSID数据集都进行随机抽样选取5000个样本，由于SSDD训练集样本数少于5000，使用了所有训练集进行降维可视化

下载: 全尺寸图片幻灯片

图 6 不同方法使用不同训练样本量对AP50精度的影响

下载: 全尺寸图片幻灯片

表 1 本文方法与典型检测方法在SSDD数据集上的性能比较

方法	监督方式	AP50	R
R-RetinaNet^[18]	有向框监督	79.9	84.0
GWD^[19]	有向框监督	82.5	85.9
S2Anet^[20]	有向框监督	88.4	90.0
R-Faster-RCNN^[3]	有向框监督	87.7	89.8
RoI Transformer^[21]	有向框监督	89.5	91.0
[14]	水平框监督	87.3	89.7
本文方法	水平框监督	89.4	95.6

下载: 导出CSV

表 2 本文方法与典型检测方法在HRSID数据集上的性能比较

方法	监督方式	AP50	R
R-RetinaNet^[18]	有向框监督	72.7	79.2
GWD^[19]	有向框监督	73.3	79.1
S2Anet^[20]	有向框监督	80.8	83.9
R-Faster-RCNN^[3]	有向框监督	77.3	82.1
RoI Transformer^[21]	有向框监督	83.8	86.4
H2Rbox-v2^[7]	水平框监督	56.2	68.6
Wholly-WOOD^[8]	水平框监督	61.5	64.5
[14]	水平框监督	81.5	85.0
本文方法	水平框监督	84.3	91.9

下载: 导出CSV

表 3 不同轴对齐检测器对最终精度的影响

数据集	轴对齐检测器	HBBs AP50	OBBs AP50	HBBs R	OBBs R
SSDD	FCOS^[22]	81.9	78.1	91.2	88.8
	CenterNet^[23]	90.2	88.5	96.7	94.7
	Faster-RCNN^[3]	90.0	88.1	93.0	90.1
	YOLOX^[4]	90.3	89.4	99.1	95.6
HRSID	FOCS^[22]	78.4	71.2	87.6	80.3
	CenterNet^[23]	86.6	77.5	92.3	85.1
	Faster-RCNN^[3]	79.7	70.3	83.5	78.7
	YOLOX^[4]	89.0	84.3	96.1	90.8

下载: 导出CSV

参考文献(26)

[1]	DI BISCEGLIE M and GALDI C. CFAR detection of extended objects in high-resolution SAR images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2005, 43(4): 833–843. doi: 10.1109/TGRS.2004.843190.
[2]	LENG Xiangguang, JI Kefeng, YANG Kai, et al. A bilateral CFAR algorithm for ship detection in SAR images[J]. IEEE Geoscience and Remote Sensing Letters, 2015, 12(7): 1536–1540. doi: 10.1109/LGRS.2015.2412174.
[3]	REN Shaoqing, HE Kaiming, GIRSHICK R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(6): 1137–1149. doi: 10.1109/tpami.2016.2577031.
[4]	GE Zheng, LIU Songtao, WANG Feng, et al. YOLOX: Exceeding YOLO series in 2021[EB/OL]. https://arxiv.org/abs/2107.08430, 2021.
[5]	CHEN Yuming, YUAN Xinbin, WANG Jiabao, et al. YOLO-MS: Rethinking multi-scale representation learning for real-time object detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025, 47(6): 4240–4252. doi: 10.1109/tpami.2025.3538473.
[6]	DOSOVITSKIY A, BEYER L, KOLESNIKOV A, et al. An image is worth 16x16 words: Transformers for image recognition at scale[C]. 9th International Conference on Learning Representations, Austria, 2021. (查阅网上资料, 未找到本条文献出版城市信息, 请确认并补充).
[7]	YU Yi, YANG Xue, LI Qingyun, et al. H2RBox-v2: Incorporating symmetry for boosting horizontal box supervised oriented object detection[C]. Proceedings of the 37th International Conference on Neural Information Processing Systems, New Orleans, USA, 2023: 2581.
[8]	YU Yi, YANG Xue, LI Yansheng, et al. Wholly-WOOD: Wholly leveraging diversified-quality labels for weakly-supervised oriented object detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025, 47(6): 4438–4454. doi: 10.1109/TPAMI.2025.3542542.
[9]	HU Fengming, XU Feng, WANG R, et al. Conceptual study and performance analysis of tandem multi-antenna spaceborne SAR interferometry[J]. Journal of Remote Sensing, 2024, 4: 0137. doi: 10.34133/remotesensing.0137.
[10]	YOMMY A S, LIU Rongke, and WU Shuang. SAR image despeckling using refined lee filter[C]. 2015 7th International Conference on Intelligent Human-Machine Systems and Cybernetics, Hangzhou, China, 2015: 260–265. doi: 10.1109/IHMSC.2015.236.
[11]	KANG Yuzhuo, WANG Zhirui, ZUO Haoyu, et al. ST-Net: Scattering topology network for aircraft classification in high-resolution SAR images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2023, 61: 5202117. doi: 10.1109/tgrs.2023.3236987.
[12]	ZHANG Yipeng, LU Dongdong, QIU Xiaolan, et al. Scattering-point-guided RPN for oriented ship detection in SAR images[J]. Remote Sensing, 2023, 15(5): 1411. doi: 10.3390/rs15051411.
[13]	PAN Dece, GAO Xin, DAI Wei, et al. SRT-Net: Scattering region topology network for oriented ship detection in large-scale SAR images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 5202318. doi: 10.1109/tgrs.2024.3351366.
[14]	YUE Tingxuan, ZHANG Yanmei, WANG Jin, et al. A weak supervision learning paradigm for oriented ship detection in SAR image[J]. IEEE Transactions on Geoscience and Remote Sensing, 2024, 62: 5207812. doi: 10.1109/TGRS.2024.3375069.
[15]	WEI Shunjun, ZENG Xiangfeng, QU Qizhe, et al. HRSID: A high-resolution SAR images dataset for ship detection and instance segmentation[J]. IEEE Access, 2020, 8: 120234–120254. doi: 10.1109/access.2020.3005861.
[16]	ZHANG Tianwen, ZHANG Xiaoling, LI Jianwei, et al. SAR ship detection dataset (SSDD): Official release and comprehensive data analysis[J]. Remote Sensing, 2021, 13(18): 3690. doi: 10.3390/rs13183690.
[17]	ZHOU Yue, YANG Xue, ZHANG Gefan, et al. MMRotate: A rotated object detection benchmark using PyTorch[C]. Proceedings of the 30th ACM International Conference on Multimedia, Lisboa, Portugal, 2022: 7331–7334. doi: 10.1145/3503161.3548541.
[18]	LIN T Y, GOYAL P, GIRSHICK R, et al. Focal loss for dense object detection[C]. 2017 IEEE International Conference on Computer Vision, Venice, Italy, 2017: 2999–3007. doi: 10.1109/iccv.2017.324.
[19]	YANG Xue, ZHANG Gefan, YANG Xiaojiang, et al. Detecting rotated objects as gaussian distributions and its 3-D generalization[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(4): 4335–4354. doi: 10.1109/tpami.2022.3197152.
[20]	LI Jianfeng, CHEN Mingxu, HOU Siyuan, et al. An improved S²A-net algorithm for ship object detection in optical remote sensing images[J]. Remote Sensing, 2023, 15(18): 4559. doi: 10.3390/rs15184559.
[21]	DING Jian, XUE Nan, LONG Yang, et al. Learning RoI transformer for oriented object detection in aerial images[C]. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, USA, 2019: 2844–2853. doi: 10.1109/cvpr.2019.00296.
[22]	TIAN Zhi, SHEN Chunhua, CHEN Hao, et al. FCOS: A simple and strong anchor-free object detector[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 44(4): 1922–1933. doi: 10.1109/tpami.2020.3032166.
[23]	DUAN Kaiwen, BAI Song, XIE Lingxi, et al. CenterNet++ for object detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 46(5): 3509–3521. doi: 10.1109/tpami.2023.3342120.
[24]	HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Deep residual learning for image recognition[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, USA, 2016: 770–778. doi: 10.1109/cvpr.2016.90.
[25]	LIU Zhuang, MAO Hanzi, WU Chaoyuan, et al. A ConvNet for the 2020s[C]. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition, New Orleans, USA, 2022: 11966–11976. doi: 10.1109/cvpr52688.2022.01167.
[26]	HEALY J and MCINNES L. Uniform manifold approximation and projection[J]. Nature Reviews Methods Primers, 2024, 4(1): 82. doi: 10.1038/s43586-024-00363-x.