大场景SAR影像舰船目标检测的轻量化研究
作者贡献:Author Contributions
张尧和张艳参与实验设计;张尧、王步云完成实验操作;张尧、张艳、王涛、王步云参与论文的写作和修改。所有作者均阅读并同意最终稿件的提交。
The study was designed by ZHANG Yao and ZHANG Yan. The experimental operation was completed by ZHANG Yao, WANG Buyun. The manuscript was drafted and revised by ZHANG Yao,ZHANG Yan, WANG Tao and WANG Buyun. All the authors have read the last version of paper and consented for submission.
张 尧(1992— ),男,河南邓州人,硕士,主要从事遥感影像智能解译等研究。E-mail: spzhangyao2011@163.com |
收稿日期: 2024-10-15
修回日期: 2024-11-28
网络出版日期: 2025-01-23
基金资助
装备综合研究科研项目(a8203)
Lightweight Research on Ship Target Detection in Large-scale SAR Images
Received date: 2024-10-15
Revised date: 2024-11-28
Online published: 2025-01-23
Supported by
Equipment Comprehensive Research Scientific Project(a8203)
【目的】基于合成孔径雷达(SAR)影像的舰船检测已广泛用于海洋搜救、港口侦察、领海防御等领域。然而,随着在轨智能处理技术的快速发展,对星载SAR影像舰船目标的实时检测能力提出了更高的要求。【方法】因此针对当前SAR影像舰船目标尺度多样,靠岸船只背景复杂,各类遥感平台硬件资源受限等问题,本文提出了一种轻量级SAR影像舰船检测模型LWM-YOLO。首先,采用优化骨干网络结合注意力机制降低网络复杂度及参数量,提出一个轻量化骨干网络(LWCA);其次,针对解决目标尺度多样问题,构建了一个轻型特征融合模块(LGS-FPN),在增强SAR影像中舰船目标细节信息提取能力的同时减少计算复杂度;然后,为优化定位精度,提出基于MPD-Head的检测架构,提升复杂环境中微小舰船目标的检测效果;最后,在LS-SSDD和SSDD舰船目标检测数据集上,将本文算法与主流算法进行对比实验。【结果】实验结果表明,本文提出的算法平均精度值(mAP)分别达到了74.7%和97.3%,较基础模型提升了1.5%和1.0%。同时,本文所提算法参数规模缩减至原模型的36%,计算复杂度降至80%。【结论】与其他方法相比,本文提出的算法不仅在精度上有所提升,而且在检测速率上也具有显著优势。这一研究成果可为智能化目标检测、空间在轨应用等领域提供有力支撑。
张尧 , 张艳 , 王涛 , 王步云 . 大场景SAR影像舰船目标检测的轻量化研究[J]. 地球信息科学学报, 2025 , 27(1) : 256 -270 . DOI: 10.12082/dqxxkx.2025.240574
[Objectives] Ship detection using Synthetic Aperture Radar (SAR) images has gained widespread recognition and application across various fields, including marine search and rescue, port reconnaissance, and territorial sea defense. Nevertheless, with the rapid advancement of on-orbit intelligent processing technologies, higher demands have emerged for real-time detection of ship targets in spaceborne SAR images. [Methods] To address challenges such as the diverse scales of ship targets in current SAR images, the complex background of shore-based vessels, and the limited hardware resources of various remote sensing platforms, this paper presents a lightweight SAR image ship detection model, LWM-YOLO. Firstly, we propose a Lightweight Backbone Network (LWCA) designed specifically for SAR image processing. The LWCA integrates an optimized backbone network with an attention mechanism, effectively reducing the model's complexity and parameter size while maintaining high performance and lowering computational demands. Secondly, to tackle the issue of diverse target scales in SAR images, we have constructed a lightweight feature fusion module, termed LGS-FPN. This module enhances the extraction of detailed information on ship targets in SAR images by efficiently fusing features from different scales, improving detection performance for ship targets of various sizes. Furthermore, the module minimizes computational complexity, ensuring that the model can operate smoothly without significant resource consumption. In addition to addressing the scale issue, we have also focused on optimizing localization accuracy. We introduce a detection architecture based on the MPD-Head, which leverages the strengths of the MPD-Head to improve detection performance for small ship targets in complex environments. Finally, we validate the proposed algorithm through comparative experiments with mainstream methods on the LS-SSDD and SSDD ship detection datasets. [Results] The results demonstrate that our algorithm achieved mean Average Precision (mAP) values of 74.7% and 97.3% on the respective datasets, representing improvements of 1.5 and 1.0 percentage points over the baseline model. Additionally, the parameter size of our model was reduced to 36% of the baseline model, and computational complexity decreased to 80%. [Conclusions] Compared to other mainstream algorithms, the proposed method demonstrates not only higher accuracy but also significant advantages in detection speed. These findings can provide robust support for intelligent target detection, space-based in-orbit applications, and related fields.
表1 环境配置Tab. 1 Environment configuration |
参数 | 配置 |
---|---|
操作系统 | Ubuntu22.04 |
GPU | Geforce RTX 2060Ti 6 G |
CPU | AMD Ryzen 7 3700 X |
内存 | 16 G |
深度学习框架 | Pytorch12.0 |
Python版本 | 3.8 |
GPU加速平台 | CUDA11.4 |
表2 LWM-YOLO对比实验Tab. 2 LWM-YOLO comparison experiments |
数据集 | 网络模型 | 精确率 | 召回率 | mAP | FLOPs/G | 参数量/M | FPS/(帧/s) |
---|---|---|---|---|---|---|---|
LS-SSDD | FreeAnchor[40] | 0.553 | 0.777 | 0.710 | 127.82 | 36.33 | 11.47 |
DCN[41] | 0.741 | 0.769 | 0.738 | 116.82 | 41.93 | 10.05 | |
EfficientDet[42] | 0.621 | 0.675 | 0.614 | 107.52 | 39.40 | 11.42 | |
YOLOv8n | 0.812 | 0.660 | 0.732 | 8.10 | 3.01 | 148.32 | |
YOLOv9t | 0.853 | 0.661 | 0.741 | 7.60 | 1.97 | 120.92 | |
YOLOv10n | 0.780 | 0.616 | 0.703 | 8.20 | 2.69 | 128.25 | |
YOLOv11n | 0.825 | 0.646 | 0.729 | 6.30 | 2.58 | 138.38 | |
本文方法 | 0.837 | 0.672 | 0.747 | 6.50 | 1.07 | 152.40 | |
SSDD | PANET[43] | 0.868 | 0.919 | 0.912 | — | — | 11.65 |
Grid R-CNN[44] | 0.878 | 0.897 | 0.890 | — | — | 9.18 | |
YOLOv8n | 0.942 | 0.956 | 0.963 | 8.10 | 3.01 | 691.51 | |
YOLOv9t | 0.959 | 0.916 | 0.969 | 7.60 | 1.97 | 707.03 | |
YOLOv10n | 0.920 | 0.964 | 0.960 | 8.20 | 2.69 | 757.44 | |
YOLOv11n | 0.961 | 0.964 | 0.970 | 6.30 | 2.58 | 725.33 | |
本文方法 | 0.941 | 0.965 | 0.973 | 6.50 | 1.07 | 729.90 |
注:加粗数值表示最优值。 |
表3 LWM-YOLO消融实验Tab. 3 LWM-YOLO ablation experiment |
网络模型 | 精确率 | 召回率 | mAP | FLOPs/G | 参数量/M | FPS/(帧/s) |
---|---|---|---|---|---|---|
YOLOv8n | 0.812 | 0.660 | 0.732 | 8.1 | 3.01 | 148.32 |
YOLOv8n-M | 0.833 | 0.644 | 0.729 | 8.1 | 3.01 | 146.92 |
YOLOv8n-LGS | 0.823 | 0.669 | 0.743 | 7.1 | 2.03 | 149.23 |
YOLOv8n-L | 0.824 | 0.654 | 0.727 | 7.5 | 2.23 | 152.28 |
YOLOv8n-L-LGS | 0.819 | 0.667 | 0.738 | 6.5 | 1.07 | 152.81 |
YOLOv8n-L-LGS-M | 0.837 | 0.672 | 0.747 | 6.5 | 1.07 | 152.40 |
注:加粗数值表示最优值。“-L”表示使用LWCA骨干网络;“-LGS”使用LGS-FPN特征融合策略;“-M”表示MPD-Head。 |
表4 更换特征融合网络的实验结果Tab. 4 Experimental results of altering the feature fusion network |
网络模型 | 精确率 | 召回率 | mAP | FLOPs/G | 参数量/M | FPS/(帧/s) |
---|---|---|---|---|---|---|
YOLOv8n-PAFPN | 0.812 | 0.66 | 0.732 | 8.1 | 3.01 | 148.32 |
YOLOv8n-HSFPN | 0.813 | 0.649 | 0.728 | 6.9 | 1.93 | 148.07 |
YOLOv8n-LGSFPN | 0.823 | 0.669 | 0.743 | 7.1 | 2.03 | 149.23 |
注:加粗数值表示最优值。 |
利益冲突:Conflicts of Interest 所有作者声明不存在利益冲突。
All authors disclose no relevant conflicts of interest.
[1] |
王桥. 中国环境遥感监测技术进展及若干前沿问题[J]. 遥感学报, 2021, 25(1):25-36.
[
|
[2] |
|
[3] |
|
[4] |
|
[5] |
|
[6] |
|
[7] |
|
[8] |
|
[9] |
|
[10] |
|
[11] |
|
[12] |
程传祥, 金飞, 林雨准, 等. 应用多尺度融合策略和改进YOLOV5的道路病害无人机检测[J]. 地球信息科学学报, 2024, 26(8):1991-2007.
[
|
[13] |
|
[14] |
|
[15] |
|
[16] |
|
[17] |
|
[18] |
|
[19] |
|
[20] |
|
[21] |
陈丁, 万刚, 李科. 多层特征与上下文信息相结合的光学遥感影像目标检测[J]. 测绘学报, 2019, 48(10):1275-1284.
[
|
[22] |
苏娟, 杨龙, 黄华, 等. 用于SAR图像小目标舰船检测的改进SSD算法[J]. 系统工程与电子技术, 2020, 42(5):1026-1034.
[
|
[23] |
贾晓雅, 汪洪桥, 杨亚聃, 等. 基于YOLO框架的无锚框SAR图像舰船目标检测[J]. 系统工程与电子技术, 2022, 44(12):3703-3709.
[
|
[24] |
高鹏飞, 曹雪峰, 李科, 等. 融合多元稀疏特征与阶层深度特征的遥感影像目标检测[J]. 地球信息科学学报, 2023, 25(3):638-653.
[
|
[25] |
杨金鹏, 黄柏圣, 陈小娇, 等. 一种改进YOLOV5的遥感和SAR图像舰船目标检测[J]. 中国电子科学研究院学报, 2023, 18(9):821-829.
[
|
[26] |
薛远亮, 金国栋, 侯笑晗, 等. 融合注意力机制与改进SSD算法的SAR舰船目标检测方法[J]. 计算机应用研究, 2022, 39(1):265-269.
[
|
[27] |
|
[28] |
李健伟, 曲长文, 彭书娟. SAR图像舰船目标联合检测与方向估计[J]. 武汉大学学报(信息科学版), 2019, 44(6):901-907.
[
|
[29] |
|
[30] |
|
[31] |
|
[32] |
|
[33] |
|
[34] |
|
[35] |
|
[36] |
|
[37] |
|
[38] |
|
[39] |
|
[40] |
|
[41] |
|
[42] |
|
[43] |
|
[44] |
|
[45] |
|
/
〈 |
|
〉 |