SAR图像舰船目标检测的轻量化和特征增强研究

doi:10.19665/j.issn1001-2400.20230407

摘要/Abstract

摘要：

针对合成孔径雷达(SAR)图像中的舰船目标的准确率易受近岸杂波的影响,且现有检测算法复杂度高,在嵌入式设备上的部署难度大的问题,提出一种适用于嵌入式设备的轻量化高精度SAR图像舰船目标检测算法CA-Shuffle-YOLO。基于YOLO v5目标检测算法,对骨干网络进行轻量化及特征精细化提取两个方面的改进,引入轻量化模块以降低网络的计算复杂度,提高推理速度,并引入协同注意力机制模块增强算法对近岸船舶目标的细节信息的提取能力。在特征融合网络中采用加权特征融合以及跨模块融合,增强模型对SAR舰船目标的细节信息的融合能力,同时,利用深度卷积模块降低计算复杂度,提高实时性。通过在SSDD舰船目标检测数据集上的测试及对比实验的结果,表明CA-Shuffle-YOLO的检测准确率约为97.4%,检测帧率为206 FPS,所需运算复杂度为6.1 GFlops,相比原始的YOLO v5,所提方法的检测帧率提升了60 FPS,所需运算复杂度降低为原来的12%。

关键词: 合成孔径雷达, 目标检测, 卷积神经网络, 特征提取

Abstract:

The accuracy of ship targets detection in sythetic aperture radar images is susceptible to the nearshore clutter.The existing detection algorithms are highly complex and difficult to deploy on embedded devices.Due to these problems a lightweight and high-precision SAR image ship target detection algorithm CA-Shuffle-YOLO(Coordinate Shuffle You Only Look Once) is proposed in this article.Based on the YOLO v5 target detection algorithm,the backbone network is improved in two aspects:lightweight and feature refinement.The lightweight module is introduced to reduce the computational complexity of the network and improve the reasoning speed,and a collaborative attention mechanism module is introduced to enhance the algorithm's ability to extract the detailed information on near-shore ship targets.In the feature fusion network,weighted feature fusion and cross-module fusion are used to enhance the ability of the model to fuse the detailed information on SAR ship targets.At the same time,the depth separable convolution is used to reduce the computational complexity and improve the real-time performance.Through the test and comparison experiments on the SSDD ship target detection dataset,the results show that the detection accuracy of CA-Shuffle-YOLO is 97.4%,the detection frame rate is 206FPS,and the required computational complexity is 6.1GFlops.Compare to the original YOLO v5,the FPS of our algorithm is 60FPS higher with the required computational complexity of our algorithm being only the 12% that of the ordinary YOLOv5.

Key words: synthetic aperture radar, object detection, convolutional neural networks, feature extraction

中图分类号:

TP391.4

龚峻扬, 付卫红, 方厚章. SAR图像舰船目标检测的轻量化和特征增强研究[J]. 西安电子科技大学学报, 2024, 51(2): 96-106.

GONG Junyang, FU Weihong, FANG Houzhang. Research on lightweight and feature enhancement of SAR image ship targets detection[J]. Journal of Xidian University, 2024, 51(2): 96-106.

图/表 15

图1

图2

图3

图4

表1

图5

表2

表3

表4

表5

表6

图6

表7

图7

图8

参考文献 18

[1]	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. doi: 10.3969/j.issn.1001-506X.2020.05.08
	SU Juan, YANG Long, HUANG Hua, et al. Improved SSD Algorithm for SAR Image Small Target Ship Detection[J]. Journal of Systems Engineering and Electronics, 2020, 42(5):1026-1034.
[2]	何楚, 张宇, 廖紫纤, 等. 基于压缩感知的SAR图像CFAR目标检测算法[J]. 武汉大学学报(信息科学版), 2014, 39(7):878-882.
	HE Chu, ZHANG Yu, LIAO Zixian, et al. SAR Image CFAR Target Detection Algorithm Based on Compressed Sensing[J]. Geomatics and Information Science of Wuhan University, 2014, 39(7):878-882.
[3]	CUI Z, LI Q, CAO Z, et al. Dense Attention Pyramid Networks for Multi-Scale Ship Detection in SAR Images[J]. IEEE Transactions on Geoscience and Remote Sensing, 2019, 57(11):8983-8997.
[4]	CHEN S Q, ZHAN R H, ZHANG J. Robust Single Stage Detector Based on Two-Stage Regression for SAR Ship Detection[C]//Proceedings of the 2nd International Conference on Innovation in Artificial Intelligence. New York: ACM,2018:169-174.
[5]	WANG J, LIN Y, GUO J, et al. SSS-YOLO:Towards More Accurate Detection for Small Ships in Sar Image[J]. Remote Sensing Letters, 2021, 12(2):93-102.
[6]	REDMON J, FARHADI A. Yolov3:An Incremental Improvement(2018)[J/OL].[2024-01-01]. https://arxiv.org/abs/1804.02767.
[7]	JIANG J, FU X, QIN R, et al. High-Speed Lightweight Ship Detection Algorithm Based on YOLO-V4 for Three-Channels RGB SAR Image[J]. Remote Sensing, 2021, 13(10):1909.
[8]	李永刚, 朱卫纲, 黄琼男, 等. 复杂背景下SAR图像近岸舰船目标检测[J]. 系统工程与电子技术, 2022, 44(10):3096-3103. doi: 10.12305/j.issn.1001-506X.2022.10.13
	LI Yonggang, ZHU Weigang, HUANG Qiongnan, et al. SAR Image Near-Shore Ship Target Detection Under Complex Background[J]. Journal of Systems Engineering and Electronics, 2022, 44(10):3096-3103.
[9]	LI J, QU C, SHAO J. Ship Detection in SAR Images Based on an Improved Faster R-CNN[C]// 2017 SAR in Big Data Era: Models,Methods and Applications.Piscataway:IEEE, 2017:1-6.
[10]	GUI Y, LI X, XUE L. A Multilayer Fusion Light-Head Detector for SAR Ship Detection[J]. Sensors, 2019, 19(5):1124.
[11]	TANG G, ZHUGE Y, CLARAMUNT C, et al. N-Yolo:A SAR Ship Detection Using Noise-Classifying and Complete-Target Extraction[J]. Remote Sensing, 2021, 13(5):871.
[12]	QING Y, LIU W, FENG L, et al. Improved Yolo Network for Free-Angle Remote Sensing Target Detection[J]. Remote Sensing, 2021, 13(11):2171.
[13]	YANG X, YAN J. On the Arbitrary-Oriented Object Detection:Classification Based Approaches Revisited[J]. International Journal of Computer Vision, 2022, 130(5):1340-1365.
[14]	REDMON J, DIVVALA S, GIRSHICK R, et al. You Only Look Once:Unified,Real-Time Object Detection[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE, 2016:779-788.
[15]	REDMON J, FARHADI A. YOLO9000:Better,Faster,Stronger[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE, 2017:7263-7271.
[16]	GLOROT X, BORDES A, BENGIO Y. Deep Sparse Rectifier Neural Networks[C]//Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics. Atlanta:JMLR, 2011:315-323.
[17]	HOU Q, ZHOU D, FENG J. Coordinate Attention for Efficient Mobile Network Design[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway:IEEE, 2021:13713-13722.
[18]	HE K, ZHANG X, REN S, et al. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 37(9):1904-1916. doi: 10.1109/TPAMI.2015.2389824 pmid: 26353135

传感器类型或环境	成像极化方式	船舶情况	成像规模
RadarSat-2 TerraSAR-X Sentinel-1	HH,VV, VH,HV	1∶1 1∶2 2∶1	不同尺寸及材料
海洋环境	分辨率	成像位置
恶劣天气和晴朗天气	1 m~5 m	远海及近岸

骨干网络/指标	P	R	F1	mAP@0.5	mAP@0.5:0.95
CSPDarkNet	0.972	0.969	0.970	0.981	0.748
Shuffle-YOLO	0.968	0.934	0.950	0.969	0.704
Shuffle-YOLO+CA(1)	0.959	0.934	0.946	0.962	0.697
Shuffle-YOLO+CA(2)	0.952	0.945	0.948	0.967	0.708
Shuffle-YOLO+CA(3)	0.97	0.938	0.954	0.967	0.705
Shuffle-YOLO+CA(4)	0.955	0.94	0.947	0.965	0.697
Shuffle-YOLO+CA(5)	0.955	0.929	0.941	0.960	0.693
Shuffle-YOLO+CA(6)	0.975	0.943	0.959	0.971	0.705

骨干网络\指标	帧率	GFLOPs	参数量	F1_density[%/M-params]
CSPDarkNet	146	50.2	21 037 638	0.046
Shuffle-YOLO	221	4.3	2 024 166	0.469
Shuffle-YOLO+CA(1)	280	4.3	2 022 846	0.468
Shuffle-YOLO+CA(2)	283	4.4	2 049 126	0.463
Shuffle-YOLO+CA(3)	267	4.4	2 049 126	0.466
Shuffle-YOLO+CA(4)	259	4.5	2 132 550	0.444
Shuffle-YOLO+CA(5)	231	4.4	2 026 014	0.464
Shuffle-YOLO+CA(6)	299	4.8	2 465 286	0.389

模型/指标	P	R	F1	mAP@0.5	mAP@0.5:0.95
CSPDarknet+PANet	0.972	0.969	0.970	0.981	0.748
CSPDarknet+BiFPN	0.981	0.965	0.973	0.982	0.747
CSPDarknet+WM-FPN	0.978	0.971	0.974	0.982	0.754

模型/指标	帧率	GFLOPs	参数量	F1_density[%/M-params]
CSPDarknet+PANet	146	50.2	21 037 638	0.046
CSPDarknet+BiFPN	152	52.1	21 523 110	0.045
CSPDarknet+WM-FPN	160	48.5	19 182 687	0.051