一种基于密集深度分离卷积的SAR图像水域分割算法

张金松; 邢孟道; 孙光才

doi:10.12000/JR19008

一种基于密集深度分离卷积的SAR图像水域分割算法

DOI: 10.12000/JR19008

西安电子科技大学雷达信号处理国家重点实验室西安 710071
西安电子科技大学信息感知技术协同创新中心西安 710071

基金项目: 国家重点研发计划(2017YFC1405600)，国家自然科学基金创新群体基金(61621005)

详细信息

作者简介:
张金松(1995–)，男，山东德州人，西安电子科技大学信号与信息处理专业博士研究生，研究方向为SAR图像解译，深度学习及SAR成像。E-mail: jinsongxd@163.com

邢孟道(1975–)，男，浙江嵊州人。西安电子科技大学教授，博士生导师，主要研究方向为雷达成像、目标识别和天波超视距雷达信号处理。E-mail: xmd@xidian.edu.cn

孙光才(1984–)，男，湖北孝感汉川人。西安电子科技大学副教授，博士生导师，主要研究方向为多通道波束指向 SAR 成像和 SAR 动目标成像。E-mail: rsandsgc@126.com

通讯作者:
张金松　jinsongxd@163.com

中图分类号: TN958
计量
- 文章访问数: 3902
- HTML全文浏览量: 1484
- PDF下载量: 279
- 被引次数: 0
出版历程
- 收稿日期: 2019-01-14
- 修回日期: 2019-04-08
- 网络出版日期: 2019-06-01

A Water Segmentation Algorithm for SAR Image Based on Dense Depthwise Separable Convolution

National Laboratory of Radar Signal Processing, Xidian University, Xi’an 710071, China
Collaborative Innovation Center of Information Sensing and Understand, Xidian University, Xi’an 710071, China

Funds: The State Key Research Development Program (2017YFC1405600), The Foundation for Innovative Research Groups of the National Natural Science Foundation of China (61621005)

More Information

Corresponding author: ZHANG Jinsong, jinsongxd@163.com

摘要

摘要: SAR图像的水域分割在舰船目标检测、灾害监测等军事和民用领域具有重要意义。针对传统水域分割算法鲁棒性差、难以准确进行分割等问题，该文首先建立了基于高分三号的SAR图像水域分割数据集，并基于深度学习技术提出了基于密集深度分离卷积的分割网络架构，该网络以SAR图像作为输入，通过密集分离卷积和扩张卷积提取图像高维特征，并构造基于双线性插值的上采样解码模块用于输出分割结果。在水域分割数据集上的实验结果表明，与传统方法相比，该方法不仅在分割准确度上有大幅提高，在算法的鲁棒性和分割速度上也具有部分优势，具备较好的工程实用价值。
- 合成孔径雷达 /
- 水域分割 /
- 深度学习 /
- 密集分离卷积 /
- 特征提取
Abstract: Water segmentation of real SAR images is of great significance in military and civilian applications such as ship target detection and disaster monitoring. To solve the issues of poor robustness and inaccurate segmentation of traditional water segmentation algorithms, this paper first establishes a SAR water segmentation dataset based on the GF3 satellite and then presents a segmentation network architecture based on depthwise separable convolution. The network takes real SAR images as inputs, extracts high-dimensional features through depthwise separable and dilated convolutions, constructs an up-sampling and decoding module based on bilinear interpolation, and then outputs the corresponding segmentation results. The segmentation results of a water segmentation dataset show that the proposed segmentation method remarkably improves the segmentation accuracy, the segmentation robustness and running speed than traditional method. Therefore, the findings demonstrate the excellent practical engineering value of the proposed algorithm.
- Synthetic Aperture Radar (SAR) /
- Water segmentation /
- Deep learning /
- Dense separable convolution /
- Feature extraction

HTML全文

图 1 水域分割对舰船检测意义

Figure 1. Significance of water segmentation for ship detection

下载: 全尺寸图片幻灯片

图 2 常规卷积和深度分离卷积结构对比图

Figure 2. Comparison of conventional convolution and depthwise separable convolution

下载: 全尺寸图片幻灯片

图 3 常规卷积和扩张卷积结构对比图

Figure 3. Comparison of conventional convolution and dilated convolution

下载: 全尺寸图片幻灯片

图 4 特征提取网络结构示意图

Figure 4. The structure of feature extraction network

下载: 全尺寸图片幻灯片

图 5 基于编码-解码结构的SAR图像水域分割网络示意图

Figure 5. The structure of encoder-decoder network for water segmentation

下载: 全尺寸图片幻灯片

图 6 分割网络训练结果示意图

Figure 6. Training results of segmentation network

下载: 全尺寸图片幻灯片

图 7 网络分割结果示意图

Figure 7. Segmentation results of segmentation network

下载: 全尺寸图片幻灯片

图 8 各方法分割结果对比图

Figure 8. Segmentation results of different methods

下载: 全尺寸图片幻灯片

图 9 不同工作模式分割结果对比图

Figure 9. Segmentation results of different imaging modes

下载: 全尺寸图片幻灯片

图 10 不同极化方式分割结果对比图

Figure 10. Segmentation results of different imaging polarizations

下载: 全尺寸图片幻灯片

表 1 高分三号成像模式

Table 1. The imaging modes of GF3 satellite

工作模式	分辨率(m)	极化方式	成像幅宽(km)
滑块聚束(SL)	1	单极化	10
超精细条带(UFS)	3	单极化	30
精细条带1(FSI)	5	双极化	50
精细条带2(FSII)	10	双极化	100
标准条带1(QPSI)	8	全极化	30

下载: 导出CSV

表 2 数据集图像组成

Table 2. The composition of dataset

图像类型	数量	图像尺寸(像素)
原始图像	10	$ \approx 10,000 \times 10,000$
裁剪图像	480	$513 \times 513$
扩充图像	21180	$513 \times 513$

下载: 导出CSV

表 3 数据扩充对分割性能的影响

Table 3. Segmentation effects of data augmentation

扩充方法	像素准确度	交并比
未扩充	0.9569	0.9497
旋转	0.9806	0.9758
翻转	0.9620	0.9603
旋转+翻转	0.9887	0.9844

下载: 导出CSV

表 4 网络结构对分割性能的影响

Table 4. Segmentation effects of network structure

连接方式	像素准确度	交并比
直连	0.9312	0.9289
仅残差	0.9703	0.9681
仅密集	0.9679	0.9638
残差+密集	0.9887	0.9844

下载: 导出CSV

表 5 各水域分割算法性能对比

Table 5. Segmentation performance of different methods

方法类别	具体方法	像素准确度	交并比	小图速度(s)	大图速度(s)
传统方法	FCM	0.6710	0.4644	8.24	206.0
	MRF	0.5961	0.5430	2.29	57.25
	OTSU	0.6303	0.6108	0.06	1.50
	Levelset	0.7134	0.6868	3.41	85.25
深度学习	Unet	0.9533	0.9496	0.07	1.75
	DeepLabv3+	0.9672	0.9566	0.10	2.50
	所提方法	0.9887	0.9844	0.14	3.50
	理想值	1.0000	1.0000	–	–

下载: 导出CSV

表 6 本文方法对多模式多极化下SAR图像的IoU分割结果

Table 6. IoU under multi-mode and multi-polarization by the proposed method

工作模式/极化方式	HH	HV	VH	VV
SL (1 m)	0.9844	–	–	–
UFS (3 m)	0.9240	–	–	–
FSI (5 m)	0.9365	0.9542	–	–
FSII (10 m)	0.9549	0.9454	–	–
QPSI (8 m)	0.9605	0.9684	0.9686	0.9717

下载: 导出CSV

参考文献(30)

[1]	吴一戎. 多维度合成孔径雷达成像概念[J]. 雷达学报, 2013, 2(2): 135–142. doi: 10.3724/SP.J.1300.2013.13047 WU Yirong. Concept of multidimensional space joint-observation SAR[J]. Journal of Radars, 2013, 2(2): 135–142. doi: 10.3724/SP.J.1300.2013.13047
[2]	艾加秋, 齐向阳, 禹卫东. 改进的SAR图像双参数CFAR舰船检测算法[J]. 电子与信息学报, 2009, 31(12): 2881–2885. doi: 10.3724/SP.J.1146.2008.01707 AI Jiaqiu, QI Xiangyang, and YU Weidong. Improved two parameter CFAR ship detection algorithm in SAR images[J]. Journal of Electronics &Information Technology, 2009, 31(12): 2881–2885. doi: 10.3724/SP.J.1146.2008.01707
[3]	牛世林, 郭拯危, 李宁, 等. 星载SAR水域分割研究进展与趋势分析[J]. 聊城大学学报: 自然科学版, 2018, 31(2): 72–86. NIU Shilin, GUO Zhengwei, LI Ning, et al. Research progress and trend analysis of water extraction by spaceborne SAR[J]. Journal of Liaocheng University:Natural Science Edition, 2018, 31(2): 72–86.
[4]	安成锦, 牛照东, 李志军, 等. 典型Otsu算法阈值比较及其SAR图像水域分割性能分析[J]. 电子与信息学报, 2010, 32(9): 2215–2219. doi: 10.3724/SP.J.1146.2009.01426 AN Chengjin, NIU Zhaodong, LI Zhijun, et al. Otsu threshold comparison and SAR water segmentation result analysis[J]. Journal of Electronics &Information Technology, 2010, 32(9): 2215–2219. doi: 10.3724/SP.J.1146.2009.01426
[5]	李智, 曲长文, 周强, 等. 基于SLIC超像素分割的SAR图像海陆分割算法[J]. 雷达科学与技术, 2017, 15(4): 354–358. doi: 10.3969/j.issn.1672-2337.2017.04.003 LI Zhi, QU Changwen, ZHOU Qiang, et al. A sea-land segmentation algorithm of SAR image based on the SLIC superpixel division[J]. Radar Science and Technology, 2017, 15(4): 354–358. doi: 10.3969/j.issn.1672-2337.2017.04.003
[6]	AMITRANO D, CIERVO F, DI MARTINO G, et al. Modeling watershed response in semiarid regions with high-resolution synthetic aperture radars[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2014, 7(7): 2732–2745. doi: 10.1109/jstars.2014.2313230
[7]	OTSU N. A threshold selection method from gray-level histograms[J]. IEEE Transactions on Systems, Man, and Cybernetics, 1979, 9(1): 62–66. doi: 10.1109/TSMC.1979.4310076
[8]	SUI H G and XU C. Automatic extraction of water in high-resolution SAR images based on multi-scale level set method and Otsu algorithm[C]. Proceedings of International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Melbourne, Australia, 2012: 453–457. doi: 10.5194/isprsarchives-XXXIX-B7-453-2012.
[9]	LIU Chun, YANG Jian, YIN Junjun, et al. Coastline detection in SAR images using a hierarchical level set segmentation[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2016, 9(11): 4908–4920. doi: 10.1109/jstars.2016.2613279
[10]	侯彪, 胡育辉, 焦李成. SAR图像水域的改进Shearlet边缘检测[J]. 中国图象图形学报, 2010, 15(10): 1549–1554. doi: 10.11834/jig.20101019 HOU Biao, HU Yuhui, and JIAO Licheng. Improved shearlet edge detection for waters of SAR images[J]. Journal of Image and Graphics, 2010, 15(10): 1549–1554. doi: 10.11834/jig.20101019
[11]	LIU Zhongling, LI Fei, LI Ning, et al. A novel region-merging approach for coastline extraction from sentinel-1A IW mode SAR imagery[J]. IEEE Geoscience and Remote Sensing Letters, 2016, 13(3): 324–328. doi: 10.1109/LGRS.2015.2510745
[12]	SHELHAMER E, LONG J, and DARRELL T. Fully convolutional networks for semantic segmentation[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(4): 640–651. doi: 10.1109/TPAMI.2016.2572683
[13]	RONNEBERGER O, FISCHER P, and BROX T. U-Net: Convolutional networks for biomedical image segmentation[C]. Proceedings of the 18th International Conference on Medical Image Computing and Computer-assisted Intervention, Munich, Germany, 2015: 234–241. doi: 10.1007/978-3-319-24574-4_28.
[14]	LIN T Y, DOLLÁR P, GIRSHICK R, et al. Feature pyramid networks for object detection[C]. Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, USA, 2017: 936–944. doi: 10.1109/CVPR.2017.106.
[15]	ZHAO Hengshuang, SHI Jianping, QI Xiaojuan, et al. Pyramid scene parsing network[C]. Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, USA, 2017: 6230–6239. doi: 10.1109/CVPR.2017.660.
[16]	CHEN L C, PAPANDREOU G, KOKKINOS I, et al. DeepLab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected CRFs[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018, 40(4): 834–848. doi: 10.1109/TPAMI.2017.2699184
[17]	CHEN L C, PAPANDREOU G, KOKKINOS I, et al. Semantic image segmentation with deep convolutional nets and fully connected CRFs[EB/OL]. arXiv: 1412.7062, 2014.
[18]	张庆君. 高分三号卫星总体设计与关键技术[J]. 测绘学报, 2017, 46(3): 269–277. doi: 10.11947/j.AGCS.2017.20170049 ZHANG Qingjun. System design and key technologies of the GF-3 satellite[J]. Acta Geodaetica et Cartographica Sinica, 2017, 46(3): 269–277. doi: 10.11947/j.AGCS.2017.20170049
[19]	丁赤飚, 刘佳音, 雷斌, 等. 高分三号SAR卫星系统级几何定位精度初探[J]. 雷达学报, 2017, 6(1): 11–16. doi: 10.12000/JR17024 DING Chibiao, LIU Jiayin, LEI Bin, et al. Preliminary exploration of systematic geolocation accuracy of GF-3 SAR satellite system[J]. Journal of Radars, 2017, 6(1): 11–16. doi: 10.12000/JR17024
[20]	CHOLLET F. Xception: Deep learning with depthwise separable convolutions[C]. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, USA, 2017: 1800–1807. doi: 10.1109/CVPR.2017.195.
[21]	YU F and KOLTUN V. Multi-scale context aggregation by dilated convolutions[EB/OL]. arXiv preprint arXiv: 1511.07122, 2015.
[22]	SZEGEDY C, VANHOUCKE V, IOFFE S, et al. Rethinking the inception architecture for computer vision[C]. Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, USA, 2016: 2818–2826. doi: 10.1109/CVPR.2016.308.
[23]	HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Deep residual learning for image recognition[C]. Proceedings of 2016 IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, USA, 2016: 770–778. doi: 10.1109/CVPR.2016.90.
[24]	HUANG Gao, LIU Zhuang, VAN DER MAATEN L, et al. Densely connected convolutional networks[C]. Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, USA, 2017: 2261–2269. doi: 10.1109/CVPR.2017.243.
[25]	NAIR V and HINTON G E. Rectified linear units improve restricted boltzmann machines[C]. Proceedings of the 27th International Conference on International Conference on Machine Learning, Haifa, Israel, 2010: 807–814.
[26]	IOFFE S and SZEGEDY C. Batch normalization: Accelerating deep network training by reducing internal covariate shift[EB/OL]. arXiv preprint arXiv: 1502.03167, 2015.
[27]	MIKOLOV T, SUTSKEVER I, CHEN Kai, et al. Distributed representations of words and phrases and their compositionality[C]. Proceedings of the 26th International Conference on Neural Information Processing Systems, Lake Tahoe, USA, 2013: 3111–3119.
[28]	HANSEN L K and SALAMON P. Neural network ensembles[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1990, 12(10): 993–1001. doi: 10.1109/34.58871
[29]	DENG Jia, DONG Wei, SOCHER R, et al. ImageNet: A large-scale hierarchical image database[C]. Proceedings of 2009 IEEE Conference on Computer Vision and Pattern Recognition, Miami, USA, 2009: 248–255. doi: 10.1109/CVPR.2009.5206848.
[30]	VESE L A and CHAN T F. A multiphase level set framework for image segmentation using the Mumford and shah model[J]. International Journal of Computer Vision, 2002, 50(3): 271–293. doi: 10.1023/a:1020874308076.