多尺度特征融合与特征通道关系校准的SAR图像船舶检测

周雪珂; 刘畅; 周滨

doi:10.12000/JR21021

多尺度特征融合与特征通道关系校准的SAR图像船舶检测

doi: 10.12000/JR21021

周雪珂^{1, 2, ,},
刘畅^2,,
周滨^{1, 2}

1.
中国科学院大学北京 100049
2.
中国科学院空天信息创新研究院北京 100190

基金项目: 国家重点研发计划(2017YFB0503001)

详细信息

作者简介:
周雪珂(1996–)，女，陕西西安人，中国科学院大学硕士研究生，研究方向为 SAR图像处理、机器学习。

刘　畅(1978–)，男，山东烟台人，研究员，博士生导师。2006年在中国科学院电子学研究所获得博士学位，现担任中国科学院空天信息创新研究院研究员、博士生导师。主要研究方向为SAR系统及其相关SAR成像处理。

周　滨(1994–)，男，江西上饶人，中国科学院大学硕士研究生。研究方向为 SAR 图像处理、目标检测。

通讯作者:
周雪珂 zhouxk96@163.com

责任主编：刘涛 Corresponding Editor: LIU Tao
中图分类号: TN957.52
计量
- 文章访问数: 2888
- HTML全文浏览量: 1102
- PDF下载量: 373
- 被引次数: 0
出版历程
- 收稿日期: 2021-03-04
- 修回日期: 2021-04-05
- 网络出版日期: 2021-08-28

Ship Detection in SAR Images Based on Multiscale Feature Fusion and Channel Relation Calibration of Features

ZHOU Xueke^{1, 2
, ,},
LIU Chang^2
,,
ZHOU Bin^{1, 2}

1.
University of Chinese Academy of Sciences, Beijing 100049, China
2.
Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100190, China

Funds: The State Key Research Development Program of China (2017YFB0503001)

More Information

Corresponding author: ZHOU Xueke, zhouxk96@163.com

摘要

摘要: 目前深度学习技术在SAR图像的船舶检测中已取得显著的成果，但针对SAR船舶图像中复杂多变的背景环境，如何准确高效地提取目标特征，提升检测精度与检测速度仍存在着巨大的挑战。针对上述问题，该文提出了一种多尺度特征融合与特征通道关系校准的 SAR 图像船舶检测算法。在Faster R-CNN的基础上，首先通过引入通道注意力机制对特征提取网络进行特征间通道关系校准，提高网络对复杂场景下船舶目标特征提取的表达能力；其次，不同于原始的基于单一尺度特征生成候选区域的方法，该文基于神经架构搜索算法引入改进的特征金字塔结构，高效地将多尺度特征进行充分融合，改善了船舶目标中对小目标、近岸密集目标的漏检问题。最后，在SSDD数据集上进行对比验证。实验结果表明，相较原始的Faster R-CNN，检测精度从85.4%提高到89.4%，检测速率也从2.8 FPS提高到10.7 FPS。该方法能够有效实现高速与高精度的SAR图像船舶检测，具有一定的现实意义。
- 合成孔径雷达 /
- Faster R-CNN /
- 船舶检测 /
- 特征融合 /
- 通道注意力
Abstract: Deep-learning technology has enabled remarkable results for ship detection in SAR images. However, in view of the complex and changeable backgrounds of SAR ship images, how to accurately and efficiently extract target features and improve detection accuracy and speed is still a huge challenge. To solve this problem, a ship detection algorithm based on multiscale feature fusion and channel relation calibration of features is proposed in this paper. First, based on Faster R-CNN, a channel attention mechanism is introduced to calibrate the channel relationship between features in the feature extraction network, so as to improve the network’s expression ability for extraction of ship features in different scenes. Second, unlike the original method of generating candidate regions based on single-scale features, this paper introduces an improved feature pyramid structure based on a neural architecture search algorithm, which helps improve the performance of the network. The multiscale features are effectively fused to settle the problem of missing detections of small targets and adjacent inshore targets. Experimental results on the SSDD dataset show that, compared with the original Faster R-CNN, the proposed algorithm improves detection accuracy from 85.4% to 89.4% and the detection rate from 2.8 FPS to 10.7 FPS. Thus, this method effectively achieves high-speed and high-accuracy SAR ship detection, which has practical benefits.
- SAR /
- Faster R-CNN /
- Ship detection /
- Features fusion /
- Channel attention

HTML全文

图 1 Faster R-CNN结构图

Figure 1. The frame structure of Faster R-CNN

下载: 全尺寸图片幻灯片

图 2 Resnet50的残差结构

Figure 2. The residual structure of Resnet50

下载: 全尺寸图片幻灯片

图 3 区域建议网络结构

Figure 3. The structure of region proposal network

下载: 全尺寸图片幻灯片

图 4 本文算法网络结构

Figure 4. The network structure of the algorithm in this paper

下载: 全尺寸图片幻灯片

图 5 RoI Align的实现原理

Figure 5. Implementation principle of RoI Align

下载: 全尺寸图片幻灯片

图 6 通道注意力模块结构

Figure 6. The structure of channel attention module

下载: 全尺寸图片幻灯片

图 7 NAS-FPN强化学习算法

Figure 7. Reinforcement learning for NAS-FPN

下载: 全尺寸图片幻灯片

图 8 FPN组合结构

Figure 8. The combination structure of FPN

下载: 全尺寸图片幻灯片

图 9 NAS-FPN结构图

Figure 9. The frame structure of NAS-FPN

下载: 全尺寸图片幻灯片

图 10 融合操作

Figure 10. Binary operation

下载: 全尺寸图片幻灯片

图 11 NAS-FPN热力图结果

Figure 11. The heatmaps of NAS-FPN

下载: 全尺寸图片幻灯片

图 12 不同算法的P-R曲线对比

Figure 12. The P-R curve of different methods

下载: 全尺寸图片幻灯片

图 13 小目标船舶图像的检测算法对比

Figure 13. Comparison of detection algorithms for small target

下载: 全尺寸图片幻灯片

图 14 近岸船舶图像的检测算法对比

Figure 14. Comparison of detection algorithms for inshore ship

下载: 全尺寸图片幻灯片

图 15 密集停靠的船舶图像检测算法对比

Figure 15. Comparison of detection algorithms for adjacent targets

下载: 全尺寸图片幻灯片

图 16 本文算法在AIR-SARShip 1.0数据上的检测结果

Figure 16. Detection result of this algorithm on AIR-SARShip 1.0

下载: 全尺寸图片幻灯片

表 1 以Resnet50作为主干网络的特征提取网络参数

Table 1. The network parameters extraction with Resnet50 as the backbone network feature

网络层名称	类型	Resnet50	SE-Resnet50
网络层名称	类型	卷积核(高度×宽度×通道数)/步长	卷积核(高度×宽度×通道数)/步长
Conv1	卷积层	$7 \times 7 \times 64/2$	$7 \times 7 \times 64/2$
max pool	池化层	$3 \times 3 \times 64/2$	$3 \times 3 \times 64/2$
Conv2_1—Conv2_9	残差结构	$\left[ {\begin{array}{*{20}{c}} {1 \times 1 \times 64/1} \\ {3 \times 3 \times 64/1} \\ {1 \times 1 \times 256/1} \end{array}} \right] \times 3$	$\left[ {\begin{array}{*{20}{c} } {1 \times 1 \times 64/1} \\ {3 \times 3 \times 64/1} \\ {1 \times 1 \times 256/1} \\ { {{{\rm{fc}}} },[16,256]} \end{array} } \right] \times 3$
Conv3_1—Conv3_12	残差结构	$\left[ {\begin{array}{*{20}{c}} {1 \times 1 \times 128/2} \\ {3 \times 3 \times 128/1} \\ {1 \times 1 \times 512/1} \end{array}} \right] \times 4$	$\left[ {\begin{array}{*{20}{c} } {1 \times 1 \times 128/2} \\ {3 \times 3 \times 128/1} \\ {1 \times 1 \times 512/1} \\ {{\rm{fc}},[32,512]} \end{array} } \right] \times 4$
Conv4_1—Conv4_18	残差结构	$\left[ {\begin{array}{*{20}{c}} {1 \times 1 \times 256/2} \\ {3 \times 3 \times 256/1} \\ {1 \times 1 \times 1024/1} \end{array}} \right] \times 6$	$\left[ {\begin{array}{*{20}{c} } {1 \times 1 \times 256/2} \\ {3 \times 3 \times 256/1} \\ {1 \times 1 \times 1024/1} \\ {{\rm{fc}},[64,1024]} \end{array} } \right] \times 6$
Conv5_1—Conv5_9	残差结构	$\left[ {\begin{array}{*{20}{c}} {1 \times 1 \times 512/1} \\ {3 \times 3 \times 512/1} \\ {1 \times 1 \times 2048/1} \end{array}} \right] \times 3$	$\left[ {\begin{array}{*{20}{c} } {1 \times 1 \times 512/1} \\ {3 \times 3 \times 512/1} \\ {1 \times 1 \times 2048/1} \\ {{\rm{fc}},[128,{\rm{2048} }]} \end{array} } \right] \times 3$

下载: 导出CSV

表 2 基于Faster R-CNN的优化算法对比

Table 2. Comparison of optimization algorithms based on Faster R-CNN

RoIAlign	CA	NAS-FPN	AP (%)	Time (s/iter)	Speed (FPS)
			85.4	0.667	2.80
√			87.2	0.727	2.78
√	√		88.2	0.741	2.67
√		√	88.0	0.493	10.72
√	√	√	89.4	0.535	10.70

下载: 导出CSV

表 3 不同检测算法的性能对比

Table 3. Comparison of different detection algorithms

Method	Image size	FLOPs (2³⁰)	Params (2²⁰)	Time (s/iter)	Speed (FPS)	SSDD	HRSID
Method	Image size	FLOPs (2³⁰)	Params (2²⁰)	Time (s/iter)	Speed (FPS)	AP (%)	AP (%)
SSD	300×300	30.49	23.75	0.061	47.20	84.7	79.6
Cascade R-CNN	300×300	59.03	68.93	0.323	13.20	88.4	80.9
PANet	300×300	59.03	68.93	0.301	14.70	88.7	81.3
本文算法	300×300	33.66	70.27	0.535	10.70	89.4	82.8

下载: 导出CSV

表 4 不同检测算法基于SSDD在近岸与离岸场景下的性能对比

Table 4. Comparison of different detection algorithms in inshore and offshore scenes of SSDD

Method	Inshore		Offshore
Method	AP (%)	R (%)	AP (%)	R (%)
SSD	73.6	92.7	88.1	95.7
Cascade R-CNN	73.7	90.7	90.4	95.0
PANet	73.7	88.0	90.7	93.6
本文算法	74.3	90.7	90.7	94.1

下载: 导出CSV

参考文献(28)

[1]	MOREIRA A, PRATS-IRAOLA P, YOUNIS M, et al. A tutorial on synthetic aperture radar[J]. IEEE Geoscience and Remote Sensing Magazine, 2013, 1(1): 6–43. doi: 10.1109/MGRS.2013.2248301
[2]	郭倩, 王海鹏, 徐丰. SAR图像飞机目标检测识别进展[J]. 雷达学报, 2020, 9(3): 497–513. doi: 10.12000/JR20020 GUO Qian, WANG Haipeng, and XU Feng. Research progress on aircraft detection and recognition in SAR imagery[J]. Journal of Radars, 2020, 9(3): 497–513. doi: 10.12000/JR20020
[3]	WACKERMAN C C, FRIEDMAN K S, PICHEL W G, et al. Automatic detection of ships in RADARSAT-1 SAR imagery[J]. Canadian Journal of Remote Sensing, 2001, 27(5): 568–577. doi: 10.1080/07038992.2001.10854896
[4]	陈慧元, 刘泽宇, 郭炜炜, 等. 基于级联卷积神经网络的大场景遥感图像舰船目标快速检测方法[J]. 雷达学报, 2019, 8(3): 413–424. doi: 10.12000/JR19041 CHEN Huiyuan, LIU Zeyu, GUO Weiwei, et al. Fast detection of ship targets for large-scale remote sensing image based on a cascade convolutional neural network[J]. Journal of Radars, 2019, 8(3): 413–424. doi: 10.12000/JR19041
[5]	李健伟, 曲长文, 邵嘉琦, 等. 基于深度学习的SAR图像舰船检测数据集及性能分析[C]. 第五届高分辨率对地观测学术年会论文集, 西安, 2018: 180–201. LI Jianwei, QU Changwen, SHAO Jiaqi, et al. Dataset and performance analysis of ship detection methods based on deep learning[C]. The 5th China High Resolution Earth Observation Conference, Xi’an, China, 2018: 180–201.
[6]	GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]. 2014 IEEE Conference on Computer Vision and Pattern Recognition, Columbus, USA, 2014: 580–587. doi: 10.1109/CVPR.2014.81.
[7]	GIRSHICK R. Fast R-CNN[C]. IEEE International Conference on Computer Vision (ICCV), Santiago, Chile, 2015: 1440–1448. doi: 10.1109/ICCV.2015.169.
[8]	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
[9]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: Unified, real-time object detection[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, USA, 2016: 779–788. doi: 10.1109/CVPR.2016.91.
[10]	LIU Wei, ANGUELOV D, ERHAN D, et al. SSD: Single shot multibox detector[C]. The 14th European Conference on Computer Vision, Amsterdam, The Netherlands, 2016: 21–37. doi: 10.1007/978-3-319-46448-0_2.
[11]	李健伟, 曲长文, 彭书娟. 基于级联CNN的SAR图像舰船目标检测算法[J]. 控制与决策, 2019, 34(10): 2191–2197. doi: 10.13195/j.kzyjc.2018.0168 LI Jianwei, QU Changwen, and PENG Shujuan. A ship detection method based on cascade CNN in SAR images[J]. Control and Decision, 2019, 34(10): 2191–2197. doi: 10.13195/j.kzyjc.2018.0168
[12]	李广帅, 苏娟, 李义红. 基于改进Faster R-CNN的SAR图像飞机检测算法[J]. 北京航空航天大学学报, 2021, 47(1): 159–168. doi: 10.13700/j.bh.1001-5965.2020.0004 LI Guangshuai, SU Juan, and LI Yihong. An aircraft detection algorithm in SAR image based on improved Faster R-CNN[J]. Journal of Beijing University of Aeronautics and Astronautics, 2021, 47(1): 159–168. doi: 10.13700/j.bh.1001-5965.2020.0004
[13]	WANG Rui, SHAO Sihan, AN Mengyu, et al. Soft thresholding attention network for adaptive feature denoising in SAR ship detection[J]. IEEE Access, 2021, 9: 29090–29105. doi: 10.1109/ACCESS.2021.3059033
[14]	ZHANG Tianwen and ZHANG Xiaoling. High-speed ship detection in SAR images based on a grid convolutional neural network[J]. Remote Sensing, 2019, 11(10): 1206. doi: 10.3390/rs11101206
[15]	ZHANG Tianwen, ZHANG Xiaoling, SHI Jun, et al. Depthwise separable convolution neural network for high-speed SAR ship detection[J]. Remote Sensing, 2019, 11(21): 2483. doi: 10.3390/rs11212483
[16]	ZHANG Tianwen and ZHANG Xiaoling. ShipDeNet-20: An only 20 convolution layers and <1-MB lightweight SAR ship detector[J]. IEEE Geoscience and Remote Sensing Letters, in press.
[17]	ZHANG Tianwen, ZHANG Xiaoling, SHI Jun, et al. HyperLi-Net: A hyper-light deep learning network for high-accurate and high-speed ship detection from synthetic aperture radar imagery[J]. ISPRS Journal of Photogrammetry and Remote Sensing, 2020, 167: 123–153. doi: 10.1016/j.isprsjprs.2020.05.016
[18]	GHIASI G, LIN T Y, and LE Q V. NAS-FPN: Learning scalable feature pyramid architecture for object detection[C]. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Long Beach, USA, 2019: 7029–7038. doi: 10.1109/CVPR.2019.00720.
[19]	LIN T Y, DOLLÁR P, GIRSHICK R, et al. Feature pyramid networks for object detection[C]. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, USA, 2017: 936–944, doi: 10.1109/CVPR.2017.106.
[20]	HE Kaiming, GKIOXARI G, DOLLÁR P, et al. Mask R-CNN[C]. 2017 IEEE International Conference on Computer Vision (ICCV), Venice, Italy, 2017: 2980–2988. doi: 10.1109/ICCV.2017.322.
[21]	BODLA N, SINGH B, CHELLAPPA R, et al. Soft-NMS—Improving object detection with one line of code[C]. 2017 IEEE International Conference on Computer Vision (ICCV), Venice, Italy, 2017: 5562–5570. doi: 10.1109/ICCV.2017.593.
[22]	HU Jie, LI Shen, GANG Sun, et al. Squeeze-and-excitation networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020, 42(8): 2011–2023. doi: 10.1109/TPAMI.2019.2913372
[23]	ZHANG Tianwen, ZHANG Xiaoling, KE Xiao, et al. LS-SSDD-v1.0: A deep learning dataset dedicated to small ship detection from large-scale sentinel-1 SAR images[J]. Remote Sensing, 2020, 12(18): 2997. doi: 10.3390/rs12182997
[24]	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
[25]	孙显, 王智睿, 孙元睿, 等. AIR-SARShip-1.0: 高分辨率SAR舰船检测数据集[J]. 雷达学报, 2019, 8(6): 852–862. doi: 10.12000/JR19097 SUN Xian, WANG Zhirui, SUN Yuanrui, et al. AIR-SARShip-1.0: High-resolution SAR ship detection dataset[J]. Journal of Radars, 2019, 8(6): 852–862. doi: 10.12000/JR19097
[26]	CAI Zhaowei and VASCONCELOS N. Cascade R-CNN: Delving into high quality object detection[C]. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, USA, 2018: 6154–6162, doi: 10.1109/CVPR.2018.00644.
[27]	LIU Shu, QI Lu, QIN Haifang, et al. Path aggregation network for instance segmentation[C]. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, Salt Lake City, USA, 2018: 8759–8768. doi: 10.1109/CVPR.2018.00913.
[28]	ZHANG Tianwen, ZHANG Xiaoling, SHI Jun, et al. Balance scene learning mechanism for offshore and inshore ship detection in SAR images[J]. IEEE Geoscience and Remote Sensing Letters, in press.