基于注意力机制特征融合网络的SAR图像飞机目标快速检测

doi:10.12263/DZXB.20200486

摘要/Abstract

摘要：

针对合成孔径雷达（Synthetic Aperture Radar，SAR）图像中飞机目标散射点离散化程度高，周围背景干扰复杂，现有算法对飞机浅层语义特征表征能力弱等问题，本文提出了基于注意力特征融合网络（Attention Feature Fusion Network，AFFN）的SAR 图像飞机目标检测算法.通过引入瓶颈注意力模块（Bottleneck Attention Module，BAM），本文在AFFN中构建了包含注意力双向特征融合模块（Attention Bidirectional Feature Fusion Module，ABFFM）与注意力传输连接模块（Attention Transfer Connection Block，ATCB）的注意力特征融合策略并合理优化了网络结构，提升了算法对飞机离散化散射点浅层语义特征的提取与判别.基于自建的Gaofen-3 与TerraSAR-X 卫星图像混合飞机目标实测数据集，实验对AFFN与基于深度学习的通用目标检测以及SAR图像特定目标检测算法进行了比较，其结果验证了AFFN对SAR图像飞机目标检测的准确性与高效性.

关键词: 注意力机制, 特征融合, 飞机目标快速检测, SAR图像, 卷积神经网络

Abstract:

Aiming at the problems of high discretization of aircraft’s backscattering points, complex background interference of surroundings in Synthetic Aperture Radar (SAR) images and weak representation of shallow semantic features of aircraft by existing algorithms, an Attention Feature Fusion Network (AFFN) was proposed for aircraft detection in SAR images. By introducing Bottleneck Attention Module (BAM), this article constructed an attention feature fusion strategy consisting of Attention Bidirectional Feature Fusion Module (ABFFM) and Attention Transfer Connection Block (ATCB) in AFFN, and rationally optimized the network structure so as to strengthen the abilities of extracting and discriminating shallow semantic features of aircraft. Based on a self-built Gaofen-3 and TerraSAR-X mixed aircraft dataset, AFFN was compared with several CNN-based general object detection methods and methods designed for specific objects in SAR images. The experimental results illustrated the accuracy and effectiveness of our method for aircraft detection in SAR images

Key words: attention feature fusion network (AFFN), feature fusion, rapid aircraft detection, synthetic aperture radar (SAR) images, convolutional neural network (CNN)

中图分类号:

TN391

赵琰, 赵凌君, 匡纲要. 基于注意力机制特征融合网络的SAR图像飞机目标快速检测[J]. 电子学报, 2021, 49(9): 1665-1674.

Yan ZHAO, Ling-jun ZHAO, Gang-yao KUANG. Attention Feature Fusion Network for Rapid Aircraft Detection in SAR Images[J]. Acta Electronica Sinica, 2021, 49(9): 1665-1674.

图/表 23

参考文献 25

1	张嘉峰, 杨子渊, 张鹏, 等. Fisher纹理分布下基于匹配滤波的极化SAR图像CFAR检测方法[J]. 电子学报, 2019, 47(12): 2533 - 2543.
	Zhang J F, Yang Z Y, Zhang P, et al. A new CFAR detection method of polarimetric SAR imagery based on matched filter under fisher texture[J]. Acta Electronica Sinica, 2019, 47(12): 2533 - 2543.(in Chinese)
2	王彬, 王国宇. 基于广义Gamma分布的高分辨率SAR图像海岸线检测[J]. 电子学报, 2018, 46(4): 827 - 833.
	Wang B, Wang G Y. A coastline detection method in high-resolution SAR images based on the generalized gamma distribution[J]. Acta Electronica Sinica, 2018, 46(4): 827 - 833.(in Chinese)
3	魏志强, 毕海霞, 刘霞. 基于深度卷积神经网络的图上半监督极化SAR图像分类算法[J]. 电子学报, 2020, 48(1): 66 - 74.
	Wei Z Q, Bi H X, Liu X. A graph-based semi-supervised PolSAR image classification method using deep convolutional neural networks[J]. Acta Electronica Sinica, 2020, 48(1): 66 - 74.(in Chinese)
4	Zhao Y, Zhao L J, Xiong B L, et al. Attention receptive pyramid network for ship detection in SAR images[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2020, 13: 2738 - 2756.
5	Hu H, Huang L Q, Yu W X. Aircraft detection for HR SAR images in non-homogeneous background using GGMD-based modeling[J]. Chinese Journal of Electronics, 2019, 28(6): 1271 - 1280.
6	Zhang H, Hao M X, Zhang C, et al. Aircraft target detection algorithm based on high resolution spaceborne SAR imagery[A] MIPPR 2017: Remote Sensing Image Processing, Geographic Information Systems, and Other Applications[C].Xiangyang, China: SPIE,2018:193 - 198.
7	Tan Y H, Li Q Y, Li Y S, et al. Aircraft detection in high-resolution SAR images based on a gradient textural saliency map[J]. Sensors, 2015, 15(9): 23071 - 23094.
8	Dou F, Diao W, Sun X, et al. Aircraft recognition in high resolution SAR images using saliency map and scattering structure features[A] 2016 IEEE International Geoscience and Remote Sensing Symposium(IGARSS)[C]. Beijing, China:IEEE, 2016.1575 - 1578.
9	王思雨, 高鑫, 孙皓, 等. 基于卷积神经网络的高分辨率SAR图像飞机目标检测方法[J]. 雷达学报, 2017, 6(2): 195 - 203.
	Wang S Y, Gao X, Sun H, et al. An aircraft detection method based on convolutional neural networks in high-resolution SAR images[J]. Journal of Radars, 2017, 6(2): 195 - 203.(in Chinese)
10	郭倩, 王海鹏, 徐丰. 星载合成孔径雷达图像的飞机目标检测[J]. 上海航天, 2018, 35(6): 57 - 64.
	Guo Q, Wang H P, Xu F. Aircraft target detection from spaceborne synthetic aperture radar image[J]. Aerospace Shanghai, 2018, 35(6): 57 - 64.(in Chinese)
11	Diao W H, Dou F Z, Fu K, et al. Aircraft detection in sar images using saliency based location regression network [A] 2018 IEEE International Geoscience and Remote Sensing Symposium [C]. Valencia, Spain: IEEE,2018.　2334 - 2337.
12	Lin T Y, Goyal P, Girshick R, et al. Focal loss for dense object detection[A].2017 IEEE International Conference on Computer Vision (ICCV)[C]. Venice, Italy: IEEE, 2017.2999 - 3007.
13	Zhao Y, Zhao L J, Li C Y, et al. Pyramid attention dilated network for aircraft detection in SAR images[J]. IEEE Geoscience and Remote Sensing Letters, 2021, 18(4): 662 - 666.
14	He C, Tu M X, Xiong D H, et al. A component-based multi-layer parallel network for airplane detection in SAR imagery[J]. Remote Sensing, 2018, 10(7): 1016.
15	Zhang S, Wen L, Bian X, et al. Single-shot refinement neural network for object detection[A]. Proceedings of the IEEE conference on computer vision and pattern recognition[C].Salt Lake City, USA: IEEE,2018.4203 - 4212.
16	Ren S Q, He K M, 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.
17	Cai Z W, Vasconcelos N. Cascade R-CNN: Delving into high quality object detection[A].2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) [C]. Salt Lake City, USA: IEEE, 2018. 6154 - 6162.
18	Lin T Y, Dollár P, Girshick R, et al. Feature pyramid networks for object detection[A]. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) [C]. Honolulu, Hawaii: IEEE, 2017. 936 - 944.
19	Liu W, Anguelov D, Erhan D, et al. SSD: Single shot multibox detector[A]. European Conference on Computer Vision (ECCV) [C]. Amsterdam, Netherlands: Springer,2016.21 - 37.
20	Liu S T, Huang D, Wang Y H. Receptive field block net for accurate and fast object detection[A]. Proceedings of the European Conference on Computer Vision (ECCV)[C]. Munich Germany: Springer,2018.385 - 400.
21	Wang T C, Anwer R M, Cholakkal H, et al. Learning rich features at high-speed for single-shot object detection[A]. 2019 IEEE/CVF International Conference on Computer Vision (ICCV)[C]. Seoul, Korea: IEEE, 2019. 1971 - 1980.
22	Redmon J, Farhadi A. YOLOv3: An Incremental Improvement[EB/OL]. , 2018.
23	Cui Z Y, Li Q, Cao Z J, 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.
24	Hu J, Shen L, Sun G. Squeeze-and-excitation networks [A].2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) [C]. Salt Lake City, USA: IEEE, 2018. 7232 - 7141.
25	Woo S, Park J, Lee J Y, et al. CBAM: Convolutional block attention module[A]. Computer Vision⁃ECCV 2018 [C]. Munich, Germany: IEEE,2018.3 - 19.

网络	锚点框尺寸	32，64，128
	锚点框比例	1，1.414，0.707
	r （BAM）	16
训练	迭代总次数（Epochs）	200
	小批次（Minibatch）	12
	初始学习率	5e-4
	优化器	Adam
	权重衰减值	1e-4

网络	锚点框尺寸	32，64，128
	锚点框比例	1，1.414，0.707
	r （BAM）	16
训练	迭代总次数（Epochs）	200
	小批次（Minibatch）	12
	初始学习率	5e-4
	优化器	Adam
	权重衰减值	1e-4

算法类别	网络名称	P	R	F1	AP	Params（M）	FLOPs （G）	FPS
通用目标两阶段算法	Faster-RCNN^［16］	0.736	0.790	0.762	0.749	136.689	149.228	14
	Cascade-RCNN^［17］	0.707	0.733	0.720	0.697	304.076	179.934	8
	FPN^［18］	0.778	0.810	0.794	0.814	120.690	134.150	16
通用目标单阶段算法	SSD^［19］	0.798	0.778	0.788	0.820	23.801	30.570	250
	RFB^［20］	0.854	0.764	0.806	0.838	31.471	34.422	100
	LRF^［21］	0.899	0.740	0.812	0.809	65.774	48.114	91
	YOLOv3^［22］	0.502	0.791	0.614	0.630	61.529	32.687	15
	RefineDet^［15］	0.856	0.806	0.830	0.843	33.915	37.426	90
	RetinaNet^［12］	0.837	0.764	0.799	0.789	135.366	111.994	28
SAR图像特定目标检测算法	PADN^［13］	0.820	0.777	0.798	0.806	102.163	123.562	33
	DAPN^［23］	0.789	0.830	0.809	0.831	128.230	167.733	12
	文献［9］	0.520	0.315	0.395	0.189	4.314	0.328	0.5
	文献［10］	0.544	0.238	0.331	0.132	6.158	11.661	0.5
	文献［11］	0.706	0.235	0.353	0.200	134.309	37.279	0.4
	文献［14］	0.644	0.807	0.716	0.741	102.357	97.549	2
本文算法	AFFN	0.856	0.873	0.864	0.871	32.190	39.215	91

算法类别	网络名称	P	R	F1	AP	Params（M）	FLOPs （G）	FPS
通用目标两阶段算法	Faster-RCNN^［16］	0.736	0.790	0.762	0.749	136.689	149.228	14
	Cascade-RCNN^［17］	0.707	0.733	0.720	0.697	304.076	179.934	8
	FPN^［18］	0.778	0.810	0.794	0.814	120.690	134.150	16
通用目标单阶段算法	SSD^［19］	0.798	0.778	0.788	0.820	23.801	30.570	250
	RFB^［20］	0.854	0.764	0.806	0.838	31.471	34.422	100
	LRF^［21］	0.899	0.740	0.812	0.809	65.774	48.114	91
	YOLOv3^［22］	0.502	0.791	0.614	0.630	61.529	32.687	15
	RefineDet^［15］	0.856	0.806	0.830	0.843	33.915	37.426	90
	RetinaNet^［12］	0.837	0.764	0.799	0.789	135.366	111.994	28
SAR图像特定目标检测算法	PADN^［13］	0.820	0.777	0.798	0.806	102.163	123.562	33
	DAPN^［23］	0.789	0.830	0.809	0.831	128.230	167.733	12
	文献［9］	0.520	0.315	0.395	0.189	4.314	0.328	0.5
	文献［10］	0.544	0.238	0.331	0.132	6.158	11.661	0.5
	文献［11］	0.706	0.235	0.353	0.200	134.309	37.279	0.4
	文献［14］	0.644	0.807	0.716	0.741	102.357	97.549	2
本文算法	AFFN	0.856	0.873	0.864	0.871	32.190	39.215	91

网络名称	P	R	F1	AP	Params（M）	FLOPs（G）	FPS
AFFN	0.856	0.873	0.864	0.871	32.190	39.215	91
AFFN-ATCB	0.850	0.869	0.859	0.870	32.172	39.201	95
AFFN-ABFFM	0.842	0.886	0.863	0.871	30.402	38.504	95
AFFN（TCB）	0.840	0.857	0.848	0.865	30.385	38.490	100
RefineDet	0.856	0.806	0.830	0.843	33.915	37.426	90