基于卷积神经网络与注意力机制的SAR图像飞机检测

doi:10.12305/j.issn.1001-506X.2021.11.20

摘要/Abstract

摘要：

在合成孔径雷达(synthetic aperture radar, SAR)图像应用领域, 对SAR图像中飞机目标的检测备受关注。针对现有检测算法模型运算复杂度高、检测性能较低的问题, 提出一种基于深度可分离卷积神经网络与注意力机制的SAR图像飞机检测算法。首先使用深度可分离卷积神经网络提取图像特征, 同时在网络中引入逆残差块, 以有效防止通道数压缩引起的特征信息丢失问题; 其次在网络中引入多尺度空洞卷积—空间注意力模块和全局上下文通道注意力模块, 通过重新分配显著区域和各特征图更有代表性的权值, 以更好地捕捉空间有效信息和通道间语义相关性, 提高模型特征表达能力; 最后在SAR飞机数据集(SAR aircraft dataset, SAD)上进行对比实验验证。实验结果表明, 所提算法具有更好的检测效果, 平均准确率达到86.3%, 检测速度达到22.4 fps/s。

关键词: 合成孔径雷达图像, 飞机检测, 深度可分离卷积, 注意力机制

Abstract:

In the application field of synthetic aperture radar (SAR) images, the detection of aircraft target in SAR images has attracted much attention. Aiming at the problems of high computational complexity and low detection performance of existing detection algorithm models, an aircraft detection in SAR images algorithm based on depthwise separable convolutional neural network and attention mechanism is proposed. Firstly, the depthwise separable convolutional neural network is used to extract image features, and the inverted residual block is introduced into the network to effectively reduce the loss of feature information caused by channel slimming. Secondly, the multi-scale dilated convolution spatial attention mechanism module and global context channel attention mechanism module are introduced into the network. By redistributing more representative weights to salient regions and each feature map, the spatial effective information and semantic correlation between channels can be better captured, and the ability of feature expression can be enhanced. Finally, the comparative experimental verification is carried out on the SAR aircraft dataset (SAD). Experimental results show that this algorithm has a better detection effect, the average precision is 86.3%, and the detection speed is 22.4 fps/s.

Key words: synthetic aperture radar (SAR) images, aircraft detection, depthwise separable convolution, attention mechanism

中图分类号:

TP751

李广帅, 苏娟, 李义红, 李响. 基于卷积神经网络与注意力机制的SAR图像飞机检测[J]. 系统工程与电子技术, 2021, 43(11): 3202-3210.

Guangshuai LI, Juan SU, Yihong LI, Xiang LI. Aircraft detection in SAR images based on convolutional neural network and attention mechanism[J]. Systems Engineering and Electronics, 2021, 43(11): 3202-3210.

图/表 15

图1

图2

图3

图4

图5

图6

图7

图8

表1

图9

表2

表3

图10

表4

表5

参考文献 31

1	FU K , DOU F , LI H , et al. Aircraft recognition in SAR Images based on scattering structure feature and template matching[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2018, 11 (11): 4206- 4217. doi: 10.1109/JSTARS.2018.2872018
2	HE C , TU M , LIU X , et al. Mixture statistical distribution based multiple component model for target detection in high resolution SAR imagery[J]. ISPRS international journal of geo-information, 2017, 6 (11): 336. doi: 10.3390/ijgi6110336
3	高君, 高鑫, 孙显. 基于几何特征的高分辨率SAR图像飞机目标解译方法[J]. 国外电子测量技术, 2015, 34 (8): 21- 28. doi: 10.3969/j.issn.1002-8978.2015.08.008
	GAO J , GAO X , SUN X . Translation of high-resolution SAR images based on geometric features[J]. ] Aircraft Targets Foreign Electronic Measurement Technology, 2015, 34 (8): 21- 28. doi: 10.3969/j.issn.1002-8978.2015.08.008
4	GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]//Proc. of the Computer Vision and Pattern Recognition, 2014: 580-587.
5	GIRSHICK R. Fast R-CNN[C]//Proc. of the International Conference on Computer Vision, 2015: 1440-1448.
6	REN S , HE K , GIRSHICK R , et al. Faster R-CNN: towards real-time object detection with region proposal networks[J]. IEEE Trans. on Pattern Analysis and Machine Intelligence, 2015, 39 (6): 1137- 1149.
7	HONG S, ROH B, KIM K H, et al. PVANet: lightweight deep neural networks for real-time object detection [EB/OL]. [2020-09-10]. https://arxiv.org/abs/1611.08588v2.
8	HE K, GKIOXARI G, DOLLAR P, et al. Mask R-CNN[C]//Proc. of the International Conference on Computer Vision, 2017: 2980-2988.
9	LIU W, ANGUELOV D, ERHAN D, et al. SSD: single shot multibox detector[C]//Proc. of the European Conference on Computer Vision, 2016: 21-37.
10	FU C Y, LIU W, RANGA A, et al. DSSD: deconvolutional single shot detector [EB/OL]. [2020-09-10]. https://arxiv.org/abs/1701.06659.
11	LI Z X, ZHOU F Q. FSSD: feature fusion single shot multibox detector[EB/OL]. [2020-09-10]. https://arxiv.org/abs/1712.00960v3.
12	REDMON J, DIVVALA S K, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]//Proc. of the Computer Vision and Pattern Recognition, 2016: 779-788.
13	REDMON J, FARHADI A. YOLO9000: better, faster, stronger[C]//Proc. of the Computer Vision and Pattern Recognition, 2017: 6517-6525.
14	REDMON J, FARHADI A. YOLOv3: an incremental improvement[EB/OL]. [2020-09-10]. https://arxiv.org/abs/1804.02767v1.
15	KONG T, SUN F, YAO A, et al. RON: reverse connection with objectness prior networks for object detection[C]//Proc. of the Computer Vision and Pattern Recognition, 2017: 5244-5252.
16	DENG J, DONG W, SOCHER R, et al. ImageNet: a large-scale hierarchical image database[C]//Proc. of the Computer Vision and Pattern Recognition, 2009: 248-255.
17	LIN T, MAIRE M, BELONGIE S, et al. Microsoft COCO: common objects in context[C]//Proc. of the European Confe-rence on Computer Vision, 2014: 740-755.
18	HE C , TU M , XIONG D , et al. Adaptive component selection-based discriminative model for object detection in high-reso-lution SAR imagery[J]. ISPRS International Journal of Geo-Information, 2018, 7 (2): 72. doi: 10.3390/ijgi7020072
19	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.
20	ZHANG L B , LI C Y , ZHAO L J , et al. A cascaded three-look network for aircraft detection in SAR images[J]. Remote Sensing Letters, 2020, 11 (1): 57- 65. doi: 10.1080/2150704X.2019.1681599
21	SHAO J , QU C , LI J , et al. A lightweight convolutional neural network based on visual attention for SAR image target classification[J]. Sensors, 2018, 18 (9): 3039. doi: 10.3390/s18093039
22	CHOLLET F. Xception: deep learning with depthwise separable convolutions[C]//Proc. of the Computer Vision and Pattern Recognition, 2017: 1800-1807.
23	ZHANG X Q , ZHENG Y G , LIU W K , et al. An improved architecture for urban building extraction based on depthwise separable convolution[J]. Journal of Intelligent and Fuzzy Systems, 2020, 38 (5): 5821- 5829. doi: 10.3233/JIFS-179669
24	SANDLER M, HOWARD A, ZHU M, et al. MobileNetV2: inverted residuals and linear bottlenecks[C]//Proc. of the Computer Vision and Pattern Recognition, 2018: 4510-4520.
25	WOO S, PARK J, LEE J, et al. CBAM: convolutional block attention module[C]//Proc. of the European Conference on Computer Vision, 2018: 3-19.
26	HU J , SHEN L , ALBANIE S , et al. Squeeze-and-excitation networks[J]. IEEE Trans. on Pattern Analysis and Machine Intelligence, 2019, 1- 1.
27	CAO Y, XU J R, LIN S, et al. Gcnet: non-local networks meet squeeze-excitation networks and beyond[C]//Proc. of the IEEE/CVF International Conference on Computer Vision Workshops, 2019.
28	DIAO W H, DOU F Z, FU K, et al. Aircraft detection in sar images using saliency based location regression network[C]//Proc. of the International Geoscience and Remote Sensing Symposium, 2018: 2334-2337.
29	LI J , JIN K , ZHOU D , et al. Attention mechanism-based CNN for facial expression recognition[J]. Neuro-computing, 2020, 411, 340- 350.
30	陈猛夫. 基于迁移学习的暴恐图像自动识别研究[EB/OL]. [2020-09-14]. http://10.13700/j.bh.1001-5965.2020.0046.
	CHEN M F. Research on automatic recognition model for terrorism related image based on transfer learning[EB/OL]. [2020-09-14]. http://10.13700/j.bh.1001-5965.2020.0046.
31	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[EB/OL]. [2020-07-22]. https://doi.org/10.13700/j.bh.1001-5965.2020.0004.
	LI G S, SU J, LI Y H. An aircraft detection algorithm in SAR image based on improved faster R-CNN[EB/OL]. [2020-07-22]. https://doi.org/10.13700/j.bh.1001-5965.2020.0004.

方法	C₃∶113×75	C₄∶57×38	C₅∶29×19
原始Faster R-CNN
融合注意力机制

方法	AP/%	R/%	P/%	F-score/%	mIoU/%	FPS
Faster R-CNN-1	79.8	79.9	73.1	76.3	71.3	16.8
Faster R-CNN-2	81.7	82.9	75.8	79.2	71.9	16.2
Faster R-CNN-3	82.8	83.8	76.7	80.1	72.5	13.7
本文方法	86.3	88.6	81.0	84.6	74.4	22.4

方法	参数量/×10⁶	模型大小/MB	训练时长/h
Faster R-CNN-1	—	1 092.5	—
Faster R-CNN-2	61.31	409.8	5.74
Faster R-CNN-3	80.30	561.9	7.07
本文方法	38.55	225.0	4.45

序号	算法				AP/%	FPS
序号	DS-Conv	MDC-SA	GC	W	AP/%	FPS
1	—	—	—	—	81.7	16.2
2	√	—	—	—	80.2	24.3
3	√	—	—	√	81.9	24.3
4	√	√	—	—	83.8	23.1
5	√	—	√	—	83.1	23.8
6	√	√	√	—	85.2	22.4
7	√	√	√	√	86.3	22.4

方法	密集分布	独立分布	大幅面场景	复杂场景
标注框
Faster R-CNN-1
Faster R-CNN-2
Faster R-CNN-3
本文方法