改进YOLOv5的合成孔径雷达图像舰船目标检测方法

doi:10.12305/j.issn.1001-506X.2023.12.04

摘要/Abstract

摘要：

针对合成孔径雷达图像目标检测易受噪声和背景干扰影响, 以及多尺度条件下检测性能下降的问题, 在兼顾网络规模和检测精度的基础上, 提出了一种改进的合成孔径雷达舰船目标检测算法。使用坐标注意力机制, 在确保轻量化的同时抑制了噪声与干扰, 以提高网络的特征提取能力; 融入加权双向特征金字塔结构以实现多尺度特征融合, 设计了一种新的预测框损失函数以改善检测精度, 同时加快算法收敛, 从而实现了对合成孔径雷达图像舰船目标的快速准确识别。实验验证表明, 所提算法在合成孔径雷达舰船检测数据集(synthetic aperture radar ship detection dataset, SSDD)上的平均精度均值达到96.7%, 相比于YOLOv5s提高1.9%, 训练时收敛速度更快, 且保持了网络轻量化的特点, 在实际应用中具有良好前景。

关键词: 合成孔径雷达, 目标检测, YOLOv5, 注意力机制, 多尺度融合

Abstract:

Aiming at the problem that target detection in synthetic aperture radar (SAR) images is easily affected by noise and background interference, and the performance of ship target detection is degraded under multi-scale conditions, an improved YOLOv5 algorithm is proposed on the basis of considering the network scale and detection accuracy. In this algorithm, coordinate attention mechanism is used to suppress noise and interference to improve the feature extraction ability of the network while ensuring its lightweight advantage. The bi-directional feature pyramid is integrated to achieve multi-scale feature fusion. A new prediction box loss function is designed to improve the detection accuracy and accelerate the convergence of the algorithm. Thus, the ship target can be recognized quickly and accurately in SAR images. Experimental verification shows that the mean average presicion (mAP) of the proposed algorithm on SSDD dataset reaches 96.7%, which is 1.9% higher than that of YOLOv5s. The convergence speed is faster during training, and the network is lightweight, which has a good prospect in practical application.

Key words: synthetic aperture radar, target detection, YOLOv5, attention mechanism, multi-scale fusion

中图分类号:

TP391

贺翥祯, 李敏, 苟瑶, 杨爱涛. 改进YOLOv5的合成孔径雷达图像舰船目标检测方法[J]. 系统工程与电子技术, 2023, 45(12): 3743-3753.

Zhuzhen HE, Min LI, Yao GOU, Aitao YANG. Ship target detection method for synthetic aperture radar images based on improved YOLOv5[J]. Systems Engineering and Electronics, 2023, 45(12): 3743-3753.

图/表 20

图1

图2

图3

图4

图5

图6

图7

图8

图9

表1

表2

表3

图10

图11

图12

表4

表5

表6

图13

图14

参考文献 31

1	谌华. SAR图像目标自动检测与识别方法研究[D]. 北京: 中国科学院大学, 2019.
	ZHAN H. A study on methods of SAR image target automatic detection and recognition[D]. Beijing: University of Chinese Academy of Sciences, 2019.
2	NOVAK L M , OWIRKA G J , BROWER W S , et al. The automatic target-recognition system in SAIP[J]. Lincoln Laboratory Journal, 1997, 10 (2): 187- 202.
3	NOVAK L M , HALVERSEN S D , OWIRKA G , et al. Effects of polarization and resolution on SAR ATR[J]. IEEE Trans.on Aerospace and Electronic Systems, 1997, 33 (1): 102- 116. doi: 10.1109/7.570713
4	ROBERTSON N, BIRD P, BROWNSWORD C. Ship surveillance using RADARSAT ScanSAR images[C]//Proc. of the Ship Detection in Coastal Water Workshop Agenda, 2000: 41-45.
5	GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2014: 580-587.
6	GIRSHICK R. Fast R-CNN[EB/OL]. [2022-8-20]. https://arxiv.org/pdf/1504.08083.pdf.
7	REN S Q , HE K M , 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, 2017, 39 (6): 1137- 1149. doi: 10.1109/TPAMI.2016.2577031
8	LIU W, ANGUELOV D, ERHAN D, et al. SSD: single shot multi-box detector[C]//Proc. of the European Conference on Computer Vision, 2016: 21-37.
9	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: unified, real-time object detection[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2016: 779-788.
10	REDMON J, FARHADI A. YOLO9000: better, faster, stronger[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2017: 6517-6525.
11	REDMON J, FARHADI A. YOLOv3: an incremental improvement[EB/OL]. [2022-08-20]. https://arxiv.org/pdf/1804.02767v1.pdf.
12	BOCHKOVSKIV A, WANG C Y, LIAO H. YOLOv4: optimal speed and accuracy of object detection[EB/OL]. [2022-08-20]. https://arxiv.org/abs/2004.10934.
13	赵江洪, 张晓光, 杨璐. 深度学习的遥感影像舰船目标检测[J]. 测绘科学, 2020, 45 (3): 110- 116.
	ZHAO J H , ZHANG X G , YANG L . Ship detection in remote sensing based on deep learning[J]. Science of Surveying and Mapping, 2020, 45 (3): 110- 116.
14	顾佼佼, 李炳臻, 刘克. 基于改进Faster R-CNN的红外舰船目标检测算法[J]. 红外技术, 2021, 43 (2): 170- 178.
	GU J J , LI B Z , LIU K . Infrared ship target detection algorithm based on improved Faster R-CNN[J]. Infrared Technology, 2021, 43 (2): 170- 178.
15	马啸, 邵利民, 金鑫, 等. 基于改进Mask R-CNN的可见光图像中舰船目标检测方法[J]. 北京理工大学学报, 2021, 41 (7): 734- 744.
	MA X , SHAO L M , JIN X , et al. Ship target detection in optical images based improved mask R-CNN[J]. Trans.of Beijing Institute of Technology, 2021, 41 (7): 734- 744.
16	唐崇武, 刘洪喜, 代长安. 基于改进YOLOv3算法的舰船目标检测识别系统[J]. 航空电子技术, 2022, 53 (2): 39- 46.
	TANG C W , LIU H X , DAI C A . Improved YOLOv3 based ship target detection and recognition system[J]. Avionics Technology, 2022, 53 (2): 39- 46.
17	胡欣, 马丽军. 基于YOLOv5的多分支注意力SAR图像舰船检测[J]. 电子测量与仪器学报, 2022, 36 (8): 141- 149.
	HU X , MA L J . Multi-branch attention SAR image ship detection based on YOLOv5[J]. Journal of Electronic Measurement and Instrumentation, 2022, 36 (8): 141- 149.
18	钱坤, 李晨瑄, 陈美杉. 基于YOLOv5的舰船目标及关键部位检测算法[J]. 系统工程与电子技术, 2022, 44 (6): 1823- 1832.
	QIAN K , LI C X , CHEN M S . Ship target and key parts detection algorithm based on YOLOv5[J]. Systems Engineering and Electronics, 2022, 44 (6): 1823- 1832.
19	LIU S, QI L, QIN H F, et al. Path aggregation network for instance segmentation[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2018: 8759-8768.
20	LIN T Y, DOLLAR P, GIRSHICK R, et al. Feature pyramid networks for object detection[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2017: 936-944.
21	ZHENG Z H , WANG P , REN D W , et al. Enhancing geometric factors in model learning and inference for object detection and instance segmentation[J]. IEEE Trans.on Cybernetics, 2022, 52 (8): 8574- 8586.
22	REZATOFIGHI H, TSOI N, GWAK J Y, et al. Generalized intersection over union: a metric and a loss for bounding box regression[C]//Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019: 658-666.
23	HOU Q, ZHOU D, FENG J. Coordinate attention for efficient mobile network design[C]//Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021: 13713-13722.
24	TAN M, PANG R, LE Q V. EfficientDet: scalable and efficient object detection[C]//Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020: 10781-10790.
25	ZHANG Y F , REN W , ZHANG Z , et al. Focal and efficient IOU loss for accurate bounding box regression[J]. Neurocomputing, 2022, 506 (9): 146- 157.
26	LI J W, QU C W, SHAO J Q. Ship detection in SAR images based on an improved faster R-CNN[C]//Proc. of the SAR in Big Data Era: Models, Methods and Applications, 2017: 1-6.
27	HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[C]//Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2016: 770-778.
28	LIN T Y, GOYAL P, GIRSHICK R, et al. Focal loss for dense object detection[C]//Proc. of the IEEE International Conference on Computer Vision, 2017: 2980-2988.
29	HU J, SHEN L, SUN G. Squeeze-and-excitation networks[C]// Proc. of the IEEE Conference on Computer Vision and Pattern Recognition, 2018: 7132-7141.
30	WOO S, PARK J, LEE J Y, et al. Cbam: convolutional block attention module[C]//Proc. of the European Conference on Computer Vision, 2018: 3-19.
31	GEVORGYAN Z. SIoU loss: more powerful learning for bounding box regression[EB/OL]. [2022-08-28] https://arxiv.org/ftp/arxiv/papers/2205/2205.12740.pdf.

网络架构	深度因子	宽度因子	总参数量	网络层数	FLOPs
YOLOv5n	0.33	0.25	1 765 270	270	4.5
YOLOv5s	0.33	0.5	7 022 326	270	16.5
YOLOv5m	0.67	0.75	20 871 318	369	49
YOLOv5l	1	1	46 138 294	468	109.1
YOLOv5x	1.33	1.25	86 217 814	567	205.7

软硬件	配置
操作系统	Ubuntu20.04
中央处理单元	Intel Xeon silver 4210R
图形处理单元	Nvidia TITAN RTX
显存大小/GB	24
内存大小/G	16
加速环境	Cuda 10.2
深度学习框架	Pytorch 1.12.0
Python版本	Python 3.7.13

算法	CA	BiFPN	REIOU	mAP	FPS
YOLOv5s	—	—	—	94.8	143
改进算法1	√	—	—	95.0	139
改进算法2	—	√	—	95.5	127
改进算法3	—	—	√	95.3	156
改进算法4	√	√	—	96.3	128
改进算法5	—	√	√	95.9	128
改进算法6	√	—	√	96.1	139
本文算法	√	√	√	96.7	127

检测算法	mAP	FPS	FLOPs/G	Para/M	Weights/M
Faster-RCNN ResNet+FPN	83.7	16.7	134.5	41.4	330.3
RetinaNet	92.1	23.4	127	53.1	257.1
SSD	86.0	41.0	15.0	13.1	105.1
YOLOv3	96.1	36.9	155.3	61.5	123.6
YOLOv5s	94.8	143	16.5	7.0	13.7
本文算法	96.7	127	16.6	7.2	14.1

嵌入方案	mAP	FPS	FLOPs/G	Para/M
方案1	96.4	195	14.8	6.6
方案2(本文方案)	96.7	127	16.6	7.2