Lightweight SAR target detection based on channel pruning and know-ledge distillation

doi:10.11947/j.AGCS.2024.20220605

Abstract

Abstract:

Lightweight SAR target detection algorithm is of great significance for the rapid detection of ground objects in SAR images. Aiming at the low precision of lightweight detection algorithm, a lightweight SAR target detection method combining channel pruning and knowledge distillation was designed. In this method, the importance of the corresponding feature channels is identified by sparse training of the scaling factor γ of the batch normalization layer in the complex network, and then the secondary channels are cut. After fine-tuning training, it is used as a teacher network to construct a knowledge distillation framework to guide the training of lightweight model and improve the detection accuracy of light weight model. The YOLOv5-6.1 algorithm was used to build a detection framework, and the training and detection experiments were carried out on the reconstituted MSAR and SSDD multi-class target datasets. The results show that the proposed method can improve the accuracy of SAR target detection while maintaining the lightweight model size of only 3.73 MB, which verifies the effectiveness of the proposed method.

Key words: SAR, target detection, lightweight, channel pruning, knowledge distillation

CLC Number:

P237

Qihao HUANG, Guowang JIN, Xin XIONG, Limei WANG, Jiahao LI. Lightweight SAR target detection based on channel pruning and know-ledge distillation[J]. Acta Geodaetica et Cartographica Sinica, 2024, 53(4): 712-723.

Figures/Tables 15

Fig. 1

Fig. 2

Fig. 3

Fig. 4

Fig. 5

Tab. 1

Tab. 2

Tab. 3

Tab. 4

Fig. 6

Fig. 7

Tab. 5

Fig. 8

Fig. 9

Tab. 6

References 28

[1]	吴良斌. SAR图像处理与目标识别[M]. 北京: 航空工业出版社, 2013.
	WU Liangbin. SAR image processing and target recognition[M]. Beijing: Aviation Industry Press, 2013.
[2]	陈鹏, 范开国, 李晓明, 等. 合成孔径雷达海上舰船遥感探测技术与应用[M]. 北京: 海洋出版社, 2019.
	CHEN Peng, FAN Kaiguo, LI Xiaoming, et al. Remote sensing detection technology and application of synthetic aperture radar for ships at sea[M]. Beijing: Ocean Press, 2019.
[3]	DU Kangning, DENG Yunkai, WANG R, et al. SAR ATR based on displacement- and rotation-insensitive CNN[J]. Remote Sensing Letters, 2016, 7(9):895-904.
[4]	靳国旺, 张红敏, 徐青. 雷达摄影测量[M]. 北京: 测绘出版社, 2015.
	JIN Guowang, ZHANG Hongmin, XU Qing. Radargrammetry[M]. Beijing: Surveying and Mapping Press, 2015.
[5]	ELDHUSET K. Automatic ship and ship wake detection in spaceborne SAR images from coastal regions[C]//Proceedings of 2002 International Geoscience and Remote Sensing Symposium. Edinburgh: IEEE, 2002: 1529-1533.
[6]	王世晞, 贺志国. 基于PCA特征的快速SAR图像目标识别方法[J]. 国防科技大学学报, 2008, 30(3):136-140.
	WANG Shixi, HE Zhiguo. The fast target recognition approach based on PCA features for SAR images[J]. Journal of National University of Defense Technology, 2008, 30(3):136-140.
[7]	李健伟, 曲长文, 彭书娟, 等. 基于卷积神经网络的SAR图像舰船目标检测[J]. 系统工程与电子技术, 2018, 40(9):1953-1959.
	LI Jianwei, QU Changwen, PENG Shujuan, et al. Ship detection in SAR images based on convolutional neural network[J]. Systems Engineering and Electronics, 2018, 40(9):1953-1959.
[8]	CHEN Sizhe, WANG Haipeng. SAR target recognition based on deep learning[C]//Proceedings of 2014 International Conference on Data Science and Advanced Analytics. Shanghai: IEEE, 2014: 541-547.
[9]	SHARIFZADEH F, AKBARIZADEH G, SEIFI KAVIAN Y. Ship classification in SAR images using a new hybrid CNN-MLP classifier[J]. Journal of the Indian Society of Remote Sensing, 2019, 47(4):551-562.
[10]	余东行, 郭海涛, 赵传, 等. 直线特征辅助的靠岸舰船检测[J]. 测绘科学技术学报, 2019, 36(3):275-280,286.
	YU Donghang, GUO Haitao, ZHAO Chuan, et al. Detection of ship docked in harbor assisted with line feature[J]. Journal of Geomatics Science and Technology, 2019, 36(3):275-280,286.
[11]	焦军峰, 靳国旺, 熊新, 等. 旋转矩形框与CBAM改进RetinaNet的SAR图像近岸舰船检测[J]. 测绘科学技术学报, 2020, 37(6):603-609.
	JIAO Junfeng, JIN Guowang, XIONG Xin, et al. SAR images nearshore ship detection based on RetinaNet algorithm with rotated rectangular box[J]. Journal of Geomatics Science and Technology, 2020, 37(6):603-609.
[12]	HAN Song, POOL J, TRAN J, et al. Learning both weights and connections for efficient neural networks[EB/OL].[2022-09-24].http://dx.doi.org/10.48550/arXiv.1506.02626.
[13]	HINTON G E, VINYALS O, DEAN J. Distilling the knowledge in a neural network[EB/OL].[2022-09-24]. https://www.semanticscholar.org/paper/0c908739fbff75f03469d13d4a1a07de3414ee19.
[14]	周舟, 王海鹏, 徐丰, 等. 基于通道剪枝的SAR图像舰船检测优化算法[J]. 上海航天(中英文), 2020, 37(4):48-54.
	ZHOU Zhou, WANG Haipeng, XU Feng, et al. Optimization algorithm for ship detection in SAR images based on channel pruning[J]. Aerospace Shanghai (Chinese & English), 2020, 37(4):48-54.
[15]	陆天宇, 徐湛, 崔红元, 等.大幅宽SAR图像嵌入式舰船实时检测系统设计[J].计算机工程与应用, 2024, 60(1):301-309.
	LU Tianyu, XU Zhan, CUI Hongyuan, et al. Design of ship real-time detection system embedded with large and wide SAR images[J]. Computer Engineering and Applications, 2024, 60(1):301-309.
[16]	LIU Zhuang, LI Jianguo, SHEN Zhiqiang, et al. Learning efficient convolutional networks through network slimming[C]//Proceedings of 2017 IEEE International Conference on Computer Vision. Venice: IEEE, 2017: 2755-2763.
[17]	王恒涛, 张上. 轻量化SAR图像舰船目标检测算法[J]. 电光与控制, 2023, 30(5):99-104,110.
	WANG Hengtao, ZHANG Shang. Lightweight ship target detection algorithm for SAR images[J]. Electronics Optics & Control, 2023, 30(5):99-104,110.
[18]	付晓雅. 基于场景分类和知识蒸馏的SAR图像舰船目标检测算法研究[D]. 天津: 河北工业大学, 2022.
	FU Xiaoya. Research on ship target detection algorithm in SAR image based on scene classification and knowledge distillation[D]. Tianjin: Hebei University of Technology, 2022.
[19]	SANDLER M, HOWARD A, ZHU Menglong, et al. MobileNetV2: inverted residuals and linear bottlenecks[C]//Proceedings of 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City: IEEE, 2018: 4510-4520.
[20]	LIU Wei, ANGUELOV D, ERHAN D, et al. SSD: single shot MultiBox detector[C]//Proceedings of 2016 Computer Vision-ECCV. Cham: Springer, 2016: 21-37.
[21]	陈诗琪, 王威, 占荣辉, 等. 特征图知识蒸馏引导的轻量化任意方向SAR舰船目标检测器[J]. 雷达学报, 2023, 12(1):140-153.
	CHEN Shiqi, WANG Wei, ZHAN Ronghui, et al. A lightweight, arbitrary-oriented SAR ship detector via feature map-based knowledge distillation[J]. Journal of Radars, 2023, 12(1):140-153.
[22]	罗杨, 卞春江, 陈红珍.基于特征解耦的SAR图像舰船检测蒸馏[J].计算机工程与应用, 2024, 60(02):171-179.
	LUO Yang, BIAN Chunjiang, CHEN Hongzhen. SAR image ship detection distillation based on feature decoupling[J]. Computer engineering and applications, 2024, 60(02):171-179.
[23]	PARK J, NO A. Prune your model before distill it[C]//Proceedings of 2022 Computer Vision-ECCV European Conference. Cham: Springer, 2022: 120-136.
[24]	CHEN Shiqi, ZHAN Ronghui, WANG Wei, et al. Learning slimming SAR ship object detector through network pruning and knowledge distillation[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2021, 14:1267-1282.
[25]	WANG Zhen, DU Lan, LI Yi. Boosting lightweight CNNs through network pruning and knowledge distillation for SAR target recognition[J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2021, 14:8386-8397.
[26]	陈杰, 黄志祥, 夏润繁, 等. 大规模多类SAR目标检测数据集-1.0[DS/OL]. [2022-09-24].https://radars.ac.cn/web/data/getData?dataType=MSAR.
	CHEN Jie, HUANG Zhixiang, XIA Runfan, et al.Large-scale multi-class SAR image target detection dataset-1.0[DS/OL]. [2022-09-24].https://radars.ac.cn/web/data/getData?dataType=MSAR.
[27]	XIA Runfan, CHEN Jie, HUANG Zhixiang, et al. CRTransSar: a visual transformer based on contextual joint representation learning for SAR ship detection[J]. Remote Sensing, 2022, 14(6):1488.
[28]	ZHANG Tianwen, ZHANG Xiaoling, LI Jianwei, et al. SAR ship detection dataset (SSDD): official release and comprehensive data analysis[J]. Remote Sensing, 2021, 13(18):3690.

目标类型	数据集	训练集/张	验证集/张	平均目标数/(个/张)
油罐	MSAR	1000	250	9.86
桥梁	MSAR	1266	316	1.17
舰船	SSDD	812	232	2.12

网络模型	平均准确率/(%)	平均召回率/(%)	平均精度/(%)	模型体积/MB	平均推理时间/ms
YOLOv5-x	90.9	93.6	93.9	165.00	23.9
YOLOv5-l	90.9	90.9	93.1	88.50	14.4
YOLOv5-m	91.4	92.5	94.2	40.20	8.9
YOLOv5-s	91.3	93.3	94.0	13.70	3.6
YOLOv5-n	91.4	91.6	92.2	3.73	1.7

学生网络	教师网络	剪枝比例/(%)	平均准确率/(%)	平均召回率/(%)	平均精度/(%)	模型体积/MB	推理时间/ms
YOLOv5-n	×	×	91.4	91.6	92.2	3.73	1.7
YOLOv5-n	YOLOv5-m	×	92.1	90.6	94.1
		10	91.6	92.5	94.7
		20	91.4	93.8	94.6
		30	91.9	92.9	95.0
		40	92.0	92.6	94.9
		50	90.2	93.6	93.8
		60	91.1	92.6	93.9
		69	90.9	93.4	94.1

网络模型	剪枝比例/(%)	平均准确率/(%)	平均召回率/(%)	平均精度/(%)	模型体积/MB	推理时间/ms
YOLOv5-m(基准)	—	91.4	92.5	94.2	40.2	8.9
教师网络A	10	91.1	93.9	94.3	36.2	6.9
教师网络B	20	91.2	94.5	94.9	31.1	6.4
教师网络C	30	91.6	94.6	94.9	26.6	6.2
教师网络D	40	91.0	93.7	94.5	22.6	6.0
教师网络E	50	92.2	94.4	95.3	19.2	5.8
教师网络F	60	90.5	94.4	94.6	16.5	5.5
教师网络G	69	91.0	94.9	94.7	14.8	5.5

方法	平均准确率/(%)	平均召回率/(%)	平均精度/(%)	模型体积/MB	推理时间/ms
YOLOv3	90.8	92.0	93.2	117.00	12.8
YOLOv3-tiny	90.9	75.1	84.1	16.60	1.9
YOLOv7	93.0	93.2	94.1	71.30	11.5
YOLOv7-tiny	91.1	88.6	89.6	11.70	3.4
YOLOX-s	92.2	93.8	94.9	34.30	7.9
YOLOX-tiny	91.5	92.7	93.0	19.40	4.8
YOLOX-nano	90.7	92.1	92.5	3.60	1.5
YOLOv5-n	91.4	91.6	92.2	3.73	1.7
本文方法(30%通道剪枝＋知识蒸馏)	91.9	92.9	95.0	3.73	1.7