A Lightweight Recognition Method for Low Altitude Targets in the Battlefield

doi:10.12382/bgxb.2024.0170

Abstract

Abstract:

In the process of intelligent radar target detection,the real-time recognition of targets by edge devices is particularly important.Considering the resource constraints inherent in embedded devices,the need for lightweight target recognition networks has become increasingly prominent.To address the challenges of limited battlefield target images and low contrast,a battlefield low-altitude target recognition algorithm named YOLOF is proposed.This algorithm is based on the YOLOv5s network model and incorporates a cycle-generative adversarial network (CycleGAN) for image enhancement.By integrating the RepVGG module and SiLU activation function,the algorithm enhances the feature extraction capability and inference speed of model through structural reparameterization and more efficient activation functions.Additionally,a pruning algorithm based on filter importance is employed to accurately evaluate and remove the filters with low weight impact,thereby optimizing the model’s structure and improving the computational and storage efficiencies.Furthermore,the knowledge distillation based on feature layers allows the transfer of knowledge from the teacher model to the student model’s feature layers,thus maintaining the high performance of the pruned model.Experimental results demonstrate that the proposed YOLOF algorithm,compared to the original YOLOv5s algorithm,achieves network lightweighting while ensuring high-precision target recognition.Specifically,the parameter count is reduced to just 3.6×10⁶,and the mean average precision (mAP) reaches 86.3% on a custom dataset,meeting the requirements for battlefield low-altitude target recognition.

Key words: target recognition, YOLOv5s network, pruning, network lightweight, knowledge distillation

CLC Number:

TP391.4

XU Tunan, GAO Ang, CHEN Yucheng, YAN Shoucheng, DENG Bin. A Lightweight Recognition Method for Low Altitude Targets in the Battlefield[J]. Acta Armamentarii, 2025, 46(2): 240170-.

Figures/Tables 15

Fig.1 Comparison of image enhancement results

Fig.2. YOLOF network model structure diagram

Fig.3 SiLU activation function curve

Fig.4 RepVGG model structure diagram

Fig.5 Specific workflow diagram of pruning

Fig.6 Knowledge distillation Workflow

Fig.7 Structure diagram of of embedded computing platform

Table 1 Predicted results of binary classification samples

真实情况	预测情况
真实情况	正例	反例
正例	TP(真正例)	FN(假反例)
反例	FP(假正例)	TN(真反例)

Table 2 Initial parameter setting

参数	数值
训练周期	100
批次训练样本数量	16
图片输入大小	640×640
是否断点续训	无
训练冻结层数	0
学习率调整算法	余弦学习率
初始学习率	0.01

Fig.8 Performance comparison of different target detection algorithms

Table 3 Experimental results of different target detection algorithms

算法	mAP@0.5	F₁	参数量/ 10⁶	模型大小/ MB	浮点运算/ GFLOPs
YOLOv5s	0.865	0.858	7.0	13.7	15.8
YOLOv5x	0.883	0.860	86	165.1	203.8
YOLOv7	0.886	0.860	37.2	71.4	51.2
YOLOv7x	0.887	0.862	70.1	135.2	102.3
YOLOF	0.863	0.861	3.6	7.0	7.1

Fig.9 YOLOF experimental effect diagram

Table 4 Experimental effect diagrams of mainstream algorithms

算法	海边	沙地	森林	平原
YOLOv7
YOLOv5s
YOLOv5x
YOLO-F (蒸馏后未剪枝)
YOLO-F

Table 5 Chart of ablation experimental results

模型	改进策略	mAP	参数量/10⁶	模型大小/MB
A	Yolov5s(初始模型)	0.865	7.2	13.7
B	A+RepVGG	0.860	6.2	12.6
C	B+剪枝	0.827	3.6	7.0
D	C+知识蒸馏	0.863	3.6	7.0

Fig.10 Comparison chart of ablation experimental performances

References 23

[1]	张凯歌, 卢志刚, 聂天常, 等. 面向无人装备的智能边缘计算软技术分析[J]. 兵工学报, 2023, 44(9):2611-2621. doi: 10.12382/bgxb.2022.1166
	ZHANG K G, LU Z G, NIE T C, et al. Analysis of soft intelligent edge computing technologies for unmanned systems[J]. Acta Armamentarii, 2023, 44(9):2611-2621. (in Chinese) doi: 10.12382/bgxb.2022.1166
[2]	陈龙, 张建林, 彭昊, 等. 多尺度注意力与领域自适应的小样本图像识别[J]. 光电工程, 2023, 50(4):60-73.
	CHEN L, ZHANG J L, PENG H, et al. Few-shot image classification via multi-scale attention and domain adaptation[J]. Opto-Electronic Engineering, 2023, 50(4):60-73. (in Chinese)
[3]	程旭, 宋晨, 史金钢, 等. 基于深度学习的通用目标检测研究综述[J]. 电子学报, 2021, 49(7):1428-1438. doi: 10.12263/DZXB.20200570
	CHENG X, SONG C, SHI J G, et al. A survey of generic object detection methods based on deep learning[J]. Acta Electonica Sinica, 2021, 49(7):1428-1438. (in Chinese)
[4]	KRIZHEVSKY A, SUTSKEVER I, HINTON G. Imagenet classification with deep convolutional neural networks[J]. Communications of the ACM, 2017, 60(6):84-90.
[5]	FANG S, ZHANG B, HU J. Improved mask R-CNN multi-target detection and segmentation for autonomous driving in complex scenes[J]. Sensors, 2023, 23(8):3853.
[6]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once:unified,Real-Time object detection[C]// Proceedings of IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas,NV,US: IEEE, 2016:779-788.
[7]	于博文, 吕明. 改进的YOLOv3算法及其在军事目标检测中的应用[J]. 兵工学报, 2022, 43(2):345-354. doi: 10.3969/j.issn.1000-1093.2022.02.012
	YU B W, LÜ M. Improved YOLOv3 algorithm and its application in military target detection[J]. Acta Armamentarii, 2022, 43(2):345-354. (in Chinese) doi: 10.3969/j.issn.1000-1093.2022.02.012
[8]	熊光明, 罗震, 孙冬, 等. 基于红外相机和毫米波雷达融合的烟雾遮挡无人驾驶车辆目标检测与跟踪[J]. 兵工学报, 2024, 45(3):893-906. doi: 10.12382/bgxb.2022.0602
	XIONG G M, LUO Z, SUN D, et al. Object detection and tracking for Unmanned Vehicles based on fusion of infrared camera and MMW radar in smoke-obscured environment[J]. Acta Armamentarii, 2024, 45(3):893-906. (in Chinese) doi: 10.12382/bgxb.2022.0602
[9]	SANDLER M, HOWARD A, ZHU M, et al. Mobilenetv2:Inverted residuals and linear bottlenecks[C]// Proceedings of the IEEE conference on computer vision and pattern recognition. Salt Lake City,UT,US: IEEE, 2018:4510-4520.
[10]	KONG L R, WANG J Z, ZHAO P. YOLO-G:a lightweight network model for improving the performance of military targets detection[J]. IEEE Access, 2022, 10:55546-55555.
[11]	WENG K, CHU X, XU X, et al. Efficientrep:an efficient Repvgg-style ConVnets with hardware-aware neural network design[J]. arxiv preprint arxiv:2302.00386, 2023.
[12]	ZHU J Y, PARK T, ISOLA P, et al. Unpaired image-to-image translation using cycle-consistent advers arialnetworks[C]// Proceedings of the IEEE International Conference on Computer Vision.Venice, Italy: IEEE, 2017:2223-2232.
[13]	HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas,NV, US: IEEE, 2016:770-778.
[14]	刘飞, 高红艳, 卫泽刚, 等. 基于Res-Net深度特征的SAR图像目标识别方法[J]. 液晶与显示. 2021, 36(4):624-631.
	LIU F, GAO L Y, WEI Z G, et al. SAR image target recognition method using based on Res-Net deep features[[J]. Chinese Journal of Liquid Crystals and Displays, 2021, 36(4):624-631. (in Chinese)
[15]	QING Y H, LIU W Y, FENG L Y, et al. Improved YOLO network for free-angle remote sensing target detection[J]. Remote Sensing, 2021, 13(11):2171.
[16]	SUN S Q, CHENG Y, GAN Z, et al. Patient knowledge distillation for BERT model compression:arXiv:1908.09355[R/OL]. Ithaca,NY,US: Cornell University, 2019(2019-08-25). https://arxiv.org/abs/1908.09355
[17]	SHEN M Y, MOLCHANOV P, YIN H X, et al. When to prune? a policy towards early structural pruning[C]// Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. New Orleans,LA,US: IEEE, 2022:12237-12246.
[18]	VADERA S, AMEEN S. Methods for pruning deep neural networks[J]. IEEE Access, 2022, 10:63280-63300.
[19]	LUO J H, WU J, LIN W. Thinet:a filter level pruning method for deep neural network compression[C]// Proceedings of the IEEE International Conference on computer vision. Venice,Italy:IEEE, 2017:5058-5066.
[20]	BA L J, CARUANA R. Do deep nets really need to be deep?[J]. Advances in neural information processing systems, 2014, 3:2654-2662.
[21]	SHU C Y, LIU Y F, GAO J F, et al. Channel-wise knowledge distillation for dense prediction[C]// Proceedings of the IEEE/CVF International Conference on Computer Vision. Montreal,QC,Canada:IEEE, 2021:5311-5320.
[22]	JI S Y, ZHANG Z Z, YING S H, et al. Kullback-Leibler divergence metric learning[J]. IEEE transactions on cybernetics, 2020, 52(4):2047-2058.
[23]	ZHOU Z, ZHUGE C, GUAN X, et al. Channel distillation:Channel-wise attention for knowledge distillation: arXiv.2006.01683[R/OL]. lthaca, NY.US:Cornell University 2020(2020-06-2).https://arxiv.org/abs/2006.01683