成像制导运动模糊目标检测算法

doi:10.12382/bgxb.2024.0376

摘要/Abstract

摘要：

为提升弹载成像制导中运动模糊图像目标检测的精确性与效率,提出一种轻量化且高效的运动模糊图像目标检测(Lighter and More Effective Motion-blurred Image Object Detection,LEMBD)网络。通过深入分析运动模糊图像的成因,基于成像机理构建了专用的运动模糊图像数据集。在不增加网络参数的前提下,采用共享权重的孪生网络设计,并引入先验知识,将清晰图像的特征学习用于模糊图像的特征提取,以同时实现对清晰与模糊图像的精准检测。此外,设计了部分深度可分离卷积替代普通卷积,显著减少了网络的参数量与计算量,并提升了学习性能。为进一步优化特征融合质量,提出跨层路径聚合特征金字塔网络,有效利用低级特征的细节信息和高级特征的语义信息。实验结果表明,所提LEMBD网络在运动模糊图像目标检测任务中的性能优于传统目标检测方法和主流运动模糊检测算法,能够为精确制导任务提供更精准的目标相对位置信息。

关键词: 精确目标检测, 运动模糊, 轻量化, 部分深度可分离卷积, 跨层路径聚合特征金字塔网络

Abstract:

To enhance the accuracy and efficiency of motion-blurred image object detection in missile-borne imaging guidance,this paper proposes a lighter and more effective motion-blurred image object detection (LEMBD) network.The causes of motion-blurred image are analyzed,and a dedicated motion-blurred image dataset is constructed based on the imaging mechanism.Without increasing network parameters,a shared-weight siamese network design is adopted,and the prior knowledge is introduced to extract the features of blurred images by the feature learning of clear images,thereby enabling the simultaneous detection of both clear and blurred images.Additionally,the partial depthwise separable convolutions are introduced to replace the standard convolutions,which significantly reduce the parameter count and computational cost while enhancing learning performance.To further improve the feature fusion quality,a cross-layer path aggregation feature pyramid network is designed to effectively leverage both the detail information of low-level features and the semantic information of high-level features.Experimental results demonstrate that the proposed LEMBD network achieves superior performance in detecting the targets within motion-blurred images compared to conventional object detection and state-of-the-art motion-blurred detection methods,which can provide more accurate relative positional information for precision guidance tasks.

Key words: accurate object detection, motion-blurring, lightweight, partial depth separable convolution, cross-layer path aggregation feature pyramid network

中图分类号:

TP391.41

赵春博, 莫波, 李大维, 赵洁. 成像制导运动模糊目标检测算法[J]. 兵工学报, 2025, 46(2): 240376-.

ZHAO Chunbo, MO Bo, LI Dawei, ZHAO Jie. Research on Motion Blur Object Detection Technology for Imaging Guidance[J]. Acta Armamentarii, 2025, 46(2): 240376-.

图/表 14

图1 运动模糊处理后的DroneVehicle数据集

Fig.1 Motion blur processed DroneVehicle dataset

图2 YOLOv5框架

Fig.2 YOLOv5 framework

图3 PConv和DSConv 结合成的基础卷积单元(PDS)

Fig.3 Basic convolution unit PDS with PConv and DSConv

图4 引入先验信息的权重共享网络主干与特征信息融合网络主干

Fig.4 Comparison between the weight sharing network backbone with prior information and the feature information fusion network backbone

图5 增加小目标检测层的原始网络与改进的特征融合网络

Fig.5 The original network with a small object detection layer and the improved feature fusion network

图6 更轻、更有效的运动模糊图像目标检测网络

Fig.6 Lighter and more efficient object detection network for motion blur images

图7 运动模糊数据集

Fig.7 Motion blur dataset

表1 CPU/GPU及其相关参数

Table 1 CPU/GPU and its related parameters

CPU参数	数值	GPU参数	数值
Intel(R) Core(TM)	i9-7980XE	NVIDIA GeForce RTX	3090
基准速度	2.60GHz	内存	24.0GB
内核	18	Pytorch版本	1.10.1
逻辑处理器	36	CUDA版本	11.3

表2 运动模糊图像目标检测网络消融实验

Table 2 Ablation test data of motion blur image object detection network

PIDSFENet	PDS	EPAFPN	CLPAFPN	轿车	货车	卡车	公交车	厢式货车	召回率	mAP0.5/%	精确率	参数量/10⁶	消耗算力/10⁹
×	×	×	×	85.7	42.7	56.7	90.6	44.5	61.1	64.1	65.4	2.01	5.0
√	×	×	×	88.7	50.4	62.4	92.2	48.0	63.7^↑2.6	68.3^↑4.2	71.3^↑5.9	2.01	5.0
√	√	×	×	89.0	47.2	64.7	92.8	49.6	62.3^↑1.2	68.7^↑4.6	75.1^↑9.7	1.19^↓40.5%	3.3^↓36%
√	√	√	×	88.0	44.0	58.5	89.7	46.9	62.5^↑1.4	65.4^↑1.3	67.9^↑2.5	1.28^↓36%	6.4^↑28%
√	√	×	√	88.5	47.2	59.1	90.5	48.4	62.0^↑0.9	66.7^↑2.6	71.8^↑6.4	1.55^↓22.5%	8.1^↑62%

图8 不同算法参数量评估结果

Fig.8 Parameters evaluated results of different algorithms

表3 运动模糊图像目标检测同类算法对比实验

Table 3 Comparative experiments on similar algorithms for object detection in motion blurred images

算法	召回率	mAP0.5/ %	精确率	参数量/10⁶	消耗算力/10⁹	帧率/ (帧·s^-1)
Spec By Exposure	53.0	64.9		41.5	67.3	20.9
TPH-YOLOv5++	67.4	74.6	79.9	7.18	29.9	59.5
LEMBD	69.6	75.2	79.0	4.90	17.8	55.5

图9 不同算法消耗算力评估结果

Fig.9 FLOPs evaluated results of different algorithms

图10 不同算法权重尺寸评估结果

Fig.10 Weight size evaluated results of different algorithms

表4 不同环境、大小目标检测结果

Table 4 Detected results of objects in different environments

情形	真实目标		基准算法		本文算法
小目标
部分缺失目标
黑暗环境下目标
黑暗环境下目标

参考文献 21

[1]	李成, 李建勋, 童中翔, 等. 红外成像制导末端局部图像识别跟踪研究[J]. 兵工学报, 2015, 36(7):1213-1221. doi: 10.3969/j.issn.1000-1093.2015.07.009
	LI C, LI J X, TONG Z X, et al. Research on partial image recognition and tracking in infrared imaging terminal guidance[J]. Acta Armamentarii, 2015, 36(7):1213-1221. (in Chinese) doi: 10.3969/j.issn.1000-1093.2015.07.009
[2]	LIU W, ANGUELOV D, ERHAN D, et al. SSD:single shot multibox detector[C]// Proceedings of the Computer Vision-ECCV 2016:14th European Conference. Amsterdam,the Netherlands: Springer, 2016:21-37.
[3]	BOCHKOVSKIY A, WANG C Y, LIAO H Y M. YOLOv4:optimal speed and accuracy of object detection[R/OL]. Ithaca,NY,US: Cornell University, 2024(2024-04-23). https://arxiv.org/abs/2004.10934
[4]	LI C Y, LI L L, JIANG H L, et al. YOLOv6:a single-stage object detection framework for industrial applications[R/OL]. Ithaca,NY,US: Cornell University, 2022(2022-09-07). https://arxiv.org/abs/2209.02976
[5]	WANG C Y, BOCHKOVSKIY A, LIAO H Y M. YOLOv7:trainable bag-of-freebies sets new state-of-the-art for real-time object detectors[C]// Proceedings of the 2023 IEEE Conference on Computer Vision and Pattern Recognition. Vancouver,Canada,IEEE, 2023:7464-7475.
[6]	ZHOU L H, MIN W D, LIN D Y, et al. Detecting motion blurred vehicle logo in IoV using filter-DeblurGAN and VL-YOLO[J]. IEEE Transactions on Vehicular Technology, 2020, 69(4):3604-3614.
[7]	KUPYN O, BUDZAN V, MYKHAILYCH M, et al. DeblurGAN:blind motion deblurring using conditional adversarial networks[C]// Proceedings of the 2018 IEEE Conference on Computer Vision and Pattern Recognition. Salt Lake City,UT,US, IEEE, 2018:8183-8192.
[8]	REDMON J, FARHADI A. YOLOv3:an incremental improvement:arXiv:1804.02767v1[R/OL]. Ithaca,NY,US: Cornell University, 2018(2018-04-08) [2022-05-17]. https://arXiv.org/abs/1804.02767v1
[9]	WANG H Z, HU C H, QIAN W J, et al. RT-Deblur:real-time image deblurring for object detection[J]. The Visual Computer, 2024, 40(4):2873-2887.
[10]	LIU B D, XU C, YANG W, et al. Motion robust high-speed light-weighted object detection with event camera[J]. IEEE Transactions on Instrumentation and Measurement, 2023, 72:1-13.
[11]	SAYED M, BROSTOW G. Improved handling of motion blur in online object detection[C]// Proceedings of the 2021 IEEE conference on computer vision and pattern recognition. Nashville,TN,US,IEEE, 2021:1706-1716.
[12]	HE K M, ZHANG X Y, REN S Q, et al. Deep residual learning for image recognition[C]// Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition. Las Vegas,NV,US,IEEE, 2016:770-778.
[13]	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 & Machine Intelligence, 2017, 39(6):1137-1149.
[14]	ZHAO Q, LIU B H, LYU S C, et al. TPH-YOLOv5++:boosting object detection on drone-captured scenarios with cross-layer asymmetric transformer[J]. Remote Sensing, 2023, 15(6):1687.
[15]	SUN Y M, CAO B, ZHU P F, et al. Drone-based RGB-infrared cross-modality vehicle detection via uncertainty-aware learning[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2022, 32(10):6700-6713.
[16]	ZHANG G J, LUO Z P, TIAN Z C, et al. Towards efficient use of multi-scale features in transformer-based object detectors[C]// Proceedings of the 2023 IEEE Conference on Computer Vision and Pattern Recognition.Vancouver,Canada, IEEE, 2023:6206-6216.
[17]	LIANG J Y, CAO J Z, FAN Y C, et al. VRT:a video restoration transformer[J]. IEEE Transactions on Image Processing, 2024, 33:2171-2182.
[18]	CHEN J R, KAO S H, HE H, et al. Run,don’t walk:chasing higher FLOPS for faster neural networks[C] // Proceedings of the 2023 IEEE Conference on Computer Vision and Pattern Recognition. Vancouver,Canada,IEEE, 2023:12021-12031.
[19]	KAISER L, GOMEZ A N, CHOLLET F. Depthwise separable convolutions for neural machine translation[R/OL]. Ithaca,NY,US: Cornell University, 2017(2017-06-16). https://arxiv.org/abs/1706.03059
[20]	BAO C, CAO J, HAO Q, et al. Dual-YOLO architecture from infrared and visible IMAGES for object detection[J]. Sensors, 2023, 23(6):2934.
[21]	YANG X, YAN J C. Arbitrary-oriented object detection with circular smooth label[C]// Computer Vision-ECCV 2020:16th European Conference.Glasgow,UK:Springer International Publishing, 2020:677-694.