基于改进快速区域卷积网络的目标检测轻量化算法

doi:10.3969/j.issn.1000-1093.2021.12.014

摘要/Abstract

摘要： 基于深度学习的目标检测算法已成为合成孔径雷达（SAR）图像目标检测任务的主流。深层网络通常具有大量参数，运行速度不能满足实时要求，难以在资源受限的设备(如移动端)上部署。考虑到对模型实时性和可移植性的要求，对双阶段目标检测算法快速区域卷积神经网络进行轻量化改进，比较不同改进方法对算法速度与精度的影响。结合SAR图像的特点，优化轻量化模型，与单阶段目标检测算法的单脉冲多盒检测网络对比。仿真实验结果表明，改进轻量化模型在保持原有精度水平下，模型占用内存和算法运算量大大减少，可有效满足SAR图像目标检测的实时性要求。

关键词: 目标检测, 快速卷积神经网络, 合成孔径雷达, 轻量化算法, 实时性

Abstract: The target detection algorithms based on deep learning has become the mainstream of target detection in synthetic aperture radar images. Deep network algorithm often has a large number of parameters and don't run fast enough to meet real-time requirements，making it difficult to deploy on resource-constrained devices such as mobile terminal. Considering the requirements of real-time performance and portability of the model，Faster-RCNN for the two-stage target detection algorithm was improved to compare the influence of different improved methods on the speed and accuracy of algorithm.The lightweight model was optimized in combination with the characteristics of synthetic aperture radar (SAR) images，and finally compared with the single shot multibox detector for one-stage target detection algorithm.The experimental results show that the speed of the improved lightweight model is greatly improved while maintaining the original accuracy level，which can effectively meet the real-time requirements of SAR image target detection.

Key words: targetdetection, fasterregion-basedconvolutionalneuralnetwork, syntheticapertureradar, lightweightalgorithm, real-time

中图分类号:

TN957.51⁺2

马月红，孔梦瑶. 基于改进快速区域卷积网络的目标检测轻量化算法[J]. 兵工学报, 2021, 42(12): 2664-2674.

MA Yuehong， KONG Mengyao. A Lightweight Target Detection Algorithm Based on the Improved Faster-RCNN[J]. Acta Armamentarii, 2021, 42(12): 2664-2674.

参考文献

［1］刘燕.高分辨率及新模式SAR成像算法研究[D].西安：西安电子科技大学，2012.
LIU Y. Research on high-resolution and new mode SAR imaging algorithm[D].Xi'an：Xidian University，2012.(in Chinese)
[2］赵明波，何峻，付强.SAR图像CFAR检测的快速算法综述[J].自动化学报，2012，38(12):1885-1895.
ZHAO M B，HE J，FU Q.Overview of fast algorithms for SAR image CFAR detection[J].Acta Automatica Sinica，2012，38(12): 1885-1895. (in Chinese)
[3］ SAHAL M，SAID A Z，PUTRA Y R，et al.Comparison of CFAR methods on multiple targets in sea clutter using SPX-Radar-Simulator[C]∥Proceedings of International Seminar on Intelligent Technology and Its Applications.Surabaya，Indonesia: IEEE，2020.
[4］ ABBADI A，ABBANEA，BENCHEIKH M L，et al.A new adaptive CFAR processor in multiple target situations[C]∥Proceedings of Seminar on Detection Systems Architectures and Technologies.Algiers，Algeria:IEEE，2017.
[5］袁湛，何友，蔡复青.基于改进扩展分形特征的SAR图像目标检测方法[J].宇航学报，2011，32(6):1379-1385.
YUAN Z，HE Y，CAI F Q.Target detection method of SAR image based on improved extended fractal features[J].Journal of Aerospace，2011，32(6):1379-1385. (in Chinese)
[6］吕本远，禚真福，韩永赛，等.基于Faster-RCNN的改进型目标检测算法[J].激光与光电子学进展，2021，58（22）：2210017.
L B Y，ZHUO Z F，HAN Y S，et al. Improved target detection algorithm based on Faster-RCNN[J].Laser & Optoelectronics Progress, 2021, 58(22): 2210017. (in Chinese)
[7］宋相法，焦李成.基于稀疏表示及光谱信息的高光谱遥感图像分类[J].电子与信息学报，2012，34(2):268-272.
SONG X F，JIAO L C.Hyperspectral remote sensing image classification based on sparse representation and spectral information[J].Journal of Electronics & Information Technology，2012，34(2):268- 272. (in Chinese)
[8］徐炎.基于相干图学习的干涉SAR图像分割[D].西安：西安电子科技大学，2014.
XU Y.Interferometric SAR image segmentation based on coherence map learning[D].Xi'an：Xidian University，2014. (in Chinese)
[9］曹磊，王强，史润佳，等.基于改进RPN的Faster-RCNN网络SAR图像车辆目标检测方法[J].东南大学学报(自然科学版)，2021，51(1):87-91.
CAO L，WANG Q，SHI R J，et al.Faster-RCNN network SAR image vehicle target detection method based on improved RPN[J].Journal of Southeast University(Natural Science Edition)，2021，51(1): 87-91. (in Chinese)
[10］韩俊，任进.基于特征增强的多尺度SSD行人检测算法[J].信息技术与信息化，2020(1):111-115.
HAN J，REN J.Multi-scale SSD pedestrian detection algorithm based on feature enhancement[J]. Information Technology and Informatization，2020(1):111-115. (in Chinese)
[11］张玉燕，李永保，温银堂，等.基于Faster-RCNN卷积神经网络的金属点阵结构缺陷识别方法[J].兵工学报，2019，40(11):2329-2335.
ZHANG Y Y，LI Y B，WEN Y T，et al.A metal lattice defect identification method based on Faster-RCNN convolutional neural network[J].Acta Armamentarii，2019，40(11):2329-2335. (in Chinese)
[12］ GIRSHICK R，DONAHUE J，DRRELL T，et al. Region-based convolutional network for accurate object detection and segmentation[J].IEEE Transcations on Pattern Analysis and Machine Intelligence，2015，38:142-158.
[13］曹之君，张良.基于Faster-RCNN的快速目标检测算法[J].航天控制，2020，38(4):49-55.
CAO Z J，ZHANG L.Fast target detection algorithm based on Faster-RCNN[J].Aerospace Control，2020，38(4):49-55. (in Chinese)
[14］ DING J，CHEN B，LIU H W，et al.Convolutional neural network with data augmentation for SAR target recognition[J].IEEE Geoscience and Remote Sensing Letters，2016，13(3):364-368.
[15］ CHOLLET F. Xception: deep learning with depthwise separable convolutions[C]∥Proceedings of 2017 IEEE Conference on Computer Vision and Pattern Recognition.Honolulu，HI，US:IEEE，2017:1800-1807.
[16］ LIU R，YU Z H，MO D L，et al.An improved Faster-RCNN algorithm for object detection in remote sensing images[C]∥Proceedings of the 39th Chinese Control Conference.Shenyang，China:IEEE，2020.
[17］王冬丽，廖春江，牟金震，等.基于特征融合的SSD视觉小目标检测[J].计算机工程与应用，2020，56(16):31-36.
WANG D L，LIAO C J，MOU J Z，et al.SSD visual small target detection based on feature fusion[J].Computer Engineering and Applications，2020，56(16):31-36. (in Chinese)
[18］ HE K M，GKIOXARI G，DOLLR P，et al.Mask-RCNN [C]∥Proceedings of the IEEE International Conference on Computer Vision. Venice，Italy:IEEE，2017.
[19］闫新庆，杨喻涵，陆桂明.基于改进Faster-RCNN的目标检测算法研究[J].中国新通信，2021，23(1):46-48.
YAN X Q，YANG Y H，LU G M.Research on target detection algorithm based on improved Faster-RCNN[J].China New Telecommunications，2021，23(1):46-48.(in Chinese)

[1]	李博，王博，韩京冶，杨宗睿，栗霖雲，傅畅之. 基于车载计算机的红外图像移动目标检测[J]. 兵工学报, 2022, 43(S1): 66-73.
[2]	徐熙毅，谭鸽伟，李彪. 基于空变分离的两步聚焦双基曲线合成孔径雷达成像[J]. 兵工学报, 2022, 43(6): 1365-1375.
[3]	王少博，张成，苏迪，冀瑞静. 基于改进YOLOv3和核相关滤波算法的旋转弹目标探测算法[J]. 兵工学报, 2022, 43(5): 1032-1045.
[4]	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.
[5]	徐英，谷雨，彭冬亮，刘俊，陈华杰. 面向合成孔径雷达图像任意方向舰船检测的改进YOLOv3模型[J]. 兵工学报, 2021, 42(8): 1698-1707.
[6]	张永梅，赖裕平，马健喆，冯超，束颉. 基于视频的装甲车和飞机检测跟踪及轨迹预测算法[J]. 兵工学报, 2021, 42(3): 545-554.
[7]	戴文君，常天庆，褚凯轩，张雷，郭理彬. 基于时空卷积特征记忆模型的坦克火控系统视频目标检测方法[J]. 兵工学报, 2020, 41(9): 1708-1718.
[8]	王健，秦春霞，杨珂，任萍. 基于导向重构与降噪稀疏自编码器的合成孔径雷达目标识别[J]. 兵工学报, 2020, 41(9): 1861-1870.
[9]	李响，苏娟，杨龙. 基于改进YOLOv3的合成孔径雷达图像中建筑物检测算法[J]. 兵工学报, 2020, 41(7): 1347-1359.
[10]	何耀民，何华锋，徐永壮，王依繁，苏敬. 基于多特征量的海杂波参数估计[J]. 兵工学报, 2020, 41(3): 517-525.
[11]	梁杰，任君，李磊，齐航，周红丽. 基于典型几何形状精确回归的机场跑道检测方法[J]. 兵工学报, 2020, 41(10): 2045-2054.
[12]	杨龙，苏娟，李响. 基于生成式对抗网络的合成孔径雷达舰船数据增广在改进单次多盒检测器中的应用[J]. 兵工学报, 2019, 40(12): 2488-2496.
[13]	伍薇, 苏娟. 一种基于双向匹配融合的合成孔径雷达景象匹配算法[J]. 兵工学报, 2019, 40(1): 124-133.
[14]	刘德鹏，李正周，曾靖杰，熊伟奇，亓波. 基于多尺度局部对比度和多尺度梯度一致性的红外小弱目标检测算法[J]. 兵工学报, 2018, 39(8): 1526-1535.
[15]	郑恩明，陈新华，宋春楠. 基于全相位预处理的低旁瓣波束形成方法[J]. 兵工学报, 2018, 39(10): 1971-1978.