基于卷积神经网络的高分辨率SAR图像飞机目标检测方法

王思雨 高鑫 孙皓 郑歆慰 孙显

王思雨, 高鑫, 孙皓, 郑歆慰, 孙显. 基于卷积神经网络的高分辨率SAR图像飞机目标检测方法[J]. 雷达学报, 2017, 6(2): 195-203. doi: 10.12000/JR17009
引用本文: 王思雨, 高鑫, 孙皓, 郑歆慰, 孙显. 基于卷积神经网络的高分辨率SAR图像飞机目标检测方法[J]. 雷达学报, 2017, 6(2): 195-203. doi: 10.12000/JR17009
Wang Siyu, Gao Xin, Sun Hao, Zheng Xinwei, Sun Xian. An Aircraft Detection Method Based on Convolutional Neural Networks in High-Resolution SAR Images[J]. Journal of Radars, 2017, 6(2): 195-203. doi: 10.12000/JR17009
Citation: Wang Siyu, Gao Xin, Sun Hao, Zheng Xinwei, Sun Xian. An Aircraft Detection Method Based on Convolutional Neural Networks in High-Resolution SAR Images[J]. Journal of Radars, 2017, 6(2): 195-203. doi: 10.12000/JR17009

基于卷积神经网络的高分辨率SAR图像飞机目标检测方法

doi: 10.12000/JR17009
基金项目: 国家自然科学基金青年基金(41501485)
详细信息
    作者简介:

    王思雨(1992–),女,山西人,中国科学院电子学研究所硕士研究生,研究方向为SAR图像飞机目标检测识别。E-mail: siyuwang92@163.com

    高 鑫(1966–),男,辽宁人,北京师范大学理学博士,现任中国科学院电子学研究所研究员,研究方向为SAR场景分类、目标检测识别、解译标注。E-mail: gaxi@mail.ie.ac.cn

    孙皓:孙   皓(1984–),男,山东人,中国科学院大学工学博士,现任中国科学院电子学研究所副研究员,研究方向为遥感图像理解。E-mail: sun.010@163.com

    郑歆慰(1987–),男,福建人,中国科学院大学工学博士,现任中国科学院电子学研究所助理研究员,研究方向为大规模遥感图像解译。E-mail: zxw_1020@163.com

    孙显:孙   显(1981–),男,浙江人,中国科学院大学工学博士,现任中国科学院电子学研究所副研究员,研究方向为计算机视觉与遥感图像理解、地理空间大数据解译。E-mail: sunxian@mail.ie.ac.cn

    通讯作者:

    高鑫   gaxi@mail.ie.ac.cn

  • 中图分类号: TP753

An Aircraft Detection Method Based on Convolutional Neural Networks in High-Resolution SAR Images

Funds: The National Natural Science Foundation of China (41501485)
  • 摘要: 传统的合成孔径雷达(Synthetic Aperture Radar, SAR)图像飞机检测方法一般利用像素对比度信息进行图像分割,从而提取待定目标。然而这些方法只考虑了像素亮度信息而忽视了目标的结构特征,进而导致目标的不精确定位和大量虚警的产生。基于上述问题,该文构建了一个全新的SAR图像飞机目标检测算法框架。首先,针对大场景SAR图像应用需求,提出了改进的显著性预检测方法,从而实现SAR图像候选飞机目标多尺度快速粗定位;然后,设计并调优了含4个权重层的卷积神经网络(Convolutional Neural Network, CNN),实现对候选目标的精确检测和鉴别;最后,因为SAR数据量有限、易导致过拟合,提出4种适用于SAR图像的数据增强方法,具体包括平移、斑点加噪、对比度增强和小角度旋转。实验证实该飞机检测算法在高分辨率TerraSAR-X数据集上效果显著,与传统的SAR飞机检测方法相比,该方法检测效率更高,泛化能力更强。

     

  • 图  1  SAR飞机检测框架的整体流程图

    Figure  1.  Overall flowchart of our SAR aircraft detection framework

    图  2  显著性变换

    Figure  2.  Saliency transformation

    图  3  改进的显著性预检测方法流程图

    Figure  3.  Flowchart of improved saliency-based pre-detection method

    图  4  数据增强示例

    Figure  4.  Demonstration of data augmentation

    图  5  本文CNN的网络结构设计

    Figure  5.  Structure of our CNN

    图  6  训练集示例

    Figure  6.  Demonstration of training set

    图  7  Selective search方法、原始显著性预检测方法和改进后的预检测方法

    Figure  7.  Comparison among the Selective search method, the basic saliency-based method and improved saliency-based method

    图  8  不同方法的ROC曲线图

    Figure  8.  ROC curves of different methods

    图  9  本文框架的检测结果

    Figure  9.  Detection result of our framework

    表  1  CNN参数

    Table  1.   Parameters of our CNN

    层结构 核结构 输出尺寸
    输入层 120×120
    C1 32@5×5 116×116
    S1 2×2 58×58
    C2 64@5×5 54×54
    S2 2×2 27×27
    C3 128@6×6 22×22
    S3 2×2 11×11
    下载: 导出CSV

    表  2  不同预检测方法性能比较

    Table  2.   The performance of different pre-detection methods

    方法 预检测 正确率(%) 候选目标 个数 预检测 时间(s)
    Selective search 100 569 47.50
    原始显著性预检测 94.12 298 6.07
    改进的显著性预检测 100 242 10.67
    下载: 导出CSV

    表  3  不同数据增强方法的检测正确率

    Table  3.   Accuracy rates of CNN with different augmentation methods

    操作 检测正确率(%)
    原始 86.33
    平移 92.01
    斑点加噪 93.74
    对比度增强 93.64
    小角度旋转 92.76
    综合4种增强方法 96.36
    下载: 导出CSV

    表  4  不同网络层数的检测正确率比较

    Table  4.   Accuracy rates of CNN with different number of layers

    网络结构 检测正确率(%)
    C1: 32@5×5, S1, C2: 64@5×5, S2, C3: 128@6×6, S3, C4: 256@6×6, S4 95.99
    C1: 32@5×5, S1, C2: 64@5×5, S2, C3: 128@6×6, S3 (本文) 96.36
    C1: 32@5×5, S1, C2: 64@5×5, S2 93.70
    下载: 导出CSV

    表  5  不同卷积核个数的检测正确率比较

    Table  5.   Accuracy rates of CNN with different number of kernels

    网络结构 检测正确率(%)
    C1: 16@5×5, S1, C2: 32@5×5, S2, C3: 64@6×6, S3 93.93
    C1: 32@5×5, S1, C2: 64@5×5, S2, C3: 128@6×6, S3 (本文) 96.36
    C1: 64@5×5, S1, C2: 128@5×5, S2, C3: 256@6×6, S3 95.05
    下载: 导出CSV

    表  6  不同卷积核大小的检测正确率比较

    Table  6.   Accuracy rates of CNN with different size of kernels

    网络结构 检测正确率(%)
    C1: 32@3×3, S1, C2: 64@3×3, S2, C3: 128@3×3, S3 95.96
    C1: 32@5×5, S1, C2: 64@5×5, S2, C3: 128@6×6, S3 (本文) 96.36
    C1: 32@5×5, S1, C2: 64@5×5, S2, C3: 128@5×5, S3 96.16
    下载: 导出CSV

    表  7  不同方法在同一数据集上的平均检测率

    Table  7.   Average detection rates of different methods on the same dataset

    方法 检测正确率(%)
    CNN 96.36
    SVM 92.64
    HOG+SVM 93.79
    AdaBoost 92.28
    下载: 导出CSV
  • [1] El-Darymli K, McGuire P, Power D, et al. Target detection in synthetic aperture radar imagery: A state-of-the-art survey[J].Journal of Applied Remote Sensing, 2013, 7(1): 071598. doi: 10.1117/1.JRS.7.071598
    [2] Steenson B O. Detection performance of a mean-level threshold[J].IEEE Transactions on Aerospace and Electronic Systems, 1968, 4(4): 529–534.
    [3] Gandhi P P and Kassam S A. Analysis of CFAR processors in homogeneous background[J].IEEE Transactions on Aerospace and Electronic Systems, 1988, 24(4): 427–445. doi: 10.1109/7.7185
    [4] Novak L M and Hesse S R. On the performance of order-statistics CFAR detectors[C]. IEEE 25th Asilomar Conference on Signals, Systems and Computers, Pacific Grove, California, USA, Nov. 1991: 835–840.
    [5] Smith M E and Varshney P K. VI-CFAR: A novel CFAR algorithm based on data variability[C]. IEEE National Radar Conference, Syracuse, NY, May 1997: 263–268.
    [6] Ai Jia-qiu, Qi Xiang-yang, Yu Wei-dong, et al. A new CFAR ship detection algorithm based on 2-D joint log-normal distribution in SAR images[J].IEEE Geoscience and Remote Sensing Letters, 2010, 7(4): 806–810. doi: 10.1109/LGRS.2010.2048697
    [7] di Bisceglie M and Galdi C. CFAR detection of extended objects in high-resolution SAR images[J].IEEE Transactions on Geoscience and Remote Sensing, 2005, 43(4): 833–843. doi: 10.1109/TGRS.2004.843190
    [8] Jung C H, Kwag Y K, and Song W Y. CFAR detection algorithm for ground target in heterogeneous clutter using high resolution SAR image[C]. 3rd International Asia-Pacific Conference on Synthetic Aperture Radar (APSAR), Seoul, South Korea, Sep. 2011: 1–4.
    [9] Olson C F and Huttenlocher D P. Automatic target recognition by matching oriented edge pixels[J].IEEE Transactions on Image Processing, 1997, 6(1): 103–113. doi: 10.1109/83.552100
    [10] Kaplan L M. Improved SAR target detection via extended fractal features[J].IEEE Transactions on Aerospace and Electronic Systems, 2001, 37(2): 436–451. doi: 10.1109/7.937460
    [11] Sandirasegaram N M. Spot SAR ATR using wavelet features and neural network classifier[R]. Defence Research and Development Canada Ottawa (Ontario), 2005.
    [12] Tan Yi-hua, Li Qing-yun, Li Yan-sheng, et al. Aircraft detection in high-resolution SAR images based on a gradient textural saliency map[J].Sensors, 2015, 15(9): 23071–23094. doi: 10.3390/s150923071
    [13] Jao J K, Lee C E, and Ayasli S. Coherent spatial filtering for SAR detection of stationary targets[J].IEEE Transactions on Aerospace and Electronic Systems, 1999, 35(2): 614–626. doi: 10.1109/7.766942
    [14] Zhao Q and Principe J C. Support vector machines for SAR automatic target recognition[J].IEEE Transactions on Aerospace and Electronic Systems, 2001, 37(2): 643–654. doi: 10.1109/7.937475
    [15] Sun Yi-jun, Liu Zhi-peng, Todorovic S, et al. Adaptive boosting for SAR automatic target recognition[J].IEEE Transactions on Aerospace and Electronic Systems, 2007, 43(1): 112–125. doi: 10.1109/TAES.2007.357120
    [16] Keydel E R, Lee S W, and Moore J T. MSTAR extended operating conditions: A tutorial[C]. Proc. SPIE 2757, Algorithms for Synthetic Aperture Radar Imagery III, Orlando, FL, Apr. 1996: 228–242.
    [17] Krizhevsky A, Sutskever I, and Hinton G E. ImageNet classification with deep convolutional neural networks[C]. International Conference on Neural Information Processing Systems, South Lake Tahoe, Nevada, USA, Dec. 2012: 1097–1105.
    [18] Ding Jun, Chen Bo, Liu Hong-wei, et al. Convolutional neural network with data augmentation for SAR target recognition[J].IEEE Geoscience and Remote Sensing Letters, 2016, 13(3): 364–368.
    [19] Chen Si-zhe, Wang Hai-peng, Xu Feng, et al. Target classification using the deep convolutional networks for SAR images[J].IEEE Transactions on Geoscience and Remote Sensing, 2016, 54(8): 4806–4817. doi: 10.1109/TGRS.2016.2551720
    [20] Diao Wen-hui, Sun Xian, Zheng Xin-wei, et al. Efficient saliency-based object detection in remote sensing images using deep belief networks[J].IEEE Geoscience and Remote Sensing Letters, 2016, 13(2): 137–141. doi: 10.1109/LGRS.2015.2498644
    [21] Chen Jie-hong, Zhang Bo, and Wang Chao. Backscattering feature analysis and recognition of civilian aircraft in TerraSAR-X images[J].IEEE Geoscience and Remote Sensing Letters, 2015, 12(4): 796–800. doi: 10.1109/LGRS.2014.2362845
    [22] Srivastava N, Hinton G, Krizhevsky A, et al. Dropout: A simple way to prevent neural networks from overfitting[J].The Journal of Machine Learning Research, 2014, 15(1): 1929–1958.
    [23] Glorot X, Bordes A, and Bengio Y. Deep sparse rectifier neural networks[C]. 14th International Conference on Artificial Intelligence and Statistics, Fort Lauderdale, FL, USA, 2011: 315–323.
    [24] Qian Ning. On the momentum term in gradient descent learning algorithms[J].Neural Networks, 1999, 12(1): 145–151. doi: 10.1016/S0893-6080(98)00116-6
  • 加载中
图(9) / 表(7)
计量
  • 文章访问数:  6355
  • HTML全文浏览量:  1493
  • PDF下载量:  1852
  • 被引次数: 0
出版历程
  • 收稿日期:  2017-01-20
  • 修回日期:  2017-03-05
  • 网络出版日期:  2017-04-28

目录

    /

    返回文章
    返回