Journal of Geo-information Science >
Review on High Resolution Remote Sensing Image Classification and Recognition
Received date: 2014-12-12
Request revised date: 2015-02-14
Online published: 2015-09-07
Copyright
Target classification and recognition (TCR) of high resolution remote sensing image is an important approach of image analysis, for the understanding of earth observation system (EOS), and for extracting information from the automatic target recognition (ATR) system, which has important values in military and civil fields. This paper reviews the latest progress and key technologies between domestic and international remote sensing image TCR in optical, infrared, synthetic aperture radar (SAR) and synthetic aperture sonar (SAS). The main research levels and the contents of high resolution remote sensing image TCR are firstly discussed. Then, the key technologies and their existing problems of high resolution remote sensing image TCR are deeply analyzed, from aspects such as filtering and noise reduction, feature extraction, target detection, scene classification, target classification and target recognition. Finally, combined with the related technologies including parallel computing, neural computing and cognitive computing, the new methods of TCR are discussed. Specifically, the main framework includes three aspects, which are detailed in the following. Firstly, the predominant techniques of high resolution remote sensing image processing are discussed based on high performance parallel computing. And the hybrid parallel architecture of high resolution remote sensing image processing based on Apache Hadoop, open multi-processing (OpenMP) and compute unified device architecture (CUDA) are also presented in this paper. Secondly, application prospects of TCR accuracy promotion are analyzed based on a thorough study of neuromorphic computing, and the method of multi-level remote sensing image target recognition based on the deep neural network (DNN) is introduced. Thirdly, the model and algorithm of big data uncertainty analysis for remote sensing images are discussed based on probabilistic graphical model (PGM) of cognitive computing, and the multi-scale remote sensing image scene description is given based on hierarchical topic model (HTM). Moreover, according to the related research of multi-media neural cognitive computing (MNCC), we discuss the development trend and research direction of TCR for remote sensing images big data in the future.
LIU Yang , FU Zhengye , ZHENG Fengbin . Review on High Resolution Remote Sensing Image Classification and Recognition[J]. Journal of Geo-information Science, 2015 , 17(9) : 1080 -1091 . DOI: 10.3724/SP.J.1047.2015.01080
Fig. 1 The research levels of high resolution remote sensing image of TCR图1 高分辨率遥感影像TCR问题的研究层次 |
Tab. 1 Typical noise and reduction method of remote sensing images表1 遥感影像的典型噪声与降噪方法 |
影像类型 | 典型噪声 | 降噪滤波方法 |
---|---|---|
光学 | 高斯加性白噪声、椒盐噪声 | 形态学滤波、中值滤波、加权滤波 |
高光谱 | 条带噪声 | 空域自适应滤波、离散剪切波、块匹配滤波 |
SAR和SAS | 相干斑乘性噪声 | PPB、SAR-BM3D、FANS、多视处理、空域滤波、局部统计特性的自适应滤波、变换域滤波 |
红外 | 热噪声、散粒噪声、1/f噪声 | 中值滤波、均值滤波、非线性各向异性滤波、选择性低通滤波 |
Tab. 2 The features of scene and typical targets of remote sensing images表2 遥感影像场景和典型目标的特征 |
特征类型 | 常见的特征 | 特点 |
---|---|---|
几何特征 | 边缘、大小、纹理、灰度、BoVW | 边缘具备不确定性,特征维数高、计算量大 |
灰度统计特征 | 矩特征、不变矩、直方图、灰度共生矩阵、HOG等 | 具备方向不变性,与空间无关,计算量大 |
变换特征 | FFT、DCT、Hough变换、Radon变换、KL变换、小波变换 | 表示能力强、特征维度低、语义不直观、计算复杂度高 |
代数特征 | PCA、ICA、CCA、MDS、LDA、流形学习低维子空间特征 | 能反映图像的内在属性,维度低、稳定,部分特征具备旋转和平移不变性 |
Tab. 3 Methods of target detection of remote sensing images表3 遥感影像目标检测方法 |
目标类型 | 典型目标示例 | 目标特征 | 主要检测算法 |
---|---|---|---|
线状目标 | 道路、机场跑道、河流等 | 边缘、大小、纹理、灰度 | 边缘提取、结构分析检测 |
团块目标 | 飞机、舰船、车辆、坦克等 | 纹理与形状特征、空间区域特性、视觉显著性 | 图像分割、结构分析检测,统计分析检测,分布模型(如CFAR)检测,变换域检测、多核学习、流形学习 |
复合目标 | 机场、车站、港口、油库、桥梁、建筑物 | 点、线和纹理结构特征,目标多特征组合 | 任务驱动和数据驱动综合策略 |
Tab. 4 Algorithms of scene classification and target classification表4 场景分类和目标分类算法 |
分类特征 | 分类算法 | 特点 | |
---|---|---|---|
场景分类 | 颜色、对象特征、方向梯度、密度、特征点、变换域纹理、局部特征 | SVM、BP、MRF、PLSA、KNN、LDA、随机森林、决策树、集成学习 | 常用场景的局部特征进行分类 |
目标分类 | SIFT、目标峰值特征、稀疏表示BoVW、极化散射特性、几何特征、矩特征和纹理特征 | KFCM、RVM、SVM、Fisher、FNN、MRF、DCNN、EM | 多用目标的全局特征进行分类 |
3.3.1 场景分类 |
Tab. 5 Target recognition algorithm of remote sensing images表5 遥感影像的目标识别算法 |
影像类型 | 特征提取 | 主要识别算法 |
---|---|---|
光学 | 几何大小、灰度统计特征、边缘形状特征、纹理特征、视觉感知特征 | 分形模型、模糊理论、粗糙集、SVM、KNN、HDR、概率生成模型、PCNN、分类判决树、RBF |
红外 | 灰度梯度特征、分形特征、LBP纹理 | Adaboost、流形学习、Fisher最佳鉴别 |
SAR | 形状特征、纹理特征、电磁散射特征、变换特征、局部不变特征 | 模板库匹配、散射图匹配、峰值与阴影相似度、Adaboost、3D模型 |
SAS | 阴影形状、活动轮廓、几何特征 | DS证据理论、流形学习、MCMC、BBA |
Fig. 2 The research direction of high resolution remote sensing image of TCR图2 高分辨率遥感影像TCR问题的研究方向 |
Fig. 3 Hybrid parallel architecture of high resolution remote sensing image processing based on Hadoop+OpenMP+CUDA图3 基于Hadoop+OpenMP+CUDA的高分辨率遥感影像混合并行处理架构 |
Fig. 4 Multi-level remote sensing image target recognition based on the deep neural network图4 基于深度神经网络的多层次遥感影像目标识别 |
Fig. 5 Multi-scale remote sensing image scene description based on hierarchical topic model图5 基于层次主题模型的多尺度遥感影像场景描述 |
The authors have declared that no competing interests exist.
[1] |
|
[2] |
|
[3] |
|
[4] |
|
[5] |
|
[6] |
|
[7] |
|
[8] |
|
[9] |
|
[10] |
|
[11] |
|
[12] |
|
[13] |
|
[14] |
|
[15] |
|
[16] |
|
[17] |
|
[18] |
|
[19] |
|
[20] |
|
[21] |
|
[22] |
|
[23] |
|
[24] |
|
[25] |
|
[26] |
|
[27] |
|
[28] |
|
[29] |
|
[30] |
|
[31] |
|
[32] |
|
[33] |
|
[34] |
|
[35] |
|
[36] |
|
[37] |
|
[38] |
|
[39] |
|
[40] |
|
[41] |
|
[42] |
|
[43] |
|
[44] |
|
[45] |
|
[46] |
|
[47] |
|
[48] |
|
[49] |
|
[50] |
|
[51] |
|
[52] |
|
[53] |
|
[54] |
|
[55] |
|
[56] |
|
[57] |
|
[58] |
|
[59] |
|
[60] |
|
[61] |
|
[62] |
|
[63] |
|
[64] |
|
[65] |
|
[66] |
|
[67] |
|
[68] |
|
/
〈 | 〉 |