基于局部多分辨特征的SAR图像自动目标识别  

SAR image automatic target recognition based on local multi-resolution features

作　　者：汪洪桥[1,2] 孙富春[1] 蔡艳宁[2] 陈宁[1] 裴得利[1] 

机构地区：[1]清华大学计算机科学与技术系,北京100084 [2]第二炮兵工程学院指挥自动化系,西安710025

出　　处：《清华大学学报（自然科学版）》2011年第8期1049-1054,共6页Journal of Tsinghua University(Science and Technology)

基　　金：国家"九七三"重点基础研究项目(2007CB311003);国家自然科学杰出青年基金资助项目(60625304;60621062)

摘　　要：合成孔径雷达(SAR)图像自动目标识别是图像识别领域的一个重要方向。受视觉细胞感受野模型的启发,该文提出了一种从图像局部点出发,对图像进行多分辨分解的图像处理方法。采用一组简单的八邻域正交基对图像进行多级滤波采样处理,得到原图像的多级类Gauss差分图像尺度空间,并将其应用到MSTAR数据集中的SAR图像目标的特征提取;同时,基于多级特征的整合思想,运用基于多尺度核方法的SVM模型,对不同级别图像特征采用不同尺度的核函数分别映射,然后进行合成,实现多类目标的分类。对MSTAR数据集的实验结果表明,该方法具有很高的正确率,并且实现简单快速。此外,该方法还可方便地应用于SAR图像场景中多类、多个目标的分割与自动目标识别,并且对相干斑噪声具有较强的鲁棒性。Synthetic aperture radar （SAR） image automatic target recognition is an important direction in image recognition domain. Inspired by the vision cell receptive field model, an image processing method was developed based on multi resolution decomposition which starts from a local point in the image. The method uses a simple 8-neighborhood orthonormal basis for image multi level filtering and sampling to obtain the difference of Gaussian liking scale space of the original image. The method was then applied to the feature extraction of the SAR image targets in the MSTAR dataset. Based on the integration of multi-level features, a multi-scale kernel method is utilized in the SVM model. The features from different levels of decomposition images are mapped into the feature spaces by kernel {unctions with different scales respectively, with the multiple kernel matrixes then integrated. Tests on the MSTAR datasel show that the method has a high correctness rate and classifies targets simply and rapidly. The method can also be conveniently used for the segmentation and automatic target recognition of multi class/multi-target in SAR image scenes, with relatively strong robustness against the speckles.

关 键 词：自动目标识别(ATR) 感受野机理 局部多分辨分析 多尺度核方法 合成孔径雷达图像 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]