基于流形和核方法的SAR自动目标识别研究

【作者】 王兵；

【导师】 杨建宇；

【作者基本信息】 电子科技大学 ， 信号与信息处理（专业学位）， 2012， 硕士

【摘要】 合成孔径雷达(Synthetic Aperture Radar, SAR)是一种高分辨成像雷达,具有全天时,全天候工作,强穿透性等特点,为目标识别提供了可靠的数据依据。SAR自动目标识别(Automatic Target Recogniton, ATR)是在没有人工干预的情况下根据SAR图像自动识别出目标位置并判断其类型属性等信息,SAR ATR在民用和军用领域具有十分重要的作用,已成为国内外研究的热门课题之一。本文从SAR图像预处理和SAR图像特征提取两个方面展开研究,具体工作如下：(1)针对SAR图像相干斑噪声严重,背景区域较大,不利于后续处理的问题,采用相干斑抑制、幂增强、图像分割和质心配准等方法,抑制SAR图像噪声分割出目标图像,降低了后续处理的数据维数,并保留了目标图像的细节信息。(2)针对采用全局线性结构的特征提取方法不利于提取高维数据特征的问题,提出了基于流形学习理论的最大异类距离嵌入(Maximum Interclass Distance Embedding, MIDE)特征提取方法,该算法可以有效解决非线性结构特征提取问题,提高特征的识别率(3)针对一维特征提取方法损失了SAR图像结构信息的问题,提出MIDE的二维扩展方法,即二维最大异类距离嵌入(Two Dimensional Maximum Interclass Distance Embedding,2DMIDE)特征提取方法；针对2DMIDE仅对SAR图像进行垂直方向压缩,特征维数较大的问题,将2DPCA和2DMIDE相结合,提出了2DPCA-based2DMIDE特征提取方法。该方法可以在水平方向和垂直方向对数据压缩,提高特征识别率的同时大幅压缩了特征维数。(4)针对MIDE方法无法对类别信息未知的训练数据提取特征信息的问题,本文提出了一种非监督特征提取算法,最大方差展开嵌入(Maximum Variance Unfolding Embedding, MVUE)。该方法结合了流形学习理论和核方法,通过该方法提取的特征具有较高的识别特性。通过基于MSTAR数据库的实验可知,本文提出的SAR ATR预处理方法和各种特征提取方法均能提高目标图像的识别特性。更多还原

【Abstract】 Synthetic Aperture Radar (SAR) is high-resolution imaging radar, with its all-time, all-time and strong penetration ability. It provides a reliable data source for target recognition. SAR automatic target recognition (ATR) is an effective SAR image interpretation tool without participation of humans. SAR ATR is widely used in civilian fields, and become a hot challenging subject.This dissertation covers two parts of SAR ATR system:SAR image pre-processing and feature extraction. Main content of this dissertation are summarized as follows:Firstly, according to the problem of SAR images, frost filter is utilized to suppress speckle, gray enhancement is used to enhance information in SAR images, two preprocessing CFAR segmentation based on Gaussian distribution is utilized to remove background clutters. Though these methods, the detail information of SAR image is kept, and data dimension can be reduced effectively.Secondly, feature extraction based on global linear structure can not extract feature of high dimensional data, Maximum Interclass Distance Embedding (MIDE) is proposed for feature extraction. This algorithm is based on manifold learning. It can solve nonlinear feature extraction of high dimensional data and improve recognition performance effectively.Thirdly, it may lose structure information of image to use MIDE, Two dimensional Maximum Interclass Distance Embedding (2DMIDE) is proposed. This algorithm can extract2D SAR image directly, it can both keep structure information and improve recognition performance. According to the problem that there is high dimension feature through2DMIDE,2D Principal Component Analysis (2DPCA)-based2DMIDE is proposed. This algorithm can compress SAR images on both horizontal and vertical, which reduce feature dimension effectively. Experiment based on MSTAR shows that2DPCA-based2DMIDE can improve recognition performance notability.Forthly, MIDE can not extract data feature without class information, Maximum Variance Unfolding Embedding (MVUE), an unsupervised feature extraction algorithm, is proposed. This algorithm combines manifold learning with kernel trick, it can improve recognition performance.Finally, experiments based on MSTAR show that SAR ATR pre-processing and several algothms proposed in this dissertation can improve SAR ATR obviously.更多还原