基于非负矩阵分解的SAR图像目标配置识别

发布时间：2018-11-11 13:47

【摘要】：合成孔径雷达(Synthetic Aperture Radar,SAR)拥有全天时、全天候等特质,是对地观测和军事侦测的重要技术措施之一。SAR图像目标配置识别作为SAR图像分析与解释的关键技术之一,具有很强的商业和军事价值,日益成为国内外研究的热点。在SAR图像目标配置识别研究中,关键技术之一为图像的特征提取。图像特征提取的主要目的是为了抑制斑点噪声对识别率的影响并且最大限度地体现SAR图像本身具有的稀疏性,所以图像特征提取的好坏将直接影响到识别精度。目前SAR图像目标配置识别的方法大都是基于灰度相关匹配和二维不变矩特征,或是基于目标边缘检测等方法。其主要思想是通过提取图像域或小波域的参数来构建特征矩阵,这种基于全局特征的方法虽然可以获得较好的识别精度,但是受噪声影响较大,运算代价高,速度慢,从而导致实用性不强。本文研究了非负矩阵分解(Nonnegative Matrix Factorization,NMF)这种有效的非负数据处理方法,具有分解速度快,分解结果有明确的物理意义,实现简单的优点,已成为高维数据降维分析领域的一个重要研究方向。本文在深入研究现有NMF算法的基础上,提出了下列三种非负矩阵分解的改进方式:1、一种稀疏约束的非负矩阵分解方法。这种方法充分利用了SAR图像本身所包含的稀疏性,通过对NMF方法进行改进从而提取出图像中的稀疏特征。有效地表征了SAR图像中的稀疏特性,在特征稀疏性和特征出图等方面,性能均优于NMF和现有的稀疏NMF。2、一种近似正交的非负矩阵分解。由于NMF有非负的约束,所以加入正交约束会给矩阵带来稀疏性,能够很好地提取出图像中的稀疏特征。确保了低维特征的非负性和局部性,减小了分解的误差,提高了对稀疏性的调节能力。3、一种平滑约束的稀疏非负矩阵分解方法。这种方法是在非平滑非负矩阵分解方法上加入了平滑约束。因为矩阵中,各个列向量是互不联系和影响的,一列图像信息的变化并不影响其之前一列或之后一列的图像信息,这种特性与马尔科夫随机过程(Markoff random process,MDP)相似,所以可以在特征提取过程中加入马尔科夫随机场(Markov Random Field,MRF)模型。该方法在稀疏特性,特征出图等方面优于现有的同类NMF方法。
[Abstract]:Synthetic Aperture Radar (Synthetic Aperture Radar,SAR) has the characteristics of all-day, all-weather and so on. It is one of the important technical measures of Earth observation and military detection. Target configuration and recognition of SAR image is one of the key techniques of SAR image analysis and interpretation. With strong commercial and military value, it has become a hot research topic at home and abroad. Feature extraction is one of the key techniques in the research of target configuration and recognition in SAR images. The main purpose of image feature extraction is to suppress the influence of speckle noise on the recognition rate and to maximize the sparsity of SAR images, so the quality of image feature extraction will directly affect the recognition accuracy. At present, the methods of object configuration and recognition in SAR image are mostly based on gray correlation matching and two-dimensional moment invariant feature, or based on object edge detection and so on. The main idea of the method is to construct the feature matrix by extracting the parameters of image domain or wavelet domain. Although the method based on global feature can obtain better recognition accuracy, it is greatly affected by noise, has high computational cost and slow speed. As a result, the practicability is not strong. In this paper, we study the nonnegative matrix factorization (Nonnegative Matrix Factorization,NMF), an effective non-negative data processing method, which has the advantages of fast decomposition speed, explicit physical significance and simple implementation. It has become an important research direction in the field of dimensionality reduction analysis of high dimensional data. In this paper, based on the research of the existing NMF algorithm, we propose the following three improved methods of nonnegative matrix factorization: 1. A sparse constraint nonnegative matrix decomposition method. This method takes full advantage of the sparsity of the SAR image itself and extracts the sparse features of the image by improving the NMF method. The sparse properties of SAR images are characterized effectively. In terms of feature sparsity and feature mapping, the performance is better than that of NMF and the existing sparse NMF.2, which is an approximate orthogonal nonnegative matrix factorization. Because NMF has non-negative constraints, adding orthogonal constraints will bring sparsity to the matrix, which can extract sparse features from images. It ensures the non-negativity and locality of the low-dimensional feature, reduces the error of decomposition, and improves the ability to adjust the sparsity. 3, a smooth constrained sparse non-negative matrix decomposition method. In this method, smooth constraints are added to the nonsmooth nonnegative matrix decomposition method. Because in the matrix, each column vector is independent and influence, the change of a column of image information does not affect the image information of the previous column or the next column, which is similar to the Markov stochastic process (Markoff random process,MDP). So we can add Markov Random Field (Markov Random Field,MRF) model in feature extraction process. The proposed method is superior to the existing NMF method in sparsity and feature mapping.
【学位授予单位】：西安电子科技大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN957.52

【相似文献】