基于双噪声相似性模型的SAR图像变化检测

发布时间：2018-03-30 01:11

本文选题：合成孔径雷达　切入点：相似性度量　出处：《西安电子科技大学》2016年博士论文

【摘要】：图像的冗余性是图像处理中最基本的性质之一。在图像处理的很多领域,如图像压缩、修复、分类、检测等应用中,如何有效利用图像的冗余性都是其中的关键问题。为了能有效的利用图像的冗余性,首先需要寻找到足够多的相似点或相似块。因此,如何准确的度量图像的相似性是一个极为重要也非常有意义的课题。本论文即是从图像的相似性度量入手,针对当前噪声块相似性度量中误差较大的问题,在贝叶斯框架下,重点研究了基于统计的双噪声相似性度量模型及其在SAR/PolSAR的降斑和多时相变化检测中的应用,取得了如下的研究成果：(1)提出了一种噪声相似性(noise similarity, NS)的概念,用于改进噪声块的相似性度量,并将其应用于非局部均值(nonlocal means, NLM)滤波器中。噪声相似性的概念表明噪声块之间的相似性不止和信号有关,与其噪声也有关系。基于噪声相似性的概念,本文提出了一种双噪声相似性(double noise similarity, DNS)模型。该模型将图像去噪的问题转化成了两种噪声的抑制问题：一种是外加的加性噪声；另一种是估计误差,即估计值与其真实值之间的误差。前一种对应于去噪问题中的噪声抑制,而后一种则对应于图像细节的恢复。为了验DNS模型的有效性,本文给出了一种迭代的NLM滤波器,在最大后验概率(maximum a posterior, MAP)框架下这两种相似性协同工作。实验表明,本文提出的基于DNS的NLM去噪结果要明显优于其它同类算法的去噪结果。(2)将加性的DNS模型推广到乘性斑点噪声情形中,提出了一种针对于SAR图像的SAR-DNS模型。和加性高斯噪声不同,SAR图像中的乘性斑点噪声是非常复杂的。为了说明这种SAR-DNS模型的有效性,其被用于了SAR图像降斑中。在SAR-DNS滤波器中,斑点噪声的相似性和估计误差的相似性在MAP框架下协同工作达到平衡,使其在斑点抑制的同时也能较好的保持细节。本文还讨论了两种“噪声相似”在SAR图像降斑中的作用,并和相关的滤波算法做了比较。实验结果表明,本文提出的基于SAR-DNS滤波器在SAR图像降斑上的性能比其它同类算法更好。(3)提出了一种基于块相似性度量的SAR变化检测算法SAR-PCD。由于SAR图像中相干斑噪声的干扰,因此降斑过程对于变化检测是非常重要的。但是,在滤波后的图像中依然会有分布不均匀的残留噪声。除此之外,包含有降斑步骤的变化检测结果往往不能通过恒虚警率(constant false alarm rate, CFAR)来调节,而不包含降斑步骤的变化检测算法在强噪声情形下效果又很差。为了克服上述缺点,本文提出了一种改进的SAR图像变化检测框架。这种新框架的主要创新点包括：1)在SAR变化检测中加入了降斑步骤,使得该算法在强噪声情况下也很鲁棒；2)提出了一种新的等效视数(equivalent number of looks, ENL)估计算法,用于估计SAR图像降斑后的ENL3)所提出的基于降斑图像的变化检测结果可以根据不同的CFAR调节。在合成数据和真实数据上的实验都证明了本文提出的SAR-PCD算法的有效性。(4)相比于单极化的SAR数据,全极化的SAR包含的信息更加丰富,应用更为广泛。在本文中,DNS模型被进一步的推广应用于了多极化SAR的环境下中,提出了一种适用于极化SAR环境的PolSAR-DNS相似性,并将其应用于降斑和变化检测中。其中,PolSAR-DNS相似性中的噪声相似性主要用于斑点抑制,而误差相似性主要用于保持细节。最后,在极化SAR降斑和变化检测上的实验都表明,本文提出的PolSAR-DNS相似性比同类型的相似性的结果更好。(5)提出了一种新的基于多层判别式Fisher自编码器(stacked Fisher autoen-coder, SFAE)的SAR变化检测算法。在SFAE框架中,网络的训练包括无监督的逐层特征训练和有监督的微调。这种训练好的网络可以用于单极化/多极化SAR数据集中的无监督实时变化检测。SFAE框架的创新点主要包括两个方面：一是将栈式自编码器推广至乘性斑点噪声环境下的SAR变化检测中。另一个创新点是,通过引入Fisher判别项,确保SFAE网络提取到的特征是有利于分类的。在合成数据和真实数据集上实验都表明SFAE算法在多时相的单极化/多极化SAR变化检测上有很明显的优势。具体来说就是,所提出的SFAE算法能在检测准确度上远高于其它实时的检测算法,而在计算复杂度上又远低于其它的非实时算法。
[Abstract]:The redundancy of the image is one of the most basic properties in image processing. In many fields of image processing, such as image compression, restoration, classification, detection and other applications, how to effectively use the redundancy of the image is one of the key issues. In order to effectively utilize the redundancy of the image, first of all need to find enough the common or similar blocks. Therefore, how to accurately measure the image similarity is a very important and very meaningful topic. This paper is from the image similarity measure of the noise problem in block similarity measurement error, in the Bayesian framework, focuses on the similarity in the SAR/PolSAR model and its speckle reduction and multi temporal change detection of dual noise based on statistics, the main research results are as follows: (1) we propose a noise similarity (noise similarity, NS). Read, for similarity measure improved noise block, and applied to the non local means (nonlocal, means, NLM) filter. Noise similarity concept that the similarity between the signal and the noise block more than, also has a relationship with noise. The concept of noise based on similarity, this paper proposes a double noise similarity (double noise similarity, DNS) model. The model of the image denoising problem is transformed into the problem of two kinds of noise suppression: an additive noise is added; another is the estimation error, namely the estimated value and the true value of the error between the former corresponds to a noise. The denoising problem in inhibition, the latter corresponds to the details of the image restoration. In order to validate the DNS model, this paper presents an iterative NLM filter, the maximum a posteriori (maximum a posterior, MAP) of the two collaborative similarity framework Work. The experiment shows that the proposed DNS based NLM denoising results to denoising results significantly better than other similar algorithms. (2) the additive DNS model is extended to the multiplicative speckle noise case, proposes a SAR-DNS model for SAR images. And the additive Gauss noise is not the same. In the SAR image speckle noise is very complex. In order to illustrate the validity of the SAR-DNS model, which is used for SAR image despeckling. In SAR-DNS filter, the similarity of speckle similarity and error estimation of collaborative work to achieve balance under the framework of MAP, so that it can better maintain at the same time the details of the speckle noise suppression. The paper also discussed two kinds of "noise like" speckle reduction in SAR images, and the related comparison and filtering algorithm. The experimental results show that the proposed SAR-DNS filter in the SAR image despeckling Can better than other similar algorithms. (3) proposed a block based similarity measure SAR change detection algorithm SAR-PCD. due to speckle noise interference in SAR image speckle reduction process, it is very important for change detection. However, the filtered image is still a residual noise uneven distribution. In addition, contains the change detection results of speckle reduction steps are often not by constant false alarm rate (constant false alarm rate, CFAR) to adjust, change and do not contain speckle reduction step detection algorithm in strong noise condition and the effect is very poor. In order to overcome these shortcomings, this paper proposes a SAR improved image change detection framework. The main points of the new framework include: 1) in the SAR change detection added speckle reduction steps, the algorithm in the condition of strong noise is also very robust; 2) put forward a new equivalent number (equiv Alent number of looks, ENL) estimation algorithm is used to estimate the SAR image speckle reduction after ENL3) proposed changes of speckle reduction of image detection results can be adjusted according to different CFAR. Based on the synthetic and real data experiments have proved the effectiveness of the proposed algorithm SAR-PCD (4) SAR data. Compared to the single polarization, fully polarimetric SAR contains more abundant information, has been widely used. In this paper, the DNS model was further popularized in the multi polarization SAR environment, a method is presented for the polarization of the SAR environment PolSAR-DNS similarity and its application in speckle reduction and change detection. Among them, PolSAR-DNS similar noise of similarity is mainly used for speckle suppression, and error of similarity are used to keep the details. Finally, the experiment of reducing speckle and change detection on the polarization of the SAR show that the proposed PolSAR-DNS is similar Than the same type of similarity results better. (5) proposed a new multi discriminant Fisher (stacked Fisher encoder based on autoen-coder, SFAE) SAR change detection algorithm. In the framework of SFAE, including the training of the network layer and the unsupervised training characteristics of the supervision of this fine-tuning. The trained network can be used for single polarization detection /.SFAE framework without changing the real-time supervision of multi polarization SAR data centralized innovation mainly includes two aspects: one is to be extended to take the SAR stack from the encoder of the speckle noise environment detection. Another innovation is that by introducing Fisher discrimination. Ensure that the characteristics of the SFAE network to extract is conducive to classification. In the synthetic data and real data experimental results show that the SFAE algorithm has obvious advantage in multi polarization single / multi polarization SAR change detection. Specifically it The proposed SFAE algorithm is far higher than other real-time detection algorithms in detection accuracy, while the computational complexity is far lower than other non real-time algorithms.

【学位授予单位】：西安电子科技大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TN957.52

【相似文献】