基于MRF模型SAR图像分割算法研究

发布时间：2018-03-08 09:18

本文选题：SAR图像　切入点：图像分割　出处：《哈尔滨工程大学》2014年硕士论文　论文类型：学位论文

【摘要】：合成孔径雷达(synthetic aperture radar, SAR)是一种二维成像微波探测雷达,具有全天时、全天候、全方位对地球表面进行探测观测的能力。因其方便、较强的观测能力,目前已广泛应用于军事和民用领域。因此,SAR技术成为各国争相研究的热点,其成像分辨率越来越高,但与其不相符的是对SAR图像极低的图像处理率和解译能力,这极大地限制了 SAR图像的利用率。而图像分割是对图像进行理解和解译的前提,因此研究能正确分割SAR图像的算法尤为重要。本文的主要工作是研究基于马尔可夫随机场(markov random field, MRF)模型的SAR图像分割算法,包括空域单尺度MRF模型和小波域多尺度MRF模型两大类。本文首先对SAR系统的成像原理和固有相干斑噪声的产生机理进行了研究,然后对SAR图像灰度的统计分布进行分析,并通过实验仿真得到不同分布模型对SAR图像灰度的拟合度,总结了 SAR图像的特点。最后介绍了图像分割质量的评价准则为后续对SAR图像分割算法的研究作准备。其次,系统且详细的研究了空域单尺度MRF模型的理论知识,包括标号场建模、观测场建模、参数估计方法以及后验分割方法。通过对多幅SAR图像实验仿真,比较了传统的k-均值聚类和空域MRF模型分割方法的优劣,并分析总结了空域MRF模型对SAR图像产生误分割的因素。再次,基于SAR图像纹理信息丰富的特点,引入纹理特征改进空域MRF模型。介绍了纹理的定义、纹理特征的提取方法以及对SAR图像纹理特征进行了分析。另外,由于SAR图像通常呈现模糊性,将模糊理论引入标号场,改善分割结果。实验证明改进的MRF模型分割结果在区域一致性和边缘保持方面表现较好。最后,研究了多尺度MRF模型,包括经典的MSRF模型和小波域分层Markov模型分割方法。实验仿真结果证明,小波域分层由于同时引入了尺度间因果模型和尺度内非因果模型,对图像全局信息和细节信息都给出了较好的描述,其分割结果在区域一致性和边缘保持性均优于空域MRF模型和MSRF模型。
[Abstract]:Synthetic aperture radar (sar) is a two-dimensional imaging microwave detection radar, which has the ability to detect and observe the earth's surface all day, all weather and all directions because of its convenience and strong observation ability. At present, it has been widely used in military and civil fields. Therefore, Sar technology has become a hot topic in many countries, and its imaging resolution is getting higher and higher. However, it is inconsistent with the very low image processing rate and interpretation ability of SAR images. This greatly limits the utilization of SAR images, and image segmentation is a prerequisite for image understanding and interpretation. Therefore, it is very important to study the algorithm to segment SAR image correctly. The main work of this paper is to study the SAR image segmentation algorithm based on Markov random field random (MRF) model. In this paper, the imaging principle of SAR system and the generating mechanism of inherent speckle noise are studied firstly, and then the statistical distribution of gray level of SAR image is analyzed. Through experimental simulation, the fitting degree of different distribution models to SAR image grayscale is obtained, and the characteristics of SAR image are summarized. Finally, the evaluation criterion of image segmentation quality is introduced to prepare for the further research on SAR image segmentation algorithm. The theoretical knowledge of single scale MRF model in spatial domain is studied systematically and in detail, including label field modeling, observation field modeling, parameter estimation and posterior segmentation. In this paper, the advantages and disadvantages of traditional K-means clustering and spatial MRF model segmentation methods are compared, and the factors that cause missegmentation of SAR image by spatial MRF model are analyzed and summarized. Thirdly, based on the rich texture information of SAR image, The definition of texture, the extraction method of texture feature and the analysis of texture feature of SAR image are introduced. In addition, fuzzy theory is introduced into label field because SAR image usually presents fuzziness. The experimental results show that the improved MRF model performs well in terms of regional consistency and edge preservation. Finally, the multi-scale MRF model is studied. The experimental results show that both the inter-scale causality model and the in-scale non-causal model are introduced in the wavelet domain hierarchy, and the simulation results show that the inter-scale causality model and the in-scale non-causality model are introduced in the wavelet domain delamination. The global information and detail information of the image are well described. The segmentation results are better than the spatial MRF model and the MSRF model in regional consistency and edge preservation.
【学位授予单位】：哈尔滨工程大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN957.52

【参考文献】