【作者】 陈晨；

【导师】 尚荣华； 岳博；

【作者基本信息】 西安电子科技大学 ， 工程硕士（专业学位）， 2019， 硕士

【摘要】 合成孔径雷达(Synthetic Aperture Radar,SAR)成像系统是应用较为广泛的遥感成像系统,具有全天时、全天候、高分辨以及信号丰富等特点,相对于其他遥感图像能提供更多有效的信息,因此在军用与民用领域中都起到了重要的作用。SAR图像在很多方面与其他图像有所不同,相干斑噪声是其固有特性,这会对SAR图像处理过程产生负面影响。SAR图像分割作为SAR图像处理中的重要步骤,可以划分、简化图像,从而为后续的识别、检测等步骤带来便利,促进SAR图像处理技术的更新与发展。本文所研究的无监督像素级SAR图像分割任务是指利用SAR图像信息,将其分割为连通但不重叠的同质区域,同一区域内的像素属于同一类。在像素级的SAR图像分割任务中,像素点数目过多导致的运算复杂度过高以及SAR图像普遍存在的相干斑噪声给图像处理过程带来了困难。本文针对这些问题进行了研究与实验,主要内容如下:(1)提出了一种基于缩略图与分层模糊聚类的SAR图像分割算法。首先,利用像素块中的多数相似像素点生成图像的缩略图,缩略图较原图尺寸明显减小,对缩略图进行模糊聚类可以减少算法在聚类运算中耗费的时间。其次,算法结合了图像中的相似近邻信息对缩略图进行模糊聚类,在一定程度上减弱了相干斑噪声的影响。在得到缩略图的聚类结果后,利用缩略图的类标信息对输入SAR图像进行分层次的分割,从而使分割结果更为准确和均匀。(2)提出了一种基于非局部超像素与分层模糊聚类的SAR图像分割算法。首先,该算法利用基于像素块的相似性度量方法将输入图像划分为非局部超像素,并在每个超像素中提取多数相似像素点生成图像的缩略图,其次,结合不规则邻域中的相似近邻信息对缩略图进行模糊聚类,得到超像素对应的类标。最后分层次地分割输入SAR图像。该算法旨在将非局部均值滤波思想与超像素思想相结合,减少两个步骤中的重复性计算,在尽量保持图像细节部分的前提下提高SAR图像分割的效率。(3)提出了一种基于双缩略图与多目标模糊聚类的算法来处理SAR图像分割问题。该算法利用不同尺度的图像信息产生了两个基于超像素的低分辨缩略图,利用求和的方式将两个缩略图的信息结合到模糊聚类的目标函数中,并建立了一个自适应系数矩阵控制两个缩略图信息在聚类操作中的求和系数。此外,引入了多目标模糊聚类对两个缩略图进行分割,有利于进一步权衡保持图像边缘与减弱噪声影响这两个冲突的需求。最后,利用缩略图分割结果确定输入SAR图像所对应的类标。该算法充分利用了不同尺度的图像信息,有效抑制了相干斑噪声的不利影响。更多还原

【Abstract】 Synthetic Aperture Radar(SAR)imaging system is widely used in remote sensing imaging field.Because of its inherent characteristics such as all-time,all-weather,high-resolution and abundant signals,SAR images provide more effective information than other remote sensing images.Therefore,SAR plays an important role in both military and civil fields.SAR images are different from other images in many aspects,and the speckle noise is also its inherent characteristic,which brings challenges to SAR image processing.As an important step in SAR image processing,SAR image segmentation can divide and simplify images to facilitate subsequent recognition,detection and some other steps.So that it help to promote the update and development of SAR image processing technology.The unsupervised pixel-level SAR image segmentation task studied in this thesis is to use SAR image information to divide it into connected but non-overlapping homogeneous regions,and the pixels within the same region belong to the same category.In the pixel-level SAR image segmentation task,the high computational complexity caused by excessive number of pixels and the common speckle noise in SAR image brings difficulties to the image segmenting process.This thesis perfoms research and experiments on these problems,and the main contents are as follows:(1)A new algorithm based on thumbnail and hierarchical fuzzy clustering is proposed for SAR image segmentation.Firstly,the major similar pixels in the pixel patches are used to generate a thumbnail of the image,whose size is obviously reduced compared with the input SAR image.Performing fuzzy clustering on the thumbnail can reduce the time consumed in the clustering operation.Secondly,the algorithm combines the similar neighbor information in the thumbnail to segment the thumbnail using fuzzy clustering algorithm,which also reduces the effect of speckle noise to some extent.After obtaining the clustering results of thumbnail,the label information of the thumbnail is used to perform hierarchical segmentation of the input SAR images,so as to make the segmentation results more accurate and uniform.(2)A new SAR image segmentation algorithm based on non-local superpixel and hierarchical fuzzy clustering is proposed.Firstly,the algorithm uses the similarity measure based on pixel patch to divide the input image into non-local superpixels,and extracts the major similar pixels in each superpixels to generate the thumbnail.Secondly,it performs fuzzy clustering on the superpixels combining with the similar neighbor information of the irregular window in the thumbnail.Finally,the input SAR image is segmented hierarchically using the label information of the superpixels.This algorithm aims to combine the idea of the non-local mean filtering and the superpixel,to reduce the repetitive calculating operation in the two steps,and improve the efficiency of SAR image segmentation on the premise of keeping the details of the image as much as possible.(3)A new algorithm based on double thumbnails and multi-objective fuzzy clustering is proposed to deal with the SAR image segmentation problem.Based on the predivided superpixels,the algorithm uses the image information of different scales to generate two low-resolution thumbnails.Then it combines the information of two thumbnail images into the objective function of fuzzy clustering by means of summation.An adaptive coefficient matrix is established to control the summation coefficient of two thumbnails in clustering operation.In addition,multi-objective fuzzy clustering is introduced to segment two thumbnail images,which is beneficial to further balance the conflicting demands of maintaining image details and reducing the impact of noise.Finally,the thumbnail segmentation result is used to traverse the pixels in the original image for three times to determine the labels of the input SAR image.This algorithm makes full use of the image information of different scales and improves the robustness to the inherent speckle noise.更多还原