基于区域的SAR分割算法及其在SAR图像分类中的应用
发布时间：2018-08-06 10:41
【摘要】：合成孔径雷达(SAR)因为独特的成像方式使得其具有全天时、全天候等特点。随着近年来SAR传感器的不断发展,高分辨率SAR图像处理逐渐成为了SAR领域的研究热点。图像的分割和分类一直以来都是图像解译的关键步骤,而SAR图像因其固有的相干斑噪声以及高分辨率下的同谱异质现象使得SAR图像自动解译非常困难。基于相似度限制区域合并的SAR图像分割算法为高分辨率下的SAR图像分析提供了新思路。基于区域的方法可以有效减少SAR图像固有的相干斑噪声对分割的影响,通过对相似度进行限制可以有效地实现SAR图像的多层次分割,以满足不同尺度的分割需求。基于区域的分割结果进行分类,可以有效地减少相干斑噪声的影响,将分割后的区域作为基本的分类单元,可以引入区域的纹理、结构以及尺寸形状等信息进行分析,有效的提高SAR图像的分类精度。 本文对遥感图像中的分割方法进行总结和研究,结合SAR图像的特点,提出并设计了一种鲁棒性高的适用于不同SAR图像的分割算法,并且将其引入到全极化SAR图像分割当中,最后基于单极化和全极化的分割结果进行SAR图像分类。本文的主要工作如下： (一)论文回顾和总结了SAR图像的分割方法和分类方法,指出了基于区域合并分割算法的优势,以及其在SAR图像分类中应用的可行性。 (二)改进了基于区域合并的分割算法。首先利用形态学分水岭算法获得边缘保持良好的过分割区域,在过分割区域的基础上进行相邻区域相似性度量,度量的方法在传统灰度均值的基础上引入了统计模型度量的方法。然后引入了最邻近图对SAR过分割区域进行表示,通过边界像素进行各个区域邻域的计算并存储。区域合并采用的是全局最优的合并方式,以保证获得的最终结果是全局最佳的。在合并过程中每一次的区域合并都要将邻接表进行一次更新。每次只对一个区域合并,无需对所有区域的邻域通过边界像素进行更新,从而加快区域合并速度。对相似度进行限制可实现SAR图像的多尺度分割。将相似度限制合并完成的分割结果进行后处理,采用全局最优的合并方式将尺寸小于某一限定区域的小区域进行合并,得到最终的分割结果。最后将分割结果进行分析,并与常用的SAR图像分割方法进行比较,验证算法的有效性。 (三)在区域合并的分割方法基础上,实现基于SVM分类器的单极化和全极化的SAR图像分类。有别于传统的基于像素的SAR图像分类方法,基于区域的分类算法可以有效降低相干斑噪声的影响。利用SVM分类方法结合单极化和全极化的分类特征进行分类,在全极化SAR图像分类中引入了多特征组合的方法来提高分类精度。最后将分类结果与基于像素的SAR图像分类结果进行比较,可以发现基于区域合并分割结果的SAR图像分类比传统的基于像素的分类结果具有更高的分类精度。
[Abstract]:Synthetic Aperture Radar (SAR) (SAR) has the characteristics of all-day, all-weather and so on because of its unique imaging mode. With the development of SAR sensors, high resolution SAR image processing has become a hotspot in the field of SAR. Image segmentation and classification have always been the key steps of image interpretation, but SAR images are very difficult to interpret automatically because of their inherent speckle noise and homospectral heterogeneity at high resolution. The SAR image segmentation algorithm based on similarity constrained region merging provides a new idea for SAR image analysis with high resolution. The region-based method can effectively reduce the effect of the inherent speckle noise on the segmentation of SAR images. By limiting the similarity, the multi-level segmentation of SAR images can be realized effectively to meet the needs of different scales of segmentation. Classification based on the segmentation results can effectively reduce the effect of speckle noise. The segmented region can be used as the basic classification unit, and the texture, structure, size and shape of the region can be analyzed. The classification accuracy of SAR image is improved effectively. In this paper, the segmentation methods in remote sensing images are summarized and studied. According to the characteristics of SAR images, a robust segmentation algorithm suitable for different SAR images is proposed and designed, and it is introduced into the segmentation of fully polarized SAR images. Finally, SAR images are classified based on the segmentation results of single polarization and full polarization. The main work of this paper is as follows: (1) the paper reviews and summarizes the segmentation methods and classification methods of SAR images, points out the advantages of region merging segmentation algorithm and the feasibility of its application in SAR image classification. (2) the segmentation algorithm based on region merging is improved. Firstly, the morphological watershed algorithm is used to obtain a well-maintained edge over-segmented region, and the similarity of adjacent regions is measured on the basis of over-segmented region. The method of statistical model measurement is introduced based on the traditional gray mean. Then the nearest neighbor graph is introduced to represent the SAR over-segmented region, and the boundary pixels are used to calculate and store the neighborhood of each region. The region merging adopts the global optimal merging method to ensure the final result is the global optimal. In the process of merging each region merge must update the adjacent table once. Only one region is merged at a time without updating all the neighborhood pixels through the boundary so as to accelerate the speed of region merging. The multi-scale segmentation of SAR images can be realized by limiting the similarity. After processing the segmentation result which is completed by similarity restriction merging, the final segmentation result is obtained by using the global optimal merging method to merge the small area smaller than a certain limited region. Finally, the segmentation results are analyzed and compared with the common SAR image segmentation methods to verify the effectiveness of the algorithm. (3) based on the segmentation method of region merging, SAR image classification based on single polarization and full polarization is realized based on SVM classifier. Different from the traditional pixel based SAR image classification algorithm, the region-based classification algorithm can effectively reduce the effect of speckle noise. The SVM classification method is used to combine the single and full polarization classification features, and the multi-feature combination method is introduced to improve the classification accuracy in the full polarization SAR image classification. Finally, by comparing the classification results with the pixel based SAR image classification results, it can be found that the classification accuracy of the SAR image based on region merging segmentation is higher than that of the traditional pixel based classification results.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN957.52
本文编号：2167481
[Abstract]:Synthetic Aperture Radar (SAR) (SAR) has the characteristics of all-day, all-weather and so on because of its unique imaging mode. With the development of SAR sensors, high resolution SAR image processing has become a hotspot in the field of SAR. Image segmentation and classification have always been the key steps of image interpretation, but SAR images are very difficult to interpret automatically because of their inherent speckle noise and homospectral heterogeneity at high resolution. The SAR image segmentation algorithm based on similarity constrained region merging provides a new idea for SAR image analysis with high resolution. The region-based method can effectively reduce the effect of the inherent speckle noise on the segmentation of SAR images. By limiting the similarity, the multi-level segmentation of SAR images can be realized effectively to meet the needs of different scales of segmentation. Classification based on the segmentation results can effectively reduce the effect of speckle noise. The segmented region can be used as the basic classification unit, and the texture, structure, size and shape of the region can be analyzed. The classification accuracy of SAR image is improved effectively. In this paper, the segmentation methods in remote sensing images are summarized and studied. According to the characteristics of SAR images, a robust segmentation algorithm suitable for different SAR images is proposed and designed, and it is introduced into the segmentation of fully polarized SAR images. Finally, SAR images are classified based on the segmentation results of single polarization and full polarization. The main work of this paper is as follows: (1) the paper reviews and summarizes the segmentation methods and classification methods of SAR images, points out the advantages of region merging segmentation algorithm and the feasibility of its application in SAR image classification. (2) the segmentation algorithm based on region merging is improved. Firstly, the morphological watershed algorithm is used to obtain a well-maintained edge over-segmented region, and the similarity of adjacent regions is measured on the basis of over-segmented region. The method of statistical model measurement is introduced based on the traditional gray mean. Then the nearest neighbor graph is introduced to represent the SAR over-segmented region, and the boundary pixels are used to calculate and store the neighborhood of each region. The region merging adopts the global optimal merging method to ensure the final result is the global optimal. In the process of merging each region merge must update the adjacent table once. Only one region is merged at a time without updating all the neighborhood pixels through the boundary so as to accelerate the speed of region merging. The multi-scale segmentation of SAR images can be realized by limiting the similarity. After processing the segmentation result which is completed by similarity restriction merging, the final segmentation result is obtained by using the global optimal merging method to merge the small area smaller than a certain limited region. Finally, the segmentation results are analyzed and compared with the common SAR image segmentation methods to verify the effectiveness of the algorithm. (3) based on the segmentation method of region merging, SAR image classification based on single polarization and full polarization is realized based on SVM classifier. Different from the traditional pixel based SAR image classification algorithm, the region-based classification algorithm can effectively reduce the effect of speckle noise. The SVM classification method is used to combine the single and full polarization classification features, and the multi-feature combination method is introduced to improve the classification accuracy in the full polarization SAR image classification. Finally, by comparing the classification results with the pixel based SAR image classification results, it can be found that the classification accuracy of the SAR image based on region merging segmentation is higher than that of the traditional pixel based classification results.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TN957.52
【参考文献】
相关期刊论文 前4条
1 薄华;马缚龙;焦李成;;基于免疫算法的SAR图像分割方法研究[J];电子与信息学报;2007年02期
2 田小林;焦李成;缑水平;;基于ICA优化空间信息PCM的SAR图像分割[J];电子与信息学报;2008年07期
3 马文萍;焦李成;尚荣华;;免疫克隆SAR图像分割算法[J];电子与信息学报;2009年07期
4 张红蕾;宋建社;张宪伟;;基于二维模糊熵的GA在SAR图像分割中的应用[J];计算机工程;2007年05期
，本文编号：2167481
本文链接：https://www.wllwen.com/kejilunwen/wltx/2167481.html
