基于快速区域合并的SAR图像分割算法

发布时间：2019-06-07 13:13

【摘要】：合成孔径雷达(synthetic aperture radar,SAR)成像系统已经被广泛应用,如目标检测与识别、海洋监视、地形绘制和自然灾情监测等。SAR图像分割是SAR图像信息提取和自动理解的一个重要问题,它通过将一幅SAR图像分割成互不重叠的同质区域来提取场景的结构信息。SAR的相干成像原理使得SAR图像中存在大量随机分布的相干斑噪声,这些相干斑噪声降低了SAR图像的质量,同时,增加了SAR图像分割的难度。本论文主要研究SAR图像分割问题的模型建立及其优化求解问题。提出几种基于区域合并技术的SAR图像分割算法。它们是:1.边缘信息引导区域合并SAR图像分割算法。针对基于区域合并技术的SAR图像分割算法中区域合并的顺序问题,提出一种由边缘信息引导的区域合并技术。首先,利用多方向比例边缘检测算子提取SAR图像的比例边缘强度映射(ratio edge strength map,RESM),提出一种新的阈值处理方法来抑制RESM的均质区域内部的极小值,进而减少了对阈值处理后的RESM进行分水岭变换获得的初始分割的区域个数。然后,利用相邻区域的面积和边缘信息设计一个区域合并优先级函数来引导区域合并的进行,该方法提高了模型参数的估计精度,同时保留图像的强边缘;最后,将边缘信息引导区域合并技术用于求解基于多边形网格和最短描述长度(minimum description length,MDL)准则的SAR图像分割模型。该方法提高了分割结果中区域边缘的定位能力与定位精度。2.基于网格编码和区域合并的MDL准则SAR图像分割算法。建立一种新的基于八邻域链码网格编码和MDL准则的SAR图像分割模型,并用区域合并技术实现模型的快速优化求解。结合比例边缘检测算子和分水岭变换获得SAR图像的初始过分割结果;递归地合并使分割模型减低最快的相邻区域实现分割模型的优化求解。利用区域邻接图(region adjacency graph,RAG)及其最近邻图(nearest neighbor graph,NNG)特性来加速区域合并过程。利用数值指标精确度(P)和命中率(R)来评价分割算法的边缘定位能力。实验结果表明,该方法具有高的边缘定位能力和低的时间复杂度。3.基于G0分布和链码网格的SAR图像分割算法。为了降低SAR图像的场景复杂度对其分割结果的影响,提出一种基于MDL准则的自适应权值SAR图像分割模型。该模型利用G0分布描述SAR图像数据,用链码网格对SAR图像中区域的边缘进行编码。提出一种利用SAR图像数据自适应地估计分割模型的权值的方法。递归地合并初始分割结果中使分割模型降低最快的相邻区域实现模型快速优化求解。实验结果表明,该方法有效地减轻了纹理区域的过分割程度。4.边缘惩罚分层区域合并SAR图像分割算法。利用方向边缘强度信息,建立一种新的边缘惩罚SAR图像分割模型,提出一种最小化该模型的分层区域合并算法。利用多方向比例边缘检测算子提取SAR图像的边缘强度信息,并结合分水岭变换获得SAR图像的高质量的初始过分割结果。利用多边形近似区域边缘,提取边缘的方向,将方向边缘强度映射融入边缘惩罚中,获得惩罚强度与边缘强度成反比的边缘惩罚项。逐渐增大边缘惩罚项的强度,获得由图像特征驱动的分层区域合并算法。利用RAG表示图像分割结果,加速区域合并。实验结果表明:该方法与其它方法相比在性能和效率上都有优势,获得更好的分割结果。5.相对公共边界长度惩罚区域合并SAR图像分割算法。提出一种基于区域合并技术的快速SAR图像分割算法。该算法对阈值处理后的比例边缘强度映射进行分水岭变换实现SAR图像的快速初始过分割,利用提出的基于相对公共边界长度惩罚的区域合并代价和用于快速搜索初始分割中最小权值相邻区域的NNG实现快速区域合并。提出一种新的度量相邻区域之间的相似性的统计相似性度量,该度量具有尺度不变性和对区域尺寸的近似恒虚警特性,将该统计相似性度量与提出的相对公共边界长度惩罚项结合,得到新的区域合并代价。利用RAG和NNG加速区域合并过程。利用数值指标精确度(P)和命中率(R)来度量最终分割结果的边缘定位能力,区域覆盖准则度量最终分割结果的区域检测性能。通过合成的和真实的SAR图像分割试验表明,与两种经典的SAR图像分割算法相比,该算法在效率和性能上均具有优势。
[Abstract]:Synthetic Aperture Radar (SAR) imaging systems have been widely used, such as target detection and identification, ocean surveillance, terrain rendering and natural disaster monitoring. SAR image segmentation is an important problem for SAR image information extraction and automatic understanding. It extracts the structure information of the scene by dividing a SAR image into a homogeneous region that does not overlap each other. The coherent imaging principle of SAR has a large number of coherent speckle noise in the SAR image, which reduces the quality of the SAR image, and increases the difficulty of the SAR image segmentation. This paper mainly studies the model establishment of SAR image segmentation and its optimization solution. In this paper, several SAR image segmentation algorithms based on region merging technology are proposed. They are:1. The edge information guide area combines the SAR image segmentation algorithm. In order to solve the order problem of the region merging in the SAR image segmentation algorithm based on the region merging technology, an area merging technique guided by the edge information is proposed. First, the ratio edge strength map (RESM) of the SAR image is extracted by the multi-directional proportional edge detection operator, a new threshold processing method is proposed to suppress the minimum value inside the homogeneous region of the RESM, And further, the number of areas of the initial segmentation obtained by the watershed transformation of the RESM after the threshold processing is reduced. then, using the area of the adjacent area and the edge information to design a region combined priority function to guide the implementation of the region combination, The edge information guidance area merging technique is used to solve the SAR image segmentation model based on the polygon mesh and the shortest description length (MDL) criterion. The method improves the positioning capability and the positioning precision of the area edge in the segmentation result. The image segmentation algorithm of the MDL based on the mesh coding and the region merging is presented. A new SAR image segmentation model based on eight-neighborhood chain code mesh coding and MDL criterion is set up, and the rapid optimization solution of the model is realized by the region merging technique. The initial over-segmentation result of the SAR image is obtained by combining the proportional edge detection operator and the watershed transform, and the optimization solution of the segmentation model is realized by recursively combining the adjacent regions with the fastest reduction in the segmentation model. The region merging process is accelerated using the region adjacency graph (rag) and its nearest neighbor graph (nng) characteristics. The edge positioning ability of the segmentation algorithm is evaluated by the numerical index precision (P) and the hit rate (R). The experimental results show that the method has high edge positioning ability and low time complexity. The algorithm of SAR image segmentation based on the G0 distribution and the chain code grid is presented. In order to reduce the influence of the scene complexity of the SAR image on the segmentation result, an adaptive weight SAR image segmentation model based on the MDL criterion is proposed. The model uses the G0 distribution to describe the SAR image data, and uses the chain code mesh to encode the edge of the region in the SAR image. A method for adaptively estimating the weight of a segmentation model using SAR image data is presented. And recursively combining the initial segmentation results to enable the segmentation model to reduce the fastest adjacent region to realize the rapid optimization solution of the model. The experimental results show that the method can effectively reduce the degree of oversegmentation of the texture region. The edge penalty layered region is combined with the SAR image segmentation algorithm. In this paper, a new edge penalty SAR image segmentation model is set up by using the direction edge strength information, and a layered region combining algorithm is proposed to minimize the model. The edge intensity information of the SAR image is extracted by the multi-directional proportional edge detection operator, and the high-quality initial over-segmentation result of the SAR image is obtained by combining the watershed transform. By using the edge of the polygonal approximation area, the direction of the edge is extracted, and the edge intensity of the direction is mapped into the edge penalty to obtain an edge penalty term which is inversely proportional to the edge intensity. The intensity of the edge penalty term is gradually increased to obtain a layered region combining algorithm driven by the image feature. The image segmentation result is represented by the RAG, and the acceleration region is combined. The experimental results show that the method is superior to other methods in terms of performance and efficiency, and a better segmentation result is obtained. A SAR image segmentation algorithm is combined with respect to the common boundary length penalty area. A fast SAR image segmentation algorithm based on region merging technology is proposed. the algorithm performs the watershed transform on the scale edge intensity mapping after the threshold processing to realize the fast initial over-segmentation of the SAR image, And fast region merging is realized by utilizing the proposed region combining cost based on the relative common boundary length penalty and the NNG used for fast searching the minimum weight adjacent area in the initial segmentation. A new statistical similarity measure of similarity between adjacent regions is proposed, which has a scale invariance and an approximate constant false alarm characteristic for the size of the region, and the statistical similarity measure is combined with the proposed relative common boundary length penalty term to obtain a new region combining cost. The region merging process is accelerated by the rag and the nng. The edge positioning ability of the final segmentation result is measured by the numerical index accuracy (P) and the hit rate (R), and the region coverage criterion measures the region detection performance of the final segmentation result. The synthetic and real SAR image segmentation test shows that the algorithm has advantages in both efficiency and performance compared to the two classical SAR image segmentation 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】：2014
【分类号】：TN957.52

【相似文献】