基于多相位的水平集遥感图像分割

发布时间：2018-06-28 16:28

本文选题：多相位 + 水平集　；参考：《新疆大学》2014年硕士论文

【摘要】：计算机视觉和模式识别的一个关键性问题就是图像分割，同时它也是图像工程中最重要的一个环节。图像分割的目的是将图像划分成一定数量的具有相同性质的子区域，以满足后期图像分析的需求。近些年来，国内外学者研究出了越来越多的图像分割方法，而基于偏微分方程的分割方法比较容易处理曲线的拓扑变化，此外它有着严格的数学理论基础，所以算法稳定性好，，计算精度高，得到了广泛的应用。基于偏微分方程的分割方法，即活动轮廓模型，主要可以分为基于边缘和基于区域两种类型。基于边缘的模型主要利用目标的边缘信息，如曲率，梯度等，但是存在一定的局限性，对于目标边缘模糊断裂的情况不能很好的处理以及对于曲线拓扑结构的变化不能很好的解决。基于区域的模型由于利用了全局的信息，如灰度，形状，颜色等，能够很好的解决基于边缘模型的不足，尤其是水平集方法的出现，它将演化曲线隐式的表达为零水平集函数，将演化的过程转化为偏微分方程的求解，从而避免了演化过程中的跟踪和参数化处理，提供了一个严谨稳定的数学模型，解决了曲线分裂合并的情况，因此基于区域的模型得到了更多的关注。遥感图像具有多灰度级、多目标的复杂结构、边界模糊等特征，上述的基于边缘的模型都不能很好的分割这类图像，所以选取基于区域的水平集方法作为研究对象。本文针对多相位的C-V模型进行了适当的改进，从而使其对遥感图像的分割效果达到更好。具体的研究内容和创新点如下：（1）首先，虽然C-V水平集模型允许初始轮廓可以在图像上任意处，但是位置的不同对分割结果的影响还是相当的巨大。初始轮廓的选择不当，一方面会降低收敛的速度，另一方面甚至可能会分割失败。为了解决这些问题，我们提出了一种新的初始化方法，即选择前面n（n为水平集函数的个数）个面积最大的连通区域的轮廓为初始轮廓，实验证明该方法能使收敛速度更快，且运行起来高效稳定。（2）其次，传统的C-V模型需在演化过程中需要不断重新初始化水平集函数为符号距离函数，这样就造成了计算量的大大增加。为此，我们引入距离惩罚项来解决重新初始化的问题。（3）最后，传统C-V模型分割遥感图像时结果中存在很多琐碎区域，这对于最终的结果来说是冗余的。因此，引入梯度信息，与全局灰度信息一起，加速收敛进程以及消除细小冗余区域。
[Abstract]:Image segmentation is a key problem in computer vision and pattern recognition, and it is also the most important part in image engineering. The purpose of image segmentation is to divide the image into a certain number of sub-regions with the same properties to meet the needs of later image analysis. In recent years, more and more image segmentation methods have been developed by scholars at home and abroad, but the segmentation method based on partial differential equation is easy to deal with the topological changes of curves. In addition, it has a strict mathematical theory foundation, so the algorithm has good stability. It has been widely used because of its high accuracy. The segmentation method based on partial differential equation, active contour model, can be divided into two types: edge-based and region-based. The edge-based model mainly uses the edge information of the target, such as curvature, gradient and so on, but it has some limitations. The fuzzy fracture of target edge can not be handled well and the change of curve topology can not be solved well. Because of the global information, such as gray scale, shape, color and so on, the region-based model can solve the problem of edge model, especially the level set method, which implicitly expresses the evolution curve as zero level set function. The evolution process is transformed into the solution of partial differential equations, which avoids the tracking and parameterization of the evolution process, provides a rigorous and stable mathematical model, and solves the problem of curve splitting and merging. Therefore, more attention has been paid to the region-based model. Remote sensing images have the characteristics of multi-grayscale, multi-target complex structure, fuzzy boundary and so on. These edge-based models can not segment these images well, so the area-based level set method is chosen as the research object. In this paper, the multi-phase C-V model is improved to make the segmentation of remote sensing image better. The specific research contents and innovations are as follows: (1) first of all, although the C-V level set model allows the initial contour to be anywhere on the image, the influence of different positions on the segmentation results is still considerable. If the initial contour is not chosen properly, the convergence speed will be reduced, on the other hand, the segmentation may even fail. In order to solve these problems, we propose a new initialization method, which is to select the contour of the largest area of connected region as the initial contour, which is the number of level set functions in front of n (n. The experimental results show that this method can make the convergence speed faster. Secondly, the traditional C-V model needs to reinitialize the level set function as the symbolic distance function in the evolution process. For this reason, we introduce the distance penalty term to solve the problem of reinitialization. (3) finally, there are many trivial regions in the traditional C-V model, which is redundant for the final result. So the gradient information is introduced to accelerate the convergence process and eliminate the small redundant regions along with the global gray level information.
【学位授予单位】：新疆大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP751

【参考文献】