基于活动轮廓模型的快速图像分割

发布时间：2018-06-14 22:34

本文选题：图像分割 + 活动轮廓模型　；参考：《西北农林科技大学》2017年硕士论文

【摘要】：图像分割作为数字图像分析的重要步骤之一。在实际的医学图像中,观测图像中经常存在复杂噪声、灰度不均匀以及低对比度的问题,使得精确、快速地分割图像面临诸多难题。而活动轮廓模型以其自适应性、亚像素精度等优点成为研究的热点。本文针对以活动轮廓模型为基础的图像分割方法进行了相对深入的探讨。首先对传统的活动轮廓模型进行了分类归纳和综述,然后重点介绍以区域信息为的几何活动轮廓模型的数学背景以及几个典型区域活动轮廓模型。最后对加速局部区域信息活动轮廓模型的收敛速度进行了重点研究,最终提出了一种基于活动轮廓模型的两步快速分割算法,主要内容如下:1.基于局部区域信息的活动轮廓模型可以抵抗未知复杂噪声的影响,对灰度非同质图像可以达到较好的分割效果,但模型在计算时采取梯度下降法求解,因而计算复杂度高、收敛速度缓慢,同时由于仅利用局部区域信息,使得模型对设定的初始轮廓敏感。本文提出并采用两步分割算法。第一步:采用下采样减少数据量,对采样后的图像进行分割,得到粗分割结果。第二步:将粗分割结果上采样到原始图像的比例,并作为细分割的初始轮廓,进行精细分割。研究结果表明,与一般的区域活动轮廓模型相比,两步分割模型由于第一步分割提供了较好的初始值,能够在极少的步数内得到更精确的结果。2.第一步分割得到的粗轮廓上采样到原始图像的比例,分割结果与真实的目标边界很接近,但仍有一定的差距。为保证第二步分割水平集函数的演化稳定、快速地进行,引入距离正则化能量泛函。在第二步分割的模型中引入一个距离函数dR,它能够校正模型与初始轮廓的偏差,保证曲线演化的稳定性,该距离函数是连接两步分割的关键。3.梯度下降流法是以函数当前点对应的梯度(或近似梯度)的反方向的规定步长距离点进行迭代探索,目前应用范围广,但其收敛速度慢,使得模型迭代过程长。本文算法中采用Bregman迭代算法快速求解,有效地提高了模型的分割效率。采用MATLAB软件进行大量仿真实验,实验结果表明,与CV模型、LBF模型、LIF模型以及LCK模型对比,本文算法能够在极少的迭代次数内分割非同质图像和噪声图像。
[Abstract]:Image segmentation is one of the important steps in digital image analysis. In actual medical images, the problems of complex noise, uneven gray scale and low contrast often exist in the observed images, which make the accurate and fast segmentation of images face many difficulties. The active contour model has become a hot research area because of its adaptability and sub-pixel accuracy. In this paper, the method of image segmentation based on active contour model is discussed. Firstly, the traditional active contour model is classified and summarized, then the mathematical background of geometric active contour model based on region information and several typical regional active contour models are introduced. Finally, a two-step fast segmentation algorithm based on active contour model is proposed. The main contents are as follows: 1. The active contour model based on local region information can resist the influence of unknown complex noise, and achieve better segmentation effect for gray non-homogeneous image. However, the model is solved by gradient descent method, so the computational complexity is high. The convergence rate is slow, and the model is sensitive to the initial profile because only the local region information is used. In this paper, a two-step segmentation algorithm is proposed and used. The first step is to use downsampling to reduce the amount of data, and then segment the sampled image to obtain coarse segmentation results. The second step: the coarse segmentation result is sampled to the original image scale, and as the initial contour of fine segmentation, fine segmentation is carried out. The results show that the two-step segmentation model can get more accurate results in a few steps because the first step segmentation provides a better initial value than the general regional active contour model. In the first step, the ratio of the coarse contour sampled to the original image is very close to the real target boundary, but there is still a certain gap between the segmentation results and the real target boundary. In order to ensure the evolution stability of the second step partition level set function, the distance regularization energy functional is introduced. A distance function, dR, is introduced into the model of the second step segmentation, which can correct the deviation between the model and the initial contour and ensure the stability of the curve evolution. The distance function is the key to connect the two-step segmentation. Gradient descent flow method is an iterative exploration based on the specified step length distance point of the gradient (or approximate gradient) corresponding to the current point of the function. At present, it has a wide range of applications, but its convergence rate is slow, so that the iterative process of the model is long. In this paper, Bregman iterative algorithm is used to solve the problem quickly, which improves the efficiency of model segmentation. A large number of simulation experiments are carried out with MATLAB software. The experimental results show that compared with the CV model, the LBF model, the LIF model and the LCK model, the proposed algorithm can segment heterogeneous and noisy images in a few iterations.
【学位授予单位】：西北农林科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】