基于模糊聚类及活动轮廓模型的图像分割技术研究

发布时间：2018-06-08 22:01

本文选题：图像分割 + 水平集　；参考：《西南交通大学》2016年博士论文

【摘要】：随着电子计算机技术的发展,数字图像处理作为一门新兴的学科已经成为信息社会中必不可少的工具。图像分割作为图像处理和计算机视觉、目标跟踪、以及医疗成像的基本课题,其主要目的是把图像分割成一系列具有均匀特性(灰度、颜色、纹理等)的子区域,进而将感兴趣的目标从背景中提取出来。在过去的几十年中,研究人员已经做出很大的努力来解决图像的分割问题,并提出了很多分割算法。然而,由于存在噪声、复杂背景、低信噪比和灰度不均匀性等问题,图像分割仍然是一项具有挑战性的任务。为了改善图像分割算法的性能,国内外学者至今仍在探索和开发新的图像分割算法和分割理论,以得到通用性更好、精度更高的分割结果,这也是本论文选题的意义所在。模糊C均值(Fuzzy C-Means, FCM)算法以最小平方误差和来衡量样本点与聚类中心之间的相似性,利用迭代法优化目标函数,从而实现图像数据的最优聚类。由于成功地将模糊关系引入到聚类方法中,使得FCM算法保留了更多的原始图像信息。基于偏微分方程的活动轮廓模型凭借其自由的拓扑和灵活的结构,得到了众多研究者们的青睐。该方法既利用了低层的图像信息,又融入了高层的理解机制,因而能获得精确的分割结果,具有较强的鲁棒性和实用性。本学位论文主要探讨图像分割领域中的模糊聚类和活动轮廓模型这两类分割方法,在原有算法的基础上进行改进,取得了如下研究成果：(1)提出了一种基于局部变异系数的模糊C均值图像分割算法。首先,利用局部窗口内所有像素点灰度的中值来代替中心点像素的灰度值,对快速广义模糊C均值(Fast Generalized Fuzzy C-Means, FGFCM)算法中的局部灰度相关性矩阵Sg_ij进行修正,提高了算法抑制噪声的能力；然后,引入局部变异系数来重新构造像素间的局部相似性度量,使其能更好地控制邻域内各点对中心像素的权重；最后,利用快速分割的思想使分割过程仅依赖于图像的灰度级,从而可以进一步提高算法的运行效率。与同类方法相比,该算法在一定程度上提升了图像的分割效果,且对噪声有很强的鲁棒性。(2)提出了一种局部交叉熵度量模糊C均值的水平集图像分割算法以及它的简化模型。首先,鉴于交叉熵准则在处理噪声方面有较大的优势,将其取代平方误差和准则来重新构造FCM_S (Fuzzy C-Means with Spatial Constraints)算法的目标函数,这样处理可以自适应地增加或减小样本点属于某个聚类的程度；其次,将改进后的聚类算法融入到变分水平集框架中,使得模型可准确地对像素点进行归类；最后,采用加权迭代法和梯度下降流法来求解本章模型。实验结果显示,相对于传统的水平集算法,该方法能成功地处理弱边缘和灰度不均匀目标,且具有一定的抗噪性。(3)提出了一种基于局部符号差和局部高斯分布拟合能量的活动轮廓模型。该模型以引入图像局部熵的局部符号差(Local Signed Difference, LSD)能量项和局部高斯分布拟合(Local Gaussian Distribution Fitting, LGDF)能量项的线性组合来构造水平集函数的演化力,并运用梯度下降流法来求解该能量泛函,从而驱使曲线向目标边缘运动。与传统的活动轮廓模型相比,新方法能正确地提取灰度不均匀图像中的目标,且对初始轮廓曲线的大小、位置和形状更不敏感。(4)提出了一种基于区域生长初始化的水平集海马图像分割算法。首先,通过自适应区域生长算法来获取海马体的大致区域；其次,对生长出的结果进行形态学操作,以消除内部斑点,并进一步利用轮廓跟踪算子得到有序的海马轮廓线；最后,将此轮廓曲线作为先验信息,运用改进的水平集方法驱动轮廓曲线向目标靠近并在海马边界处停止。实验结果显示,该算法分割出的海马结果与专家手动分割得到的结果非常相近,具有较好的准确性和分割效率。
[Abstract]:With the development of computer technology, digital image processing, as a new subject, has become an indispensable tool in the information society. Image segmentation is the basic subject of image processing and computer vision, target tracking, and medical imaging. The main purpose of image segmentation is to divide the image into a series of uniform characteristics (gray scale, In the past few decades, researchers have made great efforts to solve the problem of image segmentation and put forward a lot of segmentation algorithms. However, there are some problems, such as noise, complex background, low signal to noise ratio and gray scale inhomogeneity, and so on. Cutting is still a challenging task. In order to improve the performance of the image segmentation algorithm, scholars at home and abroad are still exploring and developing new image segmentation algorithms and segmentation theories to get better universal and more accurate segmentation results. This is also the significance of this topic. Fuzzy C mean (Fuzzy C-Means, FCM) algorithm Using the minimum square error and the similarity between the sample point and the cluster center, the optimal clustering of the image data is realized by using the iterative method to optimize the target function. The FCM algorithm preserves more original image information because of the success of introducing the fuzzy relation into the clustering method. By virtue of its free topology and flexible structure, it has been favored by many researchers. This method not only uses the low layer image information, but also integrates the high-level understanding mechanism, so it can obtain accurate segmentation results and have strong robustness and practicability. This dissertation mainly discusses the fuzzy clustering in the field of image segmentation and the fuzzy clustering in the field of image segmentation. The two segmentation methods of active contour model are improved on the basis of the original algorithm, and the following research results are obtained. (1) a fuzzy C mean image segmentation algorithm based on local variation coefficient is proposed. First, the median of all pixels in the local window is used to replace the gray value of the central point pixel, and it is fast generalized. The local gray correlation matrix Sg_ij in the fuzzy C mean (Fast Generalized Fuzzy C-Means, FGFCM) algorithm is modified to improve the algorithm's ability to suppress noise. Then, the local variation coefficient is introduced to reconstruct the local similarity measure between pixels so that it can better control the weight of the center pixels in the neighborhood. Then, the segmentation process is only dependent on the gray level of the image by fast segmentation, which can further improve the efficiency of the algorithm. Compared with the same method, the algorithm improves the image segmentation effect to some extent and has strong robustness to the noise. (2) a kind of local cross entropy measure fuzzy C mean water is proposed. The flat set image segmentation algorithm and its simplified model. First, in view of the greater advantage of the cross entropy criterion in processing noise, it replaces the target function of the FCM_S (Fuzzy C-Means with Spatial Constraints) algorithm by replacing the square error and the criterion, so that the processing can automatically increase or reduce the sample point to be one of them. Secondly, the improved clustering algorithm is incorporated into the variational level set framework, so that the model can be classified accurately. Finally, the weighted iterative method and gradient descending flow method are used to solve the model. The experimental results show that the method can successfully deal with the weak edge relative to the traditional level set algorithm. (3) an active contour model based on local symbol difference and local Gauss distribution fitting energy is proposed. The model is used to introduce local symbol difference (Local Signed Difference, LSD) energy and local Gauss distribution fitting (Local Gaussian Distribution Fittin). G, LGDF) the linear combination of energy terms to construct the evolution force of the level set function, and use the gradient descending flow method to solve the energy functional, so as to drive the curve to the edge of the target. Compared with the traditional active contour model, the new method can correctly extract the target in the gray image image, and the size and position of the initial contour curve. And the shape is more insensitive. (4) a segmentation algorithm based on the regional growth initialization is proposed. Firstly, the adaptive region growth algorithm is used to obtain the rough region of the hippocampus. Secondly, the morphological operation of the results is carried out to eliminate the internal speckles, and the contour tracking operator is further used to get the image segmentation algorithm. In the end, the contour curve is used as a priori information, and the improved level set method is used to drive the contour to the target and stop at the hippocampal boundary. The experimental results show that the results of the proposed algorithm are very close to the results obtained by the expert manual segmentation, and have better accuracy and efficiency.
【学位授予单位】：西南交通大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【参考文献】