基于活动轮廓模型的图像分割方法研究

发布时间：2018-03-09 09:22

本文选题：图像分割　切入点：灰度不均匀　出处：《西南交通大学》2017年硕士论文　论文类型：学位论文

【摘要】：随着计算机技术的日益发展,数字图像处理在信息社会中扮演着不可或缺的角色,成为一股推动社会进步的动力源。在数字图像处理中,图像分割作为关键而基础的一环,已在实际生产生活中起到越来越大的作用。图像分割的目的是将一幅图像按照不同的特征分割成具有均匀特性的子区域,并将感兴趣的目标从背景区域分离出来。目前,学者们已经提出了许多行之有效的图像分割算法。然而,由于存在灰度不均匀、噪声、背景结构复杂等问题,图像分割依旧是一项具有挑战性的任务。基于水平集的活动轮廓模型由于能处理拓扑结构变化,数值实现简单,能完成目标整体分割等特点,近年来成为图像分割领域的研究热点。本文针对图像分割中灰度不均匀、噪声和初始轮廓敏感问题,在LGDF(Local Gaussian Distribution Fitting)模型的基础上进行改进,取得了如下研究成果:(1)针对灰度不均匀和初始轮廓敏感的问题,提出一种基于多尺度局部特征的活动轮廓模型(MS_LGDF)。首先,非均匀光照引起的灰度不均匀在圆形散射状中平滑缓慢变化,该模型采用圆形区域来获取更多的局部信息;其次,针对局部区域灰度的变化程度不同,提出利用多尺度结构与均值滤波器相结合的方法获得多尺度局部灰度信息;最后,通过转换灰度不均匀模型得到一个逼近真实信息的图像,并将其融合进LGDF模型,构造出基于多尺度局部特征的能量泛函。与传统的LGDF模型相比,由于多尺度结构弱化了灰度不均匀的影响,本文提出的MS_LGDF改进模型分割速度快、精度高,对初始轮廓具有较强的鲁棒性。(2)针对灰度不均匀和噪声问题,提出一种基于于局部鲁棒统计的活动轮廓模型(LRS_LGDF)。首先,利用局部区域中像素的四分位距(IQR)、平均绝对偏差(MAD)和中中位数(MED)构造局部鲁棒统计信息,其中,IQR和MAD的作用是锐化目标边界,从而加快轮廓曲线的演化速度,MED则是用于减弱图像噪声;其次,结合LGDF与局部鲁棒统计信息构成新的数据拟合项;最后,将其与另外两个内部约束项一起融入变分水平集方法,构造出LRS_LGDF活动轮廓模型。与传统的LGDF模型相比,由于局部鲁棒统计信息的引入,本文提出的LRS_LGDF改进模型分割速度快、精度高,对噪声具有较强的鲁棒性。
[Abstract]:With the development of computer technology, digital image processing plays an indispensable role in the information society and becomes a power source to promote social progress. The purpose of image segmentation is to divide an image into subregions with uniform characteristics according to different features, and to separate the objects of interest from the background area. Many effective image segmentation algorithms have been proposed by scholars. However, there are many problems such as uneven gray scale, noise, complex background structure, etc. Image segmentation is still a challenging task. The active contour model based on the level set can deal with the topological structure change, the numerical realization is simple, and the target can be segmented as a whole. In recent years, it has become a hot topic in the field of image segmentation. In this paper, we improve the LGDF(Local Gaussian Distribution Fitting-based model for the problems of grayscale inhomogeneity, noise and initial contour sensitivity in image segmentation. A new active contour model based on multi-scale local features is proposed to solve the problem of inhomogeneity of gray scale and sensitivity of initial contour. First of all, a new active contour model based on multi-scale local features is proposed. The gray inhomogeneity caused by non-uniform illumination changes slowly and smoothly in the circular scattering form. The model uses circular region to obtain more local information. Secondly, according to the varying degree of gray level in local region, the degree of variation of gray level in local area is different. The method of combining multi-scale structure with mean value filter is proposed to obtain multi-scale local gray level information. Finally, an image approximating real information is obtained by converting gray scale non-uniform model into LGDF model. The energy functional based on multi-scale local features is constructed. Compared with the traditional LGDF model, the improved MS_LGDF model proposed in this paper has a fast segmentation speed and a high accuracy, because the multi-scale structure weakens the influence of gray inhomogeneity. A new active contour model based on local robust statistics is proposed to solve the problem of grayscale inhomogeneity and noise. Firstly, a new active contour model based on local robust statistics is proposed. The local robust statistical information is constructed by using the quartile distance of the pixels in the local region, the mean absolute deviation (MAD) and the median (MED), in which the role of IQR and MAD is to sharpen the target boundary. In order to speed up the evolution of contour curve, it is used to attenuate image noise. Secondly, combining LGDF with local robust statistical information, a new data fitting term is constructed. Finally, it is combined with the other two internal constraint terms into the variational level set method. The LRS_LGDF active contour model is constructed. Compared with the traditional LGDF model, the improved LRS_LGDF model proposed in this paper has the advantages of fast segmentation speed, high precision and strong robustness to noise due to the introduction of local robust statistical information.
【学位授予单位】：西南交通大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】