贝叶斯框架下的非参数估计Graph Cuts分割算法  

Graph Cuts segmentation based on Bayesian nonparametric estimation

作　　者：蒋建国[1,2] 郭艳蓉[1] 郝世杰[1] 詹曙[1] 李鸿[3] Ian Ross 

机构地区：[1]合肥工业大学计算机与信息学院,合肥230009 [2]安全关键工业测控技术教育部工程研究中心,合肥230009 [3]安徽医科大学第一附属医院骨科,合肥230022 [4]London Health Science Center, London Ontario Canada N6A 5A5

出　　处：《中国图象图形学报》2011年第6期947-952,共6页Journal of Image and Graphics

基　　金：教育部博士点基金项目(20060359004);教育部留学归国人员科研启动基金项目(413117);安徽省2010高校省级自然科学研究重点项目(KJ2010A193)

摘　　要：假设图像中各像素灰度值是具有一定概率分布的随机变量,由贝叶斯定理,正确分割观测图像等价于求出具有最大后验概率的实际图像估计。在此框架下,提出了一种改进型Graph Cuts图像分割算法。与传统GraphCuts分割算法相比,该算法在模型建立上有两个方面的改进:1)将模糊C均值聚类引入数据约束能量函数来得到各像素在某个标记下的概率,改善了收敛性能;2)使用非参数方法估计图像的统计分布,然后用此统计量构成图像分割的先验概率,并保证分割结果的局部平滑。由于非参数估计是由样本直接估计得到的结果,特别适用于小样本和分布函数不恒定的情况,因此拓展了算法的适用范围。实验结果表明,改进算法在遥感图像分割和医学图像分割中均提高了分割精度,证明了该算法的有效性。Suppose each pixel of an image is a random variable under some kind of probability distribution, according to the Bayes theorem, the segmentation of the original images is equivalent to their maximum a posteriori probability estimation. In this framework, we proposed an improved image segmentation algorithm based on Graph Cuts. The construction of the original Graph Cuts model is improved in two aspects. First, fuzzy C-means clustering is introduced into the energy function of data restriction. With the help of fuzzy clustering method, the energy function＇ s performanee of constringeney is improved. Second, nonparametric method is used to estimate the statistical distribution of the image, which work as the prior probability used in image segmentation. With the presented method, the results of segmentation are guaranteed to be smooth locally. Since the nonparametric estimation is directly evaluated from the samples, and is suitable for situations of small samples and variable distribution functions, the applicability of our algorithm is extended. Experimental results have shown that the proposed algorithm has good performance on segmenting remote sensing images and medical images.

关 键 词：GRAPH CUTS 贝叶斯图像分割 模糊C均值 非参数估计 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]