基于相变和似然性的多相图像分割方法  

Multiphase Image Segmentation Method Using Phase Transition and Likelihood

作　　者：刘纯平[1] CHENFu-HHa 龚声蓉[1] 崔志明[1] 刘全[1] 

机构地区：[1]苏州大学计算机科学与技术学院,江苏苏州215006 [2]佛罗里达大学数学系,佛罗里达美国

出　　处：《计算机学报》2012年第2期375-385,共11页Chinese Journal of Computers

基　　金：国家自然科学基金(60873116,60970015,61170124,61170020,61070223);江苏省自然科学基金(BK2009116,BK2009593);江苏省科技支撑计划项目(BE2009048);江苏省高校自然科学研究项目(09KJA520002);苏州市应用基础研究计划(SYG201116)资助

摘　　要：Sine-Sinc模型是一种基于材料科学中Modica-Mortola物理相变原理的多相图像分割方法.针对该模型分割结果不完全、易受噪声和亮度不均匀性影响的问题,提出了一个改进的Sine-Exp-Gauss多相图像分割模型.基于Sine-Sinc模型,Sine-Exp-Gauss模型用指数函数代替Sine-Sinc模型的Sinc函数,并从分段常数图像假设推广到高斯分布函数图像假设;模型偏微分方程的数值解采用凸函数分裂方法迭代,获得每个相的局部最优解,同时给出一种标准初始化方法使迭代过程易于收敛到理想局部极小值.与Sine-Sinc模型和偏差矫正模型相比,实验结果证明Sine-Exp-Gauss模型在噪声消除和自偏差矫正方面都更加鲁棒.The Sine-Sinc model is a novel approach to multiphase image segmentation built upon the celebrated Modica-Mortola phase transition theory in material science. This model assumes the image to be piecewise constant. In this paper, the improved model, namely the Sine-Exp- Gauss model is proposed through replacing the Sinc function by the exponential function and ex- tending the model to Gaussian-distribution-like image. Since the Sine-Exp-Gauss model is neither quadratic nor convex, for computation the implementation of the proposed model still adopts the convex-concave procedure （CCCP） that has been developed in the literatures of both computation- al nonlinear PDEs and neural computation based. Moreover, we choose normalization information of the original image as an initialization of the iterations so that it helps converge to the ＂true segmentation＂. Experiments on both synthetic and real images are presented. Comparisons are car- ried out between the Sine-Exp-Gauss model and the Sine-Sinc model, as well as the bias-correction model. Experimental results demonstrate that the Sine-Exp-Gauss model is more robust in both denoising and bias-correction.

关 键 词：Sine—Sinc模型 多相图像分割 物理相变 似然性 凸函数分裂 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]