【作者】 孔丁科；

【导师】 汪国昭；

【作者基本信息】 浙江大学 ， 应用数学， 2010， 博士

【摘要】 图像分割是图像处理领域首要解决的关键问题基于曲线演化理论和水平集方法的几何活动轮廓模型不仅能够自适应地处理拓扑结构变化、提供高精度的闭合分割曲线,而且具有对初始条件不敏感、数值实现简单等优点,目前在理论和应用方面的研究正方兴未艾.本文围绕几何活动轮廓模型在医学、自然和遥感等异质图像的分割问题进行了研究,主要获得了以下一些成果：首先,由于CV模型等传统基于区域的几何活动轮廓模型仅将灰度同质作为区域相似性的测度,致使模型在分割噪声分布大、灰度复杂的自然、遥感等异质图像时难有较好的分割结果.本文提出了一种基于Earth Mover’s Distance (EMD)的快速活动轮廓图像分割模型：一方面给出了基于EMD的区域相似性测度,并引入到模型的定义；另一方面,本文采用基于过分割的规则化和快速曲线演化方法.该模型能有效地抑制演化曲线陷入局部极小,因此具有更好的目标地物提取能力.同时,基于过分割的曲线演化提高了模型的数值实现效率.对合成图像和遥感图像的实验结果证明了模型的有效性.其次,基于局部二值拟合(Local Binary Fitting, LBF)的变分水平集模型的分割质量很大程度上取决于核带宽的选取,容易造成冗余轮廓、边界模糊等问题,本文提出了一种基于边界保持局部拟合的变分水平集模型.该模型引入图像相依的测地时间定义核函数,结合空间距离和图像梯度,自适应地选取邻域采样点；同时,采用基于多波段的图像梯度,并相应地调整图像点的相异性测度,将模型的应用范围扩展至彩色及多光谱图像.实验结果表明,该模型能选取较大核带宽并有效保留潜在的边界信息,从而避免了核带宽的选取问题,提高了演化曲线克服局部极小弱边界等问题的鲁棒性,能较好地应用于灰度异质图像的精确分割；而且,该模型对彩色及多光谱图像的分割也同样有效.接着,我们研究了基于成对相似性的图划分活动轮廓(Graph Partitioning Active Contours,GPAC)模型.传统GPAC模型的连接权函数仅与图像光谱相关,使得模型在低对比度模糊图像的应用存在较大局限,同时,成对相似性的计算量大,模型的数值实现效率不甚理想.针对上述问题,本文引入测地核函数定义连接权函数,结合多相水平集,提出了基于局部图划分的多相活动轮廓图像分割模型.大量自然图像的仿真实验表明：该模型较GPAC模型具有更好的边界获取能力,同时,模型通过测地核函数的带宽约束,实现了水平集演化的窄带加速效果,减少了成对相似性的计算和存储,从而提高了模型的计算效率.最后,为了克服传统几何活动轮廓SAR分割模型高度依赖统计分布假定的缺点,本文结合基于成对相似性的图划分方法和几何活动轮廓模型的优点,提出了基于区域相似性的活动轮廓SAR分割模型：首先将原始图像过分割成同质子区域集;然后结合强度和纹理信息真实度量子区域的成对相似性,并以此定义能量泛函；最后利用基于过分割的规则化和快速曲线演化实现SAR图像的有效分割.实测SAR图像的实验结果表明,该模型能较快、准确地得到SAR图像的分割结果.更多还原

【Abstract】 Image segmentation is a classical and crucial problem in the fields of image understanding. Geometric active contour models impletemented via curve evolution and level set method are used as a powerful tool to address a wide range of image segmentation problems. There are several desirable advantanges of geometric active contour models over classical image setmentation methods. First, geometric active contour models are capable of describing the topology change of contours, and can provide smooth and closed contours as segmentation results. Sencond, geometric active countour are numerically stable and not sensitive to initial conditions.Our works are mainly on some problems in segmenting inhomogeneous images by geometric active contour models and our major contributions in this paper are as follows.Firstly, Classical region-based geometric active contours (e.g. C-V model) only take intensity homogeneity as the similarity measure for regions, and can not obtain satisfactory segmentation results of complicated images. Thus, a fast active contour model based on Earth Mover’s Distance (EMD) is proposed and well adapted to segment images. First, a similarity measure based on EMD is proposed and employed to the segmentation model. Then, a novel regularization and curve evolution method using oversegmentation is enforced to improve the numerical accuracy and evolution efficiency. Experimental results of both synthetic and remote sensing images verify that the algorithm is efficient and accurate.Sencondly, due to the fact that the segmentation accuracy of local binary fitting energy based variational model (LBF model) is highly dependent on kernel bandwidth, and it always lead to unsatisfactory segmentation results (e.g., unnecessary contours, rough boundaries) of inhomogeneous images because of inappropriate bandwidth, an novel edge-preserving local fitting model is proposed and well adapted to segment images with intensity inhomogeneity. First, a geodesic time based kernel using spatial location and spectral gradient is defined, and it provide an adaptive geodesic neighborhood for every pixel. Then, an efficient multichannel gradient based extension combined with adjusted dissimilarity measure is enforced to segment color and multispectral images. Experiments results show that the proposed model could remain potential edge information while using larger bandwidth, and desirable segmentation results of both gray and color images can be obtained.After that, the main attention is paid on Graph Portioning Active Contours (GPAC). Recently, a new region-based active contour model based on pairwise similarity between pixels, i.e., GPAC is presented, and well adapted to segment images with intensity homogeneity. However, it only takes spectral similarity as the cost function between vertices, and can not obtain satisfactory segmentation results for low contrast images with weak boundaries. In order to overcoming this limitation of GPAC model, a novel localized graph-cuts based multiphase active contours model using geodesic kernel based cost function is proposed. Experimental results of natural images verify that the model is efficient and accurate.Finally, Due to the fact that classical active contour models for SAR image segmentation are highly dependent on statistical distributions, a novel active contour model based on pairwise region similarity is proposed and well adapted to segment SAR images. First, the image is initially divided into almost homogenous regions with high accuracy. Then, a region similarity measure using intensity and texture is defined and employed to energy functional. Finally, an efficient regularization and curve evolution method based on oversegmentation is enforced to improve the numerical accuracy and evolution efficiency. Experiments of SAR images show that the proposed model can fast and accurately obtain segmentation results of SAR images.更多还原