【作者】 林颖；

【导师】 印桂生；

【作者基本信息】 哈尔滨工程大学 ， 计算机应用技术， 2010， 博士

【摘要】 图像分割作为图像处理的基础环节,一直是图像工程领域中的热点和难点问题。几十年来,研究者们不断探索新的图像分割方法,以使求解的问题更逼近于实际应用。水平集方法作为一种新的基于偏微分方程的图像处理方法,具有自由拓扑变换以及易集成多种特征或多种方法的优点,最近受到图像分割领域众多学者的关注。但是,随着图像处理技术应用范围的日益广泛和用户需求的不断提高,图像分割所面临的问题也日趋复杂化。在此背景下,本文针对复杂场景(主要针对灰度不均匀、光谱异质、目标类型多且拓扑关系复杂的场景)中水平集图像分割方法存在的分割精度不高、局部极小解及分割速度慢等主要问题,开展了基于水平集方法的图像分割关键技术研究。论文的主要工作包括:1、为解决灰度不均匀现象对图像分割的干扰问题,开展变分单水平集图像分割方法研究。基于灰度不均匀图像在邻域范围内特征保持相对稳定的特征分析,提出一种新的邻域偏移场估计方法;并在此基础上构建变分单水平集分割模型,以缓解各种程度的灰度不均匀及伴随的弱边界和噪声等现象对图像分割产生的不利影响。此外,开展下行随机优化方法的研究以解决变分水平集模型求解中通常存在的局部极小现象,实现近似全局最优的图像划分。2、为解决多水平集图像分割中参数法概率密度估计的不足及混分现象,开展非参多水平集图像分割方法研究。首先采用Parzen窗非参数密度估计实现复杂场景下样本分析建模,并以此构建基本的非参多水平集分割模型;再利用Gabor小波的纹理分析能力改善模型的分类性能;此外,通过邻域像素的相位特征分析提出一种邻域类相似度函数,并基于此函数构造类别约束项,以缓解遥感图像中光谱异质、噪声等干扰引起的类别错分问题。3、为克服传统的基于像素图像分割方法易受噪音或冗余细节影响这一不足,开展条件随机多水平集图像分割方法研究。条件随机场理论利用概率图模型表达空间邻域关系,并具有概率推理的优势。基于条件随机场理论在数据分类中的优势,本论文将其同水平集方法结合。首先,设计一种竞争多类划分策略,并在此基础上提出新的多类划分条件随机场概率模型;再通过离散变量的连续化映射,将上述竞争条件随机场概率模型融合到多水平集图像分割方法中,以增强水平集分割模型对噪声或细节的抗干扰能力,提高分类精度。4、为解决水平集模型求解效率低的问题,开展快速水平集图像分割方法研究。通过对水平集能量变化与样本分布状态的相关性分析,本文提出一种新的加速方法,用目标样本密度的变化引导曲线快速演化,以达到加快单水平集、多水平集分割模型求解速度的目的。更多还原

【Abstract】 Image segmentation, as an important aspect in image processing, has been a hot but a formidable problem in an image project. For decades, researchers have been constantly seeking and exploring new ways to image partition, in order to the addressed problems more close to practical applications. Level set methods, as a type of novel image processing technology based on partial differential equation, have the advantages of free topology changes and easy integration of multi-cues(color, texture, shape, etc.), the kind of methods has been attracting much attentions from scholars in image segmentation field in recent years. However, problems facing to people in image segmentation are becoming increasingly complex with more extensive applications of image processing technologies and increasing users’demands.In the context, focusing on critical problems in image segmentation to be addressed including: poor accuracy of image segmentation especially in complex scene, local minima problems commonly existing in level set methods, slow speed of segmentation, in a complex scene(mainly as intensity inhomogeneity, spectral hetegeneity, multiple objects with complex topology relation), the thesis presents some subjects to be studied. The subjects are mainly as follows:1. To solve the problem caused by intensity inhomogeneity usually existing in images, this thesis develops a variational level set method for image segmentation. Based on the fact that features of intensity inhomogeneity keep relatively stable in a neighboring range of a given image, a neighbor offset field estimation method is developed. On the basis, a variational level set model for image segmentation has been constructed in order to alleviate disturbances from intensity inhomogenity, the coupled weak boundary, and noises. In addition, the thesis proposes a simple and effective downlink random optimization method so as to search an approximately global optimal solution. The purpose is to solve problems of local minimum occurring to traditional level set methods. 2. To solve the problem of misclassification existing in multiple level set segmentation, this thesis studies on a multiple level set method based on non-parameterization density estimation. Specifically, Parzen Window density estimation is firstly introduced to develop samples analysis and density modelling in complex scenes in order to build a basic non-parameterization multiple level set segmentation framework. Then, a Gabor filter bank is used in texture analysis to improve the classification performance. In addition, a neighbor similarity function is suggested and then a new class constraint energy item is constructed in order to further alleviate misclassification caused by spectral heterogeneity or noises, etc..3. To overcome the shortages of traditional pixel-based segmentation methods, a study on multiple level set method driven by Conditional Random Fields is developed. Conditional Random Fields theory has unique advantages of spatial context analysis using probability graph model and probability inference in image segmentation. Based on the advantages, the thesis combines the theory with a multiple level set method. Specifically, a strategy used to multi-class image partition based on competition between classes is proposed to create a probability model. Secondly, the proposed probability model is combined into level set framework through a mapping operation of discrete variables into continuous ones, in order to enhance the ability of resistance to noises, details interferences and to improve accuracy of image segmentation.4. In order to solve the problem of low-efficiency in solving the level set methods, a fast level set method for image segmentation is studies. Specifically, this thesis presents a novel fasten method of level set model for image segmentation through a correlation analysis between the amount of energy changes of each level set function and sample density in order to fasten curve evolution driven by a change of samples density either for two-phase level set, multi-phase level set segmentation model .更多还原