【作者】 韩守东；

【导师】 吴相林； 赵勇；

【作者基本信息】 华中科技大学 ， 控制科学与工程， 2010， 博士

【摘要】 交互式图像分割,是指按照一定的先验知识和相似性准则将图像划分成具有特殊语义的不同区域,从而在复杂的背景环境中将感兴趣的前景目标分离出来。其应用领域非常广泛,是图像分析、模式识别、计算机视觉、乃至图像理解中的一个关键问题。分割质量的好坏和分割效率的高低将直接影响其它相关应用的可用性和实用性。由于图像分割问题的重要性和复杂性,近年来关于图切分优化方法的研究受到了国内外学者的广泛关注,并成为了交互式图像分割应用的主流方法之一。它具有很多优良特性,比如多特征／约束融合能力、全局最优、数值鲁棒性强、执行效率高、分割加权图的拓扑结构自由和N-D图像分割能力等。传统图切分优化方法在分割低分辨率简单彩色图像时具有很好的分割效果和实时交互性。然而,当使用该方法进行复杂纹理图像分割、多先验模式特征混合智能分割或者高分辨率图像快速分割时,无论是方法本身的适用性、准确性、鲁棒性、还是实时性,都严重限制了其可用性。针对上述问题,本文通过对传统图像分割方法,特别是基于图切分优化的交互式图像分割方法,以及纹理特征建模的主要方法进行了较为深入的总结、分析和研究,提出了一种用于提取场景结构特征的新型纹理建模方法,并结合信息论、黎曼空问几何学、模式分类、数值解析方法等,利用统计学习方法和先验启发式信息,研究如何高质、高量、高效的对彩色及纹理图像进行解析和交互式分割。具体地,本文的创新性研究成果主要包括以下几个部分。首先,通过对传统结构张量的多尺度非线性建模,提出了一种有效的纹理特征建模方法—多尺度非线性结构张量(MSNST)模型。它同时兼顾了结构张量的全方向性压缩描述能力和Gabor小波变换的尺度空问描述能力,并具备不连续保持性滤波特性,为纹理分析的研究提供了一种新的思路。为了保证纹理提取的高效和实用,本文使用了多孔算法和加性算子分裂技术来对模型进行数值实现。另外,本文通过数学分析和实验比较,研究和探讨了关于MSNST特征的距离度量、尺度空间的概率相关性、概率分布模型的空间结构、相似性聚类等概念和方法。然后,针对图切分模型框架的诸多环节,给出了一些优化设计。具体包括优化过程加速(基于多种子图切分的快速模型优化求解、基于多层窄带闭合解的快速图切分算法、基于均值漂移预分割及高斯超像素的快速图切分算法)；基于高分辨距离度量的n-link设计(L*a*b*彩色空间中的共轭测度、MSNST纹理空间中的信息论测度和黎曼测度)；基于高准确性聚类的t-link设计(改进的K-means聚类、谱分解递归聚类、基于分量形式期望最大化的高斯混合聚类)；统计前景和背景在各特征空间的高斯混合模型(GMM),并使用GMM间的信息论距离来设计多特征的能量函数自适应融合策略；计算迭代过程中前景和背景在各特征空问的综合GMM信息论差异度,并通过判断其是否趋于稳定来改进迭代收敛准则；增强分割结果的区域一致性(在平滑能量项中引入滤噪常数、形态学后处理滤噪)等。这些优化设计为其他使用图切分框架的方法和应用提供了改进的思路,以提高它们的性能和适用范围。最后,基于以上的MSNST纹理建模方法和针对图切分模型的各类优化设计,提出了五种实用的交互式彩色及纹理图像分割方法。解决了传统分割方法中存在的诸多不足,比如模式特征信息单一、空间度量过于简化、统计模型的学习和更新准确性不够、计算复杂度和内存消耗偏高、人机交互量和参数设置过多等,对复杂自然图像的分割提供了更好的支持,具有适用范围广、分辨能力强、统计描述准、计算负担低、用户依赖少等优势特性。本文通过大量的仿真实验验证了多尺度非线性结构张量模型、图切分框架的优化和改进方法、彩色及纹理图像的分割技术等的实效性和可用性。这些研究成果不但可以深化纹理分析和图切分优化中的相关研究,而且可以在民用和军用领域扩大图像分割的应用前景。更多还原

【Abstract】 Interactive image segmentation is often described as the process of separating an image into different regions with special semantics according to some prior knowledge and similarity criterion, and then extracting the foreground objects of interest in the complex background environment. It has found a very wide range of applications, which is a key problem in the research for image analysis, pattern recognition, computer vision, and even image understanding. The quality and efficiency of segmentation will have a direct impact on the usability and practicability of other related applications.Due to the significance and complexity of image segmentation, the research about Graph Cuts optimization methods has attracted wide attention by scholars both at home and abroad in recent years, and has become one of the most popular methods used for the application of interactive image segmentation. It has many excellent features, such as ability to fuse a wide range of visual cues and constraints, global optima, numerical robustness, practically efficiency, unrestricted topological properties of weighted graph for segments, and applicability to N-D image segmentation etc. The traditional Graph Cuts optimization methods have good segmentation performance and real-time interaction when segmenting the simple color images with low resolution. However, whether the applicability, accuracy, robustness, or the real-time property of these methods, they all severely limit the usability of these methods when applying them for the complex texture image segmentation, multiple prior pattern features based hybrid intelligent segmentation, or high resolution images based fast segmentation. To address the above problems, this work provides a deeper summarization, analysis and study about the traditional image segmentation methods, especially the Graph Cuts optimization based interactive image segmentation methods, and the main methods used for modeling texture features. Additionally, a new texture modeling method used for extracting the scene structure features is proposed. By combining with the information theory, Riemannian geometry, pattern classification, and numerical analytic method, and making full use of the statistical learning methods and prior heuristic information, this work also researches on how to analyze and interactively segment the color and texture images with high quality, high amount, and high efficiency. In details, the main innovative research achievements of this work can be described as follows.Firstly, an effective texture feature modeling method (multi-scale nonlinear structure tensor, MSNST) is proposed based on the multi-scale nonlinear modeling of the traditional structure tensor. MSNST has both the omni-directional compression description ability of structure tensor and the description ability of Gabor wavelet transform in scale space. Meanwhile, it has the property of filtering in maintaining discontinuity, and provides a completely new idea for the research of texture analysis. In order to improve the efficiency and practicability of the texture extraction, this work uses the "a trous" algorithm and additive operator splitting technique to reduce the complexity of the model’s numerical implementation. In addition, by means of the mathematical analysis and experimental comparison, this work studies and discusses some concepts and methods of MSNST feature, which includes dissimilarity measure, probability correlation in scale space, space structure of probability distribution model, and similarity clustering, etc.Secondly, for several parts of the Graph Cuts model framework, this work presents some optimization designs, which detailly include the acceleration of optimization process (fast optimization and solution of models based on multi-seeds Graph Cuts, fast Graph Cuts algorithm based on multi-level banded closed-form, fast Graph Cuts algorithm based on mean shift pre-segmentation and Gaussian super-pixel);high discrimination dissimilarity measure based n-link design (conjugate measure in L*a*b* color space, information theory measure and Riemannian measure in MSNST texture space); high accuracy clustering based t-link design (improved K-means clustering, spectral decomposition based recursive clustering, component-wise expectation-maximization for Gaussian mixtures clustering); computing the Gaussian mixture model (GMM) statistics of foreground and background in each feature space, and then designing the adaptive integration strategy of energy functions for multiple kinds of features based on the information theory distance between GMMs; computing the comprehensive information theory dissimilarity of foreground GMM and background GMM in each feature space during iterating, and then improving the iteration convergence criterion by estimating whether it is stabilizing; enhancing the regional uniformity of segmentation results (including denoising constant in the smoothing energy term, morphological post-processing based denoising) and so on. These optimization designs provide some ideas for enhancing the other Graph Cuts framework based methods and applications, and help them improve the performance and applicability.Finally, based on the above MSNST texture modeling method and all the optimization designs of Graph Cuts model, this work proposes five practical interactive color and texture image segmentation methods, which solve many problems of the traditional segmentation methods. For example, single pattern feature information, simple choices of dissimilarity measure, low accuracy of learning and updating the statistical model, high computational complexity and memory consumption, many user interactions and parameter settings. However, the proposed segmentation methods provide better support for the segmentation of complex real natural scene images, and have the advantages in terms of wide range of applications, powerful discriminative ability, accurate statistical description, low computational burden, and little user dependent.In this dissertation, a large number of simulation experiments have been presented to verify the practicability and usability of multi-scale nonlinear structure tensor model, optimization and improvement methods of the Graph Cuts framework, and color and texture image segmentation techniques. These research achievements can not only enforce the related study in texture analysis and Graph Cuts optimization, but also extend the application prospect of image segmentation in both civilian and military domains.更多还原