【作者】 孙越泓；

【导师】 夏德深；

【作者基本信息】 南京理工大学 ， 模式识别与智能系统， 2010， 博士

【摘要】 图像分割是图像处理和计算机视觉领域的研究热点之一,也是图像识别的重要基础。图像分割技术的优劣决定着最终的图像分析和图像理解的结果和质量,只有通过细致精确的图像分割,才能使得更高层的图像分析和理解成为可能。图像分割问题实质上是一个在复杂的参数空间中寻求最优分割参数的问题。而各种智能优化算法可对复杂的非线性多维数据空间进行快速有效的计算,它不仅可能得到全局最优解,而且会使计算时间大大地缩短。在众多的图像分割算法中,阈值分割和基于聚类的分割是图像分割领域中最常用和应用相当广泛的方法。智能优化算法用于图像分割包括两个方面：最优阈值选取和特征空间聚类。其中最优阈值的选取就是将智能算法作为优化工具,采用迭代的方式计算在某准则下目标函数的最优值,从而求取分割图像的最优阈值。特征空间聚类则是将智能优化算法与基于聚类的图像分割技术结合起来,尽可能避免特征空间的聚类陷入局部最优,同时又尽快地获得最优聚类。本文主要对基于粒子群优化算法的图像分割方法进行研究,研究的目标是综合利用传统与现代分割技术,以粒子群优化算法为寻优工具,建立具有自适应和鲁棒性的分割方法,达到自动、精确和快速分割图像的目的。论文主要在粒子群优化算法的改进、最大类间方差法和最大熵法的推广、基于最大模糊熵和改进粒子群的阈值分割算法框架的建立、基于模糊聚类分析及粒子群优化的聚类分割算法的改进、基于互信息和类距离测度最优的图像自动聚类类别数的获取等方面展开研究工作,包括以下6个方面：(1)对图像分割算法和基于智能优化算法的图像分割研究进展进行综述,提出利用粒子群优化算法进行图像分割研究的可行性和意义；(2)对基本粒子群算法的基本理论、算法的改进思路和研究进展进行综述,分析从保持种群多样性的角度避免算法早熟收敛的优势；(3)对推广的最大类间方差法和最大熵法的基本原理和技术作系统的概述；(4)对基于最大模糊熵的阈值分割方法中的关键技术进行研究,分析最大模糊熵法在模糊集的模糊划分方式、模糊熵的定义以及优化问题的求解中面临的问题,提出解决问题的方案；(5)阐述多种改进的基于模糊C-均值聚类算法的优缺点,分析新的窗口权重构造方法对图像空间结构信息的简单有效利用、分割算法收敛速度和收敛精度的影响；(6)从聚类分割后图像的最大类内距离和平均离散度的单调性变化入手,结合互信息理论和改进的粒子群算法,研究一种新的聚类类别数自动判别方法。本文研究的主要成果及创新点为：(1)设计一种新的基于粒子空间对称分布的PSO(sdPSO)算法,算法的寻优性能得到增强；(2)提出基于邻域灰度对比度的改进二维Otsu法和改进二维最大熵法,用sdPSO算法寻找最优阈值,算法运行时间大大缩短；(3)提出一种新的基于三维最大模糊熵及sdPSO算法的阂值分割方法,能更有效地提高灰度图像的分割效果和分割精度；(4)提出基于空间距离相似性的改进FCM算法(SDSFCM)和基于灰度和空间距离相似性的改进FCM算法(GSDSFCM),采用sdPSO算法寻找最优聚类中心,算法的分割精度、抗噪能力、分割时间等综合性能均得到提高；(5)提出一种以互信息和类距离测度为优化目标,用sdPSO算法为优化技术的新的图像聚类分割算法(MIM-DIS-PSO),分割后的图像具有目标信息准确、内部特征完整和边缘连续清晰等优点,各项评价参数较好；(6)提出一种基于类内距离和平均离散度的单调性的图像聚类分割算法,能自动获得图像的最佳类别数,克服通常一些自动聚类分割算法需要在可能的类别数区间上反复聚类,出现最佳类别数太小或趋向于最大可能类别数的趋势等缺陷,使得获得的最佳类别数更加合理。更多还原

【Abstract】 Image segmentation is one of the research hotspots in the field of both image processing and computer vision, also an essential base for pattern recognition. The technology of image segmentation determines the ultimate results and quality of image analysis and interpretation. Only through thorough and precise image segmentation can the higher-level image analysis and interpretation be made possible. In essence, the image segmentation is just a problem about searching optimal segmentation parameter in the complex parameter space. Various intelligent optimization algorithms are able to perform rapid and effective computation for the complex nonlinear multi-dimensional data space, and lead to not only global optimal solution, but also far less computation time.Threshold segmentation and clustering segmentation are the most frequently and extensively used methods among numerous image segmentation algorithms. Intelligen optimization algorithms are applied to image segmentation in two aspects:Selection of the optimal threshold and feature space clustering. In selection of the optimal threshold, intelligent algorithms are used as the optimization tool, compute the optimum of objective function by iterative method according to some criteria and further obtain the optimal threshold for segmented images. Feature space clustering means combining intelligent optimization algorithms with clustering image segmentation technology to avoid local optimum and simultaneously obtain the optimal clustering as soon as possible. We focus mainly on the image segmentation algorithm based on particle swarm optimization (PSO) and try to accomplish the image segmentation in an automatic, precise and rapid way by establishing segmentation method with self-adaptability and robustness through integrated utilization of traditional and modern techniques with PSO algorithm as the optimizing tool.Our work involves mainly the improvement of PSO optimization, generalization of Otsu and KSW method; establishment of the framework for the threshold segmentation algorithm based on maximum fuzzy entropy and improved particle swarm, improvement of the clustering segmentation algorithm based on fuzzy cluster analysis and PSO, and access to categories of the image automatic clustering based on mutual information and optimal cluster distance measure. There are six aspects as follows:(1) Overview image segmentation algorithms and progress of image segmentation research based on intelligent optimization algorithms; (2) Overview the basic theory, improvement and research progress of bPSO and analyze the advantage of avoiding premature convergence by keeping particles diversified; (3) Outline systematically the basic principle and technology of generalized Otsu method and KSW method;. (4) Study the key technique in the threshold segmentation method based on maximum fuzzy entropy, analyze the fuzzy partition way of fuzzy set by maximum fuzzy entropy method, the definition of fuzzy entropy and problems in solving the optimization and propose the solution; (5) Analyze the advantage and disadvantage of much improved clustering algorithm based on fuzzy C-means as well as the simple effective utilization of image space structure information in new weighted window construction method and its impact on convergence rate and convergence precision of the segmentation; (6) Study a new method for automatically searching the number of cluster categories from monotonicity of maximum intra-category distance and mean deviation in combination with mutual information theory and improved PSO.The main achievement and innovation are summarized as follow. (1) Design a new PSO (sdPSO) algorithm based on symmetric distribution of the particle space with enhanced performance to search optimal. (2) Propose the two-dimensional improved Otsu method and improved KSW method based on neighborhood gray-scale contrast ratio, and the run time is considerably shortened with sdPSO algorithm searching the optimal threshold. (3) Propose a new threshold segmentation algorithm based on three-dimensional maximum fuzzy entropy and sdPSO algorithm, and segmentation result and precision are better for grey-scaled images. (4) Propose two improved FCM algorithms:one is SDFCM based on spatial distance similarity and the other GSDSFCM based on both grey-scale and spatial distance similarity. Algorithm sdPSO is used to search the optimal clustering center. Synthetic performances are all enhanced, such as the segmentation accuracy, anti-noise capacity and segmentation time. (5) Propose a new image clustering segmentation algorithm (MIM-DIS-PSO) taking sdPSO as optimization technology and using the mutual information and intra-category distance measure as the optimization object. The images after segmentation have such merits as accurate target information, complete interior feature as well as continuous and clear edges. Moreover, all quality evaluation parameters for images after segmentation are good. (6) Propose an image clustering segmentation algorithm based on monotonicity of maximum intra-category distance and mean deviation. The algorithm is able to automatically get the optimal number of clustering categories for the image, overcome the trend for too small optimal clustering number or only directing to the maximum clustering number, avoiding the clustering repetition usually occurring for some automatic clustering segmentation at the possible category number interval. Therefore, the optimal number clustering categories obtained is more reasonable.更多还原