【作者】 张玲；

【导师】 张彩明；

【作者基本信息】 山东大学 ， 计算机应用技术， 2012， 博士

【摘要】 图像分割是图像处理、模式识别的基础和关键技术。图像分割的结果将直接影响到后续相关处理。由于图像本身固有的许多不确定性,加大了图像分割的难度。这些不确定信息很多不具有随机性,因此不适合使用概率论、经典数学处理。1965年Zadeh提出的模糊理论是对经典二值逻辑的扩展。模糊理论对不确定性事件和不精确性知识的描述和处理具有先天的优势。基于模糊理论的模糊分割技术己成为图像分割的一个重要分支和研究热点。对基于模糊理论及其扩展的图像分割算法进行研究,以便更好地处理图像中的模糊、不确定性信息,提高图像分割的精确度,具有重要的理论意义和应用价值。本文对模糊分割技术进行了深入、系统的研究,使用经典模糊理论设计了图像分割算法,对模糊图像特别是模糊边缘进行分割识别；为满足实际应用中医学超声图像分割的高精度要求,对模糊分割算法进行了改进。本文率先将中智理论这一崭新的扩展模糊理论引入图像分割,定义了新的三域图像表达方式,不但提高了对图像中不确定性信息的表达能力,而且可以在这个统一的图像分割框架下综合使用多种信息,解决了现有模糊分割算法难以利用多种图像信息的问题,可以更好的处理一些复杂的图像分割问题。本文的研究成果不仅为解决图像分割瓶颈问题提供了新的思路和方法,还可以进一步推广到图像增强、去噪等其它图像处理领域,具有重要的理论意义和广泛的应用前景。本文的主要工作和创新点概括如下：1)设计了基于经典模糊理论的模糊分水岭算法FWST。FWST能够对模糊图像中的模糊边缘进行准确识别,解决了传统WST的过分割现象,抗噪性好,运行速度快。将FWST应用到实际的医学图像分割问题—BUS图像肿瘤全自动分割中,能够得到较理想的分割结果。传统WST方法能够生成封闭、精确的边缘,因此被广泛应用于图像分割中。然而WST方法对噪声敏感,过分割现象严重。本文在分析图像模糊性的基础上,将经典模糊理论与WST算法相结合,通过直方图和模糊最大熵原理对图像模糊化,使用INT算子增强模糊边缘信息,在模糊域上预处理后使用WST方法进行分割。实验结果表明,相比非模糊的边缘检测算法和改进的WST算法,FWST的分割结果更精确。2)定义了基于扩展模糊理论—中智理论的图像三域表示新方式。中智图像引入了新域I,可以更好的表示图像的不确定性；通过T,I,F之间的关联性和相对独立性,可以在这个统一的图像分割框架下综合使用多种信息,提高了图像分割的效果。中智图像作为新的图像表示形式和改进模糊分割算法的途径,具有一定的通用性。由于经典模糊理论的局限性,模糊分割算法对图像的不确定性表达能力不足,使用信息过于单一,无法处理空间等信息。而在图像分割中融合多种信息能够取得更合理的效果。本文概述了模糊理论的扩展方式,介绍了一种崭新的模糊理论扩展形式—中智理论。中智理论是对经典模糊理论及现有的多种扩展模糊理论的概括和泛化,目前它的相关研究主要集中在理论方面。本文率先将中智理论应用到图像处理中,并采用区间值中智集合进行特例化,定义了中智图像和区间值中智图像(统称中智图像),讨论了中智域的选取和中智元素的定义方式,为中智理论在图像分氰中的应用建立了理论基础。3)设计了基于中智图像的灰度和彩色图像区域合并算法。采用三域的表示方式,灰度图像分割算法综合使用了图像的边缘、空间等多种信息,增强了对图像不确定性的表达能力,提高了分割的精确度。在彩色图像分割中,利用T,I,F的相对独立性,算法可以同时使用多个彩色空间上的颜色等多种信息,解决了彩色图像分割中选择彩色空间的难点,综合使用多种信息的能力更加显著。验证了中智图像作为改进分割算法的通用形式,它的优越性和重要的理论意义。根据图像的特点和分割依据,定义了图像的同质性,算法利用中智图像的架构,使用了图像的边缘、全局和局部等多种信息,采用了结合全局和局部的分割策略,使用直方图确定初始聚类中心,使用I衡量像素的不确定性,设定了基于I和区域邻接关系的合并规则。灰度图像分割实验证明,相比经典模糊分割算法,该算法分割区域一致性更好,避免了错分现象,提高了分割精度。将算法推广到彩色图像分割中,T,I,F综合使用多种信息的能力更加明显。选择合适的彩色空间是彩色图像分割的难点,不同的彩色空间各有优缺点,无法相互替代,现有分割算法大多基于单彩色空间。通过应用中智图像这一新的图像表示方式和中智元素间的联系,算法可以同时使用多彩色空间及其上的颜色信息、空间特性等多种信息。采用更多的信息不但提高了分割效果,而且增强了算法的通用性。本文的彩色图像分割算法具有较好的抗噪性,大大降低了现有彩色图像分割算法的过分割现象,分割结果与人的主观视觉感知具有良好的一致性。实验结果表明,相比基于经典模糊理论的分割算法和非模糊分割算法,基于中智图像的分割算法都能获得更好的分割效果。4)设计了基于中智图像的乳腺超声图像肿瘤全自动分割算法。算法基于中智图像架构,使用了乳腺超声图像的纹理、空间等信息,改进了基于经典模糊理论的FWST算法,并结合医学先验知识,制定了肿块定位规则。改进后的算法复杂度低,实时性好,能得到精确的肿瘤边缘。验证了中智图像作为改进分割算法的通用形式,具有实际的应用价值。由于乳腺超声(BUS)图像的低对比度、肿瘤边缘模糊等特性,现有的分割算法大多需要手工标定ROI,且分割结果不理想。将中智图像应用到BUS图像全自动分割中,设计了NWST分割算法。差异性实验统计结果表明,本文的算法与其它全自动分割算法相比,与医生手工标定肿块区域相似度最好,可以准确地分割肿瘤区域。同时,相比FWST, NWST对模糊边缘具有更好的处理能力,算法的时间复杂度较低,能满足实际应用的实时性要求。更多还原

【Abstract】 Image segmentation is the basis and key technology of image processing, pattern recognition. The results of segmentation can directly affect the other steps. Image bears fuzziness in nature, which make the segmentation more difficult. Since the uncertainties are non-randomness, we cannot use classical mathematics or probability theory to handle it.Fuzzy set theory is presented by Zadeh in1965which is an extension of classic set. Fuzzy logic shows great superiority in describing, dealing with fuzzy events and imprecise knowledge. Thus, fuzzy image segmentation has been an important and hot research topic in image segmentation. In order to handle the uncertainty and improve the accuracy of image segmentation, study of fuzzy image segmentation has theoretical significance and application value.This dissertation focuses on fuzzy image segmentation. First, design the algorithm based on classical fuzzy theory and segment the fuzzy image, especially the fuzzy edge. Then, I improve the fuzzy segmentation algorithm in order to meet the high precision of the practical application of medical ultrasound image segmentation. I take the lead in using a new extended fuzzy theory--neutrosophy in segmentation field, define a new three-domain image representation, and setup a novel image segmentation framework. It can solve complex image segmentation problems with higher ability of uncertainty representation. It also solves the problem of information lacking by using varied image information under this segmentation framework. My research provides new idea and method to solve the bottleneck of segmentation, and can be further extended to image enhancement, image de-noising, and other image processing areas, which has theoretical significance and wide application prospects.The main contributions of the dissertation are as followings:1) A fuzzy watershed method (FWST) based on classical fuzzy theory is proposed. The proposed fuzzy method works well for the image with blurred edges and noise, overcomes the over-segmentation problem of traditional watershed algorithm and has fast running speed. It also can be used to breast ultrasound (BUS) segmentation and get good results.Traditional watershed algorithm can get closed、exact edge, thus be widely used in image segmentation. But it is sensitive to noise and often generates over-segmentation. After analyzing the uncertainty of image, the dissertation combines the classical fuzzy theory and watershed algorithm, transforms the image into fuzzy domain by using histogram and maximum fuzzy entropy principle, and then enhanced by the fuzzy intensifying function. The watershed method is applied in the fuzzy domain to get closed edge. Experiments show that the proposed fuzzy method can get more accurate result than other edge detection algorithms.2) A new three-domain image representation is defined which based on extended fuzzy theory--neutrosophy. Neutrosophic image can represent the image uncertainty better by introducing a new domain I; through the correlation and relative independence of T,I,F, it can use varied image information under this segmentation framework, thus improve the segmentation results. Neutrosophic image has generality as a new image representation and improved way of fuzzy segmentation algorithms.Since the limitation of fuzzy theory, fuzzy segmentation has the deficiency of uncertainty expression of image, it also use limited information and cannot handle spatial information. It will obtain more reasonable results by using variety information in image segmentation. The dissertation overviews the exist extensions of fuzzy theory, introduces a new extended fuzzy logic theory-neutrosophy. Neutrosophy is the generalization of classical fuzzy theory and variety existing extended fuzzy theory. At present, its related research mainly focuses on theory. The dissertation introduces the neutrosophy in image processing, and specified it by using interval neutrosophic set, defines neutrosophic image and interval neutrosophic image, discusses neutrosophic domainn、neutrosophic components, and establishes theoretical foundation for applying neutrosophic theory to image segmentation.3) New gray and color image region merge segmentation algorithms based on neutrosophic image is proposed. The gray segmentation algorithm uses varied image information such as edge information spatial information based on three-domain representation, enhances the expression ability of image uncertainty, improve the accuracy of segmentation. The proposed color segmentation algonthm can integrate color and other information in two color spaces, overcome the problem of selecting color space in color segmentation, the ability of using varied information is more obvious. Experiments proved the superiority and important theoretical significance of neutrosophic image as improved approach of fuzzy segmentation algorithms.According to image features and segmentation rules, the proposed algorithm defines image homogeneity, combines global information local information edge information, utilize global and local segmentation strategy. The algorithm calculates the initial cluster center by using histogram, measures the pixel indeterminacy by I, defines region merge rule based on I and region adjacency relationship. Experimental results of gray segmentation prove that the proposed approach is effective than fuzzy logic methods. It can avoid the wrong segmentation phenomenon and get more homogeneous regions. In color segmentation, the independence of T,I,F and the comprehensive ability of using varied information are more obviously. How to select a suitable color space is a problem in color segmentation, different color space has its advantages and disadvantages and cannot replace each other. However, most existing color segmentation algorithms only utilize one color space. Utilizing neutrosophic image and the relationship of neutrosophic components, the proposed algorithm can integrate color and spatial information, as well as global and local information in two color spaces. Using more information can improve the segmentation results and the algorithm’s universality. The proposed color segmentation method has noise-tolerant ability. It reduces the over-segmentation phenomenon, and the segmentation results have good consistency with the subjective visual perception. The experimental results demonstrate that the proposed approaches are more effective and powerful than fuzzy and non-fuzzy methods.4) A new fully automatic BUS tumor segmentation algorithm based on neutrosophic image is proposed. The proposed algorithm based on neutrosophic image framework, uses texture、spatial information of BUS image and improve the algorithm FWST which based on classical fuzzy theory. It combines medical knowledge and designs mass selection rules. The new algorithm can get accurate tumor boundary with less time complexity. Experiments proved the important actual application value of neutrosophic image as improved approach of fuzzy segmentation algorithms.Due to the low contrast of BUS image and blurry tumor boundary, most existing BUS image segmentation methods need region of interests to be manually selected and cannot obtain ideally results. The dissertation applies the neutrosophic image to automatic BUS image tumor segmentation and design a new algorithm NWST. Statistical results of experiment demonstrate that the proposed algorithm can segment tumors more accurately than other automatic BUS segmentation methods and has the best similarity to the manual delineation of tumor region. NWST can handle more blurry edge than fuzzy watershed method with less time complexity, and meet the requirements of real-time in practical applications.更多还原