【作者】 贺锦鹏；

【导师】 孙枫；

【作者基本信息】 哈尔滨工程大学 ， 精密仪器及机械， 2011， 博士

【摘要】 图像分割是数字图像处理领域的一个重要分支,其结果对图像分析和图像理解有重要意义。纹理是图像中普遍存在而又难以描述的特性,图像中的纹理根据物体自身的属性反映物体表面颜色和灰度的某种变化,从而纹理成为描述不同尺度下不同物体表面的一种明显特征,在图像分割中占有举足轻重的作用,纹理分割也成为一种重要的研究方向受到国内外大量研究人员重视,并在实际中得到了广泛的应用。本文以彩色纹理的分割方法为研究内容,对目前广泛使用的一些纹理图像分割算法进行了认真的学习和总结,研究如何有效地提取图像的纹理特征,并将之应用于纹理图像分割中,力图找到一种可靠、稳定、实用的纹理图像分割方法。彩色纹理分割分为彩色纹理的特征提取以及基于特征向量的一致性分割,本文围绕以上两个方面进行研究。论文首先从纹理研究的意义出发,通过对现有典型纹理定义的分析和比较,总结出纹理的共识。鉴于滤波器阵列纹理提取方法更符合人脑的认识识别系统和人眼的视觉感官系统的优势,构建了一组滤波器阵列提取图像的纹理特征,并选择HSI色彩空间提取图像的色彩特征,将纹理特征向量和色彩特征向量组合形成本文的色彩-纹理特征向量用于彩色纹理图像的特征提取。在提出本文色彩-纹理特征向量的同时,采用texton直方图描述对应像素点邻域的纹理特征,应用直方图之间的距离构造像素分类的等价关系,同时提出了一种基于像素点对应texton频道中像素距离中值的直方图邻域的计算方法,用于计算不同尺度纹理中给定像素点邻域的加窗texton直方图。图论法具有广泛的数学基础,且可以捕获全局特征,文中采用图论法中规范割框架实现纹理图像的分割。针对规范割框架运行速度慢并且不能处理高分辨率图像的缺点,分析在规范割中图的关联范围对于分割结果的影响以及总体运行时间的决定因素,设计了一种快速规范割将大尺寸图压缩为线性时间复杂度的多尺度图。在多尺度划分准则中指定贯尺度约束矩阵进行多尺度分割,同时兼顾了精细层细节清晰和粗糙层整体结构明确的优点。实验结果表明多尺度规范割算法能够大幅度缩短原始规范割运行时间,同时能够处理高分辨率图像,为解决规范割框架的实用性开辟了新思路。针对色彩.纹理特征的具体形式将多尺度规范割扩展到小波域多尺度规范割。该算法一方面使用色彩-纹理特征描述彩色纹理,并用texton直方图对特征向量进行统计,体现了纹理的区域特性；另一方面,使用规范割框架对统计信息进行分割,保证了分割能够捕获图像的全局特征,同时使用小波域多尺度图结构保证了算法快速的计算速度。实验结果表明基于色彩-纹理特征的小波域规范割在保证快速的运算时间和有效的大尺寸图像处理能力的同时,能够获得准确和稳定的纹理图像分割结果。最后将本文提出的基于色彩-纹理特征的小波域多尺度规范割应用于遥感图像分割。更多还原

【Abstract】 The image segmentation is an important branch in the image processing field, its result is very importation to the image analysing and understanding. The texture is the universal characteristic in the image and it is difficulty to describe. The texture in the image reflects a certain variety in the object surface color and gray scale according to itself attribute. So the texture becomes a kind of obvious feature to describe object surface with different object on the different scale, and it occupies an important position in the image segmentation. The texture segmentation also becomes an important direction being highly valued with a lot of domestic and foreign researchers, and it has got the extensive application in the actual.The paper is studys the algorithm of color-texture segmentation, and generalizes some texture segmentation algorithm used widely at present, then research how to effectively extract texture feature of image and apply it to texture segmentation, try to find relizble, stable and pratical texture segmentation algorithm.The color-texture segmentation is divided into the color-texture feature extraction and consistency segmentation based on feature vector. The paper is to research around the above two points.The paper at first discusses the meaning of texture research, and summarize consensus by analyzing and comparing the typical texture definitions mothod. Filter bank texture feature extraction mothod has an advantage that more suitable with human brain recognition system and eye vision systems, so the paper presents a set of filter to extract texture feature of image. And the HSI color space is chosen to extract color feature of image, the combination of texture and color feature vector is used to format color-texture feature vector, which is applied to color-texture image feature exaction.At the same time, the texture feature of corresponding image pixel neighbors in this paper is described by texton histograms, and the pairwise texture similarity is constructed by comparing windowed texton histograms. The neighbors selection of texton histograms based on the median of corresponding pixel distance in the textons channel is present, and it is used to compute windowed texton histograms of given pixel neighbors with different scale texture.According to the widely mathematical foundation and the capability of capturing global features, the Normalized Cut framework is used to segment color-texture image. For the problem of slowly running speed and being unable to deal with high-resolution image, the effect of graph connection radius on segmentation results and the deciding factor of running time are analyzed. A rapid Normalized Cut is designed to compress large image graph into multiple scales graph with linear-time complexity. In the multiscale partition criterion, the cross-scale constraint matrix is defined for multiscale segmentation, considering the benefits of detailed boundary or fine level segmentation and clearing region of coarse level segmentation. The results show, multiscale Normalized Cut markedly reduces the running time of original Normalized Cut and segments high-resolution image, which provide the new idea for solving the practicality of Normalized Cut framework.Aiming at the specific form of color-texture feature, the multiscale Normalized Cut is extended to the wavelet domain. In the algorithm, the color-texture feature is used to describe color-texture and texton histogram is used to statistic feature vector which can embody the regional property of texture. On the other hand, Normalized Cut criterion is used to segment statistic information which assures the segmentation can capture global feature of image. And the multiscale graph in the wavelet domain can assure rapid running speed. Experimental results show that the multiscale Normalized Cut in the wavelet domain can assure the rapid running time and high-resolution, at the same time can gain accurate segmentation result. At the last, the algorithm is applied to remote-image segmentation.更多还原