【作者】 王彦春；

【导师】 梁德群；

【作者基本信息】 大连海事大学 ， 通信与信息系统， 2011， 博士

【摘要】 图像分割是一种重要的图像分析技术。在图像的研究和应用中,研究人员往往仅对图像中某些部分感兴趣,这些部分常称为目标或前景(其他部分称为背景),一般对应图像中特定的、具有独特性质的区域。图像分割是计算机视觉研究中最基本的处理步骤和关键技术,是目标识别、图像理解的基础,分割结果的好坏直接影响其后续的识别和理解,因此人们一直在不断地探索新的分割算法和分割理论。本文对基于过渡区的图像分割技术进行了比较深入的研究,重点研究了直接提取过渡区的非梯度方法。提出基于方向信息测度的多尺度图像过渡区提取与分割方法；基于邻域一致性测度图像过渡区提取与分割方法以及基于模糊熵区域非一致性测度图像过渡区的提取与分割方法。分割方法抗噪声性能的本质在于测度对灰度值的变化不敏感,三种方法都具有较好的图像过渡区提取与分割性能和很好的抗噪声性能。模糊熵的引入,使模糊熵区域非一致性方法的混合噪声滤除效果更好。当图像含有多类目标时,图像过渡区直方图表现为多峰分布,每个峰对应一个分割区域。本文使用基于势函数聚类的多阈值图像分割算法,检测过渡区直方图的相邻划分势函数组,其曲线的交点即为分割阈值,然后使用差异度与代价函数决定最优类别个数及阈值。算法简单,得到的阈值准确、稳定,运算速度较快、实时性强,较好地解决了基于过渡区的图像分割方法中的多阈值分割问题。由于均值滤波和中值滤波算法本身的局限性,当图像同时受到高斯噪声和脉冲噪声污染时,此时采用单一的均值滤波或中值滤波算法滤除噪声效果都不理想。因此,本文提出一种基于区域分类的白适应混合滤波方法,利用邻域一致性测度把图像分成不同区域,并对不同区域采取不同的滤波方法,这样既保持了均值滤波算法对高斯噪声有良好滤噪能力的特性,又兼顾了中值滤波算法对图像细节有良好保持特性的优点。纹理图像是以纹理特性为主导特性的图像。纹理图像在局部内呈现不规则性,而在整体上表现出某种统计规律性。纹理表现是一种区域特性,因此纹理必然要在图像的某个区域上才能反映或测量。仅利用像素的灰度级信息并不能将其中的不同区域分割开,本文提出了基于单演相位的纹理图像分割方法。根据不同纹理之间单演相位信息的不同分布,提取特征图像。这种方法对于不同纹理之间及纹理与物体之间有很好的特征描述,取得较好的分割效果。X射线检测图像是焊缝缺陷分析和质量评定的重要依据。传统的方法是利用阶跃边缘和屋脊边缘进行圆形和条形缺陷分割,本文提出基于过渡区多尺度工业X光图像焊缝缺陷分割算法。在不同的尺度下,可以检测不同缺陷大小,通过支持向量机方法将焊缝缺陷划分为圆形和条形缺陷两类,同时对缺陷裂纹、气孔、夹渣和未焊透进行识别。更多还原

【Abstract】 Image segmentation is an important technique for image analysis. In the study and application of images, researchers usually focus on some certain parts of images, which are often named target or foreground (the rest parts are called background), generally corresponding to the given particular region. As the essential processing approach and the key point on computer vision study, image segmentation becomes the base target recognition and image understanding, As the results of segmentation have direct effect on sequent recognition and understanding, many reseaerchers are working on new segmentation algorithms and theories.This paper made an intensive study on transition region based image segmentation technique, stressing on non-gradient transition region extraction method. Three methods are presented, one of which is multi-scale transitional region extraction and segmentation method based on orientation information measure, the other is transition region extraction and segmentation method based on image neighborhood unhomogeneity measure, still another is transition region extraction and segmentation based on image fuzzy entropy region unhomogeneity measure. All the three methods are significantly resistant to salt and pepper noices. The essence of the noise resistance lies in the calculation of the measurement which is insensitive to the changes of the grey scales. The three methods are good at the image transition extraction and segmentation, and at noise resistance. The fuzzy entropy region unhomogeneity method has a better effect on mixed noices as a result of the introduction of fuzzy entropy.When an image contains several types of targets, the transition region histogram presents multi-peak distributing, each corresponding to a possible segmentation region. A potential function clustering based multi-threshold segmentation algorithm is adopted to detect adjacent division potential function group of the transition region histogram, the intersection of the curve is segmentation threshold, and then using differentiation and cost function to determine the number of prime types and the threshold. The threshold resulted from the simple algorithm will be accurate and stable. The algorithm also shows faster operating speed and better real time control. The multi-threshold segmentation problem is easily solved in the image segmentation technique based on transition region extraction.When the image is stained by both gaussian noises and pulse noises, neither average filter algorithm nor median filter algorithm can denoise effectively. This paper presents a self-adaptive mixture filter algorithm based on region classification. Using neighborhood homogeneity measurement, the image is divided into different regions where diverse filter methods are employed to deal with the noises. In this way we guaranteed that average filter algorithm has an unusual power to gaussian noises while we also ensured that median filter algorithm is effective to maintain the details of the image.Texture image is a type of complicated image whose texture is the dominate feature. Texture images present local irregularity while as a whole they take on statistical regularity. As texture has a regional characteristic texture will be reflected or measured in a given region of the image. Since grey scale information of pixel alone cannot segment different regions, this paper proposed a monogenic phase based texture segmentation method. Feature image extraction is based on the distribution of monogenic phase information. This method can obtain better segmentation results which well described the features among different textures and between textures and objects.X-ray testing images are important basis of weld defect analysis and quality evaluation, the traditional method is to make circular and bar segmentation with step edges and roof edges.This paper proposed a weld defect segmentation algorithm based on transitional region multiscale industrial X-ray images.Under different scales, the size of the defect can be inspected. Using support vector machine the weld defects are divided into circular and bar defects meanwhile the cracks, gas cavity, slag and lack of penetration can be recognized.更多还原