【作者】 何月；

【导师】 申铉京；

【作者基本信息】 吉林大学 ， 计算机应用技术， 2012， 硕士

【摘要】 图像分割的主要目的就是将感兴趣区域从背景中提取出来，以便进行进一步的分析处理。图像分割技术已广泛地应用于医学图像处理、模式识别和计算机视觉等领域。因为图片本身具有不确定性和模糊性，所以图像分割一直是图像处理中的经典问题。许多学者在图像分割方面做了大量的研究，提出很多有效实用的分割算法。其中，基于模糊聚类的图像分割方法是一类极其重要的算法，因其可以很好地描述图像的模糊性而得到众多学者的青睐。但经典的模糊聚类分割方法存在对噪音敏感、运算开销大以及对边缘数据处理较差等缺点，影响其应用。因此，本文研究了模糊聚类中颇为流行的模糊C均值算法，针对传统模糊C均值算法应用于图像分割时存在的一些缺点进行了改进，以规避其在图像处理中存在的不足，本论文主要取得了以下研究成果：(1)考虑到传统的模糊C均值算法受噪音影响大的弊端，本文提出一种包含图像空间信息和隶属度约束的模糊C均值图像分割算法。在传统的模糊C均值算法基础上，对能量函数添加约束项约束空间信息和隶属度值，可以减弱噪音点在选取聚类中心时的影响，有效地降低了椒盐噪音、高斯噪音以及混合噪声对图像分割结果的干扰。新添加的约束项对于噪音点可以适当地修正其隶属度的值并正确地归类，并且不含参数，避免人为因素对分割结果的影响。通过与FCM算法、FCM_S算法、MOP_FCM算法、鲁棒性FCM算法、RFCM算法以及FLICM算法的对比实验结果可以看到，本方法较适合处理受噪音污染的图像，得到较好的分割结果。(2)针对传统的模糊C均值算法对图片细节信息保留较少的缺点，提出一种基于塔形分层的模糊C均值图像分割算法，通过构造自适应隶属度阈值的塔形结构来避免人为设定阈值的不灵活性，引入塔形结构可以有效地减少运算时间。在传统模糊C均值算法的目标函数中添加约束项，添加塔形结构的顶层像素点隶属度对底层像素点隶属度的约束，同时约束了邻域空间信息，可以更恰当地处理边缘数据。最后进行了分割对比实验，通过与FCM算法、分层FCM算法、MS_FCM算法等六种算法对多组图像的分割结果可以看到，本文算法可以有效地保留图像细节信息，提高图像的分割质量，分割结果更加精确。(3)在彩色图像分割方面，提出一种基于S和V颜色分量的模糊C均值彩色图像分割算法。通过统计彩色图像在HSV颜色空间中的S和V颜色分量的分布情况，用基于直方图的快速模糊C均值算法对颜色分量进行聚类划分，根据得到的聚类中心及彩色图像的目标颜色数目分割图像得到最终的彩色图像分割结果。本文算法简单快速，不需要任何先验知识，可以有效提取彩色图像的目标区域。通过对多组不同颜色花朵的分割实验，以及与其他四种彩色图像分割算法的分割结果对比，证明本算法可以有效地分割目标区域颜色不同的彩色图像并提取彩色目标区域，分割结果较准确，运算速度较快，是简单快捷有效的彩色图像分割算法。更多还原

【Abstract】 The main purpose of image segmentation is to extract the interesting regions from thebackground for further analysis. Image segmentation techniques have been widely used inmedical image processing, pattern recognition, computer vision and other fields. For theimages are ambiguous and uncertain inherently, image segmentation is a classic problem inimage processing. Many scholars have done a lot researches on image segmentation andproposed plenty of segmentation algorithms. The algorithm of image segmentation basedon fuzzy clustering is very important and has been used widely. The fuzzy clusteringalgorithm can properly describe the ambiguity of images, but some drawbacks of the classicmethod affect its widely application.Therefore, some researches are proceeding on the quite popular fuzzy C-means(FCM)algorithms based on fuzzy clustering. Several improved schemes of FCM are presented toavoid the shortcomings in image segmentation. The following key findings have been madein this paper:(1) It is a big drawback that the traditional fuzzy C-means algorithm is susceptive tonoises, this paper presents a fuzzy C-means algorithm that contains the space informationand membership constraints. By adding new constrained item into the energy function ofthe traditional fuzzy C-means algorithm, the algorithm is robust to salt and pepper noise,Gaussian noise and mixed noise. The new algorithm can obtain more sensible clusteringcenters. The membership of the pixels affected by noise can be properly modified, thereforethe pixels could be correctly classified. The item newly added is full free of empiricallyadjusted parameter. Comparing with the FCM algorithm, FCM_S algorithm, MOP_FCMalgorithm, robust FCM algorithm, RFCM algorithm and FLICM algorithm, the algorithm issuitable to segment the images that affect by noises and could provide better segmentationresults.(2) The details of images can’t be reserved effectively in the traditional fuzzy C-manesalgorithm. This paper proposes pyramid based fuzzy C-means algorithm. The pyramidstructure is introduced to rapid the approach and the adaptive threshold selection ofmembership grade avoids the inflexibility of human threshold. The constraint is added intothe traditional fuzzy C-means objective function to constrain the membership of pixel in thebottom-level by taking its up-level membership into account, the constraint also containsthe spatial information. The algorithm could cluster the marginal data more appropriately.The comparisons of image segmentation experiments with several algorithms, such as FCMalgorithm, pyramid FCM algorithm, MS_FCM, etc. show that the new algorithm can effectively preserve the detail information of the image and improve the quality of imagesegmentation.(3) In order to segment color image effectively, fuzzy C-means algorithm based on Sand V color components is proposed. Firstly calculate the distribution of S, V colorcomponents of the color image in the HSV color space, then cluster the statistics using thehistogram-based fast fuzzy C-means clustering algorithm, finally segment the color imageaccording to the target color number and the cluster centers obtained. The algorithm issimple and fast, doesn’t require any priori knowledge and could effectively extract thetarget region in the color images. The segment results show that this method can segmentthe color images effectively, owns fast speed and better segment results. The algorithm issimple and quick effective.更多还原