【作者】 杨金龙；

【导师】 厉树忠；

【作者基本信息】 西北师范大学 ， 电路与系统， 2009， 硕士

【摘要】 图像分割是指把图像分解成各具特性的区域并提取出感兴趣目标的技术和过程,它是计算机视觉领域的一个重要而且基本的问题,尤其在图像处理、分析和理解等领域中是一项关键技术,分割结果的好坏将直接影响到视觉系统的性能。所以对图像分割的研究一直都是图像技术研究中的热点和焦点之一,故图像分割的算法层出不穷,而且已经在诸如计算机视觉、模式识别和医学图像处理等实际中得到了广泛的应用。到目前为止,虽然已经出现了大量的图像分割算法,但是仍然没有出现一种通用的方法,大多算法都是针对一些特定的图像而提出的。在图像分割算法的研究中,往往需要重点考虑的问题是对现实图像中含有的大量干扰噪声的抑制及算法计算速度的提高,这也是图像分割中的一个难点问题。本文也正是基于此考虑,对传统一些经典算法进行分析研究,并提出了一些适当的改进。首先,本文对传统的阈值分割算法进行了研究,提出了一种基于二维直方图的Otsu阈值分割快速算法,以及用模糊熵作后处理的图像分割算法,充分考虑了图像邻域的空间信息,并把原来的二维阈值转换为一维阈值,进一步引入快速算法,使得运算量大大降低,从而提高了图像分割的效率,最后通过实验验证了该方法的可行性及高效性,无论是在抑制噪声方面,还是在运算速度上,都明显优于传统的二维阈值分割方法;其次,本文对边缘检测算法做了深入研究。由于传统的微分算子边缘检测方法在抗噪和边缘精确定位上往往不能兼顾,于是本文提出了一种基于小波变换的多尺度边缘检测方法,该方法在选取适当变换尺度的情况下,可以抑制大量噪声,尤其本文提出把小波变换多尺度的思想和经典算子适当的结合从而可以检测出一个很好的边缘;该方法尤其是对含有大量噪声的图像分割效果更好。最后,本文对纹理图像也做了适当的研究,在理解纹理特征的基础上,提出了一种基于纹理能量特征和FCM的纹理图像分割方法,并通过实验加以实现。讨论了该方法的分割结果,得出了一些对进一步研究纹理图像分割有用的结论,为进一步深入研究纹理图像分割奠定了基础。更多还原

【Abstract】 Image Segmentation is the technique and the process to segment an image into different sub-images with different characters and to extract the interested objects from the image. It is an important and basic procedure in the field of computer vision, especially a key technique in such field as image processing, image analyzing and image understanding. The quality of image segmentation directly affects the performance of vision system. Therefore, the research into image segmentation has always been one of the central issues for researchers in the field of image process, which results in an amount of algorithms for image segmentation. And these algorithms have been put into practice broadly in areas like computer vision, pattern recognition, medical image processing, etc.So far, although many algorithms for image segmentation have emerged, there is no general one. Most of the algorithms are intended for some specific images. In the research of algorithms for image segmentation, the problem that often needs most consideration is the suppression against the numerous interference noises in real images and the improvement of calculation speed, which is also a tough problem in image segmentation. Considering this problem, we did some research into some traditional and classical algorithms, and provided some possible improvements. Firstly, this paper researches into traditional threshold segmentation algorithm, put forward a fast Otsu algorithm for threshold segmentation based on two-dimensional histogram, and an algorithm for post-processing by fuzzy entropy. Using these methods, we fully considered the spatial information of image neighborhood, transferred the original 2D threshold into 1D threshold, and furthermore introduced fast algorithm, reducing computation greatly, and thus improving the efficiency of image segmentation. At the same time, we proved the feasibility and high effectiveness of these methods through some experiments. These methods are prior to the traditional 2D Otsu method either in terms of noise suppression or in terms of operation speed. Secondly, in this paper, we did further research into edge detection algorithm. Because traditional edge detection method for differential operators cannot give consideration to both anti-noise and edge pinpoint, this paper put forward a multi-scale edge detection method based on wavelet transform, which can suppress a large amount of noises. And this paper particularly came up with an appropriate combination of the thought of multi-scale wavelet transform and classical operator, with which the edge can be detected very well. This method is especially effective in segmentation for images with a large amount of noises. Lastly, this paper also did some research on texture image. Texture image segmentation algorithm based on texture energy features and FCM has been put forward on the basis of understanding about texture energy features, and has been realized through experiments. The result of this segmentation method has been discussed, and some conclusions useful to further segmentation of texture images have been reached, which laid the foundation for more detailed study on texture image segmentation.更多还原