【作者】 林亚忠；

【导师】 陈武凡； 曹琳；

【作者基本信息】 第一军医大学 ， 生物医学工程， 2001， 硕士

【摘要】 图像分割是实现从一般图像处理到图像分析的关键步骤。它被广泛地运用于图像识别、图像配准、图像编码等研究领域之中。 近些年来有关图像分割方法的研究不断推陈出新，但均存在如下一个问题：由于对图像数据先验知识的分析不够深入，导致对图像空间分布信息的利用不足，使得难以实现自动化分割或者自动化分割的结果不够理想。为此，本文分别就图像感兴趣区与纹理特征的分割进行专门的研究并提出如下的改进： 1、在传统的MR心脏图像分割过程中，需要在每次分割时多次选择阈值才能得到比较好的分割效果。本文通过训练图像的特征阈值，提出一种有效提取和利用先验知识的方法，很好地实现了心脏分割的自动化；2、在解决多纹理图像的无监督分割中，本文在传统的最大似然标记和最大后验估计的基础上，利用Gibbs随机场作为先验知识模型，实现对多纹理图像的分割，达到了与传统的已知参数模型分割相近的分割效果。 本文通过大量的实验验证了方法的有效性。 综上所述，本文在对图像先验知识和空间分布信息的建立和应用方面做了较深入和系统地研究，提出了有效的改进算法。更多还原

【Abstract】 Image Segmentation is the key step to realize the research from general image processing into image analysis. It is widely used in many research fields like image recognition, image registration, image coding and so on. There comes up a lot of research in image segmentation methods in the recent years, but the following disadvantage was pointed out: Because of the lack of further analysis of prior knowledge of image data, which results in the insufficient usage of image spatial distribution information, the automatic segmentation becomes very difficult or the results of the automatic segment- ation are not ideal enough. So in this paper, some efficient improvements and special research are made in the segmentation of image regions of interest (ROJ) and texture feature. Firstly, during the process of the traditional segmentation method in cardiac MR image, in order to obtain the better effect, there must be threshold selections many times. In this paper, by training the feature threshold of the images, an efficient method for extracting and using prior knowledge is presented, which implements the cardiac segmentation automatically well. Secondly, during carrying out the unsupervised multi-texture image segment- ation, based on the estimation of traditional ML and MAP methods, we used the Gibbs Random Fields (GRF) as prior knowledge model to segment the images. The effect is almost equal to that of traditional method with previously known parameters model. Our methods proved efficient through many of the experiments. Briefly, in this paper, the further and systematic research was done at the aspects of how to build and use the prior knowledge of image and its spatial distribution information and the efficient improvement algorithm was presented.更多还原