QQ客服 官方微博 反馈留言

题名	基于几何主动轮廓模型和多特征集协作的图像分割研究
作者	李正龙
学位类别	工学博士
答辩日期	2008-06-02
授予单位	中国科学院研究生院
授予地点	中国科学院自动化研究所
导师	卢汉清
关键词	图像分割 几何主动轮廓模型 多特征集协作 狄利克雷过程 变分贝叶斯 image segmentation geometric active contour multiview learning dirichlet process variational bayes
其他题名	image segmentation based on geometric active contour and multiview learning
学位专业	模式识别与智能系统
中文摘要	本文主要研究了几何主动轮廓模型，和多特征集协作学习在图像分割中的应用。工作主要从以下几个方面进行：从定制矢量流的角度，讨论了集成多线索的信息到矢量场中，以用于驱动GAC模型实现较复杂场景的分割；设计了图像域内的Lennard-Jones力场，提出了基于Lennard-Jones力场的几何主动轮廓模型算法，以实现多用途的图像分割；将机器学习中的多特征集协作的思想引入图像分割，以解决分割问题的鲁棒性和可靠性，并将原始的co-EM算法扩展到co-DP；考虑了一种新颖的基于变分贝叶斯的多特征集协作图像分割方法，解决了常规EM方法无法解决的模型选择和传统类标传递过程中出现的类标变化导致的不相容问题。本文的贡献可以归纳为以下几个方面： 1. 提出了一种基于定制矢量流的GAC模型。通过引入定制矢量流的概念，将集成多线索的矢量场引入GAC模型，使之具有分割复杂场景的能力。我们首先选择集成了多线索的EdgeFlow作为候选矢量场，通过对候选矢量场的改造，使之满足我们所提出的定制矢量流的准则，然后将其集成到GAC模型中。 2. 提出了一种Lennard-Jones矢量场驱动的GAC模型。受到分子动力学中分子之间相互作用的启发，我们提出了一种新的力场驱动的GAC模型。相比较常规的GAC模型，我们的力场可以直接从原始图像数据上计算得到，不需要提取边缘图的前处理过程，并且我们的方法可以方便的集成各种有用的信息。大量的实验证明了我们方法的有效性和可靠性。 3． 将机器学习中的多特征集协作的思想引入到图像分割领域：将co-EM引入到图像分割问题中，并且扩展传统的co-EM到co-DP框架，解决了co-EM中容易出现的过拟合和欠拟合问题，实现了一种不依赖于初始化，可自动选择成份个数的分割方法。 4． 提出了一种基于概率推荐的变分多特征集协作图像分割框架。我们将变分贝叶斯的思想引入到图像分割框架中来，用它解决模型选择中容易遇到的过拟合，欠拟合问题。一种基于概率加权混合的多特征集协作的方法也被提出，集成到变分贝叶斯框架中，实现了快速，有效的图像分割。
英文摘要	In this dissertation, Geometric Active Contour (GAC) and Multiview Learning for image segmentation are studied. The research focuses on following several aspects. In the works of GAC, from the viewpoint of vector flow customization, we integrated multi-cue into vector field to drive GAC to segment complex image; we designed Lennard-Jones (L-J) force field in image domain, and proposed L-J force driven GAC for multi-purposed image segmentation. As to multiview learning in image segmentation, we introduced multiview learning to image segmentation to improve robustness of reliability of image segmentation, and extended the conventional co-EM algorithm to co-DP. Moreover, we proposed a variational Bayes segmentation approach with probability suggestion multiview strategy to handle the problem of modeling selection. The contributions of this dissertation are as follows. 1. We proposed a customized vector field based GAC framework. By introducing customization of vector flow, we integrated the multi-cue into GAC model to make the GAC able to segment complex image. We choose EdgeFlow as candidate vector flow, and modify it according to the vector flow customization criteria, then integrate the modified vector flow into GAC. 2. We proposed a Lennard-Jones force field driven GAC model. Inspired by the theory of molecular interaction, we proposed the Lennard-Jones force field for GAC. It is different from the gradient based GAC models in that the proposed model does not rely on any pre-computed edge map and is directly computed from image data. Moreover the proposed method can integrate various useful information including grayscale, color and texture etc. Many experiments validate the effectiveness and reliability of the proposed method. 3. We introduced the multiview learning in machine learning to image segmentation: the co-EM strategy was introduced into image segmentation, and we extended co-EM to co-DP framework to handle the problem of over-fitting and sub-fitting, and implemented a segmentation approach that does not rely on the initialization and can automatically determine the component number. 4. We proposed a probability suggestion multiview based variational Bayesian approach to image segmentation. The variational Bayes idea was introduced to image segmentation to deal with the problems of over-fitting and sub-fitting. And the probability suggestion multiview strategy is proposed to be integrated into the VB framework to result in a quick and efficient image segmentation framework.
语种	中文
其他标识符	200318014603013
内容类型	学位论文
源URL	[http://ir.ia.ac.cn/handle/173211/6114]
专题	毕业生_博士学位论文
推荐引用方式 GB/T 7714	李正龙. 基于几何主动轮廓模型和多特征集协作的图像分割研究[D]. 中国科学院自动化研究所. 中国科学院研究生院. 2008.

个性服务

查看访问统计

相关权益政策

暂无数据

收藏/分享

除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。