贝叶斯信息传递在计算机视觉中的应用

CASIA OpenIR > 毕业生 > 硕士学位论文

	贝叶斯信息传递在计算机视觉中的应用
其他题名	Bayesian Belief Propagation: Applications in Computer Vision
	王东升
	2003-05-01
学位类型	工学硕士
中文摘要	贝叶斯信息传递(Bayesian Belief Propagation)是用来求解概率图模型概率推断问题的重要算法。在无环的网络中,它是可以得到严格证明的。在有环的网络中,很多实验表明信息传递算法也能得到非常好的结果.论文介绍了信息传递算法的相关理论,并研究了它在几个计算机视觉问题中的应用.论文的主要工作有: ①第一章的主要内容是信息传递算法相关的基本理论:图论、概率图模型、马尔可夫网络中的信息传递算法.文章还概述了关于信息传递算法的一些最新研究进展,以及信息传递算法在计算机视觉中现有的应用. ②第二章研究了基于背景减除的运动检测问题.基于背景减除的运动检测的难点不仅在于减除本身,更在于背景模型的维持.论文首先综述并比较了现有的背景模型,然后描述了一种新的背景模型.这一模型包括两个模块:点过程模块和Markov Network(MN)模块,这两个模块交替在象素级和图象级处理输入的图象序列.点过程模块的输出作为观察输入MN模块(背景减除),MN模块的输出又作为输入来更新所有的点过程模型(背景维持).MN的最大后验估计问题可以被信息传递算法很好的求解.论文还提出了一种新的象素过程模型:自适应的HMM模型,并讨论了他的学习算法.(部分工作发表在ICVI2002) ③第四章中,为了提高虹膜识别系统的识别效果,论文针对虹膜识别提出了图象序列中计算每一时刻的场景深度.为了同时利用图象序列空间和时间上的相关性,论文提出了一个新的统计框架.在这个框架中,同时包括了时间上和空间上的信息传递.最后深度恢复问题被描述为一个MN的最大后验估计问题.信息传递可以自然的解决这个问题.试验中,真实图象的计算结果证明了算法的有效性.最后,论文介绍了一个我们实现的动态场景的实时采集绘制系统.(部分工作发表在ICIP2002) ④第四章中,为了提高虹膜识别系统的识别效果,论文针对虹膜识别提出了一种新的基于学习的图像增强方法.该算法是改进了[1]的工作.该算法通过大量的虹膜图像去学习虹膜特征图像各频带之间的先验依赖关系,这种关系被一个马尔可夫网络所描述.然后信息传递算法被用来估计给定低分辨率图像的最可能的高频细节.为了避免在学习过程中引入会严重破坏识别算法的虚假高频信息,论文引入了Gabor滤波处理使得新算法只学习对识别最重要的部分频段信息.在NLPR虹膜图像数据库上,针对这种新的算法,我们做了大量的实验,取得了非常好的效果.(部分工作投递到BMVC2002)
英文摘要	Bayesian Belief Propagation, or Belief Propagation(BP), Probability Propagation, is one of the important algorithms for probabilistic inference of probabilistic graphical model. Though it can only be proved in unloopy network, some empirical results tell us it can get excellent results in loopy network. This thesis first introduces some related basic concepts and theories, and then describes some of my work about its apphcations in computer vision. Contents of the thesis can be summarized as follows" ① Chapter 1 first introduces basic concepts and theories about graplncal model and BP algorithm, and then summarizes some recent developments about BP algorithm. At last, it gives a brief review of recent apphcations of BP in computer vision. ② Chapter 2 studies the problems about "Background Maintenance and Sub traction" "Background Maintenance and Subtraction" is an important modular m many comput, er vision applications. We first review and compare all the previous background models, and then we propose a novel one This model has two components: Pixel Process modular and Markov Net- work(MN) modular The background maintenance and subtraction is performed by processing input image sequence alternatively using these two modulars at each pixel and whole uuage respectively. The output of Pixel Process modular is input of MN modular as observation(Background Subtraction) and the output of MN is also the input of Pixel Process modular (Background Maintenance). The MAP problem of MN modular can be natually solved by BP (Published in ICVI2002) ③ Chapter 3 is my work about dynamic depth recovery using BP ie, to recover the shape of dynamic scenes from Bayesian Belief Propagation, or Belief Propagation(BP), Probability Propagation, is one of the important algorithms for probabilistic inference of probabilistic graphical model. Though it can only be proved in unloopy network, some empirical results tell us it can get excellent results in loopy network. This thesis first introduces some related basic concepts and theories, and then describes some of my work about its apphcations in computer vision. Contents of the thesis can be summarized as follows" ① Chapter 1 first introduces basic concepts and theories about graplncal model and BP algorithm, and then summarizes some recent developments about BP algorithm. At last, it gives a brief review of recent apphcations of BP in computer vision. ② Chapter 2 studies the problems about "Background Maintenance and Sub traction" "Background Maintenance and Subtraction" is an important modular m many computer vision applications. We first review and compare all the previous background models, and then we propose a novel one This model has two components: Pixel Process modular and Markov Net- work(MN) modular The background maintenance and subtraction is performed by processing input image s
关键词	信息传递马尔可夫网络贝叶斯推断背景减除与维持动态场景深度恢复图象增强 Belief Propagation Markov Network Bayesian Infeence Background Maintanance And Subtraction Dynamic Scen Depth Recovery Super-res
语种	中文
文献类型	学位论文
条目标识符	http://ir.ia.ac.cn/handle/173211/6814
专题	毕业生_硕士学位论文
推荐引用方式 GB/T 7714	王东升. 贝叶斯信息传递在计算机视觉中的应用[D]. 中国科学院自动化研究所. 中国科学院研究生院,2003.