【作者】 汪荣贵；

【导师】 张佑生；

【作者基本信息】 合肥工业大学 ， 计算机应用技术， 2004， 博士

【摘要】 本文在对Bayes网络的国内外研究现状进行深入分析的基础上，对Bayes网络的知识表示、推理、解释等基本理论和方法进行了系统的研究，并从理论上对其进行了扩展，对Bayes网络及其扩展模型在航空影像中房屋等目标的识别，以及图像中文本目标的检测与定位等方面应用进行了探讨。全文主要内容如下： 1．从图模型角度分析了随机变量的条件独立性质，讨论了Bayes网络知识表示的基本原理及性质，讨论了消息传播与消元推理方法的基本原理，在此基础上提出了一种基于扩展邻接树的消元推理算法。该算法使用扩展邻接树的深度优先确定消元次序，有效解决了选择消元次序的难题。 2．在分析Bayes网络解释机制的已有研究成果的基础上，提出一种关于Bayes网络推理结论解释的新机制。文中引入必要性和充分性因子，解释证据对推理结论的作用程度；引入概率分布变化方向的概念，用于判定证据对推理结论的作用方向，检测证据之间可能存在的冲突现象；通过定性和定量分析网络结构特征生成证据对推理结论的作用路径。通过应用实例的讨论，说明了该解释机制具有合理性。 3．在研究房屋等人造结构体的投影性质的基础上，提出了一种基于Bayes网络的感知组织算法，用于检测航空摄影图像中屋顶等目标。该算法包括提取边缘及线段、生成平行四边形、生成假设、检验假设等步骤，用Bayes网络推理实现信息融合，由Bayes网络学习实现了感知的自适应性。实验结果验证了算法的有效性。 4．针对大规模Bayes网络知识表示和推理的计算复杂度问题，提出了一种新的面向对象的概率图模型——对象概率模型(OPM)。该模型将Bayes网络分解成若干称为类的模块，每个模块中设置两个界面节点用于传递概率信息。OPM利用层次结构中所蕴含的条件独立性，有效降低了模型的构造和知识表示的复杂性。文中通过推广Bayes网络消元推理算法实现了OPM的推理机制。适当调节推理算法中的控制参数，就可有效控制OPM推理的计算复杂度。将OPM用于解决图像中文本的自动检测与定位问题，实验结果表明检测效果好、速度快。 5．对Bayes网络与确信因子模型作了对比研究。讨论了确信因子模型的理论基础，分析了确信因子模型的局限性，论证了Noisy-OR模型(Bayes网络的一种简化模型)的概率推理公式与确信因子模型的推理公式的等价性，从知识的表示、推理、获取三个方面论述了Bayes网络相对于确信因子模型的优势。更多还原

【Abstract】 After an overview on research results about Bayesian networks (BNs) in our country and the developed countries, this thesis researches systematically the knowledge representation, inference methods and explanation mechanism of BNs and makes some extension on them. The BN applications in detection of objects, such as recognition of buildings in aerophotograph images, and detection and location of texts in images, are discussed in detail.The main contents and novel parts of the thesis are as follows:1. The conditional independence of random variables in graphical models is discussed. The basic principle and methods of knowledge representation and inference for BNs are discussed. An inference algorithm is presented using variable elimination based on the extended junction-tree, which determines the sequence of variable elimination by depth-first search of the extended junction-tree and solves the problem of selecting variables for elimination.2. Based on the discussion of existing BN explanation methods, we proposed a new explanation mechanism of BN inference. In this mechanism, the necessity factor and abundance factor are inducted for explaining to what extent the evidences affect. The concept of direction of probabilistic distribution changements is inducted for determining the direction of the evidence effect on inference conclusion as well as detecting the possible conflict phenomenon of evidences. A path of evidence effect on the inference conclusion is generated for explanation by analyzing the network structure quantitatively and qualitatively. An application case shows that the explanation method proposed in the thesis is quite reasonable and effective.3. On the basis of discussing projection properties of buildings, a BN based algorithm of perceptual organization is presented, which can be used to detect objects, such as buildings in an aerophotograph image. This algorithm includes four steps: edge extraction, parallelogram generation, hypothesis generation and hypothesis examination. The information fusion is realized by BN inference, and the self- adaptability of perception is realized by BN learning. The experiment results show the effectiveness of the algorithm.4. A new object-oriented probabilistic graphical model, named object probabilistic model (OPM) is put forward, which can reduce the complexity of knowledge representation and inference of a large-scale Bayesian network. OPM can decompose a BN into many models named classes, each of which has two interface nodes for propagating probabilistic information. OPM makes full use of the conditional independence implicated by the hierarchical structural and reduces the complexity of model construction and knowledge representation. The inference mechanism of OPM is realized by generalizing the variable elimination algorithm of BN. The computing complexity of inference process of OPM can be controlled by adjust the controlling parameters in inference algorithm. Some experiment results of applying OPM to detect and localize texts in an image show that the new model and relevant algorithm are of important theoretic and practical significances.5. Some comparisons are made between Bayesian networks and the of certainty factor models. After discussing the theory foundations of the certainty factor model and its drawbacks, the equivalence of the probabilistic inference equation of a Noisy-OR model, a simplified BN model, and the inference equation of a certainty factor model is argued. The advantages of BNs over the certainty factor models are discussed in knowledge representation and acquisition and their inference.更多还原