Research On The Theory Of Bayesian Network And Its Application In Image Analysis

After the overview on research results about Bayesian networks in both developed countries and our country, this dissertation researches systematically on the construction and learning of Bayesian networks as well as the variant model. Text detection and face detection are the two applications of objection analysis by Bayesian networks. The main contents and novel parts of this thesis are as follows:1. The fundamental theory of Bayesian networks and frequently used inference algorithms are introduced. Both the accurate and approximate Inference algorithms such as Message passing algorithm, Clique tree propagation algorithm, Variable elimination algorithm and so on are summarized, and an experiment is used to verify the availability of some frequent used algorithms.2. Structure learning algorithm of Bayesian network by hybrid genetic algorithm is presented after the drawback of Genetic Algorithm-based learning algorithms is analyzed. Rudimentary equivalence theory for structure learning is presented after the equivalence structure of Bayesian network is studied. And algorithm to parameter learning is also induced. Finally the process of transferring from model-based graphical model to Bayesian network is introduced.3. The variant model of the Bayesian network is presented according to causal independence after some frequent used graphical models are analyzed. The new model has simplified the size of the probability table. And the availability of this new model is verified by analyzing the probability table of the Bayesian network of student scores.4. Scene text in images contains many uncertainty factors because of various reasons. The variant grayscale histogram for text detection and location is presented and applied to the feature extraction of text in images. Then the Bayesian network is constructed to fuse these features according to the analysis of the uncertain relationship between them, and experimental results have shown the availability of the solution.5. The difficulty of human face detection in images is caused by the many uncertainty factors. Shoulder lines detection-based Bayesian network is constructed to determine the location of face region, and multi-threshold algorithm is used to construct the variant Bayesian network for human eyes detection in the face region. Experimental results show that the solution is suitable for the detection of human face images in complicate background.