Research And Applications Of Tree-structured Graphical Models

With the rapid development of information technology, the quantity ofinformation is growing exponentially causing the problem of information explosion.For last three decades, the amount of information produced by human has surpassedthe sum of the past5000years. Facing with the huge data and information,information processing and analysis have become the bottleneck of informationtechnology. How to combine the qualitative analysis and the quantitative analysis tofind an efficient method and to fit a mathematical model of high-dimensioninformation data which can be used for decision-making, control and prediction is thefocus of the study content. Probabilistic graphical model is one of the most importantways to deal with uncertain problems and big data. It combines probability theorywith graph theory in order to explain the uncertainty and complex phenomena.Therefore, in this paper we focus graphical models. We study the tree-structuredgraphical model learning algorithms and problems among existing algorithms. At thesame time we explore the usage of tree-structured graphical model in objectrecognition. The main contents of this paper are as follows:(1) We briefly summarize the history and research status of graphical models atfirst. And on this basis, we focus on tree-structured graphical model with and withouthidden nodes, introducing the research foundation and analyzing existing problems oftree-structured graphical model.(2) A new algorithm is proposed for learning the latent tree-structured graphicalmodel based on fuzzy multi-features Recursive-Grouping. First, we transform originalobservation data to multi-features by fuzzy membership functions and constructmulti-dimensional fuzzy feature vectors. Then, we compute the distance betweeneach fuzzy feature vectors. Finally, based on the distance matrix, we construct thelatent tree graphical model by the Recursive-Grouping algorithm. In addition, wemake the simulation by actual data, validating the efficiency of our algorithm.(3) A new context model for object recognition which improved by spatialrelationships based on existing context models is proposed in this paper. The newmethod improves the context model by further considering the spatial relationships.First, we integrate global features, co-occurrence relationships, spatial relationships,and local detector outputs of images in a unified probabilistic framework. Then, weestablished a new object recognition algorithm taking full advantage of the efficient inference algorithms of tree models. It can improve the recognition result and give aconsistent scene interpretation of the image. Finally, we compare and validate thisalgorithm by actual image sets to show its effectiveness.