Research Of Image Analysis Based On Graph Model

The core contents of image understanding are object recognition and scene Analysis. In computer vision field, one of the most difficult problems is how to build a common feedback framework which can effectively recognize objects and analysis kinds of scene image. This thesis is aim at recognizing what kinds of objects are appearing in the image, finding the relationship between objects and analyzing the scene. The main task contains two aspects, one is objects recognition in the image, the other is scene analysis of the image. Objects recognition as the basis of scene analysis is benefit for explain the detail of the scene. Conversely, scene analysis which providing the prior knowledge can guide objects recognition. Therefore these two tasks are strongly connecting with each other. This thesis is mainly about as follows:1) Analysis two kinds of graph model and their characters. Then introduction how to use direct graph and probability in generative model. Meanwhile, three kinds of Dirichlet process sampling used in graph model are described.2) Achieved the single object recognition based on Hierarchical Dirichlet Process and multi-objects recognition based on Transformed Dirichlet Process. The former considers images as "object-part-region", then computers the probability distribution of higher level with lower level elements, and decides the label according highest probability. The latter method mainly integrates transformed spatial location information further, and then each kinds of object category can appear times in image.3) Research scene analysis based on graph model with hierarchical and contextual knowledge. In this thesis, the scene-level contextual information is applied for producing the prior knowledge, while the object-level contextual information is adopted for adjusting some objects appearing or not. Afterward, the recognition feedback effect on previous analysis results.