Image Scene Analysis Based On Hierarchical Generative Graph Model

Currently image scene analysis is the central issue in image understanding. The essential task is to interpret the acquired image scene and its contents accurately. Image understanding has the distinct layer property, the visual information in lower layer as the contextual knowledge information in higher one, visual data and contextual knowledge are two types of information through analysis scene images.Considering key issues about data-driven knowledge and knowledge-guiding data in image scene analyzing, this paper research on visual feature extraction and description in image analysis, by combine flat directed graph model, focuses on object recognize graph model; Meanwhile, build a novel scene analysis model based on object recognize graph model to represent corresponding restriction and feedback mechanisms, ultimately achieved tasks as object recognition and scene understanding.This paper includes the following contents:1. We describe research contents of image scene analysis, including recognize object in scene and scene understanding, analysis two kinds of graph model and their characters. Then we discuss three kinds of Dirichlet process probability distribution and their learning and inference method used in graphical model.2. We study the extraction and storage on visual information to solve the graphic models as carriers in image scene analysis.Then we discuss the theories for parameter estimation and probability inference with hierarchical dirichlet process, obtain the Gibbs iterative equations to analyze single object in scene image. We integrated transformed spatial location information in hierarchical dirichlet process to form transformed dirichlet process to solve the problems in multi-object categories and instances recognition.3. We analyze the basic method for scene analysis graph model which considers images as "scene-object-part" to construct the relationship between visual feature, object categories and scene categories. At the some time,we also validate the scene analysis results as prior knowledge have strong guidance and restrict to improve accuracy in object recognition and reveal feedback in scene analyzing.