A New Hierarchical Probability Generative Model And Its Application To Scene Analysis

With the rapid development of multimedia and information and technology, the image data increases rapidly. Because of vast amounts of data, the effective management in computer has become an urgent task. Image scene analysis system based on image content information can divide images into categories and recognize target objects automatically. It plays an important role in image retrieval, object recognition, computer vision and other fields. Traditional ways to learn the scenes of images is to use the supervised machine learning algorithms with Bag of Visual Words(BoV). These methods are computationally simple but not enough to present abundant appearance visual feature of image region and neglect spatial information among visual words.This paper applys probability generative model to scene learning and focus on the methods about the semantic scene modeling and the presentation of visual feature of region and spatial information among visual words. The main content of this dissertation is summarized as follows:1) According to human’s cognitive habits of a scene, this paper viewed a scene image as three layers:scene-class-layer, object-layer and feature-layer and acquired a unified framework including the three layers’ information. Then we modeled and infered model parameters by use of probabilistic constraint relation among the three layers.2) A super-pixel mixture (SPM) is used to present overall appearance of image region. That is to say, each region of an image is further segmented into the mixture of some smaller regions which is usually called super-pixels, from which color and texture features are extracted. Last, together with SIFT features which were extracted from image patches, each image region was represented by the joint distribution of three kinds of visual words. And so, the stability and richness of visual-feature representation of region were enhanced.3) In order to consider spatial information among different visual words, all visual words in the same region are generated from a same object in the process of generative model. So, our model is subject to some certain spatial coherence.As a self-contained generate model, our model does not depend on the particular classifier. It means that using the own model we can carry out scene classification as well as pixel-level object recognition and segmentation. On the testing experiments run on database UIUC-Sport, we compare our results to these from earlier methods including Li-HPGM、Spatial-LTM etc.al., and shows that our SP-HPGM model obtained higher accuracy of scene classification and object recognition.