Image Content Representation And Classification Based On Bayesian Networks

Nowadays, image classification and video Image analysis are active areas attended widely in the fields of science research and application. They play an important role for organization of image and video collections due to the explosion of visual information. It is still a challenging task because of the uncertainties and ambiguities inherent to visual data in application case. The use of multiple features and contextual information is potential trend to alleviate this problem significantly. This thesis focuses on modeling both the spatial context in images and the spatio-temporal context in videos using Bayesian networks, and provides a unified framework to incorporate multiple features for image representation, classification and video behavior analysis.In the aspect of description of local image content, the thesis mainly focuses the local semantic modeling methods with contextual information for image annotation and classification. (1) A contextual Bayesian network for semantic objects (CBN_SO) is presented, which can reduce those unreasonable annotation errors in the neighborhood caused by low-level based approaches for image annotation. (2) An automatic Learned Bayesian network is developed, which is capable of learning specific spatial relation sets as contextual nodes for each object. (3) A method to classify high resolution remote sensing (HRRS) images based on LBN is offered, which is capable of learning adaptively the specific context information for each ground cover type and resulting in intelligent and efficient classification.?In the aspect of semantic modeling of scenes, a novel approach based on contextual Bayesian network for scenes (CBN_S) is presented for natural scene modeling and classification. To deal with the lack or hardness of modeling spatial configuration for scenes in the literature, the proposed model introduces the key semantic regions and their adjacent regions as the contextual nodes and makes use of their spatial arrangements to represent spatial configuration for scenes simply and flexibly. Furthermore, it integrates both local materials of images and spatial configurations of scenes into a unified probabilistic framework, which has ability to distinguish some confused natural scenes images.In the aspect of video behavior analysis, the thesis focuses on the DBN-based behavior modeling methods for composite behaviors with multi-object in real scenes. Based on the hierarchical definition and scene-event description for behaviors, a hierarchical multi-observation dynamic Bayesian network (HMO-DBN) is proposed for modeling the human-object interacting behaviors The proposed model put the spatial relation between different objects as a hidden node of the network, and takes account of its dynamic transformation with time , which is the critical part effectively to use the spatio-temporal context for modeling interacting behaviors . The propose models are applied to natural scenes, remote sensing images and surveillance videos, and expectant results are obtained.