Research On Key Technologies Of Hierarchical Scene Understanding

Scene understanding can be applied to a wide field of military and civil aspects,and give benefits to the future of human in many ways.Hierarchical scene understanding aims to coherently interpreting a scene from parts,to objects,and to scene.To this end,it is necessary to firstly compose multiple layers of entities from part to scene in the bottom-up way,and then build a probabilistic graphical model(PGM),which at least includes part level,object level and scene level.In the model,according to the prior knowledge about multiple layers of entities in scenes the bottom-up composed results are verified by top-down,and the scene is understood through inference.Recently,as knowledge base,visual dictionary is widely applied to recognizing and understanding tasks.Additionally,as general stochastic grammar for representing visual knowledge,And-Or-Graph is successfully applied to object recognization,and scene understanding.Both of two methods for visual knowledge representation are based upon automatic clustering.However,it is yet an unresolved problem to find the number of clusters in multi-dimensional dataset.To this end,we try to solve this problem through combining local search with global analysis.In the bottom-up composition,it is critical to map low-level features into high-level semantics.Generally,the gap between two levels can be bridged by object detection.When detecting objects affine invariance is generally required,since objects in scenes may be across lots of scales and poses.That how to invariantly represent object shape with feature for object detection is the second aspect to be explored,and in the view of affine invariance model-based representation methods are introdued.In PGM,message passing scheme can be used to get a coherent interpretation of scene.To this end,Junction-tree as the general method for exact inference and NBP as the widely-used approximate inference technique are considered.In the process of combining above two methods,the main problem is that how to downsize the cliques computed during graph triangulation for improving efficiency.Although considerable experiments show that two existing methods for removing the redundant fill edges in the triangulated graph are effective for downsizing the maximum clique in Junction-tree,their popularities are determined by efficiency.Through considering the removal ordering strategies,we shall improve their time-complexities.The fourth aspect to be discussed in this paper is the method for building the hierarchical scene understanding model.To this end,we propose a generative scene understanding framework,including hierarchical representation module,multi-class object detection module and hierarchical probabilistic graph module.The representation module is used to represent knowledge about multi-layers of entities in scenes,and the object detection module can detect multiple classes of objects with a unified scheme.In the probabilistic graph module,the PGM for understanding scene is constructed for certain scene with the aid of the representation module and the object detection module,which includes three types of nodes,such as parts,objects and scene.All of four aspects are theoretically and experimentally verified.For the method finding the number of clusters,its convergence is theoretically analyzed,and its effectiveness is validated with multi-kinds of datasets.With regard to shape object representation,the affine invariance of Zernike moments is measured theorectically and experimentally,and the performance of the proposed shape descriotor is experimentally tested.The proposed ordering strategies for removing redundant fill edges are proved to be effective for improving the time-complexities of the two existing methods.Finally,the scene understanding framework is tested,and our experimental results show that object detection and part representation are two critical factors influencing upon its performance.