Research On Object Representation And Recognition Techniques Based On Skeleton

Image resources that can be obtained and utilized are rapidly increasing with the development of modern techniques. It is urgent demand to detect and recognize targets from images by using the theories and techniques of the image analysis and understanding in both industry and military fields.Object representation and recognition techniques are kernel issues in the image analysis and understanding, in which the appropriate object representation is the groundwork and different representation forms will result in different recognition strategy. At present, a lot of representation approaches have been presented, but they have some shortcoming, respectively. The representation and recognition approaches based on skeleton are recently paid more attention since skeleton has following characteristics: hierarchical, multi-scale, topology of uniformity and adaptability of variety. This thesis aims to develop basic theories and techniques of representing and recognizing two-dimension and three-dimension objects based on skeleton representation of objects.The thesis focuses on theories and techniques of modeling and matching, which include the extraction of object skeletons in images, extraction of skeletons' structure elements, skeletons' graph representing and their matching. A multi-scale skeletonization algorithm of gray images with the non-ridge points lowering operation is developed by means of analyzing the skeletonization algorithms. On the basis of analyzing the skeleton basic structures, two methods are developed: one is for the skeleton elements acquired by hierarchical, iterative decomposition, and the other is for forming the elements into attributed graph or attributed tree. In order to match and recognize, the similarity measure approach is offered for a pair of skeletons' attributed graphs or a pair of attributed trees, in which the technique based on weighted optimal bipartite graph matching and maximal weighted cliques of the attributed trees' association graph are used, respectively. Also, the basal representation and recognition framework of 2-D objects are built in virtue of the flexible representation of skeletons. A large amount of experiments demonstrate the theory approaches adaptability.A novel approach of modeling 3-D objects built on the foundation of 2-D objectrepresentation is advanced. Firstly, a relative 2-D aspect set of an object is formed based on view-independent 3-D geometry representation, which reflects the different appearances of the object in various views. Then, the object's aspect model is established after the similarity of aspects is defined by the skeletons' attributed tree. Finally, the 3-D objects are modeled as a set of view regions, which are obtained with hierarchically clustered mode by merging similar and adjoining aspects. Experiments show validity of the approach presented, and possibility in some extent to revolve the ill-posed problem of reconstructing the object's 3-D structure from the 2-D images.