Based On The Symmetry Of The Target Description And Identification Technology,

Object representation and recognition techniques are one of the key issues in pattern recognition and image understanding areas. The current object representation methods are usually based on the image attributes and the geometric attributes of the objects. The thesis is emphasized on the techniques conducted from the ubiquitous inherent symmetric properties of the man-made and natural objects.The basic symmetry and the generalized symmetry and their applications are the main research content of this thesis. It includes the following research works.First, the basic theories and detection techniques for the basic symmetries and the skewed symmetries are systematically summarized. In order to overcome the difficulties to detect the basic reflectional axes of the objects directly from a gray image, a new approach based on the distributed active agents is presented in this thesis. This method enables to detection the local symmetry axes, growing and grouping them by simulating the perception, communication and action behaviors of the agents.Next, the basic theories and the key techniques of the skeleton-based object representation and recognition are studied in this thesis, which include the skeleton extraction, skeleton decomposition, skeleton based object representation and matching techniques and so on.In the skeleton extraction, an novel approach for extracting the skeletons directly from the gray images is developed by simulating the procedure of water eroding the topography surface. The resultant skeletons are single-pixel width and connected. It also lies in the middle line of the object and keeps the topological properties of the original image unchanged.In the skeleton decomposition and representation, an efficient skeleton decomposition and vectorization method is presented. This method first decomposes the skeleton into its meaningful component (such as branch and loop) set, then vectorizes each components to simple structure elements (such as straight line segments and circles) and finally arrange them into an attributed relation graph. In order to construct the attributed relation graph, a new real time method is developed, which computes the K-NN of the arbitrary planar sets using the distance transformation technique. This idea can also be applied to the higher dimensional point sets.In the skeleton matching, an novel approach based on weighted optimal bipartite graph matching theory is also presented. This method measure the distance of the skeletons-based on the maximal common sub-graph of their corresponding attributed relation graphs.Lastly, a skeleton-based experiment system for image database retrieval is designedand implemented on the basis of the symmetry based object representation and recognition theories. The prospective result is shown.