Research On Mutiscale Representation And Geometry Reconstruction Of Shape Skeleton

As a special expression of the shape, skeleton acts well in compressing shape data and catching features. Thus, the shape of the skeleton extraction has become a hot research problem in computer vision field. However, the presence of skeleton extraction is sensitive to noise which leads redundant branches.The existing skeleton extraction methods are noise-sensitive, so that there is a problem of redundant skeleton branches. In the thesis, we concentrate on solving the problem of skeleton pruning better. The following two are the main points.1. A skeleton pruning method based on Bayesian model is proposed. Firstly, we extract the original skeleton from input shape. Basing on feature clustering and structure division of the original skeleton, the bayesian framework is introduced to establish the likelihood model. Secondly, we take the simplicity and accuracy into consideration to determine the pruning sequence of candidate branches.In order to get the ideal expression of skeleton structure, we don't stop iterating method until the post posterior probability gets maximal.2. A multi-scale representation algorithm based on mass distribution of skeleton is proposed. The algorithm extracts the muti-scale skeleton with the purpose which simulates the multi-granularity visual properties of the observer. Firstly, we need describe skeleton as a tree data structure. Secondly, we take the sum of sub node numbers as the quality value of the skeleton point. Basing on this, it is simple to find the quality value of each node is the value of prior node plus one. Through the linear superposition method, the quality distribution of entire skeleton is obtained. The quality distribution is introduced to measure the importance of each skeleton point. Finally, we adjust the root node by computing the difference of the quality between each nodes.Basing on balancing the quality distribution, the muti-scale skeleton presentation of shape could be obtained with threshold filtering.To verify the effectiveness of algorithm, we perform a series of experiments in skeleton pruning, shape reconstructions and computing the quality of the distribution.The experiments show that the proposed algorithm is robust to the shape noise and acts better in simplifying the skeleton structure and representing shape accurately.