Research On Co-Segmentation For Digital Geometry Model

3D digital geometry models have nowadays become an emerging type of digital media after sound, images, and video, which have been intensively used in many fields such as industrial manufacturing, digital entertainment, biology, medicine and digital protection. Techniques for efficient and robust processing of digital geometry have also become an active area of research in computer graphics, and are developing towards the high-level and semantic direction. As a key ingredient in many digital geometry processing tasks, segmentation for 3D models has attracted many researchers’attention. However, traditional segmentation methods mainly pay attention to segmenting an individual model. They are difficult to consistently segment a set of models. Thus, they are unable to satisfy the semantic processing requirement. Recently, researchers have observed that segmenting a set of 3D models as a whole into consistent parts can infer more knowledge than from an individual model, which leads to a new research problem of co-segmentation and opens up a completely new train of research thought for 3D model segmentation.In this thesis, we mainly studied several key technologies of 3D model co-segmentation, such as the influence of per-object pre-segmentation, the interactive intention expression and the updating strategy.Regarding these issues, this paper engages in the following work:1) We proposed a co-segmentation method by combining Fuzzy C-Means (FCM) and Random Walks. As an efficient soft clustering algorithm, FCM is firstly used to cluster directly all the facets in the set in terms of their shape descriptors to improve the efficiency of clustering on face level. Random Walks model is then incorporated into the iterations of FCM clustering to combine the segmentation properties to the clustering measurement. It avoids the results of clustering on face level becoming discrete and unnatural, and makes geometric structure of the segmented parts reasonable all through the clustering. Finally, the combination of Fuzzy C-Means and Random Walks eliminates the influence of per-object segmentation in discovery-based methods and makes the segmentation more stable.2) We proposed a progressive segmentation method using online learning. In this method, region-based interaction is firstly used to simplify users’efforts. A segmentation model is then trained/updated in an online way by using Online Multi-Class LPBoost (OMCLP) and Online Random forests (ORFs). And graph cuts optimization is further performed to segment the 3D model. This method achieves the progressively interacting and accumulating in segmentation, and incrementally updating when new models added. So, it solves the problems of interactive intention expression and the updating strategy in guidance-based co-segmentation methods.3) We proposed an incremental segmentation method by blended learning. Co-segmentation method by combining Fuzzy C-Means and Random Walks is performed to analyze the shape set and obtain the predefined segmentation. It reduces the interactions required during the initial stage of the progressive segmentation. Then, progressive segmentation method using online learning is performed to achieve the progressively interacting and accumulating during the segmentation. It makes the discovery-based method interactive. Finally, weighted online learning is used to blend these two methods, which achieves the incremental accumulating segmentation on the basis of the predefined segmentation. And it achieves the unified incremental updating scheme for these two methods for composite-based co-segmentation methods. It has formed a co-segmentation framework, which satisfies four characteristics:predefined, interactive, accumulative and updatable.