Hypergraph based visual categorization and segmentation

摘要: This dissertation explores original techniques for the construction of hypergraph models computer vision applications. A is a generalization pairwise simple graph, where an edge can connect any number vertices. The expressive power places special emphasis on relationship among three or more objects, which has made hypergraphs better choice in lot problems. sharp contrast with conventional graph representation visual patterns only connectivity between objects described. contribution this thesis fourfold: (i) For first time advantage neighborhood structure analyzed. We argue that summarized local grouping information contained causes ‘averaging’ effect beneficial to clustering problems, just as image smoothing may be segmentation task. (ii) discuss how build incidence structures and solve related unsupervised semi-supervised problems different scenarios: video object segmentation, categorization retrieval. compare our algorithms state-of-the-art methods effectiveness proposed demonstrated by extensive experimentation various datasets. (iii) application retrieval, we propose novel model — probabilistic exploit data manifold considering not information, but also similarities vertices hyperedges. (iv) In all applications mentioned above, conduct depth comparison based algorithms, other