Sketch Symbol Extraction Based On Bayesian Network

Sketch symbol extraction from continuously drawn strokes provides the basis for semantic processing in Sketching Understanding, which involves two sub-problems: strokes grouping determining which strokes constitute the same symbol and symbol recognition identifying what these strokes represent. Strokes grouping present a challenge to the problem of sketch symbol extraction.Currently two popular methods are employed to extract symbols from sketches. The first relies on hints from users such as pushing a button after finishing one symbol, pausing awhile between two consecutively drawn symbols, or demands that symbol drawing must be done with one stroke, which not only result in a less than natural drawing environment but fail to group strokes. Although some system can group strokes by virtual of stroke spatial information and the existence of special marker symbol, strokes are often grouped wrongly.To address this problem, Bayesian network approach to symbol extraction is proposed in this thesis, in which most likely hypothesis generated from partial strokes is employed to aid stroke grouping. Because stroke grouping is guided by semantic knowledge, error rate decreases. By representing uncertainty present in sketches, this approach can support imprecision and high variability in drawing style. The main achievements this thesis reaches can be summarized as follows:1) Bayesian model for sketch symbols and the process of symbol extraction based on Bayesian network is discussed.To facilitate Bayesian network constructing, K2 algorithms is adopted to learn network structure instead of manually constructed from samples provided by user.2) A prototype for UML domain based on Intel? PNL Bayesian network tool is built to demonstrate the feasibility of this method.