Chinese Journal of Computers   Full Text
Title	Adaptive Symbol Recognition for Sketch-Based Interfaces Based on Template Matching and SVM
Authors	JIANG Ying-Ying1),2) TIAN Feng1) WANG Xu-Gang1) DAI Guo-Zhong1),2)
Address	1)(Intelligence Engineering Laboratory, Institute of Software, Chinese Academy of Sciences, Beijing 100190) 2)(State Key Laboratory of Computer Science, Institute of Software, Chinese Academy of Sciences, Beijing 100190)
Year	2009
Issue	No.2(252—260)
Abstract &   Background	Abstract During adaptive learning of symbols in sketch-based interfaces, the number of training samples may be different for different users and it is challenging for recognition methods to learn with flexible sample numbers. This paper proposes an adaptive symbol recognition method for sketch-based interfaces. It combines template matching method that could learn with few samples and SVM method that could learn with more samples by a strategy related to sample numbers. Both online information and offline information are utilized. Thus it could learn and recognize with different sample numbers. Based on the proposed method, the authors build a symbol widget that supports adaptive recognition. At last, a prototype system, IdeaNote, is built based on the extended PIBG Toolkit. Evaluation shows that when there are 24 kinds of symbols, the method could achieve high recognition accuracy and good time performance with 1 to 20 training samples. Keywords symbol recognition; adaptive learning; template matching; SVM; classifier combination; widget Background This research is supported by the National Natural Science Foundation of China under grant Nos.60503054, U0735004, 60603073, and the National High Technology Research and Development Program (863 Program) of China under grant No.2007AA01Z158. Sketch-based interface is useful for early design, idea capturing and idea sharing. Currently, there’re some popular sketch-based systems, such as Silk, Electronic Cocktail Napkin and MathPad. These systems often predefined symbols that can be recognized, and users need to draw the symbol in the way the system defined. However, as different users often have different preferences for symbol definition and drawing, this predefined way is not very good. To allow users to define their own symbols would be a good solution to this problem. Adaptive symbol recognition is essential for recognizing users’ personalized symbols. When defining a new symbol, different users tend to use different numbers of samples. This requires the symbol recognition method could learn with adaptive sample numbers. This paper proposes an adaptive symbol recognition method that could learn with either small number of samples or large number of samples. Template matching method that could support small sample learning and SVM that could learn with more samples are combined to achieve adaptive symbol recognition. In addition, both online stroke information and offline image information are utilized to improve the recognition accuracy. A widget built on the proposed method is incorporated into PIBG Toolkit which was developed by the authors’ lab and could be used for building pen-based applications. Thus, the extended PIBG Toolkit could support fast building of adaptive pen-based systems. Evaluation shows that the method could achieve high recognition accuracy and good time performance. It could effectively build recognition methods according to users’ sample number.