Using data mining for digital ink recognition

• Main Author: Blagojevic, Rachel Venita
• Other Authors: Plimmer, Beryl
• Published: ResearchSpace@Auckland 2011
• Online Access: http://hdl.handle.net/2292/7526

Computational recognition of hand-drawn diagrams has come a long way but is still inadequate for general use. This research uses data mining techniques to improve the accuracy of recognition. We focus on text-shape division as a challenging example that benefits from this approach. Surprisingly, although text is a fundamental part of diagrams it has been largely ignored. A review of the literature will show that feature-based recognisers are ideal candidates for solving these types of problems. Such recognisers require a good feature set and a suitable algorithm. For recognition to be successful, the features fed into the algorithms must provide good distinguishing characteristics between classes of interest. While small feature sets have been reported, currently there is no extensive survey of existing features employed for sketch recognition. Such a survey could act as a library for algorithms to employ for a given problem in sketch recognition. In addition, while various algorithms have been tried, there has been no extensive study of algorithms to determine the most optimal fit for accurate text-shape dividers. To build our text-shape dividers, we have assembled a comprehensive library of ink features that can be used for sketch recognition problems and compiled a large repository of labelled sketch data. To collect this data we built our own tool, DataManager, which includes support for collecting and labelling sketches as well as automatically generating datasets. Using this feature library and data repository a systematic investigation and tuning of machine learning algorithms has identified the algorithms best suited to text-shape division. The extensive evaluation on diagrams from six different domains has shown that our resulting dividers, using LADTree and LogitBoost, are significantly more accurate than three existing dividers. To our knowledge, these algorithms have not been used for text-shape division before.