Global feature for online character recognition

• We propose global features that extract global structure of online character pattern.
• Global features represent the relationship between distantly-positioned points.
• We validate global features through feature selection and recognition experiments.
• Global features are more often selected as important classifiers than local features.
• Global features achieve higher recognition rates than local features.

This paper focuses on the importance of global features for online character recognition. Global features represent the relationship between two temporally distant points in a handwriting pattern. For example, it can be defined as the relative vector of two xy  -coordinate features of two temporally separated points. Most existing online character recognition methods do not utilize global features, since their non-Markovian property prevents the use of the traditional recognition methodologies, such as dynamic time warping and hidden Markov models. However, we can understand the importance of, for example, the relationship between the starting and the ending points by attempting to discriminate “0” and “6”. This relationship cannot be represented by local features defined at individual points but by global features. Since O(N2)O(N2) global features can be extracted from a handwriting pattern with N points, selecting those that are truly discriminative is very important. In this paper, AdaBoost is employed for feature selection. Experiments prove that many global features are discriminative and the combined use of local and global features can improve the recognition accuracy.