Prototype Pruning by Feature Extraction for Handwritten Mathematical Symbol Recognition

Stephen M. Watt and Xiaofang Xie

Summary

The authors present a mathematical symbol gesture recognizer using prototype pruning by feature extraction. Initially, a preprocessing stage smooths out strokes, resamples, size normalizes, and chops off heads and tails. The features used include a number of geometric features (number of loops, number of intersections, and number of cusps), ink related features (number of strokes, point density), directional features, and global features. Using the features, a prototype pruning process reduces the number of potential classes for an unrecognized symbol. An elastic matching recognizer is the final step, where an unknown symbol is classified.

The recognizer was evaluated on a test set of 227 symbols. The author's compared their results with those derived by J. Kurtzberg. While the accuracy of the authors' recognizer was 1% less than that of J. Kurtzberg's, the number of pruned prototypes was significantly greater for the author's approach. The author's approach pruned 85.8% of prototypes, while J. Kurtzberg's approach pruned 61.5%.

Discussion

The features selected by the authors are interesting. I can see a uniqueness in symbols coming from the number of loops, intersections, and cusps, especially with letters and numbers. The basis for their selection is what they say humans use in reconizing symbols. They don't present any evidence of this, though. Not that I necessarily disagree with them, but I think to make this statement without evidence or at least a reason for this insight seems bold.