Abstract
Single-character recognition of mathematical symbols poses challenges from its two-dimensional pattern, the variety of similar symbols that must be recognized distinctly, the imbalance and paucity of training data available, and the impossibility of final verification through spell check. We investigate the use of support vector machines to improve the classification of InftyReader, a free system for the OCR of mathematical documents. First, we compare the performance of SVM kernels and feature definitions on pairs of letters that InftyReader usually confuses. Second, we describe a successful approach to multi-class classification with SVM, utilizing the ranking of alternatives within InftyReader’s confusion clusters. The inclusion of our technique in InftyReader reduces its misrecognition rate by 41%.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have