Offline Cursive Character Recognition: A state-of-the-art comparison

Recent research has demonstrated the superiority of SVM-based approaches for offline cursive character recognition. In particular, Camastra’s 2007 study showed SVM to be better than alternative LVQ and MLP approaches on the large C-Cube data set. Subsequent work has applied hierarchical vector quantization (HVQ) with temporal pooling to the same data set, improving on LVQ and MLP but still not reaching SVM recognition rates. In the current paper, we revisit Camastra’s SVM study in order to explore the effects of using an alternative modified direction feature (MDF) vector representation, and to compare the performance of a RBF-based approach against both SVM and HVQ. Our results show that SVMs still have the better performance, but that much depends on the feature sets employed. Surprisingly, the use of more sophisticated MDF feature vectors produced the poorest results on this data set despite their success on signature verification problems.