Orientational features with the SNT-grid

The Scanning N-Tuple Grid (SNT-Grid) has been demonstrated to be a fast classifier for 2-dimensional images. The high speed is accomplished by scanning separately along rows and columns to extract features and can process thousands of pre-segmented characters per second in training and recognition. This paper proposes the use of orientational features within the SNT-Grid and makes a comparison in performance with features previously reported in literature. In terms of training the classifier, it explores cross entropy training and concludes that it outperforms more conventional maximum likelihood training. Finally, zoned orientational features offer a better implementation with an additional cost in computational time for training and recognition. The best accuracy reported has reduced the error rate of the system by 70% on the same dataset.