Fast Convolutional OCR with the Scanning N-Tuple Grid

This paper introduces a novel high speed convolutional character recognition system. Convolutional mode oper-ation means that no prior localization or segmentation of characters is required, making this mode extremely robust. The method uses a 2-d n-tuple grid to sample the im-age, but decomposes the address calculations into two one-dimensional scans. This simple innovation leads to a very fast system, and speeds in excess of 100,000 recognitions per second have been achieved for a 10-class character recognition problem, when operated in convolutional mode. Quantitative performance results show an error rate of 4.3 % on the MNist dataset of isolated hand-written charac-ters. Qualitative results are presented on museum archive card images, indicating that the method has great potential for the character recognition component in a document im-age analysis system for images of this type.