Boosting OCR for Some Important Mutations

Optical Character Recognition (OCR) systems have to regularly deal with mutations. These mutations arise due to changes in the shape or effect of environment on the image. In this paper the effect of mutations on OCR has been investigated. The mutations investigated include dilation, skeletization, noising, rotation and outlining or erosion (by taking contour). The effect of these mutations on recognition accuracy of K-Nearest Neighbour and Support Vector Machine Approaches has been investigated. The results show that neither of the two techniques is efficient in recognition of mutations. Two approaches have been proposed in this paper which lead to better recognition for mutations. These include combining dimensionally reduced sets and removal of less relevant vectors from clusters of each class. These approaches lead to increase in recognition accuracy of the above said classifiers. The datasets used for the experimental investigations include USPS (Latin digits) and MADBase (Arabic digits).