An OCR system for printed Kannada using k-means clustering

We address the problem of Kannada character recognition, and propose a recognition mechanism based on k-means clustering. The large dataset of Kannada characters and their similarity makes the problem one order of magnitude more difficult than for a standard language like English. We propose a segmentation technique to decompose each character into components from 3 base classes, thus reducing the magnitude of the problem. k-means provides a natural degree of font independence and this is used to reduce the size of the training database to about a tenth of those used in related work. Consequently, recognition proceeds an order of magnitude faster. We present accuracy comparisons with related work, showing the proposed method to yield a better peak accuracy. We also discuss the relative merits of probabilistic and geometric seeding in k-means.