Jan Rupnik - A Hidden Markov Model based approach for solving the Optical Character Recognition Problem

Optical Character Recognition is the translation of scanned images of text to machine-editable text. The problem is largely solved for Latin script OCR but other more complex scripts, such as Arabic, still present a challenge. Such scripts typically include a large set of possible characters, the shapes of the characters are context dependent, characters can be connected and many approaches suitable for Latin scripts are not applicable. Inspired by the success of Hidden Markov Model (HMM) based models in the speech recognition community this is the current paradigm in developing systems for solving the Optical Character Recognition (OCR) problem for connected scripts and multiple fonts. The HMM can be used to build a model of the stochastic process of generating words and their images and that model can be used for pattern recognition. I will present the basics of HMM theory and the ideas used in my approach.

Jan's slides are available here.