Pseudo Two-Dimensional Hidden Markov Models for Document Recognition
Oscar E. Agazzi, Shyh-shiaw Kuo · AT&T Technical Journal · 1993
Hidden Markov models (HMM) have become the most popular technique for automatic speech recognition. Extending this technique to the two-dimensional domain is a promising approach to solving difficult problems in optical character recognition (OCR), such as recognizing poorly printed text. Hidden Markov models are robust for OCR applications due to: — Their inherent tolerance to noise and distortion, — Their ability to segment blurred and connected text into words and characters as an integral part of the recognition process, — Their invariance to size, slant, and other transformations of the basic characters, and — The ease with which contextual information and language models can be incorporated into the recognition algorithms. We give a brief overview of OCR algorithms based on two-dimensional hidden Markov models, and we present three case studies that show their remarkable strengths.