Character Recognition with Hidden Markov Models (HMM)
Anthony Dipirro and Ji Mei - Undergraduate Symposium March 2012

qqq


Character recognition algorithms generally fall inot one of two categories: neural network (NN) or Hidden Markov Models HMM).  We adopt the perspective from HMM's , namely, that character recognition is an example of a noisy channel problem.  The noisy channel problem assumes that a  standard form of information is being transmitted over a noisy channel and that the noisy channel will distort the information received. The Viterbi algorithm , employing probabilistic methods, determines the most likely form of the original data.

Anthony and Ji used bit mapped images of characters and developed a modified Viterbi algorithm for recognition. Their model achieved over an 80% accuracy rate. To determine the generality of the algorithm, we expanded the test-bed of characters to include Chinese characters and again achieved the same level of accuracy.

Powerpoint slides from the presentation are here.