Machine recognition of off–line handwriting can be achieved by either recognising words as individual symbols (word level recognition) or by segmenting a word into parts, usually letters, and classifying those parts (letter level recognition). Whichever method is used, current handwriting recognition systems cannot overcome the inherent ambiguity in writingwithout recourse to contextual information.
This thesis presents a set of experiments that use Hidden Markov Models of language to resolve ambiguity in the classification process. It goes on to describe an algorithm designed to recognise a document written by a single–author and to improve recognition by adaptingto the writing style and learning new words. Learning and adaptation is achieved by
reading the document over several iterations. The algorithm is designed to incorporate contextual processing, adaptation to modify the shape of known words and learning of new words within a constrained dictionary.
Adaptation occurs when a word that has previously been trained in the classifier is recognised at either the word or letter level and the word image is used to modify the classifier. Learning occurs when a new word that has not been in the training set is recognised at the
letter level and is subsequently added to the classifier.
Words and letters are recognised using a nearest neighbour classifier and used features based on the two–dimensional Fourier transform. By incorporating a measure of confidence based on the distribution of training points around an exemplar, adaptation and learning is constrained to only occur when a word is confidently classified.
The algorithm was implemented and tested with a dictionary of 1000 words. Results show that adaptation of the letter classifier improved recognition on average by 3.9% with only 1.6% at the whole word level. Two experiments were carried out to evaluate the learning in the system. It was found that learning accounted for little improvement in the classification results and also that learning new words was prone to misclassifications being propagated.
