Teaching a Computer to Read

📄 Project Abstract

This report outlines an investigation into the ideal information that a handwriting recogniser would need to be able to extract to perform accurate sentence recognition. Research into the area suggests that the length of a word, its shape and the letters at either end of a word should be investigated for this purpose. Two models are implemented to test these variables. The brute force method finds results of little value, and thus the modified Viterbi algorithm is recommended for further investigation in this field. The results suggest that word length is not useful for word recognition unless it can be fully accurate. Surprisingly, the recommended number of letters to be found at either end of a word is only one, which is due largely to the small average word length in the English language. The results show the need for a fairly sophisticated shape classifier, but indicate that it is possible to have redundant information from too many levels of classifications. Improvements to the sy...