Recognition of handwritten text: basic concepts of a new approach

An approach to off-line recognition of scanned cursive handwriting is presented in the paper. Character HMMs and lexicon are used to generate HMMs for every word in the lexicon which are then matched with input images. Results are reported for the recognition of Cyrillic lower case cursive words using a small lexicon of 40 potentially confusing words. Using the three best hypothesis the recognition rate of 94.12% is achieved with simple features. The system is not restricted to any language. Roman orthography can be used with no system modifications.