Word Recognition by MLP-based Character Spotting and Dynamic Programming

This paper describes a method for handprinted word recognition, with the following characteristics: traditional pre-processing (relevant to single characters, obtained by word segmentation) is replaced by pre-processing based on piecewise normalization applied at whole words; feature extraction and character classification by MLP are performed in a sliding window fashion; the output string is matched with an ASCII word vocabulary by Dynamic Programming with the Levenshtein distance; a list of word candidates is issued. Afterwards, when the language is formally known, an appropriate parser can be applied to full Sentence Recognition. Preliminary tests on a medium size vocabulary show extremely promising results.