5 Using Character Prototypes in Text Recognition 4 Extraction of Character Prototypes 3 Recognition of Stop Words

the words containing missing characters to the 43,300 word lexicon. Table 1: Results of regular expression matching with a 43,300 word lexicon (averaged over 200 pages). No. of words that are present on a test page 248.75 100.00% have missing characters 42.04 16.90% have a unique match of same length 26.74 10.75% have multiple matches of same length 11.27 4.53% have no matches of same length 4.13 1.66% 7 Conclusions We propose an adaptive text recognition strategy that is initiated by recognizing some common (stop) words. The recognized words are then aligned for extraction of common characters, and then recursively segmented using accumulated character prototypes. Character prototypes obtained from processing the stop words can then be used to train a classiier to recognize the rest of the text. By focusing on a relatively small set of words initially, more expensive word recognition techniques such as holistic word shape matching can be applied, so that the method can be applied to lower quality images. In several stages reject criteria are introduced to defend against unreliable recognition and character or word matches. The strategy uses a characteristic mix of whole word and isolated character based methods. Also, other techniques such as string matching and dynamic programming are found useful in the process. Moreover , we do not intend to completely abandon multi-font classiiers. Instead, we consider them as a supplementary tool for ambiguities that cannot be solved by the proposed procedure. Those include text in a non-dominant font (such as italics, boldfaces, numbers, or headers) and words with missing characters that cannot be matched uniquely to words in a general purpose lexicon. The study of the eeect of missing characters shows that the use of a multi-font classiier can be reduced to a minimum. Acknowledgements The author thanks George Nagy for reference 2], and him and YiHong Xu for helpful discussions and sharing of their prior work; Henry Baird for the defect model; John Hobby for trials of his bitmap matching and averaging algorithms; and Dan Lopresti and Raf-faele Giancarlo for discussions on string matching. Figure 3: Two pairs of words aligned for extraction of common characters (\a" in the left pair, \n" in the right pair). baseline of the both words, and a small local search is performed for the best vertical shift. The column-wise scores are summed over the width of the region which is dependent on the estimated character …