Exploration of Contextual Constraints for Character Pre.pdfVIP

Exploration of Contextual Constraints for Character Pre-Classification Tin Kam Ho George Nagy Bell Labs, Lucent Technologies Rensselaer Polytechnic Institute tkh@ nagy@ Abstract In most previous work - selectively cited below - such information has been used for character or word recogni- We pr esent strategies and results for identifying the tion. In contrast, here we propose to use the information symbol type (lower-case, upper-case, digit, and punctu- in a preprocessor. Recognition is much easier if the sym- ation or special symbols) of every character in a text bol shapes can be pre-classified into character-types, i.e., document by using various kinds of information fr om lower-case, upper-case, digit, and punctuation or special neighboring characters. In the expectation of reason- symbols. able word and character segmentationfor shape cluster- OCR based on context alone has already been ing, we designed several type recognition methods that demonstrated when the text is segmented at the word and depend on cluster n-grams, shape codes, and within- character level, restricted to a single type-case, and all word context. On an ASCII test corpus of 925 arti- digits, special symbols and punctuation are suppressed. cles that simulates perfect image-level pr ocessing, these The clusters can then be mapped to the alphabet either methods achieve a substantial improvement over default by considering the frequencies of N-grams of cluster- assignment of all characters to lower case.