OCR: A Lomg Way fr@m the O!d Dajs

Over the past 10 years, OCR has come a long way. Early desktop-based OCR solutions required tedious user training. The software would build a database, assigning an ASCII value to the bit-map image of each character it came across. Once trained, the OCR engine compared scanned character images to the images in its database, returning an ASCII character that represented the best match. Besides consuming time and resources, trainable packages did not work well when faced with a variety of fonts.

Instead of bit-map matching, omnifont technology uses feature extraction to distinguish characters, breaking down a character into its component parts and recognizing an image based on its uniqu e combination of shapes. Omnifont requires no training, consumes less d isk space, and, because a character retains the same basic features regardless of the font, works well across a broad range of typefaces.

Technology developed by Caere (maker of OmniPage Pro 6.0, WordScan Plus 4.0, and OmniPage Lite OCR, which comes bundled with the PaperPort software) not only recognizes single characters but evaluates entire words to make better predictions. The technology is based on the hidden Markov model, a complex mathematical model used in predictive technologies such as speech recognition and DNA matching.

The OCR engine looks at all the pixels that make up a word and then uses probability algorithms, including contextual evidence and language rules, to select the most probable word. Trigram analysis and dictionaries further improve accuracy. Trigram algorithms consider the probabilities of three words appearing together and can adjust the prediction of a single word based on the trigram probabilities.