Optical Character Recognition

Optical character recognition (OCR) uses machines—computers and sometimes other peripherals such as scanners—and software to recognize printed text characters. OCR enables users to digitally “read” and store printed text, thereby reducing the amount of typing required for inputting text. Scanning and interpreting text may seem simple, but the technologies required are complex; the potential uses are broad, important, and may be scaled toward increasingly general and difficult recognition problems.

The first general-use OCR system was created by futurist Ray Kurzweil in 1976. His company, Computer Products, Inc., released a commercial version two years later.

OCR is an analog-to-digital process; it begins with analog materials, and converts them to digital data. In order to scan printed words (typically, black text printed on a white page), a computer must utilize a charged coupled device (CCD)—a scanner. The CCD is charged by light, and successively codes and records the reflections at each point on the image. Scanning a text [Page 360]document turns it into a bitmap. The OCR software then analyzes the light and dark areas of the bitmap, and translates the results to a computer for storage and output. The success of the process depends in large part on the quality of the source material (clarity of the image being scanned), the effectiveness of the hardware (for example, the resolution ability of the scanner being used), and the quality of the software (its level of sophistication and accuracy).

The most difficult part of the process is the translation from light/dark images to words, a procedure that requires sophisticated pattern recognition. The OCR software must match shapes to character definitions. Since there is an abundance of fonts (letter and number character-shape sets) and languages (many using a wide variety of letter and number character-shape sets), there are many complexities and ambiguities that must be recognized, interpreted, and translated. Introducing non-regularized characters, like those produced by handwriting, further complicates the process.

OCR software must first find and note the boundary and region of a pattern; it must then match the object to items with which the object is conceptually related. The context in which the item is used informs the machine's ability to make choices about its identity. For example, one does not often find two instances of the letter “y” together in English words, so if there are two adjacent shapes that both look like “y,” at least one of them is probably something else. The software must then select the proper representation pattern. Since pattern recognition is a sort of under-constrained problem—one in which there isn't enough information to know that the arrived-at solution is uniquely correct—the system must also include other means of evaluating the accuracy of the solution.

The algorithms in OCR and other character-recognition software work using complex trial-and-error functions that model and simplify human cognitive systems. Using fuzzy logic and concepts of neural networking, computers (and programs) can be “taught” to recognize ambiguous or previously unknown characters (via repetitive trial and error using complex algorithms), thereby enabling scanners and OCR software to recognize handwriting or exotic languages, and/or to respond to idiosyncratic written input.

...