Sharper `Eyes' Through Neural Networks
Using optical character recognition (OCR) software to "read" print into computers is one way to cut down on paper clutter. But OCR programs can't interpret faxes with greater than 80% accuracy. "Noisy" phone lines produce smudged text that makes it difficult to distinguish an "l" from a "1," for example. To change that, OCR developers are turning to neural-network technology.
Mimicking the brain's circuits, neural nets seem to learn from past mistakes. So for OCR tasks, the software takes into account difficult text it has previously been exposed to, in addition to mathematical descriptions of characters. Image-In Inc., an OCR software developer in Minneapolis, says its forthcoming Image-In/Read program, based on neural-net techniques, has an accuracy rate of up to 99.03%. Other software makers, including market leader Caere Corp. in Los Gatos, Calif., are also working on neural-net OCR programs.