Basic OmniPage OCR Technologies
Understanding OCR 253
AnyPage
AnyPage technology is Caere’s proprietary dynamic auto-thresholding
technology. It is in many ways similar to the HP AccuPage 2™ technology
available with Hewlett-Packard scanners. AnyPage works with grayscale
images. This dynamic thresholding technology requires a great deal of
interaction between the scanner and the OCR software to vary the
definition of “background” on a dynamic, real-time basis.
On a page with multiple shades of background, from a colored side bar in
a magazine or newspaper, to a coffee stain on a business letter, AnyPage
automatically varies which shades of gray will be considered
“background” and which will be considered “type.” AnyPage does this
not just once for the entire page, or large blocks at a time, but over
thousands of regions per page, sensing the optimum level for each region.
The remaining levels are dropped out, leaving the image very clean and
sharp and transforming it into a bi-modal black and white image to be
delivered to the AnyFont OCR engine.
This technology allows OmniPage to recognize documents and parts of
documents that other OCR packages literally can’t even see. Black print on
gray stationary, sidebars in the Wall Street Journal, and stained, yellowed,
and faded documents. All these can now be seen, cleaned up, and
recognized.
An important benefit to Dynamic Automatic Thresholding is that not only
fringe or difficult documents need contrast adjustment; a variety of
regular documents do too. With AnyPage operating with a sheet feeder,
the user is assured that each document in the pile will be properly
recognized, with no need to stop and change the scanner settings as the
documents in the stack vary.
Compound Neural System
Just as our brain can recognize letters on a page even when part of the
letter is missing or misprinted, Caere has developed a compound neural
system that has “learned” to recognize characters even when the
characters have been damaged.
This neural system is used for higher accuracy on characters that are
distorted through printing and faxing and for text printed in small point
sizes.
Caere’s neural system consists of several neural networks whose
components are modeled after the nerve cells in a brain. Each network
consists of rows and columns of software-simulated neurons. Each neuron
(Caere actually employs “perceptrons” as the active neurons in
OmniPage) weighs evidence provided by the image and other neurons to
