Saturday, June 19, 2004

How to Design a Form for Effective Automatic Character Recognition

Technoflak missed last month's TAWPI meeting. Fortunatley TAWPI has good notes on their site:

Charlie Montague introduced our presenter, Dr. Bernard Leikind, Doctus Solutions Manager for Mitek Systems, Inc. Dr. Leikind deals in sales, sales engineering, project management, training and customer support for Doctus and for Mitek's other toolkit products.

Dr. Leikind divided accuracy enhancement into four areas of concern: Registration, Data Fields, Form ID and Other Examples.
Throughout the evening, Dr. Leikind offered several hints to improve recognition accuracy on forms. For example, to perform a quick quality control check on forms printed in dropout ink, design the form with a dropout color circle around a black circle. If the circles aren't concentric, it's immediately obvious that the forms were misprinted and fields may not align properly.
Proper design of data fields can eliminate some ambiguity. For instance, try to avoid data fields with “I” and “1”, “O” and “0”. Also, the software must be able to pick out one character, big or small, felt tip or fine drafting point, slanted or straight. Noise, (such as speckles and boxes) must be removed, and ideally, the form should have enough room for writing without touching the boxes. Noise removal can be tricky for faxed forms, where dropout ink boxes become black also.

Form ID is another area where Mitek has worked hard to excel. Mitek software recognizes the line pattern of forms and tries to match it to a list of similar forms. It can handle photocopied forms and faxed forms, where the registration varies from one area of the paper to another.

When using Bar Codes for form ID where the form may be faxed, remember to orient the bar code so that the verticals are parallel to the length (scan direction) of the form, since faxes typically have 203 pixels horizontal by 93 pixels vertical. For bar code testing, Mitek uses

On forms, a rule of thumb for picking font sizes is that each recognized character should have 50 pixels.

When considering the ability of Neural Nets to intelligently identify data, one must remember that the smartest Neural Net algorithm is still pretty dumb. People can instantly recognize a five (5) for instance, no matter if stretched, squashed, turned sideways, broken, etc. Dr. Leikind remarked that programmers had still not come up with the characteristics of ‘fiveness', as people understand it.

Other factors influencing ICR accuracy: Image quality and resolution – 200 DPI is usually OK, as long as the forms are designed big enough.

For choosing a font for machine print, if you have control, pick a font that you and your system read well. OCR-A and OCR-B please the software, but are hard to read for people. Verdana is a good choice for ICR, people also like to read it, and it comes standard with MicroSoft, so most people have it.

A good form is clear, open, and attractive to computers. A good form is clear, open, and attractive to people. A good form assists people in filling it out accurately and quickly, and a good form can be accurately and quickly read by computers. Since we rarely can start from a blank slate in the design of a form, software attempts to overcome problems with image quality and form design, but this lowers accuracy and speed. Mitek's product DynaFind, for instance, studies the LAR amount on a check, and attempts to match it to the CAR amount to improve accuracy.

In concluding, Dr. Leikind gave some examples of humorously bad form design, and also of clean, roomy, and easy to read forms.

Jim Everett thanked Dr. Leikind for an excellent and witty presentation.

Note- CAR stands for Courtesy Amount Recognition: Machine reading of the hand written numerical amount of a check. Technoflak was unable to find a definition of LAR.

No comments: