New developments in character recognition

Today, Oct 1999 by Chrismer, Denny

Despite the growth of computers, most people still fill out hundreds of forms a year. Processing them is a difficult and time-consuming. Natural handwriting recognition can make the task easier.

CR and ICR are decades old technologies that reduce the need to keyenter data. The problem is they both have limitations - they can only recognize machine- or constrained handprinted characters.

OCR (Optical Character Recognition) is a pattern-matching technology that translates the shapes and patterns of machine-made characters into computer codes by reading one character at a time. On the other hand, ICR converts hand-printed characters to their machine (ASCII) equivalent.

OCR

Consider the letter "Y" Visual cues let humans recognize it in a variety of fonts ranging from "Y" to "Y." When visual cues fail, we perform a series of highly complex steps including:

1. Adding more light.

2. Holding the document closer or further away.

3. Using linguistic rules to determine what the hard to read character is.

4. Repeating portions of this process until we have an answer.

Our mind performs these complicated adjustments simultaneously in a split-second. Computers perform these processes sequentially.

OCR typically needs a near-perfect image in terms of contrast, character, clarity, etc. It then applies a patternmatching scheme, and, finally, uses the linguistic rules that are embedded in its engine technology.

With machine print, these adjustments are limited because the individual characters have common design elements or visual cues. Letters are evenly spaced across and up and down the page and generally printed with an eye to maximum foreground and background contrast separation.

While most OCR engines can recognize just about any type of machine print, some fonts are harder to read quickly and more accurately than others.

ICR

ICR (Intelligent Character Recognition) is more complicated than OCR because there are many more details to consider when converting handprint than machine print. ICR must ignore everything from individual handwriting styles to the type of pen used. This requires a different pattern matching technology. The visual and linguistic models also use a completely different set of elements.

The differences are significant. An ICR engine recognizes that people don't write the same letter, the same way, every time they write it. And that they write a letter differently at the beginning of a word than in the middle or at the end of a word. ICR uses visual cues to recognize these variations. This can be done using a fairly impressive system of logic, pattern matching and intelligence.

NHR Technology

NHR technology has many advantages over both OCR and ICR technologies because no matter how hard you try to constrain, restrict and predict handwriting, people still write the way they normally do - often a combination of print and script.

NHR breaks writing into a unique set of cursive moves that characterize all handwriting and treat input "holistically," executing linguistic and context rules on the fly, instead of recognizing character patterns and then running them against a context database or executing and spelling out the rules.

To understand the various letters in a word, the NHR engine uses clues to determine individual characters and letters. The "visual cues" for matching ICR characters with patterns rely on individual letter shapes. Rather than plotting a series of points for a letter and trying to understand the letter shape (like a dot-to-dot puzzle), NHR uses the swooping motions people make when they write to create all of the different letter shapes used in handwriting.

Parascript's Method

Companies do this in different ways. Parascript uses a set of formative symbols -- called XR elements - that embody the essence of cursive handwriting. These XR elements are a set of eight swooping motions that can be used in any combination to form all of the letters used in handwriting.

These elements can be combined to form a variety of letter shapes (figure 1). The NHR technology uses these elements to read an entire word, and parse out the individual letters within the word.

This is similar to the way humans read handwriting. Swoops are recognized as representing symbols in our written language. If NHR was primarily a pattern-matching technology like OCR and ICR, this would be enough.

Ambiguities

However, handwriting contains many ambiguities, making this process slow and inaccurate. That's why we haven't read words as a series of letters since the second grade. In order to read at speed, we read entire words, simultaneously processing the other recognition aspects after that, performing linguistic and context analysis in parallel.

NHR technology lets a computer mimic the cognitive processes we use when we read. It performs all recognition process steps simultaneously.

The following example can be read as "clear" or "dear." On a character recognition basis both are correct. This ambiguity can't exist in machine print - it's caused by the nuances of this person's cursive scripting. From a pattern matching perspective, we would be able to see the differences more easily with machine print, than with handwriting. The same is true for a computer.