Monday, November 18, 2013

Recognizing handwritten letters: the state of the art with ICR technologies

Today, we’re going to talk a bit about one type of document capture technology: Intelligent Character Recognition (ICR) technology. ICR software which has been created to convert handwritten letters into text that can be recognized and read by computers.

This technology isn’t as advanced as OCR technology – it has a lot of problems and the results aren’t as accurate. The degree of difficulty in reading words that are written by hand isn’t even something that humans are good at mastering: when was the last time you tried to make out the content of a doctor’s prescription?

Nonetheless, this technology looks like it’s moving forward. Some companies such as Parascript are talking about ICR systems being able to read up to 95% of handwritten texts, with an error rate of about 2%. 

But what have been the developments of solutions when it comes time to read handwritten characters in documents and convert them into digital information? 

  1. The first solution that came to the minds of humans was to have people read digitized information on paper and then put it into the computer themselves. This system is still being used in government organizations, hospitals, banks, educational institutions, etc. It goes without saying that because this solution is manual, it’s also expensive; but it solves an even greater problem – recovering information – for businesses. 
  2. Next, we started using boxes in documents to force people to write in a specific space, as you can see in the example below:
    This was because the ICR technology that was available at the time couldn’t recognize the characters if they were touching each other. 
  3. After that, the idea of printing documents in "drop-out ink" (pastel colors, most of which would block reading by OCR) came along. With this, it became possible to make the ICR only read handwritten characters without added noise. According to Imerge Consulting, this solution alone could eliminate 60% of the workers dedicated to data entry. Up until now, ICRs worked by identifying letters one by one (box by box), but what’s being looked for in the industry is a “criteria of usability” at the field level (or, put another way, that the complete word or sentence is correct and makes sense).
  4. Truth is, we don’t live in a world where people always write inside the lines. That’s why ICR technologies have been reinforced over the past few years for freely-written text. Since the letters aren’t constricted within boxes, we face an endless number of additional problems; for example, what happens when the width of the letters varies, or the letters touch or overlap, etc. New algorithms are currently being used, some of which compare the handwritten characters against an immense database of images, analyzing the parts, linguistic patterns, etc. With all of this, results still aren’t as good as those obtained by using OCR. Those defending the use of OCR affirm that whatever accuracy rate reached still translates into reduced labor costs. What’s certain is that aspects like the reading of cursive written, still don’t have a solution that could fix them. 


No comments:

Post a Comment