Around $70 billion is spent globally on handwriting records; scanning, storing and ‘reading’ them requires handwriting recognition software.
How often have you looked at your meeting or study notes and given up saying, “Aaargh! I can’t figure out what I wrote?” Now think of this: How often have you scanned your handwritten notes desperately looking for a particular name, word or figure? Most of us do this fairly often in order to locate details of a particular meeting, a study class, someone’s phone number taken down in a hurry, a to-do list, process steps for a task or just a handy recipe dictated by mom.
While we don’t think of decoding handwriting as a task, it is one we do each day. Combined with this is the act of “searching” through handwritten documents, looking for data, ideas, names, places, etc.
Can scanning handwriting and locating critical words be made easier? Can computers be used for the task? And, more importantly, are there areas in which computers need to be used for the task? A hospital administrator will answer that question fairly fast. Show him a reliable solution to handwriting recognition and he will beg for it. In hospitals across the world, doctors, nurses, technicians and administrative staff fill millions of forms each day – and in many cases, once filled, they are difficult to read or scan for critical data. In many instances, the inability to do this can become a threat to a patient’s life.
Similarly, can a railway clerk scan thousands of forms looking for handwritten data without making a mistake? Can handwritten insurance claim forms be read clearly and scanned for critical data? Can FIRs in a police station be scanned quickly for important material? Can students share their notes in a manner that can be understood by other classmates and friends?
About $ 70 billion is spent globally each year on handwritten forms of the kind used for medical records, insurance, police complaints, railway reservations, university work, employee records, invoices and financial details alone. In many cases, to meet statutory legal requirements the handwritten record needs to be scanned and stored electronically.
“The amount of paper that needs ‘reading' is huge,” says Thomas Binford, the founder, Chairman and Chief Technology Officer of Read-Ink, a Bangalore-based start up that has been developing handwriting recognition software for the last eight years. Binford, who has supervised more than 40 theses at Stanford University, while leading research in computer vision, artificial intelligence, medical image processing, robotics and industrial inspection, has been a research scientist at MIT and a Fulbright Scholar at TIFR, Mumbai.
Today, with the help of top-flight Indian talent, Binford is on the threshold of prototyping his solution to the problem of handwriting recognition, making any squiggle readable, searchable and shareable. The most common understanding we have of handwriting and character recognition is “ORC” or Optical Character Recognition — where a machine scans handwriting or text, examines it, matches it against patterns stored in its memory and spews it out as electronic text which can be stored, displayed, searched, edited and printed using a computer.
The problem is that ORC has only 95 per cent accuracy – and that’s an accuracy rate not acceptable in areas such as medicine, insurance, finance and banking. “At the moment, our product is able to achieve 98 per cent accuracy,” says Ione Binford, CEO of Read-Ink, “Our target is 99 per cent.” The Read-Ink system makes ambiguous guesses about characters and then makes a lexical match, it self corrects, learns with each use and becomes better. “People have different ways of forming characters and no two people are the same,” says Ione Binford. “Some people use shorthand that only they understand. Our system is geared to manage all these variants.”
“Only toilet paper is uniform in size and structure. But take the case of invoices — every invoice is different,” says Thomas Binford. This is a critical and valuable market to address. Put another way, the big market to address is the enterprise market.
On the other hand, the device that will lend itself the most to such hand writing recognition technology is the mobile phone. With mobile screens becoming touch sensitive, the input for natural writing is available. Now, you just need software that can recognise your squiggle, turn it into text and send it to your bank or to your passport officer.
Although there are several methods and devices that address the need for handwriting recognition, countries such as India, China, Brazil and many African nations will leapfrog these expensive solutions and adopt the answer that becomes available on cheap mobile phones.
Read-Ink thinks its technology has arrived at just the right time to encash that need – clearly, it has read the writing on the wall.
Brief history of handwriting recognition
CalliGrapher: The first two commercially available PDAs that used handwriting recognition, Apple MessagePad and Tandy Zoomer, used CalliGrapher. High expectations and low performance killed the product.
IBMs ThinkWrite: Used character patterns, strokes and timing to recognize handwriting – Met with limited success.
Graffiti: A Palm Computing product that found success because it completely sidestepped the problem of handwriting recognition. Instead, it forced users to write characters using certain strokes. Brilliant, but people still want to use “natural” writing. CEDAR Penman: An early system that tried to read naturally-written handwriting using an algorithm based on visual clues, running them over a neural network trained to figure out words.
OCR: Optical Character Recognition, the most common form of making text machine readable, storable and searchable.
MyScript: Uses a smart pen and paper to take down notes electronically and records the audio as well. The notes can be stored in a computer and made searchable and sharable using OCR.
Others: There are several pen and paper based devices priced between $100 and $150 that use scanning techniques to store handwriting and read it for further processing. None are as accurate as we want them to be.