Something that comes-up from time to time is the question “how accurate is your handwriting recognition technology”? It’s not a terribly surprising question as accuracy IS important. Also, the conceptual similarity to recognition of machine print (often called OCR) and business familiarity with that technology and its high accuracies makes folks curious. Just how good is handwriting recognition?
The answer is “it depends”. Ok, now stop rolling your eyes. The reason why I say “it depends” is because accuracy of handwriting recognition is heavily dependent upon the type of information you want to recognize, the scope of that information, and whether or not you have the ability to constrain possible recognition results.
In another blog I wrote, I talked about how handwriting recognition is drastically different from OCR largely due to the fact that handwriting essentially has a character font for each person on this planet. Your letter “A” is different from my letter “A”. Even my letter “A” is different depending on my mood, the type of writing instrument I’m using, and a number of other factors.
So the practical tools and techniques used by OCR for typed characters and words cannot be applied to handwriting recognition. While OCR works on fonts at a very basic level, and has since evolved to support any font through intelligent character analysis and mapping, handwriting recognition cannot make the same use of fonts or character mapping. We have to use a completely different set of techniques.
These techniques are based upon the theory that handwritten characters can be deconstructed into basic elements. But it doesn’t stop there. We have to gather and use large sample sets of real handwriting in order to train the system on how to associate these elements with different words – basic letters are just not enough information to successfully recognize handwriting. Because we use samples, results form our recognition rely heavily upon statistics.
It is the combination of use of statistical sampling and inability to map characters that leads to so much confusion. Here are some common questions:
I notice that your technology can successfully recognize the handwritten word “dictionary” on this form but not on this other form. Yet the second form is much easier to read. Why can’t it read the easier-to-read version?
Remember, handwriting recognition is based upon statically-valid sample sets. The objective is achieving your target recognition results over the entire collection of documents, not just one or two. Individual results of recognition should not be evaluated on a result-by-result basis. Rather the overall performance of recognition on your entire collection of documents should be what you are evaluating.
Can you train your handwriting recognition?
We do indeed train our underlying algorithms on large sample sets of handwriting. But training cannot be accomplished in the field by individual companies. Because handwriting recognition is based upon sophisticated statistical models (among other things), training cannot be accomplished on particular documents or small samples. The results would be something called “over fitting” that effectively makes the system work very well on the trained documents but not well on anything else. Rather we train our systems with the overall objective of achieving a target recognition and accuracy rate on a large sample of representative handwritten data. The result is that the system works very well on handwritten data across a broader set of documents.
Can I add new samples of handwriting to help with accuracy?
Hopefully after reading the answers to the first two questions, you’ll understand that the answer is “no”. Because our technology is not font based (remember, every person on the planet has his or her own font) and developed using large statistically-valid sample sets, the notion of improving recognition by adding examples of new writing is not viable or desirable. By introducing one or two new examples at a time, the result would be very good recognition “fitted” to those specific samples and a general degradation of the recognition overall.
The very good news is that, even though handwriting recognition is dramatically different from OCR and is not “field trainable”, the technology can yield dramatic labor savings and efficiency improvements if you have specific needs and scope for recognition.
Do you have any questions about recognition? Send them over!