As ROI increases OCR read fails to separate characters

[iniazi]

Occasionally OCR fails. During OCR training, when the ROI increases (in X direction; the background is same) the OCR does not seprate the characters. However, when ROI region is made smaller OCR appears to be working.

[vicky]

Hi,
Basically OCR works by the principle of template matching ,what i mean by this is when you do a ocr on a particular area OCR searched for charaters in this area and shows it up,so if you give a bigger ROI then it will search for more and charaters and give false results,if you give a larger ROI then you will have to set the % of matching in your program more so that the results are accurate.

[iniazi]

Thanks for the response. Sorry, I'm new, so how does one go about setting matching percent in the program.
In the image, I am trying to OCR the Credit field with ROI set as 652,65,753,88; which unfortunately fails. In OCR training tool with the ROIs 690,64,753,88 OCRing succeed. However ROI 670,64,753,88 and lrager fail. Maybe the magic is in the threshholding? Any Comments!!

[Jim Laudie]

To whom it may concern:

There are a couple of things you can try. The thresholding can play a part in what is going on. There are some other parameters that you can play with as well. Vicky mentioned the percentage that you can alter. I believe that she means the "Acceptance Level". "Read Strategy" and "Read Resolution" are other parameters that you will find helpful. Refer to the online help for OCR in order to get a better explanation of what these do. I'm linking here a document in our KnowledgeBase on our website which talks about how you can make your OCR application a little more robust. You'll find it at http://digital.ni.com/public.nsf/websearch/FC9B745C16429E1D86256C2A0069AFE2?OpenDocument.

One other approach is to
resample your image and achieve a higher resolution. I've attached a little segment of code below that will allow you to take your original image bufferPoll000 and specify some new dimensions in pixels and save it as a new image. In taking the dimensions up to 2000x2000, I noticed a great jump in performance. You probably wouldn't even have to go that far. You'll need NI-Vision on your computer to run this code as it makes a call to *.dll's installed with that package.

Hopefully this is of help to you. I really think that online help will give you some good direction. Best of luck...

Jim Laudie
Applications Engineer, National Instruments