Hello Georg, I see a big problem in these training images: the background is not clearly separated from the digits.
My suggestion here is to apply a background removal filter first. If you have "purely white" digits on "purely black" background in both training and testing images, you will have better results. Generally, to perform the training, first remove the background (all colors similar to black should be replaced with black). All colors similar to white should be replaced with white. I don't know the name of this filter it is easy-to-implement in any programming language. Next, on the images for OCR, do the same - apply specific filters to remove the background and keep the digits in white. Svetlin Nakov Managing Partner Consulting and Information Technology Agency http://www.citagency.eu -----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of Georg Oberth Sent: Wednesday, September 30, 2009 11:37 AM To: [email protected] Subject: AW: OCR Digits only Hi Svetlin, Thanks for the advice. Using Arial 10pt is not easily implemented, since the displayed digits are no a meter and they are not Arial. I did assemble a tiff image of the digits in Photoshop and tried to use it, but the results were not that great. During the training I got a lot of errors in the log file about boxes overlapping, etc. which they were not (I attached a sample tiff). Do you know of some info as to how the actual training is done? This would allow me to assemble a tiff image which would not freak out the tess engine. Thank you & have a nice day! Georg Oberth [email protected] www.godata.at 0043-676-692-6070 0043-3127-88121 -----Ursprüngliche Nachricht----- Von: [email protected] [mailto:[email protected]] Im Auftrag von Svetlin Nakov Gesendet: Tuesday, September 29, 2009 6:58 PM An: [email protected] Betreff: RE: OCR Digits only Most people use for training font Arial, size 10pt. This is the official recommendation I found in the documentation. Attached is a sample training file for Bulgarian. I use a self-made tool to generate the TIF and BOX files, then print the page and scan it. Finally the training is performed. I cannot send the source code of the training image generator (it is under NDA). Svetlin Nakov Managing Partner Consulting and Information Technology Agency http://www.citagency.eu -----Original Message----- From: [email protected] [mailto:[email protected]] On Behalf Of georg Sent: Tuesday, September 29, 2009 12:41 PM To: tesseract-ocr Subject: OCR Digits only I am trying to OCR digits only and I made boxfiles for the font, but my results are not really good. I use softifree OCR as a reference program and it seems that their original files (the 8 files used for dictionary, etc) work best. My questions: How big (pixels) should my digits be in the tiff file? Are there any special issues I should be aware of when creating box files? How many samples should I box per page? Is it a problem that my digits are white on black background? Thanks in advance for your help. Regards, Goerg Oberth --~--~---------~--~----~------------~-------~--~----~ You received this message because you are subscribed to the Google Groups "tesseract-ocr" group. To post to this group, send email to [email protected] To unsubscribe from this group, send email to [email protected] For more options, visit this group at http://groups.google.com/group/tesseract-ocr?hl=en -~----------~----~----~----~------~----~------~--~---
