SynapseDirect

Another factor behind the variable vertical dimension ... how my lab prints on paper creates variability.

Idea: the small selection areas are what introduce errors,  a bigger selection area would by definition be easier to make sure the important text is within the image. 

Q: are there different algorithms for "Finding" the text to be OCR'd? 

Build 15 fixes all these errors.

I ran another test using build 15. I had 25 of my own documents, 8 of Jason's and a total of 13 rules.  Managed 100% recognition.

Tesseract seems to do a fairly decent job on some documents and poorly on others.

Another optimization might be to allow the user to specify which OCR engine to use .. 

• Option to use Tesseract on the regions after the form type has been recognized.  Use this on high quality deskewed scans.  Will speed up the recognition quite considerably.
• Regions are now cropped before being OCR'd and any regions that are pure white are no longer submitted.
• Renamed and copy the source files to the local directory before running ghostscript/imagemagick ... as some filenames generated by Paperport would cause an error  

Graham:

Build 18

• Regions are now cropped before being OCR'd
  

What's the cropping process. 

Removes any partial words fragments, or junk, above and below the area of interest.

Graham:

Removes any partial words fragments, or junk, above and below the area of interest.

sorta like I depecit here ?

?

Would I make my "Zonal OCR boxes" bigger for more accuracy ?   

That's the intent .. to allow taller boxes so we can be sure we are getting the desired text.

But if we get all the text of the line above, then that'll defeat the purpose. 

Zonal OCR: a fine balance.  Some differential equation should allow us to figure out the idea size :)

Bumped to build 19

If the date OCR fails on Tesseract, now switches back to Web OCR for another try.

Build 20

Switches back to Web OCR if name contains non-alpha characters.

Build 21

If a rule is selected in the rules table, it will be used as the first rule.  This is where you know which document type you are importing.  Good also for testing a particular rule.

Build 22

Crops the image horizontally as welll as vertically.

This means if your scan box cuts thru a letter on the left or right edges, these partial letter fragments are now discarded.