Hey, this is pretty cool. I actually tried something similar. (Keeping a list of shop names and matching it with tesseract's results)
I was trying hough transform for slight image rotations. I wasn't aware of imagemagick's textcleaner script. That could have save me a lot of trouble :)
I got roadblocked by the problem of having various kinds of receipts with absolutely no layout in common. I figured it would need a lot of training for the system to have a decent accuracy and left it for another day.