3

I have a paper document that will be scanned, and then I'll want to perform image registration (image alignment) on different scans of different copies of the document.

I've noticed that paper forms often have a "registration mark" printed in the four corners of the piece of the paper (a cross-hairs: a circle with a plus intersecting in it). It looks something like this:

In my case, I have the freedom to choose the exact shape of the registration mark, to make it as easy as possible for the image processing code to detect the location of the four registration marks. My goal is for the code to detect registration marks as efficiently and robustly as possible, given that the image might be slightly rotated/skewed/translated. Is the "cross-hairs" shape shown above optimal? Are there better marks that are easier to algorithmically locate?

Black-and-white only; I can't print or scan in color, unfortunately.

D.W.
  • 3,382
  • 7
  • 44
  • 110
  • Related (but different, in that it focuses on algorithms to detect the cross-hairs rather than whether some other shape might enable algorithms that are faster or more robust): http://stackoverflow.com/q/9809769/781723 – D.W. Apr 14 '17 at 02:07
  • If colour is a possibility, I would have thought that two highly contrasting colours (cyan & magenta maybe) for the diagonally opposite quadrants of your circle, which do not normally occur in your documents would be easier to find faster than circles and crosshairs. Also, thinking about QR codes, it may be better to use 3 (or 5) markers rather than 4 so you know the correct orientation too. – Mark Setchell Apr 14 '17 at 07:26
  • Another, nearly random, thought. If your pages have diagrams with borders and tables, there will potentially be many crosses in your image, so maybe rotate some or all your crosshairs 45 degrees to differentiate them as diagonal crosshairs are no harder to locate than vertical ones during processing. – Mark Setchell Apr 14 '17 at 07:46
  • @MarkSetchell, thanks for the suggestions! Unfortunately I need a solution for black-and-white. Your other ideas are helpful -- thanks. – D.W. Apr 14 '17 at 16:36
  • any high contrast feature with unique shape will do for template matching. usually you know very well where to look for those features so finding them should be very fast and robust anyway. – Piglet Apr 14 '17 at 17:10

0 Answers0