1

I am developing an automatic number plate verification system. I have managed to locate the number plate area(rectangles) on the vehicle image but I need to filter out the non-number plate rectangular areas on the image.

I am using criterias such as

  1. Minimum number plate width x height so that smaller non-number plate areas can be filtered out

  2. Number plate width/ height ratio

these two criterion has helped me reduce the number of non number plate areas. but still I have a few candidate regions which are not number plate areas. Can someone suggest me some criteria that i could use to return the exact number plate region.

I am using C# and aforge.net

Thanks

Mr.Noob
  • 1,005
  • 3
  • 24
  • 58

1 Answers1

2

It will be impossible to guarantee anywhere close to a 100% matching record in real-world scenarios. Be careful with size and ratio issues because some legal number plates can be dramatically different sizes and ratios. Such as the "Q" plates, (Qld) and things like Trailer/Bike Rack plates out on the roads.

If you're getting a reasonable hit rate, and ensuring that you're are getting nearly all plates plus a few false-positives, then process/OCR all hits and pick the "best" match. In cases where you detect false-positives, but find a single match for plate suspect, flag them for review. (low urgency) Cases where you get no match, or multiple matches, flag for review at high urgency.

You can prioritize placement in the image (depending on whether you're capturing front or back images, front should be easier for placement) but again this cannot be too strict as trucks and bike-racks can have plates in less expected regions of the image, plus people that put them in the rear window. (no idea how legal that is.)

On a non-technical note, if you have control over the hardware then be sure to use an Infra-red camera. Plates are manufactured using IR reflective inks. (usually the background) This aides OCR contrast, but also filters out personalized backgrounds from the images. (So Daffy's face doesn't mess up the OCR.)

Steve Py
  • 26,149
  • 3
  • 25
  • 43
  • Thank you @Steve Py. I was thinking the same because there is no way in the world I would be able to get a 100% accurate result. but I have managed to bring the number of rectangular areas detected to a minimum of like from 2-7 false areas including the number plate which I think is good enough. Therefore I guess feeding them directly to the OCR engine wouldn't harm it's performance as well. also just add to this comment. I read in a research paper which said that the number of connected components in a candidate region would be between 2-15. but I am kind of lost trying it with aforge.net – Mr.Noob Jul 24 '12 at 06:18
  • just to add to my previous comment. what if i do a colour filtering on the final few candidate regions? – Mr.Noob Jul 24 '12 at 07:03
  • 1
    Colour filtering probably won't work too well simply because of the variety of plate designs that will likely be available. Sorry, for a minute I was thinking this was an Aussie site (Whirlpool) so my reply was based on Aussie licence plate designs. Depending on where you are and where your code will be expected to be running you may have as much variety, or maybe even more in plate designs. This is where using an infra-red camera comes in because you get a black on white (or white on grey) image. Colour filtering might work for local general issue plates, but what about custom / out of state? – Steve Py Jul 24 '12 at 11:30
  • 1
    With the high contrast of infra-red, OCR hit-rates can be pretty reliable, and OCR is generally pretty fast. A reasonably low percentage of false-positive regions screened out at OCR shouldn't affect your throughput noticeably. – Steve Py Jul 24 '12 at 11:34
  • Yeah you are right. this system is being built targeting the UK number plates. OK I guess I'll go ahead just with those two filter criterion and it should do the job good enough I suppose. I got infra-red cameras but the images I am using are ones obtained from a digital camera so I guess once I set it up with an infra-red camera it should work fine. Thanks a lot for your input Steve! – Mr.Noob Jul 24 '12 at 11:45