1

I'm developing an ANPR for Persian plates, I've found the way to find plate, and with some methods I have reached the image below, now I need to remove all non-character objects from the image to process them later. there are some similar questions on SO but they have different image noise and also different aim. I have also tried Erode and Dilate but since characters are small and has low resolution, it destroys characters.

I don't want to use counters features because of performance. I need to remove these noises with some effects/filters. So this is not a duplicate question.

Here is some input images and outputs I need.

input: enter image description here

output:enter image description here

input:enter image description here

output: enter image description here

Matt
  • 12,848
  • 2
  • 31
  • 53
alireza_fn
  • 884
  • 1
  • 8
  • 21
  • Instead of removing Non-character objects, why not taking Character objects and process them? Since you already have training images in your database no? so you can match – Khalil Khalaf Mar 25 '16 at 13:32
  • @FirstStep I 'm developing a handwritten OCR which needs a clean text image to process. – alireza_fn Mar 25 '16 at 13:35
  • I am not sure about that. I know there is an OpenCV function that retrieves the best match contour out of a database. So break down the image into columns maybe (characters part of the image - one character every column), retrieve largest contour in every column and compare it with your data base using that function and good luck that's a cool application – Khalil Khalaf Mar 25 '16 at 13:37
  • 2
    have you tried text detection strategies like [MSER](https://en.wikipedia.org/wiki/Maximally_stable_extremal_regions) or [SWT](http://research.microsoft.com/pubs/149305/1509.pdf)? – Daniel Albertini Mar 25 '16 at 14:11

1 Answers1

0

At least in the western world licence plates have a fixed layout. Having this preknowledge it is sufficient to localize the plate and get its orientation.

Then simply crop the regions you are interested in.

We also have standardized characters optimized for machine readability. I don't know if this is the case for your characters as well. You should be able to apply any decent OCR to read the plates contents.

Another option would be to search for blobs. Then delete everything that is too small or too big too eccentric or whatever to be a character. Not sure if this dot above the U shaped character is important or can be omitted.

Piglet
  • 27,501
  • 3
  • 20
  • 43