1

I would like to match two license plate images, sample images given below

License plate 1
License plate 2

Here these two license plate belong to same vehicle, hence they should give match. There may be zoom and slight rotation in these images, also only a part of the original may be visible as given in the example. If the License plate belong to different vehicle algorithm should say it is different.

Which is best algorithm for doing this ?

scap3y
  • 1,188
  • 10
  • 27
  • Perhaps OCR can do the trick. Will that work for you? – GilLevi Dec 15 '13 at 08:50
  • In most of the cases i am getting wrong results during segmentation, in such cases i am not getting correct results in OCR –  Feb 06 '14 at 11:46

2 Answers2

1

Frankly this is a non-trivial question.

Just to list some obvious options:

  • Implement one of the numerous character recognition softwares, and get the string of characters, and then do a search for the substring in another string.
  • For images with almost no difference in zoom level, Use edge detection filters, like canny edge detection, to enhance the image, then use ICP (Iterative Closest Point), letting each edge pixel provide a vector to the closest edge pixel in the other image, with a similar value. this typically aligns images if they are similar enough. The final score tells you how similar they are.
  • For very large zoom levels, use multiple rotation and zoom hypothesis, and for each, scale the images and do cross correlation of the two images. select the hypothesis, that provides the coordinates with the best correlation, and use the point of correlation, as the x and y offset. The value of the correlation tells you how good a fit you have..

many other smarter algorithms have been produced for image fitting. However, you have much larger problems. The two example images you provide does not show the entire licenseplate, so you will not be able to say anything better than, "the probabillity of a match is larger than zero", as the number of visible characters increase, so does the probabillity of a match.

you could argue that small damages to a license plate also increases the probabillity, in that case cross correlation or similar method is needed to evaluate the probabillity of a match.

Henrik
  • 2,180
  • 16
  • 29
1

I would suggest you use openCV functions from Features2D Framework, and Homography method to handle the scaling and rotation problem. Specifically, in Features2D, there are classes that may be helpful for your to detect the letter, extract them, and match your two templates after extraction.

lennon310
  • 12,503
  • 11
  • 43
  • 61
  • What all features that I can use from Features2D, for detecting and extracting the letters –  Feb 02 '14 at 15:20
  • Use Scale-invariant feature transform to extract the lines at different orientations for each letter. Besides,feature point extraction might help, although it is used in OCR problem, which seems much more complicated than the matching problem. Basically you can investigate the possible pixel neighborhoods: http://www.ccs.neu.edu/home/feneric/charrec.html – lennon310 Feb 02 '14 at 15:52