Revision history [back]

The following approach will work only for relatively big and clear license plates:

1. Use adaptive thresholding to get the binary image of your license plate
2. Smooth them with median filter or morphology
3. Find contours
4. Filter them according to their area, perimeter, bounding box etc.
5. Calculate some features of your contours (or even just use raw pixel values as features, but normalize first)
6. Use some classifier (SVM, NN, RandomTrees, whatever) to classify your characters, alternatively use Tesseract

If your plate is not big enough or not thresholded well or dirty etc, some characters will merge and you will not be able to recognize them this way.