I am looking for a method that does not need thousands of images, with various illumination levels. Maybe something like this : https://espace.library.uq.edu.au/view/UQ:151921/ivcnz07lovellpaper1.pdf
It should be able to get a 100% recognition rate.
Did anyone do any work on this? I'm afraid my maths are not up to par to implement such a proposal. Alternatively i am happy to pay for your time if you could come up with a solution. It should be non patented, and non GPL.