Hi,
I have eagerly followed many of the different courses on fast.ai and applied the lesssons to my own ML problems with amazing results 
There is one problem however that is quite resistant where I am not really able to get beyond the performance provided by more traditional image processing and ML techniques.
The problem is the following: I have images of car license plates taken at different locations, in for instance parking facilities. The challenge is to match DIFFERENT images of the same license plate while not matching images of different license plates. So if the system sees two different images of the SAME license plate it should declare a match, if the the images are of two truly different (in the sense that the text they contain are different) license plates , then the system shuld declare a mismatch.
My attempt at solving this was to train an autoencoder that I applied separately to the two images of license plates constituting a possible pair. I then had a simple series of dense layers comapre the two feature vectors extracted by the auto encoder. Training happened against a set of carefully labelled images of license plates.
This solution performs reasonably well, but is actutally outperformed by a more traditional system based on SIFT features and matching of local features.
Could anyone suggest improvements to my basic setup or point me in the direction of previous work in this domain?
Kind regards from locked down Oslo,
Lars