Step 1: complete »
The clustering algorithm has been tuned and it’s working reliably—even on all sorts of odd foreign plates!
On top of that, I’ve added a de-skewing and extraction function. These are the results extracted from the previous image.
I’m happy with the speed too – these two were extracted in 0.2s. The only downside at the moment is that the clustering algorithm is roughly O(n^2) in terms of the number of contours, so on busier images it can take a second or two to do the clustering stage (and it’ll produce an occasional false positive).