Even April Tags Can't Escape DL :)
I'm looking at this paper:
Nothing fancy but its the type of stuff that common ML tasks look like in the wild. Some interesting pointers: - This uses DL to identify the area that may have a tag. - Once we are sure that an area is of interest we "zoom" in and hand that over to the classic methods. - you are burning a lot more CPU but in exchange you get a substantial improvement in accuracy.
Code seems to be at least partially available on github