[droz]

Well written piece, kudos to the author.

If we are uploading cheques and licenses, why not just have a form with the respective upload sections for each document. Sure, people might get the two backward by accident, but is that error rate really any worse than the false positive rate of the proposed solution.

If you are going to go the machine learning route, then it really only makes sense if you go all the way and do some OCR on the documents to pick out the meaningful information to prepopulate some input screen that would likely appear after submitting those documents.

[wicknicks]

Yes, content analysis can be greatly simplified using symbolic inputs. GPS Tags in images, User input in OP's check example and most industrial pipelines used sophisticated barcode scanners to scan objects very quickly, but the cameras are at a fixed distance/height from the object and the objects are aligned in a particular direction. Some purists would argue that this is not computer vision, but it makes for some excellent engineering, and solves some hard problems very well.