Comment by samus

Output really doesn't have to be the same datatypes as the input. Text tokens are good enough for a lot of interesting applications, and transforming percels (name suggested by another commenter here) into text tokens is exactly what an OCR model is anyway trained to do.