Digitising a marked-up document means separating two kinds of ink that share the page: what was printed and what a person later wrote on top. Schools and offices want this constantly, because the value is usually in the annotations.

Deep learning handles it well and expensively. The models that do the job need more computation than the devices where the job naturally happens, which is a scanner, a phone or some other modest piece of hardware sitting on a desk. Accuracy that only exists in the data centre is not much use to a teacher with a stack of worksheets.

The framework here is built for the constraint rather than around it, targeting devices with severely limited computational capacity. Its descriptors are engineered specifically for the visual signature of human writing, which is a way of buying accuracy through knowing the problem instead of through raw model size.