Rely on a common/proven/maintained models retrieval logic

[drzraf]

Every new OCR solution seems to rely on its own set of model. EasyOCR, DocTR, OpenMMLab's MMOCR, ...
It's even worst when model are not retrieved from their upstream/official location leading to all sort of question about performances, training, ... (Dozens of issues in the project bugtracker about dbnet18, dbnet50, custom model, ...)

MMOCR seems to provide many models (and a clear list) https://mmocr.readthedocs.io/en/dev-1.x/modelzoo.html
Don't you think all the zip/download/config could be removed/unified so that model list/choice/selection is abstracted instead of being repeated with as many hardcoded-list as there are Python OCR projects?

The immediate benefit is that one can keep its usual codebase / library and switch/compare models with little to no changes involved.