An Introduction to Donut

An OCR-Free, Small Model-Based PDF Parsing Method

The previously introduced pipeline-based PDF parsing method primarily uses OCR engines for text recognition. However, it results in high computational costs, inflexibility with regard to language and document type, and potential OCR errors that may impact subsequent tasks.

Donut is an OCR-free model that circumvents these issues. It eliminates OCR dependency by directly mapping the original input image to the desired output.