Teaching a Transformer to Read Azerbaijani Handwriting
Two-stage fine-tuning took character error rate from 17.23% to 3.47% on a language with almost no training data. The interesting part is not the architecture — it is that 5,000 real lines were worth as much as twenty thousand synthetic ones, and that most of the error left over sits on seven letters.
OCRLow-Resource NLPTransformersTrOCR