PARSeq model pre-trained on various real
STR datasets at image size 128x32 with a patch size of 8x4.
PARSeq (Permuted Autoregressive Sequence) models unify the prevailing modeling/decoding schemes in Scene Text Recognition (STR). In particular, with a single model, it allows for context-free non-autoregressive inference (like CRNN and ViTSTR), context-aware autoregressive inference (like TRBA), and bidirectional iterative refinement (like ABINet).
You can use the model for STR on images containing Latin characters (62 case-sensitive alphanumeric + 32 punctuation marks).
1@InProceedings{bautista2022parseq,
2 author={Bautista, Darwin and Atienza, Rowel},
3 title={Scene Text Recognition with Permuted Autoregressive Sequence Models},
4 booktitle={Proceedings of the 17th European Conference on Computer Vision (ECCV)},
5 month={10},
6 year={2022},
7 publisher={Springer International Publishing},
8 address={Cham}
9}
