method: e2e-mlt + bug fix2019-04-11

Authors: Yash Patel, Michal Busta, Jiri Matas

Description: E2E-MLT, an end-to-end trainable unconstrained method for multi-language
scene text localization and recognition. The method is based on a single fully convolutional network (with shared layers for both tasks.
The results are plain network outputs - no dictionary has been used.

method: DIY FOTS2019-04-28

Authors: Qiaozhi

Description: This my own implementation of FOTS

method: TextProposals + DictNet2016-03-02

Authors: Lluis Gomez-Bigorda, Dimosthenis Karatzas

Description: Uses TextProposals [1] (a Text-specific Selective Search Algorithm for Word Spotting in the Wild) in combination with the DictNet [2] CNN.

Source code of the complete end-to-end system is available at: https://github.com/lluisgomez/TextProposals

[1] Lluis Gomez-Bigorda and Dimosthenis Karatzas "TextProposals: a Text-specific Selective Search Algorithm for Word Spotting in the Wild", arXiv:1604.02619 2016.

[2] M. Jaderberg, K. Simonyan, A. Vedaldi, A. Zisserman
"Synthetic Data and Artificial Neural Networks for Natural Scene Text Recognition", Workshop on Deep Learning, NIPS, 2014.

Ranking Table

Description Paper Source Code
DateMethodRecallPrecisionHmean
2019-04-11e2e-mlt + bug fix55.71%58.49%57.07%
2019-04-28DIY FOTS54.65%58.15%56.34%
2016-03-02TextProposals + DictNet37.89%89.84%53.30%
2018-10-25e2e-mlt49.45%55.13%52.13%
2021-05-25Char-level PTSD on SynthText18.73%33.19%23.95%

Ranking Graphic