search

robertknight / ocrs-models

PyTorch models for the ocrs OCR engine
29 stars 5 forks source link

Integrate synthetic data for text recognition training #10

Open robertknight opened 5 months ago

robertknight commented 5 months ago

The main improvement needed for Ocrs to be more useful is higher text recognition accuracy / lower error rate, especially with longer lines. Also for multilingual support, examples in more languages will be needed. The main plan to improve this is to expand the training data with synthetic images. There are a number of existing text generation projects that might be useful:

  1. https://github.com/ankush-me/SynthText
  2. https://github.com/Belval/TextRecognitionDataGenerator (forked here to add Pillow v10 support)
  3. https://github.com/clovaai/synthtiger
robertknight commented 4 months ago

The main improvement needed for Ocrs to be more useful is higher text recognition accuracy / lower error rate, especially with longer lines.

This was partly addressed in https://github.com/robertknight/ocrs/pull/32. Long line images are still squashed to 800px during training though, which needs to be fixed.