Hangul Single Text Recognition model

Structure

This repository uses pytorch-lightning hevaily.

• model
• dataset
• config
• static
• utils

In model/, there are two models now: SimpleCNN and SATNet. You can specify which model you want to use with the arguments.

There are no straightforward inference / visualization options yet.

Instructions

You first need a dataset to run HaSTeR. Refer to KoTDG for details.
Dataset should be in the following format:

• <name>
  • train
    • %d_%s.jpg % (index, text) where index is interger and text is single hangul character. (e.g. 3885_알.jpg)
  • valid
    • same as train
  • tests
    • same as train

---

Install all python packages written in requirements.txt.
You might need to use venv, conda or Docker.

Execute the following to train the NN.

python3 run.py --config config/default --model

Run python3 run.py -h for a small help.

The logs and checkpoints will be save in logs folder.

Real World Data

If you want to train or test on real world data, you can use utils/scale.py to rescale & rename files.

Refer to utils/scale.py about the usage. It is recommended to process data in static/files/ directory.

Example: python utils/scale.py static/files/raw/*.jpg --out_path static/files/processed

---

Reference

https://github.com/sgrvinod/a-PyTorch-Tutorial-to-Image-Captioning