dikubab / amharic_ocr Goto Github PK

MMOCR is an open-source toolbox based and for details of installation and relate information see (https://github.com/open-mmlab/mmocr.

Geʽez/Abugida/Ethiopic script has up 519 characters. For Amharic, we use 289-319 characters depending on whether we use Ethiopic numerals and punctuation. There are no capital or small letters.

We have two datasets for the detection task. HUST-ART is the real word dataset, and HUST-AST is the synthetic dataset. HUST-ART consists of 1500 training images and 700 test images. HUST-AST comprises 75,904 training images. To convert the datasets labels to MMOCR format, use tools/data/textdet/icdar_converter.py as follows

python tools/data/textdet/icdar_converter.py det_datasets/HUST-ART -o det_datasets/HUST-ART -d icdar2015 --split-list training test

We have two training sets and two test sets datasets. Tana (TN) and Waliya (WL) training set consist of 2.85 and 4M cropped words, respectively. HUST-ART and ABE test sets consist of 4039 and 5218 text images. We also have a validation dataset consisting of 14835 text images, which is the training part of HUST-ART and ABE. All five datasets are in LMDB format.

MMOCR usage

1. In the directory configs/base/recog_pipelines/, you have different pipelines you must change dict(type='LoadImageFromFile') to dict(type='LoadImageFromLMDB'),
2. In the directory configs/base/recog_datasets/, you need to modify the path of test and train datasets.
3. In the directory mmocr/models/textrecog/convertors/ base.py define the dictionary using the 314 Amharic characters. No need to worry we have modified it. Based on your character set, modify dict_type in all other related files. We have modified the configs/textrecog/satrn/satrn_small.py settings. You can use it as an example.
   The datasets for both detection and recognition can be downloaded from the website https://dk-liang.github.io/HUST-ASTD/.

The Waliya related LMDB dataset link will be provided very soon.

1. Test and Validation sets LMDB https://mega.nz/folder/Ub0SnBBa#Fh6pFqbvXVxsa7OJPfJEwA
2. Tana(TN) LMDB https://mega.nz/folder/NGcC1DaQ#soagog8p_LgOnm6Gx9wdCQ

If you find our datasets are useful in your research, please consider cite:

@article{dikubab2022comprehensive,
  title={Comprehensive benchmark datasets for Amharic scene text detection and recognition},
  author={Dikubab, Wondimu and Liang, Dingkang and Liao, Minghui and Bai, Xiang},
  journal={Science China Information Sciences, Vol. 65, Special Focus on Deep Learning for Computer Vision, Article number: 160106},
  year={2022}
}

This project is released under the Apache 2.0 license.