Cross-Lingual Learning in Multilingual Scene Text Recognition

Official PyTorch implementation | paper | training and evaluation data | pretrained model |

Preliminaries:

1. Multilingual STR: A task to recognize text from multiple languages in a word- or line-level scene image.
2. Cross-Lingual Learning (CLL): A methodology for transferring knowledge from one language to another.

What we did:

1. Verified two general insights about CLL and showed that the two general insights may not be applied to multilingual STR.
2. Showed that CLL is simple and highly effective for improving performance in low-resource languages.
3. Found the condition where CLL works well: the crucial condition for CLL is the dataset size of high-resource languages, regardless of the kind of high-resource languages.

More detail:

1. In this work, we investigate CLL for multilingual scene text recognition (STR).
   CLL transfers knowledge from one language to another. Often, the number of resources in each language is not even. There are high-resource languages such as English and low-resource languages such as Korean. Generally, the performance on a low-resource language is low, and improving the performance requires more resources. In this case, instead of obtaining more resources, applying CLL from an existing high-resource language to the low-resource language can improve the performance, as shown in the figure below.

(a) Monolingual training on a low-resource language Korean (KOR) results in incorrect predictions, whereas (b) CLL from a high-resource language English (ENG) to KOR leads to correct predictions.

2. We aim to find the condition where CLL works well. To do so, we first verify two general insights about CLL discussed in previous works (MRN and M-BRET): (a) joint learning with high- and low-resource languages may reduce performance on low-resource languages, and (b) CLL works best between similar languages. By using a representative multilingual scene text dataset MLT19, we show that the two general insights may not be applied to multilingual STR.

Two general insights about CLL related to (a) joint learning and (b) cascade learning (CL).

3. After that, through extensive experiments with additional data SynthMLT, we empirically show that the crucial factor is the number of samples in high-resource languages rather than the similarity between languages. In other words, CLL works well when we have sufficiently large samples in high-resource languages regardless of the kind of high-resource languages. We assume this is because the essential knowledge of STR is distinguishing text in the image and can be learned from any language.

We hope this work helps get more insights into multilingual STR.

Getting Started

Installation

This work was tested with PyTorch 1.12.1, CUDA 11.3, Python 3.8, and Linux.
Install conda >= 4.11.0 (python >= 3.8) and then run the following commands.

conda create -n CLL python=3.8 -y
conda activate CLL

git clone https://github.com/ku21fan/CLL-STR.git
cd ./CLL-STR
pip install torch==1.12.1+cu113 torchvision==0.13.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113
pip install -r requirements.txt

Run demo with pretrained model

1. Download pretrained model
   For example, download SVTR_joint_All.pth for demo.

   wget -O SVTR_joint_All.pth https://www.dropbox.com/scl/fi/q9r2ebbnr94xv4azyl7o1/SVTR_joint_All.pth?rlkey=wsonhqkffwyb4jixva44p0s5y
   

2. Add image files to test into demo_image/
3. Run demo.py

   CUDA_VISIBLE_DEVICES=0 python3 demo.py --model_name SVTR --image_folder demo_image/ \
   --saved_model SVTR_joint_All.pth
   

Prediction on demo images

Here, LAT is considered a high-resource language (28K) compared to other languages. The results below show the effectiveness of CLL.

demo images	SVTR_joint_All_but_LAT	SVTR_joint_All	Ground Truth
	اﻟﺤﻢ	اﻟﺒﺎب	الباب
	বান	রাখিবেন	রাখিবেন
	か本	北京西	北京西
	ज	मंडप	मंडप
	٥り	くすり	くすり
	상온	냉면은	냉면은
	디센다	CUISINE	CUISINE

Download preprocessed lmdb datasets for training and evaluation

We use MLT19 and SynthMLT datasets. MLT19 and SynthMLT consist of 7 scripts: Arabic (ARA), Bengali (BEN), Chinese (CHI), Hindi (HIN), Japanese (JPN), Korean (KOR), and Latin (LAT). LAT consists of English, French, German, and Italian.

Download the preprocessed lmdb datasets from here, or use wget as follows.

wget -O CLL-data.zip https://www.dropbox.com/scl/fi/js1y8dd6y0sethbuge1zl/CLL-data.zip?rlkey=bca3631p91mw1f98qhh79qprg
unzip CLL-data.zip -d data/

The structure of data/ is as follows:

data
├── training
│   ├── ARA
│   ├── BEN
│   ├── CHI
│   ├── HIN
│   ├── JPN
│   ├── KOR
│   ├── LAT
│   ├── SynthMLT_ARA
│   ├── SynthMLT_BEN
│   ├── SynthMLT_CHI
│   ├── SynthMLT_HIN
│   ├── SynthMLT_JPN
│   ├── SynthMLT_KOR
│   └── SynthMLT_LAT
├── validation
│   ├── ARA
│   ├── BEN
│   ├── CHI
│   ├── HIN
│   ├── JPN
│   ├── KOR
│   └── LAT
└── evaluation
    ├── ARA
    ├── BEN
    ├── CHI
    ├── HIN
    ├── JPN
    ├── KOR
    └── LAT

Training

1. Train SVTR with only one language data (LAT). For detail, check --select_data part in train.py.

   CUDA_VISIBLE_DEVICES=0 python3 train.py --model_name SVTR --exp_name SVTR_lang_L --select_data lang_L
   

2. Joint learning with all seven languages (Related to Table 2 in the paper)

   CUDA_VISIBLE_DEVICES=0 python3 train.py --model_name SVTR --exp_name SVTR_joint_All --select_data lang_CBHJAKL
   

3. Cascase learning: Pre-train SVTR on LAT then HIN (Related to Table 3 in the paper)

   CUDA_VISIBLE_DEVICES=0 python3 train.py --model_name SVTR --exp_name SVTR_L_to_H \
   --num_iter 10000 --saved_model saved_models/SVTR_lang_L/best_score.pth --select_data lang_H
   

4. Joint learning with all seven languages but using only 2K LAT instead of all LAT = Base setting of Table 4 in the paper.

   CUDA_VISIBLE_DEVICES=0 python3 train.py --model_name SVTR --exp_name SVTR_joint_Base \
   --data_usage_LAT 2000 --select_data lang_CBHJAKL
   

   Joint learning with Base setting + 2K SynthMLT KOR

   CUDA_VISIBLE_DEVICES=0 python3 train.py --model_name SVTR --exp_name SVTR_joint_Base_2KSynthMLT_K \
   --data_usage_LAT 2000 --data_usage_SynthMLT 2000 --select_data lang_CBHJAKL-SynthMLT_K
   

   Joint learning with Base setting + all SynthMLT KOR

   CUDA_VISIBLE_DEVICES=0 python3 train.py --model_name SVTR --exp_name SVTR_joint_Base_allSynthMLT_K \
   --data_usage_LAT 2000 --select_data lang_CBHJAKL-SynthMLT_K
   

Evaluation

Test SVTR model.

CUDA_VISIBLE_DEVICES=0 python3 test.py --model_name SVTR --saved_model saved_models/SVTR_joint_All/best_score.pth

Main arguments

train.py

• --train_data: folder path for training lmdb dataset. default: data/training/
• --valid_data: folder path for validation lmdb dataset. default: data/validation/
• --eval_data: folder path for evaluation lmdb dataset. default: data/evaluation/
• --model_name: select model 'CRNN' or 'SVTR'.
• --saved_model: assign the saved model for fine-tuning

test.py

• --eval_data: folder path for evaluation lmdb dataset. default: data/evaluation/
• --model_name: select model 'CRNN' or 'SVTR'.
• --saved_model: assign the saved model for evaluation.

demo.py

• --image_folder: path for image_folder which contains text images. default: demo_image/
• --model_name: select model 'CRNN' or 'SVTR'.
• --saved_model: assign the saved model to use.

Citation

Please consider citing this work in your publications if it helps your research.

@inproceedings{baek2024CLLSTR,
  title={Cross-Lingual Learning in Multilingual Scene Text Recognition},
  author={Baek, Jeonghun and Matsui, Yusuke and Aizawa, Kiyoharu},
  booktitle={IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year={2024}
}

Contact

Feel free to contact us if there is any question: Jeonghun Baek [email protected]

Acknowledgements

This implementation has been based on the repositories STR-Fewer-Labels and MRN.

License

MIT