ML-CrAIST: Multi-scale Low-high Frequency Information-based Cross Attention with Image Super-resolving Transformer

Paper Link: arXiv, Springer.

This paper has been accepted at the 27th International Conference on Pattern Recognition (ICPR 2024).

The official repository with Pytorch

Installation

Python 3.9.12

• create a virtual environment

python3 -m venv ./venv_name

• activate virtual environment

source venv_name/bin/activate

• install dependencies

pip3 install torch torchvision opencv-python matplotlib pyyaml tqdm tensorboardX tensorboard einops thop

Train

• Train the ML-CrAIST (Ours)

python train.py -v "CrAIST_X2_V1" -p train --train_yaml "trainSR_X2_DIV2K.yaml"
python train.py -v "CrAIST_X3_V1" -p train --train_yaml "trainSR_X3_DIV2K.yaml"
python train.py -v "CrAIST_X4_V1" -p train --train_yaml "trainSR_X4_DIV2K.yaml"

• Train the lighter version of ML-CrAIST (Ours-Li)

python train.py -v "CrAIST_X2_48" -p train --train_yaml "trainSR_X2_DIV2K_48.yaml"
python train.py -v "CrAIST_X3_48" -p train --train_yaml "trainSR_X3_DIV2K_48.yaml"
python train.py -v "CrAIST_X4_48" -p train --train_yaml "trainSR_X4_DIV2K_48.yaml"

Fine-tune

python train.py -v "CrAIST_X2_V1" -p finetune --ckpt 79

Test

Use version "CrAIST_X2_V1" for ML-CrAIST model (Ours) and "CrAIST_X2_48" for lighter model (Ours-Li).

--	Ours	--		--	Ours-Li	--
Scale	Version	Epoch		Scale	Version	Epoch
2x	CrAIST_X2_V1	414		2x	CrAIST_X2_48	761
3x	CrAIST_X3_V1	584		3x	CrAIST_X2_48	911
4x	CrAIST_X4_V1	682		4x	CrAIST_X2_48	766

• e.g.,

python test.py -v "CrAIST_X2_V1" --checkpoint_epoch 414 -t tester_Matlab --test_dataset_name "Urban100"

• provide dataset path in env/env.json file
• other configurations are done using yaml files

Citation

@inproceedings{pramanick2025ml,
  title={ML-CrAIST: Multi-scale Low-High Frequency Information-Based Cross Attention with Image Super-Resolving Transformer},
  author={Pramanick, Alik and Bheda, Utsav and Sur, Arijit},
  booktitle={International Conference on Pattern Recognition},
  pages={291--307},
  year={2025},
  organization={Springer}
}

Acknowledgement

• https://github.com/Francis0625/Omni-SR