Not All Patches Are Equal: Hierarchical Dataset Condensation for Single Image Super-Resolution

The pytorch implementation of ''Not All Patches Are Equal: Hierarchical Dataset Condensation for Single Image Super-Resolution'', SPL 2023.

Method

Requirements

Python 3.7
PyTorch == 1.7.0
numpy
skimage
imageio
matplotlib
cv2

Dataset condensation

1. Prepare condensation dataset

1.1 Download the DIV2K

1.2 Crop the HR images into image patches in 'data_path/DIV2K/HR'

2. Texture complexity-based condensation strategy

2.1 Measure the complexity of image patches

python texture_complexity_measurement.py

2.2 Remove the low-complexity image patches

python remove_low_complexity.py

3. Texture diversity-based condensation strategy

3.1 Cluster the image patches

python cluster_patches.py

3.2 Remove the image patches with similar textures in each cluster

python sample_patches.py

Performance evaluation

1. Train super-resolution network with condensed dataset

python main.py --model EDSR --scale 2 --save DCSR_X2 --patch_size 96 --batch_size 16

2. Test model performance

2.1 Prepare test data Download benchmark datasets (e.g., Set5, Set14 and other test sets) and prepare HR/LR images in /benchmark following the example of benchmark/Set5 2.1 Test the performance

python main.py --model EDSR --data_test Set5 --scale 2  --test_only

Citation

@ARTICLE{10305246,
  author={Ding, Qingtang and Liang, Zhengyu and Wang, Longguang and Wang, Yingqian and Yang, Jungang},
  journal={IEEE Signal Processing Letters}, 
  title={Not All Patches Are Equal: Hierarchical Dataset Condensation for Single Image Super-Resolution}, 
  year={2023},
  volume={30},
  number={},
  pages={1752-1756},
  doi={10.1109/LSP.2023.3329754}}

Acknowledgements

This code is built on EDSR (PyTorch) and ClassSR. We thank the authors for sharing the codes.

QingtangDing / DCSR

readme