Awesome
Dual Regression Compression (DRC)
Yong Guo, Jingdong Wang, Qi Chen, Jiezhang Cao, Zeshuai Deng, Yanwu Xu, Jian Chen, Mingkui Tan
This repository contains the official Pytorch implementation and the pretrained models of Towards Lightweight Super-Resolution with Dual Regression Learning.
Contents
Catalog
- We release the pre-trained models of some large SR models enhanced by our Dual Regression (DR).
- We release the pre-trained models of the non-blind SR models compressed by our Dual Regression Compression (DRC).
- We release the pre-trained models of the blind SR models compressed by our Dual Regression Compression (DRC).
- We release the code of evaluation and training.
Datasets
Used training and testing sets can be downloaded as follows:
| Training Set | Non-Blind Testing Set | Blind Testing Set |
|---|---|---|
| DIV2K (800 training images, 100 validation images) + Flickr2K (2650 images) | You can evaluate our non-blind models on several widely used benchmark datasets, including Set5, Set14, B100, Urban100, Manga109. | You can evaluate our blind models on DIV2KRK dataset. |
Please organize the datasets using the following hierarchy.
- datasets/
- DIV2K
- DIV2K_train_HR
- DIV2K_train_LR_bicubic
- DF2K
- DF2K_HR
- DF2K_LR_bicubic
- benchmark
- Set5
- Set14
- B100
- Urban100
- Manga109
- DIV2KRK
- gt
- lr_x2
- lr_x4
Models
You can download the pre-trained large SR models enhanced by our Dual Regression (DR) scheme for 4x SR. More pretrained models can be found in the released assets of this repository.
| Method | Params | FLOPs (G) | Dataset | PSNR (dB) | SSIM | Model Zoo |
|---|---|---|---|---|---|---|
| DRN-S | 4.8M | 109.9 | Set5 | 32.68 | 0.901 | Download |
| DRN-L | 9.8M | 224.8 | Set5 | 32.74 | 0.902 | Download |
| SwinIR-DR | 11.9M | 121.1 | Set5 | 33.03 | 0.904 | Download |
| DAT-DR | 14.8M | 155.1 | Set5 | 33.17 | 0.906 | Download |
You can download the compressed non-blind SR models (remove 30% parameters) obtained by our Dual Regression Compression (DRC) approach for 4x SR.
| Method | Params | FLOPs (G) | Dataset | PSNR (dB) | SSIM | Model Zoo |
|---|---|---|---|---|---|---|
| DRN-S30 | 3.1M | 72.3 | Set5 | 32.66 | 0.900 | Download |
| SwinIR-light-DRC | 635K | 6.8 | Set5 | 32.44 | 0.896 | Download |
You can download the compressed blind SR models (remove 30% parameters) obtained by our DRC approach for 4x SR.
| Method | Params | FLOPs (G) | Dataset | PSNR (dB) | SSIM | Model Zoo |
|---|---|---|---|---|---|---|
| DCLS-DRC | 14.2M | 57.1 | DIV2KRK | 29.01 | 0.798 | Download |
Evaluating and Training
We put the detailed explanations about the code of evaluating and training in the corresponding folders. Please refer to more details in the README.md file within these folders.
Results
We achieved competitive performance. Detailed results can be found in the paper.
<details> <summary>Click to expand</summary>- Comparison results with SOTA SR methods for 4x SR in Table 1.
- Comparison results with lightweight SR methods for 4x SR in Table 2.
- Comparison results with blind SR methods for 4x SR in Table 3
- Visual comparison (x4) with SOTA methods for 4x SR.
- Visual comparison (x4) with compression methods for 4x SR.
Citation
If you find this repository helpful, please consider citing:
@article{guo2022towards,
title={Towards lightweight super-resolution with dual regression learning},
author={Guo, Yong and Wang, Jingdong and Chen, Qi and Cao, Jiezhang and Deng, Zeshuai and Xu, Yanwu and Chen, Jian and Tan, Mingkui},
journal={arXiv preprint arXiv:2207.07929},
year={2022}
}