Mutual Affine Network for Spatially Variant Kernel Estimation in Blind Image Super-Resolution (MANet, ICCV2021)
This repository is the official PyTorch implementation of Mutual Affine Network for Spatially Variant Kernel Estimation in Blind Image Super-Resolution (arxiv, supplementary).
๐ ๐ ๐ News:
- Aug. 7, 2021: We add an online Colab demo for MANet kernel estimation
- Sep.06, 2021: See our recent work SwinIR: Transformer-based image restoration.
- Aug. 17, 2021: See our previous work on blind SR: Flow-based Kernel Prior with Application to Blind Super-Resolution (FKP), CVPR2021
*
- Aug. 17, 2021: See our recent work for generative modelling of image SR: Hierarchical Conditional Flow: A Unified Framework for Image Super-Resolution and Image Rescaling (HCFlow), ICCV2021
- Aug. 17, 2021: See our recent work for real-world image SR: Designing a Practical Degradation Model for Deep Blind Image Super-Resolution (BSRGAN), ICCV2021
Existing blind image super-resolution (SR) methods mostly assume blur kernels are spatially invariant across the whole image. However, such an assumption is rarely applicable for real images whose blur kernels are usually spatially variant due to factors such as object motion and out-of-focus. Hence, existing blind SR methods would inevitably give rise to poor performance in real applications. To address this issue, this paper proposes a mutual affine network (MANet) for spatially variant kernel estimation. Specifically, MANet has two distinctive features. First, it has a moderate receptive field so as to keep the locality of degradation. Second, it involves a new mutual affine convolution (MAConv) layer that enhances feature expressiveness without increasing receptive field, model size and computation burden. This is made possible through exploiting channel interdependence, which applies each channel split with an affine transformation module whose input are the rest channel splits. Extensive experiments on synthetic and real images show that the proposed MANet not only performs favorably for both spatially variant and invariant kernel estimation, but also leads to state-of-the-art blind SR performance when combined with non-blind SR methods.
- Python 3.7, PyTorch >= 1.6.0, scipy >= 1.6.3
- Requirements: opencv-python
- Platforms: Ubuntu 16.04, cuda-10.0 & cuDNN v-7.5
Note: this repository is based on BasicSR. Please refer to their repository for a better understanding of the code framework.
Download stage3_MANet+RRDB_x4.pth
from release and put it in ./pretrained_models
. Then, run following command. Or you can go to our online Colab demo for MANet kernel estimation to have a try.
cd codes
python test.py --opt options/test/test_stage3.yml
To prepare data, put training and testing sets in ./datasets
as ./datasets/DIV2K/HR/0801.png
. Commonly used datasets can be downloaded here.
Step1: to train MANet, run this command:
python train.py --opt options/train/train_stage1.yml
Step2: to train non-blind RRDB, run this command:
python train.py --opt options/train/train_stage2.yml
Step3: to fine-tune RRDB with MANet, run this command:
python train.py --opt options/train/train_stage3.yml
All trained models can be downloaded from release. For testing, downloading stage3 models is enough.
To test MANet (stage1, kernel estimation only), run this command:
python test.py --opt options/test/test_stage1.yml
To test RRDB-SFT (stage2, non-blind SR with ground-truth kernel), run this command:
python test.py --opt options/test/test_stage2.yml
To test MANet+RRDB (stage3, blind SR), run this command:
python test.py --opt options/test/test_stage3.yml
Note: above commands generate LR images on-the-fly. To generate testing sets used in the paper, run this command:
python prepare_testset.py --opt options/test/prepare_testset.yml
To explore spaitally variant kernels on an image, use --save_kernel
and run this command to save kernel:
python test.py --opt options/test/test_stage1.yml --save_kernel
Then, run this command to creat an interactive window:
python interactive_explore.py --path ../results/001_MANet_aniso_x4_test_stage1/toy_dataset1/npz/toy1.npz
We conducted experiments on both spatially variant and invariant blind SR. Please refer to the paper and supp for results.
@inproceedings{liang2021mutual,
title={Mutual Affine Network for Spatially Variant Kernel Estimation in Blind Image Super-Resolution},
author={Liang, Jingyun and Sun, Guolei and Zhang, Kai and Van Gool, Luc and Timofte, Radu},
booktitle={IEEE International Conference on Computer Vision},
year={2021}
}
This project is released under the Apache 2.0 license. The codes are based on BasicSR, MMSR, IKC and KAIR. Please also follow their licenses. Thanks for their great works.