Authors: Xiangtao Kong, Chao Dong and Lei Zhang
While single task image restoration (IR) has achieved significant successes, it remains a challenging issue to train a single model which can tackle multiple IR tasks. In this work, we investigate in-depth the multiple-in-one (MiO) IR problem, which comprises seven popular IR tasks. We point out that MiO IR faces two pivotal challenges: the optimization of diverse objectives and the adaptation to multiple tasks. To tackle these challenges, we present two simple yet effective strategies. The first strategy, referred to as sequential learning, attempts to address how to optimize the diverse objectives, which guides the network to incrementally learn individual IR tasks in a sequential manner rather than mixing them together. The second strategy, i.e., prompt learning, attempts to address how to adapt to the different IR tasks, which assists the network to understand the specific task and improves the generalization ability. By evaluating on 19 test sets, we demonstrate that the sequential and prompt learning strategies can significantly enhance the MiO performance of commonly used CNN and Transformer backbones. Our experiments also reveal that the two strategies can supplement each other to learn better degradation representations and enhance the model robustness. It is expected that our proposed MiO IR formulation and strategies could facilitate the research on how to train IR models with higher generalization capabilities.
⭐ If MiOIR is helpful to your images or projects, please help star this repo. Thanks! 🤗
More results of Restormer, Uformer and PromptIR (including the single task models) can be found in paper.
## git clone this repository
git clone https://github.com/Xiangtaokong/MiOIR.git
cd MiOIR
# create an environment with python >= 3.8
conda create -n MiOIR python=3.8
conda activate MiOIR
conda install -r requirements.txt
OR refer to the environment of BasicSR
Baidu Drive. Key: yxuw
We provide 27 pre-trained models (including SRResNet, SwinIR, Restormer, Uformer and PromptIR) that appear in the paper.
Baidu Drive. Key: yxuw
Please download MiO_test.zip and unzip it to MiOIR/data.
Edit the yml files in MiOIR/options/test that you need.
Please modify the path of your model and data.
cd MiOIR
CUDA_VISIBLE_DEVICES=0 python basicsr/test.py --launcher none -opt MiOIR/options/test/xxxx.yml
The results will be put in MiO/results.
Baidu Drive. Key: yxuw
Download all the data (including sr, blur, noise, jpeg, rain, haze, dark, GT .zip), intotal 120G for dowmload.
OR only download GT.zip and depth.zip, then generate the rest data by MiOIR/data_script/add_MiO_train_degradation.py, intotal 16G for dowmload.
Use data_script/gen_sub.py to crop the data to 300x300, modify some path and make sure save them with floder name of xxx_sub300.
Use data_script/gen_meta.py to generate training meta file, modify some path and the order of the sequential learning (_S) you want.
Edit the yml files in MiOIR/options/train that you need.Please modify the path of your model and data.
Note that:
Because the sequential learning (_S) need a series of yml files, please edit mixed learning (_M) first and then use options/gen_train.py to generate the yml files.
cd MiOIR
CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=2 --master_port=4176 basicsr/train.py -opt xxxxxx.yml
Note to modify the GPU card index and number --nproc_per_node= GPU number (correspond to it in yml).
If you use options/gen_train.py to generate yml files, it will also generate a .sh script in MiOIR/run_sh.
You could train Sequential Learning models through the script conveniently.
This project is based on BasicSR.
If you have any questions, please feel free to contact: xiangtao.kong@connect.polyu.hk
If our code helps your research or work, please consider citing our paper. The following are BibTeX references:
@article{kong2024towards,
title={Towards Effective Multiple-in-One Image Restoration: A Sequential and Prompt Learning Strategy},
author={Kong, Xiangtao and Dong, Chao and Zhang, Lei},
journal={arXiv preprint arXiv:2401.03379},
year={2024}
}
This project is released under the Apache 2.0 license.