SYENet

This repository contains the official implementation for ICCV 2023 paper:

SYENet: A Simple Yet Effective Network for Multiple Low-Level Vision Tasks with Real-Time Performance on Mobile Device

Weiran Gou, Ziyao Yi, Yan Xiang, Shaoqing Li, Zibin Liu, Dehui Kong, Ke Xu. [arxiv]

SYENet is an efficient network that could handle multiple low-level vision (isp, lle and sr) tasks. SYENet utilises re-parameterization for fast inference and got the highest score in MAI 2022 Learned Smartphone Challenge.

Citation

If you find our work useful in your research, please cite:

@InProceedings{Gou_2023_ICCV,
    author    = {Gou, Weiran and Yi, Ziyao and Xiang, Yan and Li, Shaoqing and Liu, Zibin and Kong, Dehui and Xu, Ke},
    title     = {SYENet: A Simple Yet Effective Network for Multiple Low-Level Vision Tasks with Real-Time Performance on Mobile Device},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {12182-12195}
}

Environment

python 3.8
pytorch == 1.12.1
numpy == 1.23.3
cv2 == 4.7.0
PIL == 9.2.0
tqdm == 4.64.1
yaml == 6.0

Configuration

Edit the yaml files (isp.yaml, lle.yaml, sr.yaml) in ./config.

You are recommended to use basicsr to train our sr models to get higher PSNR. We put the train/test configuration files for training/testing our sr models using basicsr in ./config, which are sr_basicsr_train.yaml and sr_basicsr_test.yaml.

Train

If you want to re-parameterize the model and save it, please set 'train $\rightarrow$ save_slim' parameter in the configuration yaml file to be true. And hence, the re-parameterized small model for fast inference will be saved.

For isp and lle tasks, we utilise a warmup phase which is a self-supervised training stage. This phase could be cancalled by setting 'train $\rightarrow$ warmup' parameter in the configuration yaml file to be false.

python main.py -task train -model_type original -model_task isp/lle/sr -device cuda

Test

Set the 'model $\rightarrow$ type' parameter in the configuration yaml file to be original if you are loading an original pretrained model, otherwise, set it to be re-parameterized for loading a re-parameterized model.

If you are loading an original pretrained model, but you want to re-parameterize it before inference, you could set 'model $\rightarrow$ need_slim' parameter in the configuration yaml file to be true. Notice that you cannot re-parameterize a re-parameterized model.

You could save the images generated in the test by setting 'test $\rightarrow$ save' parameter in the configuration yaml file to be true.

python main.py -task test -model_type original -model_task isp/lle/sr -device cuda

Demo

You could save the images generated in the demonstration by setting 'demo $\rightarrow$ save' parameter in the configuration yaml file to be true.

python main.py -task demo -model_type original -model_task isp/lle/sr -device cuda

liuguoyou / syenet