PowerQE

This repository supports some quality enhancement approaches for compressed image/video based on PyTorch and MMEditing.

Image approaches:

• MPRNet @ CVPR'21: Multi-stage structure.
• RBQE @ ECCV'20: Multi-exit structure and early-exit mechanism.
• CBDNet @ CVPR'19: Noise estimation. Originally for image denoising.
• ESRGAN @ ECCVW'18: Relativistic discriminator. Originally for image super resolution. PIRM'18 winner.
• RDN @ CVPR'18: Residual dense network. Originally for image super resolution.
• DnCNN @ TIP'17: Pioneer of CNN-based image denoising.
• DCAD @ DCC'17: Pioneer of HEVC compression artifacts reduction.
• U-Net @ MICCAI'15: Multi-scale structure. Originally for biomedical image processing.
• AR-CNN @ ICCV'15: Pioneer of CNN-based image compression artifacts reduction.

Video approaches:

• ProVQE @ CVPRW'22: Key-frame propagation. NTIRE'22 winner.
• BasicVSR++ @ CVPR'22: Flow-guided deformable alignment. Originally for video super resolution. NTIRE'21 winner.
• STDF @ AAAI'20: Deformable alignment.
• MFQEv2 @ TPAMI'19: Key-frame alignment.
• EDVR @ CVPR'19: Deformable alignment. Originally for video super resolution. NTIRE'19 winner.

Content:

• PowerQE
  • Install dependency
  • Prepare data
  • Training
  • Testing
  • Changelog

Resources:

• Pretrained models
• Document

Install dependency

MMEditing is a submodule of PowerQE. One can easily upgrade the MMEditing, and add their models to PowerQE without modifying the MMEditing repository. One should clone PowerQE along with MMEditing like this:

git clone --depth 1 --recurse-submodules --shallow-submodules\
 https://github.com/ryanxingql/powerqe.git

Install dependency:

• environment.yml
• PyTorch v1 + MMCV v1 + MMEditing v0

Please refer to the document for detailed installation.

Prepare data

mkdir data

Place your data like this:

powerqe/data
`-- {div2k,div2k_lq/bpg/qp37}
    `-- train
    `   `-- 0{001,002,...,800}.png
    `-- valid
        `-- 0{801,802,...,900}.png

Please refer to the document for detailed preparation.

Training

#chmod +x tools/dist_train.sh  # for the first time

conda activate pqe &&\
 CUDA_VISIBLE_DEVICES=0\
 PORT=29500\
 tools/dist_train.sh\
 configs/<config>.py\
 1\
 <optional-options>

• Activate environment.
• Use GPU 0.
• Use port 29500 for communication.
• Training script.
• Configuration.
• Use one GPU.
• Optional options.

Optional options:

• --resume-from <ckp>.pth: To resume the training status (model weights, number of iterations, optimizer status, etc.) from a checkpoint file.

Testing

#chmod +x tools/dist_test.sh  # for the first time

conda activate pqe &&\
 CUDA_VISIBLE_DEVICES=0\
 PORT=29510\
 tools/dist_test.sh\
 configs/<config>.py\
 work_dirs/<ckp>.pth\
 1\
 <optional-options>

• Activate environment.
• Use GPU 0.
• Use port 29510 for communication.
• Test script.
• Configuration.
• Checkpoint.
• Use one GPU.
• Optional options.

Optional options:

• --save-path <save-folder>: To save output images.

Changelog

Version	PyTorch	MMEditing	Video approaches
V3	V1	V0	Supported
V2	V1	V0	N/A
V1	V1	V0	N/A