splinter21 / Invertible-Image-Rescaling

Implementation of paper: Invertible Image Rescaling

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Invertible Image Rescaling

This is the PyTorch implementation of paper: Invertible Image Rescaling (ECCV 2020 Oral). arxiv.

Dependencies and Installation

  • Python 3 (Recommend to use Anaconda)
  • PyTorch >= 1.0
  • NVIDIA GPU + CUDA
  • Python packages: pip install numpy opencv-python lmdb pyyaml
  • TensorBoard:
    • PyTorch >= 1.1: pip install tb-nightly future
    • PyTorch == 1.0: pip install tensorboardX

Dataset Preparation

Commonly used training and testing datasets can be downloaded here.

Get Started

Training and testing codes are in 'codes/'. Please see 'codes/README.md' for basic usages.

Invertible Architecture

Invertible Architecture

Quantitative Results

Quantitative evaluation results (PSNR / SSIM) of different downscaling and upscaling methods for image reconstruction on benchmark datasets: Set5, Set14, BSD100, Urban100, and DIV2K validation set. For our method, differences on average PSNR / SSIM from different z samples are less than 0.02. We report the mean result over 5 draws.

Downscaling & Upscaling Scale Param Set5 Set14 BSD100 Urban100 DIV2K
Bicubic & Bicubic 2x / 33.66 / 0.9299 30.24 / 0.8688 29.56 / 0.8431 26.88 / 0.8403 31.01 / 0.9393
Bicubic & SRCNN 2x 57.3K 36.66 / 0.9542 32.45 / 0.9067 31.36 / 0.8879 29.50 / 0.8946
Bicubic & EDSR 2x 40.7M 38.20 / 0.9606 34.02 / 0.9204 32.37 / 0.9018 33.10 / 0.9363 35.12 / 0.9699
Bicubic & RDN 2x 22.1M 38.24 / 0.9614 34.01 / 0.9212 32.34 / 0.9017 32.89 / 0.9353
Bicubic & RCAN 2x 15.4M 38.27 / 0.9614 34.12 / 0.9216 32.41 / 0.9027 33.34 / 0.9384
Bicubic & SAN 2x 15.7M 38.31 / 0.9620 34.07 / 0.9213 32.42 / 0.9028 33.10 / 0.9370
TAD & TAU 2x 38.46 / – 35.52 / – 36.68 / – 35.03 / – 39.01 / –
CNN-CR & CNN-SR 2x 38.88 / – 35.40 / – 33.92 / – 33.68 / –
CAR & EDSR 2x 51.1M 38.94 / 0.9658 35.61 / 0.9404 33.83 / 0.9262 35.24 / 0.9572 38.26 / 0.9599
IRN (ours) 2x 1.66M 43.99 / 0.9871 40.79 / 0.9778 41.32 / 0.9876 39.92 / 0.9865 44.32 / 0.9908
Downscaling & Upscaling Scale Param Set5 Set14 BSD100 Urban100 DIV2K
Bicubic & Bicubic 4x / 28.42 / 0.8104 26.00 / 0.7027 25.96 / 0.6675 23.14 / 0.6577 26.66 / 0.8521
Bicubic & SRCNN 4x 57.3K 30.48 / 0.8628 27.50 / 0.7513 26.90 / 0.7101 24.52 / 0.7221
Bicubic & EDSR 4x 43.1M 32.62 / 0.8984 28.94 / 0.7901 27.79 / 0.7437 26.86 / 0.8080 29.38 / 0.9032
Bicubic & RDN 4x 22.3M 32.47 / 0.8990 28.81 / 0.7871 27.72 / 0.7419 26.61 / 0.8028
Bicubic & RCAN 4x 15.6M 32.63 / 0.9002 28.87 / 0.7889 27.77 / 0.7436 26.82 / 0.8087 30.77 / 0.8460
Bicubic & ESRGAN 4x 16.3M 32.74 / 0.9012 29.00 / 0.7915 27.84 / 0.7455 27.03 / 0.8152 30.92 / 0.8486
Bicubic & SAN 4x 15.7M 32.64 / 0.9003 28.92 / 0.7888 27.78 / 0.7436 26.79 / 0.8068
TAD & TAU 4x 31.81 / – 28.63 / – 28.51 / – 26.63 / – 31.16 / –
CAR & EDSR 4x 52.8M 33.88 / 0.9174 30.31 / 0.8382 29.15 / 0.8001 29.28 / 0.8711 32.82 / 0.8837
IRN (ours) 4x 4.35M 36.19 / 0.9451 32.67 / 0.9015 31.64 / 0.8826 31.41 / 0.9157 35.07 / 0.9318

Qualitative Results

Qualitative results of upscaling the 4x downscaled images

Acknowledgement

The code is based on BasicSR, with reference of FrEIA.

Contact

If you have any questions, please contact mingqing_xiao@pku.edu.cn.

About

Implementation of paper: Invertible Image Rescaling

License:Apache License 2.0


Languages

Language:Python 93.3%Language:MATLAB 6.5%Language:Shell 0.2%