hdrnet-pytorch-2023

The repository contains a alternative implementation of HDRNet (Deep Bilateral Learning for Real-Time Image Enhancement, SIGGRAPH 2017) in 2023.

environment

python 3.8

pytorch 1.13.1

cuda 11.6

description

Since I needed to study the usability of HDRNet for my research, I developed an HDRNet-like network. To develop efficiency, I used 1 * 1 convolutional layers instead of pixel-wise network and applying coefficient calculations, two 4 * 4 deconvolutional layers were used instead of bilateral grid upsampling.

train

dataset

Fivek dataset image. Import it into the folder dataset/train.

request

Images with a resolution of more than 1024 * 1024 -> dataset/train/full -> crop/flip/rotate -> input-full

Align with the upper left corner of Full Img -> dataset/train/gt -> crop/flip/rotate -> input-gt

data augmentation

Each image will be randomly cropped, rotated, and flipped 5 times in an epoch

then

python3 train.py 

output

epoch_xxx.tar

test

dataset

Crop the Fivek dataset image and import it into the folder dataset/test/img.

request

1024 * 1024 -> dataset/test/img

then

python3 test.py

output

1024 * 1024 -> dataset/test/output

code by

Wu F.Y

If

Modify the graphics resolution and network size : network :: FullNet & LowNet

Modify the channal : train & train

about

@article{gharbi2017deep,

    title={Deep bilateral learning for real-time image enhancement},

    author={Gharbi, Micha{\"e}l and Chen, Jiawen and Barron, Jonathan T and Hasinoff, Samuel W and Durand, Fr{\'e}do},

    journal={ACM Transactions on Graphics (TOG)},

    volume={36},

    number={4},

    pages={118},

    year={2017},

    publisher={ACM}
  
}