🔥🔥🔥 Multi-scale Shifted Dilated Convolutional Neural Network for Low-light Image Enhancement

• ⭐ - Arxiv (Coming Soon)
• ⭐ - DATESET: 👉 LOL 👉 FIVEK 👉 LSRW
• ⭐ - Pretrained Models: Mini_weight.pkl Max_weight.pkl

Introduction

MSDC-NET aims to become the deep learning model for low-light image enhancement with the smallest number of parameters and minimal floating-point operations in recent years. Both the model and its variants have parameter counts of ≤ 240 bytes. Experimental validation demonstrates that MSDC-NET achieves state-of-the-art performance compared to existing algorithms across multiple metrics, including SSIM, PSNR, parameters, and FLOPs

🏃 TODO

• Release project page
• Release max / min model weights on Github
• Release paper link
• Release inference code
• Release training code
• Release evaluation code

▶️ Getting Started

Installation

1. Clone MSDC-NET.

git clone --recursive https://github.com/chenyuhan1997/MSDC-NET
cd MSDC-NET
# git submodule update --init --recursive

2. Create the environment, here we show an example using conda.

conda create -n MSDC-NET python=3.11
conda activate MSDC-NET
conda install pytorch torchvision pytorch-cuda=12.1 -c pytorch -c nvidia  # use the correct version of cuda for your system
pip install opencv, kornia, pytorch_msssim, matplotlib, PIL, scikit-image, scipy, einops, math, typing

Train & Test

1. Train

python train.py

2. Test

python test.py

🌆 Results on Low-light Image Enhancement

⚠️ The current image is for demonstration purposes only and is not the final product.