trellixvulnteam/psd-aecr-net_26ct

My NAPI 2022 Project! Enjoy!

PSD-AECR-NET

Welcome! This repository contains my output for my NAPI 2022 internship where I combined PSD & AECR-Net for dehazing foggy images. See the files in the docs folder for details. Have fun experimenting!

AECR-Net

Contrastive Learning for Compact Single Image Dehazing, CVPR2021. Official Pytorch based implementation.
arxiv

Please share them some love on their GitHub

Principled S2R Dehazing

PSD: Principled Synthetic to Real Dehazing Guided by Physical Priors
Zeyuan Chen, Yangchao Wang, Yang Yang, Dong Liu
CVPR 2021 (Oral)

Please share them some love on their GitHub

Environment

• Python 3.9.9
• Pytorch 1.10.1+cu113

Downloads

Below are download links to the datasets, image outputs, pre-trained models respectively since they can't fit in this repository. Simply paste the files into their respective folders and it should work as intended.

Folder Name	File size	Download
Images	20GB	Google Drive
Outputs	812MB	Google Drive
Pre-trained Models	567MB	Google Drive

Notebooks

Notebook	Description
AECRNet_train.ipynb	Contains training code for AECR-Net using OTS dataset
PSD CVPR Testbench.ipynb	Compares all of the model's image results using CVPR dataset
PSD SOTS Testbench.ipynb	Compares all of the model's image results using SOTS dataset
PSD Validate.ipynb	Validates all of the models using SOTS dataset
UNET_train.ipynb	Contains training code for U-Net using OTS dataset
napi_student_2021.ipynb	Initial practice exercise containing cropping & resizing code of OTS dataset

Testing

Runs the CVPR dataset to the PSD AECR-Net model and saves the output images to output/AERCNET/

python test_aecr.py

Runs the CVPR dataset to the PSD FFA-Net model and saves the output images to output/FFA/

python test_ffa.py

Runs the CVPR dataset to the PSD GCA-Net model and saves the output images to output/GCA/

python test_gca.py

Runs the CVPR dataset to the PSD MSBDN-Net model and saves the output images to output/MSBDN/

python test_msbdn.py

Trains the PSD AECR-Net model to with the OTS dataset and outputs the trained model at output/

python train_aecr.py

• To run tests, I highly recommend running command prompt as administror since some modules won't work without admin privileges. You can change settings such as the number of images to output or training values by editing the code.