Multi-scale Network for debluring

Prerequisite

Dowload data set from the link: : https://www.dropbox.com/s/u842yorwmap7xij/GOPRO_Large.zip?dl=0
Create data folder:

mkdir data

Unzip GoPro dataset to data folder such that we have:

• data/train : train data
• data/test : test data

Set up environment:

conda create -n deblur python=3.6
conda activate deblur
pip install -r requirement.txt

How to train

Train the network by run corresponding command below:

One scale

./one_scale_no_lsc.sh

One scale with long skip connection

./one_scale__lsc.sh

Multi scale

./multi_scale_no_lsc.sh

Multi scale with long skip connection

./multi_scale_with_lsc.sh

How to test

I provide pretrained model at url: https://drive.google.com/file/d/1OrtRLABEVb-nLHf39CamDKp4ayrxIDi9/view?usp=sharing

upzip the pretrained model to src folder such that we have these folders:

• one_scale1
• one_scale_lsc1
• multi_scale1
• multi_scale_lsc1
• multi_scale_lsc1000

Run test:

./test_model.sh

All the result will be store in val folder

In case that you want to test your model, read the test_model.sh and modify the pretrained_model path.

PSNR, SSIM, MS-SSIM

I used SKIMAGE library for calculate PSNR and SSIM

from skimage.measure import compare_ssim as ssim
from skimage.measure import compare_psnr as psnr

For MS-SSIM, I used Tensorflow code which is available at: https://github.com/tensorflow/models/blob/master/research/compression/image_encoder/msssim.py The code is hard copy to utils.py, so we don't need to worry about the dependency.