danimlk/blockcnn

BlockCNN: A Deep Network for Artifact Removal and Image Compression

BlockCNN

BlockCNN: A Deep Network for Artifact Removal and Image Compression

This repository containing the implementation of BlockCNN which published in CVPR Workshop 2018. The implementation is in Pytorch. Original paper can be found in this link http://openaccess.thecvf.com/content_cvpr_2018_workshops/w50/html/Maleki_BlockCNN_A_Deep_CVPR_2018_paper.html

Requirments:

• Python 3.6
• Pytorch 0.3
• Torchvision
• Opencv
• Tensorflow 1.3
• Matplotlib

Instalation:

git clone https://github.com/DaniMlk/BlockCNN.git
cd BlockCNN
# [Option 1] To replicate the conda environment:
conda env create -f environment.yml
source activate pytorch
# [Option 2] Install everything globaly.

Using

We used Pascal_VOC 2012 to train our network. To run the code firstly you should put your dataset in the root path which mentioned in the main.py

python main.py

Results

In the below image you can see the output of our network, the input image and the original of it. The result of it shows an amazing fact that our network can enhance the quality of the image significantly.

In the below image you can compassion our output with other states of the art compression methods which shows that our network works in a better way.

TODO List:

For now we are working on this project to improve our results. we got promising results and we plan to publish our new algoirthm for ICCV 2019. You will shoke with our results 🔥

Citing

@InProceedings{Maleki_2018_CVPR_Workshops,
author = {Maleki, Danial and Nadalian, Soheila and Mahdi Derakhshani, Mohammad and Amin Sadeghi, Mohammad},
title = {BlockCNN: A Deep Network for Artifact Removal and Image Compression},
booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {June},
year = {2018}
}

blockcnn's Issues

how to prepare the train data ?

Is there any description on how to prepare the train data, thank you !

Artifact Removal code?

Hi Danial!

I am trying to reproduce Artifact Removal part of your paper :)

Looks like there are some code changes required, as the code in this repository focus on Compression part.

Will it be possible to point to what is changing?

Do you change the model? (at the moment it take 24x16 input and produce 12x12 output.
If I understand the paper correctly - for Artifact Removal you take 24x24 input and still produce 12x12?
     Or you only use first two block rows as in the Compression code?

Thank you.
nStyler

about train dataset

Hi, May i ask how do you create you train dataset, do you do jpg compression for ASCAL VOC 2007 ?

Can't Find Pre-Trained Model and Inference Code

Hi,

I check your README file, and your result looks great!

Can you provide your pre-trained model and give some hints about how to use your code for inference， so that I can try it by myself?

Best Regard,
Even.