Deep Structured Energy-Based Image Inpainting
Fazil Altinel, Mete Ozay, Takayuki Okatani - http://www.vision.is.tohoku.ac.jp/
If you make use of this code, please cite the following paper:
@INPROCEEDINGS{altinel2018dsebii,
author={F. Altinel and M. Ozay and T. Okatani},
booktitle={2018 24th International Conference on Pattern Recognition (ICPR)},
title={Deep Structured Energy-Based Image Inpainting},
year={2018},
volume={},
number={},
pages={423-428},
doi={10.1109/ICPR.2018.8546025},
ISSN={1051-4651},
month={Aug},}
Overview
This repository contains TensorFlow implementation of "Deep Structured Energy-Based Image Inpainting" paper (accepted to ICPR 2018).
- Network Architecture:
Input(x) -> CONV1(KernelSize=8, NumFilter= 32, Stride=4) -> CONV2(KernelSize=4, NumFilter= 64, Stride=2) -> CONV3(KernelSize=3, NumFilter= 64, Stride=1) -> FC1(512)
> Energy_x(y^)
Input(y^) -> CONV1(KernelSize=8, NumFilter= 32, Stride=4) -> CONV2(KernelSize=4, NumFilter= 64, Stride=2) -> CONV3(KernelSize=3, NumFilter= 64, Stride=1) -> FC1(512)
- Learning rates that used during training:
For energy update: Learning rate = 0.01, momentum = 0.9.
For parameter update: Learning rate = 0.001.
Files
files/
├── imgs/ - Test images folder
├── model/ - Model files folder
└── results/ - Test results folder
inpaint.py - Loads the model file and generates inpainted image(s) for given image(s).
utils.py - Various utilities for 'inpaint.py'
Dependencies
Tests are performed with following version of libraries:
- Python 3.4
- Numpy 1.11.3
- TensorFlow 1.0.1
Running
Download CelebA dataset (Align&Cropped Images): http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html.
Download the model file trained on CelebA dataset: https://drive.google.com/open?id=1ulpms4ni4ydJJ2NDm9YIl8zSliulFo7t. Extract and locate the files under files/model/.
Run the command below for all testing set of CelebA dataset:
$ python inpaint.py --allTest 1 --allImagesPath /path/to/all/dataset/folder/
Run the command below for testing images under files/imgs/:
$ python inpaint.py --allTest 0 --allImagesPath /path/to/all/dataset/folder/ --testImagesPath files/imgs/
Result images will be located under files/results/.
License
The source code is licensed under GNU General Public License v3.0.