# Denoising with GAN
[Paper](https://uofi.box.com/shared/static/s16nc93x8j6ctd0ercx9juf5mqmqx4bp.pdf) | [Video](https://www.youtube.com/watch?v=Yh_Bsoe-Qj4)
## Introduction
Animation movie companies like Pixar and Dreamworks render their 3d scenes using a technique called Pathtracing which enables them to create high quality photorealistic frames. Pathtracing involves shooting 1000’s of rays into a pixel randomly(Monte Carlo) which will then hit the objects in the scene and based on the reflective property of the object the rays reflect or refract or get absorbed. The colors returned by these rays are averaged to get the color of the pixel and this process is repeated for all the pixels. Due to the computational complexity it might take 8-16 hours to render a single frame.
We are proposing a neural network based solution for reducing 8-16 hours to a couple of seconds using a Generative Adversarial Network. The main idea behind this proposed method is to render using small number of samples per pixel (let say 4 spp or 8 spp instead of 32K spp) and pass the noisy image to our network, which will generate a photorealistic image with high quality.
# Demo Video [Link](https://www.youtube.com/watch?v=Yh_Bsoe-Qj4)
[![IMAGE ALT TEXT HERE](https://img.youtube.com/vi/Yh_Bsoe-Qj4/0.jpg)](https://www.youtube.com/watch?v=Yh_Bsoe-Qj4)
#### Table of Contents
* [Installation](#installation)
* [Running](#running)
* [Dataset](#dataset)
* [Hyperparameters](#hyperparameter)
* [Results](#results)
* [Improvements](#improvements)
* [Credits](#credits)
## Installation
To run the project you will need:
* python 3.5
* tensorflow (v1.1 or v1.0)
* PIL
* [CKPT FILE](https://uofi.box.com/shared/static/21a5jwdiqpnx24c50cyolwzwycnr3fwe.gz)
* [Dataset](https://uofi.box.com/shared/static/gy0t3vgwtlk1933xbtz1zvhlakkdac3n.zip)
## Running
Once you have all the depenedencies ready, do the folowing:
Download the dataset extract it to a folder named 'dataset' (ONLY if you want to train, not needed to run).
Extract the CKPT files to a folder named 'Checkpoints'
Run main.py -- python3 main.py
Go to the browser, if you are running it on a server then [ip-address]:8888, if you are on your local machine then localhost:8888
## Dataset
We picked random 40 images from pixar movies, added gaussian noise of different standard deviation, 5 sets of 5 different standard deviation making a total of 1000 images for the training set. For validation we used 10 images completely different from the training set and added gaussian noise. For testing we had both added gaussian images and real noisy images.
## Hyperparameters
* Number of iterations - 10K
* Adversarial Loss Factor - 0.5
* Pixel Loss Factor - 1.0
* Feature Loss Factor - 1.0
* Smoothness Loss Factor - 0.0001
## Results
3D rendering test data:
<img src="https://github.com/manumathewthomas/CS523Project3/blob/master/result1.PNG" alt="alt text" width="960" height="480">
Real noise images:
<img src="https://github.com/manumathewthomas/CS523Project3/blob/master/result2.png" alt="alt text" width="960" height="480">
CT-Scan:
<img src="https://github.com/manumathewthomas/CS523Project3/blob/master/result3.PNG" alt="alt text" width="960" height="480">
## Improvements
* Increase the num of iteration to 100K.
* Train the network for different noises.
* Make it work on a real-time app.
## Credits
* [SRGAN](https://arxiv.org/pdf/1609.04802.pdf)
* [Image De-raining using conditional generative adversarial network](https://arxiv.org/pdf/1701.05957.pdf)
* [Creating photorealistic images from gameboy camera](http://www.pinchofintelligence.com/photorealistic-neural-network-gameboy/)
* [CS20SI](cs20si.stanford.edu)
* [CS231n](https://cs231n.github.io/)
没有合适的资源?快使用搜索试试~ 我知道了~
生成对抗网络降噪算法
共32个文件
py:8个
pyc:4个
png:4个
4星 · 超过85%的资源 需积分: 50 182 下载量 175 浏览量
2018-09-07
16:30:38
上传
评论 19
收藏 19.51MB ZIP 举报
温馨提示
使用tensorflow框架写的生成对抗网络用于图像降噪,降噪效果在测试集上表现非常好,可以参看https://blog.csdn.net/xiaoxifei/article/details/82498705 记载的效果
资源推荐
资源详情
资源评论
收起资源包目录
ImageDenoisingGAN.zip (32个子文件)
ImageDenoisingGAN-master
paper.pdf 14.01MB
train.py 3KB
conv_helper.py 2KB
test.py 2KB
utils.py 6KB
main.py 614B
model.py 1KB
Project Proposal with summary of papers reviewed.pdf 951KB
result3.PNG 1.22MB
README.md 4KB
libs
utils.py 21KB
__pycache__
utils.cpython-35.pyc 21KB
vgg16.cpython-35.pyc 8KB
utils.cpython-34.pyc 21KB
vgg16.cpython-34.pyc 8KB
vgg16.py 9KB
result1.PNG 1.97MB
templates
index.html 5KB
static
output.png 66KB
js
jquery.js 94KB
bootstrap.js 68KB
bootstrap.min.js 36KB
fonts
glyphicons-halflings-regular.ttf 44KB
glyphicons-halflings-regular.woff2 18KB
glyphicons-halflings-regular.eot 20KB
glyphicons-halflings-regular.svg 106KB
glyphicons-halflings-regular.woff 23KB
css
bootstrap.min.css 118KB
4-col-portfolio.css 442B
bootstrap.css 143KB
style.css 26B
result2.png 1.32MB
共 32 条
- 1
资源评论
- 杰骜不逊_6662019-06-22谢谢谢谢谢谢
xiaoxifei
- 粉丝: 207
- 资源: 23
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功