SlideShare a Scribd company logo
1 of 15
Photo-Realistic Single Image Super -Resolution
Using a Generative Adversarial Network
2017 CVPR (oral)(arXiv 버전)
Christain Ledig 외 10명 트위터
발표자 : 정우진
한양대학교 컴퓨터 비전 및 패턴 인식 연구실
4x SRGAN (proposed) origianl
/ 20
• Introduction
– Related Work
– Contribution
• Proposed Method
– Network Architecture
– Loss Function
• Experiments
2
Contents
/ 20
기존 SR(Super-Resolution) 기법의 문제점
• MSE(Mean Squared Error) 사용
• PSNR 평가에 유리
• 흐릿한 영상
• 텍스쳐 손실
• 실사진과는 다름
해결책
• GAN(Generative Adversarial
Network) 사용
3
Introduction
MSE 기반 기법과 GAN 기
반 기법의 비교
/ 20
SRCNN
• Image Super-Resolution Using Deep Convolutional Networks
• (아마도)최초의 DNN기반 슈퍼 레졸루션(SR)
• 기존의 기계학습 방법을 능가
• 3층의 convolution layer
– SR의 전 과정이 가능함을 증명
4
Related Work
Introduction
SRCNN의 구조
/ 20
DRCN
• Deeply-Recursive Convolutional Network for Image Super-
Resolution
• FOV(Field of View)를 확장 (Receptive field와 같은 개념)
• 깊게 구성
– SRCNN : 3층, DRCN : 20층
• 반복구성
– 가중치 공유
– 훈련난이도
하락
5
Related Work
Introduction
DRCN의 구조
/ 20
• SRResNet
– ResNet 구조를 차용
– 좋은 PSNR, SSIM
– MSE loss
• SRGAN
– GAN 기반의 네트워크 학습
– 사진에 적합한 loss 기법 사용 : VGG loss
• MOS
– Mean Option Score
– 새로운 평가 방법
6
Contribution
Introduction
/ 20
• 제안하는 방법이 2개
– SRResNet
– ResNet 구조를 응용
– GAN 사용하지 않음
– SRGAN
– SRResNet에 추가 학습함
– GAN 사용
– 실사진과 같은 복원이 목표
7
Proposed Method
/ 20
Generator and Discriminator 구조
8
Network Architecture
Proposed Method
k : 커널크기
n : 채널크기
s : stride
/ 20
전체에러
• Generator에러 + Discriminator에러
Discriminator 에러
• 틀리면 페널티
9
Loss Function
Proposed Method
/ 20
Generator 에러 1 :MSE
• L2 놈
Generator 에러 2 : VGG특징 사용
• 사전학습 된 VGG사용
• 특징을 비교
• Feature of j-th convolution before the i-th maxpooling layer
10
Loss Function
Proposed Method
/ 20
객관적 지표
– SRResNet이 우수
11
Experiments
/ 20
실험결과
12
Experiments
/ 20
주관적 지표
• 화질을 평가하기 위
해 MOS도입
• Mean Opinion
Score
– 1점~5점 투표
• SRGAN이 우수
– 차이가 많이남
13
Experiments
/ 20
MSE loss vs. VGG loss
• VGG loss의 결과는 실사진과 유사
14
Experiments
/ 20
PSNR/time
15
Experiments

More Related Content

What's hot

[OSGeo-KR Tech Workshop] Deep Learning for Single Image Super-Resolution
[OSGeo-KR Tech Workshop] Deep Learning for Single Image Super-Resolution[OSGeo-KR Tech Workshop] Deep Learning for Single Image Super-Resolution
[OSGeo-KR Tech Workshop] Deep Learning for Single Image Super-ResolutionTaegyun Jeon
 
A Certain Slant of Light - Past, Present and Future Challenges of Global Illu...
A Certain Slant of Light - Past, Present and Future Challenges of Global Illu...A Certain Slant of Light - Past, Present and Future Challenges of Global Illu...
A Certain Slant of Light - Past, Present and Future Challenges of Global Illu...Electronic Arts / DICE
 
A Deep Journey into Super-resolution
A Deep Journey into Super-resolutionA Deep Journey into Super-resolution
A Deep Journey into Super-resolutionRonak Mehta
 
Object detection with deep learning
Object detection with deep learningObject detection with deep learning
Object detection with deep learningSushant Shrivastava
 
Deep Learning in Computer Vision
Deep Learning in Computer VisionDeep Learning in Computer Vision
Deep Learning in Computer VisionSungjoon Choi
 
Image classification on Imagenet (D1L4 2017 UPC Deep Learning for Computer Vi...
Image classification on Imagenet (D1L4 2017 UPC Deep Learning for Computer Vi...Image classification on Imagenet (D1L4 2017 UPC Deep Learning for Computer Vi...
Image classification on Imagenet (D1L4 2017 UPC Deep Learning for Computer Vi...Universitat Politècnica de Catalunya
 
Variational Autoencoders For Image Generation
Variational Autoencoders For Image GenerationVariational Autoencoders For Image Generation
Variational Autoencoders For Image GenerationJason Anderson
 
Image super resolution
Image super resolutionImage super resolution
Image super resolutionAkshay Hazare
 
Generative Adversarial Networks
Generative Adversarial NetworksGenerative Adversarial Networks
Generative Adversarial NetworksMark Chang
 
Generative Adversarial Networks
Generative Adversarial NetworksGenerative Adversarial Networks
Generative Adversarial NetworksMustafa Yagmur
 
Image segmentation with deep learning
Image segmentation with deep learningImage segmentation with deep learning
Image segmentation with deep learningAntonio Rueda-Toicen
 
Deep Learning: Application Landscape - March 2018
Deep Learning: Application Landscape - March 2018Deep Learning: Application Landscape - March 2018
Deep Learning: Application Landscape - March 2018Grigory Sapunov
 
State of transformers in Computer Vision
State of transformers in Computer VisionState of transformers in Computer Vision
State of transformers in Computer VisionDeep Kayal
 
Enhanced Deep Residual Networks for Single Image Super-Resolution
Enhanced Deep Residual Networks for Single Image Super-ResolutionEnhanced Deep Residual Networks for Single Image Super-Resolution
Enhanced Deep Residual Networks for Single Image Super-ResolutionNAVER Engineering
 
CVPR2019読み会 "A Theory of Fermat Paths for Non-Line-of-Sight Shape Reconstruc...
CVPR2019読み会 "A Theory of Fermat Paths  for Non-Line-of-Sight Shape Reconstruc...CVPR2019読み会 "A Theory of Fermat Paths  for Non-Line-of-Sight Shape Reconstruc...
CVPR2019読み会 "A Theory of Fermat Paths for Non-Line-of-Sight Shape Reconstruc...Hajime Mihara
 
NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis
NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis
NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis taeseon ryu
 

What's hot (20)

[OSGeo-KR Tech Workshop] Deep Learning for Single Image Super-Resolution
[OSGeo-KR Tech Workshop] Deep Learning for Single Image Super-Resolution[OSGeo-KR Tech Workshop] Deep Learning for Single Image Super-Resolution
[OSGeo-KR Tech Workshop] Deep Learning for Single Image Super-Resolution
 
AlexNet
AlexNetAlexNet
AlexNet
 
Super resolution
Super resolutionSuper resolution
Super resolution
 
Object detection
Object detectionObject detection
Object detection
 
A Certain Slant of Light - Past, Present and Future Challenges of Global Illu...
A Certain Slant of Light - Past, Present and Future Challenges of Global Illu...A Certain Slant of Light - Past, Present and Future Challenges of Global Illu...
A Certain Slant of Light - Past, Present and Future Challenges of Global Illu...
 
A Deep Journey into Super-resolution
A Deep Journey into Super-resolutionA Deep Journey into Super-resolution
A Deep Journey into Super-resolution
 
Object detection with deep learning
Object detection with deep learningObject detection with deep learning
Object detection with deep learning
 
Deep Learning in Computer Vision
Deep Learning in Computer VisionDeep Learning in Computer Vision
Deep Learning in Computer Vision
 
Image classification on Imagenet (D1L4 2017 UPC Deep Learning for Computer Vi...
Image classification on Imagenet (D1L4 2017 UPC Deep Learning for Computer Vi...Image classification on Imagenet (D1L4 2017 UPC Deep Learning for Computer Vi...
Image classification on Imagenet (D1L4 2017 UPC Deep Learning for Computer Vi...
 
Variational Autoencoders For Image Generation
Variational Autoencoders For Image GenerationVariational Autoencoders For Image Generation
Variational Autoencoders For Image Generation
 
Image super resolution
Image super resolutionImage super resolution
Image super resolution
 
Generative Adversarial Networks
Generative Adversarial NetworksGenerative Adversarial Networks
Generative Adversarial Networks
 
Generative Adversarial Networks
Generative Adversarial NetworksGenerative Adversarial Networks
Generative Adversarial Networks
 
Image segmentation with deep learning
Image segmentation with deep learningImage segmentation with deep learning
Image segmentation with deep learning
 
Deep Learning: Application Landscape - March 2018
Deep Learning: Application Landscape - March 2018Deep Learning: Application Landscape - March 2018
Deep Learning: Application Landscape - March 2018
 
State of transformers in Computer Vision
State of transformers in Computer VisionState of transformers in Computer Vision
State of transformers in Computer Vision
 
Enhanced Deep Residual Networks for Single Image Super-Resolution
Enhanced Deep Residual Networks for Single Image Super-ResolutionEnhanced Deep Residual Networks for Single Image Super-Resolution
Enhanced Deep Residual Networks for Single Image Super-Resolution
 
CVPR2019読み会 "A Theory of Fermat Paths for Non-Line-of-Sight Shape Reconstruc...
CVPR2019読み会 "A Theory of Fermat Paths  for Non-Line-of-Sight Shape Reconstruc...CVPR2019読み会 "A Theory of Fermat Paths  for Non-Line-of-Sight Shape Reconstruc...
CVPR2019読み会 "A Theory of Fermat Paths for Non-Line-of-Sight Shape Reconstruc...
 
Deep Learning for Computer Vision: Data Augmentation (UPC 2016)
Deep Learning for Computer Vision: Data Augmentation (UPC 2016)Deep Learning for Computer Vision: Data Augmentation (UPC 2016)
Deep Learning for Computer Vision: Data Augmentation (UPC 2016)
 
NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis
NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis
NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis
 

Similar to Review SRGAN

[paper review] 손규빈 - Eye in the sky & 3D human pose estimation in video with ...
[paper review] 손규빈 - Eye in the sky & 3D human pose estimation in video with ...[paper review] 손규빈 - Eye in the sky & 3D human pose estimation in video with ...
[paper review] 손규빈 - Eye in the sky & 3D human pose estimation in video with ...Gyubin Son
 
[PR12] image super resolution using deep convolutional networks
[PR12] image super resolution using deep convolutional networks[PR12] image super resolution using deep convolutional networks
[PR12] image super resolution using deep convolutional networksTaegyun Jeon
 
Deep learning super resolution
Deep learning super resolutionDeep learning super resolution
Deep learning super resolutionNAVER Engineering
 
Simple Review of Single Image Super Resolution Task
Simple Review of Single Image Super Resolution TaskSimple Review of Single Image Super Resolution Task
Simple Review of Single Image Super Resolution TaskMYEONGGYU LEE
 
Photo realistic single image super-resolution using a generative adversarial ...
Photo realistic single image super-resolution using a generative adversarial ...Photo realistic single image super-resolution using a generative adversarial ...
Photo realistic single image super-resolution using a generative adversarial ...soul8085
 
Deep Learning & Convolutional Neural Network
Deep Learning & Convolutional Neural NetworkDeep Learning & Convolutional Neural Network
Deep Learning & Convolutional Neural Networkagdatalab
 
"From image level to pixel-level labeling with convolutional networks" Paper ...
"From image level to pixel-level labeling with convolutional networks" Paper ..."From image level to pixel-level labeling with convolutional networks" Paper ...
"From image level to pixel-level labeling with convolutional networks" Paper ...LEE HOSEONG
 
Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
Faster R-CNN: Towards Real-Time Object Detection with Region Proposal NetworksFaster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
Faster R-CNN: Towards Real-Time Object Detection with Region Proposal NetworksOh Yoojin
 
[Paper] shuffle net an extremely efficient convolutional neural network for ...
[Paper] shuffle net  an extremely efficient convolutional neural network for ...[Paper] shuffle net  an extremely efficient convolutional neural network for ...
[Paper] shuffle net an extremely efficient convolutional neural network for ...Susang Kim
 
"Dataset and metrics for predicting local visible differences" Paper Review
"Dataset and metrics for predicting local visible differences" Paper Review"Dataset and metrics for predicting local visible differences" Paper Review
"Dataset and metrics for predicting local visible differences" Paper ReviewLEE HOSEONG
 
[Pix2 pix] image to-image translation with conditional adversarial network re...
[Pix2 pix] image to-image translation with conditional adversarial network re...[Pix2 pix] image to-image translation with conditional adversarial network re...
[Pix2 pix] image to-image translation with conditional adversarial network re...JaeYeongKo
 
247 deview 2013 이미지 분석 - 민재식
247 deview 2013 이미지 분석 - 민재식247 deview 2013 이미지 분석 - 민재식
247 deview 2013 이미지 분석 - 민재식NAVER D2
 
Lab_Study_0615.pptx
Lab_Study_0615.pptxLab_Study_0615.pptx
Lab_Study_0615.pptxssuser0e717a
 
Deep neural networks cnn rnn_ae_some practical techniques
Deep neural networks cnn rnn_ae_some practical techniquesDeep neural networks cnn rnn_ae_some practical techniques
Deep neural networks cnn rnn_ae_some practical techniquesKang Pilsung
 
Faster R-CNN
Faster R-CNNFaster R-CNN
Faster R-CNNrlawjdgns
 
DeepAR:Probabilistic Forecasting with Autogressive Recurrent Networks
DeepAR:Probabilistic Forecasting with Autogressive Recurrent Networks DeepAR:Probabilistic Forecasting with Autogressive Recurrent Networks
DeepAR:Probabilistic Forecasting with Autogressive Recurrent Networks pko89403
 
Unsupervised learning for real-world super-resolution review presentation
Unsupervised learning for real-world super-resolution review presentationUnsupervised learning for real-world super-resolution review presentation
Unsupervised learning for real-world super-resolution review presentationSeoung-Ho Choi
 

Similar to Review SRGAN (19)

[paper review] 손규빈 - Eye in the sky & 3D human pose estimation in video with ...
[paper review] 손규빈 - Eye in the sky & 3D human pose estimation in video with ...[paper review] 손규빈 - Eye in the sky & 3D human pose estimation in video with ...
[paper review] 손규빈 - Eye in the sky & 3D human pose estimation in video with ...
 
[PR12] image super resolution using deep convolutional networks
[PR12] image super resolution using deep convolutional networks[PR12] image super resolution using deep convolutional networks
[PR12] image super resolution using deep convolutional networks
 
Review EDSR
Review EDSRReview EDSR
Review EDSR
 
Deep learning super resolution
Deep learning super resolutionDeep learning super resolution
Deep learning super resolution
 
Simple Review of Single Image Super Resolution Task
Simple Review of Single Image Super Resolution TaskSimple Review of Single Image Super Resolution Task
Simple Review of Single Image Super Resolution Task
 
Photo realistic single image super-resolution using a generative adversarial ...
Photo realistic single image super-resolution using a generative adversarial ...Photo realistic single image super-resolution using a generative adversarial ...
Photo realistic single image super-resolution using a generative adversarial ...
 
Deep Learning & Convolutional Neural Network
Deep Learning & Convolutional Neural NetworkDeep Learning & Convolutional Neural Network
Deep Learning & Convolutional Neural Network
 
"From image level to pixel-level labeling with convolutional networks" Paper ...
"From image level to pixel-level labeling with convolutional networks" Paper ..."From image level to pixel-level labeling with convolutional networks" Paper ...
"From image level to pixel-level labeling with convolutional networks" Paper ...
 
Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
Faster R-CNN: Towards Real-Time Object Detection with Region Proposal NetworksFaster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
 
Detecting fake jpeg images
Detecting fake jpeg imagesDetecting fake jpeg images
Detecting fake jpeg images
 
[Paper] shuffle net an extremely efficient convolutional neural network for ...
[Paper] shuffle net  an extremely efficient convolutional neural network for ...[Paper] shuffle net  an extremely efficient convolutional neural network for ...
[Paper] shuffle net an extremely efficient convolutional neural network for ...
 
"Dataset and metrics for predicting local visible differences" Paper Review
"Dataset and metrics for predicting local visible differences" Paper Review"Dataset and metrics for predicting local visible differences" Paper Review
"Dataset and metrics for predicting local visible differences" Paper Review
 
[Pix2 pix] image to-image translation with conditional adversarial network re...
[Pix2 pix] image to-image translation with conditional adversarial network re...[Pix2 pix] image to-image translation with conditional adversarial network re...
[Pix2 pix] image to-image translation with conditional adversarial network re...
 
247 deview 2013 이미지 분석 - 민재식
247 deview 2013 이미지 분석 - 민재식247 deview 2013 이미지 분석 - 민재식
247 deview 2013 이미지 분석 - 민재식
 
Lab_Study_0615.pptx
Lab_Study_0615.pptxLab_Study_0615.pptx
Lab_Study_0615.pptx
 
Deep neural networks cnn rnn_ae_some practical techniques
Deep neural networks cnn rnn_ae_some practical techniquesDeep neural networks cnn rnn_ae_some practical techniques
Deep neural networks cnn rnn_ae_some practical techniques
 
Faster R-CNN
Faster R-CNNFaster R-CNN
Faster R-CNN
 
DeepAR:Probabilistic Forecasting with Autogressive Recurrent Networks
DeepAR:Probabilistic Forecasting with Autogressive Recurrent Networks DeepAR:Probabilistic Forecasting with Autogressive Recurrent Networks
DeepAR:Probabilistic Forecasting with Autogressive Recurrent Networks
 
Unsupervised learning for real-world super-resolution review presentation
Unsupervised learning for real-world super-resolution review presentationUnsupervised learning for real-world super-resolution review presentation
Unsupervised learning for real-world super-resolution review presentation
 

More from Woojin Jeong

Super resolution-review
Super resolution-reviewSuper resolution-review
Super resolution-reviewWoojin Jeong
 
2017 on calibration of modern neural networks
2017  on calibration of modern neural networks2017  on calibration of modern neural networks
2017 on calibration of modern neural networksWoojin Jeong
 
Review Wide Resnet
Review Wide ResnetReview Wide Resnet
Review Wide ResnetWoojin Jeong
 

More from Woojin Jeong (6)

Review MLP Mixer
Review MLP MixerReview MLP Mixer
Review MLP Mixer
 
Super resolution-review
Super resolution-reviewSuper resolution-review
Super resolution-review
 
2017 on calibration of modern neural networks
2017  on calibration of modern neural networks2017  on calibration of modern neural networks
2017 on calibration of modern neural networks
 
Review Wide Resnet
Review Wide ResnetReview Wide Resnet
Review Wide Resnet
 
Review Dense net
Review Dense netReview Dense net
Review Dense net
 
Review DRCN
Review DRCNReview DRCN
Review DRCN
 

Review SRGAN

  • 1. Photo-Realistic Single Image Super -Resolution Using a Generative Adversarial Network 2017 CVPR (oral)(arXiv 버전) Christain Ledig 외 10명 트위터 발표자 : 정우진 한양대학교 컴퓨터 비전 및 패턴 인식 연구실 4x SRGAN (proposed) origianl
  • 2. / 20 • Introduction – Related Work – Contribution • Proposed Method – Network Architecture – Loss Function • Experiments 2 Contents
  • 3. / 20 기존 SR(Super-Resolution) 기법의 문제점 • MSE(Mean Squared Error) 사용 • PSNR 평가에 유리 • 흐릿한 영상 • 텍스쳐 손실 • 실사진과는 다름 해결책 • GAN(Generative Adversarial Network) 사용 3 Introduction MSE 기반 기법과 GAN 기 반 기법의 비교
  • 4. / 20 SRCNN • Image Super-Resolution Using Deep Convolutional Networks • (아마도)최초의 DNN기반 슈퍼 레졸루션(SR) • 기존의 기계학습 방법을 능가 • 3층의 convolution layer – SR의 전 과정이 가능함을 증명 4 Related Work Introduction SRCNN의 구조
  • 5. / 20 DRCN • Deeply-Recursive Convolutional Network for Image Super- Resolution • FOV(Field of View)를 확장 (Receptive field와 같은 개념) • 깊게 구성 – SRCNN : 3층, DRCN : 20층 • 반복구성 – 가중치 공유 – 훈련난이도 하락 5 Related Work Introduction DRCN의 구조
  • 6. / 20 • SRResNet – ResNet 구조를 차용 – 좋은 PSNR, SSIM – MSE loss • SRGAN – GAN 기반의 네트워크 학습 – 사진에 적합한 loss 기법 사용 : VGG loss • MOS – Mean Option Score – 새로운 평가 방법 6 Contribution Introduction
  • 7. / 20 • 제안하는 방법이 2개 – SRResNet – ResNet 구조를 응용 – GAN 사용하지 않음 – SRGAN – SRResNet에 추가 학습함 – GAN 사용 – 실사진과 같은 복원이 목표 7 Proposed Method
  • 8. / 20 Generator and Discriminator 구조 8 Network Architecture Proposed Method k : 커널크기 n : 채널크기 s : stride
  • 9. / 20 전체에러 • Generator에러 + Discriminator에러 Discriminator 에러 • 틀리면 페널티 9 Loss Function Proposed Method
  • 10. / 20 Generator 에러 1 :MSE • L2 놈 Generator 에러 2 : VGG특징 사용 • 사전학습 된 VGG사용 • 특징을 비교 • Feature of j-th convolution before the i-th maxpooling layer 10 Loss Function Proposed Method
  • 11. / 20 객관적 지표 – SRResNet이 우수 11 Experiments
  • 13. / 20 주관적 지표 • 화질을 평가하기 위 해 MOS도입 • Mean Opinion Score – 1점~5점 투표 • SRGAN이 우수 – 차이가 많이남 13 Experiments
  • 14. / 20 MSE loss vs. VGG loss • VGG loss의 결과는 실사진과 유사 14 Experiments

Editor's Notes

  1. 의문이 있다. Discriminator는 맞추면 보상을 틀리면 페널티를 받아야 되는데, 페널티만 있다.