SlideShare a Scribd company logo
CGLAB 이명규Simple Review of Super Resolution Task (1/34) Presented by CGLAB 이명규
2020/05/27
Simple Review of
Super Resolution Task
CGLAB 이명규Simple Review of Super Resolution Task (2/34)
I N D E X
01
02
Introduction
Featured Papers
CGLAB 이명규Simple Review of Super Resolution Task (3/34)
Introduction
Part 01
1. Super Resolution 소개
2. 관련 연구 소개
CGLAB 이명규Simple Review of Super Resolution Task (4/34)
↳
Super Resolution 소개1-1
• LR(Low-resolution) image 또는 video를
HR(High-resolution)으로 복구하는 문제
• SISR(Single Image SR)과 MISR(Multiple Image SR)로 구분
Problem Definition
𝒚 𝑳𝑹 = 𝒙⨂𝒌 𝒔 + 𝒏.
GT HR image
Blur Kernel
Down
Sampling Noise
CGLAB 이명규Simple Review of Super Resolution Task (5/34)
↳
Super Resolution 소개1-1
• SISR Algorithms
• Interpolation based method
(Bicubic, bilinear, Nearest neighbor etc….)
=> Just “upscaling” image
• Reconstruction based method
• Deep Learning based method
Problem Definition
https://bskyvision.com/531
CGLAB 이명규Simple Review of Super Resolution Task (6/34)
↳
Super Resolution 소개1-1
Applications
Super Resolution Applications in Modern Digital Image Processing (IJCA 2016)
▲ SR for Satellite Image ▲ SR for Medical Imaging
▲ SR for Astrological Studies ▲ SR for Microscopy
Image Processing
CGLAB 이명규Simple Review of Super Resolution Task (7/34)
↳
Related Works1-2
CGLAB 이명규Simple Review of Super Resolution Task (8/34)
↳
Related Works1-2
Deep Learning for Image Super-resolution: A Survey
CGLAB 이명규Simple Review of Super Resolution Task (9/34)
↳
Related Works1-2
• DL을 SSIR에 사용한 첫 논문
→ 3-layer CNN, MSE loss
• 전통적인 방식 대비 월등한 성능
SRCNN (ECCV 2014)
Image Super-Resolution Using Deep Convolutional Networks (ECCV 2014)
CGLAB 이명규Simple Review of Super Resolution Task (10/34)
↳
Related Works1-2
• Too Shallow했던 SRCNN의 단점 개선,
안정적이면서 빠른 학습
→ VGGNet based deep residual learning+MSE
“Adjustable gradient clipping for maximal boost
in speed while suppressing exploding gradients”
• 단순 LR→HR mapping 학습이 아닌,
bicubic upsampled LR 이미지와
HR 이미지의 잔차(residual) 혹은 noise를
학습하는 방식
VDSR (CVPR 2016)
Image Super-Resolution Using Deep Convolutional Networks (ECCV 2014)
이후 DRCN(Deeply-recursive CNN), SRResNet,
DRRN의 제안으로 이어짐
CGLAB 이명규Simple Review of Super Resolution Task (11/34)
↳
Related Works1-2
• GAN을 활용해 인간이 보기에 그럴싸한 이미지를 생성하려는 시도
• MSE loss는 blurry한 이미지를 생성하므로, Perceptual loss 사용을 제안
→ 𝑷𝒆𝒓𝒄𝒆𝒑𝒕𝒖𝒂𝒍 𝒍𝒐𝒔𝒔 = 𝑮𝑨𝑵 𝒍𝒐𝒔𝒔 + 𝑪𝒐𝒏𝒕𝒆𝒏𝒕 𝒍𝒐𝒔𝒔
MSE loss 대신 style transfer에서 사용되는 VGG loss로 교체
SRGAN (CVPR 2017)
Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network (CVPR 2017)
CGLAB 이명규Simple Review of Super Resolution Task (12/34)
↳
Related Works1-2
SRGAN (CVPR 2017)
Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network (CVPR 2017)
CGLAB 이명규Simple Review of Super Resolution Task (13/34)
Featured Papers
Part 02
1. Paper 1 (“EDSR”)
(Enhanced Deep Residual Networks for
Single Image Super-Resolution)
2. Paper 2 (“SAN”, SOTA)
(Second-order Attention Network for
Single Image Super-Resolution)
CGLAB 이명규Simple Review of Super Resolution Task (14/34)
Paper 1:
Enhanced Deep Residual Networks for
Single Image Super-Resolution(“EDSR”)
Part 2-1
1. Introduction
2. Architecture Overview
3. Experiment & Conclusion
CGLAB 이명규Simple Review of Super Resolution Task (15/34)
↳
Introduction2-1-1
Limitations of Previous Works
• ResNet을 그대로 사용 시(SRResnet)SISR에
좋은 성능을 보여 주지만 다음 이슈가 존재
• ResNet과 같은 분류기들은 classification과 같은 high-level 문제를 위한 구조
(SR은 low-level task)
• ResNet에 적용된 BN으로 인해 네트워크의 flexibility 저하
→ 학습 시간이 오래 걸리는 문제
▲ Batch Normalization
Enhanced Deep Residual Networks for Single Image Super-Resolution (CVPR 2017)
CGLAB 이명규Simple Review of Super Resolution Task (16/34)
↳
Introduction2-1-1
Contributions
• 학습 시 메모리 사용량 40% 절감
→ BN layer를 제거한 new residual block 제안
→ 더 깊은 네트워크를 학습할 수 있게 됨
• Single Scale Model(EDSR)과 Multi Scale Model(MDSR) 제안
→ 각각 single scale(x2, x3, x4)을 따로따로 학습(EDSR)하거나 동시에 여러 scale에 대해
학습(MDSR)
Enhanced Deep Residual Networks for Single Image Super-Resolution (CVPR 2017)
CGLAB 이명규Simple Review of Super Resolution Task (17/34)
↳
Architecture Overview2-1-2
Model Overview
▲ EDSR ▲ MDSR
Enhanced Deep Residual Networks for Single Image Super-Resolution (CVPR 2017)
CGLAB 이명규Simple Review of Super Resolution Task (18/34)
↳
Architecture Overview2-1-2
Model Overview
• BN이 빠진 새 residual block 구조 제안
• Residual block 최종 feature matrix에
0.1의 constant를 곱해 줌
→ 안정적인 학습 도모
• x3, x4배 scaling 학습에서는 x2 모델을
기반으로 transfer learning
Enhanced Deep Residual Networks for Single Image Super-Resolution (CVPR 2017)
CGLAB 이명규Simple Review of Super Resolution Task (19/34)
↳
Evaluation results
2-1-3 Experiment & Conclusion
Enhanced Deep Residual Networks for Single Image Super-Resolution (CVPR 2017)
CGLAB 이명규Simple Review of Super Resolution Task (20/34)
↳
• Conclusion
• 종전 방법 대비 학습 시 메모리를 40%까지 줄이는 방법을 제안
• Single Scale(EDSR)과 Multi Scale(MDSR)에서 작동하는
SR 모델의 제안
• Limitations
• CNN의 receptive field size가 상대적으로 작음. (즉, local patch에만 집중)
→ 이미지에서 보다 넓은 영역을 고려하지 못함.
• Feature들이 담고 있는 local 혹은 global 정보가 동등하게 처리됨
→ 이후 Dilated conv, spatial or channel-wise attention 등의 제안으로 이어짐
Conclusion & Limitations
2-1-3 Experiment & Conclusion
Enhanced Deep Residual Networks for Single Image Super-Resolution (CVPR 2017)
CGLAB 이명규Simple Review of Super Resolution Task (21/34)
Paper 2:
Second-order Attention Network for
Single Image Super-Resolution(“SAN”)
Part 2-2
1. Introduction
2. Architecture Overview
3. Experiment & Conclusion
CGLAB 이명규Simple Review of Super Resolution Task (22/34)
↳
Introduction2-2-1
Limitations of Previous Works
• 기존 모델들은 깊거나 넓은 구조 디자인에만 초점을 맞춤
→ 따라서 레이어 간의 관계를 탐색하지 않아 네트워크 전체의 표현력 저하
• 대부분 LR 이미지의 모든 정보를 사용하지 않아 낮은 성능을 보여 옴
• 학습 속도가 본 논문 대비 상대적으로 느림
Enhanced Deep Residual Networks for Single Image Super-Resolution (CVPR 2017)
CGLAB 이명규Simple Review of Super Resolution Task (23/34)
↳
Introduction2-2-1
Contributions
• Second order statistics를 활용해 레이어 간 feature 상호 의존성 학습
• LSRAG(local-source residual attention groups) 구조를 제안해 LR 이미지
정보를 적극 활용
→ 풍부한 low-frequency 정보
CGLAB 이명규Simple Review of Super Resolution Task (24/34)
↳
Architecture Overview2-2-2
Model Overview
CGLAB 이명규Simple Review of Super Resolution Task (25/34)
↳
Architecture Overview2-2-2
Detailed view of Model:
RL-NL Module
• 목적 : 영상을 잘게 쪼개어 영상 전체를 보기 위함
→ 영상을 4등분 후 각 영역에 대해 non-local module 적용 (큰 해상도에서도 유리)
• SSRG모듈 전후로 수행되며, high-level에서 넓은 범위의 정보들을 모으는 역할
→ Global level non-local 연산은 인풋 사이즈가 클 경우 연산량 증가
(Region-Level Non-Local)
CGLAB 이명규Simple Review of Super Resolution Task (26/34)
↳
Architecture Overview2-2-2
Detailed view of Model:
LSRAG Module
• 목적 : Feature inter-dependencies를 잘 유지하기 위함
→ 영상을 4등분 후 각 영역에 대해 non-local module 적용 (큰 해상도에서도 유리)
• Simplified residual block들로 구성
(Local Source
Residual Attention Group)
CGLAB 이명규Simple Review of Super Resolution Task (27/34)
↳
Architecture Overview2-2-2
Detailed view of Model:
SOCA Module
• 목적 : 공분산 정규화를 통한 모델의 Discriminative representation 능력 향상
→ Attention을 통해 네트워크가 더욱 중요한 정보를 갖는 feature에 가중치를 더 두도록 함
• 기존 Attention 알고리즘들은 GAP을 이용한 1차 statistics 정보만 활용
→ 기존 방식들은 1st order statistics(=average) 이상의 정보를 활용하지 않아
모델의 Discriminative representation 능력이 저하됨
→ 따라서 Covariance Normalization을 거쳐 channel attention 수행
(Second-Order
Channel Attention)
CGLAB 이명규Simple Review of Super Resolution Task (28/34)
↳
Architecture Overview2-2-2
Detailed view of Model:
SOCA Module
• GAP(Global Average Pooling) :
각 feature map의 노드들로부터
단순히 평균을 추출해 차원 축소
e.g.)
(𝟏+𝟗+𝟔+𝟒+𝟓+𝟒+𝟕+𝟖+𝟓+𝟏+𝟐+𝟗+𝟔+𝟕+𝟔+𝟎)
𝟏𝟔
= 𝟓
Replace to GCP(Global Covariance Pooling)
CGLAB 이명규Simple Review of Super Resolution Task (29/34)
↳
Architecture Overview2-2-2
Detailed view of Model:
SOCA Module
• Covariance Normalization:
1. Reshape 𝑯 × 𝑾 × 𝑪 feature map 𝑭 = [𝐟 𝟏, … , 𝐟𝒄] to feature matrix 𝑿.
(𝑿 has 𝒔 = 𝑾𝑯 features of 𝑪-dim)
2. Compute sample covariance matrix 𝚺.
𝚺 = 𝐗 ҧ𝐈𝐗 𝑻
, where ҧ𝐈 =
𝟏
𝒔
(𝐈 −
𝟏
𝒔
𝟏). (𝑰=𝒔 × 𝒔 Identity matrix, 𝟏=matrix of all ones)
CGLAB 이명규Simple Review of Super Resolution Task (30/34)
↳
Architecture Overview2-2-2
Detailed view of Model:
SOCA Module
3. 𝚺 is symmetric positive semi-definte, thus has EIG as follows:
𝚺 = 𝐔𝚲𝐔 𝐓, 𝒘𝒉𝒆𝒓𝒆 𝑼 = 𝒐𝒓𝒕𝒉𝒐𝒈𝒐𝒏𝒂𝒍 𝒎𝒂𝒕𝒓𝒊𝒙, 𝜦 = 𝒅𝒊𝒂𝒈 𝒎𝒂𝒕𝒓𝒊𝒙 𝒘𝒊𝒕𝒉 𝒆𝒊𝒈𝒆𝒏𝒗𝒂𝒍𝒖𝒆𝒔.
4. So, covariance normalization can be coverted to power of 𝒆𝒊𝒈𝒆𝒏𝒗𝒂𝒍𝒖𝒆𝒔.
෡𝐘 = 𝚺 𝜶 = 𝐔𝚲 𝜶 𝐔 𝑻
- 𝜶는 양의 실수로, 1일 경우 정규화를 수행하지 않음. ( 𝛼 =½에서 잘 작동함을 확인)
- 𝜶<1인 경우 1.0보다 큰 eigenvalue는 non-linear하게 축소하고, 반대는 늘림.
CGLAB 이명규Simple Review of Super Resolution Task (31/34)
↳
Architecture Overview2-2-2
Detailed view of Model:
SOCA Module
• Normalized covariance matrix ෠𝐘는 channel-wise feature들 간의
상관 관계를 characterize하는 역할
• Covariance normalized ෡𝐘 를 이용해 channel level에서 pooling을 수행
𝒍𝒆𝒕 ෠𝐘 = 𝒚 𝟏, … , 𝒚 𝑪 , 𝒄𝒉𝒂𝒏𝒏𝒆𝒍𝒘𝒊𝒔𝒆 𝒔𝒕𝒂𝒕𝒊𝒔𝒕𝒊𝒄𝒔 𝒛 = 𝑯 𝑮𝑪𝑷 𝒚 𝒄 =
𝟏
𝑪
෍
𝒊
𝑪
𝒚 𝑪(𝒊)
CGLAB 이명규Simple Review of Super Resolution Task (32/34)
↳
Evaluation results
2-2-3 Experiment & Conclusion
CGLAB 이명규Simple Review of Super Resolution Task (33/34)
↳
• 다양한 모듈을 제안해 SISR의 PSNR 성능을 높임
• SSRG 모듈을 통해 low-frequency 정보들을 충분히 활용
• RL-NL 모듈을 통해 long-distance spatial contextual information을 활용
• SOCA 모듈을 통해 global covariance pooling을 수행하고, 레이어 간의
dependencies를 학습
• Second-order channel attention을 통해 discriminative representation학습에 집중
• 네트워크 규모 대비 낮은 수의 파라미터 개수
Conclusion & Limitations
2-2-3 Experiment & Conclusion
CGLAB 이명규Simple Review of Super Resolution Task (34/34)
Thank you for Listening.
Email : brstar96@naver.com (or brstar96@soongsil.ac.kr)
Mobile : +82-10-8234-3179

More Related Content

PDF
007 20151214 Deep Unsupervised Learning using Nonequlibrium Thermodynamics
PDF
004 20151116 deep_unsupervisedlearningusingnonequlibriumthermodynamics
PDF
Performance Analysis and Troubleshooting Methodologies for Databases
PDF
How Safe is Asynchronous Master-Master Setup?
PPTX
Auditoria de código de aplicaciones
PPTX
Diffusion models beat gans on image synthesis
PDF
MySQL InnoDB Cluster - A complete High Availability solution for MySQL
PDF
Derivation of the gradient descent rule.
007 20151214 Deep Unsupervised Learning using Nonequlibrium Thermodynamics
004 20151116 deep_unsupervisedlearningusingnonequlibriumthermodynamics
Performance Analysis and Troubleshooting Methodologies for Databases
How Safe is Asynchronous Master-Master Setup?
Auditoria de código de aplicaciones
Diffusion models beat gans on image synthesis
MySQL InnoDB Cluster - A complete High Availability solution for MySQL
Derivation of the gradient descent rule.

What's hot (12)

PDF
Databricks: A Tool That Empowers You To Do More With Data
PDF
MySQL Shell for DBAs
PDF
Büyük veri teknolojilerine giriş v1l
PDF
Deep Learning for Computer Vision: ImageNet Challenge (UPC 2016)
PPTX
Introduction to Deep learning
PDF
Technical Deep Dive: Using Apache Kafka to Optimize Real-Time Analytics in Fi...
PDF
OpenGLES Android Graphics
PDF
Semantic Segmentation AIML Project
PPTX
Prendre la data par le bon sens
PPT
적외선 치료
PDF
SSII2021 [SS2] Deepfake Generation and Detection – An Overview (ディープフェイクの生成と検出)
PPTX
bigquery.pptx
Databricks: A Tool That Empowers You To Do More With Data
MySQL Shell for DBAs
Büyük veri teknolojilerine giriş v1l
Deep Learning for Computer Vision: ImageNet Challenge (UPC 2016)
Introduction to Deep learning
Technical Deep Dive: Using Apache Kafka to Optimize Real-Time Analytics in Fi...
OpenGLES Android Graphics
Semantic Segmentation AIML Project
Prendre la data par le bon sens
적외선 치료
SSII2021 [SS2] Deepfake Generation and Detection – An Overview (ディープフェイクの生成と検出)
bigquery.pptx
Ad

Similar to Simple Review of Single Image Super Resolution Task (20)

PDF
[PR12] image super resolution using deep convolutional networks
PDF
Deep learning super resolution
PPTX
Photo realistic single image super-resolution using a generative adversarial ...
PPTX
Review SRGAN
PPTX
Unsupervised learning for real-world super-resolution review presentation
PDF
Alexnet paper review
PPTX
Review EDSR
PDF
제 13회 보아즈(BOAZ) 빅데이터 컨퍼런스 - [경기는 계속되어야 한다] : 실시간 축구 중계 영상 화질개선 모델
PDF
History of Vision AI
PPTX
Image net classification with deep convolutional neural networks
PDF
Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)
PPTX
Progressive Growing of GANs for Improved Quality, Stability, and Variation Re...
PDF
"Learning transferable architectures for scalable image recognition" Paper Re...
PPTX
Segment Anything
PPTX
Super resolution
PDF
Summary in recent advances in deep learning for object detection
PDF
Summary in recent advances in deep learning for object detection
PPTX
Image Deep Learning 실무적용
PDF
[부스트캠프 Tech Talk] 배지연_Structure of Model and Task
PDF
"Dataset and metrics for predicting local visible differences" Paper Review
[PR12] image super resolution using deep convolutional networks
Deep learning super resolution
Photo realistic single image super-resolution using a generative adversarial ...
Review SRGAN
Unsupervised learning for real-world super-resolution review presentation
Alexnet paper review
Review EDSR
제 13회 보아즈(BOAZ) 빅데이터 컨퍼런스 - [경기는 계속되어야 한다] : 실시간 축구 중계 영상 화질개선 모델
History of Vision AI
Image net classification with deep convolutional neural networks
Semantic Image Synthesis with Spatially-Adaptive Normalization(GAUGAN, SPADE)
Progressive Growing of GANs for Improved Quality, Stability, and Variation Re...
"Learning transferable architectures for scalable image recognition" Paper Re...
Segment Anything
Super resolution
Summary in recent advances in deep learning for object detection
Summary in recent advances in deep learning for object detection
Image Deep Learning 실무적용
[부스트캠프 Tech Talk] 배지연_Structure of Model and Task
"Dataset and metrics for predicting local visible differences" Paper Review
Ad

More from MYEONGGYU LEE (17)

PDF
(Paper Review) Abnormal Event Detection in Videos using Generative Adversaria...
PDF
Survey of HDR & Tone Mapping Task
PDF
Survey of Super Resolution Task (SISR Only)
PDF
ICCV 2019 Review
PDF
(Paper Review)Few-Shot Adversarial Learning of Realistic Neural Talking Head ...
PDF
(Paper Review)U-GAT-IT: unsupervised generative attentional networks with ada...
PDF
(Book Summary) Classification and ensemble(book review)
PDF
(Paper Review) Reconstruction of Monte Carlo Image Sequences using a Recurren...
PDF
(Paper Review)A versatile learning based 3D temporal tracker - scalable, robu...
PDF
(Paper Review)Neural 3D mesh renderer
PDF
(Paper Review)3D shape reconstruction from sketches via multi view convolutio...
PDF
(Paper Review)Image to image translation with conditional adversarial network...
PDF
(Book summary) Ensemble method 2018summerml_study
PDF
(Paper Review)Towards foveated rendering for gaze tracked virtual reality
PDF
(Paper Review)Geometrically correct projection-based texture mapping onto a d...
PDF
(Papers Review)CNN for sentence classification
PDF
(Paper Review)Kernel predicting-convolutional-networks-for-denoising-monte-ca...
(Paper Review) Abnormal Event Detection in Videos using Generative Adversaria...
Survey of HDR & Tone Mapping Task
Survey of Super Resolution Task (SISR Only)
ICCV 2019 Review
(Paper Review)Few-Shot Adversarial Learning of Realistic Neural Talking Head ...
(Paper Review)U-GAT-IT: unsupervised generative attentional networks with ada...
(Book Summary) Classification and ensemble(book review)
(Paper Review) Reconstruction of Monte Carlo Image Sequences using a Recurren...
(Paper Review)A versatile learning based 3D temporal tracker - scalable, robu...
(Paper Review)Neural 3D mesh renderer
(Paper Review)3D shape reconstruction from sketches via multi view convolutio...
(Paper Review)Image to image translation with conditional adversarial network...
(Book summary) Ensemble method 2018summerml_study
(Paper Review)Towards foveated rendering for gaze tracked virtual reality
(Paper Review)Geometrically correct projection-based texture mapping onto a d...
(Papers Review)CNN for sentence classification
(Paper Review)Kernel predicting-convolutional-networks-for-denoising-monte-ca...

Simple Review of Single Image Super Resolution Task

  • 1. CGLAB 이명규Simple Review of Super Resolution Task (1/34) Presented by CGLAB 이명규 2020/05/27 Simple Review of Super Resolution Task
  • 2. CGLAB 이명규Simple Review of Super Resolution Task (2/34) I N D E X 01 02 Introduction Featured Papers
  • 3. CGLAB 이명규Simple Review of Super Resolution Task (3/34) Introduction Part 01 1. Super Resolution 소개 2. 관련 연구 소개
  • 4. CGLAB 이명규Simple Review of Super Resolution Task (4/34) ↳ Super Resolution 소개1-1 • LR(Low-resolution) image 또는 video를 HR(High-resolution)으로 복구하는 문제 • SISR(Single Image SR)과 MISR(Multiple Image SR)로 구분 Problem Definition 𝒚 𝑳𝑹 = 𝒙⨂𝒌 𝒔 + 𝒏. GT HR image Blur Kernel Down Sampling Noise
  • 5. CGLAB 이명규Simple Review of Super Resolution Task (5/34) ↳ Super Resolution 소개1-1 • SISR Algorithms • Interpolation based method (Bicubic, bilinear, Nearest neighbor etc….) => Just “upscaling” image • Reconstruction based method • Deep Learning based method Problem Definition https://bskyvision.com/531
  • 6. CGLAB 이명규Simple Review of Super Resolution Task (6/34) ↳ Super Resolution 소개1-1 Applications Super Resolution Applications in Modern Digital Image Processing (IJCA 2016) ▲ SR for Satellite Image ▲ SR for Medical Imaging ▲ SR for Astrological Studies ▲ SR for Microscopy Image Processing
  • 7. CGLAB 이명규Simple Review of Super Resolution Task (7/34) ↳ Related Works1-2
  • 8. CGLAB 이명규Simple Review of Super Resolution Task (8/34) ↳ Related Works1-2 Deep Learning for Image Super-resolution: A Survey
  • 9. CGLAB 이명규Simple Review of Super Resolution Task (9/34) ↳ Related Works1-2 • DL을 SSIR에 사용한 첫 논문 → 3-layer CNN, MSE loss • 전통적인 방식 대비 월등한 성능 SRCNN (ECCV 2014) Image Super-Resolution Using Deep Convolutional Networks (ECCV 2014)
  • 10. CGLAB 이명규Simple Review of Super Resolution Task (10/34) ↳ Related Works1-2 • Too Shallow했던 SRCNN의 단점 개선, 안정적이면서 빠른 학습 → VGGNet based deep residual learning+MSE “Adjustable gradient clipping for maximal boost in speed while suppressing exploding gradients” • 단순 LR→HR mapping 학습이 아닌, bicubic upsampled LR 이미지와 HR 이미지의 잔차(residual) 혹은 noise를 학습하는 방식 VDSR (CVPR 2016) Image Super-Resolution Using Deep Convolutional Networks (ECCV 2014) 이후 DRCN(Deeply-recursive CNN), SRResNet, DRRN의 제안으로 이어짐
  • 11. CGLAB 이명규Simple Review of Super Resolution Task (11/34) ↳ Related Works1-2 • GAN을 활용해 인간이 보기에 그럴싸한 이미지를 생성하려는 시도 • MSE loss는 blurry한 이미지를 생성하므로, Perceptual loss 사용을 제안 → 𝑷𝒆𝒓𝒄𝒆𝒑𝒕𝒖𝒂𝒍 𝒍𝒐𝒔𝒔 = 𝑮𝑨𝑵 𝒍𝒐𝒔𝒔 + 𝑪𝒐𝒏𝒕𝒆𝒏𝒕 𝒍𝒐𝒔𝒔 MSE loss 대신 style transfer에서 사용되는 VGG loss로 교체 SRGAN (CVPR 2017) Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network (CVPR 2017)
  • 12. CGLAB 이명규Simple Review of Super Resolution Task (12/34) ↳ Related Works1-2 SRGAN (CVPR 2017) Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network (CVPR 2017)
  • 13. CGLAB 이명규Simple Review of Super Resolution Task (13/34) Featured Papers Part 02 1. Paper 1 (“EDSR”) (Enhanced Deep Residual Networks for Single Image Super-Resolution) 2. Paper 2 (“SAN”, SOTA) (Second-order Attention Network for Single Image Super-Resolution)
  • 14. CGLAB 이명규Simple Review of Super Resolution Task (14/34) Paper 1: Enhanced Deep Residual Networks for Single Image Super-Resolution(“EDSR”) Part 2-1 1. Introduction 2. Architecture Overview 3. Experiment & Conclusion
  • 15. CGLAB 이명규Simple Review of Super Resolution Task (15/34) ↳ Introduction2-1-1 Limitations of Previous Works • ResNet을 그대로 사용 시(SRResnet)SISR에 좋은 성능을 보여 주지만 다음 이슈가 존재 • ResNet과 같은 분류기들은 classification과 같은 high-level 문제를 위한 구조 (SR은 low-level task) • ResNet에 적용된 BN으로 인해 네트워크의 flexibility 저하 → 학습 시간이 오래 걸리는 문제 ▲ Batch Normalization Enhanced Deep Residual Networks for Single Image Super-Resolution (CVPR 2017)
  • 16. CGLAB 이명규Simple Review of Super Resolution Task (16/34) ↳ Introduction2-1-1 Contributions • 학습 시 메모리 사용량 40% 절감 → BN layer를 제거한 new residual block 제안 → 더 깊은 네트워크를 학습할 수 있게 됨 • Single Scale Model(EDSR)과 Multi Scale Model(MDSR) 제안 → 각각 single scale(x2, x3, x4)을 따로따로 학습(EDSR)하거나 동시에 여러 scale에 대해 학습(MDSR) Enhanced Deep Residual Networks for Single Image Super-Resolution (CVPR 2017)
  • 17. CGLAB 이명규Simple Review of Super Resolution Task (17/34) ↳ Architecture Overview2-1-2 Model Overview ▲ EDSR ▲ MDSR Enhanced Deep Residual Networks for Single Image Super-Resolution (CVPR 2017)
  • 18. CGLAB 이명규Simple Review of Super Resolution Task (18/34) ↳ Architecture Overview2-1-2 Model Overview • BN이 빠진 새 residual block 구조 제안 • Residual block 최종 feature matrix에 0.1의 constant를 곱해 줌 → 안정적인 학습 도모 • x3, x4배 scaling 학습에서는 x2 모델을 기반으로 transfer learning Enhanced Deep Residual Networks for Single Image Super-Resolution (CVPR 2017)
  • 19. CGLAB 이명규Simple Review of Super Resolution Task (19/34) ↳ Evaluation results 2-1-3 Experiment & Conclusion Enhanced Deep Residual Networks for Single Image Super-Resolution (CVPR 2017)
  • 20. CGLAB 이명규Simple Review of Super Resolution Task (20/34) ↳ • Conclusion • 종전 방법 대비 학습 시 메모리를 40%까지 줄이는 방법을 제안 • Single Scale(EDSR)과 Multi Scale(MDSR)에서 작동하는 SR 모델의 제안 • Limitations • CNN의 receptive field size가 상대적으로 작음. (즉, local patch에만 집중) → 이미지에서 보다 넓은 영역을 고려하지 못함. • Feature들이 담고 있는 local 혹은 global 정보가 동등하게 처리됨 → 이후 Dilated conv, spatial or channel-wise attention 등의 제안으로 이어짐 Conclusion & Limitations 2-1-3 Experiment & Conclusion Enhanced Deep Residual Networks for Single Image Super-Resolution (CVPR 2017)
  • 21. CGLAB 이명규Simple Review of Super Resolution Task (21/34) Paper 2: Second-order Attention Network for Single Image Super-Resolution(“SAN”) Part 2-2 1. Introduction 2. Architecture Overview 3. Experiment & Conclusion
  • 22. CGLAB 이명규Simple Review of Super Resolution Task (22/34) ↳ Introduction2-2-1 Limitations of Previous Works • 기존 모델들은 깊거나 넓은 구조 디자인에만 초점을 맞춤 → 따라서 레이어 간의 관계를 탐색하지 않아 네트워크 전체의 표현력 저하 • 대부분 LR 이미지의 모든 정보를 사용하지 않아 낮은 성능을 보여 옴 • 학습 속도가 본 논문 대비 상대적으로 느림 Enhanced Deep Residual Networks for Single Image Super-Resolution (CVPR 2017)
  • 23. CGLAB 이명규Simple Review of Super Resolution Task (23/34) ↳ Introduction2-2-1 Contributions • Second order statistics를 활용해 레이어 간 feature 상호 의존성 학습 • LSRAG(local-source residual attention groups) 구조를 제안해 LR 이미지 정보를 적극 활용 → 풍부한 low-frequency 정보
  • 24. CGLAB 이명규Simple Review of Super Resolution Task (24/34) ↳ Architecture Overview2-2-2 Model Overview
  • 25. CGLAB 이명규Simple Review of Super Resolution Task (25/34) ↳ Architecture Overview2-2-2 Detailed view of Model: RL-NL Module • 목적 : 영상을 잘게 쪼개어 영상 전체를 보기 위함 → 영상을 4등분 후 각 영역에 대해 non-local module 적용 (큰 해상도에서도 유리) • SSRG모듈 전후로 수행되며, high-level에서 넓은 범위의 정보들을 모으는 역할 → Global level non-local 연산은 인풋 사이즈가 클 경우 연산량 증가 (Region-Level Non-Local)
  • 26. CGLAB 이명규Simple Review of Super Resolution Task (26/34) ↳ Architecture Overview2-2-2 Detailed view of Model: LSRAG Module • 목적 : Feature inter-dependencies를 잘 유지하기 위함 → 영상을 4등분 후 각 영역에 대해 non-local module 적용 (큰 해상도에서도 유리) • Simplified residual block들로 구성 (Local Source Residual Attention Group)
  • 27. CGLAB 이명규Simple Review of Super Resolution Task (27/34) ↳ Architecture Overview2-2-2 Detailed view of Model: SOCA Module • 목적 : 공분산 정규화를 통한 모델의 Discriminative representation 능력 향상 → Attention을 통해 네트워크가 더욱 중요한 정보를 갖는 feature에 가중치를 더 두도록 함 • 기존 Attention 알고리즘들은 GAP을 이용한 1차 statistics 정보만 활용 → 기존 방식들은 1st order statistics(=average) 이상의 정보를 활용하지 않아 모델의 Discriminative representation 능력이 저하됨 → 따라서 Covariance Normalization을 거쳐 channel attention 수행 (Second-Order Channel Attention)
  • 28. CGLAB 이명규Simple Review of Super Resolution Task (28/34) ↳ Architecture Overview2-2-2 Detailed view of Model: SOCA Module • GAP(Global Average Pooling) : 각 feature map의 노드들로부터 단순히 평균을 추출해 차원 축소 e.g.) (𝟏+𝟗+𝟔+𝟒+𝟓+𝟒+𝟕+𝟖+𝟓+𝟏+𝟐+𝟗+𝟔+𝟕+𝟔+𝟎) 𝟏𝟔 = 𝟓 Replace to GCP(Global Covariance Pooling)
  • 29. CGLAB 이명규Simple Review of Super Resolution Task (29/34) ↳ Architecture Overview2-2-2 Detailed view of Model: SOCA Module • Covariance Normalization: 1. Reshape 𝑯 × 𝑾 × 𝑪 feature map 𝑭 = [𝐟 𝟏, … , 𝐟𝒄] to feature matrix 𝑿. (𝑿 has 𝒔 = 𝑾𝑯 features of 𝑪-dim) 2. Compute sample covariance matrix 𝚺. 𝚺 = 𝐗 ҧ𝐈𝐗 𝑻 , where ҧ𝐈 = 𝟏 𝒔 (𝐈 − 𝟏 𝒔 𝟏). (𝑰=𝒔 × 𝒔 Identity matrix, 𝟏=matrix of all ones)
  • 30. CGLAB 이명규Simple Review of Super Resolution Task (30/34) ↳ Architecture Overview2-2-2 Detailed view of Model: SOCA Module 3. 𝚺 is symmetric positive semi-definte, thus has EIG as follows: 𝚺 = 𝐔𝚲𝐔 𝐓, 𝒘𝒉𝒆𝒓𝒆 𝑼 = 𝒐𝒓𝒕𝒉𝒐𝒈𝒐𝒏𝒂𝒍 𝒎𝒂𝒕𝒓𝒊𝒙, 𝜦 = 𝒅𝒊𝒂𝒈 𝒎𝒂𝒕𝒓𝒊𝒙 𝒘𝒊𝒕𝒉 𝒆𝒊𝒈𝒆𝒏𝒗𝒂𝒍𝒖𝒆𝒔. 4. So, covariance normalization can be coverted to power of 𝒆𝒊𝒈𝒆𝒏𝒗𝒂𝒍𝒖𝒆𝒔. ෡𝐘 = 𝚺 𝜶 = 𝐔𝚲 𝜶 𝐔 𝑻 - 𝜶는 양의 실수로, 1일 경우 정규화를 수행하지 않음. ( 𝛼 =½에서 잘 작동함을 확인) - 𝜶<1인 경우 1.0보다 큰 eigenvalue는 non-linear하게 축소하고, 반대는 늘림.
  • 31. CGLAB 이명규Simple Review of Super Resolution Task (31/34) ↳ Architecture Overview2-2-2 Detailed view of Model: SOCA Module • Normalized covariance matrix ෠𝐘는 channel-wise feature들 간의 상관 관계를 characterize하는 역할 • Covariance normalized ෡𝐘 를 이용해 channel level에서 pooling을 수행 𝒍𝒆𝒕 ෠𝐘 = 𝒚 𝟏, … , 𝒚 𝑪 , 𝒄𝒉𝒂𝒏𝒏𝒆𝒍𝒘𝒊𝒔𝒆 𝒔𝒕𝒂𝒕𝒊𝒔𝒕𝒊𝒄𝒔 𝒛 = 𝑯 𝑮𝑪𝑷 𝒚 𝒄 = 𝟏 𝑪 ෍ 𝒊 𝑪 𝒚 𝑪(𝒊)
  • 32. CGLAB 이명규Simple Review of Super Resolution Task (32/34) ↳ Evaluation results 2-2-3 Experiment & Conclusion
  • 33. CGLAB 이명규Simple Review of Super Resolution Task (33/34) ↳ • 다양한 모듈을 제안해 SISR의 PSNR 성능을 높임 • SSRG 모듈을 통해 low-frequency 정보들을 충분히 활용 • RL-NL 모듈을 통해 long-distance spatial contextual information을 활용 • SOCA 모듈을 통해 global covariance pooling을 수행하고, 레이어 간의 dependencies를 학습 • Second-order channel attention을 통해 discriminative representation학습에 집중 • 네트워크 규모 대비 낮은 수의 파라미터 개수 Conclusion & Limitations 2-2-3 Experiment & Conclusion
  • 34. CGLAB 이명규Simple Review of Super Resolution Task (34/34) Thank you for Listening. Email : brstar96@naver.com (or brstar96@soongsil.ac.kr) Mobile : +82-10-8234-3179