[pr12] image super resolution using deep convolutional networks

31
Image Super-Resolution using Deep Convolutional Neural Networks (2016) Paper reviewed by Taegyun Jeon Dong, Chao, et al. "Image super-resolution using deep convolutional networks." IEEE transactions on pattern analysis and machine intelligence. 38.2 (2016): 295-307.

Upload: taegyun-jeon

Post on 21-Jan-2018

1.249 views

Category:

Engineering


0 download

TRANSCRIPT

Page 1: [PR12] image super resolution using deep convolutional networks

Image Super-Resolution using Deep Convolutional Neural Networks (2016)

Paper reviewed by Taegyun Jeon

Dong, Chao, et al. "Image super-resolution using deep convolutional networks." IEEE transactions on pattern analysis and machine intelligence. 38.2 (2016): 295-307.

Page 2: [PR12] image super resolution using deep convolutional networks

연구소개 | 문제 정의● Single Image Super-Resolution

○ ILL-posed problem (many candidate solutions)

저해상도 이미지(Low resolution)

고해상도 이미지(High resolution)

Page 3: [PR12] image super resolution using deep convolutional networks

연구소개 | 파급효과 및 지향점

위성영상 의료 영상

현미경영상 천체영상

Page 5: [PR12] image super resolution using deep convolutional networks

배경지식 | 평가방식● PSNR (Peak Signal-to-Noise Ratio)

○ 최대 신호 대 잡음비 (영상 화질 손실정보에 대한 평가)

● SSIM (structural similarity index)○ “밝기, 명암, 구조”를 조합하여 두 영상의 유사도 평가

● MS-SSIM (Multi-scale SSIM)○ 다양한 배율 (scale)에 대해 SSIM 평가

● IFC (Information fidelity criterion)○ 서로 다른 두 확률분포에 대한 의존도 평가

Page 6: [PR12] image super resolution using deep convolutional networks

배경지식 | SISR 관련 선행 연구

Page 7: [PR12] image super resolution using deep convolutional networks

방법론 | 핵심 아이디어 ● Convolutional Neural Networks with 3-layers

Page 8: [PR12] image super resolution using deep convolutional networks

방법론 | 핵심 아이디어

ReLU

ReLU

Page 9: [PR12] image super resolution using deep convolutional networks

주요결과 | 실험 결과

Page 10: [PR12] image super resolution using deep convolutional networks

주요결과 | 실험 결과● Training data

Page 11: [PR12] image super resolution using deep convolutional networks

주요결과 | 실험 결과● Number of filters

Page 12: [PR12] image super resolution using deep convolutional networks

주요결과 | 실험 결과

Page 13: [PR12] image super resolution using deep convolutional networks

주요결과 | 실험 결과● Color channels

Page 14: [PR12] image super resolution using deep convolutional networks

리뷰의견 | ● 방법과 결과가 유효한가?

○ Sparse-coding과 같은 맥락의 표현을 CNN으로 구현

● 효율적인 방법인가?○ Representation learning 관점에서 end-to-end learning이 가능하므로 효율적

● 독창적인가?○ Single Image Super-Resolution 문제에 대해 최초로 딥러닝 적용

● 중요한 결과를 만들어냈는가?○ 2014년 이후 거의 모든 Super-Resolution 문제에 대한 연구들이 SRCNN을 기반으로 함.○ Super-Resolution 에 대한 breakthrough

Page 15: [PR12] image super resolution using deep convolutional networks

Accurate Image Super-Resolution using Very Deep Convolutional Networks (2016)

Paper reviewed by Taegyun Jeon

Kim, Jiwon, Jung Kwon Lee, and Kyoung Mu Lee. "Accurate image super-resolution using very deep convolutional networks." Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016.

Page 16: [PR12] image super resolution using deep convolutional networks

방법론 | 핵심 아이디어

SRCNN VDSR

Architecture end-to-end learning end-to-end learning

Receptive fields 13 x 13 41 x 41

Scale factors x3 x2, x3, x4

Learning rate 10-5 learning rate decay (10-2, 10-6)

Depth 4 (Up to) 20

Page 17: [PR12] image super resolution using deep convolutional networks

방법론 | 핵심 아이디어

Page 18: [PR12] image super resolution using deep convolutional networks

방법론 | 핵심 아이디어 ● Convolutional Neural Networks with 20 layers

○ Large receptive fields (41 x 41) with consecutive 3x3 convolution kernels○ Convolutional layers with zero-padding

● Residual learning ○ Original loss:

○ Residual loss:

● Adjustable gradient clipping with high learning rates○ Common gradient clipping:

○ Adjustable gradient clipping:

y: outputx: inputf(x): prediction

θ: predefined range: current learning rate

Page 19: [PR12] image super resolution using deep convolutional networks

방법론 | 핵심 아이디어 ● Single CNN for multi-scale factor super-resolution

○ Size of input patch = Size of receptive fields (41 x 41)○ Images are divided into sub-images with no overlap.○ Mini-batch size is 64.○ Sub-images from different scales can be in the same batch.

Page 20: [PR12] image super resolution using deep convolutional networks

방법론 | 핵심 아이디어 ● Hyperparameters (Final model)

○ depth = 20○ batch_size = 64○ momentum = 0.9○ weight_decay = 0.0001○ weight_initialization = he_normal()○ activations = ReLU○ epochs = 80○ learning_rate = 0.1○ learning_rate was decreased by a factor 10 every 20 epochs.

Page 21: [PR12] image super resolution using deep convolutional networks

주요결과 | 실험 결과

Page 22: [PR12] image super resolution using deep convolutional networks

주요결과 | 실험 결과●

Page 23: [PR12] image super resolution using deep convolutional networks

주요결과 | 실험 결과●

Page 24: [PR12] image super resolution using deep convolutional networks

리뷰의견 | ● 방법과 결과가 유효한가?

○ SRCNN을 20개의 레이어로 확장. 표현력 증가

● 효율적인 방법인가?○ Residual learning과 adjustable gradient clipping을 통해 빠른 학습 달성

● 독창적인가?○ Image restoration 문제에서 input-output의 유사성을 residual learning 적용

● 중요한 결과를 만들어냈는가?○ 학습 속도 대폭 개선.○ SRCNN과 비교하여 확연히 개선된 결과

Page 25: [PR12] image super resolution using deep convolutional networks

부록 | 1. 구현 코드 (Caffe and Matlab)

Caffe and Matlab Implementation (Author): http://mmlab.ie.cuhk.edu.hk/projects/SRCNN.html

Page 26: [PR12] image super resolution using deep convolutional networks

부록 | 2. CNN 기반 관련 논문 리스트 (2014-2016)● Deep Networks for Image Super-Resolution with Sparse Prior (ICCV2015),

Zhaowen Wang et al.● Robust Single Image Super-Resolution via Deep Networks with Sparse Prior (TIP2016),

Ding Liu et al.● Accurate Image Super-Resolution Using Very Deep Convolutional Networks (CVPR2016),

Jiwon Kim et al.● Deeply-Recursive Convolutional Network for Image Super-Resolution (CVPR2016),

Jiwon Kim et al.● Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel

Convolutional Neural Network (CVPR2016), Wenzhe Shi et al.

● Accelerating the Super-Resolution Convolutional Neural Network (ECCV2016), Dong Chao et al.

Page 27: [PR12] image super resolution using deep convolutional networks

부록 | 2. GAN 기반 관련 논문 리스트 (2014-2016)● Perceptual Losses for Real-Time Style Transfer and Super-Resolution (ECCV2016),

Justin Johnson et al.● Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network,

Christian Ledig et al.● Amortised Map Inference for Image Super-Resolution,

Casper Kaae Sønderby et al.● EnhanceNet: Single Image Super-Resolution through Automated Texture Synthesis,

Mehdi S. M. Sajjadi et al.

Page 28: [PR12] image super resolution using deep convolutional networks

부록 | 3. Challenge on Super-Resolution

Page 29: [PR12] image super resolution using deep convolutional networks

부록 | Open Questions● CNN 기반 모델은 GAN 기반 모델보다 무조건 성능이 나쁘다고 볼 수 있는가?

Page 30: [PR12] image super resolution using deep convolutional networks

부록 | Open Questions● PSNR은 여전히 유효한성능 지표인가?

Page 31: [PR12] image super resolution using deep convolutional networks

감사합니다

fb.com/taegyun.jeon

github.com/[email protected]

Paper reviewed by Taegyun Jeon