[논문리뷰] SRGAN : Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network

Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network

논문 링크 : https://arxiv.org/abs/1609.04802

GAN(Generative Adversarial Network) 구조를 이용한 Super-Resolution(초해상화) 방법론입니다.

(수정 중 입니다.)

---

Key Points

: It’s frist framework capable of inferring photo-realistic natural images for 4X upscaling factors.
- 4배 업스케일링이 가능한 첫 번째 프레임워크.

: To achieve upscaling, developers propose a perceptual loss function which consists of an adversarial loss and a content loss.
- Loss Function은 Adversarial loss, Content loss를 사용하여 구성.

---

Abstract

Despite the accuracy and speed of Neural Networks with using faster and deeper convolutional neural networks, one central problem remains largely unsolved
: how to recover the finer texture details when we super-resolve at large upscaling factors?

Existing Super-Resolution methods are optimized by using MSE minimizing
-> PSNR is high, but high-frequency details are often lacking.

So, they suggest SRGAN for image SR, with using Adversarial Loss & Content Loss.

Adversarial Loss : discriminator network train.
Content Loss : Perceptual similarity train at the pixel space.

---

Introduction

The task of estimating a high-resolution(HR) image from its low-resolution(LR) counterpart is referred to as super-resolution(SR)

SR algorithms is commonly the minimization of the MSE between the recovered HR image and the ground truth
-> But the ability of MSE and PSNR to capture perceptually relevant differences, such as high texture detail, is very limited as they are defined based on pixel-wise image differences.

In this work, they propose a SRGAN for which they employ a ResNet with skip-connection and diverge from MSE as the sole optimization target.
And define a novel perceptual loss using high-level feature maps of the VGG Network

 

---

Methods

In SISR(Single Image Super-Resolution) aim is to estimate a high-resolution, super-resolved image
from a low-resolution input image.

It’s almost same the usual GAN Loss that Goodfellow suggested

The Loss value is obtained using the Fake high-resolution images which is
created from the LR image and HR image (ground truth)

Generator & Discriminator Network Structure

---

Results

---

개인적인 생각
SRGAN의 경우 PSNR이나 SSIM은 수치적으로 낮음에도 불구하고 MOS평가에서 높은 점수를 받았는데,
수치적으로 더 나은 결과를 가져오지 않았더라도 논문이 accept된 것이 특이한 케이스가 아닌가 생각됩니다.