Zero-Shot Super-Resolution using Deep Internal Learning

[flrngel]

https://arxiv.org/abs/1712.06087

Abstract

• many paper has restriction on their data

1. Introduction

• this paper talks about "Real LR" means low-rate image on wild

Feature of ZSSR(Zero-Shot Super Resolution)

• train small CNN at test time
• uses CNN to infer HR-LR relation

2. The Power of Internal Image Statistics

• evidence from same image
• predictive
• gets low entropy of internal information

3. Image-Specific CNN

• Pair "I_down_scale" and "I"

3.1. Architecture & Optimization

• Model uses 8 Layer and 64 channels
• uses ReLU
• with 1 increment of Scale(S), 54 seconds takes more to test

3.2. Adapting to the Test Image

• other model's hyperparameter can not be change after train

4.2. The 'Non-ideal' Case

• they made their own dataset

Checkpoints

• Why model uses residual between the interpolated LR and its HR parent?
• What does non-synthesized goes with reliability in this paper?
• This paper can be read with "Deep Prior"
• What effect "Gaussian noise" does?
• Check "Nonparametric blind super-resolution" paper