Skip to content

Latest commit

 

History

History
36 lines (27 loc) · 680 Bytes

EMBSR.md

File metadata and controls

36 lines (27 loc) · 680 Bytes
Raw

Short introduction

Moduled based architecture with modified CNN.

Main contributions

  • Modular approach
  • Improve EDSR with pyramid pooling
  • DnResNet

Architecture

Module approach

alt text

EDSR-PP

alt text

DnResNet

text

Loss

Training strategy

  • Fine tune existing modular for different track

Experiments

  • Dataset: DIV2k
  • Evaluation metric: PSNR, SSIM
  • Patchsie: 48 × 48
  • Learning rate from 0.0001, 100 epochs, 0.00001 for next 200 epochs

Final summary

Pros:

  • Great performance achieved in track 2 and 3 evaluated by PSNR

Cons:

  • Details are missing after denoising

Tips: