README.md

Improving-Closed-loop-Matters(Super Resolution)

Presentation

Goal

By adding a small idea, tried to improve Closed-loop Matters: Dual Regression Networks for Single Image Super-Resolution(CVPR 2020) baseline code

Overview

Inspired by SamsungSDS 's FrePGAN: Robust Deepfake Detection Using Frequency-level Perturbations,

Images in Frequency domain seems to contain extra infomations that are not in rgb images.

Thus I used images in frequency domain and rgb images at the sametime to train model for Super Resolution task

I simply calcuated new Floss(frequency loss), which is loss between GT images in frequency doamin and created images in frequency domain

As result there was a slight improvement in PSNR score.

Results

Tried simple tests altering weights of Floss, and there were slight improvement, but it was hard to tell with bare eyes

Ground Truth

Baseline

With New Floss

Reference

Closed-loop Matters: Dual Regression Networks for Single Image Super-Resolution [arXiv] [github]