Analysis of performance difference when using knowledge distillation of efficient CNN-based super-resolution algorithm

In this paper, we used RLFN, the NTIRE 2022 Efficient Super-Resolution Challenge winning model, to implement a lightweight and efficient super-resolution algorithm. We also apply knowledge distillation in a partially modified form of the PISR framework and analyze the qualitative and quantitative results to improve the performance while maintaining the cost of RLFN.

I. Introduction

Recently, various studies on super resolution are being actively conducted. Among them, studies are being conducted to make the SR algorithm light but efficient with good performance and operate it in real time even in low-power devices such as edge devices.