FourierSR: A Fourier Token-Based Plugin for Efficient Image Super-Resolution.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Image super-resolution (SR) aims to reconstruct high-resolution images from low-resolution inputs.
• Current SR methods using convolutions or window-based Transformers face limitations due to restricted receptive fields, hindering efficiency under computational constraints.

Purpose of the Study:

• To develop an efficient SR plugin that overcomes the limitations of existing receptive field methods.
• To introduce a novel approach for improving SR performance with global receptive fields and reduced complexity.

Main Methods:

• Proposed FourierSR, a Fourier token-based plugin inspired by the convolution theorem.
• Utilized Fourier transform and multiplication operations, avoiding complex token mixing techniques.
• Integrated FourierSR as a plugin into existing efficient SR models.

Main Results:

• FourierSR achieved an average PSNR gain of 0.34dB on the Manga109 dataset at 4x scale.
• The plugin introduced minimal overhead, with only a 0.6% increase in parameters and 1.5% in FLOPs.
• Demonstrated improved SR efficiency and performance across various existing SR methods.

Conclusions:

• FourierSR offers an effective and efficient solution for image super-resolution.
• The plugin's global receptive field and low computational complexity make it suitable for resource-constrained SR tasks.
• FourierSR uniformly improves SR performance without the instability or inefficiency of other plugin methods.