FGA-SR: Fourier-Guided Attention Upsampling for Image Super-Resolution

Daejune Choi, Youchan No, Jinhyung Lee, Duksu Kim
Korea University of Technology and Education (KOREATECH)
Image and Vision Computing
Volume 172, August 2026, 106049
Paper arXiv Code Visual Results Dataset Weights
Pre/Post spectrum & feature maps across upsamplers

Comparison of upsampling heads on Urban100 (img_007) with an EDSR backbone (×4 SR). Top row: pre-/post-upsampling feature maps; bottom row: corresponding Fourier magnitudes for Interp+Conv, Deconv, PixelShuffle, and FGA. FGA concentrates energy along true orientations, suppresses folding near Nyquist, and yields cleaner post-upsampling features with fewer checkerboard artifacts.

Abstract

We propose Fourier-Guided Attention (FGA), a lightweight and general-purpose upsampling module for single image super-resolution. While conventional approaches such as sub-pixel convolution remain efficient and widely adopted, they often fail to recover high-frequency details and are prone to aliasing artifacts such as checkerboard patterns. To address these limitations, FGA incorporates three complementary components: (1) a Fourier feature-based Multi-Layer Perceptron (MLP) for positional frequency encoding, (2) a cross-resolution Correlation Attention Layer for adaptive spatial alignment, and (3) a frequency-domain L1 loss for spectral fidelity supervision. Despite introducing only 0.3M additional parameters, FGA consistently improves reconstruction quality across seven representative super-resolution backbones under both lightweight and full-capacity settings. Experimental results demonstrate average PSNR gains of 0.04–0.12 dB and improved frequency-domain consistency by up to 29\%, particularly evident on texture-rich datasets. Visual and spectral evaluations confirm FGA's effectiveness in reducing aliasing and preserving fine details, establishing it as a practical, scalable alternative to traditional upsampling methods.

Method Overview

FGA architecture: FF-MLP + CAL + FL1

FF-MLP. Inject Fourier-encoded (sin/cos) spatial coordinates into an MLP aligned with sub-pixel groups, conditioning high-frequency behavior during PixelShuffle.

CAL (in XRA, Cross-Resolution Attention). Window attention uses HR queries and LR keys/values to correct misalignment and suppress aliasing, with efficient overlapping windows.

FL1 Loss. Direct supervision in the Fourier domain (half-spectrum), constraining both amplitude and phase to reduce ringing/checkerboard artifacts.

Spectral & Feature Analysis

FRC-AUC: Measuring High-Frequency Fidelity

Average FRC curves and equal-pixel ring indexing map

We compute Fourier Ring Correlation (FRC) between the SR output \(\hat{Y}\) and the ground truth \(Y\). The 2D spectrum is partitioned into \(N=64\) equal-pixel rings \(\{\mathcal{R}_i\}_{i=0}^{N-1}\). Using your paper’s formulation, for a spatial frequency radius \(q\) we define:

\[ \mathrm{FRC}(q)= \frac{\displaystyle\sum_{|\mathbf f|=q}F_{1}(\mathbf f)\,F_{2}^{*}(\mathbf f)} {\displaystyle\sqrt{\left(\sum_{|\mathbf f|=q}|F_{1}(\mathbf f)|^{2}\right) \left(\sum_{|\mathbf f|=q}|F_{2}(\mathbf f)|^{2}\right)}} \, . \]

We summarize the high-frequency band by averaging the top quartile of rings. With \(N=64\), the high-frequency index starts at \(i_{\mathrm{HF}}=\lceil 0.75N\rceil = 48\). The FRC-AUC is:

\[ \mathrm{FRC\text{-}AUC} = \frac{1}{N_{\text{HF}}} \sum_{i=i_{\mathrm{HF}}}^{N - 1} \mathrm{FRC}_i, \quad i_{\mathrm{HF}} = \lceil 0.75N \rceil = 48, \quad N_{\text{HF}} = 16 \; (N - i_{\mathrm{HF}}). \]

Top-Quartile Ring–Masked Reconstructions (Rings 48–63)

Only rings 48–63 are retained in the Fourier domain; all other frequencies are zeroed before iFFT. Colormap: viridis (purple/blue → yellow = larger magnitude of the band-passed spatial signal).

This visualization is not an FFT magnitude map. We keep only the top-quartile rings (48–63), zero out all other frequencies, and invert back to the spatial domain. FGA suppresses checkerboard-like high-frequency components near the Nyquist region while concentrating energy along true structural directions, yielding spatial patterns closer to the GT.

Empirically, FGA consistently boosts the high-frequency FRC bands across architectures/datasets, indicating closer spectral agreement with GT textures and fewer aliasing artifacts.

FRC-AUC (Top-25% Rings)

Open demo notebook

Quantitative Results

Qualitative Comparisons

Paper

BibTeX

@article{CHOI2026106049,
title = {Fourier-guided attention upsampling for image super-resolution},
journal = {Image and Vision Computing},
volume = {172},
pages = {106049},
year = {2026},
issn = {0262-8856},
doi = {https://doi.org/10.1016/j.imavis.2026.106049},
url = {https://www.sciencedirect.com/science/article/pii/S0262885626001563},
author = {Daejune Choi and Youchan No and Jinhyung Lee and Duksu Kim},
keywords = {Super-resolution, Attention, Upsampler},
abstract = {We propose Fourier-Guided Attention (FGA), a lightweight and general-purpose upsampling module for single image super-resolution. While conventional approaches such as sub-pixel convolution remain efficient and widely adopted, they often fail to recover high-frequency details and are prone to aliasing artifacts such as checkerboard patterns. To address these limitations, FGA incorporates three complementary components: (1) a Fourier feature-based Multi-Layer Perceptron (MLP) for positional frequency encoding, (2) a cross-resolution Correlation Attention Layer for adaptive spatial alignment, and (3) a frequency-domain L1 loss for spectral fidelity supervision. Despite introducing only 0.3M additional parameters, FGA consistently improves reconstruction quality across seven representative super-resolution backbones under both lightweight and full-capacity settings. Experimental results demonstrate PSNR gains by up to 0.42 dB and improved frequency-domain consistency by up to 29%, particularly evident on texture-rich datasets. Visual and spectral evaluations confirm FGA’s effectiveness in reducing aliasing and preserving fine details, establishing it as a practical, scalable alternative to traditional upsampling methods. Our code and trained weights are publicly available.11http://github.com/HPC-Lab-KOREATECH/FGA-SR.}
}