Fourier-Guided Attention Upsampling for Image Super-Resolution

2025-08-14

AI Summary

Traditional sampling methods (e.g., sub-pixel convolution) in single-image super-resolution often suffer from high-frequency detail loss and aliasing artifacts. To address this, we propose the Frequency-Guided Attention (FGA) module—the first to integrate Fourier feature encoding with cross-resolution correlation attention. FGA employs a Fourier-based MLP for explicit frequency-domain positional encoding and introduces an L1 loss in the Fourier domain to enforce spectral consistency of reconstructions. This lightweight design incurs no additional parameters and can be seamlessly embedded into mainstream super-resolution architectures. Evaluated on five benchmark models, FGA consistently improves PSNR by 0.12–0.14 dB on average and enhances spectral consistency by up to 29%. The method significantly strengthens texture recovery fidelity and aliasing suppression capability, demonstrating superior performance in preserving fine-grained structural details.

Technology Category

• Computer Vision: Multi-modal Vision
• Machine Learning: Multimodal Learning
• Natural Language Processing: Generation

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

We propose Frequency-Guided Attention (FGA), a lightweight upsampling module for single image super-resolution. Conventional upsamplers, such as Sub-Pixel Convolution, are efficient but frequently fail to reconstruct high-frequency details and introduce aliasing artifacts. FGA addresses these issues by integrating (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. Adding merely 0.3M parameters, FGA consistently enhances performance across five diverse super-resolution backbones in both lightweight and full-capacity scenarios. Experimental results demonstrate average PSNR gains of 0.12~0.14 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.

Problem

• Reconstruct high-frequency details in super-resolution images
• Reduce aliasing artifacts in upsampling processes
• Improve frequency-domain consistency in image reconstruction

Innovation

• Fourier feature MLP for frequency encoding
• Cross-resolution attention for spatial alignment
• Frequency-domain L1 loss for spectral fidelity