Dual Domain Learning with Fourier Transforms

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2282 papers356 conceptsfrozen 2026-10-04

Standard convolutional neural networks (CNNs) are inherently local. A $3 \times 3$ kernel only "sees" its immediate neighbors, requiring many stacked layers to capture the global structure of an image. While Transformers address this with self-attention, they often suffer from quadratic computational costs. A growing body of research suggests a more efficient alternative: processing information simultaneously in the spatial domain (pixels) and the frequency domain (spectral coefficients).

By leveraging the Fast Fourier Transform (FFT), researchers are building "dual-domain" architectures that achieve a global receptive field in a single layer while maintaining the fine-grained precision of local convolutions.

The Global Receptive Field via Frequency Duality

The core intuition behind this research direction is the Convolution Theorem: a convolution in the spatial domain is equivalent to point-wise multiplication in the frequency domain. More importantly, a single point in the frequency domain represents a periodic pattern that spans the entire spatial input.

When Jiacheng Ruan and colleagues developed MEW-UNet, they replaced standard self-attention with Multi-axis External Weights in the frequency domain. This allows the model to capture long-range dependencies—like the overall shape of an organ in a medical scan—without the heavy overhead of Transformers. Similarly, Chenyang Wang and others proposed SFMNet, which uses the Fourier Transform to capture global facial geometry, ensuring that a super-resolved face remains structurally coherent even as local textures are sharpened.

The Dual-Branch Architecture

The most successful implementation of this idea is the parallel dual-branch structure. Instead of choosing between pixels and frequencies, models like MD-Recon-Net by Maosong Ran process both in tandem.

  1. The Spatial Branch: Uses standard convolutions to preserve local edges, textures, and fine-grained details.
  2. The Frequency Branch: Uses FFT to transform features into the spectral domain, applies learnable filters (often called "spectral filters" or "Fourier kernels"), and transforms them back.

This interaction solves a common trade-off in image restoration. In tasks like dehazing, Han Zhou demonstrated that Fast Fourier Convolutions can account for large-scale atmospheric haze patterns (global) while standard convolutions handle the underlying sharp objects (local).

Decoupling Amplitude and Phase

A unique advantage of the frequency domain is the ability to separate an image into its Amplitude and Phase components. In the Fourier domain, the amplitude typically carries information about illumination and color, while the phase carries structural and edge information.

Jie Huang 0017 exploited this in FECNet for exposure correction. By modulating the amplitude, the network can adjust global lighting without distorting the structural integrity preserved in the phase. This decoupling allows models to be significantly more lightweight; FECNet achieves state-of-the-art results with a fraction of the parameters of spatial-only models.

Beyond Images: Physics and Signals

The power of dual-domain learning extends to scientific computing and signal processing. Standard Fourier Neural Operators (FNOs) often act as low-pass filters, struggling with sharp transitions in fluid dynamics. Xuanle Zhao addressed this by introducing complementary convolutions to capture high-frequency oscillations that global spectral methods might smooth over.

In medical imaging, Nassir Navab and Azade Farshad showed that their Y-Net architecture could identify fluid pockets in retinal scans more accurately by looking for specific spectral signatures that are nearly invisible in the spatial pixel grid.

Why It Works: Efficiency and Priors

The reason this direction is gaining momentum is twofold:

  • Computational Efficiency: FFT-based operations provide a global receptive field with $O(N \log N)$ complexity, compared to the $O(N^2)$ of standard global attention.
  • Inductive Priors: Many real-world degradations—like motion blur, moiré patterns, or JPEG artifacts—have distinct periodic signatures. As Wei Shen showed in Deep Residual Fourier Transformation, applying a simple ReLU in the frequency domain can act as a "frequency selector" that isolates and removes blur kernels that are otherwise tangled with image content in the spatial domain.

Papers in this version

Showing 60 of 282 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 195 papers
  2. Version 1 · 2026-10-04 · 282 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_dual_domain_feature_extraction_0607733e,
  title = {dual-domain feature extraction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dual-domain-feature-extraction/rdr_0607733e0cc14fd182153f602e0c997e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_dual_domain_learning_with_fourier_transf_0607733e_v1,
  title = {Dual Domain Learning with Fourier Transforms},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dual-domain-learning-with-fourier-transforms/rdr_0607733e0cc14fd182153f602e0c997e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}