Dual-Domain Learning for Image Restoration

This direction is merged into another direction: dual-domain image restoration. Its last version stays readable and citable here.

Version 1 of 155 papers66 conceptsfrozen 2026-10-04

Most image restoration tasks—like removing blur, noise, or compression artifacts—are traditionally fought in the "spatial domain." We treat images as grids of pixels and apply convolutional filters to fix local errors. However, many degradations are not local. A JPEG "blocky" artifact is a mathematical byproduct of frequency quantization, and a camera blur is a global spreading of light.

Recent research has converged on a powerful alternative: Dual-Domain Learning. By processing images simultaneously in both the spatial (pixel) domain and the frequency (spectral) domain, models can fix structural details and global patterns far more efficiently than spatial-only networks.

The Frequency Domain as a Mathematical Shortcut

In the spatial domain, a neural network must use many layers or large kernels to "see" global structures. In the frequency domain—accessed via the Fast Fourier Transform (FFT) or Discrete Cosine Transform (DCT)—global patterns become local. A repeating moiré pattern that covers an entire image in pixels becomes a single, identifiable point in the frequency spectrum.

Zhangyang Wang, Ding Liu, and their colleagues demonstrated this with the D³ model. Instead of forcing a CNN to learn how to "un-block" a JPEG image from scratch, they "unfolded" the JPEG compression math into the network architecture. By operating in the DCT domain first, the model addresses frequency-specific artifacts where they are most visible, then uses a spatial branch to refine the final pixels. This dual-domain approach allowed them to outperform previous models by 1 dB in PSNR while running 30 times faster.

Decoupling Amplitude and Phase

A key intuition in this field is that an image's frequency representation can be split into amplitude (which often corresponds to lighting and contrast) and phase (which contains the structural "skeleton" and edges).

For tasks like low-light enhancement, Daniel Feijoo and Marcos V. Conde in DarkIR argue that global illumination is best handled in the frequency domain, while local structural issues like blur and noise are better handled in the spatial domain. Similarly, the UHDRes framework achieves state-of-the-art results on 4K images by decoupling restoration into explicit amplitude modulation (for lighting) and implicit phase refinement (for structure).

This separation prevents the "checkerboard" artifacts and color distortions common in purely spatial models. By treating restoration as a signal processing task, researchers like Kai Shang in the MDMS Diffusion Model can refine global brightness and fine textures in parallel, ensuring that brightening an image doesn't accidentally amplify high-frequency noise.

Frequency-Aware Architectures

The shift toward dual-domain processing has led to several novel architectural components:

  • Frequency Selectors: Wei Shen discovered that simply applying a ReLU activation in the frequency domain acts as a "frequency selector" that can isolate hidden blur patterns, leading to the Res FFT-ReLU Block.
  • Wavelet-Spatial Hybrids: While FFT is global, Wavelet transforms provide a middle ground by capturing localized frequency information. Honggang Chen and Truong Q. Nguyen used this in DPW-SDNet to preserve high-frequency textures that are usually smoothed out by standard denoising filters.
  • Fourier-Mamba Integration: Newer models like FourierMamba by Dong Li combine the efficiency of State Space Models (Mamba) with Fourier transforms. This allows the model to capture long-range dependencies across the frequency spectrum, which is particularly effective for removing complex, non-uniform degradations like rain streaks.

Why It Works: Efficiency and Precision

The primary reason dual-domain learning is becoming the standard for high-resolution restoration is efficiency. Transformers are excellent at capturing global context but suffer from quadratic computational costs as image resolution increases.

In contrast, the FFT has a complexity of $O(N \log N)$. By using frequency-domain branches to handle global context, models like FUSION by Pavithra L. K. can achieve high-fidelity results with as few as 0.28 million parameters. This makes it possible to run sophisticated image restoration on edge devices—like underwater drones or smartphones—where heavy Transformer models would fail.

By bridging the gap between classical signal processing and modern deep learning, dual-domain architectures allow us to fix images not just by looking at pixels, but by understanding the underlying "rhythm" of the light and the degradation itself.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_dual_domain_learning_for_image_restorati_167a5e78,
  title = {Dual-Domain Learning for Image Restoration},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dual-domain-learning-for-image-restoration/rdr_167a5e78eeaa4e06898b01e38307148b},
  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_for_image_restorati_167a5e78_v1,
  title = {Dual-Domain Learning for Image Restoration},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/dual-domain-learning-for-image-restoration/rdr_167a5e78eeaa4e06898b01e38307148b/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}