UHDRes: Ultra-High-Definition Image Restoration via Dual-Domain Decoupled Spectral Modulation

TL;DR

UHDRes addresses the challenge of restoring ultra-high-definition images under degradations with a lightweight, dual-domain approach. It explicitly models the amplitude spectrum in the frequency domain through the spectral amplitude modulation unit (SAMU) and implicitly refines the phase via spatial-domain refinement with the structural refinement unit (SRU), all within a spatio-spectral fusion module (SSFM). The method leverages three key components—MSCA for multi-scale spatial features, a decoupled spectral modulation block (DSMB), and a shared gated feed-forward network (SGFN)—organized into dual-domain adaptive enhancement blocks (DAEBs) in a three-level encoder–decoder, enabling residual learning to produce high-quality UHD restorations with about 400K parameters. Across five UHD benchmarks covering low-light, dehazing, deblurring, and deraining tasks, UHDRes achieves state-of-the-art restoration while reducing memory and latency, demonstrating strong practical appeal for real-time edge deployment. The work highlights that concentrating on amplitude-domain information in the frequency domain, while letting spatial refinement handle phase-related details, can efficiently balance restoration quality and computational cost for UHD imagery.

Abstract

Ultra-high-definition (UHD) images often suffer from severe degradations such as blur, haze, rain, or low-light conditions, which pose significant challenges for image restoration due to their high resolution and computational demands. In this paper, we propose UHDRes, a novel lightweight dual-domain decoupled spectral modulation framework for UHD image restoration. It explicitly models the amplitude spectrum via lightweight spectrum-domain modulation, while restoring phase implicitly through spatial-domain refinement. We introduce the spatio-spectral fusion mechanism, which first employs a multi-scale context aggregator to extract local and global spatial features, and then performs spectral modulation in a decoupled manner. It explicitly enhances amplitude features in the frequency domain while implicitly restoring phase information through spatial refinement. Additionally, a shared gated feed-forward network is designed to efficiently promote feature interaction through shared-parameter convolutions and adaptive gating mechanisms. Extensive experimental comparisons on five public UHD benchmarks demonstrate that our UHDRes achieves the state-of-the-art restoration performance with only 400K parameters, while significantly reducing inference latency and memory usage. The codes and models are available at https://github.com/Zhao0100/UHDRes.

Figures (14)

• Figure 1: Visualization of the effects of perturbing the amplitude and phase spectra of a given image. (Top) After frequency-domain decomposition, the amplitude and phase spectra are independently perturbed (0.1–0.3) and then inverse-transformed to obtain the visualized reconstructions. (Bottom) PSNR scores of the reconstructed images under different perturbation levels applied to the decomposed amplitude and phase spectra. Under the same degree of direct interference, the phase spectrum is more sensitive.
• Figure 2: Model performance and computational efficiency comparison on $1024 \times 1024$ resolution images from the UHD-Haze dataset. PSNR, inference latency, and parameter count are visualized. UHDRes achieves the best trade-off between restoration quality, inference speed, and model size.
• Figure 3: Overall network architecture of UHDRes. The model consists of stacked dual-domain adaptive enhancement blocks (DAEBs), each comprising a spatio-spectral fusion module (SSFM) and a shared gated feed-forward network (SGFN). SSFM combines multi-scale feature extraction with frequency-domain decoupled modulation. The multi-scale context aggregator (MSCA) captures local-global textures, while the decoupled spectral modulation block (DSMB) adjusts amplitude via the spectral amplitude modulation unit (SAMU) and refines structure using the structural refinement unit (SRU). SGFN promotes efficient and consistent feature learning through shared dual-branch gating.
• Figure 4: Visual comparison to SOTA methods on the UHD-LL dataset. UHDRes restores more realistic colors.
• Figure 5: Visual comparison to SOTA methods on the UHD-Haze dataset. UHDRes removes thick fog and restores richer details.