Skip to main navigation Skip to search Skip to main content

Wavelet-Fourier frequency decoupling network for underwater image enhancement

  • Zhanyou Cui
  • , Peirui Ji
  • , Zhuang Miao
  • , Shuming Yang
  • , Deguang Li
  • , Weibo Zhang
  • , Wenyi Zhao
  • , Zheng Liang
  • , Weidong Zhang
  • School of Engineering
  • Xi’an International University
  • Xi'an Jiaotong University
  • School of Information Science
  • Luoyang Normal University
  • College of Oceanography and Space Informatics
  • China University of Petroleum (East China)
  • School of Artificial Intelligence
  • Beijing University of Posts and Telecommunications
  • Anhui University
  • School of Information Engineering
  • Henan Institute of Science and Technology

Research output: Contribution to journal › Article › peer-review

Abstract

Underwater image degradation exhibits strong frequency dependence. Low-frequency components tend to suffer from color and illumination shifts, while high-frequency components exhibit blurred edges and attenuated textures, resulting in simultaneous structural distortion and detail loss. Existing methods mainly perform unified modeling in the spatial domain and lack targeted processing of different frequency components, which leads to mutual interference between low- and high-frequency information during joint optimization and limits coordinated recovery of structure and detail. To address this issue, we propose a wavelet-Fourier frequency decoupling network for underwater image enhancement (UIE). Specifically, we first apply a wavelet transform to decompose the input features into low-frequency structural components and high-frequency detail components. We then design a frequency-separated dual-branch module that introduces Fourier modulation into the low-frequency branch and performs dynamic frequency modeling to restore global structure, color, and illumination. Meanwhile, the high-frequency branch employs multi-scale spatial modeling and attention mechanisms to enhance edges and texture details, and the module performs cross-frequency fusion between low- and high-frequency features to achieve collaborative enhancement of structure and detail. On this basis, we further develop a frequency dynamic feed forward network, which utilizes frequency positional encoding and dynamic frequency filtering to improve frequency representation and structural consistency. Extensive qualitative and quantitative results demonstrate that the proposed method outperforms representative UIE methods in terms of visual quality and detail restoration and exhibits greater robustness and generalizability in complex underwater scenes. The code is available at WFFDNet.

Original languageEnglish
Article number134365
JournalExpert Systems with Applications
Volume333
DOIs
StatePublished - 15 Jan 2027

Keywords

  • Cross-frequency fusion
  • Frequency decoupling
  • Underwater image enhancement
  • Wavelet-Fourier modeling

Fingerprint

Dive into the research topics of 'Wavelet-Fourier frequency decoupling network for underwater image enhancement'. Together they form a unique fingerprint.

Cite this