Scholarlink

<p>Underwater images often suffer from color distortion, low contrast, and blurred details due to light absorption and scattering, which severely affect image quality and limit their practical applications. Although existing deep learning-based methods have achieved certain progress, they still face challenges in balancing local detail restoration and global contextual modeling. To address these issues, this paper proposes UFMFormer, a hybrid Transformer-Mamba architecture for underwater image enhancement. Specifically, the model employs Transformer modules at high-resolution stages to enhance local feature extraction and fine detail restoration, while integrating Mamba modules at low-resolution stages to achieve global modeling with reduced computational complexity. In addition, an RGB channel separation strategy, a Frequency-Spatial Hybrid Attention Transformer (FSHAT), and an Underwater Color Cast Correction Module (UCCM) are designed to further improve color restoration and multi-scale feature fusion capabilities. Experimental results on multiple datasets, including UIEB, LSUI, U45, and C60, demonstrate that UFMFormer achieves competitive performance compared with state-of-the-art methods, obtaining the best PSNR and perceptual scores while maintaining comparable SSIM performance. Ablation studies further validate the effectiveness of each proposed module. The proposed UFMFormer provides an efficient and robust solution for underwater image enhancement and shows great potential for practical applications. Our source code is available at:<a href="https://github.com/DingBC/UFMFormer">https://github.com/DingBC/UFMFormer</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

UFMFormer: a hybrid transformer-Mamba network for underwater image enhancement via FFT

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Ning Li,
  • Bocheng Ding,
  • Yi Yang,
  • Bo Li,
  • Gaochao Yang

摘要

Underwater images often suffer from color distortion, low contrast, and blurred details due to light absorption and scattering, which severely affect image quality and limit their practical applications. Although existing deep learning-based methods have achieved certain progress, they still face challenges in balancing local detail restoration and global contextual modeling. To address these issues, this paper proposes UFMFormer, a hybrid Transformer-Mamba architecture for underwater image enhancement. Specifically, the model employs Transformer modules at high-resolution stages to enhance local feature extraction and fine detail restoration, while integrating Mamba modules at low-resolution stages to achieve global modeling with reduced computational complexity. In addition, an RGB channel separation strategy, a Frequency-Spatial Hybrid Attention Transformer (FSHAT), and an Underwater Color Cast Correction Module (UCCM) are designed to further improve color restoration and multi-scale feature fusion capabilities. Experimental results on multiple datasets, including UIEB, LSUI, U45, and C60, demonstrate that UFMFormer achieves competitive performance compared with state-of-the-art methods, obtaining the best PSNR and perceptual scores while maintaining comparable SSIM performance. Ablation studies further validate the effectiveness of each proposed module. The proposed UFMFormer provides an efficient and robust solution for underwater image enhancement and shows great potential for practical applications. Our source code is available at:https://github.com/DingBC/UFMFormer.