Rethinking underwater image enhancement via a hybrid Mamba–transformer network
Abstract
Underwater imaging is simultaneously affected by non-uniform illumination and multi-path light scattering, often resulting in severe degradations such as color distortion, contrast reduction, and detail blurring. These degradations substantially hinder the performance of high-level vision tasks, including underwater object detection. To address these challenges, this paper proposes a hybrid Mamba–Transformer network for underwater image enhancement, centered on a Hybrid Mamba–Transformer Block that couples local state-space modeling with phase-guided frequency-domain self-attention. Specifically, the local state-space branch employs four-directional selective scanning to capture local structural information and texture continuity, while the frequency-domain self-attention branch exploits phase information to model complementary frequency-aware contextual correlations. Through their collaborative modeling, the proposed framework provides a unified representation of local structural degradation and frequency-related feature corruption in underwater images. Complementing this core design, a hybrid depth convolutional feed-forward network and a wavelet-based downsampling block are incorporated to further refine cross-channel features, suppress noise interference, and preserve structural details during feature transformation. Extensive experiments on multiple benchmark datasets demonstrate that the proposed method consistently achieves competitive or superior enhancement performance with relatively low model complexity, validating its effectiveness and robustness for underwater image enhancement.
Related
- Interpretable attention-deficit hyperactivity disorder diagnosis from functional magnetic resonance imaging via hybrid convolutional neural network–transformer learning and multi-level graph-based comprehensibility
- MTF-Net: a brain tumor segmentation network with multi-scale transformer and feature fusion enhancement
- Rethinking Input Complexity in Transformer-Based Clinical Prediction: Implications for Feature Dimensionality and Sequence Length in Longitudinal EHR Data
- Underwater hybrid ESD versus conventional ESD for colorectal laterally spreading tumors: a retrospective study
- Parrot dandelion optimizer-based routing and spiking neural network enabled intrusion detection in underwater IoT
BibTeX
@article{li2026rethinking,
title = {{Rethinking underwater image enhancement via a hybrid Mamba–transformer network}},
author = {Yufeng Li and Ruochen Shi and Chuanlong Xie},
journal = {Scientific Reports},
year = {2026},
month = oct,
doi = {10.1038/s41598-026-73332-x},
url = {https://www.nature.com/articles/s41598-026-73332-x},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Rethinking underwater image enhancement via a hybrid Mamba–transformer network
AU - Yufeng Li
AU - Ruochen Shi
AU - Chuanlong Xie
PY - 2026
DA - 2026/10/05
JO - Scientific Reports
DO - 10.1038/s41598-026-73332-x
UR - https://www.nature.com/articles/s41598-026-73332-x
AB - Underwater imaging is simultaneously affected by non-uniform illumination and multi-path light scattering, often resulting in severe degradations such as color distortion, contrast reduction, and detail blurring. These degradations substantially hinder the performance of high-level vision tasks, including underwater object detection. To address these challenges, this paper proposes a hybrid Mamba–Transformer network for underwater image enhancement, centered on a Hybrid Mamba–Transformer Block that couples local state-space modeling with phase-guided frequency-domain self-attention. Specifically, the local state-space branch employs four-directional selective scanning to capture local structural information and texture continuity, while the frequency-domain self-attention branch exploits phase information to model complementary frequency-aware contextual correlations. Through their collaborative modeling, the proposed framework provides a unified representation of local structural degradation and frequency-related feature corruption in underwater images. Complementing this core design, a hybrid depth convolutional feed-forward network and a wavelet-based downsampling block are incorporated to further refine cross-channel features, suppress noise interference, and preserve structural details during feature transformation. Extensive experiments on multiple benchmark datasets demonstrate that the proposed method consistently achieves competitive or superior enhancement performance with relatively low model complexity, validating its effectiveness and robustness for underwater image enhancement.
SN - 2045-2322
ER -