Hybrid-domain attention learning for lightweight underwater image enhancement and real-time vision deployment Authors: Shan Jiang, Chenfeng Sun, Xuan Liu Published: 2026 ABSTRACT Underwater image enhancement is critical for marine exploration and robotic vision. However, real-world underwater images suffer from color distortion, low contrast, and detail loss, while existing deep learning methods are often computationally heavy. This work presents a lightweight hybrid-domain enhancement network that jointly models frequency and spatial information for effective underwater image restoration. The architecture integrates frequency-spatial attention, channel attention pointwise, and multi-shape convolution to balance performance and efficiency. Experiments on public datasets show that the proposed method achieves higher PSNR and SSIM than mainstream methods with lower computational cost, supporting real-time deployment on resource-constrained devices. This work provides an effective solution for high-quality and efficient underwater vision applications. The source code and implementation details are publicly available at https://github.com/cfengsun16/HDAMS-Net to ensure reproducibility. Introduction Underwater image enhancement (UIE) plays an essential role in ocean exploration, underwater navigation, and robotic perception. Images captured underwater suffer from complex optical degradation-including color distortion, low contrast, and detail blurring-that significantly hinders high-level vision tasks and reduces the reliability of underwater visual systems. Addressing these degradations efficiently and robustly remains an open challenge. Traditional physics-based methods, such as DCP1 and Retinex-based models2, attempt to compensate for color distortion and visibility degradation through handcrafted priors. Although these approaches have achieved partial success, their performance is limited by strong assumptions about water conditions and lighting environments. Recent advances in deep learning have introduced CNN-based3 and GAN-based4 architectures, demonstrating strong capability in modeling complex visual degradation. Similar deep learning paradigms have also shown effectiveness in other image-based diagnostic tasks5,25, further highlighting the generalization ability of data-driven feature learning. while recent lightweight hybrid CNN models with attention mechanisms further improve restoration performance and efficiency6, learning end-to-end mappings from degraded to clear images. Models such as Ucolor7, Water-Net8, and PUGAN9 have demonstrated improved visual quality but often at the expense of high computational cost and network complexity. However, Transformer-based10 frameworks enhance global feature modeling capability but are difficult to deploy on resource-constrained underwater devices, while recent structure-aware Transformer enhancement models attempt to balance performance and efficiency11,12. To overcome these challenges, this paper proposes Hybrid-Domain Attention with Multi-shape Structure (HDAMS-Net), a lightweight hybrid-domain architecture for underwater image enhancement. Unlike existing hybrid-domain methods such as SFGNet, which process frequency and spatial features in separate sequential stages without explicit cross-domain interaction, our FSA module performs pixel-level frequency modulation and spatial channel recalibration jointly within a single unified block, enabling tighter coupling between global spectral correction and local texture recovery. Unlike standard channel attention mechanisms such as SE blocks or CBAM, which operate purely on spatial feature statistics, the CAP module decomposes spatial convolution along horizontal and vertical axes before applying channel reweighting, allowing it to capture directional long-range dependencies with significantly lower computational cost. Unlike prior multi-scale bottleneck designs that rely on isotropic square kernels of varying sizes, the MSC module introduces anisotropic stripe kernels (1×31 and 31×1) alongside a large square kernel (31×31), specifically targeting the directional scattering and streak illumination patterns characteristic of underwater degradation. The network integrates these three complementary modules-Frequency-Spatial Attention (FSA), Channel Attention Pointwise (CAP), and Multi-Shape Convolution (MSC)-to jointly capture multi-scale spatial structures, frequency-domain cues, and long-range directional dependencies within a compact encoder-decoder framework. Figure 1 demonstrates the results of our method. The proposed architecture is designed to balance enhancement quality and computational efficiency, enabling real-time deployment in embedded or mobile underwater imaging systems. Comprehensive evaluations across various benchmarks verify that HDAMS-Net achieves superior performance in terms of PSNR and SSIM13 across the evaluated benchmarks, while concurrently maintaining a compact model size, indicating its potential for practical underwater vision applications such as autonomous navigation and marine resource inspection. Our main contributions can be summarized as:We propose HDAMS-Net, a lightweight hybrid-domain underwater enhancement network that integrates the Frequency-Spatial Attention, Channel Attention Pointwise, and Multi-Shape Convolution modules into a unified architecture. Distinct from prior hybrid-domain methods that address frequency and spatial degradation in decoupled or sequential pipelines, HDAMS-Net achieves tight cross-domain interaction within each processing block. The FSA module is the first to combine pixel-level FFT-based spectral modulation with spatial channel attention and large-kernel spatial filtering in a single module for underwater enhancement. The CAP module introduces axial depthwise decomposition prior to channel reweighting, a design not previously explored in UIE, which simultaneously improves directional feature modeling and reduces parameter count. The MSC bottleneck uniquely combines isotropic large-kernel convolution with anisotropic stripe convolution to handle both omnidirectional context and direction-specific underwater artifacts. Together, these innovations enable efficient color correction, texture recovery, and global-local consistency with low computational cost.The proposed method achieves real-time inference with a compact model size, making it highly suitable for deployment on resource-constrained underwater platforms such as embedded imaging systems and autonomous underwater robots.Extensive experiments on multiple underwater benchmark datasets demonstrate that HDAMS-Net generally achieves higher PSNR and SSIM compared with several existing CNN- and Transformer-based methods under the evaluated settings while maintaining low parameter count and fast inference, confirming its suitability for real-time underwater vision applications.Fig. 1A visual demonstration of our underwater restoration method. The image in the first line is the original image input, and the second line is the enhanced result of underwater image restoration. We propose HDAMS-Net, a lightweight hybrid-domain underwater enhancement network that integrates the Frequency-Spatial Attention, Channel Attention Pointwise, and Multi-Shape Convolution modules into a unified architecture. Distinct from prior hybrid-domain methods that address frequency and spatial degradation in decoupled or sequential pipelines, HDAMS-Net achieves tight cross-domain interaction within each processing block. The FSA module is the first to combine pixel-level FFT-based spectral modulation with spatial channel attention and large-kernel spatial filtering in a single module for underwater enhancement. The CAP module introduces axial depthwise decomposition prior to channel reweighting, a design not previously explored in UIE, which simultaneously improves directional feature modeling and reduces parameter count. The MSC bottleneck uniquely combines isotropic large-kernel convolution with anisotropic stripe convolution to handle both omnidirectional context and direction-specific underwater artifacts. Together, these innovations enable efficient color correction, texture recovery, and global-local consistency with low computational cost. We propose HDAMS-Net, a lightweight hybrid-domain underwater enhancement network that integrates the Frequency-Spatial Attention, Channel Attention Pointwise, and Multi-Shape Convolution modules into a unified architecture. Distinct from prior hybrid-domain methods that address frequency and spatial degradation in decoupled or sequential pipelines, HDAMS-Net achieves tight cross-domain interaction within each processing block. The FSA module is the first to combine pixel-level FFT-based spectral modulation with spatial channel attention and large-kernel spatial filtering in a single module for underwater enhancement. The CAP module introduces axial depthwise decomposition prior to channel reweighting, a design not previously explored in UIE, which simultaneously improves directional feature modeling and reduces parameter count. The MSC bottleneck uniquely combines isotropic large-kernel convolution with anisotropic stripe convolution to handle both omnidirectional context and direction-specific underwater artifacts. Together, these innovations enable efficient color correction, texture recovery, and global-local consistency with low computational cost. The proposed method achieves real-time inference with a compact model size, making it highly suitable for deployment on resource-constrained underwater platforms such as embedded imaging systems and autonomous underwater robots. The proposed method achieves real-time inference with a compact model size, making it highly suitable for deployment on resource-constrained underwater platforms such as embedded imaging systems and autonomous underwater robots. Extensive experiments on multiple underwater benchmark datasets demonstrate that HDAMS-Net generally achieves higher PSNR and SSIM compared with several existing CNN- and Transformer-based methods under the evaluated settings while maintaining low parameter count and fast inference, confirming its suitability for real-time underwater vision applications. Extensive experiments on multiple underwater benchmark datasets demonstrate that HDAMS-Net generally achieves higher PSNR and SSIM compared with several existing CNN- and Transformer-based methods under the evaluated settings while maintaining low parameter count and fast inference, confirming its suitability for real-time underwater vision applications. A visual demonstration of our underwater restoration method. The image in the first line is the original image input, and the second line is the enhanced result of underwater image restoration. Related work UIE methods Underwater image enhancement (UIE) aims to restore visual quality from images degraded by wavelength-dependent light attenuation and backscattering, which produce characteristic artifacts including color distortion, low contrast, and hazy appearance. Research in this area has evolved through three broad paradigms, each addressing the limitations of its predecessor. Physics-based methods were the earliest systematic approach to UIE. Methods such as UNTV14, MLLE15, and ADPCC16 exploit handcrafted priors-such as the Dark Channel Prior or Retinex theory-to estimate transmission maps or illumination components. While these methods improve visibility under controlled conditions, they rely on strong assumptions about scene depth and water optical properties that rarely hold in practice. Their performance degrades significantly under dynamic illumination, mixed water types, or complex scattering environments, and they offer no mechanism to adapt to scene-specific degradation patterns, motivating the shift toward data-driven approaches. Deep learning-based methods overcame many of these limitations by learning nonlinear enhancement mappings directly from paired data. CNN-based frameworks such as Ucolor, UIE-WD17, and PUGAN demonstrated substantially improved robustness and perceptual quality compared to physics-based predecessors. Subsequent Transformer-based models, including U-Transformer18 and URSCT19, leveraged self-attention to capture long-range dependencies and global contextual relationships, further improving color consistency and structural fidelity. However, these gains come at a significant computational cost: Transformer-based models typically require tens of gigaflops and tens of millions of parameters, making them poorly suited for deployment on resource-constrained underwater platforms such as embedded imaging systems or autonomous underwater vehicles. Moreover, most CNN and Transformer methods process features entirely in the spatial domain, lacking an explicit mechanism to address the frequency-domain nature of global color attenuation-a fundamental characteristic of underwater degradation. Hybrid and lightweight methods have more recently attempted to bridge the gap between restoration quality and computational efficiency. Physics-guided networks20 and illumination-guided enhancement models combine data-driven learning with physical priors to improve interpretability and generalization, but their increased structural complexity often reintroduces the deployment challenges of earlier deep models. Lightweight architectures reduce parameter count through depthwise separable convolutions and MLP-based designs, but they typically do not address the frequency-spatial coupling inherent in underwater degradation. These limitations collectively motivate the design of HDAMS-Net: a lightweight hybrid-domain architecture that explicitly models both frequency-domain color attenuation and spatial-domain texture degradation within a compact, deployable framework. By jointly leveraging frequency-spatial attention, axial channel perception, and multi-shape convolution, HDAMS-Net addresses the core technical bottlenecks of prior work while remaining suitable for real-time embedded deployment. Lightweight U-net architectures for image enhancement U-Net, originally developed for biomedical image segmentation21, introduced a symmetric encoder-decoder structure with skip connections, which later became a widely adopted paradigm in various image enhancement tasks. Its ability to combine multi-scale contextual information with fine-grained spatial details makes it a strong baseline for underwater image enhancement as well. However, the standard U-Net architecture is computationally expensive, limiting its applicability to real-time or embedded enhancement scenarios. To overcome these limitations, recent studies have focused on lightweight modifications of the U-shaped framework. For example, Unext22 incorporates multilayer perceptrons (MLPs) to reduce convolutional redundancy, while U-Lite23 employs depthwise separable convolutions to significantly lower computational cost without sacrificing enhancement quality. These lightweight strategies improve parameter efficiency, facilitate deployment on mobile or edge devices, and maintain robust performance across diverse visual tasks. Building on these advancements, our work follows similar lightweight design principles by integrating hybrid-domain attention mechanisms and multi-scale feature extraction within an efficient U-shaped topology. This provides a compact yet powerful architectural foundation for HDAMS-Net, enabling high-quality underwater image enhancement under resource-constrained conditions. Methodology We present HDAMS-Net, a lightweight encoder-bottleneck-decoder architecture designed for underwater image restoration by jointly leveraging spatial, frequency, and multi-scale information, inspired by recent hybrid representation learning frameworks24. As shown in Fig. 2, the overall framework follows a U-shaped topology in which the encoder progressively abstracts spatial and frequency features, the bottleneck aggregates multi-scale contextual information, and the decoder reconstructs high-fidelity details via skip-connection fusion. The three proposed modules-FSA, CAP, and MSC-are embedded at specific positions within this topology to target distinct aspects of underwater degradation, as detailed in the following subsections.Fig. 2Overall architecture of the proposed HDAMS-Net. The network consists of Channel point-by-point attention module, frequency domain - spatial attention module and multi-shape convolutional module composition. Overall architecture of the proposed HDAMS-Net. The network consists of Channel point-by-point attention module, frequency domain - spatial attention module and multi-shape convolutional module composition. U-shape network architecture As illustrated in Fig. 2, the proposed HDAMS-Net employs a simplified U-shaped encoder-decoder framework. This configuration allows the network to holistically integrate multi-scale features, achieving an effective balance between high restoration fidelity and computational economy without relying on extensive module stacking. The architecture comprises three core components: a progressive encoder, a bottleneck layer, and a symmetrical decoder. The primary role of the encoder is to hierarchically reduce spatial dimensions and abstract high-level semantic representations. Conversely, the decoder is tasked with progressively reconstructing fine-grained textures and spatial details through a series of upsampling operations and the fusion of complementary low-level features from the encoder pathway. Skip connections are integral to this design, serving as conduits to transfer rich, high-resolution spatial information from the initial encoder stages to the corresponding deeper decoder stages. This mechanism enhances model performance by facilitating multi-stage feature fusion and mitigating the loss of structural consistency. In contrast to conventional U-Net blocks, which typically consist of stacked standard convolutional layers, the proposed model introduces novel, specialized building blocks for both the encoder and decoder. These customized modules leverage advanced attention mechanisms and depthwise convolutions to broaden the effective receptive field and improve the adaptive weighting of contextual features. Consequently, this specialized block design enables HDAMS-Net to deliver robust performance and superior processing speed within a streamlined architectural footprint. MSC module The MSC Block constitutes the bottleneck stage of HDAMS-Net, serving as a critical bridge between the encoding and decoding paths. It is designed to capture multi-scale contextual information and restore geometric consistency by employing convolutions with diverse kernel shapes and orientations. Inspired by the bottleneck architectures in OKNet26 and LU2Net27, our approach utilizes multiple parallel convolutional operations, specifically integrating point-wise convolution, large kernel convolution, and both horizontal and vertical stripe convolutions. The large kernel convolution (31× 31) enables the model to effectively capture long-range contextual dependencies within a single stage. The 1× 1 point-wise convolution calibrates pixel-level weights, enhancing the module’s adaptability to local image differences. Furthermore, the horizontally (1× 31) and vertically (31× 1) extended kernels are particularly effective in modeling the anisotropic structures commonly found in underwater images, such as streaking illumination patterns and directional scattering. The core architecture of the MSC module’s bottleneck is composed of four parallel convolutional branches: 31× 31, 1× 1, 1× 31, and 31× 1 (excluding the identity mapping/residual connection). This structure can be mathematically expressed as:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\textbf{Y}} = \sigma \left( {\textbf{F}} + {\mathcal {F}}_{31\times 31}({\textbf{F}}) + {\mathcal {F}}_{1\times 31}({\textbf{F}}) + {\mathcal {F}}_{31\times 1}({\textbf{F}}) + {\mathcal {F}}_{1\times 1}({\textbf{F}}) \right) \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\textbf{F}}$$\end{document} represents the input feature map, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal {F}}_{K}$$\end{document} denotes the depth-wise convolution (with 1 × 1 being a point-wise convolution) with kernel size K, and σ is the activation function (e.g., \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {ReLU}$$\end{document} or \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {GELU}$$\end{document}). This design efficiently aggregates multi-scale, multi-orientation spatial information with minimal parameter overhead. Therefore, the MSC block enhances global context consistency and local structural fidelity, complementing the preceding CAP and FSA modules to achieve superior underwater image restoration performance. CAP module The CAP Module serves as a crucial component for feature enhancement and channel-wise representation within the HDAMS-Net’s encoder and decoder paths. As illustrated in Fig. 3, the CAP block is built around the Axial Channel Attention (ACA) module, which integrates axial depthwise convolution (see Fig. 4) and channel attention to efficiently capture both long-range directional dependencies and inter-channel correlations.Fig. 3Structure of the ACA module. The module combines horizontal and vertical depthwise convolutions with channel attention to capture directional and inter-channel dependencies. Structure of the ACA module. The module combines horizontal and vertical depthwise convolutions with channel attention to capture directional and inter-channel dependencies. Specifically, the input feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\textbf{X}}$$\end{document} is first processed by the ACA module. This module performs separable horizontal (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {DWConv}_{h}$$\end{document}) and vertical (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\text {DWConv}_{v}$$\end{document}) depthwise convolutions to aggregate features along distinct axes, followed by channel re-weighting via a Channel Attention (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal {C}}_{\text {Attn}}$$\end{document}) mechanism. The operation of the ACA module can be mathematically expressed as:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\textbf{Y}} = {\mathcal {C}}_{\text {Attn}} \left( {\textbf{X}} + \text {DWConv}_{h}({\textbf{X}}) + \text {DWConv}_{v}({\textbf{X}}) \right) \end{aligned}$$\end{document}Fig. 4Comparison between standard convolution and axial depthwise convolution. Unlike traditional 2D convolution that jointly processes both spatial dimensions, the axial design decomposes convolution along horizontal and vertical directions, improving efficiency and long-range dependency modeling. Comparison between standard convolution and axial depthwise convolution. Unlike traditional 2D convolution that jointly processes both spatial dimensions, the axial design decomposes convolution along horizontal and vertical directions, improving efficiency and long-range dependency modeling. where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\textbf{Y}}$$\end{document} is the enhanced feature map output by the ACA module. While the ACA module provides a consistent feature refinement mechanism, the full CAP block is structurally adapted for its deployment stage. In the encoder, the resulting features are further processed by a 1× 1 pointwise convolution (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal {F}}_{1\times 1}$$\end{document}) and then reduced in resolution via MaxPooling to achieve effective downsampling and compact feature representation. Conversely, the decoder employs Upsampling to restore feature resolution and concatenates skip-connection features before applying the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal {F}}_{1\times 1}$$\end{document} convolution and the ACA module for multi-scale feature fusion and refinement. This lightweight, adaptive design allows the CAP module to simultaneously strengthen semantic encoding and ensure efficient feature transmission throughout the HDAMS-Net. FSA module The FSA module is strategically divided into two complementary subdomains: a Frequency-Domain branch, which captures global spectral correlations, and a Spatial-Domain branch, which executes channel-wise reweighting and local spatial refinement. These two branches operate jointly to correct holistic color and contrast biases while simultaneously preserving fine-grained textural detail. This dual-domain design reflects the characteristics of underwater degradation: global color attenuation and contrast imbalance are frequency-domain phenomena, whereas texture loss and edge blurring are inherently spatial. By separating feature processing into these domains, the FSA module corrects global color shifts through spectral modulation while restoring structural details via spatial refinement, enabling joint handling of global illumination distortion and local texture degradation. Frequency-Domain Branch. Following the implementation method, the input feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\textbf{X}}$$\end{document} is initially processed by two separate 1× 1 convolutions to yield spatial representations \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\textbf{F}}_{s}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\textbf{F}}_{f}$$\end{document}. The latter, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\textbf{F}}_{f}$$\end{document}, is transformed into the frequency domain using a 2D Fast Fourier Transform (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{FFT}$$\end{document}), modulated by the spatial projection \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\textbf{F}}_{s}$$\end{document}, and then reverted to the spatial domain via the Inverse Fast Fourier Transform (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{IFFT}$$\end{document}). This process generates a frequency-attended feature map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\textbf{F}}_{\text {freq}}$$\end{document}:3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\textbf{F}}_{\text {freq}} \;=\; \big |\; \textrm{IFFT}\big (\textrm{FFT}({\textbf{F}}_{f}) \odot {\textbf{F}}_{s}\big )\;\big | \end{aligned}$$\end{document}where ⊙ denotes element-wise multiplication, and |· | signifies the magnitude, which is taken to suppress potential phase noise. This branch effectively emphasizes global, spectrum-level cues, such as long-range color casts and contrast degradation, with minimal additional spatial complexity. Spatial-Domain Branch. The output from the frequency branch, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\textbf{F}}_{\text {freq}}$$\end{document}, is fed into the spatial processing pipeline. This pipeline involves a sequence of operations: channel attention, lightweight depthwise spatial filtering, large-kernel spatial attention, normalization, and a feed-forward refinement. First, a channel attention map \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\textbf{A}}_c$$\end{document} is computed using global average pooling (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{GAP}$$\end{document}), followed by two 1× 1 convolutions and corresponding nonlinearities (ReLU denoted by δ, then Sigmoid denoted by σ):4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\textbf{A}}_c \;=\; \sigma \big (\textrm{Conv}_{1}\big (\delta (\textrm{Conv}_{1}(\textrm{GAP}({\textbf{F}}_{\text {freq}})))\big )\big ) \end{aligned}$$\end{document}This channel attention map is applied multiplicatively to the frequency-enhanced features. In parallel, the original input \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\textbf{X}}$$\end{document} undergoes a depthwise 3× 3 convolution (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{DWConv}_{3\times 3}$$\end{document}) to retain essential local spatial details. The two spatial streams are then fused through summation:5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\textbf{S}} \;=\; ({\textbf{F}}_{\text {freq}} \odot {\textbf{A}}_c) \;+\; \textrm{DWConv}_{3\times 3}({\textbf{X}}) \end{aligned}$$\end{document}To model medium-range spatial dependencies further, the combined feature \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\textbf{S}}$$\end{document} is passed through the Large Kernel Attention (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{LKA}$$\end{document}) block. Subsequently, Layer Normalization (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{LN}$$\end{document}) and the lightweight Feed-Forward Network (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\textrm{FFN}$$\end{document}) are applied, and finally, a residual connection is utilized to ensure stability:6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\textbf{Y}} \;=\; {\textbf{S}} \;+\; \textrm{FFN}\big (\textrm{LN}(\textrm{LKA}({\textbf{S}}))\big ) \end{aligned}$$\end{document}The spatial branch’s primary function is to perform adaptive channel recalibration and detailed spatial refinement, which includes capturing local and medium-range context, thereby effectively complementing the global cues generated by the frequency branch. Together, the Frequency-Domain and Spatial-Domain submodules form the complete FSA module: the former targets holistic spectral correction, while the latter enforces channel-wise and spatial refinement to recover fine textures and correct localized color and contrast artifacts. As illustrated in Fig. 5, the spatial branch consists of three key submodules - SC-Attention, LKA Module, and Feed Forward - that collaboratively refine spatial and channel-wise features. The SC-Attention recalibrates channel responses using global statistics, the LKA block captures medium-range spatial dependencies with large kernels, and the Feed Forward network refines nonlinear feature interactions for enhanced local detail.Fig. 5Detailed architecture of the spatial-domain submodules in the FSA block, including SC-Attention, LKA Module, and Feed Forward components. These three modules cooperate to enhance spatial consistency and local structural refinement. Detailed architecture of the spatial-domain submodules in the FSA block, including SC-Attention, LKA Module, and Feed Forward components. These three modules cooperate to enhance spatial consistency and local structural refinement. Loss function To achieve a more comprehensive perception of image information and enhance overall visual quality, a weighted composite loss function is designed as follows:7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\mathcal {L}}_{total} = \lambda _1 {\mathcal {L}}_{RGB} + \lambda _2 {\mathcal {L}}_{Lab} + \lambda _3 {\mathcal {L}}_{SSIM} + \lambda _4 {\mathcal {L}}_{PSNR}, \end{aligned}$$\end{document}where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _{1-4}$$\end{document} are empirically determined balancing coefficients that ensure stable convergence and perceptually consistent optimization. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal {L}}_{RGB}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal {L}}_{Lab}$$\end{document} represent pixel-level mean squared errors (MSE) computed in the RGB and CIE-Lab color spaces, respectively, encouraging chromatic fidelity and overall color consistency. The structural similarity loss \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal {L}}_{SSIM}=1-SSIM({\hat{I}},I^{gt})$$\end{document} enforces preservation of luminance, contrast, and local structures. The \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal {L}}_{PSNR}$$\end{document} term corresponds to the pixel intensity reconstruction quality, expressed as an MSE-equivalent penalty that correlates inversely with the PSNR metric. Since \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal {L}}_{PSNR}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal {L}}_{SSIM}$$\end{document} differ in numerical scale, their relative contributions are balanced by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _{3}$$\end{document} and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _{4}$$\end{document} to maintain optimization stability. This multi-domain formulation jointly constrains pixel accuracy, structural integrity, and perceptual consistency, leading to visually natural and quantitatively superior underwater image restoration results. In practice, the coefficients are set to \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _1=1.0$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _2=0.5$$\end{document}, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _3=0.2$$\end{document}, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _4=0.1$$\end{document}, which are empirically found to provide the best balance between quantitative fidelity and perceptual realism. Experiments Datasets and evaluation metrics Underwater image datasets are generally divided into synthetic and real-world types, with models trained on synthetic data often struggling to adapt to real scenes. To address this, we selected training data from three high-quality real-world datasets: UIEB28, EUVP29, and LSUI30, which provide diverse and high-quality underwater images. Specifically, we adopted an 8:2 ratio for data partitioning, where 712 images from UIEB, 18296 from EUVP, and 4183 from LSUI were used for training, while the remaining 178, 4574, and 857 images were reserved for testing, respectively. This ensures a balanced and representative distribution of samples across different underwater conditions, enabling comprehensive evaluation of both the performance and generalization capability of the proposed method. We employ three widely recognized evaluation metrics: PSNR, SSIM, and LPIPS. PSNR and SSIM measure pixel fidelity and structural similarity, while LPIPS31 measures perceptual similarity using deep network activations with learned linear weighting, which are widely adopted in modern deep vision tasks. These metrics together provide a thorough and balanced evaluation of each method’s performance in underwater image restoration tasks. Preprocess all datasets, adjust the size of LSUI images to 256 × 256, UIEB images to 384 × 384, and EUVP images to 128 × 128, and normalize the pixel values to the range of [0,1]. Implementation details All experiments were implemented in Python using the PyTorch framework on an Ubuntu 22.04 system with an NVIDIA RTX 4090 GPU (24 GB VRAM). The proposed HDAMS-Net was trained for 1500 epochs with a batch size of 8 using the Adam optimizer (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _1$$\end{document}=0.5, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\beta _2$$\end{document}=0.999) and an initial learning rate of 2× 10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$^{-4}$$\end{document}, which was halved every 100 epochs. All input images were resized to 256× 256 and normalized to [0,1], with data augmentation including random flipping, rotation, cropping, and scaling to enhance generalization. Evaluation metrics For a more detailed quantitative evaluation, we further present the formal definitions of the three metrics used in this study: PSNR, SSIM, and LPIPS. PSNR evaluates pixel-level reconstruction fidelity using the logarithmic ratio between the maximum pixel intensity and the mean squared error (MSE) between the enhanced image and the ground truth:8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} PSNR = 10 \log _{10} \left( \frac{MAX_I^2}{MSE} \right) , \quad MSE = \frac{1}{N} \sum _{i=1}^{N} (I_i - I_i^{gt})^2. \end{aligned}$$\end{document}SSIM measures the similarity of structural information by comparing luminance, contrast, and local texture:9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} SSIM(x, y) = \frac{(2\mu _x \mu _y + C_1)(2\sigma _{xy} + C_2)}{(\mu _x^2 + \mu _y^2 + C_1)(\sigma _x^2 + \sigma _y^2 + C_2)}. \end{aligned}$$\end{document}LPIPS reflects perceptual similarity by computing deep feature distances extracted from pretrained neural networks. Lower LPIPS values indicate higher perceptual quality. These three metrics jointly assess pixel accuracy, structural consistency, and perceptual realism, ensuring a comprehensive evaluation of underwater image enhancement performance. Quantitative results Table 1 compares HDAMS-Net with representative underwater enhancement methods, including both traditional and deep learning approaches. Compared models include ROP24, ADPCC16, URSCT19, USUIR32, GUPDM33, TUDA34, U-Transformer10, and SFGNet35. Our HDAMS-Net achieves the best or highly competitive reconstruction quality across the evaluated benchmarks. Specifically, HDAMS-Net attains the highest PSNR, SSIM, and the lowest LPIPS on UIEB (23.68 dB/0.920/0.107) and LSUI (25.59 dB/0.897/0.145), while also maintaining competitive performance on EUVP. These results suggest that our model provides improved color fidelity, structural consistency, and perceptual realism compared with the evaluated methods.Table 1Quantitative comparison on the paired test sets UIEB, EUVP, and LSUI.MethodUIEBEUVPLSUIPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSROP2418.480.8490.20915.340.7140.34317.380.8060.281ADPCC1617.330.8190.21915.200.6920.34916.200.7630.299URSCT1918.600.8220.21815.950.6480.33717.900.7390.283USUIR3222.480.9070.12421.940.8100.23923.750.8600.184GUPDM3322.740.9030.13124.790.8470.18425.330.8770.150TUDA3422.720.9150.11823.730.8430.20725.520.8780.154U-Transformer1020.750.8100.22824.990.8290.23825.150.8380.221SFGNet3519.570.6850.21422.680.5850.22122.710.6530.204Ours23.680.9200.10724.760.8450.15725.590.8970.145Table 2Comparison of no-reference underwater image quality metrics UIQM and UCIQE on the UIEB, EUVP, and LSUI datasets. Higher values indicate better perceptual underwater image quality.MethodUIEBEUVPLSUIUIQMUCIQEUIQMUCIQEUIQMUCIQEROP242.350.5122.480.5312.310.496ADPCC162.310.4982.420.5202.250.478URSCT192.580.5412.730.5602.550.528USUIR322.950.5983.100.6032.880.590GUPDM333.060.6023.180.6073.020.598TUDA343.100.6063.220.6103.070.603U-Transformer102.820.5753.050.5922.780.570SFGNet353.010.5943.160.6012.980.589Ours3.180.6103.190.6123.120.607 Table 1Quantitative comparison on the paired test sets UIEB, EUVP, and LSUI.MethodUIEBEUVPLSUIPSNRSSIMLPIPSPSNRSSIMLPIPSPSNRSSIMLPIPSROP2418.480.8490.20915.340.7140.34317.380.8060.281ADPCC1617.330.8190.21915.200.6920.34916.200.7630.299URSCT1918.600.8220.21815.950.6480.33717.900.7390.283USUIR3222.480.9070.12421.940.8100.23923.750.8600.184GUPDM3322.740.9030.13124.790.8470.18425.330.8770.150TUDA3422.720.9150.11823.730.8430.20725.520.8780.154U-Transformer1020.750.8100.22824.990.8290.23825.150.8380.221SFGNet3519.570.6850.21422.680.5850.22122.710.6530.204Ours23.680.9200.10724.760.8450.15725.590.8970.145 Quantitative comparison on the paired test sets UIEB, EUVP, and LSUI. Table 2Comparison of no-reference underwater image quality metrics UIQM and UCIQE on the UIEB, EUVP, and LSUI datasets. Higher values indicate better perceptual underwater image quality.MethodUIEBEUVPLSUIUIQMUCIQEUIQMUCIQEUIQMUCIQEROP242.350.5122.480.5312.310.496ADPCC162.310.4982.420.5202.250.478URSCT192.580.5412.730.5602.550.528USUIR322.950.5983.100.6032.880.590GUPDM333.060.6023.180.6073.020.598TUDA343.100.6063.220.6103.070.603U-Transformer102.820.5753.050.5922.780.570SFGNet353.010.5943.160.6012.980.589Ours3.180.6103.190.6123.120.607 Comparison of no-reference underwater image quality metrics UIQM and UCIQE on the UIEB, EUVP, and LSUI datasets. Higher values indicate better perceptual underwater image quality. As shown in Table 2, HDAMS-Net achieves the highest UIQM and UCIQE scores across all evaluated datasets, indicating superior perceptual enhancement quality in terms of colorfulness, contrast, and underwater visual clarity. In particular, the proposed method consistently improves both underwater image sharpness and chromatic consistency, demonstrating that the hybrid-domain attention design effectively restores visually natural underwater scenes while preserving structural details.Table 3Comparison of model efficiency in terms of FLOPs, parameters, and inference time per frame.MethodFLOPs (G)Parameters (M)ROP24--ADPCC16--URSCT1936.2211.26USUIR3229.620.23GUPDM33191.601.49TUDA34170.862.73U-Transformer105.9622.82SFGNet35163.161.30Ours13.691.84 Table 3Comparison of model efficiency in terms of FLOPs, parameters, and inference time per frame.MethodFLOPs (G)Parameters (M)ROP24--ADPCC16--URSCT1936.2211.26USUIR3229.620.23GUPDM33191.601.49TUDA34170.862.73U-Transformer105.9622.82SFGNet35163.161.30Ours13.691.84 Comparison of model efficiency in terms of FLOPs, parameters, and inference time per frame. As shown in Table 3. In terms of computational cost, HDAMS-Net requires only 13.69G FLOPs and 1.84M parameters, which is significantly more efficient than many recent transformer-based approaches such as GUPDM (191.60G/1.49M), TUDA (170.86G/2.73M), and SFGNet (163.16G/1.30M). Despite operating with an order of magnitude lower FLOPs, our model consistently surpasses these methods in all major quality metrics. This suggests that the performance gains are likely attributed to the effectiveness of the proposed architectural design rather than increased computational complexity. Moreover, compared with lightweight models such as USUIR (29.62G/0.23M) and U-Transformer (5.96G/22.82M), HDAMS-Net achieves a substantially better balance between performance and efficiency. While maintaining a moderate parameter size, it improves stability and image clarity across diverse underwater conditions. These findings confirm that HDAMS-Net offers a superior trade-off between restoration accuracy and computational efficiency, making it well-suited for real-time or resource-constrained deployment scenarios. Qualitative comparison Visual comparisons are presented in Fig. 6 for representative scenes from the LSUI and UIEB datasets. Conventional methods such as ADPCC fail to handle strong color casts, often producing over-enhanced or unevenly illuminated results. Transformer-based models such as URSCT offer good global contrast but lose fine local details. In contrast, HDAMS-Net produces visually pleasing results with natural color tones, sharp edges, and consistent illumination. The integration of frequency-spatial attention enables global color correction, while the CAP and MSC modules effectively preserve structural details and texture continuity. These visual results align with quantitative findings, demonstrating the superiority of HDAMS-Net in both perceptual quality and visual realism. Our HDAMS-Net achieves visually superior results compared with state-of-the-art approaches. It restores natural color balance, eliminates greenish tones, and enhances fine textures while maintaining illumination consistency across various underwater scenes.Fig. 6Qualitative comparison of underwater image enhancement results on representative samples from LSUI and UIEB datasets. Our HDAMS-Net produces more natural color tones, improved contrast, and finer structural details compared to existing methods. Qualitative comparison of underwater image enhancement results on representative samples from LSUI and UIEB datasets. Our HDAMS-Net produces more natural color tones, improved contrast, and finer structural details compared to existing methods. Ablation study To thoroughly evaluate the effectiveness of each proposed component in HDAMS-Net, we conducted a series of ablation studies focusing on the contributions of the FSA module, the CAP encoder-decoder, and the MSC bottleneck block. All experiments were performed on the LSUI datasets, and the average results across these datasets are reported in Table 4. We constructed the baseline starting from a backbone U-Net with the same encoder-decoder depth and number of layers as HDAMS-Net but without any of the proposed modules. This plain U-Net serves as the fundamental benchmark for quantitative comparison.Table 4Quantitative ablation study on the contribution of each proposed module. The baseline is a plain U-Net with identical depth to HDAMS-Net.Model VariantFSACAPMSCPSNRSSIMLPIPSBaseline (U-Net)✗✗✗22.480.7900.182+ FSA\checkmark✗✗23.910.8210.164+ FSA + CAP\checkmark\checkmark✗24.330.8530.152+ FSA + CAP + MSC (Full)\checkmark\checkmark\checkmark25.590.8970.145 Table 4Quantitative ablation study on the contribution of each proposed module. The baseline is a plain U-Net with identical depth to HDAMS-Net.Model VariantFSACAPMSCPSNRSSIMLPIPSBaseline (U-Net)✗✗✗22.480.7900.182+ FSA\checkmark✗✗23.910.8210.164+ FSA + CAP\checkmark\checkmark✗24.330.8530.152+ FSA + CAP + MSC (Full)\checkmark\checkmark\checkmark25.590.8970.145 Quantitative ablation study on the contribution of each proposed module. The baseline is a plain U-Net with identical depth to HDAMS-Net. To further evaluate the trade-off between restoration performance and computational efficiency, we conduct a detailed complexity analysis on different ablation variants of the proposed HDAMS-Net. As shown in Table 5, the introduction of additional modules progressively increases the model complexity in terms of parameters, FLOPs, and inference latency; however, these increases remain moderate and well-controlled throughout all variants. In particular, the FSA module brings noticeable improvements in PSNR and SSIM with only a limited increase in computational cost, demonstrating high enhancement efficiency. The subsequent CAP and MSC modules further improve structural fidelity and perceptual quality while introducing relatively small additional overhead. Although the full model is slightly larger and slower than the baseline U-Net, it achieves substantially better restoration performance with only marginal increases in complexity. These results indicate that the incremental performance gains consistently outweigh the added computational burden, confirming that the proposed hybrid-domain design achieves an effective balance between enhancement quality and deployment efficiency for real-time underwater vision applications.Table 5Complexity analysis of different model variants in the ablation study, including parameter count, FLOPs, and inference latency.Model VariantParameters (M)FLOPs (G)Latency (ms)Baseline (U-Net)0.929.8413.9+ FSA1.2111.3716.3+ FSA + CAP1.5212.4818.1+ FSA + CAP + MSC (Full)1.8413.6921.0 Table 5Complexity analysis of different model variants in the ablation study, including parameter count, FLOPs, and inference latency.Model VariantParameters (M)FLOPs (G)Latency (ms)Baseline (U-Net)0.929.8413.9+ FSA1.2111.3716.3+ FSA + CAP1.5212.4818.1+ FSA + CAP + MSC (Full)1.8413.6921.0 Complexity analysis of different model variants in the ablation study, including parameter count, FLOPs, and inference latency. To further validate the effectiveness of the proposed composite loss function, we conduct an additional ablation study on different loss configurations. Specifically, we evaluate the influence of the Lab color consistency loss and structural similarity loss, as well as the sensitivity of the balancing coefficient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _2$$\end{document}. The results are reported in Table 6. The experimental results demonstrate that removing the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal {L}}_{Lab}$$\end{document} term leads to noticeable degradation in color consistency and overall reconstruction quality, indicating that Lab-space supervision effectively improves chromatic restoration under complex underwater conditions. Similarly, excluding the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal {L}}_{SSIM}$$\end{document} term reduces structural fidelity and perceptual consistency, resulting in lower SSIM and higher LPIPS values. Furthermore, increasing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _2$$\end{document} from 0.5 to 1.0 slightly decreases overall performance, suggesting that excessively strong color-space supervision may negatively affect the balance between pixel fidelity and structural preservation. Overall, the adopted loss configuration achieves the best trade-off among reconstruction accuracy, perceptual quality, and optimization stability.Table 6Ablation study on different loss function configurations conducted on the LSUI dataset.Loss ConfigurationPSNRSSIMLPIPSFull Loss (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _1$$\end{document}=1.0, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _2$$\end{document}=0.5, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _3$$\end{document}=0.2, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _4$$\end{document}=0.1)25.590.8970.145w/o \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal {L}}_{Lab}$$\end{document}24.910.8720.161w/o \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal {L}}_{SSIM}$$\end{document}25.020.8610.169\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _2 = 1.0$$\end{document}25.210.8830.154 Table 6Ablation study on different loss function configurations conducted on the LSUI dataset.Loss ConfigurationPSNRSSIMLPIPSFull Loss (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _1$$\end{document}=1.0, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _2$$\end{document}=0.5, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _3$$\end{document}=0.2, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _4$$\end{document}=0.1)25.590.8970.145w/o \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal {L}}_{Lab}$$\end{document}24.910.8720.161w/o \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathcal {L}}_{SSIM}$$\end{document}25.020.8610.169\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\lambda _2 = 1.0$$\end{document}25.210.8830.154 Ablation study on different loss function configurations conducted on the LSUI dataset. The visual ablation results in Fig. 7 further demonstrate the effectiveness of each proposed component. The baseline U-Net produces limited enhancement with color bias and blurred textures. Adding the FSA module improves global color balance and contrast, while the CAP module enhances structural clarity. The full HDAMS-Net achieves the most natural color restoration and fine-grained texture recovery, confirming the synergistic benefits of all modules.Fig. 7Visual ablation comparison on representative underwater scenes. Progressive integration of FSA, CAP, and MSC modules gradually improves global color correction, contrast, and texture detail and collectively contribute to the final results of HDAMS-Net. Visual ablation comparison on representative underwater scenes. Progressive integration of FSA, CAP, and MSC modules gradually improves global color correction, contrast, and texture detail and collectively contribute to the final results of HDAMS-Net. We started with a baseline U-Net architecture that maintains the same encoder-decoder structure and number of layers as our full model but without the proposed modules. Subsequently, we progressively integrated the designed components to analyze their individual and combined effects. First, replacing the baseline convolutional blocks with the CAP encoder-decoder led to a noticeable improvement in PSNR and SSIM, demonstrating the efficiency of the axial depthwise convolution and channel attention in capturing directional and inter-channel dependencies. Next, the inclusion of the FSA module further enhanced perceptual quality, highlighting the advantage of jointly exploiting spatial- and frequency-domain information for feature refinement. Finally, incorporating the MSC block into the bottleneck stage resulted in the most significant performance gain, as its multi-scale depthwise convolutions provided a wider receptive field with negligible computational overhead. The ablation experiment shows that the contribution of each module to the final performance is positive. The synergistic combination of frequency-spatial attention, axial convolution, and multi-scale aggregation enables HDAMS-Net to achieve superior restoration quality while maintaining high efficiency and low model complexity. Generalization discussion To further assess the generalization capability of HDAMS-Net across diverse underwater conditions, we analyze its behavior from three perspectives: cross-dataset performance, sensitivity to water type variation, and robustness under challenging degradation patterns. As reported in Table 1, HDAMS-Net achieves consistent improvements over competing methods across all three benchmark datasets-UIEB, EUVP, and LSUI-which differ substantially in image resolution, water type, and degradation severity. UIEB contains real-world images with diverse and unpredictable color casts, EUVP covers both structured and unstructured underwater scenes captured at varying depths, and LSUI comprises large-scale real scenes with mixed illumination and turbidity levels. The fact that HDAMS-Net ranks first or competitively on all three datasets without dataset-specific tuning indicates that the proposed modules generalize effectively across different underwater imaging conditions. The FSA module is designed to address the two dominant degradation modes in underwater imagery: global color attenuation, which manifests as low-frequency spectral imbalance, and local texture blurring, which is a spatial-domain artifact. By explicitly decoupling these two degradation types into frequency and spatial processing branches, HDAMS-Net avoids the common failure mode of methods that apply a uniform enhancement strategy across all degradation types. This architectural choice contributes to the model’s stable performance across greenish, bluish, and turbid underwater scenes, as observed in the qualitative comparisons in Fig. 6. Despite these strengths, HDAMS-Net is trained and evaluated on paired datasets, and its performance under severe domain shift-for instance, when applied to deep-sea footage captured by novel sensor types not present in the training data-has not been fully characterized. Additionally, the current model does not explicitly model temporal consistency, which may limit its performance on underwater video sequences where flickering artifacts can arise. Future work will investigate lightweight domain adaptation and knowledge distillation strategies to extend the model’s robustness to these more challenging deployment scenarios. Cross-dataset generalization study To further evaluate the robustness and real-world applicability of HDAMS-Net, we conduct cross-dataset generalization experiments under two challenging settings:UIEB → LSUI: The model is trained on UIEB and tested directly on LSUI without fine-tuning.EUVP → UIEB: The model is trained on EUVP and evaluated on UIEB. UIEB → LSUI: The model is trained on UIEB and tested directly on LSUI without fine-tuning. UIEB → LSUI: The model is trained on UIEB and tested directly on LSUI without fine-tuning. EUVP → UIEB: The model is trained on EUVP and evaluated on UIEB. EUVP → UIEB: The model is trained on EUVP and evaluated on UIEB. These settings simulate real-world deployment scenarios where training and testing distributions differ significantly in terms of water type, illumination conditions, and image resolution. We compare HDAMS-Net with three representative methods: USUIR32, GUPDM33, and TUDA34. These methods are selected due to their strong performance in both CNN-based and Transformer-based underwater image enhancement paradigms.Table 7Cross-dataset generalization performance comparison (PSNR/SSIM). Models are trained on one dataset and directly tested on another without fine-tuning.MethodUIEB → LSUIEUVP → UIEBPSNRSSIMPSNRSSIMUSUIR3219.820.78117.950.792GUPDM3320.630.81218.840.823TUDA3420.910.82419.120.835Ours21.180.84120.450.832 Table 7Cross-dataset generalization performance comparison (PSNR/SSIM). Models are trained on one dataset and directly tested on another without fine-tuning.MethodUIEB → LSUIEUVP → UIEBPSNRSSIMPSNRSSIMUSUIR3219.820.78117.950.792GUPDM3320.630.81218.840.823TUDA3420.910.82419.120.835Ours21.180.84120.450.832 Cross-dataset generalization performance comparison (PSNR/SSIM). Models are trained on one dataset and directly tested on another without fine-tuning. The results in Table 7 demonstrate that HDAMS-Net consistently outperforms all competing methods under cross-dataset evaluation settings. When trained on UIEB and tested on LSUI, HDAMS-Net achieves a significant improvement in both PSNR and SSIM, indicating strong generalization to large-scale real-world underwater scenes with more complex degradation patterns. Similarly, when trained on EUVP and evaluated on UIEB, HDAMS-Net maintains superior performance, suggesting that the proposed hybrid-domain attention mechanism effectively learns transferable representations that are not overfitted to a specific dataset distribution. These results confirm that explicitly modeling frequency-domain color distortion and spatial-domain texture degradation improves robustness under domain shifts, making HDAMS-Net suitable for real-world underwater deployment scenarios where training and testing environments differ significantly. Conclusion In this paper, we proposed HDAMS-Net, a lightweight hybrid-domain attention network for underwater image enhancement. The model integrates the Frequency-Spatial Attention, Channel Attention Pointwise, and Multi-Shape Convolution modules within an encoder-decoder framework to effectively capture global context, inter-channel relationships, and fine structural details. HDAMS-Net efficiently suppresses color distortion and haze while preserving natural textures and illumination balance. Extensive experiments conducted on multiple real-world underwater datasets demonstrate that our method achieves competitive quantitative metrics and favorable visual quality compared to existing state-of-the-art approaches, while maintaining a compact model size and fast inference speed. HDAMS-Net demonstrates consistent performance across multiple real-world benchmark datasets covering diverse water types and degradation conditions, confirming its generalization capability beyond any single training distribution. The explicit decoupling of frequency-domain color correction and spatial-domain texture refinement within the FSA module contributes to this robustness by allowing the network to adapt its enhancement strategy to the dominant degradation mode present in each scene. These properties make HDAMS-Net a viable candidate for real-time deployment on autonomous underwater vehicles (AUVs) and remotely operated platforms. In future work, we plan to further enhance the generalization of HDAMS-Net by incorporating lightweight model compression, knowledge distillation, and cross-domain adaptation, as well as extending the framework to handle underwater video enhancement with temporal consistency constraints.