LPD-Net: lightweight adaptive weight-driven dual-stream network guided by physics-inspired priors for structure-preserving underwater image enhancement

1 Introduction

Reliable optical sensing is the cornerstone of Autonomous Underwater Vehicles (AUVs), particularly in high-precision missions such as seabed resource exploration and the health monitoring of hydraulic infrastructures (Anwar and Li, 2020; Jian et al., 2021). Unlike terrestrial imaging, underwater environments present a complex optical medium where light undergoes wavelength-selective attenuation and volumetric scattering. These phenomena result in a dominant blue-greenish color cast and a loss of contrast, often referred to as the “hazing effect.” Due to the extremely irregular non-linear degradation of underwater images, finding a stable balance between color restoration and texture enhancement remains a persistent challenge (Qian et al., 2024). For engineering applications such as micro-concrete crack detection (Qi et al., 2022), these degradations are particularly detrimental as they obliterate the fine-grained geometric and structural details essential for reliable defect quantification. Furthermore, robust and lightweight enhancement algorithms are increasingly demanded as pre-processing modules for high-fidelity underwater imaging systems deployed on embedded platforms (Liu et al., 2024).

To address these visibility issues, enhancement algorithms have generally evolved into two paradigms: physical model-based methods and data-driven deep learning (DL) approaches. Traditional physical methods attempt to invert the optical imaging model to stretch contrast and recover color. These are typically grounded in Retinex theory (Land, 1977), which serves as the foundation for the widely utilized Multi-Scale Retinex with Color Restoration (MSRCR) algorithm (Jobson et al., 1997). Following the Dark Channel Prior (DCP) (He et al., 2011), various modifications have been proposed to improve restoration, including light absorption modeling (Peng and Cosman, 2017), transmission map estimation (Drews et al., 2013), and multi-scale fusion strategies (Ancuti et al., 2018). More recently, advanced physical regularization schemes, such as the Rank-One Prior (ROP) (Liu et al., 2023), have been developed to achieve unified, real-time physical scene recovery across various scattering media. Additionally, non-physical methods like CLAHE (Zuiderveld, 1994) are employed to optimize local contrast through histogram equalization. However, these approaches often suffer from inherent side effects, such as the significant amplification of high-frequency noise in low-light regions (Peng and Cosman, 2017). Essentially, a fundamental conflict exists between global chromatic restoration and local structural preservation, where the excessive compensation of physical signals often introduces irreversible noise artifacts.

In response, deep learning methods have achieved remarkable progress. Lightweight architectures like LPS-Net (Li et al., 2020) have been developed for resource-constrained platforms, alongside established benchmarks such as the UIEB dataset (Jiang et al., 2023a). Frameworks including generative architectures like FUnIE-GAN (Islam et al., 2020) and multi-color space embedding schemes like U-Color (Li et al., 2021) have delivered high scores in pixel-wise metrics. To explicitly integrate physical properties into deep representations, wavelength-attributed networks, such as Deep-WaveNet (Sharma et al., 2023), have been introduced to reconstruct precise spectral channels. Recent architectures have explored homology perspectives (Fu et al., 2022), zero-reference curve estimation (Guo et al., 2020), and Transformer-based global modeling (Peng et al., 2023). Feature extraction is further enhanced by polymorphic kernels in UIR-Net (Guo et al., 2025), semi-supervised learning strategies (Huang et al., 2023), and extremely lightweight network designs (Jiang et al., 2023b). Furthermore, joint optimization frameworks, such as Joint-ID (Yang et al., 2024), have been proposed to address both illumination decay and chromatic casts simultaneously. Additionally, a physical imaging model embedded architecture, designated as DPF-Net (Mei et al., 2025), has been recently developed to integrate deterministic optical constraints into data-driven frameworks for improved restoration reliability. Nevertheless, existing physics-embedded networks often treat physical priors as static inputs without considering their spatial reliability across diverse water types. This lack of an adaptive signal selection mechanism can lead to over-smoothed results or the propagation of prior-induced artifacts, which is unacceptable for applications requiring precise structural analysis.

To address these limitations, a Lightweight Physics-Inspired Dual-Stream Network (LPD-Net) is proposed for structure-preserving enhancement, following a motivation-driven paradigm that explicitly maps physical degradations to specific neural operators. The task is decomposed into two distinct stages to ensure robust restoration. Firstly, an Adaptive Noise-Gated Fusion (ANGF) module is introduced to dynamically regulate the contribution of the physics-inspired prior stream. Instead of using rigid concatenation, it is recognized that dynamic weighting strategies are highly effective for handling complex underwater light attenuation (Awan and Mahmood, 2024). Inspired by the success of attention mechanisms in feature selection and contextual modeling (Woo et al., 2018; Guo et al., 2022), the proposed ANGF module implicitly perceives scene degradation and adaptively modulates the influence of physics-inspired priors. Rather than treating physical enhancement as a fixed preprocessing operation, the proposed framework jointly optimizes physics-inspired priors within an end-to-end restoration architecture, enabling adaptive suppression of prior-induced artifacts while preserving structural boundaries. Secondly, a CNN refinement module based on polymorphic large kernels is employed (Guo et al., 2025). By integrating global residual learning (He et al., 2016) and advanced optimization via AdamW (Loshchilov and Hutter, 2017), the network focuses on suppressing noise and reconstructing geometric edges rather than synthesizing images from scratch.

To optimize the enhancement process, the proposed two-stage physics-inspired framework combines the contrast restoration advantages of physical models with the robust denoising capabilities of CNNs. Specifically, the adaptive fusion strategy acts as an intelligent “noise gate” to prevent the propagation of prior-induced artifacts. To further enforce structural preservation, a composite loss function incorporating Laplacian-based supervision (Lai et al., 2017) is utilized during the training phase. Extensive experiments demonstrate that the proposed method significantly outperforms state-of-the-art (SOTA) methods across standard metrics, including PSNR, SSIM (Horé and Ziou, 2010; Wang et al., 2004), UCIQE (Yang and Sowmya, 2015), UIQM (Panetta et al., 2016), and perceptual similarity (LPIPS) (Zhang et al., 2018). These results validate the effectiveness of the proposed approach for structure-preserving underwater imaging, ensuring that fine-grained geometric features remain intact under diverse degradation conditions. Overall, the proposed framework establishes a degradation-aware restoration paradigm that explicitly bridges underwater optical physics and adaptive neural reconstruction, enabling simultaneous optimization of chromatic fidelity, structural preservation, and computational efficiency. The main contributions of this work are summarized as follows:

(1) A novel physics-inspired dual-stream framework is proposed to jointly optimize chromatic signal restoration and local structural preservation. By decoupling color guidance and texture retention, the framework effectively alleviates the long-standing trade-off between visual aesthetics and structural fidelity in underwater imaging.

(2) An adaptive noise-gated fusion mechanism is designed to dynamically regulate the influence of physics-inspired priors. By functioning as an intelligent signal selector, it adaptively suppresses prior-induced artifacts in high-visibility regions while leveraging color guidance in turbid scenarios, thereby safeguarding the clarity of structural boundaries.

(3) A lightweight and resolution-agnostic architecture is developed with only 1.84 M parameters. Unlike existing fixed-resolution models that rely on mandatory resizing operations, the proposed framework preserves high-frequency structural details directly on high-resolution sensor data, thereby avoiding interpolation-induced structural degradation and enabling efficient structure-aware underwater imaging.

The remainder of this paper is organized as follows. Section 2 presents the proposed physical foundation and network methodology. Section 3 describes the experimental settings, followed by a comprehensive presentation of quantitative and qualitative evaluations. A detailed analysis regarding computational efficiency and engineering deployment is provided in Section 4. Finally, the conclusions of this work are summarized in Section 5.

2 Proposed methodology

2.1 Physical foundation and problem formulation

The physical process of underwater image degradation serves as the foundational basis for the proposed network architecture. According to the simplified Jaffe-McGlamery model, the observed intensity I(x) at pixel x is formulated in Equation 1 as:

I(x)=J(x)·t(x)+B·(1-t(x))(1)

where J(x) represents the clear scene radiance to be recovered, B denotes the global background light, and t(x) is the medium transmission map. As illustrated in Figure 1, the degradation is primarily governed by wavelength-dependent absorption, forward scattering, and backscattering.

Specifically, chromatic signals are non-uniformly attenuated due to absorption, leading to the dominant greenish or bluish color casts observed in Figure 1b. Forward scattering induces small-angle diffusion of light, which acts as a low-pass filtering process that blurs geometric contours and reduces resolution (Figures 1c, e). Furthermore, backscattering introduces a global veiling light B that significantly degrades the signal-to-noise ratio (SNR) and obscures structural details (Figures 1d, f).

From a signal processing perspective, these combined effects create a fundamental conflict between global color restoration and local structure preservation. Consequently, a single-stream recovery paradigm often fails to simultaneously rectify spectral biases and reconstruct high-frequency textures without amplifying background noise. Rather than explicitly solving the ill-posed physical parameters t(x) and B from a single degraded input, the design of the proposed network focuses on utilizing a physics-inspired prior to guide the neural operators, thereby decoupling chromatic restoration and geometric edge preservation.

2.2 Design motivation and physical mapping

Guided by the physical analysis in Section 2.1, the proposed architecture is developed following a motivation-driven paradigm. Instead of adopting a black-box approach, a direct mapping is established between specific physical degradation factors and targeted network components, as summarized in Table 1.

Table 1. Mapping from physical degradations to network components.

Instead of explicitly solving the Jaffe-McGlamery imaging equation through parameterized inversion, we employ the MSRCR algorithm as a physically-motivated color correction prior. Although MSRCR does not estimate the transmission map t(x) or background light B, its multi-scale illumination-reflection decomposition partially compensates for wavelength-dependent attenuation and uneven ambient lighting. Therefore, it is incorporated as a degradation-aware empirical prior rather than a strict optical inversion model.

Based on the analytical mapping in Table 1, the proposed network is designed to explicitly decouple the restoration of global radiance and local geometry. Specifically, a dual-stream paradigm is introduced where the raw stream serves as a “fidelity anchor” to retain original statistical distributions, while the prior stream provides “physics-inspired prior” for chromatic correction.

Central to this architecture is the adaptive noise-gated fusion mechanism. The weights γ and δ are not static parameters but are jointly optimized with the reconstruction backbone through an end-to-end training process. This ensures that the physical-prior serves as a dynamic guidance rather than a fixed pre-processing step, enabling the network to perceive the reliability of physical knowledge across diverse water types. The bridge between the two streams–the adaptive fusion module–functions as a signal selection mechanism that prevents the propagation of scattering-induced noise and artifacts from the physical prior into the final reconstruction. This rigorous alignment between physical principles and neural operators ensures that the enhancement process remains both high-fidelity and physically interpretable. The analytical mapping from physical degradation to network components is further elaborated as follows:

(1) Absorption: The chromatic signal energy loss caused by wavelength-dependent absorption is addressed via the physics-inspired prior branch, where the MSRCR algorithm is utilized to compensate for global color attenuation and restore the spectral balance.

(2) Forward scattering: The diffusion of high-frequency components and resultant blurring induced by forward scattering are specifically mapped to the Hybrid Domain Attention (HDA) and Large Kernel Attention (LKA) modules. These components are dedicated to reconstructing the lost geometric structures and sharpening object boundaries.

(3) Backscattering: The veiling light and feature noise introduced by backscattering are mitigated through the adaptive weighting module. By functioning as a “noise gate,” this mechanism ensures that background noise embedded in the physical prior is adaptively suppressed, thereby preventing the corruption of the raw information stream during the fusion process.

Based on the mapping relationship established in Table 1, a Lightweight physics-inspired prior Guided Dual-Stream Network (LPD-Net) is developed to achieve structure-preserving restoration. The schematic representation of the proposed architecture is depicted in Figure 2. The framework adopts a cascaded two-stage optimization strategy to decouple the restoration of chromatic signals and geometric textures.

In the first stage, designated as the Adaptive Physical Prior Module (Stage I), an intelligent fusion mechanism is implemented to integrate the raw sensor data with a physically-guided representation. Instead of a rigid pre-processing approach, this stage dynamically modulates the contribution of the physical knowledge based on the inferred scene degradation level. Subsequently, the fused features are transmitted to the CNN Refinement Module (Stage II), which functions as a high-fidelity reconstructor. A U-shaped backbone, enhanced by polymorphic large kernels (Guo et al., 2025), is utilized to suppress residual noise and sharpen structural edges. To ensure that fine-grained geometric details are not lost during the recursive convolution process, a global residual connection is established, anchoring the final output directly to the original raw intensities.

2.3 Stage I: adaptive physical prior module

Stage I is designed to provide a robust chromatic baseline while preventing the introduction of scattering-induced noise. The module consists of three sub-components: physics-inspired prior generation, spatial weight inference, and feature-level noise-gated fusion.

2.3.1 Physical prior generation

The raw underwater observation is denoted as Iraw∈ℝH×W×3. To embed empirical domain knowledge into the network, a physics-inspired prior I_prior is generated via the Multi-Scale Retinex with Color Restoration (MSRCR) algorithm (Jobson et al., 1997). Grounded in Retinex theory (Land, 1977), the core operation for each color channel i∈{R, G, B} is defined as the difference between the logarithmic image and the weighted sum of several Gaussian surround functions in Equation 2:

Ri(x,y)=∑k=1Kwk(logIi(x,y)-log(Fk(x,y)*Ii(x,y)))(2)

where F_*k* represents the Gaussian kernels at varying scales, and w_*k* is the corresponding weight. While I_prior provides a global color baseline to mitigate absorption-induced shifts, it inherently carries backscattering-induced noise that can be detrimental in high-visibility conditions.

2.3.2 Adaptive noise-gated fusion mechanism

To effectively manage the trade-off between chromatic restoration and noise propagation, an adaptive noise-gated fusion mechanism is implemented at the feature level. Rather than utilizing a rigid pixel-level addition, this module is designed to dynamically predict a pixel-wise weighting map to regulate the contribution of the physics-inspired prior.

As illustrated in the top-left of Figure 2, the raw feature F_*raw* and the prior feature F_*prior* are concatenated along the channel dimension. For visual simplicity in the schematic diagram, the generation of the joint feature map is represented by these two primary streams; however, in the formal implementation, the spatial difference feature F_*diff* is implicitly calculated and appended within the concatenation layer to drive the gate generator. The generation of the spatial weight map γ(H, W) is formulated in Equation 3 as:

γ(H,W)=σ(G(Concat(Fraw,Fprior,Fdiff)))(3)

where G(·) denotes the sequential convolutional layers of the gate generator, consisting of a primary 3 × 3 convolution with a ReLU activation to extract localized spatial features from the concatenated channels, followed by a secondary 3 × 3 convolution to compress the features and map them into a single-channel spatial weight representation. Additionally, σ(·) denotes the Sigmoid activation function that constrains each pixel value within [0, 1] to produce the final normalized gating coefficients.

The resulting 2D weight map γ(H, W) and its spatial complement 1 − γ(H, W) represent the confidence levels for the raw and prior features, respectively. The final fused representation F_fused is subsequently computed through a pixel-wise weighted linear combination in Equation 4:

Ffused= γ·Fraw+(1− γ)·Fprior(4)

This formulation enables the network to dynamically regulate the influence of the physical prior according to localized scene degradation levels. Specifically, in severely degraded or highly turbid regions, smaller γ values (and thus larger 1−γ values) are assigned to enhance the chromatic compensation capability of the network. Conversely, in relatively clear regions where the physical prior may contain prior-induced artifacts, the gate adaptively increases the weight γ. This noise-gating property effectively suppresses prior-induced distortions and ensures that original structural boundaries are preserved from the raw signal stream. By optimizing γ through end-to-end joint training, the fusion layer functions as an intelligent signal selector, providing a high-fidelity feature base for the subsequent Stage II refinement.

2.4 Stage II: CNN refinement module

Following the signal selection in Stage I, the fused feature representation F_fused is processed by the CNN Refinement Module for high-fidelity structural reconstruction. The backbone architecture is extended from the polymorphic kernel design (Guo et al., 2025), specifically optimized for multi-scale feature extraction in complex underwater environments. Three specialized signal-processing components are integrated to ensure robust restoration:

(1) Hybrid domain attention (HDA): Features are analyzed simultaneously in both the spatial and frequency domains via the Fast Fourier Transform (FFT). This dual-domain processing allows the network to effectively separate low-frequency background haze (signal bias) from high-frequency structural textures. By performing spectral filtering, the module enhances the global signal clarity while suppressing scattering-induced noise.

(2) Large kernel attention (LKA): Long-range pixel dependencies are established by decomposing large kernels into depth-wise and dilated convolutions. This mechanism acts as an adaptive scale-aware filter, effectively correcting global color shifts and non-uniform illumination that conventional local convolutions fail to address.

(3) Composite shape convolution (CSC): Non-square kernels (horizontal, vertical, and square) are employed in the decoder to precisely reconstruct the irregular geometric structures of underwater objects, such as concrete cracks or coral reefs. This geometric-aware design ensures that the anisotropy of structural edges is preserved during the feature-to-image mapping process.

The final output is generated by learning the residual relative to the raw input, formulated in Equation 5 as:

Ioutput=σ(Nstage2(Ffused)+Iraw)(5)

where Nstage2 represents the mapping function of the refinement network and σ denotes the Sigmoid activation. By anchoring the reconstruction to I_raw within the Sigmoid function, the refinement stage focuses its capacity on correcting color signal biases while maintaining the inherent integrity of original structural boundaries without pixel overflow.

2.5 Optimization objective

To achieve an optimal balance between pixel-level accuracy, geometric integrity, and visual vibrancy, a multi-term composite loss function Ltotal is constructed. Following the optimization framework established by Guo et al. (2025), the basic objective incorporates pixel-wise fidelity, structural similarity, and underwater color quality. To enforce our structure-preserving constraints, we further augment this formulation with a dedicated Laplacian boundary regularization term, yielding the total joint loss in Equation 6:

Ltotal=Lpixel+λ1Lssim+λ2Luciqe+λ3Llap(6)

where λ₁ = 0.2, λ₂ = 0.01, and λ₃ = 0.1 are empirically determined balancing hyper-parameters. The individual loss terms are defined as follows:

(1) Structural preservation loss (Lpixel): The smooth L₁ loss (also known as Huber loss) (Lai et al., 2017) is employed to measure pixel-level accuracy. For a residual x = I_output(p)−I_gt(p) at pixel p, the loss is formulated in Equation 7 as:

Lpixel(x)={0.5x2,if|x|<1|x|-0.5,otherwise(7)

Compared to the standard L₂ loss, a more stable gradient is provided, preventing gradient explosion for large errors while ensuring smooth convergence for small errors, which demonstrates superior robustness against outlier noise (e.g., sensor noise in Figure 1f).

(2) Structural loss (Lssim): To maximize the geometric consistency between the enhanced result and the ground truth, an SSIM-based loss (Wang et al., 2004) is utilized, defined as 1 − SSIM(I_output, I_gt). This term enforces the retention of structural similarity, ensuring that fine-grained edges remain sharp.

(3) Metric-aware quality loss (Luciqe): Diverging from standard practices where quality metrics are used solely for post-hoc evaluation, the UCIQE measure (Yang and Sowmya, 2015) is explicitly integrated into the training objective. It is formulated as Luciqe=1-UCIQE(Ioutput), where the UCIQE metric is defined in Equation 8 as:

UCIQE=c1·σc+c2·Ccon+c3·μs(8)

where σ_*c* is the standard deviation of chroma (a^*, b^* components in CIELAB space), C_con is the contrast of luminance (L^*), and μ_*s* is the average of saturation. The coefficients are empirically set as c₁ = 0.4680, c₂ = 0.2745, and c₃ = 0.2576. By minimizing this term, the network is guided to actively optimize the distribution of color saturation and contrast, ensuring alignment with human visual perception and underwater-specific quality standards.

(4) Laplacian boundary loss (Llap): To explicitly enforce boundary preservation and satisfy the structure-preserving constraint, a Laplacian loss is introduced to supervise edge alignment in the derivative space. It is formulated in Equation 9 as:

Llap=∥Δ(Ioutput)-Δ(Igt)∥1(9)

where Δ(·) denotes the standard discrete Laplacian operator, which acts as a high-pass filter to extract second-order spatial gradients. By minimizing the L₁ distance between the Laplacian-filtered maps, the network is penalized for blurring fine-grained geometric transitions.

3 Experimental results

3.1 Experimental settings

The performance of the proposed framework was evaluated on two widely recognized large-scale benchmarks: the Underwater Image Enhancement Benchmark (UIEB) (Li et al., 2020) and the Large-Scale Underwater Image (LSUI) dataset (Peng et al., 2023). The UIEB dataset contains 890 paired images covering diverse water types, including green, blue, and turbid scenes. Following the standard split, the first 800 pairs were utilized for training, while the remaining 90 pairs were reserved for testing. The LSUI dataset comprises 4,279 real-world deep-sea image pairs characterized by low light and uniform color casts. To verify the model's robustness in massive underwater environments, 3,850 pairs were employed for training and 429 pairs for testing.

The proposed network was implemented using the PyTorch framework and trained on NVIDIA GeForce RTX series GPUs. The AdamW optimizer (He et al., 2016) was utilized with an initial learning rate of 2 × 10⁻⁴ and a minimum learning rate of 1 × 10⁻⁶, employing a cosine annealing strategy for learning rate decay. To guarantee deterministic reproducibility, the random seed was standardized to 3,407 across all experiments, following the protocol in Guo et al. (2025). The training patch size was set to 256 × 256 with a batch size of 2 due to physical hardware constraints. To ensure complete convergence, the training process was consistently executed for 500 epochs on both the UIEB and LSUI datasets.

To provide a fair and standardized engineering comparison, the computational complexity metrics, frames per second (FPS), and inference latency of all SOTA models were benchmarked on a single NVIDIA GeForce RTX 3050 GPU at a fixed resolution of 256 × 256. Unless otherwise specified, all quantitative results, including the ablation study, are re-evaluated offline using a unified evaluation script. The reported metrics (PSNR, SSIM, LPIPS, and UCIQE) are consistently computed from the saved restored PNG images under identical mathematical formulations to eliminate potential discrepancies in on-the-fly metric implementations across different frameworks.

To provide a multi-dimensional assessment, standard fidelity metrics including Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) (Horé and Ziou, 2010; Wang et al., 2004) were employed. Additionally, Learned Perceptual Image Patch Similarity (LPIPS) (Zhang et al., 2018) was utilized to evaluate perceptual quality, while the Underwater Color Image Quality Evaluation (UCIQE) (Yang and Sowmya, 2015) was adopted to measure color vibrancy and contrast.

3.2 Quantitative and qualitative evaluation on full-resolution images

A comprehensive comparison is made between the proposed LPD-Net and five state-of-the-art (SOTA) approaches, including physical methods [CLAHE (Zuiderveld, 1994), Pure-MSRCR (Jobson et al., 1997)] and deep learning methodologies [Zero-DCE (Guo et al., 2020), FUnIE-GAN (Islam et al., 2020), Joint-ID (Yang et al., 2024), Deep-WaveNet (Sharma et al., 2023), and UIR-Baseline (Guo et al., 2025)]. The quantitative results evaluated on native full-resolution images are summarized in Table 2. For visual and layout clarity, six representative methods (including the newly added Deep-WaveNet) are visualized in Figure 3, while the complete numerical benchmarks of all seven compared methods (including Joint-ID) are detailed in Table 2.

Table 2. Quantitative comparison of representative methods on UIEB and LSUI datasets (full-resolution).

The bold values indicate the best performance for each metric.

As presented in Table 2, the proposed LPD-Net achieves dominant performance on the UIEB dataset. Specifically, a significant leap in PSNR is observed, reaching 22.4800 dB, which represents a 0.29 dB improvement over the competitive UIR-Baseline (22.1921 dB). More importantly, the SSIM index is improved to 0.9146, indicating a superior capability in reconstructing fine geometric structures. Although the baseline model yields a slightly lower LPIPS of 0.1267, the proposed method maintains an extremely competitive perceptual score of 0.1334 while substantially enhancing the global contrast and chromatic naturalness.

To facilitate a highly detailed visual inspection of the fine-grained textures and prevent scaling-induced rendering artifacts in the paper layout, localized regions of interest (ROIs) are extracted from the natively processed 1, 920 × 1, 080 inference outputs. Specifically, as presented in Figure 3, a 320 × 320 ROI is extracted for the top row (diver scene) under coordinate offsets of (Δx = −40, Δy = 0), and a 512 × 512 ROI is extracted for the bottom row (reef scene) under coordinate offsets of (Δx = 50, Δy = −100) relative to the respective center of each full-resolution frame. As observed in these results, no structural boundary degradation or blocky artifacts are introduced when processing large-scale images, confirming the spatial translation invariance of our model. While traditional methods like Pure-MSRCR (Jobson et al., 1997) often introduce excessive granular noise and over-exposure, and FUnIE-GAN (Islam et al., 2020) achieves limited contrast enhancement, more natural color distributions and superior edge sharpness are produced by the proposed method. Notably, the red spectrum is effectively recovered without compromising the structural integrity of the background, which aligns with the superior PSNR and SSIM scores reported in Table 2.

In the large-scale LSUI scenarios, the advantages of the proposed dual-stream framework are even more pronounced. While traditional Pure-MSRCR suffers from severe noise amplification (collapsing to a PSNR of 13.9853 dB), the proposed model intelligently suppresses these artifacts through the adaptive spatial gating mechanism. Consequently, the highest PSNR (26.5045 dB) and SSIM (0.8833) are simultaneously achieved on the LSUI dataset, outperforming the UIR-Baseline by 0.67 dB and 0.0112, respectively. This stable improvement across distinct benchmarks demonstrates the robust generalization capability of LPD-Net.

Regarding the non-reference UCIQE metric, although the highest scores are obtained by the empirical Pure-MSRCR (Jobson et al., 1997), this occurs at the severe cost of pixel-saturation and image quality degradation, as reflected by its collapsed fidelity metrics. In contrast, the proposed method focuses on a balanced trade-off where structural fidelity is prioritized over aggressive, unnatural color saturation. While purely physical methods often achieve higher UCIQE scores through excessive contrast stretching–which frequently introduces color distortions–a stable balance between chromatic realism and edge sharpness is maintained by the proposed network. This engineering priority is further validated by the leading SSIM and LPIPS results, which are more critical indicators for downstream visual perception and object localization than superficial colorfulness.

3.3 Quantitative and qualitative evaluation on fixed-resolution images

To ensure a fair and comprehensive evaluation, an additional benchmark is established under a fixed resolution of 256 × 256, as presented in Table 3. The corresponding qualitative assessment under this unified 256 × 256 resolution protocol is presented in Figure 4. This experimental setting incorporates recent state-of-the-art methods that are originally designed for low-resolution inputs, thereby enabling a standardized comparison under consistent computational constraints.

Table 3. Quantitative comparison of state-of-the-art underwater image enhancement methods under a fixed 256 × 256 resizing setting.

The bold values indicate the best performance for each metric.

In comparison with the physics-inspired prior SOTA, DPF-Net (Mei et al., 2025), a significant performance enhancement is consistently observed. Specifically, the proposed LPD-Net achieves a PSNR of 22.7348 dB on UIEB and 26.7107 dB on LSUI, outperforming DPF-Net by 1.15 and 5.63 dB, respectively. This substantial margin is primarily attributed to the resolution-agnostic design of the proposed network, which effectively circumvents the information loss typically introduced by the downsampling-upsampling pipeline inherent in fixed-resolution models.

Furthermore, it is observed that Transformer-based methods, such as U-Shape (Peng et al., 2023), suffer from severe performance degradation under fixed-resolution constraints, yielding a PSNR of only 19.8543 dB on UIEB and 24.6777 dB on LSUI. This phenomenon is driven by the fact that Transformer models tend to induce over-smoothing, leading to the noticeable obliteration of fine-grained structural details, which are critical for precise underwater application. In contrast, our convolution-dominant dual-stream architecture successfully preserves sharp edges and texture boundaries.

In summary, as evidenced by the quantitative results in Table 3, the proposed LPD-Net consistently secures the leading position across all evaluated metrics (PSNR, SSIM, and LPIPS) on both benchmarks. Most notably, LPD-Net outperforms the backbone model, UIR-Baseline (Guo et al., 2025), by a margin of +0.12 dB on UIEB and +0.62 dB on LSUI, while achieving the lowest perceptual distortion (0.1012 and 0.1129 LPIPS). These results validate that superior structural fidelity and perceptual alignment are successfully achieved by the proposed physics-inspired dual-stream architecture, outperforming contemporary SOTA methods in overall restoration quality.

3.4 Robustness analysis on high-resolution image patches

In the patch-wise evaluation presented in Table 4, the highest PSNR values on the UIEB dataset are achieved by the U-Shape model (Peng et al., 2023). This performance characteristic is attributed to the Transformer architecture's extensive parameter capacity, which enables superior localized pixel-level mapping when global resizing distortions are eliminated. However, it is observed that high PSNR scores under this localized protocol do not consistently translate to superior structural accuracy. Despite the higher PSNR attained by U-Shape, significantly lower LPIPS and higher SSIM scores are maintained by the proposed LPD-Net.

Table 4. Patch-wise evaluation using cropped 256 × 256 patches from high-resolution images (without resizing).

The bold values indicate the best performance for each metric.

Most importantly, on the large-scale LSUI dataset (Table 4), the proposed model completely dominates SOTA methods across all metrics, achieving the highest PSNR (26.2864 dB) and SSIM (0.8879) alongside the lowest LPIPS (0.1236). These results indicate that the proposed framework prioritizes the restoration of perceptually relevant structural features over mere pixel-wise intensity matching, ensuring higher fidelity for practical tasks where edge definition is paramount.

A qualitative comparison of the cropped 256 × 256 patches and their corresponding Canny edge maps is provided in Figure 5 to further evaluate structural fidelity under localized conditions. As shown in the top row, different methods exhibit distinct behaviors in local enhancement. Pure-MSRCR and DPF-Net generate visually enhanced outputs but introduce noticeable over-enhancement and noise artifacts. In contrast, UIR-Baseline and Semi-UIR exhibit over-smoothing effects, leading to the loss of fine-grained textures. Although U-Shape produces visually appealing results, structural continuity remains inconsistent. This observation is further corroborated by the Canny edge maps in the bottom row of Figure 5, where Pure-MSRCR and DPF-Net generate dense but noisy edge responses, while the baseline and Semi-UIR produce sparse edges due to over-smoothing. In contrast, the proposed method achieves a balanced representation, preserving clean, continuous, and well-localized structural contours, demonstrating superior robustness in maintaining high-frequency details at the patch level.

Regarding the visual characteristics in Figure 5, a relative increase in brightness is observed in the proposed method under patch-wise processing. This is attributed to the content-adaptive enhancement mechanism, which adjusts the intensity based on local statistics, resulting in improved local contrast and visibility.

While Figure 5 focuses on local robustness at the patch level, further mechanism-level analysis is conducted at the full-image level in the following section.

3.5 Mechanism analysis

To further investigate the underlying mechanism of structural preservation beyond localized patch analysis, a full-resolution case study is conducted.

As illustrated in Figure 6, significant differences are observed in the structural representations at the full-image level. Pure-MSRCR produces a cluttered edge map dominated by high-frequency noise, resulting in numerous spurious edges. In contrast, the Baseline generates sparse and incomplete edge structures due to over-smoothing, failing to capture the global reef contours. The proposed method achieves a more balanced structural representation, preserving both global structural integrity and local texture details. Clear and continuous edge boundaries are extracted, closely aligning with the reference. This observation suggests that global residual learning anchored to the raw input facilitates effective preservation of fine structural details, which is critical for downstream inspection tasks such as crack detection (Qi et al., 2022).

To further provide a quantitative interpretation of the observed visual improvements, the grayscale histogram of the same representative sample (Image 8056) is analyzed, as shown in Figure 7.

The histogram of the Raw image is highly narrow and skewed toward the low-intensity range, corresponding to its hazy and “flat” visual appearance. In contrast, the baseline and the proposed method (red curve) exhibit a significantly expanded dynamic range and a much smoother, more uniform distribution. Crucially, pixel accumulation at the extreme values (0 and 255) is avoided, proving that the adaptive mechanism achieves controlled contrast enhancement while safely preventing under-exposure or over-exposure artifacts seen in traditional Retinex-based methods. Furthermore, the proposed method presents a more continuous and stable transition across the mid-to-high intensity range compared to the baseline (blue curve), indicating that quantization noise and artificial color banding are effectively suppressed, thereby preserving the natural structural gradients of the underwater scene.

3.6 Ablation study

Ablation experiments were conducted to validate the necessity of the cascaded two-stage architecture and to evaluate the individual contributions of the proposed components. The quantitative results across five configurations evaluated on both the UIEB and LSUI datasets are summarized in Table 5.

Table 5. Ablation study on the effectiveness of different configurations on the UIEB and LSUI datasets.

The bold values indicate the best performance for each metric.

It is observed that relying solely on the handcrafted empirical prior (Pure-MSRCR) yields extremely low fidelity metrics, collapsing to a PSNR of only 14.78 dB on UIEB and 13.99 dB on LSUI. Although it achieves high UCIQE scores due to aggressive local contrast stretching, severe background noise amplification and pixel clipping artifacts are introduced due to the lack of regularization. With the integration of the CNN Refinement backbone (UIR-Baseline), a substantial performance leap is achieved, with the PSNR rising to 22.19 dB on UIEB and 25.84 dB on LSUI. This confirms that the learnable refinement network is essential for rectifying structural distortions and suppressing prior-induced noise.

To isolate the contribution of the proposed adaptive noise-gated fusion mechanism from the main backbone, two alternative feature-level configurations were implemented to replace the proposed gate: Average Fusion (replaced by addition) and Concat Fusion (replaced by concatenation). As presented in Table 5, replacing the gate with a static Concat Fusion causes the PSNR to drop by 0.29 dB (to 22.19 dB) on UIEB and by 0.90 dB (to 25.60 dB) on LSUI, indicating that a lack of adaptive filtering permits the propagation of prior-induced noise into the reconstruction pipeline. Furthermore, while Average Fusion yields moderate PSNR metrics, its structural similarity (SSIM) drops significantly to 0.9097 on UIEB and 0.8552 on LSUI, confirming that a spatially-invariant linear combination fails to safeguard fine-grained geometric boundaries from prior-induced blur. In contrast, the proposed Adaptive Gate (Ours) consistently achieves the best trade-off across both datasets, outperforming the non-adaptive alternatives while maintaining the highest structural preservation, validating its role as an intelligent signal selector for structure-preserving underwater image enhancement.

3.7 Adaptive weight analysis

To visually demonstrate the region-adaptive signaling and reliability-aware capabilities of the proposed framework, the learned 2D spatial weight map γ(H, W) is visualized as a pseudo-color heatmap in Figure 8. To further verify the resolution-agnostic generalization of LPD-Net, these qualitative and statistical evaluations are conducted directly on the native dimensions of representative scenes: a clear but blue-attenuated scene (Image 925, 500 × 333) and a severely turbid green-hazed scene (Image 8056, 1, 280 × 720). The statistical properties of the inferred spatial weights are summarized in Table 6.

Table 6. Statistical properties of the learned spatial weight map γ(H, W) under varying scene degradations.

As illustrated in the comparative matrix of Figure 8, LPD-Net exhibits an outstanding capacity for localized feature selection. In the relatively clear but severely color-shifted diver scene (Image 925), the physical prior provides high-quality chromatic clues without introducing severe structural noise. Consequently, the network adaptively reduces the average γ weight to 0.6478 (Table 6), shifting more reliance to the prior stream (represented by the dominant cool/blue regions in the top row of Figure 8c) to achieve efficient color correction.

In contrast, for the highly turbid coral reef scene (Image 8056), the empirical MSRCR prior inevitably introduces localized scattering artifacts and high-frequency noise. To satisfy the structure-preserving constraint, our spatial gate intelligently assigns a higher average raw weight of 0.7053 (represented by the warm/orange regions in the bottom row of Figure 8c), thereby anchoring the reconstruction pipeline to the original edge features and shielding the output from prior-induced distortions.

Furthermore, the standard deviations of the learned spatial weight maps for the clear and turbid scenes are 0.033 and 0.037, respectively (Table 6). These non-zero statistics quantitatively demonstrate that the proposed spatial gate performs localized pixel-wise modulation rather than collapsing into a spatially invariant global weighting factor. Such spatial heterogeneity allows the network to adaptively balance the contributions of the raw-image and prior streams across different image regions, helping suppress prior-induced artifacts while preserving fine structural details.

4 Discussion

The experimental results and efficiency benchmarks presented in Section 3 suggest that the proposed LPD-Net successfully achieves a superior balance between visual restoration and computational economy. This section further interprets these findings from the perspectives of signal selection, hardware constraints, and practical scalability.

4.1 Interpretability of the adaptive gating mechanism

The performance gains observed in Tables 2, 3 validate the effectiveness of the motivation-driven dual-stream paradigm. Unlike conventional black-box CNNs or static fusion methods, the proposed adaptive noise-gated mechanism functions as an intelligent signal selector. It is observed that the network dynamically adjusts the influence of the physical prior based on local degradation intensity. This behavior resolves the long-standing “restoration-enhancement conflict” in underwater imaging: by assigning higher weights to the raw stream in clear regions, the model effectively prevents the injection of scattering-induced artifacts from the MSRCR prior. This interpretability ensures that the enhancement process is not merely a data-fitting exercise but is guided by the underlying physical laws of light attenuation.

4.2 Analysis of the efficiency-performance trade-off

To evaluate the feasibility of deploying the proposed framework on resource-constrained underwater platforms, a comprehensive benchmark of computational efficiency was conducted. Metrics including trainable parameters (Params), multiply-accumulate operations (MACs), frames per second (FPS), and inference latency were measured on a single NVIDIA GeForce RTX 3050 GPU at a resolution of 256 × 256. To ensure statistical stability, each model was evaluated five times following 100 warmup iterations; the average results are summarized in Table 7, and the corresponding efficiency-performance trade-off is visualized in Figure 9. Crucially, as demonstrated in the Pareto front analysis, LPD-Net occupies the preferred “High FPS / Low Latency” region in terms of pure neural acceleration, achieving a competitive core throughput of 27.70 FPS (36.11 ms) while significantly outperforming SOTA methods such as DPF-Net (Mei et al., 2025) and Semi-UIR (Huang et al., 2023) in both latency and parameters.

Table 7. Computational complexity and efficiency comparison under different inference modes at 256 × 256 resolution.

A paradoxical phenomenon is observed in the U-shape Transformer (Peng et al., 2023): despite possessing relatively low theoretical MACs (2.98 G), its actual neural inference throughput is limited to 15.81 FPS, with latency exceeding 63 ms. Such results confirm that the extensive operator-switching and memory-copying overheads inherent in Transformer architectures significantly hinder hardware acceleration. In contrast, the proposed CNN-based architecture, with only 1.838 M parameters, maintains high efficiency. This demonstrates that our structural and quality gains are achieved without compromising the operational efficiency required for the sensing-control loop of Autonomous Underwater Vehicles (AUVs).

Beyond MACs, the practical bottleneck for edge devices often lies in activation memory usage. While LPD-Net maintains a consistent memory footprint, Transformer models exhibit high peak memory demands during the self-attention computation, which frequently leads to Out-of-Memory (OOM) errors on devices with restricted RAM.

4.3 Engineering viability and resolution scalability

The resolution-agnostic nature of LPD-Net provides substantial flexibility for diverse engineering scenarios. To address the requirements of practical edge deployment and provide a transparent evaluation, a comprehensive end-to-end latency benchmark was executed (Table 7). Unlike conventional studies that report only the neural backbone speed, our pipeline benchmark computes the total operational cycle, incorporating: (1) empirical prior generation via host-CPU-based MSRCR, (2) host-to-device memory copy, and (3) neural inference with GPU synchronization.

At a standardized resolution of 256 × 256, the complete end-to-end processing throughput of LPD-Net under a sequential Python prototype is measured as 6.58 frames/s (with a total latency of 152.06 ms) on a workstation-grade GPU (NVIDIA RTX 3050) and a mid-range host CPU (Intel Core i5-10500H @ 2.50GHz). Crucially, the latency breakdown reveals that the pure CNN backbone inference requires only 35.72 ms (28.00 FPS), while the unoptimized Python-based MSRCR generation on the host CPU consumes the majority of the cycle (115.25 ms, or over 75% of the total latency). The data transfer and concatenation overhead remains negligible at 1.09 ms.

To demonstrate the true engineering upper-bound of the proposed framework, a step-by-step optimization profile is conducted in Table 8. By converting the MSRCR operators into compiled C++ (compiled with −O3 optimization on the same i5-10500H CPU), and utilizing an eight-bit Logarithm Lookup Table (LUT) alongside a capped maximum Gaussian kernel size of 127 to alleviate Cache Miss, the preprocessing latency is dramatically reduced to 5.50 ms (equivalent to 181.85 FPS). Consequently, when integrated with our 35.72 ms neural backbone, the total end-to-end latency of LPD-Net converges to 42.31 ms, supporting a near real-time throughput of 23.63 FPS.

Table 8. Step-by-step latency optimization profile of the proposed pipeline at 256 × 256 resolution.

The bold values indicate the best performance for each metric.

From a system-level engineering perspective, maintaining the non-learning MSRCR prior generation on the host CPU/ISP rather than migrating it to the GPU is a deliberate design choice aimed at resource-contention mitigation. In heterogeneous computing systems deployed on AUVs, GPU resources must be strictly reserved for high-complexity deep learning tasks (such as our Stage II refinement and downstream object detection). Offloading the deterministic, scale-invariant MSRCR filter to the host CPU or dedicated on-chip Image Signal Processors (ISPs) enables a concurrent pipeline parallelism. By eliminating the CPU-side bottleneck, the proposed co-processing paradigm ensures that the primary GPU acceleration is fully leveraged for high-dimensional feature reconstruction without being bottlenecked by low-level, non-learning image processing operations.

4.4 Limitations and future work

Despite its efficiency, certain limitations remain. The current implementation of the Hybrid Domain Attention (HDA) module involves Fast Fourier Transforms (FFT), which exhibit non-linear computational peaks when processing non-power-of-two resolutions. Future work will focus on optimizing the spectral filtering branch through specialized hardware kernels or pruning strategies. Furthermore, future efforts will involve integrating the proposed framework into downstream autonomous defect detection systems to evaluate its efficacy in practical engineering scenarios, as well as conducting evaluations in more extreme underwater environments, such as deep-sea hydrothermal vents, where the light attenuation model becomes more complex.

5 Conclusions

In this paper, a physics-inspired adaptive dual-stream network is proposed to address the trade-off between color restoration and structural noise suppression in underwater image enhancement. By incorporating an adaptive “noise gate” mechanism, a dynamic balance between raw image fidelity and physics-inspired guidance is achieved under varying degradation conditions. Extensive evaluations on the UIEB and LSUI datasets indicate that competitive performance is achieved by the proposed method, with consistent improvements observed over representative data-driven baselines in both objective metrics (PSNR and SSIM) and perceptual quality. The effectiveness of the framework in preserving structural details is further supported by Canny edge analysis and patch-wise evaluation, demonstrating its suitability for high-precision inspection tasks. In addition, the proposed approach is designed to operate directly on high-resolution inputs, avoiding the interpolation artifacts commonly introduced by fixed-resolution processing pipelines. As a result, structural integrity is better maintained, which is critical for real-world engineering applications. Despite its effectiveness, the model maintains a compact size of 1.84 M parameters, offering an outstanding balance between visual enhancement quality and computational efficiency.

Generative AI statement

The author(s) declared that Generative AI was used in the creation of this manuscript. The author(s) utilized Gemini for language polishing and proofreading to enhance the readability of the manuscript. The AI tool assisted in refining the phrasing and sentence structure of the original draft. No original research data, figures, or core scientific claims were generated by AI; the author(s) take full responsibility for the integrity and accuracy of the content.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Xue-Cheng Tai, Norwegian Research Institute (NORCE), Norway

Reviewed by: Daoping Zhang, Nankai University, China

Preben Johnsen Bentdal, Uni Research, Norway