Adaptive Multi-Branch Feature Fusion for Low-Light Image Enhancement

Abstract

Low-light image enhancement (LLIE) remains a challenging problem due to spatially varying illumination degradation, compressed tonal distributions, and structural detail loss. This paper presents Adaptive Multi-Branch Feature Fusion (AMBFF), a unified framework that formulates LLIE as a multi-domain representation alignment task. The proposed architecture explicitly models complementary feature domains, including hierarchical spatial context, luminance–chrominance decoupling, edge–texture structures, frequency-domain information, and differentiable tonal histogram representations. A spatially adaptive gating mechanism dynamically weights multi-feature branches through a convex fusion strategy, enabling location-aware illumination correction while preserving structural integrity and color fidelity. Extensive evaluations on widely used benchmark datasets demonstrate that AMBFF consistently outperforms representative conventional and deep learning-based approaches in terms of PSNR, SSIM, and LPIPS. Ablation analyses confirm the complementarity of the proposed feature domains and the robustness benefits of adaptive fusion. Despite its multi-branch design, AMBFF maintains a favorable performance–complexity trade-off, highlighting the effectiveness of structured multi-domain modeling for low-light image enhancement.

1. Introduction

Images captured in poor-light conditions are generally lack of natural color tones and details. These images are not only visually unpleasant but also diminish computer vision-based applications’ performance. Due to that reason low-light image enhancement (LLIE) has been played a critical role in computer vision as a preprocessing step. LLIE implementations has found place for itself in various fields, such as photography [1], remote sensing [2], autonomous driving [3], underwater imaging [4], medical image analysis [5], and other vision-based tasks.

There are two primary approaches to enhancing low-light images: modifying hardware configurations or applying algorithmic processing. With hardware modifications, one can adjust the ISO (the camera’s sensitivity) or increase the exposure time. However, these adjustments can introduce noise and may saturate pixel values. In algorithmic processing, a low-light image is provided as input to a linear or non-linear function designed to enhance the image’s contrast and brightness. Most studies on LLIE primarily focus on algorithmic processing.

Traditional image enhancement methods can generally be classified into four categories: transformation-based, histogram-based, Retinex-based, and learning-based approaches. In transformation-based approaches, a linear or non-linear function is applied to the pixel values, resulting in a contrast-stretched output image. In histogram-based methods, the color distribution of the image is reorganized to expand the dynamic range. Retinex-based methods decompose an image into its reflectance and illumination components, and, by estimating the illumination, reconstruct the reflectance component to enhance the image. Learning-based methods obtain optimized weights that transform low-light pixel values into well-exposed ones.

Learning-based approaches leverage diverse image cues, including illumination characteristics, color distributions, and structural features, to learn optimal enhancement mappings. While many existing methods focus on a single dominant representation, more recent efforts explore multi-feature integration to improve robustness under complex low-light conditions. To address the limitations of single-path modeling, an Adaptive Multi-Branch Feature Fusion (AMBFF) framework is introduced for the LLIE task. The proposed model explicitly decomposes feature extraction into complementary domains—spatial context, luminance–chrominance components, edge–texture structures, frequency-domain information, and differentiable tonal histogram representations—and integrates them through a learnable attention-based weighting mechanism. The enhanced image is reconstructed from the adaptively fused representation, enabling balanced illumination correction while preserving structural integrity and color fidelity.

The contributions of this study are as follows:

• An Adaptive Multi-Branch Feature Fusion (AMBFF) framework is introduced for low-light image enhancement, jointly modeling spatial context, luminance–chrominance decoupling, edge–texture structure, frequency-domain cues, and differentiable tonal histogram representations within a unified end-to-end architecture.
• A content-adaptive, spatially varying gating mechanism combined with channel and spatial attention dynamically weights multi feature branches, enabling illumination correction while preserving structural details and color fidelity.
• A unified Residual Attention Block (RAB) integrating channel recalibration, spatial modulation, and residual learning is systematically deployed across backbone, branch, fusion, and reconstruction stages to ensure stable and parameter-efficient feature refinement.
• A systematic ablation study with zero-branch substitution verifies that adaptive fusion enhances robustness to branch exclusion and confirms the complementarity of the proposed multi-domain feature representations.
• Experiments on LOL, LOL-V2-Real, and LOL-V2-Synthetic datasets demonstrate consistent improvements over conventional and recent deep learning methods in terms of PSNR, SSIM, and LPIPS metrics.

The remainder of this paper is organized as follows. Section 2 reviews related work on low-light image enhancement and multi-branch feature learning. Section 3 presents the proposed AMBFF framework in detail. Section 4 describes the experimental setup, benchmark datasets, evaluation metrics, and comparative results. Section 5 provides a discussion of the obtained results and analyzes the strengths and limitations of the proposed method. Finally, Section 6 concludes the paper and outlines potential directions for future research.

2. Related Work

Low-light image enhancement (LLIE) aims to restore images captured under poor lighting conditions to a state that is both visually accurate and pleasing. This task is an ill-posed problem, as inadequate illumination can compromise brightness, contrast, color fidelity, texture clarity, and introduce noise. LLIE research has progressed significantly over time, moving from traditional handcrafted methods to advanced deep learning approaches. Recent work particularly highlights the importance of diverse feature representation, effective disentanglement, and adaptive fusion for optimal results.

Early studies on LLIE primarily rely-on transformation-based techniques, such as logarithmic transformation, gamma correction, and contrast stretching. These methods operate directly on pixel intensity values, using predefined intensity mapping functions to boost the visibility of underexposed pixels. Despite their computational efficiency and straightforward implementation, transformation-based algorithms typically lack the flexibility needed to handle complex lighting conditions. As a result, they are prone to amplifying noise or introducing visual artifacts, especially in challenging, severely underexposed regions encountered in real-world scenarios.

Transformation-based enhancement techniques often exhibit limited adaptability under spatially varying illumination. Consequently, histogram-based approaches have been developed to exploit the statistical distribution of pixel intensities for contrast enhancement. Histogram Equalization (HE) remains a foundational technique, employing the cumulative distribution function to globally remap intensity values and improve overall contrast. However, its global adjustment strategy may lead to over-enhancement and visible artifacts, particularly in images with non-uniform lighting. To mitigate these limitations, Adaptive Histogram Equalization (AHE) [6] was introduced, partitioning the image into local regions and enhancing contrast within each region independently. Although this localized mechanism improves detail visibility, it can also amplify noise in homogeneous areas, thereby reducing robustness. Contrast-Limited Adaptive Histogram Equalization (CLAHE) [7] further refines this strategy by incorporating a clipping function that constrains the histogram distribution, effectively suppressing noise amplification while preserving local contrast.

Retinex theory, first proposed by Land [8], draws inspiration from human visual perception by decomposing an image into illumination and reflectance components. This conceptual framework enables Retinex-based algorithms to selectively enhance scene illumination while maintaining the integrity of reflectance, leading to more visually realistic results under diverse lighting conditions. Single-Scale Retinex (SSR) [9] achieves contrast enhancement by applying a logarithmic transformation to the ratio between the original image and its Gaussian-blurred counterpart. Although SSR is effective in boosting contrast, its performance is heavily dependent on the chosen scale and it can introduce color artifacts. To address these limitations, Multi-Scale Retinex (MSR) [10] aggregates SSR outputs computed at various spatial scales to simultaneously preserve global contrast and fine image details. Building upon this, Multi-Scale Retinex with Color Restoration (MSRCR) [11] integrates a color restoration function to mitigate chromatic distortions introduced by independent channel processing.

Recent advancements in deep learning (DL) have driven LLIE research toward data-driven paradigms that learn end-to-end mappings from degraded inputs to well-exposed outputs [12,13,14,15]. DL-based LLIE methods are typically divided into supervised approaches, which require paired low-light and normal-light image sets, and unsupervised approaches, which do not depend on paired reference images. Within supervised methods, Convolutional Neural Networks (CNNs) [16,17] are widely utilized for their effectiveness in capturing spatial dependencies and modeling complex nonlinear transformations. Notably, encoder-decoder architectures such as LLNet [18] and the Dual Autoencoder Network [19] encode input images into compact latent representations and reconstruct enhanced outputs while maintaining structural information.

RetinexNet [20] represents one of the earliest supervised learning-based frameworks, embedding Retinex theory within a deep learning context by decomposing input images into illumination and reflectance components. The illumination component is enhanced using a dedicated subnetwork to restore intrinsic brightness and color fidelity. Building on this approach, KinD [21] introduced a unified architecture that simultaneously performs image decomposition, illumination adjustment, and reflectance restoration, which reduces color distortions and improves perceptual consistency. To reduce reliance on paired training data and improve generalization, subsequent research has investigated unsupervised and self-supervised learning strategies. EnlightenGAN [22] employs adversarial and self-supervised objectives, utilizing global-local discriminators and attention mechanisms to generate visually coherent results without paired supervision. In a related development, Zero-DCE [23] formulates illumination enhancement as a learnable curve estimation task, allowing single-image training through specifically designed self-regularized loss functions.

Among illumination estimation-based methods, LIME [24] generates pixel-wise illumination maps through channel-wise maximum operations and refines them using structure-aware smoothing. This approach enables adaptive brightness enhancement while minimizing visual artifacts. Distort-and-Recover [25] conceptualizes low-light enhancement as a sequential decision-making task utilizing deep reinforcement learning (DRL), in which synthetic low-light images are produced by systematically degrading well-exposed inputs. DSLR [26] employs a stacked Laplacian pyramid architecture to incrementally restore illumination and texture information at multiple spatial scales. In [27], a semi-supervised band-representation framework is introduced that decomposes images into linear bands and reconstructs them using perceptual and adversarial learning objectives, thereby achieving a balance between signal fidelity and perceptual realism.

Recent LLIE methods increasingly employ multi-feature fusion strategies to jointly address illumination correction, texture preservation, and color consistency. Fusion-based architectures integrate complementary information across multiple feature levels, scales, or domains, thereby reducing detail loss and artifact amplification that are common in single-path models. A representative deep feature fusion approach [28] employs hierarchical feature extraction and cross-level feature integration to generate pixel-wise enhancement functions in a reference-free manner, improving robustness under complex illumination conditions. To further overcome the limited dynamic range of single low-light observations, a multi-exposure generation and fusion strategy has been proposed [29]. Beyond the spatial domain, spatial–frequency dual-domain fusion has been explored to capture global context and long-range dependencies more effectively. For instance, the Dual-Domain Feature Fusion Network (DFFN) [30] leverages Fourier-domain amplitude for illumination restoration and phase information for detail preservation, integrating these components through cross-stage and cross-scale fusion to mitigate information loss in complex scenes. Meanwhile, FusionNet [31] introduces a fully parallel linear fusion framework that integrates CNN-, Transformer-, and color-space-based representations, enabling efficient inference and stable optimization with theoretical guarantees derived from Hilbert space formulations. Attention-guided multi-scale fusion has emerged as an effective strategy for low-light image enhancement. MSFFNet [32] integrates channel–spatial attention with cascade fusion within a unified encoder–decoder architecture, facilitating cross-scale information propagation while maintaining low computational complexity. MBLLIE-Net [33] further introduces a multi-branch feature specialization framework combined with iterative refinement, where features extracted from different depths and scales are processed in parallel and progressively enhanced through recurrent update stages. This design incorporates spatial recurrent units and adaptive channel attention to strengthen long-range dependency modeling and improve perceptual consistency. Collectively, these studies indicate that coordinated multi-feature fusion across scales, domains, and iterative stages has become a central paradigm for robust low-light image enhancement in real-world scenarios. There are other domain-adaptive methods that enhance images with effective computing [34,35,36,37,38,39,40].

Despite the significant progress achieved by fusion-based LLIE methods, several limitations remain. Many existing models rely on fixed or loosely coupled fusion strategies and single-stage enhancement pipelines, which restrict their ability to adaptively balance complementary features under spatially varying and complex low-light conditions. Moreover, limited interaction among feature representations across domains and scales often leads to suboptimal illumination restoration and detail preservation. To address these challenges, this paper proposes Adaptive Multi-Branch Feature Fusion (AMBFF), a unified framework that integrates multiple complementary feature representations within an adaptive fusion architecture. The proposed design enables dynamic feature weighting and residual refinement, resulting in more robust illumination enhancement while preserving structural details and color fidelity in real-world low-light scenarios.

3. Proposed Method

In this study an Adaptive Multi-Branch Feature Fusion (AMBFF) network is introduced for low-light image enhancement. The primary motivation is that low-light degradation simultaneously impacts multiple image characteristics, including global brightness, local contrast, texture sharpness, frequency content, and chromatic consistency. These aspects cannot be adequately addressed by a single representation or processing path. Therefore, the proposed framework explicitly decomposes the enhancement process into complementary feature domains and integrates them using an adaptive, learnable fusion mechanism. The model pipeline is illustrated in Figure 1.

3.1. Residual Attention Block (RAB)

The Residual Attention Block (RAB) constitutes the core feature-refinement unit employed throughout the AMBFF network—in the encoder, bottleneck, decoder, and all spectral/statistical branches. Its design integrates two complementary attention mechanisms, channel-wise squeeze-and-excitation [41] and spatial attention [42], within a residual learning framework, enabling the network to adaptively recalibrate both which features and where to attend at each processing stage.

RAB first performs two successive 3 × 3 convolutions to extract local spatial features. Let the input feature map be $\mathrm{F}\in {\mathbb{R}}^{B\times C\times H\times W}$, where B denotes the batch size, C represents the number of feature channels, and H and W correspond to the spatial height and width of the feature map, respectively. The intermediate refined feature ${\mathrm{F}}_{loc}$ is expressed as:

$${\mathrm{F}}_{loc}={\mathrm{Conv}}_{3\times 3}\left(\mathrm{SiLU}\left({\mathrm{Conv}}_{3\times 3}\left(\mathrm{F}\right)\right)\right)$$

(1)

where SiLU denotes the Sigmoid Linear Unit activation function. SiLU is utilized here for its smooth gradient properties and self-gated mechanism, which have been shown to improve convergence in deep architectures [43,44].

The local features are subsequently passed through a dual-attention pipeline to model inter-dependencies across channels and spatial locations. A Channel Attention mechanism ${\mathcal{A}}_c$ is first applied, inspired by the Squeeze-and-Excitation (SE) [41] framework. This module models global channel inter-dependencies by first squeezing the spatial dimensions into a global descriptor $\mathrm{z}\in {\mathbb{R}}^{B\times C\times 1\times 1}$ via global average pooling (GAP):

$$\mathrm{z}=\mathrm{GAP}\left({\mathrm{F}}_{loc}\right)$$

(2)

The descriptor is then processed by two 1 × 1 convolutions to generate channel-wise weights:

$${\mathcal{A}}_c\left({\mathrm{F}}_{loc}\right)={\mathrm{F}}_{loc}\otimes \mathrm{SiLU}\left({\mathrm{Conv}}_{1\times 1}\left(\mathrm{SiLU}\left({\mathrm{Conv}}_{1\times 1}\left(\mathrm{z}\right)\right)\right)\right)$$

(3)

where ⊗ denotes element-wise multiplication. To focus on structurally significant regions, such as edges and high-frequency textures, a Spatial Attention [42] mechanism ${\mathcal{A}}_s$ is applied. By concatenating the mean and maximum descriptors across the channel dimension, the module identifies salient spatial locations:

$${\mathcal{A}}_s\left(\mathrm{F}\right)=\mathrm{F}\otimes \mathrm{SiLU}\left({\mathrm{Conv}}_{7\times 7}\left([\mathrm{Mean}\left(\mathrm{F}\right);\mathrm{Max}\left(\mathrm{F}\right)]\right)\right)$$

(4)

where $[·;·]$ denotes the concatenation operator along the channel axis. The RAB mapping can be summarized as

$$\mathrm{RAB}\left(\mathrm{F}\right)=\mathrm{F}+{\mathcal{A}}_s\left({\mathcal{A}}_c\left({\mathrm{F}}_{loc}\right)\right)$$

(5)

The RAB utilizes a base channel width of $C=48$, a value determined through hyperparameter optimization using the Optuna framework. The optimization was conducted over the candidate set $C\in \{32,48,64\}$, where 48 was identified as the optimal configuration for balancing representational capacity and computational efficiency. For the remainder of this work, the channel dimension C is fixed to 48 across all modular components.

3.2. Multi-Branch Feature Extraction

The AMBFF framework decomposes feature extraction into multiple complementary branches, with each branch designed to capture a specific aspect of the degradation process. By modeling spatial context, luminance-chrominance relationships, structural information, frequency-domain characteristics, and intensity distribution in parallel, the multi-branch design enables specialized feature learning that is challenging to achieve within a unified single-path architecture. These domain-specific features are then integrated through an adaptive fusion mechanism to produce a comprehensive and balanced representation for LLIE task.

3.2.1. Spatial Feature Branch

The spatial feature branch adopts a U-Net–style encoder–decoder architecture with residual attention refinement to model multi-scale spatial dependencies. This branch serves as the primary structural backbone and provides dense feature maps aligned with the input resolution.

The input RGB image $\mathrm{X}\in {\mathbb{R}}^{B\times 3\times H\times W}$ is first projected into the base feature space:

$${\mathrm{F}}_{\mathrm{enc}}^{\left(0\right)}=\mathrm{RAB}\left(\mathrm{SiLU}\left({\mathrm{Conv}}_{3\times 3}\left(\mathrm{X}\right)\right)\right),$$

(6)

with

$${\mathrm{F}}_{\mathrm{enc}}^{\left(0\right)}\in {\mathbb{R}}^{B\times C\times H\times W}.$$

The encoder–decoder consists of D hierarchical stages representing the number of resolution levels. In the implementation $D\in \{2,3,4\}$, and $D=3$ is selected via hyperparameter optimization. The encoder consists of $D-1$ stages and at stage i, spatial resolution is reduced by a factor of two using a stride-2 convolution, followed by SiLU activation and residual attention refinement.

For $i=1,\dots ,D-1$:

$${\mathrm{F}}_{\mathrm{enc}}^{\left(i\right)}=\mathrm{RAB}\left(\mathrm{SiLU}\left({\mathrm{Conv}}_{3\times 3}^{(\mathrm{stride}=2)}\left({\mathrm{F}}_{\mathrm{enc}}^{(i-1)}\right)\right)\right),$$

(7)

with

$${\mathrm{F}}_{\mathrm{enc}}^{\left(i\right)}\in {\mathbb{R}}^{B\times 2^{i}C\times {\displaystyle \frac{H}{2^i}}\times {\displaystyle \frac{W}{2^i}}}.$$

Each encoder feature map ${\mathrm{F}}_{\mathrm{enc}}^{\left(i\right)}$ is retained for skip connections. At the coarsest resolution ($i=D-1$), the encoder output ${\mathrm{F}}_{\mathrm{enc}}^{(D-1)}$ forms the bottleneck representation. Additional refinement is performed using stacked residual attention blocks:

$${\mathrm{F}}_{\mathrm{bn}}=\mathrm{RAB}\left(\mathrm{RAB}\left({\mathrm{F}}_{\mathrm{enc}}^{(D-1)}\right)\right),$$

(8)

where

$${\mathrm{F}}_{\mathrm{bn}}\in {\mathbb{R}}^{B\times 2^{D-1}C\times {\displaystyle \frac{H}{2^{D-1}}}\times {\displaystyle \frac{W}{2^{D-1}}}}.$$

This bottleneck feature aggregates global contextual information with the largest effective receptive field. The decoder progressively restores spatial resolution. Initialization is given by

$${\mathrm{F}}_{\mathrm{dec}}^{(D-1)}={\mathrm{F}}_{\mathrm{bn}}.$$

For $i=D-1,\dots ,1$:

$${\widehat{\mathrm{F}}}_{\mathrm{dec}}^{(i-1)}=\mathrm{RAB}\left(\mathrm{SiLU}\left({\mathrm{Conv}}_{3\times 3}\left(\mathrm{Up}\left({\mathrm{F}}_{\mathrm{dec}}^{\left(i\right)}\right)\right)\right)\right),$$

(9)

where $\mathrm{Up}(·)$ denotes bilinear upsampling by a factor of two. The decoder incorporates additive skip connections from the encoder. After each upsampling block $\mathrm{Up}(·)$, the corresponding encoder feature map ${\mathrm{F}}_{\mathrm{enc}}^{\left(i\right)}$ is added element-wise to the decoder feature map $\mathrm{Up}\left({\mathrm{F}}_{\mathrm{dec}}^{\left(i\right)}\right)$:

$${\mathrm{F}}_{\mathrm{dec}}^{(i-1)}={\widehat{\mathrm{F}}}_{\mathrm{dec}}^{(i-1)}+{\mathrm{F}}_{\mathrm{enc}}^{(i-1)}.$$

(10)

After the final decoding stage ($i=0$),

$${\mathrm{F}}_{\mathrm{Spatial}}={\mathrm{F}}_{\mathrm{dec}}^{\left(0\right)}\in {\mathbb{R}}^{B\times C\times H\times W}.$$

(11)

3.2.2. Luma–Chrominance Feature Branch

Low-light degradation affects luminance and chrominance components differently. While insufficient illumination primarily compresses the dynamic range of the luminance channel, chrominance channels often exhibit color shifts and reduced saturation. Joint processing of RGB channels may entangle brightness restoration with color correction, leading to suboptimal enhancement. The luma–chrominance branch is therefore designed to explicitly decouple intensity and color information while enabling controlled feature fusion.

The RGB input $\mathrm{X}\in {\mathbb{R}}^{B\times 3\times H\times W}$, converted to the YCbCr color space, separating luminance component (Y) from chrominance components (${\mathrm{C}}_b,{\mathrm{C}}_r$),

$$\mathrm{Y}\in {\mathbb{R}}^{B\times 1\times H\times W},{\mathrm{C}}_b,{\mathrm{C}}_r\in {\mathbb{R}}^{B\times 1\times H\times W}.$$

The luminance channel is processed independently through a convolutional refinement module:

$${\mathrm{F}}_Y=\mathrm{RAB}\left(\mathrm{SiLU}\left({\mathrm{Conv}}_{3\times 3}\left(\mathrm{Y}\right)\right)\right),$$

(12)

yielding ${\mathrm{F}}_Y\in {\mathbb{R}}^{B\times C\times H\times W}$. This stream captures illumination structure and brightness-dependent features. The chrominance channels $C_b,C_r\in {\mathbb{R}}^{B\times 1\times H\times W}$, representing the blue-difference and red-difference chroma channels, are concatenated:

$${\mathrm{C}}_{cat}=\mathrm{Concat}({\mathrm{C}}_b,{\mathrm{C}}_r)\in {\mathbb{R}}^{B\times 2\times H\times W},$$

and processed analogously:

$${\mathrm{F}}_{cat}=\mathrm{RAB}\left(\mathrm{SiLU}\left({\mathrm{Conv}}_{3\times 3}\left({\mathrm{C}}_{cat}\right)\right)\right),$$

(13)

producing ${\mathrm{F}}_{cat}\in {\mathbb{R}}^{B\times C\times H\times W}$. This stream focuses on color distribution and saturation cues. The luminance and chrominance embeddings are concatenated and projected back to the common embedding space using a 1 × 1 convolution followed by residual attention refinement:

$${\mathrm{F}}_{\mathrm{LC}}=\mathrm{RAB}\left(\mathrm{SiLU}\left({\mathrm{Conv}}_{1\times 1}\left(\mathrm{Concat}({\mathrm{F}}_Y,{\mathrm{F}}_{cat})\right)\right)\right).$$

(14)

The branch outputs a feature embedding:

$${\mathrm{F}}_{\mathrm{LC}}\in {\mathbb{R}}^{B\times C\times H\times W}.$$

(15)

3.2.3. Edge–Texture Feature Branch

This branch extracts structure-aware features that highlight object boundaries and local texture patterns, enabling the network to differentiate meaningful details from noise. Integration with spatial and color-aware representations via the adaptive fusion mechanism enhances sharpness and structural consistency, while mitigating common artifacts such as halos and excessive smoothing.

To capture prominent structural boundaries, the Sobel operator is employed to compute the gradient magnitude. Given the luminance channel Y, the horizontal and vertical gradients are computed via convolution with discrete Sobel kernels $K_x$ and $K_y$:

$${\mathrm{K}}_x=\left[\begin{array}{ccc}-1& 0& 1\\ -2& 0& 2\\ -1& 0& 1\end{array}\right],{\mathrm{K}}_y=\left[\begin{array}{ccc}-1& -2& -1\\ 0& 0& 0\\ 1& 2& 1\end{array}\right].$$

$${\mathrm{G}}_x=\mathrm{Y}\ast {\mathrm{K}}_x,{\mathrm{G}}_y=\mathrm{Y}\ast {\mathrm{K}}_y,$$

(16)

where ∗ denotes spatial convolution, and ${\mathrm{G}}_x$ and ${\mathrm{G}}_y$ horizontal and vertical gradients, respectively. Some references swap the notations of ${\mathrm{K}}_x$ and ${\mathrm{K}}_y$ (or use sign-flipped variants); such differences only affect gradient orientation, while the magnitude in (17) remains unchanged. The gradient magnitude map is then computed as

$$\mathrm{E}=\sqrt{{\mathrm{G}}_x^2+{\mathrm{G}}_y^2+ϵ},$$

(17)

with $ϵ$ ensuring numerical stability. To capture second-order intensity variations associated with fine textures, a Laplacian operator is applied:

$${\mathrm{K}}_{\Delta }=\left[\begin{array}{ccc}0& 1& 0\\ 1& -4& 1\\ 0& 1& 0\end{array}\right].$$

The texture response is defined as

$$\mathrm{T}=\left|\mathrm{Y}\ast {\mathrm{K}}_{\Delta }\right|.$$

(18)

The gradient magnitude E and texture response T are concatenated along the channel dimension:

$$\mathrm{S}=\mathrm{Concat}(\mathrm{E},\mathrm{T})\in {\mathbb{R}}^{B\times 2\times H\times W}.$$

(19)

This two-channel structural map is projected into the common embedding space using a convolutional refinement module:

$${\mathrm{F}}_{\mathrm{ET}}=\mathrm{RAB}\left(\mathrm{SiLU}\left({\mathrm{Conv}}_{3\times 3}^{(d=2)}\left(\mathrm{SiLU}\left({\mathrm{Conv}}_{3\times 3}\left(\mathrm{S}\right)\right)\right)\right)\right),$$

(20)

yielding ${\mathrm{F}}_{\mathrm{ET}}\in {\mathbb{R}}^{B\times C\times H\times W}$ where d = 2 denotes dilation factor 2, which expands the effective receptive field of the 3 × 3 kernel to 5 × 5 without increasing parameter count.

3.2.4. Frequency Feature Branch

The frequency feature branch is designed to capture global illumination characteristics that are challenging to represent within the spatial domain. In low-light images, illumination degradation primarily impacts low-frequency components, whereas high-frequency details are frequently attenuated or distorted by noise. The frequency-aware features complement spatial, structural, and color representations. When adaptively fused, they enhance the robustness and consistency of image enhancement under uneven and extreme lighting conditions.

To obtain frequency-domain features, the two-dimensional Discrete Fourier Transform (DFT) is applied to the luminance map Y:

$$\mathcal{F}\left(\mathrm{Y}\right)(u,v)={\displaystyle \sum _{m=0}^{H-1}\sum _{n=0}^{W-1}}\mathrm{Y}(m,n)e^{-j2\pi \left({\displaystyle \frac{um}{H}}+{\displaystyle \frac{vn}{W}}\right)},$$

(21)

where $\mathcal{F}\left(Y\right)(u,v)\in {\mathbb{C}}^{B\times 1\times H\times W}$ denotes the Fourier transform of the luminance channel, with $\mathbb{C}$ representing the complex domain. The indices $(m,n)$ denote spatial coordinates, whereas $(u,v)$ correspond to frequency indices. The variables H and W indicate the spatial height and width of the image, respectively, and $j=\sqrt{-1}$ denotes the imaginary unit. In practice, this transform is efficiently computed using the Fast Fourier Transform implementation with orthonormal normalization.

In low-light image enhancement, the phase component often contains degraded structural information due to sensor noise. In contrast, the magnitude spectrum $\left|\mathcal{F}\right(\mathrm{Y}\left)\right(u,v\left)\right|$ provides a stable representation of global illumination and contrast energy. The zero-frequency coefficient $\mathcal{F}(0,0)$, commonly referred to as the DC component, corresponds to the average brightness of the image and typically has significantly larger magnitude than higher-frequency components. To prevent the DC component from dominating gradient flow during optimization, logarithmic dynamic range compression is applied:

$$\mathrm{L}=ln\left(1+\left|\mathcal{F}\right(\mathrm{Y}\left)\right|+ϵ\right),$$

(22)

where $ϵ$ is a small constant for numerical stability. This transformation compresses high-energy low-frequency responses while relatively enhancing mid- and high-frequency components that encode structural details.

The log-amplitude map L is projected to the common branch feature space through a lightweight convolutional refinement module consistent with the overall AMBFF design:

$${\mathrm{F}}_{\mathrm{Fourier}}=\mathrm{SiLU}\left({\mathrm{Conv}}_{3\times 3}\left(\mathrm{RAB}\left(\mathrm{SiLU}\left({\mathrm{Conv}}_{3\times 3}\left(\mathrm{L}\right)\right)\right)\right)\right),$$

(23)

yielding ${\mathrm{F}}_{\mathrm{Fourier}}\in {\mathbb{R}}^{B\times C\times H\times W}$.

3.2.5. Histogram Feature Branch

Low-light images often suffer from severe tone compression, where pixel intensities cluster near the lower dynamic range, resulting in a highly skewed brightness distribution. Convolutional layers effectively capture local structures but do not explicitly encode global tonal statistics. The histogram branch addresses this limitation by transforming brightness distributions into dense feature maps, providing each pixel with information about its relative position within the global tonal landscape. Acting as a learnable and differentiable form of histogram equalization, this branch guides the network to correct tonal imbalance in low-light images.

The input RGB image $\mathrm{X}\in {\mathbb{R}}^{B\times 3\times H\times W}$ is converted into the YCbCr color space, and the luminance channel $\mathrm{Y}\in {\mathbb{R}}^{B\times 1\times H\times W}$ is extracted. The branch operates exclusively on the luminance channel Y and employs a differentiable soft histogram similar in [45].

To balance tonal resolution and computational cost, different bin counts $N_b\in \{16,32,64\}$ were experimentally evaluated. Although larger values increase tonal granularity, they also increase computational complexity. Consequently, $N_b=32$ was selected as the optimal trade-off between performance and efficiency.

Using Gaussian kernel density estimation on evenly spaced bin centers within the normalized intensity range $[0,1]$, the bin centers are defined as

$${\mu }_k={\displaystyle \frac{k-1}{N_b-1}},k=1,2,\dots ,N_b.$$

(24)

The centers are stacked into a tensor

$$\mathit{\mu }\in {\mathbb{R}}^{1\times N_b\times 1\times 1},$$

(25)

which is broadcast along the batch and spatial dimensions during computation. For each pixel value $y_{b,i,j}$, the soft membership to bin k is computed as

$${\widehat{h}}_{b,k,i,j}=exp\left(-{\displaystyle \frac{{(y_{b,i,j}-{\mu }_k)}^2}{2{\sigma }^2}}\right),$$

(26)

where the bandwidth parameter $\sigma$ is defined relative to the uniform bin spacing and controls the smoothness of the assignment. To ensure normalization across bins, the following operation is applied:

$$h_{b,k,i,j}={\displaystyle \frac{{\widehat{h}}_{b,k,i,j}}{{\displaystyle {\sum }_{t=1}^{N_b}}{\widehat{h}}_{b,t,i,j}+\epsilon }}.$$

(27)

This produces the spatial histogram tensor

$$\mathrm{H}\in {\mathbb{R}}^{B\times N_b\times H\times W}.$$

(28)

The $N_b$-channel tonal representation H is projected into the common branch feature space ${\mathbb{R}}^{B\times C\times H\times W}$ through a lightweight learned module:

$${\mathrm{F}}_{\mathrm{Hist}}=\mathrm{RAB}\left(\mathrm{SiLU}\left({\mathrm{Conv}}_{1\times 1}\left(\mathrm{H}\right)\right)\right).$$

(29)

3.3. Fusion Block

The proposed AMBFF architecture integrates information from five distinct domains to reconstruct high-quality images from low-light inputs. The fusion process is governed by a dynamic gating mechanism that adaptively modulates the contribution of each specialized branch.

Let ${\mathrm{F}}_{Spatial},{\mathrm{F}}_{LC},{\mathrm{F}}_{ET},{\mathrm{F}}_{Fourier}$, and ${\mathrm{F}}_{Hist}$ represent the feature maps generated by the Spatial, Luma-Chroma, Edge-Texture, Fourier, and Histogram branches, respectively. These maps are first concatenated along the channel dimension:

$${\mathrm{F}}_{cat}=\mathrm{Concat}({\mathrm{F}}_{Spatial},{\mathrm{F}}_{LC},{\mathrm{F}}_{ET},{\mathrm{F}}_{Fourier},{\mathrm{F}}_{Hist})\in {\mathbb{R}}^{B\times 5C\times H\times W}$$

(30)

where C is the base channel capacity of the network. A lightweight gating network produces spatially varying fusion logits:

$$\mathrm{Z}={\mathrm{Conv}}_{1\times 1}\left(\mathrm{SiLU}\left({\mathrm{Conv}}_{1\times 1}\left({\mathrm{F}}_{cat}\right)\right)\right),\mathrm{Z}\in {\mathbb{R}}^{B\times 5\times H\times W}.$$

(31)

For each batch index b and spatial location $(i,j)$, softmax is applied along the branch dimension ($K=5$) to obtain normalized weights:

$${\alpha }_{b,m,i,j}={\displaystyle \frac{exp\left(Z_{b,m,i,j}\right)}{{\displaystyle {\sum }_{t=1}^5}exp\left(Z_{b,t,i,j}\right)}},m=1,\dots ,5.$$

(32)

The weights satisfy

$${\displaystyle \sum _{m=1}^5}{\alpha }_{b,m,i,j}=1,{\alpha }_{b,m,i,j}\ge 0.$$

(33)

The fused feature map is computed as a convex combination of branch features at each spatial location:

$${\mathrm{F}}_{Fused}={\displaystyle \sum _{m=1}^5}{\alpha }_m\odot {\mathrm{F}}_m,$$

(34)

where ${\alpha }_m\in {\mathbb{R}}^{B\times 1\times H\times W}$ is obtained by taking the m-th slice of $\alpha \in {\mathbb{R}}^{B\times 5\times H\times W}$ and broadcasting it along the channel dimension, and ⊙ denotes element-wise multiplication, yielding:

$${\mathrm{F}}_{Fused}\in {\mathbb{R}}^{B\times C\times H\times W}.$$

Following the weighted summation, a RAB is applied to refine the fused representation and suppress potential inter-branch inconsistencies:

$${\mathrm{F}}_{Final}=\mathrm{RAB}\left({\mathrm{F}}_{Fused}\right).$$

(35)

3.4. Reconstruction Block

The reconstruction block converts the fused feature representation into the final enhanced image. After the adaptive fusion stage, the integrated features are further refined and projected into the image space using a lightweight reconstruction head consisting of convolutional layers and residual attention refinement. The final output image is obtained by combining the predicted global residual (${\mathrm{F}}_{Final}$) with the original input (X) and constraining the result to a valid intensity range. This reconstruction block enables effective illumination enhancement while preserving structural integrity and color consistency across diverse low-light conditions. The final enhanced output is computed as:

$$\widehat{\mathrm{X}}=\mathrm{Clip}(\mathrm{X}+{\mathrm{F}}_{Final},0,1),$$

(36)

where $\mathrm{Clip}(·)$ constrains pixel values to the valid intensity range $[0,1]$.

3.5. Training Objective

The pixel-wise loss enforces accurate luminance and color reconstruction by measuring the discrepancy between the predicted enhanced image and the corresponding ground-truth at the intensity level. Specifically, the pixel fidelity loss is defined using the ${\mathcal{l}}_1$ distance between the enhanced output $\widehat{\mathrm{X}}$ and the ground-truth image ${\mathrm{X}}_{GT}$:

$${\mathcal{L}}_{\mathrm{pix}}={∥\widehat{\mathrm{X}}-{\mathrm{X}}_{GT}∥}_1.$$

(37)

To further preserve structural consistency and fine-grained textures beyond strict pixel alignment, a perceptual loss is incorporated. This loss is computed in the deep feature space extracted from a selected layer of a pretrained VGG16 network [46]. Let ϕ(·) denote the feature extraction function of the selected VGG16 layer. The perceptual loss is defined as

$${\mathcal{L}}_{\mathrm{per}}={∥\varphi \left(\widehat{\mathrm{X}}\right)-\varphi \left({\mathrm{X}}_{GT}\right)∥}_1.$$

(38)

During training, the network parameters $\theta$ are optimized using a weighted combination of pixel fidelity (${\mathcal{L}}_{\mathrm{pix}}$) and perceptual consistency (${\mathcal{L}}_{\mathrm{per}}$) losses:

$$\mathcal{L}\left(\theta \right)={\lambda }_{\mathrm{pix}}{\mathcal{L}}_{\mathrm{pix}}+{\lambda }_{\mathrm{per}}{\mathcal{L}}_{\mathrm{per}},$$

(39)

where ${\lambda }_{\mathrm{pix}}$ and ${\lambda }_{\mathrm{per}}$ denote weighting coefficients determined through hyperparameter optimization to balance reconstruction accuracy and perceptual quality.

4. Experiments

This section presents a comprehensive comparison between the proposed AMBFF framework and representative state-of-the-art low-light image enhancement methods on widely used benchmark datasets. The datasets used for evaluation are first described, followed by the implementation details of the proposed model. Quantitative and qualitative comparisons are then conducted using standard evaluation metrics to assess enhancement performance. An ablation study is further performed to analyze the contribution of each feature branch within the adaptive fusion mechanism. Finally, the computational complexity of the model is examined in terms of the number of parameters, multiply–accumulate operations (MACs), and floating-point operations (FLOPs).

4.1. Datasets

The proposed AMBFF framework is trained and evaluated on two widely adopted benchmark datasets for low-light image enhancement: LOL [20] and LOL-V2 [47] (including both the real and synthetic subsets). These datasets encompass a diverse range of illumination conditions, scene categories, and exposure variations, providing a comprehensive evaluation environment for assessing the robustness and generalization capability of the proposed approach under challenging low-light scenarios.

The LOL dataset [20] contains 500 paired low-light and normal-light images, of which 485 pairs are used for training and the remaining 15 pairs are reserved for testing. All image pairs are captured in real-world environments, where the low-light observations are obtained by adjusting camera exposure settings while preserving consistent scene content.

The LOL-V2 dataset [47] extends the original LOL benchmark by introducing two complementary subsets that further increase illumination diversity. The LOL-V2-Real subset contains 689 training pairs and 100 testing pairs acquired using multiple camera devices under varying lighting conditions. In addition, the LOL-V2-Synthetic subset includes 900 training pairs and 100 testing pairs generated through synthetic illumination manipulation, enabling the evaluation of enhancement performance across a broader spectrum of lighting degradations.

4.2. Implementation Details

All experiments are implemented in PyTorch 2.10.0 version and executed on the TRUBA [48] High-Performance Computing (HPC) infrastructure using CUDA-enabled GPUs. Input images are normalized to the [0, 1] range. During training, random cropping together with horizontal and vertical flipping are applied for data augmentation, while full-resolution images are used during validation and testing.

Model hyperparameters are optimized using the Optuna [49] framework. The search space includes batch size, learning rate, optimizer {AdamW, Adam, RMSprop, SGD}, base channel width, and network depth. Among the evaluated optimizers, AdamW consistently provided the best performance.

A two-stage training protocol is adopted. In the first stage, Optuna-based hyperparameter optimization is conducted using reduced folds and training epochs to efficiently explore the search space. In the second stage, the best configuration obtained from the tuning phase is used for full 10-fold cross-validation. Early stopping is employed based on validation PSNR, and the final model is selected according to the highest validation performance.

4.3. Comparison with State-of-the-Art

The effectiveness of the proposed AMBFF network is quantitatively evaluated by comparing its performance with several conventional image enhancement techniques, including Histogram Equalization (HE), Adaptive Histogram Equalization (AHE) [6], Contrast Limited Adaptive Histogram Equalization (CLAHE) [7], and Single-Scale Retinex [9]. In addition, representative deep learning-based approaches such as RetinexNet [20], EnlightenGAN [22], and AEHM [50] are included for comparison.

The evaluation is conducted using three widely adopted image quality metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [51], and Learned Perceptual Image Patch Similarity (LPIPS) [52].

The PSNR measures the reconstruction fidelity between an enhanced image and its corresponding reference image by quantifying the error between them. It is defined as

$$\mathrm{PSNR}(\widehat{X},X)=10{log}_{10}\left({\displaystyle \frac{MA{X}^2}{\mathrm{MSE}(\widehat{X},X)}}\right),$$

(40)

where $MAX$ denotes the maximum possible pixel intensity value, and $\mathrm{MSE}(\widehat{X},X)$ is the mean squared error between the enhanced image $\widehat{X}$ and the ground-truth image X, given by

$$\mathrm{MSE}(\widehat{X},X)={\displaystyle \frac{1}{CHW}}{\displaystyle \sum _{c=1}^C\sum _{i=1}^H\sum _{j=1}^W}{\left({\widehat{X}}_{c,i,j}-X_{c,i,j}\right)}^2.$$

(41)

A higher PSNR value indicates lower reconstruction error and thus better enhancement quality.

The SSIM evaluates perceptual similarity by jointly comparing luminance, contrast, and structural information between the enhanced image $\widehat{X}$ and the reference image X. It is defined as

$$\mathrm{SSIM}(\widehat{X},X)={\displaystyle \frac{(2{\mu }_{\widehat{X}}{\mu }_X+C_1)(2{\sigma }_{\widehat{X}X}+C_2)}{({\mu }_{\widehat{X}}^2+{\mu }_X^2+C_1)({\sigma }_{\widehat{X}}^2+{\sigma }_X^2+C_2)}},$$

(42)

where ${\mu }_{\widehat{X}}$ and ${\mu }_X$ denote the mean intensities of the two images, ${\sigma }_{\widehat{X}}^2$ and ${\sigma }_X^2$ represent their variances, and ${\sigma }_{\widehat{X}X}$ denotes the covariance between the two images. The constants $C_1$ and $C_2$ are small positive values introduced to ensure numerical stability. A higher SSIM value indicates greater structural similarity and thus better perceptual quality.

The LPIPS metric measures perceptual discrepancy using deep feature representations extracted from a pretrained neural network. In this study, LPIPS is computed using an AlexNet [53] backbone. Unlike PSNR and SSIM, lower LPIPS values indicate smaller perceptual distance between the enhanced image and the reference image, corresponding to better visual similarity and more natural enhancement results.

Table 1. Method comparison across datasets. For PSNR/SSIM, higher is better (↑); for LPIPS, lower is better (↓). Best values are highlighted in bold, and second-best values are underlined.

Method	LOL			LOL-V2-Real			LOL-V2-Syn
	PSNR ↑	SSIM ↑	LPIPS ↓	PSNR ↑	SSIM ↑	LPIPS ↓	PSNR ↑	SSIM ↑	LPIPS ↓
HE	14.2356	0.4502	0.3354	12.9819	0.4289	0.2765	15.7284	0.7742	0.1672
AHE	13.2092	0.3129	0.4622	14.1108	0.3308	0.3819	13.1813	0.5914	0.2314
CLAHE	9.6548	0.4394	0.2805	12.2986	0.4943	0.2191	12.7368	0.5888	0.2001
Retinex	14.9772	0.6855	0.2067	15.0225	0.6565	0.1613	14.4796	0.8427	0.1703
RetinexNet	16.7740	0.5364	0.2833	16.0972	0.5123	0.3352	17.1365	0.7943	0.1596
EnlightenGAN	17.4834	0.7164	0.1851	18.6396	0.7284	0.1828	16.5726	0.8054	0.1313
AEHM	17.8859	0.6451	0.1775	16.7033	0.6193	0.1533	19.2727	0.8818	0.1013
AMBFF (Proposed)	20.1490	0.8211	0.0671	18.6617	0.8098	0.1211	22.5394	0.9055	0.0507

reports quantitative comparisons of PSNR, SSIM, and LPIPS for all methods evaluated on the benchmark datasets. Figure 2 presents qualitative comparisons with representative state-of-the-art methods. As shown in Figure 2, the proposed AMBFF framework produces enhanced images with consistent luminance, improved structural clarity, and accurate color reproduction.

4.4. Ablation Study

A comprehensive ablation study is conducted to analyze the contribution of each feature branch and to evaluate the effectiveness of the adaptive fusion mechanism. Six configurations are considered: the full model and five variants obtained by excluding one branch at a time (w/o Spatial, w/o LC, w/o ET, w/o Fourier, w/o Hist). To ensure architectural consistency, excluded branches are replaced with zero tensors of identical dimensions rather than being physically removed from the network, preserving identical fusion structure and parameterization across configurations.

Table 2. Ablation study comparison between Fixed and Adaptive fusion strategies. For PSNR/SSIM, higher is better (↑); for LPIPS, lower is better (↓). Best and worst results are highlighted in bold and underlined, respectively.

	Fixed			Adaptive
Configuration	PSNR ↑	SSIM ↑	LPIPS ↓	PSNR ↑	SSIM ↑	LPIPS ↓
w/o Spatial	18.5000	0.8009	0.0744	19.7324	0.8046	0.0754
w/o Luma-Chroma	18.7067	0.7985	0.0732	20.0387	0.7983	0.0769
w/o Edge-Texture	19.6244	0.8027	0.0781	20.6979	0.8031	0.0830
w/o Fourier	17.5993	0.7843	0.0912	20.6594	0.8057	0.0700
w/o Histogram	18.9747	0.8123	0.0729	20.2409	0.8125	0.0725
Full	20.4983	0.8180	0.0749	20.1490	0.8211	0.0671

reports quantitative results under both fixed-weight and adaptive fusion strategies. The results demonstrate that the spatial and Fourier branches provide substantial performance gains under fixed fusion, indicating their importance in modeling global illumination and contextual structure. However, performance degradation becomes more pronounced when fixed fusion is employed and individual branches are removed.

In contrast, the adaptive fusion mechanism significantly mitigates performance drops across all ablation settings. By dynamically reallocating representational importance to the remaining active branches, adaptive fusion preserves structural fidelity and perceptual quality even when specific feature domains are excluded. This behavior confirms the complementarity of the proposed multi-domain representations and validates the necessity of spatially adaptive feature weighting.

Figure 3 further provides qualitative comparisons between fixed and adaptive fusion under branch exclusion. The adaptive model consistently produces more stable luminance correction and fewer structural artifacts, particularly in severely underexposed regions. These findings demonstrate that adaptive multi-branch fusion enhances robustness and prevents over-reliance on any single feature domain.

4.5. Computational Complexity

The computational efficiency of the proposed AMBFF framework is evaluated in terms of the number of learnable parameters, Multiply–Accumulate operations (MACs), and Floating-Point Operations (FLOPs) per forward pass. All complexity measurements are computed under identical input resolution to ensure fair comparison with competing methods.

Table 3. Comparison of model complexity in terms of parameters, MACs, and FLOPs per forward pass. Lower values indicate more efficient models. The best (lowest) results are shown in bold, and the second-best results are underlined.

Model	Parameters	MACs	FLOPs
RetinexNet	555.21 K	21.73 G	43.46 G
EnlightenGAN	22.53 M	59.34 G	118.68 G
AEHM	8.66 M	8.65 M	17.30 M
AMBFF (Proposed)	3.26 M	51.98 G	103.95 G

presents the complexity comparison against representative state-of-the-art approaches. The proposed AMBFF model contains 3.26M parameters and requires 51.98G MACs (approximately 103.95G FLOPs) per inference. Compared with large generative models such as EnlightenGAN, AMBFF achieves significantly lower parameter count while maintaining competitive computational cost. Although AMBFF involves multiple feature branches, the shared backbone and lightweight branch designs prevent excessive parameter growth.

These results demonstrate that structured multi-domain feature modeling combined with adaptive fusion provides strong enhancement performance without disproportionately increasing model size. The complexity–performance trade-off indicates that AMBFF remains computationally practical for high-resolution low-light enhancement while achieving superior reconstruction fidelity and perceptual quality.

5. Discussion

The experimental results demonstrate that the proposed AMBFF framework consistently achieves superior performance across LOL, LOL-V2-Real, and LOL-V2-Synthetic benchmarks in terms of PSNR, SSIM, and LPIPS. These improvements indicate that adaptive multi-domain feature modeling enables both high reconstruction fidelity and enhanced perceptual realism under diverse illumination conditions.

The effectiveness of AMBFF can be attributed to its explicit decomposition of low-light degradation into complementary feature domains. While spatial convolutional features capture local structural patterns, they do not explicitly encode global intensity statistics or frequency characteristics. The inclusion of frequency-domain and differentiable histogram representations enables the network to model both global tonal compression and high-frequency detail attenuation, which are typical in severely underexposed regions. The adaptive gating mechanism further enhances robustness by dynamically reweighting feature contributions according to local illumination severity, mitigating the limitations of fixed or loosely coupled fusion strategies.

Compared with traditional histogram-based and Retinex-inspired methods, AMBFF avoids common artifacts such as over-enhancement, halo effects, and color distortion. Unlike generative adversarial approaches, which may emphasize perceptual sharpness at the cost of structural consistency, the proposed framework maintains a balanced optimization between distortion-based and perceptual metrics. The ablation study with zero-branch substitution confirms that each domain contributes complementary information, and that adaptive fusion substantially improves resilience to branch exclusion, highlighting the structural redundancy embedded within the architecture.

From a computational perspective, although AMBFF incorporates multiple branches, the shared backbone and lightweight branch designs prevent excessive parameter growth. The resulting complexity–performance trade-off indicates that structured multi-domain modeling can yield meaningful gains without resorting to substantially larger backbones or iterative enhancement loops.

Nevertheless, certain limitations remain. The reliance on supervised training constrains performance to the distribution of available datasets, and extremely low signal-to-noise scenarios may still challenge stable reconstruction. In addition, while the histogram branch improves global tonal correction, it does not explicitly model sensor-specific noise characteristics, which could be beneficial for real-world deployment. Future work may explore domain adaptation strategies, noise-aware modeling, or transformer-based global context integration to further enhance robustness and scalability.

Overall, the findings suggest that adaptive multi-branch feature fusion constitutes a principled and effective paradigm for low-light image enhancement, offering improved stability, interpretability, and generalization across challenging illumination conditions.

6. Conclusions

This paper presented AMBFF, an Adaptive Multi-Branch Feature Fusion framework for low-light image enhancement. The proposed approach formulates enhancement operation as a multi-domain representation problem, explicitly modeling spatial context, luminance–chrominance decoupling, edge–texture structure, frequency-domain cues, and differentiable tonal distributions within a unified architecture. A spatially adaptive fusion mechanism enables dynamic integration of multiple feature domains, allowing the network to balance illumination correction, structural preservation, and color fidelity across diverse lighting conditions.

Extensive evaluations on LOL and LOL-V2 benchmarks demonstrated consistent improvements over conventional and recent deep learning methods in terms of PSNR, SSIM, and LPIPS. The ablation study confirmed the complementarity of the proposed feature branches and highlighted the robustness gains provided by adaptive fusion. Furthermore, computational analysis showed that AMBFF achieves a favorable performance–complexity trade-off despite its multi-branch design.

Future work will investigate lightweight model compression strategies and real-time deployment, as well as extending adaptive multi-domain fusion to other low-level vision tasks such as deblurring, denoising, and image restoration under adverse conditions.