Alpha-DehazeNet: single image dehazing via RGBA haze modeling and adaptive learning

Physics-driven haze estimation

The atmospheric scattering model proposed by Koschmieder (1925) has served as the cornerstone of physics-based dehazing methods. Building on this framework, He, Sun & Tang (2010) introduced the dark channel prior (DCP), which estimates transmission maps by exploiting the statistical dominance of low-intensity pixels in local haze-free image patches. While DCP demonstrates efficacy in homogeneous haze conditions, its fundamental assumptions collapse in sky regions and dense urban fog due to invalid intensity distributions (Zhu, Mai & Shao, 2015). Subsequent approaches, such as haze-line clustering (Berman, treibitz & Avidan, 2016) and non-local regularization techniques (Sulami et al., 2014), attempted to address spatially varying haze at the expense of computational practicality. A critical limitation persists across these methods: the implicit assumption of uniform haze density directly contradicts empirical observations of layered fog dynamics and localized atmospheric pollutants in real-world environments (Zheng et al., 2022).

Data driven dehazing architectures

The advent of deep learning has shifted the paradigm toward data-driven solutions. Early convolutional neural networks (CNNs) like DehazeNet and AOD-Net (Li et al., 2017) pioneered direct mappings from hazy to clean images, achieving notable success on synthetic benchmarks. However, their reliance on artificially generated training data introduces significant domain adaptation challenges, as evidenced by performance degradation on real-world hazy images with complex illumination encompassing daytime varying light and nighttime low-light conditions and aerosol distributions (Ancuti et al., 2019a). Most prior research has focused on daytime dehazing, where methods such as the multi-scale CNN with holistic edges effectively handle diverse lighting to preserve scene details (Ren et al., 2020). In contrast, nighttime dehazing presents unique challenges due to non-uniform light sources and color distortion, spurring specialized models like the hybrid-regularized variational framework (Liu et al., 2025), the unified variational Retinex approach (Liu et al., 2022), and the prior query Transformer in to enhance visibility in low-light scenarios (Liu et al., 2023). Transformer-based architectures (Song et al., 2023) later emerged to model long-range dependencies more effectively, yet their computational complexity hinders deployment in latency-sensitive applications. Recent 4K dehazing research introduces specialized methods and datasets to preserve details in high-resolution scenes (Zheng & Jia, 2023). A persistent trade-off plagues these methods: aggressive haze suppression often erodes high-frequency textures, while conservative approaches retain residual artifacts—a dichotomy rooted in the lack of physics-aware constraints during network optimization.

Attention mechanisms and hybrid frameworks

Recent hybrid methodologies seek to bridge the gap between physical interpretability and data-driven performance. DCNet (Yi et al., 2022) cascades transmission estimation modules with refinement networks, while PAD-Net (Yi et al., 2024a) integrates edge-aware filtering to mitigate artifact generation. Parallel developments in attention mechanisms, exemplified by channel attention of FFA-Net (Qin et al., 2020) and dual-branch spatial-frequency designs (Wang et al., 2024), selectively enhance haze-dense regions.

Contrastive learning and multi-stage frameworks

Recent advancements leverage contrastive learning to bridge synthetic-to-real domain gaps. SID-Net (Yi et al., 2024b) combines adversarial training with contrastive regularization, improving generalization on real-world haze. A multi-scale dehazing network with dark channel priors (Yang et al., 2023) progressively refine haze removal through global-local feature interaction. However, these methods still rely on fixed atmospheric parameter assumptions, limiting adaptability to spatially varying haze.

Nevertheless, three fundamental limitations remain unaddressed: (1) static parameterization of atmospheric light and transmission values in hybrid frameworks prevents adaptation to dynamic haze layers; (2) existing attention strategies treat haze localization and detail preservation as decoupled objectives, leading to feature conflicts in partially occluded regions; and (3) no current method explicitly models the intrinsic correlation between haze transparency and density gradients—a critical factor for preserving perceptual depth cues in complex scenes.

Architecture of the Alpha-DehazeNet network

To allow for dynamic parameter adjustment and incorporate atmospheric physical model parameters into the dehazing framework during training, this study introduces a haze effect generation model based on the RGBA color space. RGBA extends the traditional RGB format by adding an alpha channel, which serves as an opacity parameter. The alpha channel facilitates combining images with backgrounds, enabling the generation of partially or fully transparent visual effects (Tan, Lien & Gingold, 2016). In the context of dehazing, incorporating the alpha channel allows precise control over haze layer transparency, making it possible to analyze haze effects with varying densities.

In real-world environments, haze density often varies with distance and surrounding conditions. The RGBA format provides the flexibility to adjust the transparency of individual pixels, enabling more realistic haze simulations. Compared to the standard RGB format, RGBA offers enhanced representation of transparency and depth. An example of the RGBA effect is illustrated in Fig. 2.

As illustrated in Fig. 2, the Alpha channel is used to achieve layer compositing. The exact implementation effect is described by Eq. (2):

(2) $$S=\alpha F+(1-\alpha )B.$$

In this equation, $\mathit{F}$ represents the foreground image, while $\mathit{B}$ represents the background image. By applying Eq. (2), the two images can be combined to form the composite image $\mathit{S}$. This process allows the RGBA format to be converted into a haze effect image for simulating the haze layer. Compared to Eq. (1), $\mathit{J}(\mathbf{x})$ can be considered the original haze-free image, while $\mathit{A}$ represents the global atmospheric light value. In Eq. (1), $\mathit{A}$ is represented as a three-dimensional array; however, $\mathit{F}$ can denote the atmospheric light value at each specific location in the image, which can be understood as the haze layer image. Consequently, can be expressed as $1-t(x)$, representing the non-transmissivity or opacity of the haze layer.

Based on the RGBA color space, this study introduces the Alpha-DehazeNet dehazing network. The overall architecture of the network is shown in Fig. 3.

As illustrated in Fig. 3, the model network structure comprises two main components. The first is RGBA-Net, a model designed to generate haze effect images. This model utilizes the original hazy image $I_{gt}$ and a randomly generated non-transparent RGBA image $R_{rgba}$ to produce the haze effect of the original image. The second component is REC-Net, a model aimed at optimizing the dehazing effect. Using Eq. (2), the haze effect image $F_{rgba}$ and the original hazy image $I_{gt}$ can be combined to generate a preliminary dehazed image $J_{gen}$, which cancan generate an optimized dehazed image $J_{rec}$, thereby enhancing the dehazing effectiveness of the model. Furthermore, by applying the inverse formula of Eq. (2), the optimized dehazed image $J_{rec}$ and haze effect image $F_{rgba}$ are used to generate a hazy image $I_{rec}$. The model enhances the quality of the haze effect image $F_{rgba}$ by comparing the optimized dehazed image $J_{rec}$ with the real haze-free image $J_{gt}$ and comparing the generated hazy image $I_{rec}$ with the original hazy image $I_{gt}$.

RGBA-Net: a model for generating haze effect images

In this study, a deep learning model is designed based on the RGBA color space to generate haze effect images. The detailed structure of the RGBA-Net is illustrated in Fig. 4.

As depicted in Fig. 4, the model starts by defining a random, opaque gray-white image in the RGBA space as the initial haze effect image $R_{rgba}$. The original hazy image $I_{gt}$ is then incorporated, and a U-Net is employed to generate the potential haze effect $F_{rgba}$ for the original image.

The resulting haze effect image $F_{rgba}$ is also in RGBA format. The transparency information provided by the alpha channel enables the model to more accurately simulate the semi-translucent effect of haze, allowing the generated image to not only retain color details but also exhibit the gradual fading or thickening of haze.

The inclusion of the alpha channel offers an additional information dimension during model training, which aids the model in comprehensively understanding the haze generation process. By processing both RGB and alpha data simultaneously, the model captures features related to haze layer thickness, opacity, and the distribution of haze across different locations, thus improving its ability to generate high-quality haze effects. The alpha channel also assists the model in learning how to generate smooth transitions in areas with varying transparency levels, further enhancing the realism of the generated images.

To improve the model’s ability to recognize different haze layers, the original hazy image is incorporated into the network. A simple neural network structure is used to extract and integrate its features, enhancing the model’s capacity to handle various haze intensities.

Moreover, to generate complex haze effects, the model combines channel attention (Zhang et al., 2020) and spatial attention mechanisms (Zhou et al., 2022). These mechanisms enhance and filter features across different dimensions (Jiang et al., 2021; Li, Hua & Li, 2023). The channel attention mechanism automatically selects the most relevant feature channels and assigns them higher weights. By weighting feature responses across different channels, the model highlights haze-related features while suppressing irrelevant or redundant ones. This enables the model to focus on the most significant features for haze generation, improving its ability to differentiate between different haze effects.

The spatial attention mechanism, in contrast, identifies critical locations within the feature map. Unlike channel attention, which focuses on feature channels, spatial attention emphasizes the spatial dimension. By identifying key areas with noticeable haze distribution, the model generates a spatial attention map that is multiplied with the original feature map, highlighting regions where the haze effect is more prominent.

Through the combination of channel and spatial attention mechanisms, the model effectively selects global features for haze generation, while also concentrating on local areas with more prominent haze, ensuring the generated haze effect maintains global consistency while achieving precise detail.

REC-Net: the dehazing optimization model

We propose the dehazing optimization model REC-Net, which is employed to refine the initially generated dehazed image $J_{gen}$, transforming it into a more precise and enhanced dehazed image $J_{rec}$. The architecture of the REC-Net is illustrated in Fig. 5.

As illustrated in Fig. 5, the dehazing optimization model also adopts Adversarial architecture. To enhance optimization accuracy, the model integrates a residual structure to improve the dehazing effect (Dudhane, Singh Aulakh & Murala, 2019; Qiu, Cheng & Wang, 2022). The output at each layer is not only combined with the feature map of the subsequent layer but also fused with the feature map from earlier layers through skip connections, which helps preserve the richness of low-level features. This approach allows the model to more effectively understand and reconstruct the image content, optimizing the initially generated dehazed image, and improving its detail and contrast.

Loss function

To ensure the performance of both the haze effect image generation model and the dehazing optimization model, we introduce a loss function, as presented in Eq. (3).

(3) $$L_{total}=\alpha L_{mse}+\beta L_{gan}+\gamma L_{dp}.$$

As illustrated in Eq. (3), the loss function $L_{total}$ is composed of the mean squared error loss $L_{mse}$, adversarial loss $L_{gan}$, and a custom deep consistency loss $L_{dp}$, where $\alpha$, $\beta$, $\gamma$ represents the weights for each of the losses. These losses work collaboratively to optimize the generated image from different angles.

The mean squared error loss is primarily used to measure the discrepancy between the generated dehazed image and the ground truth clear image, ensuring the accuracy of the generated result (Liu et al., 2018). Its calculation is presented in Eq. (4).

(4) $$L_{t}mse=(J_{rec}-J_{gt}{)}^2+(I_{rec}-I_{gt}{)}^2.$$

Furthermore, we incorporate adversarial loss to enhance the realism of the generated image (Goodfellow et al., 2020), as presented in Eq. (5).

(5) $$L_{gan}=E[\log D(J_{gt})]+E[\log (1\text{-}D(J_{rec}))]+E[\log D(I_{gt})]+E[\log (1\text{-}D(I_{rec}))].$$

Through a generative-adversarial interplay, the model enhances the realism of the generated image. The generator’s objective is to create dehazed images that are as realistic as possible, making it difficult for the discriminator to distinguish between real and generated images, while the discriminator strives to correctly identify the differences. This adversarial process allows the model to progressively improve its generation capabilities, resulting in dehazed images that are more natural and lifelike.

Finally, the depth consistency loss function is computed using histogram analysis, which helps ensure that the generated dehazed images maintain a more realistic geometric structure and preserve the three-dimensional spatial integrity of the scene (Jayanthi, Rajput & Indu, 2020). The first step is to convert the RGB image into a grayscale image, and apply as channel-weighting coefficients, as shown in Eq. (6).

(6) $$G_{image}(x,y)=w_r\cdot I_{image}(x,y,0)+w_g\cdot I_{image}(x,y,1)+w_b\cdot I_{image}(x,y,2).$$

Next, the grayscale image is converted into a histogram, and the calculation formula is provided in Eqs. (7), (8).

(7) $$H_{gray}(i)=\frac{{\sum }_{(x,y)}1(G_{image}(x,y)\in [i,i+1])}{N_{image}}$$

(8) $$D(i)=|H_{rec}(i)-H_{orig}(i)|.$$

Here, $N_{image}$ denotes the pixel values of the image. This function calculates the proportion of pixels that fall within the same interval, and the difference between the original and generated images is computed using the formula.

The depth consistency loss is ultimately calculated using the formula provided in Eq. (9).

(9) $$L_{\mathrm{d}\mathrm{p}}=1-\sum _{i=0}^{\mathrm{i}\mathrm{m}\mathrm{a}\mathrm{g}\mathrm{e}\mathrm{\_}\mathrm{s}\mathrm{i}\mathrm{z}\mathrm{e}}D(i).$$

The depth consistency loss function helps preserve the geometric consistency of the scene, ensuring that the depth information in the generated dehazed image is consistent with that of the real image. This prevents objects from appearing unnatural or distorted as a result of dehazing. Furthermore, it aids in maintaining the blur of distant objects and the clarity of nearby objects, thereby enhancing the overall visual effect.

In conclusion, the mean squared error loss, adversarial loss, and depth consistency loss collectively ensure that the generated dehazed images achieve optimal performance in terms of pixel-level accuracy, perceptual quality, and visual realism. These loss functions impose constraints on the model from different perspectives, improving the overall dehazing capability.

Datasets

To ensure a fair comparison, we evaluate the proposed model using both open-source synthetic datasets and real-world scene datasets.

RESIDE is a widely recognized benchmark dataset (Li et al., 2018). Among its five subsets, we select ITS and OTS as the training datasets and use SOTS-indoor and SOTS-outdoor as the testing datasets for synthetic image dehazing.

Two real-world scene datasets, Dense-Haze and NH-HAZE (each comprising 55 image pairs), are used to validate the model. Dense-Haze primarily contains dense, uniform haze scenes from various outdoor settings (Ancuti et al., 2019b), while NH-HAZE features non-uniform haze with significant depth variations (Ancuti, Ancuti & Timofte, 2020). These datasets provide complementary benchmarks for evaluating the model’s performance under diverse real-world haze conditions: Dense-Haze focuses on homogeneous fog, and NH-HAZE targets complex, layered distributions with spatial opacity gradients.

Ablation study

This section analyzes the role and impact of each component of the proposed Alpha-DehazeNet model. The model is divided into two parts: the first part, RGBA-Net, is responsible for haze layer recognition; the second part, REC-Net, is used for dehazing optimization. We trained each module using the ITS dataset and evaluated the dehazing results of each module on the SOTS-indoor and SOTS-outdoor datasets. The experimental results are shown in Table 2.

We analyze the loss function of Alpha-DehazeNet, where mean squared error (MSE) loss and adversarial loss act as the fundamental loss functions. To evaluate the impact of the depth consistency loss, an ablation study is performed. The ITS dataset is used for training, and the dehazing performance is tested on the SOTS-indoor and SOTS-outdoor datasets. The experimental results are presented in Table 3.