Prior-based techniques rely on assumptions derived from atmospheric optics. The Dark Channel Prior (DCP), proposed by He et al. [5], remains one of the most influential methods in this category. Other approaches, such as Retinex-based enhancement [12] and non-local priors [13], have also been widely used. However, despite their computational efficiency, these methods often fail in sky regions and under dense haze conditions.

CNN-based methods learn haze-related features directly from training data. DehazeNet [6] estimates transmission maps using learned features, while Multi-Scale CNN [7] incorporates multi-scale learning to handle varying haze densities. AOD-Net [8] reformulates the atmospheric model to enable end-to-end training. Although effective, these models often suffer from color distortion and limited generalization to real-world haze.

GAN-based methods utilizing adversarial loss to enhance perceptual realism. Conditional GANs (cGANs) [14][15] have been widely used for paired image dehazing tasks. DHSGAN [16] demonstrated better robustness under haze and smoke conditions, while CycleGAN-based methods [17][18] enabled unpaired training at the cost of structural consistency. Another GAN-based dehazing models that used attention-driven feature fusion to produce superior image [19]-[21]. Extensive experimental evidence indicates that GAN-based approaches outperform traditional CNN-based dehazing models in perceptual quality, although their training stability and architectural flexibility remain challenging [10][11].

Despite significant progress in CNN- and GAN-based image dehazing, the review of existing literature reveals several unresolved challenges. Most CNN-based methods rely on single-stream architectures that inadequately capture the hierarchical nature of haze formation, while many GAN-based approaches prioritize perceptual realism without explicitly integrating multi-level haze-aware representations or physically motivated constraints. In addition, existing studies often evaluate models under limited experimental settings, with insufficient discussion on architectural design rationale, benchmarking strategy, and practical limitations. To address these gaps, the present manuscript introduces a multi-CNN-based GAN framework that explicitly integrates hierarchical feature learning within an adversarial setting based on the atmospheric scattering model and informed by insights from prior comparative analyses. Furthermore, this work emphasizes a good experimental design, including dataset selection, implementation strategy, evaluation metrics, and benchmark comparisons, thereby providing a structured foundation for implementation and extension of the proposed framework.

The majority of dehazing techniques are based on the atmospheric scattering model [1]-[4], expressed as

therefore, the haze-free image is

where I ( x ) denotes the observed hazy image at pixel position x , J ( x ) represents the haze-free scene radiance, t ( x ) is the transmission map, and A denotes global atmospheric light. The transmission map is commonly modeled through an exponentially decaying function as

where β represents the atmospheric scattering coefficient and d ( x ) represents scene depth. Figure 1 shows the flow diagram of haze-free image from the observed hazy image.

Figure 1. A simple image dehazing flow diagram based on scattering model.

Single image dehazing is an inherently ill-posed problem due to unknown depth information, spatially varying haze density, and complex illumination conditions [5]-[7]. An effective dehazing model must balance visibility enhancement with color fidelity and structural preservation.

The proposed framework comprises a multi-CNN generator and a discriminator trained under an adversarial learning system. The generator integrates three parallel CNN branches followed by hierarchical feature fusion and reconstruction layers. The three parallel CNNs comprising 1) a low-level feature extraction network that captures edge information, color attenuation, and haze density cues; 2) a mid-level transmission-aware network that learns representations related to atmospheric scattering and depth-dependent haze; and 3) a high-level semantic refinement network that preserves global structure and semantic consistency. The feature maps from all three CNNs are fused through a hierarchical feature fusion module and passed to reconstruction layers to generate the final dehazed image. A PatchGAN discriminator with a patch size (70 × 70) is used to evaluate local realism by distinguishing generated dehazed images from ground-truth haze-free images, thereby guiding the generator through adversarial feedback.

The proposed framework is theoretically grounded in the atmospheric scattering model, the current GAN implementation is data-driven rather than explicitly estimating t ( x ) and A as separate outputs for the final pass. The “mid-level transmission-aware network” is designed to implicitly learn these representations. In fact, the physics-based parameters provide the conceptual motivation for the multi-branch architecture rather than being explicitly supervised auxiliary outputs.

In the preprocessing, the images are normalized and resized to a consistent resolution (256 × 256). Figure 2 shows our dehazing framework.

Figure 2. The framework for image dehazing using multi-CNN-based GAN.

The working algorithm of this proposed model is shown in Algorithm 1.

Algorithm 1:CNN-GAN-based dehazing

Input: Hazy image I.

Output: Clear dehazed image J through minimizing haze while preserving naturalness and fine details.

Step 1: Preprocessing

Normalize and preprocess I.

Step 2: Multi-CNN GeneratorG

Input Layer: Accept the preprocessed hazy image.Feature Extraction Layers: Use multiple parallel convolutional layers with varying kernel sizes to capture multi-scale features.Attention Mechanism: Integrate a self-attention module to focus on regions with haze.Residual Blocks: Include residual blocks with skip connections to enhance feature learning and prevent vanishing gradients.Multi-scale Fusion: Fuse features extracted at multiple scales to generate dehazed image Ĵ = G(I) with fine details.

Step 3: CNN DiscriminatorD

Discriminator D evaluates Ĵ against ground truth.Compute reconstruction, perceptual, and adversarial losses.Update G to minimize total loss.Update D to maximize adversarial discrimination.Iterate until convergence, that means, the generator produces visually appealing dehazed images.Return clear dehazed image J.

Step 4: Optional post-processing

Use a sharpening filter to enhance J through contrast enhancement and artifact removal.

The proposed framework adopts an adversarial learning strategy based on the standard min-max formulation of Generative Adversarial Networks. Let G denotes the generator network and 𝐷 denotes the discriminator network. The adversarial objective is defined by the standard min–max formulation [9]

where I represents a hazy input image sampled from the probability distribution of hazy images p_I, and Jdenotes the corresponding ground-truth haze-free image sampled from the probability distribution p_J. The generator G maps a hazy image I to a dehazed image G(I), while the discriminator D outputs the probability that an input image is a real haze-free image.

During training, the discriminator aims to correctly distinguish real haze-free images J from generated images G(I), whereas the generator attempts to produce dehazed images that are indistinguishable from real haze-free images. This adversarial process encourages the generator to learn realistic image distributions and improves the perceptual quality of the dehazed results.

D(G(I)) is the discriminator’s estimated probability that the generated image G(I) is a real haze-free image.

E[⋅] is the expectation operator, computed over the data distribution of real haze-free images J and hazy images I.

To ensure both pixel-level accuracy and perceptual realism [22], the generator is optimized using a combined loss function defined as

where L_G is the combined loss function, L_adv is the adversarial loss,L_rec is the reconstruction loss and L_perc is a perceptual loss computed using deep feature representations extracted from a pre-trained network, such as using a VGG-16 or VGG-19 backbone on ImageNet, specifically targeting the feature maps from the deeper layers to ensure semantic consistency [7][19] and λ₁, λ₂, and λ₃ are positive weighting coefficients (For example, λ₁ = 1, λ₂ = 100 and λ₃ = 1 is a good choice. The reconstruction loss is weighted more heavily) that balance the contribution of each loss component for better image-to-image translation.

The adversarial loss for the generator is formulated as

This term encourages the generator to produce dehazed images that the discriminator classifies as real, thereby enhancing visual realism and reducing perceptual artifacts.

The reconstruction loss enforces pixel-wise similarity between the generated image and the ground-truth haze-free image. It is defined as

where ‖ ⋅ ‖ denotes the L₁ or L₂ norm. This loss preserves structural details, color consistency, and low-level image fidelity, which are critical for effective haze removal. However, the use of the L₁ norm preserve the edge sharpness better.

The perceptual loss [22] measures the difference between high-level feature representations of the generated and ground-truth images, computed using a pre-trained deep network as mentioned earlier. It is expressed as

where ϕ_l(⋅) denotes the feature map extracted from the l-th layer of the pre-trained network. This loss promotes semantic consistency, texture preservation, and global perceptual coherence.

The combined loss formulation enables the generator to jointly optimize adversarial realism, pixel-level accuracy, and perceptual quality. Such a balanced objective is particularly important for single image dehazing, where excessive reliance on reconstruction loss may cause over-smoothing, while adversarial loss alone may introduce structural distortions. The proposed composite loss thus ensures robust and visually convincing dehazing performance.

To evaluate the effectiveness and generalization capability of the proposed CNN-GAN framework, both synthetic and real-world image dehazing datasets should be employed. Diverse datasets are essential to capture variations in haze density, scene depth, illumination conditions, and real atmospheric effects. Synthetic datasets provide paired hazy and haze-free images, enabling supervised training and objective quantitative evaluation. Among these, the RESIDE benchmark [23] is the most widely adopted dataset for single image dehazing research and consists of multiple subsets, including the Outdoor Training Set (OTS), Synthetic Objective Testing Set (SOTS), and Hybrid Subjective Testing Set (HSTS). RESIDE has been extensively used to evaluate CNN- and GAN-based dehazing models due to its controlled haze synthesis and availability of ground-truth images.

In addition to RESIDE, other synthetic datasets commonly used in the literature include the I-HAZE [24] and O-HAZE [25] datasets, which provide real haze scenes with corresponding haze-free ground truth captured under controlled conditions. These datasets bridge the gap between fully synthetic data and real atmospheric environments. Furthermore, Dense-Haze [26] is another dataset characterized by dense and homogeneous hazy scenes, which were generated by professional haze machines. The hazy and haze-free corresponding scenes contain the same visual content captured under the same illumination parameters.

Real-world datasets, which lack ground-truth haze-free images, are critical for assessing model robustness and practical applicability. Unpaired real-world hazy images collected from outdoor environments and traffic scenes are frequently used for qualitative evaluation in recent studies [6]-[8][10][11][15]-[17][19]-[21]. Such real-world datasets expose models to non-uniform haze, complex lighting, and color distortions that are difficult to simulate synthetically.

By training and evaluating the proposed framework across a combination of synthetic and real-world datasets, a balanced assessment of reconstruction accuracy, perceptual quality, and generalization performance can be achieved.

The current framework is primarily designed for paired (supervised) learning.

Training Datasets: For supervised training, the framework utilizes paired datasets such as the RESIDE Outdoor Training Set (OTS), I-HAZE, and O-HAZE.Testing Datasets: Performance is evaluated using the RESIDE Synthetic Objective Testing Set (SOTS) and Hybrid Subjective Testing Set (HSTS), as well as Dense-Haze.

While training is conducted in a paired setting, however, to assess generalization, unpaired real-world images are used solely for qualitative (no-reference) testing.

The proposed CNN-GAN framework is intended to be implemented using a deep learning platform such as PyTorch or TensorFlow, which supports flexible construction of adversarial architectures and efficient GPU-based training. The generator network adopts a multi-branch (3-branch) CNN design, where each branch is specialized for learning complementary representations at different semantic levels.

1) Generator Architecture

Each CNN branch in the generator is composed of a sequence of convolutional layers with small kernel sizes (e.g., 3 × 3) to preserve spatial resolution while capturing local haze-related features followed residual blocks. Residual learning is adopted through the residual blocks for stable gradient propagation and speedy convergence. Batch normalization, followed by non-linear activation functions such as ReLU or Leaky ReLU, is used in the generator branches to stabilize training and prevent vanishing gradients.

The three branches do not share weights; they are specialized to learn complementary representations at different semantic levels (low-level, mid-level, and high-level). The low-level feature extraction branch focuses on capturing edge information, color attenuation, and local contrast degradation. The mid-level transmission-aware branch is used to learn depth-dependent haze characteristics and atmospheric scattering effects. The high-level semantic refinement branch emphasizes global structure and semantic consistency to prevent distortion of scene content. The fusion of feature maps from all branches is performed via a hierarchical feature fusion module that employs concatenation followed by additional refinement convolutional layers. The fused representation is then passed to reconstruction layers to generate the final dehazed image.

2) Discriminator Architecture

The discriminator is implemented as a PatchGAN [15], which classifies local image patches rather than the entire image as real or fake. This process ensures local realism and sharpness in the generated dehazed images while reducing the number of model parameters. The discriminator consists of a series of convolutional layers with increasing feature depth and stride convolutions for spatial down sampling, followed by Leaky ReLU activations. Model training can be done through Adam optimizer with learning rate scheduling. Training continues until the loss curves for both G and D reach a stable equilibrium.

To ensure a comprehensive and fair assessment of the proposed CNN-GAN framework, both full-reference and no-reference image quality metrics should be employed. Since image dehazing aims to improve not only pixel-level accuracy but also perceptual quality and visual realism, reliance on a single metric is not enough. Therefore, multiple complementary evaluation criteria are adopted, as commonly practiced in prior dehazing studies [10][11][27][28].

1) Full-Reference Metrics

Full-reference metrics [11] require access to corresponding ground-truth haze-free images and are primarily used for evaluating performance on synthetic datasets. There are some full-reference metrics: Peak Signal-to-Noise Ratio (PSNR) is used to measure pixel-wise reconstruction fidelity between the generated dehazed image and the ground-truth image. Higher PSNR values indicate better reconstruction accuracy and lower distortion. Structural Similarity Index Measure (SSIM) evaluates perceptual similarity by considering luminance, contrast, and structural information. SSIM is more consistent with human visual perception than PSNR and is widely adopted in image dehazing research. For a model, PSNR and SSIM together express the insight of the detailed restoration and structural consistency preservation in haze-free regions.

2) No-Reference (Blind) Image Quality Metrics

No-reference metrics [27] are essential for evaluating real-world hazy images where ground-truth references are unavailable. There are some full-reference metrics: Natural Image Quality Evaluator (NIQE) measures deviations from natural scene statistics without requiring reference images. Lower NIQE scores correspond to better perceptual quality and have been widely used to assess real-world dehazing results. Blind Image Spatial Quality Evaluator (BRISQUE) assesses image quality based on locally normalized luminance coefficients and natural scene statistics. It complements NIQE by focusing on spatial distortions introduced during dehazing. These blind metrics are particularly important for assessing the generalization capability of dehazing models under uncontrolled atmospheric conditions.

3) Perceptual Quality Metrics

To better quantify perceptual similarity aligned with human judgment, Learned Perceptual Image Patch Similarity (LPIPS) [28] can be employed. LPIPS computes distances between deep feature representations extracted from pre-trained networks and has been shown to correlate well with human perceptual assessments. Lower LPIPS values indicate higher perceptual similarity.

4) Subjective Evaluation

In addition to objective metrics, subjective visual evaluation is recommended to assess color fidelity, residual haze, halo artifacts, and overall visual realism. User studies involving human observers have been commonly used in recent dehazing literature to complement quantitative results [11].

By jointly considering full-reference, no-reference, perceptual, and subjective metrics, a balanced and reliable evaluation of dehazing performance can be achieved across both synthetic and real-world datasets.