Physics-Driven Image Dehazing from the Perspective of Unmanned Aerial Vehicles

Abstract

Drone vision is widely used in change detection, disaster response, and military reconnaissance due to its wide field of view and flexibility. However, under haze and thin cloud conditions, image quality is usually degraded due to atmospheric scattering. This results in issues like color distortion, reduced contrast, and lower clarity, which negatively impact the performance of subsequent advanced visual tasks. To improve the quality of unmanned aerial vehicle (UAV) images, we propose a dehazing method based on calibration of the atmospheric scattering model. We designed two specialized neural network structures to estimate the two unknown parameters in the atmospheric scattering model: the atmospheric light intensity A and medium transmission t. However, calculation errors always occur in both processes for estimating the two unknown parameters. The error accumulation for atmospheric light and medium transmission will cause the deviation in color fidelity and brightness. Therefore, we designed an encoder-decoder structure for irradiance guidance, which not only eliminates error accumulation but also enhances the detail in the restored image, achieving higher-quality dehazing results. Quantitative and qualitative evaluations indicate that our dehazing method outperforms existing techniques, effectively eliminating haze from drone images and significantly enhancing image clarity and quality in hazy conditions. Specifically, the compared experiment on the R100 dataset demonstrates that the proposed method improved the peak signal-to-noise ratio (PSNR) and structure similarity index measure (SSIM) metrics by 6.9 dB and 0.08 over the second-best method, respectively. On the N100 dataset, the method improved the PSNR and SSIM metrics by 8.7 dB and 0.05 over the second-best method, respectively.

1. Introduction

During drone image acquisition, atmospheric conditions such as haze and thin clouds significantly impact image quality [1,2,3]. These conditions reduce the signal-to-noise ratio, cause color distortion, and even lead to information loss, limiting the application of drone images. Therefore, developing an effective dehazing method for drone images is crucial, as high-quality images are essential for the accuracy and reliability of subsequent tasks [4] such as change detection, disaster response, and military reconnaissance.

Early researchers proposed methods based on statistic priors to mitigate the impact of haze, such as the dark channel prior [5] and haze line prior [6]. However, these methods have poor generalization and struggle to handle the diverse distribution of haze in different scenarios. Nowadays, image restoration algorithms are mainly based on both convolutional neural network (CNN)- [7,8,9,10] and Transformer-based methods [11,12,13,14]. These methods are generally divided into two main categories.

The first category comprises neural networks without physical models, which perform dehazing by directly learning the mapping between hazy and clear images. Cai et al. [15] proposed DehazeNet, which uses a neural network to learn the direct mapping between hazy images and medium transmission and then estimates clear images using the atmospheric scattering model. Ren et al. [16] also used CNNs to predict medium transmission. However, the atmospheric scattering model is a simplified approximation of haze effects, and solely relying on medium transmission often fails to achieve high-quality reconstruction. Qin et al. [7] incorporated attention mechanisms to construct a deep CNN for directly restoring clear images. Despite their ability to directly obtain clear images, these non-physical model neural network methods may perform poorly when faced with non-uniform haze conditions.

The second category involves neural networks with physical models, which enhance the interpretability of the network by incorporating physical attributes. Li et al. [17] introduced the all-in-one dehazing network (AOD-Net), which uses a variant of the atmospheric scattering model to implicitly learn the atmospheric light A and medium transmission t through a neural network to solve for clear images. AOD-Net is a lightweight network which reduces the runtime effectively, but its dehazing capability is limited. Zhang et al. [18] estimated the atmospheric light A and medium transmission t using two CNN models and combined them with a generative adversarial network (GAN) to obtain a final clear image. Similarly, Li et al. [19] used comparable approaches. These methods leverage physical models to improve the stability and interpretability of dehazing results. However, stepwise parameter estimation can lead to accumulating errors, resulting in suboptimal outcomes. Moreover, due to complex haze distributions like haze clusters and non-uniform haze in drone images, dehazing is often more challenging than in natural scene images.

To address these issues, we propose a dehazing neural network based on the atmospheric scattering model. This network adheres to physical principles to enhance interpretability. In addition, to reduce the accumulation of errors in stepwise learning, we design a coding-decoding structure to guide the results obtained from the atmospheric scattering model, resulting in higher detail quality and more thorough dehazing. The main contributions of this paper are as follows:

1.

We propose a dual-branch dehazing architecture based on the atmospheric scattering model specifically designed to enhance the accuracy and stability of dehazing results for aerial images captured by drones, which often suffer from varying haze densities.

2.

We design a dual-branch residual attention module (DBRA) which generates spatial attention maps tailored to each feature channel. This allows for more precise handling of haze non-uniformity, particularly in drone images, where environmental factors can lead to significant variations in visibility.

3.

We design an encoder-decoder structure which simulates irradiance-guided strategies by inputting the dehazing results from the atmospheric scattering model into a neural network. This approach effectively guides irradiance, minimizes error accumulation, and enhances the quality of drone imagery.

2. Related Works

2.1. Dehazing Networks Based on Physical Models

Physical model-based dehazing networks aim to restore clear images from hazy ones by understanding and simulating the physical processes of haze formation. The core model is the atmospheric scattering model, which is expressed as follows:

$$I\left(x\right)=J\left(x\right)t\left(x\right)+A\left(1-t\left(x\right)\right)$$

(1)

where $I\left(x\right)$ represents the observed hazy image, $J\left(x\right)$ is the latent clear image, $t\left(x\right)$ is the medium transmission, and A is the global atmospheric light. The transmission map $t\left(x\right)$ encodes the portion of the scene radiance which reaches the camera without being scattered, while A represents the intensity of the ambient light scattered by the atmosphere. By accurately estimating $t\left(x\right)$ and A, the original scene radiance $J\left(x\right)$ can be effectively recovered, resulting in a haze-free image.

Early methods primarily relied on statistical priors and strong constraints related to haze or image degradation to estimate transmission maps and global atmospheric light. However, these methods exhibit significant limitations when addressing the complex nonlinear problems associated with image dehazing. Specifically, they encounter issues in several aspects. First, these methods often assume uniform lighting and color in a scene, making them ineffective in dealing with the variations in illumination and complex structures found in actual images. Second, the strong constraints restrict the flexibility of the models, leading to suboptimal restoration results when processing regions with complex textures or high contrast. Additionally, early methods are prone to noise and artifacts in the transmission map estimation, which further impact the final quality of the images. Recent advancements in dehazing have integrated the atmospheric scattering model directly into neural networks, allowing an end-to-end learning process to optimize the dehazing parameters and the final image quality simultaneously. This integration facilitates the joint estimation of medium transmission, atmospheric light, and the haze-free image, enabling the network to learn the intrinsic relationships between these parameters during training. A typical example of a densely connected pyramid dehazing network (DCPDN) [18], which directly embeds the atmospheric scattering model into the network, allows for joint optimization of the transmission map, atmospheric light, and dehazing results, effectively capturing the intrinsic relationships between these three components. This kind of dehazing method, which incorporates some physical priors into deep networks by leveraging large-scale datasets and the powerful learning capabilities of neural networks, can obtain a certain degree of a haze removal effect. However, this method also has a limited generalization capability for specific scenes, and errors in the estimation of the three parameters may lead to cumulative errors in the final results. The encoder-decoder structure, which extracts and compresses image features through the encoder and then gradually upsamples them in the decoder to restore both the image resolution and illumination information, can effectively reduce the estimation error accumulation of the three unknown parameters. Hence, we draw inspiration from illumination-guided strategies with an encoder-decoder structure in image restoration and then design an irradiance guidance module to eliminate cumulative errors and improve the accuracy and effectiveness of dehazing networks.

2.2. Hybrid Attention Mechanism

Convolutional attention mechanisms have garnered significant attention in image processing tasks, particularly in the domain of image dehazing. By enabling networks to selectively focus on specific areas of an image, these mechanisms help improve image restoration quality. Early approaches primarily used spatial attention mechanisms, which applied attention maps at various network layers to enhance the capture of local details while minimizing irrelevant noise. Subsequent studies have proposed hybrid attention mechanisms [20] which combine spatial and channel attention, further refining detail processing performance. Hu [21] introduced channel attention, with recalibration of feature maps through learning to emphasize more informative channels, thus improving feature representation. This concept has also been applied to dehazing tasks, such as in FFA-Net [7], which employs sequential spatial and channel attention mechanisms to enhance dehazing by focusing on the most relevant features. Integrating these mechanisms into dehazing networks enables models to more accurately remove haze based on their spatial distribution while preserving image details better. Based on hybrid attention, the feature-level convolutional attention (FLCA) module proposed in this work further improves the overall dehazing performance.

3. Methods

Our dehazing network architecture, illustrated in Figure 1, comprises two stages: the restoration phase based on the atmospheric scattering model and the irradiance guidance phase. In the first stage, we design two subnetworks to estimate the medium transmission t and atmospheric light A, which are then used in the atmospheric scattering model to produce an initial clear image J. In the second stage, we design an irradiance-guided structure to recover high-fidelity results for the dehazed image.

3.1. Medium Transmission Estimation Network

Images captured by UAV cameras are usually affected by different densities of haze in one image. To effectively restore the image, the dehazing network needs to adapt to the non-uniform haze densities in images. Our approach achieves the goal of realizing non-uniform haze removal by designing fine feature-level attention to obtain the weights of each channel and the spatial attention weights within the channel, which in turn achieves non-uniform treatment of each channel and pixel during computation.

Existing CNN-based methods typically perform feature extraction at either full resolution or progressively lower resolutions. The former provides spatially accurate results but lacks robust contextual information, while the latter offers semantic reliability but with less spatial precision. In this paper, we introduce a novel architecture aimed at maintaining spatially precise high-resolution representations throughout the network while also integrating strong contextual information from low-resolution features. The core of this approach is the full-scale residual group (FSRG), which includes several key elements: (1) multi-scale features are extracted by a MRCB, and (2) the attention weights of each channel and the spatial attention weights within the channel are captured by a FLCA module.

Our medium transmission estimation network T-net, the structure of which is shown in Figure 2, consists of three FSRGs. The input for T-net is a hazy image $h\in R^{H\times W\times 3}$. Initially, T-net applies a convolutional layer to extract the underlying feature $X_0\in R^{H\times W\times C}$. Next, the feature map $X_0$ passes through three FSRGs to obtain the depth features $X_d\in R^{H\times W\times C}$. Each FSRG consists of multiple MRCBs, and inside the MRCB, we design three parallel branches with different sizes of convolutional kernels, each of which incorporates FLCA for attention-weight computation. In the MRCB, the multi-scale information flow complements the original feature stream by feeding full-resolution features into branches with different kernel sizes. This process can be described as follows:

$$\begin{array}{cc}\hfill & f_{1\times 1}=FLCA\left(Con{v}_{1\times 1}\right)\hfill \\ \hfill & f_{3\times 3}=Con{v}_{3\times 3}\left(Con{v}_{3\times 3}\left(FLCA\left(Con{v}_{3\times 3}\right)\right)\oplus Con{v}_{5\times 5}\left(FLCA\left(Con{v}_{5\times 5}\right)\right)\right)\hfill \\ \hfill & f_{5\times 5}=Con{v}_{5\times 5}\left(Con{v}_{5\times 5}\left(FLCA\left(Con{v}_{5\times 5}\right)\right)\oplus Con{v}_{3\times 3}\left(FLCA\left(Con{v}_{3\times 3}\right)\right)\right)\hfill \\ \hfill & f_{fuse}=Con{v}_{3\times 3}\left(f_{3\times 3}\otimes f_{5\times 5}\right)\hfill \\ \hfill & f_{\mathrm{agg}}=\left({\mathrm{Conv}}_3\left(f_{\mathrm{fuse}}\right)\right)\oplus \left({\mathrm{f}}_{1\times 1}\right))\hfill \end{array}$$

(2)

where the FLCA module, discussed in detail in Section 3.2, is used to enhance feature representation.

Next, we apply a convolutional layer to the deep features $X_d$ to obtain the residual image $t\in R^{H\times w\times 3}$. Finally, the transmission map is computed as follows:

$$\widehat{t}=t+I$$

(3)

3.2. Feature-Level Convolutional Attention

Transformers are typically computationally intensive and architecturally more complex than CNNs, making them unsuitable for deployment in UAVs and edge devices without further modification. We designed a convolution-based attention mechanism to replace the transformer’s self-attention mechanism and address the challenge of non-uniform haze removal. We retained the transformer’s architecture except for the self-attention part, where we implemented a feature-level convolutional attention mechanism to accurately represent the non-uniform distribution of haze. Specifically, we introduced dual branching residual attention (DBRA), which generates dedicated spatial importance maps for each input feature channel from coarse to fine while effectively mixing channel and spatial attention weights to ensure thorough information interaction. The detailed process of FLCA is shown in Figure 3.

In the DBRA structure, $r\in R^{H\times W\times C}$ is the input feature. The goal of DBRA is to produce channel-specific spatial importance maps (SIMs) with the same dimensions as X, denoted as $W_s\in R^{H\times W\times C}$. We first compute the corresponding channel attention map $W_c$ and spatial attention map $W_s$ using the following equations:

$$W_s=Con{v}_{7\times 7}\left(\left[X_{GAP}^s,X_{GMP}^s\right]\right)$$

(4)

$$W_c=Con{v}_{1\times 1}\left(max\left(0,Con{v}_{1\times 1}\left(X_{GAP}^c\right)\right)\right)$$

(5)

where $GAP$ denotes the global average pooling and $GMP$ denotes the global max pooling.

The sum of $W_c$ and $W_s$ produces a channel-specific spatial weight map W for each input feature channel. Initially, the input features and W are concatenated to yield a preliminary attention result. Instead of simply concatenating, we perform a channel alignment to seamlessly integrate different channels’ spatial importance maps with the input features. This involves adjusting the spatial importance map of each input channel to align it adjacently, followed by passing the aligned weight map through a convolution layer and a sigmoid function, resulting in a feature-level attention map.

3.3. Atmospheric Light Estimation Network

Accurate estimation of atmospheric light is crucial in image dehazing tasks, as it directly impacts the quality of the final dehazing result. However, traditional methods often suffer from errors in atmospheric light estimation, which can affect the precision and consistency of the dehazing outcome [22]. To address this issue, we have developed a network specifically for atmospheric light estimation, referred to as A-Net. Figure 4 illustrates its structural diagram.

A-Net is a lightweight network designed for atmospheric light estimation. Its feature extraction layer consists of a convolutional layer and FLCA in tandem, which capture the local features and texture information of the input image and provide a high-dimensional feature representation. Considering that atmospheric light is often located in regions distant from the observation point, and these areas are usually significantly affected by haze [7], A-Net incorporates FLCA modules to allow the network to treat different pixels differently, focusing more on regions affected by haze.

We used global maximum pooling at the end of A-Net to intuitively pool individual maximum intensities. Unlike previous encoder-decoder frameworks [18,23,24], this approach associates ambient light with high-intensity pixels, consistent with the high luminance properties of the ambient light. The extracted features are then passed through a fully connected network to estimate the RGB values of the atmospheric light A. Finally, these RGB values are expanded to match the dimensions of the input image, allowing them to participate in the dehazing process effectively.

3.4. Irradiance Guidance Module

Deep learning-based dehazing methods which use the atmospheric scattering model typically involve two separate steps to accurately simulate the recovery process: estimating the medium transmission and estimating the atmospheric light. Estimating the medium transmission relies on the local brightness information of the image, while estimating the atmospheric light is based on the global brightness distribution. Since these steps are performed independently, errors in the estimates of atmospheric light and medium transmission can accumulate and be magnified, impacting the quality of the dehazing results and making it challenging to restore image details and contrast. To address this issue, we propose an irradiance-guided method based on an encoder-decoder architecture. This approach aims to minimize the impact of errors and enhance the recovery of irradiance.

Our encoder-decoder architecture aims to minimize error accumulation and restore high-fidelity color through progressive feature-level irradiance guidance. As shown in Figure 5, the encoder consists of four convolutional layers and two downsampling operations, culminating in a bottleneck structure. Each convolutional layer applies an ReLU activation function for nonlinear transformation to extract local features from the image. Downsampling reduces the spatial resolution of feature maps to lower the computational complexity while preserving global information. The bottleneck structure comprises four residual blocks, each with two 3 × 3 convolutional layers and skip connections to address the vanishing gradient problem [25,26].

The decoder mirrors the encoder, progressively upsampling the feature maps to restore them to the original image resolution. This upsampling process increases the spatial resolution of the feature maps to recover image details and integrates high-level features extracted by the encoder to restore the original image’s resolution and detail. The irradiance guidance module adjusts the irradiance information at the feature level to mitigate the impact of errors and enhance the color restoration process. This module aims to improve the quality of dehazing results by progressively recovering the irradiance information in the feature maps.

4. Results

4.1. Experimental Set-Up

4.1.1. Implementation Details

Our network was implemented using PyTorch (v1.10.1)and trained on an NVIDIA (Santa Clara, CA, USA) RTX 3090 GPU. We employed the ADAM optimizer with an initial learning rate set to s $5\times {10}^{-4}$. The batch size was configured to be 4, and the training was conducted over 500 epochs. In addition, the UAV used for data collection was the P200 basic frame from Amu Lab, and the onboard camera was a G1 fixed-focus three-axis pod camera. The UAV is shown in Figure 6. The inference speed for a single image was 55 ms on the NVIDIA RTX 3090 GPU, while on the onboard computer (NVIDIA Orin NX), the inference speed was about 100 ms per image.

4.1.2. Experiment Settings

Datasets: We used the RSID dataset, the NID dataset [27], and a self-built dataset with 500 clear images captured by a UAV’s onboard camera and the corresponding hazy images synthesized by the atmospheric scattering model, which covered the image content of different scenes. The ratio of the training, validator, and test sets is 8:1:1. During the training process, the training set images were cropped to 128 × 128 pixels to increase the diversity of the data. For ease of representation, we refer to the test set in the RSID as the R100 dataset and the test set in the NID as the N100 dataset. In addition, we evaluated our method on the UAV dataset [28], which consists of 150 real-world haze images captured by UAV cameras.

Comparison Methods: We compared Ours-Net with several state-of-the-art methods, including supervised methods such as DehazeNet, AOD, FFA, Dehamer, and MB-Taylor, as well as unsupervised methods like DAD and PSD.

Evaluation Metrics: For the synthetic R100 and N100 datasets, we used PSNR and SSIM. Higher PSNR and SSIM values indicate that the dehazed images are closer to the ground truth. For the real-world UAV dataset [28], we conducted no-reference evaluation using MUSIQ [29] and the DBCNN [30]. Higher DBCNN scores indicate better fidelity, while higher MUSIQ scores reflect a higher resolution and more detailed image presentation.

4.2. Visual Comparisons

For visual comparison, we first present the results on the synthetic R100 dataset in Figure 7. PSD [31] achieved relatively good contrast recovery by adjusting the color in the fine-tuning stage, but the experimental results show that it could not completely eliminate the uneven haze. Other comparison methods either failed to remove non-uniform haze or introduced undesirable color distortions, suggesting that non-uniform haze removal from the UAV viewpoint is a problem to be solved. In contrast, Ours-Net produced results which most closely matched the ground truth images in terms of detail and color. Although Dehamer [11] and MB-Taylor [12] use transformer-based deep model architectures, their dehazing performance was inferior to Ours-Net, highlighting the importance of considering the spatial distribution of haze, which is effectively addressed by our feature-level convolutional attention design.

The results on the synthetic dense haze N100 dataset are presented in Figure 8. While DAD [32] effectively removes haze, the GAN may exhibit color reproduction biases, resulting in slight color discrepancies between the haze removal results and the ground truth images. PSD [31], though fine-tuning for color and brightness, might not comprehensively optimize the haze removal mechanism, leading to noticeable residual haze. Methods such as FFA [7], Dehamer [11], and MB-Taylor [12] achieved relatively good dehazing results but fell short in restoring local texture details. In contrast, our approach integrates domain knowledge into the network structure and introduces an irradiance guidance module to address error accumulation from stepwise parameter estimation. The dehazing results show that our method not only effectively removed dense haze but also restored colors more accurately.

Figure 9 illustrates the test results on the real-world UAV dataset [28]. The DehazeNet method [15], which learns the transmission, achieved a notable dehazing effect but tended to produce results which were too dark. Similarly, AOD [17], which also learns the parameters of the atmospheric scattering model, demonstrated effective dehazing on real-world haze images captured by UAVs. However, AOD did not match the detailed processing of our method, as evidenced by the metrics in

Table 1. Comparison of methods across different datasets. Red highlights indicate the best results, while purple highlights denote the second-best results. The values in the table represent the average metrics across all datasets.

Methods	Publication	R100		N100		UAV
		PSNR ↑	SSIM ↑	PSNR ↑	SSIM ↑	MUSIQ ↑	DBCNN ↑
Dehaze	TIP’16	15.6341	0.4273	11.1996	0.5503	29.6630	0.2733
AOD	ICCV’17	16.0922	0.5208	10.7617	0.5208	31.5178	0.2806
FFA	AAAI’20	18.7787	0.8628	18.5184	0.8744	25.6996	0.2808
DAD	CVPR’20	16.7190	0.7573	18.9910	0.8618	28.8767	0.2875
PSD	CVPR’21	11.1533	0.6634	11.7369	0.7755	29.7870	0.2961
Dehamer	CVPR’22	18.8173	0.8542	22.2579	0.9186	25.7494	0.2874
MB-Taylor	ICCV’23	17.2736	0.8268	22.8777	0.9300	25.0852	0.2813
Ours	–	25.7046	0.9487	31.6056	0.9881	31.8629	0.3302

. The haze residue remained evident in the results of the end-to-end image mapping dehazing methods based on learning [7,11,12]. Overall, the experiments with real-world data show that physical model-driven methods provide superior dehazing performance in actual haze environments.

4.3. Quantitative Comparisons

Table 1 reports the quantitative results for both the synthetic and real-world datasets. In terms of full-reference metrics, Ours-Net outperformed the competing models, achieving the highest PSNR and SSIM scores on the R100 and N100 datasets. This indicates that our designed dehazing network for UAV images provided results closest to the ground truth, demonstrating strong dehazing capabilities. For no-reference metrics, Ours-Net also achieved the highest scores in MUSIQ and the DBCNN on the UAV dataset compared with other models. This suggests that our method delivers high-fidelity, high-resolution dehazing results with excellent image detail under real-world haze conditions. These results confirm that our approach effectively enhances the visibility of UAV images in hazy environments.

4.4. Evaluation in Other Scenarios

To further validate the applicability of the proposed method beyond UAV imagery, we conducted experiments across four distinct scenarios: satellite imagery, underwater imagery, outdoor natural scenes, and indoor scenes. For each scenario, we utilized well-established public datasets, and the model achieved satisfactory results in all restoration tasks, demonstrating its robustness and broad applicability. The restoration results are presented in Figure 10, where (a) corresponds to satellite imagery, (b) corresponds to underwater imagery, (c) corresponds to outdoor natural scenes, and (d) corresponds to indoor scenes.

Satellite Imagery: Using the Haze1k [33] dataset, the model successfully restored details in high-resolution satellite images, particularly in handling complex textures and terrain variations, proving its excellent performance on high-complexity images.

Underwater Imagery: With the UIEB dataset [34], the model performed well under the challenging conditions of varying light and water turbidity. It effectively removed noise and blur from underwater images, producing clear and structurally accurate reconstructions.

Outdoor Natural Scenes: Using the HSTS dataset [35], the model exhibited outstanding performance in outdoor environments with diverse lighting and structural variations, achieving significant results in restoring edge details and complex textures.

Indoor Scenes: With the I-Haze dataset [36], the model performed high-fidelity restoration of indoor images affected by artificial haze, recovering the integrity of objects and spatial relationships and especially excelling in processing light and shadow details.

These results, illustrated in Figure 10, further confirm the model’s applicability across different image types and its versatility across multiple scenarios. This demonstrates its broad potential for practical applications in various imaging contexts.

4.5. Computational Efficiency

In this section, we compare the computational efficiency of our proposed method with five mainstream methods by analyzing their computational costs and parameter sizes.

Table 2. Model Parameter and computational cost comparison.

Methods	Params (M)	FLOPs (GMac)
FFA	4.46	287.80
PSD	33.11	182.50
DAD	54.59	789.11
Dehamer	132.50	60.30
MB-Taylor	7.43	88.10
Ours	20.139	30.783

presents a detailed comparison of the computational costs and parameter sizes for each model. Our model performed the best in terms of computational cost, achieving 30.783 GMac, while the parameter size was 20.139 M. Although it ranked third in parameter size, it had the best dehazing effect among all of the models. This makes our model particularly suitable for real-time applications and scenarios with limited computational resources.

Furthermore, the lower computational cost resulted in reduced energy consumption and faster inference times, which are crucial for large-scale deployments. In resource-constrained environments, our method not only effectively reduced the demand for computational resources but also maintained excellent performance and scalability. Overall, despite ranking third in parameter size, our model, with its optimal computational cost and outstanding dehazing effect, demonstrated high practicality for real-world applications, achieving an ideal balance between accuracy and efficiency.

4.6. Ablation Experiment

To evaluate the effectiveness of our proposed full model with the FLCA module, we conducted ablation experiments comparing the performance of three different models: the version without the FLCA module, the version using Transformer self-attention, and the full model including FLCA. The experiment consisted of two parts: a visual comparison and a quantitative comparison.

Visual Comparison: We show the performance of the different models in the dehazing task. Figure 11 shows the results of the three models processing the same haze image. It can be clearly seen that the version without the FLCA module had some defects in recovering image details and its dehazing ability, while the version with Transformer self-attention improved the dehazing ability but still failed to achieve the best result. Our full model performed optimally in terms of dehazing, with significant improvements in both dehazing and image contrast, showing the most natural and clear results.

Quantitative Comparison:

Table 3. Quantitative comparison of ablation experiments on R100. The values in the table represent the average metrics across all datasets.

	PSNR	SSIM
Without FLCA	18.6132	0.7090
With Transformer	22.6664	0.9125
Ours	25.7046	0.9487

lists the quantitative comparison of three models on evaluation metrics, including the PSNR and SSIM. The results show that the full model with the FLCA module outperformed the other two models in all metrics, with significant improvements in the PSNR and SSIM, demonstrating enhanced dehazing performance.

These results indicate that the introduction of the FLCA module significantly improved the dehazing effect, showing the best performance both visually and in terms of quantitative metrics. This validates the effectiveness of the FLCA module in improving the quality of dehazing.

Impact of Model Parameters: To enrich our analysis, we also evaluated the impact of different parameter configurations on the model’s performance. Specifically, we tested variations in the number of layers in the FLCA module and the use of different network types. The results show that increasing the number of layers up to an optimal threshold enhanced the model’s ability to capture fine-grained details, while excessively increasing the layers led to diminishing returns due to overfitting. Similarly, using different network types (such as ResNet variants) demonstrated that our proposed structure with convolutional attention was more effective for dehazing tasks compared with standard architectures, offering a better balance between computational efficiency and performance. These findings underline the importance of optimizing the model’s architecture for achieving the best results.

In the proposed DBRA structure, the final step is to concatenate the features from the two branches and perform channel alignment. This design aims to assign different spatial attention weight maps to different channels, allowing for pixel weight assignments within each channel. To validate the effectiveness of this design, this study conducted ablation experiments to investigate the model’s performance without channel alignment. This approach further confirms the role of channel alignment in enhancing model performance, as illustrated in Figure 12, which presents qualitative comparison results.

In the ablation experiment results under non-uniform haze scenarios, the dehazing results without channel alignment exhibited slight haze residuals in heavily hazy areas. In contrast, with the implemented channel alignment, the system achieved complete removal of non-uniform haze. This significant improvement underscores the importance of the model architecture, particularly the effectiveness of channel alignment, in accurately addressing varying haze densities across different regions, leading to clearer and more visually appealing outcomes.

5. Disscussion

Although our proposed dehazing method demonstrated its effectiveness across various experiments and outperformed existing techniques in complex hazy conditions, there are still some limitations which warrant further exploration. The method’s performance may degrade in extreme haze conditions, such as in cases of highly dense or highly non-uniform haze. In particular, when the haze concentration is high and unevenly distributed, the model may struggle to accurately estimate global illumination and transmission maps. Future research could focus on optimizing computational efficiency and improving the model’s performance in extreme haze scenarios while also exploring its generalization capabilities across different environments. This would enhance its applicability and robustness in real-world settings.

6. Conclusions

In this study, we proposed a novel dehazing method for drone imagery which integrates a physical atmospheric scattering model with deep learning techniques. Our experimental results demonstrate that combining physical models with deep networks effectively addresses issues such as color distortion and reduced contrast encountered in traditional dehazing methods. Specifically, we introduced a joint optimization strategy based on the atmospheric scattering model which not only relies on data-driven approaches but also incorporates the physical characteristics of atmospheric scattering, thereby enhancing the accuracy of dehazing and improving image quality. Additionally, by designing an irradiance-guided coding and decoding structure, we effectively reduced the accumulation of errors in parameter estimation, further enhancing the clarity of dehazed images. Evaluation results on multiple datasets show that our method outperformed existing dehazing technologies, validating its effectiveness under complex hazy conditions. In summary, our approach, by integrating theoretical foundations with practical applications, significantly improved the precision and reliability of image dehazing, demonstrating its strong capability in handling images with haze in complex environments.