Infra Aerial Dehaze: Dehazing of Infrared Aerial Images Using Cycle Generative Adversarial Network

Abstract

Hazy environments create substantial difficulties in infrared (IR) aerial imaging by reducing the quality of image making it difficult for human eye to see the object features clearly. This work proposes a robust CycleGAN based haze removal model, called Infra Aerial Dehaze, specifically for infrared aerial images with good generalization to various haze densities. The proposed model is implemented using the encoder-decoder-transformer based dehazing architecture and its performance is tested using two datasets HIT-UAV and Transpetro-Train containing various scenarios and hazy conditions. Moreover, a new dataset of 3604 aerial infrared hazy images is synthetically produced by an atmospheric scattering model to further validate the proposed method. The performance of the proposed dehazing model is assessed using qualitative and quantitative measures, with PSNR and SSIM metrics used for quantitative assessment. Additionally, CycleGAN was not previously used for infrared image dehazing, therefore we present first CycleGAN-based generative model, Infra Aerial Dehaze which can remove haze from infrared aerial images with varying haze intensities. Experimental results confirm that the proposed model is capable of restoring details of the texture and enhancing the visibility of the image. The proposed model can be applied effectively for dense haze removal and demonstrate superior results to other prior based dehazing techniques.

Keywords

generative adversarial network CycleGAN IR image dehazing atmospheric scattering model dark channel prior relevant haze feature

1. Introduction

Infrared imagery spans a wide range of electromagnetic spectrum and has unique significance and applications. Infrared images exhibit unique characteristics like signal attenuation, reduced sharpness and image clarity during hazy condition. The presence of haze increases problems in interpretation of IR imagery because of scattering of infrared radiation by dust particles. This leads to difficulties in contrast, resolution, temperature determination and object recognition in infrared images (Erlenbusch et al., 2023). Consequently object detection, tracking and segmentation are affected by the presence of atmospheric haze (Erlenbusch et al., 2023; Li et al., 2018).

Image dehazing is primarily used to restore image visual quality by eliminating the effects that arise from atmosphere haze or fog. Most of the dehazing algorithms are based on the estimation of the atmospheric haze component to get the better and more clear images in terms of visibility. Several approaches have been presented in the past literature to address the challenges of removing haze in images. The dark channel prior (DCP) is one of commonly used technique for dehazing the hazy images (Hua et al., 2021; Nawaz et al., 2021; Parihar et al., 2020; Qasim & Raja, 2022; Yan et al., 2022). DCP is based on the observation that in natural images, the pixels with the lowest intensity value, often referred to as the “dark channel,” tend to contain minimal atmospheric haze. This prior information is leveraged to estimate and remove the haze from the entire image. Figure 1 shows the result of dehazed images using dark channel prior encompassing a mountainous region veiled by heavy clouds (Nawaz et al., 2021). Prior-based techniques require manual tuning and are therefore less adaptive in handling diverse hazy conditions. Consequently, recent haze removal methods rely on learning-based approaches. Learning-based haze removal techniques train a CNN architecture to acquire haze-free transmission maps from input image (Cai et al., 2016; Hovhannisyan et al., 2022; Li et al., 2017). Learning-based approaches for image dehazing increasingly use Generative Adversarial Networks (GANs) (Chaitanya & Mukherjee, 2021; Engin et al., 2018; Fu et al., 2021; Kan et al., 2022). GANs is the type of Deep Neural Networks that are employed for generating artificial data particularly images from noises or from exiting images. GAN-based techniques usually use a hazy image and ground truth haze free images (paired data) during the training process. These networks effectively learn to produce haze-free images even when there are significant variations between the two domains because they are trained to transfer the hazy input domain to the clear output domain (Chaitanya & Mukherjee, 2021; Engin et al., 2018; Fu et al., 2021; Kan et al., 2022; Wang et al., 2022). However, paired data is not required while utilizing a CycleGAN network (Chaitanya & Mukherjee, 2021; Engin et al., 2018; Wang et al., 2022).

Figure 1.

Dehazed Images Using Dark Channel Prior (DCP).

CycleGAN is useful for tasks like dehazing, style transfer, and other image modification applications because it can learn mappings between two distinct picture domains (for example, hazy and haze-free images) without the need for paired datasets (Chaitanya & Mukherjee, 2021; Engin et al., 2018; Kan et al., 2022). The cycle-consistency loss in CycleGAN guarantees that the dehazed image can be returned to its original hazy state, reducing artifacts and enhancing image realism. Furthermore, CycleGAN is very adaptive for real-world dehazing applications since it preserves the image structure while generalizing well across different degrees of haze (Engin et al., 2018; Wang et al., 2022). Considering these characteristics, we proposed end-to-end Infra Aerial Dehaze network for infrared aerial image dehazing. Previous methods used for dehazing of infrared (IR) images are mostly prior based without using deep learning methods like CycleGAN. Prior-based methods do not perform well in handling complicated haze distribution. Our proposed model not only eliminates the haze but also preserves the critical detail in the image.

Our paper makes the following significant contributions:

The unavailability of infrared hazy datasets in many real-world applications forces the researchers often to rely on visible spectrum images, for image guidance and evaluation of the real time infrared imaging system. Due to the lack of infrared hazy aerial dataset, first varying degrees of haze is introduced in infrared aerial images using atmospheric scattering model. A dataset of 3,604 artificially created infrared hazy images is obtained. This dataset overcomes difficulty related to lack of availability of labelled hazy infrared images.

The generative models like CycleGAN have not yet been explored for infrared image dehazing, primarily due to the domain gap between visible and infrared spectrum. To address this, an end-to-end lightweight, and effective Infra Aerial Dehaze, architecture for dehazing in infrared aerial images is proposed. The proposed framework allows to reduce the domain gap between visible and infrared imagery and allows an efficient dehazing based on the spectral and textural properties of the infrared aerial imagery.

Results of the proposed model are compared with those obtained from existing approaches in image dehazing problems. The proposed model is first tested on the HIT-UAV and Transpetro-Train infrared aerial datasets which contains broad sets of scenes and with different levels of haze. The quantitative and qualitative comparisons is conducted with state-of-the-art approaches on both reference-based (PSNR, SSIM) and no-reference (NIQE, PIQE) image quality metrics. Experimental has confirm that our approach gives better dehazing results and strong generalization to different haze intensities and real-world hazy situations. The performance of model is also evaluated on night infrared hazy images. Although the model was not trained on night hazy images yet it gives outstanding results depicting the generalizability of the proposed network.

In the next section, we discuss the studies that have been done to tackle the problem of image dehazing.

2. Related Work

Addressing single image dehazing is a complex challenge with an extensive research background (Erlenbusch et al., 2023; Li et al., 2018; Kan et al., 2022; Chaitanya & Mukherjee, 2021; Engin et al., 2018). Traditional methods adopted for dehazing single images involve hand crafted features such as the dark channel prior, color attenuation prior and non-local prior that shows simplicity and effectiveness across different scenes (Nawaz et al., 2021; Hua et al., 2021; Yan et al., 2022; Zhu et al., 2015; Gao et al., 2018). Recent methods use multi-spectral priors, which are the combination of spectral and learnt features, to improve the performance over different fog densities and wavelength bands (Qasim & Raja, 2022; Wang et al., Nov. 2021). N Wang et al. presents Prior-Guided Multiscale Network (PGMNet) integrates prior-based dehazed images into a multiscale CNN framework to enhance real-scene image dehazing. PGMNet outperforms the state of the art when trained on both synthetic and real-world datasets, by integrating the advantages of both traditional priors and deep learning, as well as attention and feature aggregation modules that minimize artifacts (Wang et al., Nov. 2021). M Qasim et al. SPIDE-Net proposes a spectral prior-based network image dehazing and enhancement network integration of spectral and multi-scale prior information for better performance across varying haze densities and wavelength bands. By Using the SID-Net and MPD-Net sharing the generative decoder, it achieved better results both in a qualitative and quantitative evaluation of SHIA dataset (Qasim & Raja, 2022).

Over the past few years, there has been a notable transition towards the utilization of learning-based methods in the field of single image dehazing problem. Comparing with other approaches, the learning-based methods present better flexibility and higher accuracy especially for dealing with various hazy scenes (Cai et al., 2016; Li et al., 2017; Kan et al., 2022; Chaitanya & Mukherjee, 2021; Engin et al., 2018; Wang et al., 2022). Hovhannisyan et al. (Hovhannisyan et al., 2022) introduce AED-Net an end-to-end adaptive enhancement dehazing network for single image dehazing. Compared to other works, based on PSNR, SSIM and other indicators, AED-Net is superior and improves the contrast, color, texture and edge details of the image. Boyi Li et al. (Li et al., 2017) proposes the All-in-One Dehazing Network (AOD-Net), this is an image dehazing model that is built using a convolutional neural network (CNN). AOD-Net can directly generate the clean image using lightweight CNN network. In terms of PSNR, SSIM and subjective quality test on both synthesized and real hazy image the AOD-Net has better performance than the state of the art. W Ren et al. (Ren et al., 2020) proposed a multi-scale CNN-dehazing model that learns the hazy image and transmission maps using coarse- and fine-scale networks. Moreover, a holistic edge-guided refinement is presented to boost the edges of the transmission map. The trained on synthetic data derived from NYU Depth, the method performs very well in both quality and efficiency on synthetic and real-world images. Y Liu et al. (Liu et al., 2025) presented a variational nighttime dehazing framework, VNDHR, which tackles the specific noise, weak illumination and glows of the nighttime hazy images. The method employs a hybrid regularization model to construct illumination that is structure-aware and noise-free reflectance with a combination of of ℓp norm, ℓ2 norm and total variation. VNDHR gives state of the art performance on a range of degraded conditions such as night, low-light, and underwater images.

The SCL-Dehaze (Cui et al., 2024) proposes a semi-supervised learning of the codebook to learn about the real-world dehazing process. It tackles latent feature matching blunders by means of a UDART module, and enhances fidelity by using a haze-density governed adjustable module. The method exhibits good generalization and highest hazy performance on real-world datasets. Y Huang et al. (Huang et al., 2025) introduces DCD-Net a weakly supervised decomposition-based method to dehaze real-world images, which bridges the domain gap between synthetic and real haze. It does not use paired data because it models the process of hazing by interactions between clean images and illumination that takes the form of haze. The model trained on the URD dataset and it has better generalization and realism on real-world tasks.

Xuejie Cao et al. (Cao et al., 2020) have proposed a single-image dehazing algorithm that employed GAN with Feature Pyramid Network (FPN). This is an end-to-end image dehazing approach and avoids dependence on a particular physical model. The similar approach was used in (Chaitanya & Mukherjee, 2021; Suárez et al., 2018; Long et al., 2020) an end-to-end network for single-image dehazing. Anil Singh et al. (Parihar et al., 2022) Anil proposed learning-based approach for producing dehazed images from the hazy input images. Their method uses an end-to-end encoder-decoder based Generative Adversarial Network (GAN) with high connection density for improved features extraction from images and optimize their utilization in the process of image dehazing. This technique is useful in the elimination of haze in images and provides a perfect solution for enhancing the visibility of image with high quality. In (Fu et al., 2021) the authors highlighted a new approach to handle non-homogeneous dehazing problem by employing a two-branch end to end learnable Generative Adversarial (GAN) architecture. Their method introduces a new idea which is the integration of 2D discrete wavelet transform into the network and the primary objective of this is to retain the important high frequency information and reconstruct lost textural details of the dehazed images. The presented work not only explains how the proposed solution is innovative but also provide how it is effective based on careful design consideration.

Yu Dong et al. (Dong et al., 2020) presented an extensive Generative Adversarial Network (GAN) for image dehazing known as Fusion Discriminator GAN (FD-GAN). This unique approach includes a Fusion Discriminator which uses frequency information as additional prior information. These priors are beneficial in improving the dehazing model that produces images which have not only more realistic naturalness but are also less distorted in terms of color and are less likely to have artifacts. Yongzhen Wang et al. (Wang et al., 2022) has proposed new technique called Cycle Spectral Normalized Soft Likelihood Estimation Patch Generative Adversial Network (Cycle-SNSPGAN) for image dehazing. This method is designed an unsupervised manner and it mainly focused on the generalizability to the real-world hazy images.

In (Mehta et al., 2020) the authors present a deep learning model called hyperspectral guided generative adversarial network (HIDEGAN) for image dehazing. HIDEGAN's architecture is carefully constructed by combining two essential elements: a conditional generative adversarial network H2RGAN and a second one is a CycleGan known as R2HCYCLE. Notably this work is the first effort to integrate Hyperspectral Imaging (HSI) in the context of GAN to address the haze removal task specifically. In (Mehta et al., 2020) HIDEGAN achieves a significant improvement over existing state-of-the-art approaches in terms of D-Hazy, HazeRD, RESIDE-Standard (SOTS), and RESIDE- β (HSTS) datasets. The performance evaluation highlights HIDEGAN as a highly effective solution for image dehazing, surpassing existing methods in terms of overall efficacy.

In (Gan et al., 2020) an innovative multilevel image dehazing algorithm is introduced, leveraging Conditional Generative Adversarial Networks (CGAN). The proposed methodology involves utilizing a hazy image to jointly estimate the composed image K incorporating a transmission map and atmospheric light value through a generator network. Subsequently, a dehazed image is computed using an enhanced atmospheric scattering model. The adversarial training and reconstruction constraints are applied to both the generator network and the joint discriminator network. The experimental results validate the effectiveness of the proposed method, showcasing commendable dehazing outcomes in both synthetic hazy images and real-world hazy images. Guodong Fan et al. (Fan et al., 2022) present a novel approach in their work, introducing a multiscale cross-connected dehazing network with scene depth fusion. The primary emphasis is placed on understanding the correlation between a hazy image and its corresponding depth image. This design enables the direct generation of a clean image in an end-to-end fashion. The work described in (Sim et al., 2018) offers a simple but effective network designed specifically for high-resolution image dehazing using a Conditional Generative Adversarial Network referred to as DHGAN. The methodology involves the use of hazy patches obtained from scaled down training input images as inputs to the generator network of DHGAN. In (Sim et al., 2018) by constructing these hazy patches the DHGAN is capable of learning more fine grained haziness to facilitate an enhanced dehazing capacity and performance.

Dwij Mehta et al. (Mehta et al., 2022) presented a deep learning model called AerialGAN to reduce the effects of haze on aerial images. The main aim is to improve the quality of ill-defined images obtained from UAVs and come up with better images with less or no haze effects. The introduced dehazing technique has potential to contribute to the improvement of surveillance systems, especially in case when UAVs are used for protection of the critical infrastructures. Similar approach is presented in (Mehta et al., 2021) where a deep learning-based model named SkyGAN is introduced for haze removal in aerial images. SkyGAN consist of two essential components: a domain-aware hazy-to-hyperspectral (H2H) module, and a conditional Generative Adversarial Network (cGAN) multi-cue image-to-image translation module (I2I) for dehazing purpose. This is a dual-module architecture in SkyGAN showing a holistic approach to tackling the problem of haze in aerial scenes from both domains.

Finally, it can be noted that the latest methods in image dehazing are based on deep learning specifically generative adversarial networks. These methods perform well in visible image dehazing but are not utilized for infrared image dehazing, therefore the current work proposes Infra Aerial Dehaze network for haze removal in infrared aerial images. Next, we provide the detailed description of the proposed method.

3. Proposed Method

3.1 Hazy Image Formation Model

Due to the unavailability of infrared hazy image datasets, we first simulate synthetic haze in infrared images based on the atmospheric scattering model (Li et al., 2018; Nawaz et al., 2021; Yan et al., 2022; Li et al., 2017; Kan et al., 2022; Chaitanya & Mukherjee, 2021; Engin et al., 2018; Fu et al., 2021; Zhu et al., 2015; Parihar et al., 2022; Sarker et al., 2019). This technique allows us to create a controlled set of hazy images (varying level of haze) which can be applied to infrared imaging research and algorithm development. Image dehazing typically involves the use of an image degradation model known as the atmospheric scattering or koschmieder model. A lot of earlier dehazing techniques rely on the traditional atmospheric scattering model which can be stated mathematically as

\begin{aligned} I (x) & = J (x) t (x) + A (1 - t (x)) \end{aligned}

(1)

\begin{aligned} t (x) & = e^{- β d (x)} \end{aligned}

(2)

The first term in the right-side J(x)t(x) of the preceding equation (1) is referred to as direct attenuation, while the second term in the right-side A(1-t(x)) is referred to as air light. I(x) represent the hazy image and J(x) represent the scene radiance, which is the “clear image” that we want to recover. ‘A’ stands for the global atmospheric light, and t(x) for the transmission map. The pixel-by-pixel depth map is represented by d(x), while the medium's scattering coefficient is denoted by $β$ . We produce a dataset of 3,604 artificially created infrared hazy photos with different degrees of haze strength. The source original clear infrared images datasets are publicly available in Kaggle (HIT-UAV: A High-altitude Infrared Thermal Dataset, 2025) and roboflow (Transpetro_train Dataset > Overview, 2025) respectively. The atmospheric scattering model with different values of the parameter of β represents the various haze densities. We will apply various densities of haze to high-resolution clear infrared photos using a controlled atmospheric scattering model. We particularly used β = 0.20 for light haze, 0.42 for medium, 0.60 for high, 0.79 for very high and 0.90 for dense haze. Figure 2 represents the generated hazy images with various degree of haze. Next, we describe our proposed dehazing method.

Figure 2.

Generated Dataset of Hazy Images with Various Degree of Haze.

3.2 Proposed CycleGAN Dehazing Architecture

CycleGAN is ideal for dehazing tasks that involve with a focus on image quality compared to other methods because of flexibility, consistency preservation and unpaired image translation. It effectively gathers haze characteristics and reproduces perceptible images demonstrated by encoder-decoder blocks and it also gathers important tiny details necessary for the visibility of objects. Also, the CycleGAN has not been previously applied for dehazing of infrared (IR) images, for this reason, we proposed a novel method that utilize CycleGAN for the haze removal of infrared aerial images. The proposed dehazing architecture follows the original CycleGAN and is trained on infrared aerial images. It incorporates residual-transformer block within the generator, that is specifically for deep features extraction (Zhu et al., 2017).

The proposed Infra Aerial Dehaze architecture includes two generators, A and B and two discriminators $D_{A}$ and $D_{B}$ . Generator A generates images from domain A, similarly the generator B generates images from Domain B. In the dehazing architecture, each generator contains three primary components, an encoder block, a transformer block, and a decoder block. The generator architecture utilized in our suggested dehazing model is shown in Figure 3.

Figure 3.

Generator Architecture of Proposed Dehazing Model.

The dehazing process begins start with the hazy image has an input from Domain A which is firstly passing through the encoder block. This block contains convolutional layers to extract and then compress useful features or characteristics of hazy image. The encoded representation is then passed through to the transformer block, which includes six residual blocks with skip connections that learn and transform the features into target domain B, and removes relevant components of haze. Each residual block starts with 1 × 1 convolution that decreases the channels from 256 to 128 followed by batch normalization and ReLU activation function. The output of every block is directly added to its input through a skip connection that allow residual learning and stable flows of the gradient during training. The residual transformer block is a bottleneck between encoder-decoder and uses several grouped convolutional layers inside of residual connections, and it is based on convolutional operations instead of the self-attention mechanism. The residual transformation block improves the mid-level feature representation by learning complex spatial patterns with grouped of convolutional layers and support efficient feature transformation. Finally, a Decoder is used to decode that transformed embedding into the image from corresponding domain B. The overall architecture of the proposed dehazing approach is illustrated in Figure 4. The discriminator in our proposed method of dehazing is a general discriminator structure similar to the Patch GAN discriminator (Wang et al., 2022; Zhu et al., 2017) that takes an image of shape (H, W, C) and determines whether it is fake or real by passing it through the Convolution Layers as shown in Figure 5. Finally, the classification is being made by Sigmoid Layer.

Figure 4.

Overall Architecture of Proposed Dehazing Model.

Figure 5.

Discriminator Architecture of Proposed Dehazing Model.

The discriminator in the proposed architecture takes an input image of size 256 × 256 × 3 and passes through six layers. The first three layers use a stride of 2, progressively reducing the spatial dimensions by half up to the fourth layer, following the sequence: 256→128→64→32. Every layer of the discriminator is a convolutional layer followed by batch normalization to enhance the model and to enable input adjustment which is then followed by LeakyReLU to introduce non-linearity. The first layer does not use normalization since it directly gets the input. The final layer is a binary classifier with a Sigmoid function that classify images weather fake or real. This layer takes a 32 × 32 × 2 feature map and produces an (m, 1) vector, where each vector in the batch size of m samples is either 0 or 1 indicating real or fake images. Next, we discuss loss function of proposed dehazing model.

3.3 Loss Function

The objective or loss function for the proposed Infra Aerial Dehaze architecture is based on the loss function presented in Cycle Dehaze (Engin et al., 2018) the combination of both adversarial loss and cycle consistency loss. The least squares loss, or adversarial loss (A, Dy, X) of generator A is represented by equation (3).

\begin{aligned} AL (A, D_{y,} X) = \frac{1}{m} \sum_{i = 0}^{m} [1 - D_{y} (A (x_{i}))]^{2} \end{aligned}

(3)

where

A (x_{i})

is an image generated by generator A and

D_{y}

distinguishes between the actual ground truth image y from the generated image

A (x_{i})

and

x_{i}

is an image from a domain X. Likewise, the adversarial loss for generator B is defined by the (B, Dx, Y) represented as follows in equation (4).

\begin{aligned} AL (B, D_{x,} Y) = \frac{1}{m} \sum_{i = 0}^{m} [1 - D_{x} (B (y_{i}))]^{2} \end{aligned}

(4)

where

B (y_{i})

is an image generated by generator B and

D_{x}

distinguishes between the actual ground truth image x from the generated image

B (Y_{i})

and

Y_{i}

is an image from a domain Y. Consequently, the complete adversarial loss (AL) is expressed as follows in equation (5).

\begin{aligned} A L = A L (A, D y, X) + A L (B, D, x, Y) \end{aligned}

(5)

Cyclic consistency loss guarantees that, for an image to be transformed to another domain and then converted back to the original domain it should retain the same similarity necessary to retain image quality. This loss can be represented by equation (6) with batch size m, where information obtained from Domain X is translated into Domain Y by function A and the same information is returned to Domain X by using inverse function B.

\begin{aligned} L_{c y c} (A, B, X, Y) = \frac{1}{m} \sum_{i = 0}^{m} ([B (A (x_{i})) - x_{i}]) + ([A (B (y_{i})) - y_{i}]) \end{aligned}

(6)

The final loss function for the proposed Infra Aerial Dehaze, is the combination of both the losses, i.e., the adversarial Loss and the cyclic consistency loss represented as follows in equation (7).

\begin{aligned} L_{f i n a l} = A L + λ L_{c y c} (A, B, X, Y) = \frac{1}{m} \sum_{i = 0}^{m} ([B (A (x_{i})) - x_{i}]) + ([A (B (y_{i})) - y_{i}]) \end{aligned}

(7)

The cyclic consistency loss is controlled by the hyperparameter λ, this particular parameter helps in ensuring that the generated images appear natural and realistic to human perception at the same time the structure and content of the input images is retained (cycle consistency loss). Next, we provide implementation details for the suggested approach.

3.4 Implementation Details

The implementation of the proposed model was done using PyTorch as the primary deep learning platform. Training was performed on Google Colaboratory with T4 GPU, 12 GB VRAM and locally on Nvidia GTX 1660 Ti graphic card having 6 GB VRAM. The model was trained on images of size 256 × 256 for 400 epochs where the batch size is one. The optimization was performed using the Adam optimizer for learning rate of 0.0002, which was further decayed by the learning rate scheduler after every 100 epochs. The proposed model has 112 GFLOPs, 11.37 M trainable parameters and average inference time per image is 0.14 s. Next, we provide a qualitative and quantitative analysis of our suggested dehazing technique.

4. Experiments and Results

This section presents a qualitative and quantitative evaluation of the proposed model, with PSNR and SSIM (Sara et al., 2019; Engin et al., 2018) metrics used for quantitative assessment. We begin with a brief summary of the datasets used to validate our proposed technique.

4.1 Dataset and Experiment

We utilize the Long-wave infrared HIT-UAV aerial dataset i (Suo et al., 2023) along with the infrared aerial Transpetro_Train dataset (Transpetro_train Dataset > Overview, 2025) to carry out our experiments. The HIT-UAV dataset is composed of 2,898 infrared thermal images acquired from 43,470 frames captured by UAVs at several scenarios such as schools, parking lots, roads, and playgrounds. This dataset is related to different conditions such as objects (person, bicycle, car, and other vehicles), flight height (60–130 meters), camera angle (30–90 degrees), and light intensity during both daytime and night time. The Transpetro_Train dataset contains 706 infrared aerial photographs in different environmental conditions and perspectives. These datasets helped in the development as well as evaluation of dehazing model for thermal IR images with both daytime and night-time illumination to ensure the effectiveness of the proposed technique. After the simulation of synthetic haze, the proposed model is initially trained using the two datasets listed above. The dataset was split into train and test sets with a percentage of 80% and 20% respectively.

4.2 Quantitative Analysis

4.2.1 Results on HIT-UAV dataset

The proposed dehazing model is evaluated using two metrics: the Structural Similarity Index (SSIM) and the Peak Signal-to-Noise Ratio (PSNR) defined in equations (8) and (9). SSIM is a perceptual metric, where two images are compared with basic image properties such as brightness, contrast and structural information. SSIM value is normally between 0 and 1 with 1 meaning that the two images are perfectly similar in structure. PSNR is measured in decibels (dB) which is a commonly used metric in reconstructed or compressed images. Higher PSNR means precisely quality image and nearly real image with least distortion.

\begin{aligned} S S I M (x, y) & = \frac{(2 μ_{x} μ_{y} + C_{1}) . (2 σ_{x y} + C_{2})}{(μ_{x}^{2} + μ_{y}^{2} + C_{1}) . (σ_{x}^{2} + σ_{y}^{2} + C_{2})} \end{aligned}

(8)

\begin{aligned} P S N R & = 10 l o g_{10} (\frac{M A X^{2}}{M S E}) \end{aligned}

(9)

Table 1 shows the performance of the proposed model using the HIT-UAV dataset in terms of PSNR and SSIM scores obtained at different haze levels. Table 1 shows that the proposed model attains high PSNR values on the whole and exhibits particularly strong SSIM performance in both light and high haze conditions. Table 1 clearly indicates that the suggested model does not distort the features of the images by maintaining high PSNR values as well as fine texture details.

Table 1.

Results on HIT-UAV Dataset.

Metrics	Light Haze	Medium Haze	High Haze	Very High Haze	Dense Haze
PSNR	30.13	30.63	30.65	28.25	28.86
SSIM	0.89	0.85	0.85	0.60	0.75

4.2.2 Results on Transpetro_Train dataset

Table 2 presents the performance of the proposed dehazing model on the Transpetro_Train dataset under various haze intensities. The results demonstrate that the model consistently achieves high PSNR values and exhibits strong SSIM performance, particularly under very high and medium haze conditions. Figure 6 shows the graphical representation of PSNR and SSIM values on both the HIT-UAV and the Transpetro_Train datasets. The graph on the left, indicates the trend of SSIM in different iterations and shows a progressive improvement in the curve with slight fluctuations. On the other hand, the graph on the right showing PSNR vs iterations, indicates a considerable improvement with some fluctuations and noise indicating the model improvement with increasing iterations.

Figure 6.

Graphical Representation of Fluctuations in the PSNR, SSIM Values When Testing the Model on HIT-UAV and Transpetro_Train Dataset.

Table 2.

Results on Transpetro_Train Dataset.

Metrics	Light Haze	Medium Haze	High Haze	Very High Haze	Dense Haze
PSNR	28.59	28.36	28.12	29.95	28.15
SSIM	0.74	0.82	0.68	0.78	0.64

4.2.3 Comparison with state-of-the-art techniques

A quantitative evaluation was performed on 721 test images, multiple state-of-the-art existing dehazing methods were used in experiments to assess the effectiveness of the proposed model. In Table 3 it can be observed that the proposed dehazing model delivers improved average PSNR and SSIM values than the existing state-of-the-art methods. The proposed technique delivers a PSNR of 30.17 which is better than the existing techniques, the next best score is 29.41. This shows a marked improvement in the quality of image reconstruction and realism of the image due to the inclusion of residual-transformer block within the generator to prevent relevant component of haze. Also, the use of adversarial and cycle consistency loss enhanced the performance of the proposed model compared to the standard deep learning architecture (IDD-Net, IASSF, PSD Dehaze, UCL Dehaze, SPIDE-Net, DA-Net, AOD-Net and Dehaze-Net) and prior based dehazing techniques, such as SLP, DCP, and CAP. The SSIM of the suggested Infra Aerial Dehaze approach is 0.80 which is the better than all other techniques, the proposed technique exhibits its ability to retain the visual quality and structural features in dehazed images. Table 3 demonstrates that, the proposed method delivers better results than traditional prior based dehazing approaches as well as deep learning methods through its average PSNR and SSIM values. The incorporation of residual-transformer block and adversarial learning improved the performance of proposed model for infrared aerial image dehazing.

Table 3.
Comparison with State-of-the-art Techniques.

Metrics

Method Ref PSNR SSIM

IDD-Net S Sun et al. (Sun et al., 2025) 28.82 0.67

IASSF H Li et al. (Li et al., 2025) 27.80 0.56

PSD Dehaze Z Chen et al. (Chen et al., 2021) 27.87 0.60

UCL Dehaze Y Wang et al. (Wang et al., 2024) 29.41 0.75

SPIDE-Net M Qasim et al. (Qasim & Raja, 2022) 28.91 0.69

DA-Net N Kim et al. (Kim et al., 2024) 27.44 0.64

SLP P Ling et al. (Ling et al., 2023) 27.95 0.63

DCP S Yan et al. (Yan et al., 2022) 27.91 0.71

CAP Q Zhuet al. (Zhu et al., 2015) 27.94 0.68

AOD-Net B Li et al. (Li et al., 2017) 28.06 0.66

Dehaze-Net B Cai et al. (Cai et al., 2016) 27.91 0.66

Infra Aerial Dehaze Ours 30.17 0.80

		Metrics
IDD-Net	S Sun et al. (Sun et al., 2025)	28.82	0.67
IASSF	H Li et al. (Li et al., 2025)	27.80	0.56
PSD Dehaze	Z Chen et al. (Chen et al., 2021)	27.87	0.60
UCL Dehaze	Y Wang et al. (Wang et al., 2024)	29.41	0.75
SPIDE-Net	M Qasim et al. (Qasim & Raja, 2022)	28.91	0.69
DA-Net	N Kim et al. (Kim et al., 2024)	27.44	0.64
SLP	P Ling et al. (Ling et al., 2023)	27.95	0.63
DCP	S Yan et al. (Yan et al., 2022)	27.91	0.71
CAP	Q Zhuet al. (Zhu et al., 2015)	27.94	0.68
AOD-Net	B Li et al. (Li et al., 2017)	28.06	0.66
Dehaze-Net	B Cai et al. (Cai et al., 2016)	27.91	0.66
Infra Aerial Dehaze	Ours	30.17	0.80

In Table 4, a quantitative analysis is performed with different state-of-the-art dehazing techniques using no-reference image quality metrics such as Natural Image Quality Evaluator (NIQE) and Perception Based Image Quality Evaluator (PIQE). These metrics provide the perceptual quality score of dehazed images and the lower values indicate the better visual quality and effective haze removal. The proposed Infra Aerial Dehaze achieves NIQE score of 6.13 and PIQE score of 16.47, this suggests the more effective haze removal and improved perceptual image quality. The traditional approaches DCP and CAP achieve lower NIQE scores of 8.10 and 9.52 but higher PIQE scores of 44.29 and 42.64 which indicates less effective haze removal. The UCL Dehaze achieves the lowest NIQE score of 4.81 but has a higher PIQE score of 21.44, these results indicate that the proposed Infra Aerial Dehaze offers effective balance between the improvement in perceptual quality and the removal of haze.

Table 4.

Comparison with State-of-the-art Techniques Using no Reference Metrics NIQE and PIQE.

		Metrics
Method	Ref	NIQE	PIQE
IDD-Net	N Kim et al. (Sun et al., 2025)	8.43	10.46
IASSF	Y Du et al. (Li et al., 2025)	8.37	23.40
PSD Dehaze	Z Chen et al. (Chen et al., 2021)	4.87	23.83
UCL Dehaze	Y Wang et al. (Wang et al., 2024)	4.81	21.44
SPIDE-Net	M Qasim et al. (Qasim & Raja, 2022)	10.67	19.54
DA-Net	N Kim et al. (Kim et al., 2024)	7.41	22.59
SLP	P Ling et al. (Ling et al., 2023)	6.93	43.20
DCP	S Yan et al. (Yan et al., 2022)	8.10	44.29
CAP	Q Zhu et al. (Zhu et al., 2015)	9.52	42.64
AOD-Net	B Li et al. (Li et al., 2017)	9.53	54.56
Dehaze-Net	B Cai et al. (Cai et al., 2016)	9.61	41.07
Infra Aerial Dehaze	Ours	6.13	16.47

Table 5 provides the comparison of average run time (in seconds) between the proposed Infra Aerial Dehaze model and some of the most recent state-of-the-art approach to image dehazing methods on GPU and CPU. The proposed model has competitive inference speed (0.14 s on GPU), better than the traditional technique like DCP and complicated CNN-based models such as Multi-Scale CNN. This is due to the generative nature of our CycleGAN based network that leads to increased forward pass efficiency on inference because no explicit transmission map estimation is needed. The UCL-Dehaze model has a shortest run time, our model has a balance between both run time and quality of dehazing and it is computationally efficient in real time, which makes it suitable in infrared aerial applications. This performance is also due to its lightweight generative architecture of proposed Infra Aerial Dehaze which bypasses the multi-stage processing and makes the fast image translation. Also, Figure 7 shows the graphical representation of average inference time per image for different dehazing methods. The graphical representation in Figure 7 shows the average inference time of the proposed Infra Aerial Dehaze method is 0.14 s, which highlights that the proposed model is efficient in real-time applications.

Figure 7.

Graphical Representation of Average Inference Time per Image.

Table 5.

Average Running Times (Seconds) of Different Methods on Test Images.

Method	Ref	Platform	Average Run Time
DCP	S Yan et al. (Yan et al., 2022)	Python (CPU)	1.40
SPIDE-Net	M Qasim et al. (Qasim & Raja, 2022)	Py-Torch (GPU)	0.44
PSD Dehaze	Z Chen et al. (Chen et al., 2021)	Py-Torch (GPU)	0.39
Multi-Scale CNN	W Ren al. (Ren et al., 2020)	Py-Torch (GPU)	0.84
UCL-Dehaze	Y Wang et al. (Wang et al., 2024)	Py-Torch (GPU)	0.08
Infra Aerial Dehaze	Ours	Py-Torch (GPU)	0.14

4.3 Qualitative Analysis

To evaluate the effectiveness of the proposed image dehazing method, we experimented on 10 different types of infrared (IR) aerial outdoor images including; Light, Medium, High, Very High, and Dense haze from both the HIT-UAV and Transpetro_Train datasets. The performance evaluation carried out on these datasets clearly indicates that the proposed method exhibits good robustness to images of outdoor aerial IR images of varying haze levels and the fine texture information of the image is preserved while performing the dehazing process on the images.

Figure 8 shows the qualitative results produced by the proposed model when testing the model on ten samples of images from the HIT-UAV dataset. As shown in Figure 8 in case of dense haze the proposed model performs more effective dehazing and preserving the finer texture details of images. Similarly, Figure 9 shows the qualitative results from the Transpetro_Train dataset, demonstrating the model's ability to preserve the minute texture information contained in all ten dehazed images of Figure 9 each with varying levels of haze intensity. These findings demonstrate the suggested method's robustness in addressing a variety of hazy circumstances.

Figure 8.

Qualitative Results of Proposed Model on HIT-UAV Dataset.

Figure 9.

Qualitative Results of Proposed Model on Transpetro_Train Dataset.

Figure 10 represents the qualitative analysis of the proposed method with the other existing methods for image dehazing, the evaluation is conducted on of 721 test images to assess the effectiveness of the proposed approach. As indicated in Figure 10 the suggested method provides more clarity and detail recovery of image compared to other the dehazing methods. Each image in Figure 10 is labeled by red bounding boxes, which point out the specific key regions and allowing a focused comparison of fine details and structural clarity of proposed approach with existing techniques. These regions illustrate that our proposed approach efficiently maintains the structural characteristics of a scene, while other approaches produce over-smoothing or introduce artifacts in dehazed images. This visual analysis supports the effectiveness of our approach in producing improved quality dehazed infrared images. Compared to deep learning algorithms such as IDD-Net, IASSF, UCL Dehaze, SPIDE-Net, AOD-Net, Dehaze-Net, PSD Dehaze and traditional methods including SLP, DCP and CAP, the proposed method is found to be more effective in different haze conditions, as the obtained dehazed images are very close to the ground truth and have higher contrast and image visibility. Figure 10 shows that PSD Dehaze and UCL also suffers with the residual haze and artifacts, whereas the proposed model significantly enhances image visibility while preserving structural information in image without artifacts. The results from both quantitative and qualitative assessments demonstrate that the proposed method eliminates haze under different conditions, particularly in case of dense haze.

Figure 10.

Qualitative Results of Proposed Model with Existing Dehazing Methods.

Figure 11 shows the qualitative evaluation of the proposed dehazing model on real infrared hazy images from the IRSTD-1 K dataset (Zhang et al., 2022) that mainly contains ground to ground scenes. Although due to unavailability of infrared hazy aerial datasets, the proposed model still demonstrates good generalization capability and improve the visual image quality under the ground-based hazy conditions. In Figure 11 left column shows the original haze-degraded infrared hazy inputs, while the right column displays the corresponding dehazed outputs images generated by our proposed Infra aerial dehaze method. The red bounding boxes in Figure 11 indicate that the model is able to restore important structural and semantic information and improving the scene visibility in dehazed images. These results also confirm the proposed model robustness on real infrared hazy conditions.

Figure 11.

Qualitative Results of Proposed Model on Real Infrared Hazy IRSTD-1 K Dataset.

Figure 12 shows Qualitative the results of the proposed model on nighttime infrared hazy images from the LLVIP dataset (LLVIP: A Visible-infrared Paired Dataset for Low-light Vision, 2025). LLVIP dataset is a benchmark dataset used to deal with low-light and infrared-visible image processing tasks. It includes paired visible and infrared images acquired in different illumination conditions such as low-light and night scenarios. In Figure 12 the first column contains the visible-spectrum night images, the second column indicates the ground truth (GT) infrared images under clear conditions, the third column indicates the infrared night hazy images, and then finally the fourth column shows the dehazed outputs images generated by the proposed Infra Aerial Dehaze model. As indicated by the red boxes, the proposed technique efficiently manages to retrieve the object visibility and structural information under the difficult night time and low-contrast conditions with high level of generalization ability in complex environment by infrared haze.

Figure 12.

Qualitative Results of Proposed Model on Night Hazy Images.

Although the proposed model shows a high level of performance in recovering visibility in night-time infrared hazy images, there exist some limitations that can be noted in Figure 12. As depicted in Figure 12, the proposed model struggles to work in cases with extremely dense haze or too low thermal contrast, where the fine objects details like faint pedestrians are partially lost or blurred. This demonstrates the limitation of proposed model to tackle severe occlusion conditions and low thermal emission, which can be further enhanced by many structural priors or temporal cues.

5. Conclusions

We proposed a CycleGAN based dehazing model named as Infra Aerial Dehaze, for dehazing in infrared aerial imagery, that can handle different levels of haze intensity and maintain fine details of texture in image. We also created a new dataset comprising of 3604 aerial infrared hazy images synthesized with the help of the atmospheric scattering model. The experimental results on HIT-UAV and Transpetro_Train datasets, shows that our method produces visually better images and achieves higher PSNR and SSIM values across the different degrees of haze. The reconstructed images quality depicts the superiority of proposed model over existing methods in terms of improving picture clarity and retaining fine details and structure. Although the proposed model performs well on synthetically generated infrared hazy aerial images, yet the unavailability of real infrared hazy aerial datasets leaves a small space for further evaluation of the proposed model. Despitethis limitation, the qualitative and quantitative results demonstrate the feasibility of the proposed model in real hazy scenarios. The proposed approach can be used in poor environmental conditions like fog, smoke, dust, mist, pollution etc. It can also be utilized in real-time processing systems for UAV tracking by effectively restoring visibility of image in hazy conditions.

Footnotes

Ethical Approval

No ethical approval was required for the study.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

ORCID iDs

Farhan Mushtaq Hashmi

Gulistan Raja

References

Cai

Jia

Qing

Tao

(2016). Dehazenet: An end-to-end system for single image haze removal. IEEE transactions on Image Processing, 25(11), 5187–5198. https://doi.org/10.1109/TIP.2016.2598681

Cao

Dong

Zhou

Zhao

Zhang

(2020). Single image dehazing algorithm using generative adversarial network based on feature pyramid network. In 2020 International Conference on Computer Vision, Image and Deep Learning (CVIDL), IEEE, 48–53. Accessed: Sept. 16, 2024. [Online]. Available: https://ieeexplore.ieee.org/abstract/document/9270528/

Chaitanya

Mukherjee

. (2021) Single image dehazing using improved cycleGAN. Journal of Visual Communication and Image Representation, 74, 103014. https://doi.org/10.1016/j.jvcir.2020.103014

Chen

Wang

Yang

Liu

(2021). PSD: Principled synthetic-to-real dehazing guided by physical priors. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 7180–7189. Accessed: June 08, 2025. [Online]. Available: http://openaccess.thecvf.com/content/CVPR2021/html/Chen_PSD_Principled_Synthetic-to-Real_Dehazing_Guided_by_Physical_Priors_CVPR_2021_paper.html

Cui

Dai

Zhang

(Jan. 2024). SCL-Dehaze: Toward real-world image dehazing via semi-supervised codebook learning. Electronics, 13(19), Art. no. 19. https://doi.org/10.3390/electronics13193826

Dong

Liu

Zhang

Chen

Qiao

(2020). FD-GAN: Generative adversarial networks with fusion-discriminator for single image dehazing. In Proceedings of the AAAI conference on artificial intelligence, 10729–10736. Accessed: Sept. 16, 2024. [Online]. Available: https://aaai.org/ojs/index.php/AAAI/article/view/6701

Engin

Genç

Kemal Ekenel

(2018). Cycle-dehaze: Enhanced cyclegan for single image dehazing. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, 825–833. Accessed: Sept. 15, 2024. [Online]. Available: https://openaccess.thecvf.com/content_cvpr_2018_workshops/w13/html/Engin_Cycle-Dehaze_Enhanced_CycleGAN_CVPR_2018_paper.html

Erlenbusch

, et al. (2023). Thermal infrared single image dehazing and blind image quality assessment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 459–469. Accessed: Sept. 15, 2024. [Online]. Available: https://openaccess.thecvf.com/content/CVPR2023W/PBVS/html/Erlenbusch_Thermal_Infrared_Single_Image_Dehazing_and_Blind_Image_Quality_Assessment_CVPRW_2023_paper.html

Fan

Gan

Fan

Chen

C. P.

(2022). Multiscale cross-connected dehazing network with scene depth fusion. IEEE Transactions on Neural Networks and Learning Systems, 35(2), 1598–1612. https://doi.org/10.1109/TNNLS.2022.3184164

10.

Liu

Chen

Wang

(2021). Dw-gan: A discrete wavelet transform gan for nonhomogeneous dehazing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 203–212. Accessed: Sept. 16, 2024. [Online]. Available: https://openaccess.thecvf.com/content/CVPR2021W/NTIRE/html/Fu_DW-GAN_A_Discrete_Wavelet_Transform_GAN_for_NonHomogeneous_Dehazing_CVPRW_2021_paper.html

11.

Gan

Zhao

Chen

(2020). Multilevel image dehazing algorithm using conditional generative adversarial networks. IEEE Access, 8, 55221–55229. https://doi.org/10.1109/ACCESS.2020.2981944

12.

Gao

H.-M.

Guo

(2018). Detail preserved single image dehazing algorithm based on airlight refinement. IEEE Transactions on Multimedia, 21(2), 351–362. https://doi.org/10.1109/TMM.2018.2856095

13.

HIT-UAV: A High-altitude Infrared Thermal Dataset . (2025). “HIT-UAV: A High-altitude Infrared Thermal Dataset,” Accessed: Aug. 05, 2025. [Online]. Available: https://www.kaggle.com/datasets/pandrii000/hituav-a-highaltitude-infrared-thermal-dataset

14.

Hovhannisyan

S. A.

Gasparyan

H. A.

Agaian

S. S.

Ghazaryan

(2022). AED-Net: A single image dehazing. IEEE Access, 10, 12465–12474. https://doi.org/10.1109/ACCESS.2022.3144402

15.

Hua

Ding

(2021). Image dehazing using near-infrared information based on dark channel prior. Procedia Computer Science, 187, 18–23. https://doi.org/10.1016/j.procs.2021.04.028

16.

Huang

Dong

Huang

Yan

Chen

(May 2025). DCD-Net: Weakly supervised decomposition learning for real-world image dehazing. Signal Processing, 230, 109826. https://doi.org/10.1016/j.sigpro.2024.109826

17.

Kan

Zhang

Cen

(2022). A GAN-based input-size flexibility model for single image dehazing. Signal Processing: Image Communication, 102, 116599. https://doi.org/10.1016/j.image.2021.116599

18.

Kim

Choi

I.-S.

Han

S.-S.

Jeong

C.-S.

(2024). DA-Net: Dual attention network for haze removal in remote sensing image. IEEE Access, 12, 136297–136312. https://doi.org/10.1109/ACCESS.2024.3459588

19.

, et al. (2018). Benchmarking single-image dehazing and beyond. IEEE Transactions on Image Processing, 28(1), 492–505. https://doi.org/10.1109/TIP.2018.2867951

20.

Peng

Wang

Feng

(2017). Aod-net: All-in-one dehazing network. In Proceedings of the IEEE international conference on computer vision, 4770–4778. Accessed: Sept. 16, 2024. [Online]. Available: http://openaccess.thecvf.com/content_iccv_2017/html/Li_AOD-Net_All-In-One_Dehazing_ICCV_2017_paper.html

21.

Fang

Zhang

Liu

(2025, February 8) Infrared-assisted single-stage framework for joint restoration and fusion of visible and infrared images under hazy conditions. arXiv: arXiv:2411.12586, 1–13. https://doi.org/10.48550/arXiv.2411.12586

22.

Ling

Chen

Tan

Jin

Chen

(2023). Single image dehazing using saturation line prior. IEEE Transactions on Image Processing, 32, 3238–3253. https://doi.org/10.1109/TIP.2023.3279980

23.

Liu

Wang

Shiri

Lin

(July 2025). VNDHR: Variational single nighttime image dehazing for enhancing visibility in intelligent transportation systems via hybrid regularization. IEEE Transactions on Intelligent Transportation Systems, 26(7), 10189–10203. https://doi.org/10.1109/TITS.2025.3550267

24.

LLVIP: A Visible-infrared Paired Dataset for Low-light Vision . (2025). Accessed: Aug. 08, 2025. [Online]. Available: https://bupt-ai-cz.github.io/LLVIP/

25.

Long

Zha

Zhang

(2020). PHC-GAN: Physical constraint generative adversarial network for single image dehazing. In 2020 IEEE 32nd International Conference on Tools with Artificial Intelligence (ICTAI), IEEE, 545–549. Accessed: Sept. 16, 2024. [Online]. Available: https://ieeexplore.ieee.org/abstract/document/9288256/

26.

Mehta

Sinha

Mandal

Narang

(2021). Domain-aware unsupervised hyperspectral reconstruction for aerial image dehazing. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, 413–422. Accessed: Sept. 17, 2024. [Online]. Available: http://openaccess.thecvf.com/content/WACV2021/html/Mehta_Domain-Aware_Unsupervised_Hyperspectral_Reconstruction_for_Aerial_Image_Dehazing_WACV_2021_paper.html

27.

Mehta

Sinha

Narang

Mandal

(2020). Hidegan: A hyperspectral-guided image dehazing gan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, 212–213. Accessed: Sept. 17, 2024. [Online]. Available: http://openaccess.thecvf.com/content_CVPRW_2020/html/w14/Mehta_HIDeGan_A_Hyperspectral-Guided_Image_Dehazing_GAN_CVPRW_2020_paper.html

28.

Mehta

Narang

Chamola

Zeadally

(2022). Deep learning enhanced uav imagery for critical infrastructure protection. IEEE Internet of Things Magazine, 5(2), 30–34. https://doi.org/10.1109/IOTM.004.2200006

29.

Nawaz

Raja

Qasim

(2021). A new approach for dehazing and enhancement of infrared images. In 2021 international Bhurban conference on applied sciences and technologies (ibcast), IEEE, 545–551. Accessed: Sept. 15, 2024. [Online]. Available: https://ieeexplore.ieee.org/abstract/document/9393238/

30.

Parihar

A. S.

Gupta

Y. K.

Singodia

Singh

(2020). A comparative study of image dehazing algorithms. In 2020 5th International Conference on Communication and Electronics Systems (ICCES), IEEE, 766–771. Accessed: Sept. 15, 2024. [Online]. Available: https://ieeexplore.ieee.org/abstract/document/9138037/

31.

Parihar

A. S.

Singh

Ganotra

Yadav

(2022). Contrast aware image dehazing using generative adversarial network. In 2022 2nd international conference on intelligent technologies (CONIT), IEEE, 1–6. Accessed: Sept. 17, 2024. [Online]. Available: https://ieeexplore.ieee.org/abstract/document/9847710/

32.

Qasim

Raja

(2022). Spide-net: Spectral prior-based image dehazing and enhancement network. IEEE Access, 10, 120296–120311. https://doi.org/10.1109/ACCESS.2022.3221992

33.

Ren

Pan

Zhang

Cao

Yang

M.-H.

(Jan. 2020). Single image dehazing via multi-scale convolutional neural networks with holistic edges. Int J Comput Vis, 128(1), 240–259. https://doi.org/10.1007/s11263-019-01235-8

34.

Sara

Akter

Uddin

M. S.

(2019). Image quality assessment through FSIM, SSIM, MSE and PSNR—a comparative study. Journal of Computer and Communications, 7(3), 8–18. https://doi.org/10.4236/jcc.2019.73002

35.

Sarker

Akter

Uddin

M. S.

(2019). Simulation of hazy image and validation of haze removal technique. Journal of Computer and Communications, 7(2), 62–72. https://doi.org/10.4236/jcc.2019.72005

36.

Sim

Choi

J.-S.

Seo

Kim

(2018). High-resolution image dehazing with respect to training losses and receptive field sizes. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 912–919. Accessed: Sept. 17, 2024. [Online]. Available: http://openaccess.thecvf.com/content_cvpr_2018_workshops/w13/html/Sim_High-Resolution_Image_Dehazing_CVPR_2018_paper.html

37.

Suárez

P. L.

Sappa

A. D.

Vintimilla

B. X.

Hammoud

R. I.

(2018). Deep learning based single image dehazing. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, 1169–1176. Accessed: Sept. 16, 2024. [Online]. Available: http://openaccess.thecvf.com/content_cvpr_2018_workshops/w21/html/Suarez_Deep_Learning_Based_CVPR_2018_paper.html

38.

Sun

, et al. (2025). IDDNet: Infrared object detection network based on multi-scale fusion dehazing. Sensors, 25(7), 2169. https://doi.org/10.3390/s25072169

39.

Suo

Wang

Zhang

Chen

Zhou

Shi

(2023). HIT-UAV: A high-altitude infrared thermal dataset for unmanned aerial vehicle-based object detection. Scientific Data, 10(1), 227. https://doi.org/10.1038/s41597-023-02066-6

40.

Transpetro_train Dataset > Overview . (2025). “Transpetro_train Dataset > Overview,” Roboflow. Accessed: Nov. 09, 2024. [Online]. Available: https://universe.roboflow.com/usp-xkiuj/transpetro_train

41.

Wang

Cui

Lan

(Nov. 2021). Prior-guided multiscale network for single-image dehazing. IET Image Processing, 15(13), 3368–3379. https://doi.org/10.1049/ipr2.12333

42.

Wang

, et al. (2022). Cycle-snspgan: Towards real-world image dehazing via cycle spectral normalized soft likelihood estimation patch gan. IEEE Transactions on Intelligent Transportation Systems, 23(11), 20368–20382. https://doi.org/10.1109/TITS.2022.3170328

43.

Wang

, et al. (2024). UCL-Dehaze: Toward real-world image dehazing via unsupervised contrastive learning. IEEE Transactions on Image Processing, 33, 1361–1374. https://doi.org/10.1109/TIP.2024.3362153

44.

Yan

Zhu

Yun

Wang

(2022). An infrared image dehazing method based on modified dark channel prior. In International Conference on Biometrics, Microelectronic Sensors, and Artificial Intelligence (BMSAI), SPIE , 132–138. Accessed: Sept. 15, 2024. [Online]. Available: https://www.spiedigitallibrary.org/conference-proceedings-of-spie/12252/122520N/An-infrared-image-dehazing-method-based-on-modified-dark-channel/10.1117/12.2640246.short

45.

Zhang

Yang

Bai

Zhang

Guo

(2022). ISNet: Shape matters for infrared small target detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 877–886. Accessed: June 07, 2025. [Online]. Available: http://openaccess.thecvf.com/content/CVPR2022/html/Zhang_ISNet_Shape_Matters_for_Infrared_Small_Target_Detection_CVPR_2022_paper.html

46.

Zhu

J.-Y.

Park

Isola

Efros

A. A.

(2017). Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, 2223–2232. Accessed: Nov. 06, 2024. [Online]. Available: http://openaccess.thecvf.com/content_iccv_2017/html/Zhu_Unpaired_Image-To-Image_Translation_ICCV_2017_paper.html

47.

Zhu

Mai

Shao

(2015). A fast single image haze removal algorithm using color attenuation prior. IEEE transactions on Image Processing, 24(11), 3522–3533. https://doi.org/10.1109/TIP.2015.2446191