Single-Image Dehazing of High-Voltage Power Transmission Line Based on Unsupervised Iterative Learning of Knowledge Transfer

Abstract

Single-image dehazing of high-voltage power transmission lines (HPTLs) using deep learning methods confronts two critical challenges: the non-homogeneous haze distribution in HPTL images and the unavailability of paired clear images for supervised training. To overcome these issues, this paper proposes a novel dehaze neural network, named FIF-RSCT-Net, that employs a hybrid supervised-to-unsupervised iterative learning approach according to the characteristic of HPTL single images. The FIF-RSCT-Net incorporates the Spatial–Channel Feature Intersection modules and Residual Separable Convolution Transformers to enhance the feature representation capability. Crucially, this novel architecture could learn more generalized dehazing knowledge that can be transferred from the original image domain to HPTL scenarios. In the dehazing knowledge transformation, an unsupervised iterative learning mechanism based on the Line Segment Detector is designed to optimize the restoration of power transmission lines. The effectiveness of FIF-RSCT-Net on the original image domain is demonstrated in the comparative experiments of the I-Haze, O-Haze, NH-Haze, and SOTS datasets. Our methodology achieves the best average PSNR of 24.647 dB and SSIM of 0.8512. And the qualitative evaluation of unsupervised iterative learning results shows that the missed line segments are exhibited during progressive training iterations.

1. Introduction

In winter, the images of high-voltage power transmission lines (HPTLs) often suffer from haze, which makes the following vision tasks hard to perform effectively. Therefore, single-image dehazing of HPTL is an important preprocessing for the power system.

Image dehazing consistently remains a significant research challenge in machine vision. However, for the HPTL images, the heavy fog is non-homogeneous due to the spatially random distribution and varying concentration. Therefore, a strong nonlinear model is necessary to restore clear images from non-homogeneous haze. A more challenging issue is that the hazy images of HPTL lack paired clear labels, which are essential for supervised learning.

Previous works on domain adaptation for single-image dehazing show that the unsupervised learning approach is often tailored to the specific scenarios of the hazy images by designing specialized neural network architectures and knowledge transfer mechanisms. The key distinguishing characteristics of HPTL images are the structural features, which are represented as lines, corners, and topological features. In addition to these fundamental challenges, the structural characteristics of HPTL images introduce further complexity to the dehazing task.

To address the dehazing challenges, this paper proposes a novel dehazing neural network named FIF-RSCT-Net. Leveraging the characteristics of HPTL images, the network incorporates a Feature Intersection Fusion (FIF) mechanism and a Residual Separable Convolution Transformer (RSCT) module to enhance feature representation and reconstruction performance. The FIF-RSCT-Net acquires the dehazing knowledge through two stages: a supervised dehazing stage using labeled images from the original domain, and an unsupervised iterative learning stage that transfers the learned dehazing knowledge to unlabeled HPTL images. The principal contributions of this work are summarized as follows:

(1) In the FIF-RSCT-Net, an RSCT module is presented to model the topological relationships of the image. This architecture employs separable convolutions to generate Query, Key, and Value parameters of the Transformer, significantly improving computation efficiency. Furthermore, a FIF module is introduced to deeply fuse the features extracted by RSCT, which can improve the transferability of the model across diverse domains.

(2) A novel Spatial–Channel Feature Intersection (SCFI) module is proposed for FIF-RSCT Net to enhance the extraction of line and point features in shallow convolutional layers. This module incorporates a Channel Feature Intersection (CFI) method, which stacks the weighted channel features alternately to improve the network convergence.

(3) An iterative learning mechanism guided by a Line Segment Detector (LSD) loss function is designed to effectively transfer the learned dehazing knowledge from the labeled domain to unlabeled HPTL images. The unsupervised iterative learning method can improve the restoration of hazy HPTL images.

The remaining sections are structured as follows: Section 3 details the architecture of the proposed FIF-RSCT-Net; Section 4 explains the unsupervised iterative learning mechanism for dehazing knowledge transfer; Section 5 presents the experiments and results analysis; and the conclusion is provided in Section 6.

2. Background and Related Works

In the field of single-image dehazing, traditional methods typically operate without requiring paired labels for model training. A significant milestone in single-image dehazing was established by He et al. in 2009. They proposed the Dark Channel Prior (DCP) [1] method, which is based on the classical atmospheric scattering model (ASM) and exploits the dark channel prior to estimate the transmission map for ASM. Subsequently, many methods inherited from DCP and introduced other information, such as context [2], Rank-One Prior [3], and saturation line prior [4]. The traditional methods typically rely on manually physical models; hence, they are hard to describe the strong non-linear feature of the dehazing process.

Deep learning demonstrates powerful nonlinear modeling capacities, which have led to tremendous success [5,6]. Hence, it has been introduced into single-image dehazing. The fundamental architectures include convolutional neural networks (CNNs) [7,8,9,10,11,12,13,14] and transformers [15,16]. Researchers have developed multitudinous frameworks, DehazeNet [17], multi-scale CNN [18,19,20,21], Ranking-CNN [22], cascaded CNN [23,24], Color Transferred CNN [25], All-in-One Dehazing Network [26], Light-DehazeNet [27], Quaternion neural networks [27], and the application scenarios [28,29,30] encompass remote sensing hazy image [31,32,33,34,35], under water image [36,37,38], autonomous driving [39,40,41], coal mining [42], and image restoration [43,44].

The aforementioned deep learning methods all require clear labels for supervised training. However, obtaining the clean image to serve as a labels for hazy HPTL images is often impractical in real-world scenarios. Consequently, the core challenge in HPTL single-image dehazing lies in developing a deep neural network-based dehazing method to restore images without paired labels.

In recent years, the research has progressively shifted toward unsupervised dehazing techniques to overcome this limitation. The methods primarily fall into two categories: the generator-discriminator network (GAN)-based framework and domain adaptation approaches.

Within the GAN framework, the clear images are synthesized by the generator, while the discriminator evaluates their authenticity against the characteristics of real clear images. However, the generator produces plausible, clear images without being rigidly bound to the details of the input hazy images. Hence, the prior knowledge is introduced into GAN to guide the clear image production process. DFP-Net [45] incorporates dark channel prior, and UIDF-Net [46] applies the frequency information. Similar approaches have been extensively explored in [47,48,49,50,51]. Essentially, the GAN-based methods are supervised learning. They need a clear image as the reference criterion.

Domain adaptation for image dehazing follows a two-stage process. At first, the model is trained on hazy images with paired ground truth labels in the original domain. Then, the learned dehazing knowledge is transferred into the unlabeled target domain. Hence, the core challenge lies in designing an effective knowledge transfer mechanism to bridge the domain gap. To address this issue, the DCM-dehaze method [52] establishes a bidirectional contour constraint optimization to enhance the feature capture ability from the pretrained domain to the target domain. TOLPnet [53] presents a Typhoon Optimization algorithm to optimize the hyperparameters of a CNN model for knowledge transfer. In reference [54], the transfer learning is separated into two stages, combining synthetic-to-real domain adaptation with degraded-to-clear domain transformation. UCL-Dehaze [55] applies the patch and pixel- wise contrastive learning for unsupervised transfer. In Haze Style Transfer (HST) [56], the multiple domains of hazy styles are defined based on K-means clustering, and the dehazing process is considered as transferring the special instance of HST into the no-haze style. In ContourDCP [57], the training labels for the dehazed model are reformulated using the dark channel prior through a contour-based approach.

Obviously, for the non-homogeneous hazy images of HPTL, the neural network architecture can incorporate structural features into the knowledge transfer mechanism to enhance the restoration of power lines. However, in the hazy conditions of HPTL, these critical features frequently suffer from severe degradation or even complete disappearance. Consequently, the  core challenge in unsupervised HPTL image dehazing lies in augmenting a pre-trained model’s capacity to (1) effectively extract and reconstruct these critical line, point, and topological features, and (2) maintain strong domain adaptability for HPTL scenarios. These two aspects represent the key contributions that distinguish our work from existing methods.

3. Architecture

The proposed novel HPTL image dehazing neural network, FIF-RSCT-Net, is illustrated in Figure 1. The acquisition of dehazing knowledge is implemented through supervised and unsupervised iterative learning stages.

Firstly, in the supervised learning stage, the FIF-RSCT-Net is trained in the labeled original image domain. In the unsupervised iterative learning stages, the final convolution layers are fine-tuned through iterative learning, which enables effective transfer of dehazing knowledge to the HPTL image domain under the guidance of the LSD loss function.

The critical challenge of FIF-RSCT-Net for HPTL image dehazing lies in enhancing the model’s capability to capture generalizable knowledge for image dehazing; meanwhile, the learned dehazing knowledge should be more readily transferred to HPTL image domain.

To address this issue, in the FIF-RSCT-Net, the SCFI modules, RSCT modules, and FIF modules are proposed and integrated into the classical encoder–decoder structure. In the encoder, the paired SCFI module constitutes a residual network to jointly extract spatial and channel features from shallow layers. While the deep layers are composed of RSCT modules. The FIF modules are applied to fuse the extracted features and restore the clear image.

3.1. SCFI Module

In a hazy HPTL image, the shallow features are typically explicit representations of line structures and point features. These low-level features serve as crucial guidance for the decoder during clear image reconstruction. Notably, such features demonstrate strong domain adaptability, making them particularly transferable to the target domain. Motivated by this observation, this paper proposes the SCFI module, whose structure is detailed in Figure 2.

Since a convolutional network provides a straightforward approach to mapping spatial representations into channel dimensions, the primary concept of the SCFI module is to integrate spatial and channel-wise features, for which a weighted fusion mechanism is proposed.

The channel weights are generated through a spatial global average pooling (SGAP) followed by a 1D encoder–decoder. The SGAP computes the mean activation of each channel to represent the average energy distribution, which is decomposed into the channels. Subsequently, the channel mean features are transferred into weight coefficients through the 1D encoder–decoder network.

The spatial weights are derived from a dual-branch pooling mechanism composed of channel global maximum pooling (CGMP) and channel global average pooling (CGAP). These complementary operations extract both peak and mean activations of the channels. Then, the features are concatenated to form the spatial energy descriptors that will be converted into the weight coefficients by a convolution and sigmoid activation function. Assigned coefficients serve as spatial and channel descriptors, which are then fused by the CFI module.

3.2. CFI Module

The CFI module is motivated by the deep feature fusion. The channel-weighted features $F_C\in {\mathbb{R}}^{h\times w\times c}$ and spatial-weighted features $F_s\in {\mathbb{R}}^{h\times w\times c}$ constitute distinct complementary representations of the original feature in different feature spaces. It implies that the feature fusion operation essentially performs a mapping from both channel and spatial spaces ${\mathbb{R}}_C\times {\mathbb{R}}_S$ to an integrated feature space ${\mathbb{R}}_F$.

The common fusion approaches typically employ simple concatenation, where $F_C$ is appended behind $F_S$ to form the combined feature $F_F\in {\mathbb{R}}^{h\times w\times 2c}$. However, this method would lead to abrupt gradient transitions at feature boundaries. From a theoretical perspective, such a discontinuous characteristic poses significant mathematical modeling challenges; hence, the fusion network would be hard to converge during training process.

To address these limitations, this paper introduces the CFI module, which implements a more sophisticated fusion strategy as illustrated in Figure 3.

The feature fusion is implemented through an interleaved channel stacking strategy. This mechanism smoothens the gradual transitions between the distinct feature spaces $F_C$ and $F_S$. The smooth fusion features provide more stable gradient flow during backpropagation to improve the convergence of the network.

3.3. RSCT Module

The deep features in the encoder inherently describe global relationships. Therefore, the topological features of the HPTL image could be explicitly represented in the deep layers of the encoder. Transformer is excellent at modeling such global dependencies effectively through the self-attention mechanism, which operates via trainable Query (Q), Key (K), and Value (V) matrices. Generally, these matrices are initialized randomly. However, for the single-image dehazing of HPTL, the Q, K, and V demands more sophisticated feature representation to ensure the dehazing knowledge transfer. Consequently, in this section, the RSCT module is proposed to achieve this purpose. The structure is shown in Figure 4.

The primary innovation of the RSCT module lies in the reconstruction of the Q, K, and V matrices, which are replaced by the separable convolution.

The convolution operations inherently capture local relationships of the input information. It implies that the modified Q, K, and V can represent the topology information of the HPTL image. Furthermore, the dual residual connections structure in the RSC-Transformer can avoid the gradient vanishing and explosion during training. Therefore, the RSCT module can converge more quickly. That implies the dehazing knowledge represented by the RSCT module would be more easily transferred from the original domain to the HPTL image domain.

3.4. FIF Module in Decoder

The decoder in the proposed architecture is designed to reconstruct a clear image from the extracted multi-scale features. As illustrated in Figure 1, the restoration process is guided by four distinct scale features. To enable effective feature fusion, each scale feature should be upsampled to the same spatial resolution using transposed convolution layers with stride = 2. The proposed FIF module is depicted in Figure 5. Firstly, the FIF module computes the spatial weights for each position in the feature maps. Then, the weights are multiplied by the input feature spatially to emphasize spatially relevant features. Following the same principle as the CFI module, the fusion process employs an intersection mechanism to optimally combine features while maintaining the gradient stability.

It is important to note that the decoder incorporates the final convolution layers to restore the image. These layers will be fine-tuned during the dehazing knowledge transfer with unsupervised iterative learning. The final convolution layers follow a conventional structure that is illustrated in Figure 6a.

The FIF-RSCT-Net training in the original domain employs a supervised learning approach. Let the dehazed image and label be denoted by $I_{de}$ and $I_{GT}$. The loss function is defined using the smooth L1 loss, formulated as follows:

$$L_s=\left\{\begin{array}{cc}{\displaystyle \frac{1}{6\times H\times W}}{\displaystyle \sum _{c=1,2,3}}{\displaystyle \sum _{i=1}^{H\times W\times 3}}{\left|I_{de}(C,i)-I_{GT}(C,i)\right|}^2,\hfill & \left|I_{de}(C,i)-I_{GT}(C,i)\right|<1\hfill \\ {\displaystyle \frac{1}{3\times H\times W}}{\displaystyle \sum _{c=1,2,3}}{\displaystyle \sum _{i=1}^{H\times W\times 3}}\left(\left|I_{de}(C,i)-I_{GT}(C,i)\right|-0.5\right),\hfill & \left|I_{de}(C,i)-I_{GT}(C,i)\right|\ge 1\hfill \end{array}\right.$$

(1)

where the symbol C and $\mathit{i}$ are, respectively, the channel and pixel index; H and W are the height and width of the image.

4. Unsupervised Iterative Learning for Dehazing Knowledge Transfer

The learned dehazing knowledge from the original domain requires effective transfer to the unlabeled HPTL image domain. To address this, this paper proposes an iterative learning mechanism that fine-tunes the final convolution layers of the dehazing neural network, by  which the learned dehazing knowledge can be transferred to HPTL image domain.

For the HPTL images, the critical information is a topological structure that is always represented through line structure. Consequently, line features in a hazy HPTL image can serve as the primary targets for dehazing. Hence, we employ the Line Segment Detector (LSD) [58] to monitor line visibility in the network output. The iterative learning process continues until the visible lines stop increasing.

For the $k-th$ iteration, suppose network output is $I_{de}^k\in {\mathbb{R}}^{H\times W\times 3}$, the result of LSD is denoted by $p^k\in {\mathbb{R}}^{H\times W}$, written as follows:

$$p^k\left(i\right)=LSD\left(I_{de}^k\left(i\right)\right)=\left\{\begin{array}{cc}1,\hfill & I_{de}^k\left(i\right)\in line\hfill \\ 0,\hfill & I_{de}^k\left(i\right)\notin line\hfill \end{array}\right.$$

(2)

where $\mathit{i}=\{1,2,\dots ,H\times W\}$ denotes the pixel index. $LSD(·)$ is the function of LSD. We define the loss function for fine tuning the final convolutional layers as follows:

$$L_{LSD}^k=\sum _{i=1}^{H\times W}\left|p^{k-1}\left(i\right)-p^k\left(i\right)\right|,k=1,2,\dots ,N$$

(3)

Theoretically, if the lines in the HPTL image are perfectly reconstructed, the $L_{LSD}$ is close to zero. In practice, we empirically set a maximum iteration number $N=50$ and minimum threshold value $\theta =5$, which provides a good balance between visual line completeness and computational efficiency. If $k\ge N$ or $|L_{LSD}^k|<\theta$, the iterative learning stops. The unsupervised iterative learning mechanism is illustrated in Figure 6b and Algorithm 1.

Algorithm 1 Unsupervised Iterative Learning for Dehazing Knowledge Transfer

Input: A hazy HPTL image I, iterative number $k=0$, maximum iteration number $N=50$, minimum threshold value $\theta =5$. NN: the final convolutional layers of FIF-RSCT-Net will be fine-tuned, while the other layers weights are frozen. Output: Dehazed HPTL image $I_{de}$ For $k=0$; $k++$; $k<N$  Get the neural network output $I_{de}^k$=NN(I)  If $k<1$:     Check the lines in the image by Equation (2), and get $p^k\left(i\right)$     Continue For-loop  Else     Check the lines in the image by Equation (2), and get $p^k\left(i\right)$     Calculate the loss $L_{LSD}^k$ for NN by Equation (3)     If $|L_{LSD}^k|<\theta$:        Break For-loop     Else:         Loss Back Propogation to adjust weights of the final convolution layers      End if   End if End For loop Output $I_{de}=I_{ed}^k$

5. Experiment and Result Analysis

5.1. The Dataset and Experiment Procedures

The proposed methodology comprises two key components. The first is the developed FIF-RSCT-Net to enhance the feature extraction in the supervised learning, and the second is the dehazing knowledge transfer through the iterative learning method based on LSD loss. Therefore, in the experiment, the two different kinds of datasets are respectively adopted to validate the effectiveness of each component.

The supervised learning datasets include I-Haze [59], O-Haze [60], NH-Haze [61], and SOTS [62]. Among which, I-Haze, O-Haze, and NH-Haze are the natural hazy environment datasets, while SOTS is the synthetic hazy image dataset. In the dehazing knowledge transfer step, the natural HPTL images with different concentrations are applied to iteratively fine-tuning the dehazing neural network. In the experiment, five classical dehazing algorithms, DCP, FFANet [63], GCANet [64], GDNet [65], and MSBDN [66], are introduced for comparative analysis.

In the supervised learning stage, because the training samples have clear labels, the performance evaluation indicators are Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM). In unsupervised iterative learning, the subjective assessment is adopted to evaluate the results, since the paired ground truth of the HPTL images is lacking.

5.2. Training Details

The configuration of computer server applied for training is as follows: CPU: CEOX Gold 6146 × 2, RAM: 192 GB, GPU RTX 2080Ti × 6, video memory: 66 GB, OS: Ubuntu 20.04, and Pytorch version: 1.8.0. In the supervised training, the optimizer adopts Adadelta. The learning rate scheduling is the cosine annealing learning rate, and the initial is 0.001. The epoch is set as 100. During the unsupervised iterative learning, the weights except the final convolution layers are frozen. It should be noted that the gradient backpropagation of loss function Equation (3) cannot directly adopt the loss back in PyTorch if the LSD is implemented by numpy. In our code, the loss gradient is computed by the custom function. And the parameters are set as maximum iteration $N=50$ and the threshold $\theta =5$.

5.3. The Dehazing Experiment on Synthetic Foggy Image Dataset of SOTS

The SOTS dataset contains synthetically generated haze effects applied to both indoor and outdoor scenes. The image dehazing results for indoor images are shown in Figure 7.

As illustrated in Figure 7, all the methods demonstrate effectiveness in image dehazing. However, the restored image obtained by DCP exhibits an overall dark appearance and generates noticeable artifacts around object boundaries. In contrast, FFANet, GCANet, GDNet, MSBDN, and our proposed method produce significantly milder artifacts. The primary differences among these algorithms lie in the finer details, as evidenced by the close-up comparisons in Figure 8.

As shown in Figure 8, the white wall region, marked with a red box, exhibits noticeable dark blocks in the dehazed results of GDNet, GCANet, and FFANet. Although these artifacts are significantly reduced in MSBDN, the brightness is elevated slightly. Therefore, our method achieves the best dehazing performance.

The dehazing results for outdoor images from the SOTS dataset are presented in Figure 9. It is obvious that the DCP fails to restore the sky region naturally, introducing noticeable artifacts. The details comparative analysis of the other algorithms are provided in Figure 10.

In the dehazed details, FFANet exhibits conspicuous dark block artifacts. Notably, both the tree (Region A) and stormwater drainage grate (Region B) are inadequately restored in most methods. While GDNet and our algorithm demonstrate superior performance for the tree reconstruction, GDNet still presents a dark region near the drainage grate. In contrast, our algorithm achieves optimal restoration for both structural elements, demonstrating robust dehazing capability.

The quantitative results of PSNR and SSIM about the dehazing effectiveness on SOTS are reported in

Table 1. PSNR and SSIM of dehazing image on SOTS.

Algorithm	Indoor		Outdoor
	PSNR(dB)↑	SSIM↑	PSNR (dB)↑	SSIM↑
DCP	19.31	0.860	15.54	0.826
FFANet	30.12	0.968	19.55	0.804
GCANet	29.78	0.964	21.67	0.880
GDNet	29.22	0.975	17.70	0.807
MSBDN	28.29	0.966	18.25	0.809
Our	31.01	0.973	21.74	0.881

. Our algorithm obtains the best PSNR both in indoor and outdoor dehazing images and achieves the best and second-best SSIM in the outdoor and indoor image dehazing, respectively.

The Dehazing Experiment on Nature Foggy Image Datasets

(1) The experiment on the homogeneous hazy image

For the natural hazy image datasets I-Haze and O-Haze, the fog in the two datasets could be considered as homogeneous because the fog concentration remains minimal fluctuation.

The comparative experiment results are presented in Figure 11 and Figure 12 respectively. As we can find, FFANet, GCAet, MSBDN, and our method demonstrate effective dehazing performance. However, in Figure 11, MSBDN and FFANet exhibit slight darkened patches in the white background regions. In Figure 12, the four algorithms show no significant differences in dehazing performance through subjective evaluation.

Quantitative evaluations further confirm the superiority of our approach. As shown in

Table 2. PSNR and SSIM of dehazing image on natural foggy image datasets.

Algorithms	I-Haze		O-Haze		NH-Haze
	PSNR (dB)↑	SSIM↑	PSNR (dB)↑	SSIM↑	PSNR (dB)↑	SSIM↑
DCP	13.21	0.622	15.32	0.581	13.29	0.577
FFANet	21.28	0.842	24.69	0.759	20.37	0.747
GCANet	21.45	0.841	24.32	0.759	19.95	0.734
GDNet	20.37	0.824	23.08	0.745	20.93	0.752
MSBDN	22.07	0.860	24.21	0.764	20.07	0.754
Our	23.13	0.863	25.80	0.774	21.55	0.765

, our method achieves the highest PSNR and SSIM scores on both I-Haze and O-Haze datasets, outperforming all comparative algorithms.

(2) The experiment on the non-homogeneous hazy image

Significant variations in dehazing performance are observed on the NH-Haze dataset, where the non-homogeneous fog distribution exhibits random spatial patterns and uneven concentration levels. As demonstrated in Figure 13, only FFANet and our proposed method achieve satisfactory dehazing results.

The detailed comparison between these two approaches is presented in Figure 14. Our method demonstrates superior performance in both structural preservation (particularly for fine leaf details) and color fidelity. Quantitative evaluations further confirm this advantage, with our method achieving state-of-the-art metrics of 21.55 dB PSNR and 0.765 SSIM, outperforming all comparative methods.

5.4. The Dehazing Experiment on HPTL Images

The dehazing knowledge of the neural network is transferred by the proposed unsupervised iterative learning. As demonstrated by the training loss convergence shown in Figure 15, the LSD loss can converge to the optimal station. It indicates that the final convolution layers undergo gradual fine-tuning until the dehazing knowledge transfer is completed.

The lines detected by LSD are illustrated in Figure 16, where the “NN Output” image denotes the dehazed result of the neural network, and the “Transferred” means the output of unsupervised iterative learning. It is obvious that some missed line segments can be identified during iterations and the fog covering HPTL has been dehazed.

To comprehensively evaluate the proposed method’s performance, we conducted extensive experiments on the natural HPTL dataset containing images with varying fog densities (light, moderate, and dense). We compared our approach with five dehazing algorithms: DCP, FFANet, GCANet, GDNet, and MSBDN. All methods were applied directly to the hazy images under identical conditions for fair comparison. The experimental results are presented in Figure 17, Figure 18 and Figure 19, which demonstrate the comparative performance across different fog concentrations.

As presented in the results, DCP fails to restore the sky regions, exhibiting significant artifacts. The dark patches can be found in GDNet and FFANet, as shown in Figure 16 and Figure 17. Among comparative methods, GCANet, MSBDN, and our algorithm demonstrated superior performance on subjective perception. The GCANet results have a yellow color cast. While the unsupervised iterative learning method demonstrates superior capability in preserving and enhancing linear structures, maintaining natural color balance, and achieving visually pleasing results without common artifacts.

6. Conclusions

This paper has presented a FIF-RSCT-Net for HPTL image dehazing under the hybrid supervised and unsupervised iterative learning approach. In the network, the novelty SCFI, FIF, and RSCT modules have been proposed according to the HPTL image characteristics. It has learned the generalized dehazing knowledge that can be more efficiently transferred to the HPTL image domain. In the unsupervised iterative learning, the LSD loss function has been defined to guide the dehazing knowledge transfer, by which the power transmission line segments can be restored effectively.

Comparative experiments have validated the effectiveness of our approach on I-Haze, O-Haze, NH-Haze, SOTS, and HPTL image datasets. The proposed method achieves the best dehazing performance on supervised learning and knowledge transfer. In the non-homogeneous haze datasets, the method achieves PSNR scores of 23.13 dB, 25.80 dB, and 21.55 dB, alongside SSIM values of 0.863, 0.774, and 0.765 on I-Haze, O-Haze, and NH-Haze, respectively.

Future work will incorporate the specific features of HPTL images, such as local topological and global consistency, into the unsupervised iterative learning stage to enhance the naturalness of restored images.