A power transmission line remote sensing image defogging method and device

Abstract

The application provides a power transmission line remote sensing image defogging method and device. The method comprises the following steps: obtaining a training data set, constructing an unsupervised generative adversarial network, performing cyclic adversarial training on the unsupervised generative adversarial network, calculating a cyclic consistency loss and an adversarial loss to update network parameters of a first generator, a second generator, a first discriminator and a second discriminator; iteratively performing until the network model converges, and obtaining a training complete defogging model; inputting a foggy power transmission line remote sensing image to be defogged into the first generator in the training complete defogging model, and outputting a final target power transmission line remote sensing image without fog. The application can not only effectively remove the fog of the power transmission line remote sensing image, improve the definition and visual effect of the power transmission line remote sensing image, but also better restore the power transmission line body and the surrounding environment target, better retain the detail features of the power transmission line body, and provide high-quality images for subsequent target detection.

Description

Technical Field

[0001] This application relates to the fields of computer vision and image processing technology, and in particular to a method and apparatus for dehazing remote sensing images of power transmission lines. Background Technology

[0002] The safe operation of power grids is crucial to national economic and social stability. Currently, satellite remote sensing technology has become an important means of power transmission line inspection due to its advantages such as wide coverage, lack of terrain limitations, and stable data update cycles. However, the quality of remote sensing images acquired in smoggy weather is severely degraded, with low image contrast and blurred details, directly affecting the accuracy of subsequent AI-based automatic identification of defects in power transmission equipment (such as insulators and clamps) and potential hazards in the surrounding environment (such as construction machinery and illegal buildings).

[0003] Existing deep learning-based image dehazing methods are mostly focused on natural scene images, with the core objective of improving the overall contrast and color fidelity of the image. These methods are typically based on atmospheric scattering models, using deep learning networks to optimize the global statistical properties of the image. However, when applied to remote sensing images of power transmission lines, they have significant shortcomings: the power transmission equipment itself usually appears as a small, localized target in the image, and its color features are not distinct. Existing general dehazing methods, while removing global haze, easily weaken or blur the detailed information of these key targets. As a result, although the visual effect of the dehazed image is improved, the clarity of the power transmission line itself decreases, failing to meet the needs of subsequent accurate detection.

[0004] Therefore, there is an urgent need for a professional defogging method that can effectively remove fog from remote sensing images of power transmission lines while preserving and enhancing the details of the power transmission equipment itself. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method and apparatus for defogging remote sensing images of power transmission lines, which can effectively remove fog and haze while significantly improving the clarity of details of the power transmission line itself and its surrounding environment.

[0006] A first aspect of this application provides a method for dehazing remote sensing images of power transmission lines, the method comprising: Obtain a training dataset, which includes a remote sensing image sample set of transmission lines in fog and a remote sensing image sample set of transmission lines without fog; Construct an unsupervised generative adversarial network, wherein the unsupervised generative adversarial network includes a first generator and a second generator with the same structure, and a first discriminator and a second discriminator with the same structure; The unsupervised generative adversarial network is subjected to cyclic adversarial training, the training process of which includes: First training loop: Input the remote sensing image of the foggy transmission line into the first generator to obtain the first generated fog-free image, and input the first generated fog-free image into the first discriminator for authenticity judgment; at the same time, input the first generated fog-free image into the second generator to obtain the first reconstructed foggy image; Second training cycle: Input the fog-free transmission line remote sensing image into the second generator to obtain a second generated foggy image, and input the second generated foggy image into the second discriminator for authenticity judgment; at the same time, input the second generated foggy image into the first generator to obtain a second reconstructed fog-free image; Based on the first generated hazy image, the first reconstructed hazy image, the second generated hazy image, the second reconstructed hazy image, and the discrimination results of the first discriminator and the second discriminator, the cycle consistency loss and adversarial loss are calculated to update the network parameters of the first generator, the second generator, the first discriminator, and the second discriminator; the process is iteratively executed until the network model converges, and a fully trained dehazing model is obtained. The remote sensing image of the foggy transmission line to be defogged is input into the first generator in the fully trained defogging model, and the final target fog-free remote sensing image of the transmission line is output.

[0007] A second aspect of this application provides an unsupervised generative adversarial network for performing the method described in the first aspect, including a first generator and a second generator with identical structures, and a first discriminator and a second discriminator with identical structures. Both the first generator and the second generator include an encoder, a semantic feature extraction module, and a decoder connected in sequence; The encoder is configured to downsample the input image and extract multi-scale features; The semantic feature extraction module is configured to perform semantic enhancement and feature fusion on the features output by the encoder. The semantic feature extraction module also integrates a high-frequency feature spatial attention module and a multi-scale dilated convolutional spatial attention module. The decoder is configured to upsample the features output by the semantic feature extraction module to reconstruct the output image; Both the first discriminator and the second discriminator are configured to extract local feature maps of the input image for authenticity determination.

[0008] A third aspect of this application provides a dehazing device for remote sensing images of power transmission lines, comprising: The training dataset acquisition module is used to acquire the training dataset, which includes a remote sensing image sample set of transmission lines with fog and a remote sensing image sample set of transmission lines without fog. An unsupervised generative adversarial network construction module is used to construct an unsupervised generative adversarial network, wherein the unsupervised generative adversarial network includes a first generator and a second generator with the same structure, and a first discriminator and a second discriminator with the same structure; The training module is used to perform cyclic adversarial training on the unsupervised generative adversarial network. The training process includes: First training loop: Input the remote sensing image of the foggy transmission line into the first generator to obtain the first generated fog-free image, and input the first generated fog-free image into the first discriminator for authenticity judgment; at the same time, input the first generated fog-free image into the second generator to obtain the first reconstructed foggy image; Second training cycle: Input the fog-free transmission line remote sensing image into the second generator to obtain a second generated foggy image, and input the second generated foggy image into the second discriminator for authenticity judgment; at the same time, input the second generated foggy image into the first generator to obtain a second reconstructed fog-free image; The dehazing model acquisition module is used to calculate the cycle consistency loss and adversarial loss based on the first generated hazy image, the first reconstructed hazy image, the second generated hazy image, the second reconstructed hazy image, and the discrimination results of the first discriminator and the second discriminator, so as to update the network parameters of the first generator, the second generator, the first discriminator, and the second discriminator; iteratively executes until the network model converges to obtain a fully trained dehazing model; The target fog-free transmission line remote sensing image output module is used to input the foggy transmission line remote sensing image to be defogged into the first generator in the fully trained defogging model, and output the final target fog-free transmission line remote sensing image.

[0009] A fourth aspect of this application provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0010] A fifth aspect of the embodiments of this application provides a readable storage medium storing a computer program that, when executed by a processor, implements the steps of the above-described method.

[0011] The beneficial effects of this application's embodiments compared to existing technologies are as follows: This application's embodiments are highly targeted: the network structure is specifically designed to address the problem of easily lost details of small local targets (transmission equipment) in remote sensing images of power transmission lines. By integrating multiple attention mechanisms, it can effectively preserve and enhance the detailed features of the line itself while dehazing. It adopts an unsupervised recurrent adversarial training paradigm, eliminating the need for paired foggy-free image data for training, reducing the difficulty and cost of data acquisition, and enhancing the practicality and applicability of the method. It can not only effectively remove haze from remote sensing images of power transmission lines, improving the clarity and visual effect of the images, but also better recover the power transmission line itself and surrounding environmental targets, better preserving the detailed features of the power transmission line itself, and providing high-quality images for subsequent target detection. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a schematic flowchart of a method for dehazing remote sensing images of power transmission lines provided in an embodiment of this application; Figure 2 This is a network structure diagram of the first generator in an embodiment of the present invention;

[0014] Figure 3 This is a structural diagram of the progressive feature extraction module in an embodiment of the present invention;

[0015] Figure 4 This is a structural diagram of the semantic feature extraction module in an embodiment of the present invention;

[0016] Figure 5 This is a network structure diagram of the first discriminator in an embodiment of the present invention;

[0017] Figure 6 This is a schematic diagram of a defogging device for remote sensing images of power transmission lines provided in an embodiment of this application;

[0018] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0019] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0020] The following will describe in detail, with reference to the accompanying drawings, a method and apparatus for defogging remote sensing images of power transmission lines according to embodiments of this application.

[0021] Figure 1 This is a flowchart illustrating a method for dehazing remote sensing images of power transmission lines provided in an embodiment of this application. Figure 1 As shown, the method for dehazing remote sensing images of the transmission line includes: S101, Obtain the training dataset, which includes a remote sensing image sample set of transmission lines with fog and a remote sensing image sample set of transmission lines without fog. S102, Construct an unsupervised generative adversarial network, which includes a first generator and a second generator with the same structure, and a first discriminator and a second discriminator with the same structure; S103, perform cyclic adversarial training on the unsupervised generative adversarial network; S104: Based on the first generated hazy-free image, the first reconstructed hazy image, the second generated hazy image, the second reconstructed hazy-free image, and the discrimination results of the first discriminator and the second discriminator, calculate the cycle consistency loss and adversarial loss to update the network parameters of the first generator, the second generator, the first discriminator, and the second discriminator; iterate until the network model converges to obtain a fully trained dehazing model. S105: Input the remote sensing image of the foggy transmission line to be defogged into the first generator in the well-trained defogging model, and output the final target fog-free remote sensing image of the transmission line.

[0022] Specifically, step 101 is the cornerstone of building the intelligent dehazing model, and its core task is to prepare the data raw materials required for model learning. The training dataset is a large amount of classified data used to train the neural network model. In this embodiment, the dataset is specifically divided into two subsets: the "foggy transmission line remote sensing image sample set" refers to a collection of blurry images of power facilities such as transmission lines, towers, and insulators taken by satellites or drones under adverse weather conditions such as fog, haze, and smog; while the "fog-free transmission line remote sensing image sample set" is a collection of clear images of the same area acquired under clear weather conditions with excellent visibility. These represent the ideal visual effect that dehazing processing is expected to achieve. The reason for needing these two types of data is that this invention adopts an advanced paradigm called "unsupervised learning." Compared with traditional methods that require tens of thousands of strictly paired images of "the same scene, the same angle, only different weather," this invention does not require such demanding and costly data. In reality, it is almost impossible to find a fog-free image taken at the exact same position and angle for every foggy image. In this way, the difficulty and cost of data acquisition are significantly reduced. For example, historical meteorological data can be easily used to select all power transmission line images taken in North China during hazy days over the past five years, forming a "foggy set"; simultaneously, clear images taken during the same period in different regions such as East and South China under clear weather conditions can be formed a "foggy set". By learning from these two large, unpaired sets, the model can autonomously understand the low contrast and grayish tones commonly found in "foggy images", as well as the rich detail and vivid colors that "foggy images" should possess. This lays a solid data foundation for subsequently training a powerful dehazing model with strong generalization ability that does not require precisely paired data.

[0023] Furthermore, the unsupervised generative adversarial network in step 102 is a complex machine learning architecture consisting of a "generator" and a "discriminator." In this embodiment, two such pairs are constructed. The first generator learns to restore a hazy input image to a clear one; the corresponding first discriminator determines whether the received clear image is the "true original" (from the hazy dataset) or a "fake created by the first generator." To form a complete training loop, a functionally symmetrical second generator and second discriminator are also set up. The former learns to "add" a haze effect to the clear image, while the latter is responsible for identifying the authenticity of the hazy image. This complex adversarial network structure is mainly used to address the "oversmoothing" and "loss of detail" problems that are easily caused by traditional dehazing methods. If there is only a simple generative network, it is easy to learn a simple strategy, namely, to generate an image that looks clear overall, but all key objects (such as thin power lines and complex insulators) become blurry. Because from the perspective of overall pixel error, such an output may not be bad. However, with the introduction of a discriminator, the generator must continuously improve its generation technology, striving to produce clear images with realistic details that even the discriminator cannot distinguish; meanwhile, the discriminator is also constantly upgrading its "fake detection" capabilities. This forces the generator to focus on restoring those subtle and realistic structures in the image, thus effectively solving the core problem of insufficient preservation of details in the transmission line itself.

[0024] Furthermore, the recurrent adversarial training includes a first training cycle: inputting a remote sensing image of a foggy transmission line into a first generator to obtain a first generated fog-free image, and inputting the first generated fog-free image into a first discriminator for authenticity judgment; simultaneously, inputting the first generated fog-free image into a second generator to obtain a first reconstructed foggy image; Second training loop: Input the fog-free transmission line remote sensing image into the second generator to obtain the second generated foggy image, and input the second generated foggy image into the second discriminator for authenticity judgment; at the same time, input the second generated foggy image into the first generator to obtain the second reconstructed fog-free image.

[0025] This embodiment introduces the concept of "cycle consistency" to ensure the accuracy of image transformation and the fidelity of content. The training process comprises two closed-loop cycles. The first is a "foggy → fog-free → foggy" cycle: a real foggy image is input into the first generator to obtain a "first generated fog-free image" (i.e., an imagined clear image); then, this generated clear image is fed into the second generator to attempt to reconstruct the fog effect, resulting in a "first reconstructed foggy image." Ideally, this reconstructed image should be very similar to the original input real foggy image. The second is a reverse "foggy → foggy → fog-free" cycle: a real fog-free image is input into the second generator to obtain a "second generated foggy image"; then, it is input into the first generator for defogging to obtain a "second reconstructed fog-free image," which should also be largely consistent with the original fog-free image. This step primarily addresses the "pattern collapse" and "arbitrary translation" problems that may occur during generative adversarial network training. Without this cycle constraint, the first generator might learn a "cheating" behavior—outputting the same seemingly clear but irrelevant image regardless of the input foggy image. Cyclic consistency enforces that an image, after a complete "dehazing-re-fogging" cycle, must be able to "return to its original state" as much as possible. This constraint ensures that the generator cannot arbitrarily change the image's basic content (such as the shape of the transmission tower, the direction of the power lines, or the location of buildings) when removing fog; it must retain all key structural information and only remove visual interference caused by fog. If the first generator haphazardly performs dehazing, for example, by adding complex structures to the tower out of thin air, the second generator will be unable to accurately restore the original simple outline when re-fogging. The cycle consistency loss will immediately detect this inconsistency and penalize the first generator. Through this strict constraint, the final dehazing model achieves improved clarity while maintaining extremely high content fidelity, making it ideal for engineering applications requiring precise dimensional measurements or defect identification.

[0026] Further, step 104 evaluates the model's performance through quantification and guides its improvement. "Loss" is a numerical value used to measure the difference between the model's current output and the ideal state. In this invention, two types of loss are mainly calculated: First, "adversarial loss," which stems from the game between the generator and the discriminator. For the first generator, its adversarial loss measures the degree to which the image it generates is misclassified as a "realistic and clear image" by the first discriminator; the generator wants this loss to be as small as possible (successfully fooling the discriminator), while the discriminator wants to accurately distinguish between them, resulting in a large loss. Second, "cyclic consistency loss," which directly quantifies the difference between the images before and after the cyclic process mentioned in S103, for example, calculating the average absolute error of all pixel values ​​between the "original foggy image" and the "first reconstructed foggy image." Using adversarial loss alone, while encouraging the generator to output realistic images, may lead to outputs that do not match the input (i.e., "indistinguishable from reality" but not "fidelity-assured"). Using pixel-level cyclic loss alone, on the other hand, can easily lead to results that are too smooth and lack texture detail. Adversarial loss drives the model to generate detailed and visually stunning images; cycle-consistent loss ensures that the model does not arbitrarily alter the content of the original image in pursuit of realism. When the model's output no longer shows significant improvement over a long period, it can be considered that the model has "converged," and a reliable dehazing model has been trained.

[0027] Furthermore, the "fully trained defogging model" specifically refers to the first generator in S104 that has been fully trained, has stable performance, and meets practical requirements. At this point, its parameters are fixed, making it a dedicated system for defogging transmission line images. Its task is very clear: to receive a new, unprocessed "remote sensing image of the transmission line to be defogged" (e.g., a real-time downlink image of a hazy area from a meteorological satellite) and output a clear "remote sensing image of the target fog-free transmission line." After the model is deployed, all images of transmission line areas marked as "affected by fog" transmitted via satellite links are automatically fed into this fully trained first generator for processing. For example, an image that was originally hazy, with only faint outlines of towers visible, can, after model processing, clearly show the main structure of the towers, the integrity of the insulator strings, and even allow for preliminary observation of any hanging foreign objects on the conductors. Ultimately, this provides strong technical support for the safe and stable operation of the power grid.

[0028] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0029] This application also provides an unsupervised generative adversarial network (GAN). The GAN includes a first generator and a second generator with identical structures, as well as a first discriminator and a second discriminator with identical structures. Both the first and second generators include an encoder, a semantic feature extraction module, and a decoder connected in sequence. The encoder is configured to downsample the input image and extract multi-scale features. The semantic feature extraction module is configured to perform semantic enhancement and feature fusion on the features output by the encoder; the semantic feature extraction module also integrates a high-frequency feature spatial attention module and a multi-scale dilated convolutional spatial attention module. The decoder is configured to upsample the features output by the semantic feature extraction module to reconstruct the output image. Both the first and second discriminators are configured to extract local feature maps of the input image for realism judgment.

[0030] The encoder consists of a primary feature extraction section and three cascaded high-level feature extraction sections. The primary feature extraction section consists of a 3×3 convolutional layer, a batch normalization layer, and a LeakyReLU activation function layer connected in sequence. Each high-level feature extraction section consists of a progressive feature extraction module, a 4×4 convolutional layer with a stride of 2, a batch normalization layer, and a LeakyReLU activation function layer connected in sequence.

[0031] The progressive feature extraction module includes a first convolutional branch, a feature fusion node, a second convolutional branch, and a hybrid attention module connected in sequence. The first convolutional branch includes a parallel 7×7 convolutional layer and a 5×5 convolutional layer, used to extract global information of the input features. The feature fusion node is used to add the output of the first convolutional branch to the input of the progressive feature extraction module element-wise. The second convolutional branch includes a sequentially connected 3×3 convolutional layer and a 1×1 convolutional layer, used to extract detailed information from the fused features. The hybrid attention module, connected after the second convolutional branch, is used to enhance key information in the processed feature map.

[0032] The semantic feature extraction module includes a semantic preliminary extraction unit, a parallel downsampled high-frequency feature extraction branch and a full-frequency feature extraction branch, and a feature fusion unit. The semantic preliminary extraction unit is used to perform preliminary semantic extraction on the input features. The semantic preliminary extraction unit includes two sequentially connected 3×3 convolutional layers, each followed by a batch normalization layer and a LeakyReLU activation function layer. The downsampled high-frequency feature extraction branch receives the output of the semantic preliminary extraction unit and processes it through the built-in high-frequency feature spatial attention module to extract and enhance high-frequency feature information. The full-frequency feature extraction branch receives the output of the semantic preliminary extraction unit and processes it through the built-in multi-scale dilated convolutional spatial attention module to extract multi-scale deep semantic information. The feature fusion unit is used to concatenate and fuse the input features of the semantic feature extraction module, the upsampled features of the output features of the downsampled high-frequency feature extraction branch, and the output features of the full-frequency feature extraction branch.

[0033] Both the first and second discriminators include: a multi-scale shallow feature extraction module, comprising a 1×1 convolutional layer and a LeakyReLU activation function layer for channel upscaling of the input image, as well as parallel 1×1, 3×3, 5×5 and 7×7 convolutional layers, each followed by a batch normalization layer and a LeakyReLU activation function layer, to extract multi-scale shallow features; a feature fusion and downsampling module, which concatenates the multi-scale shallow features and then passes them sequentially through a 3×3 convolutional layer and three 4×4 convolutional layers with a stride of 2 for equalization, extraction and downsampling to obtain a small-scale feature map; and an output layer, comprising a 1×1 convolutional layer, which adjusts the channels of the small-scale feature map and outputs a local feature map.

[0034] The decoder consists of three cascaded upsampling feature extraction sections and one image generation section. Each upsampling feature extraction section includes a 4×4 deconvolutional layer, a batch normalization layer, a LeakyReLU activation function layer, and a progressive feature extraction module connected in sequence. The image generation section includes a 3×3 convolutional layer, a batch normalization layer, and a Tanh activation function layer connected in sequence, used to generate an output image with the same size as the input image.

[0035] The unsupervised generative adversarial network provided in this embodiment has the following structure for its core first generator: Figure 1 As shown, it consists of an encoder, a semantic feature extraction module, and a decoder connected in sequence.

[0036] Encoders are used for feature extraction and downsampling. For example... Figure 2As shown, it comprises a primary feature extraction section and three advanced feature extraction sections. The primary feature extraction section consists of a 3×3 convolutional layer, a batch normalization layer, and a LeakyReLU activation function layer connected sequentially. Each advanced feature extraction section is connected sequentially to a progressive feature extraction module, a 4×4 convolutional layer with a stride of 2, a batch normalization layer, and a LeakyReLU activation function layer. Through these three downsampling layers, the feature map output by the encoder is scaled to 1 / 8 of the input image.

[0037] The structure of the progressive feature extraction module is as follows: Figure 3 As shown, it first extracts global information from the input features through a first convolutional branch consisting of parallel 7×7 and 5×5 convolutional layers. Then, through a feature fusion node (i.e., a residual connection), the output of this branch is added element-wise to the original input of the module to reduce the loss of shallow information. The added features then enter a second convolutional branch consisting of sequentially connected 3×3 and 1×1 convolutional layers to extract finer details. Finally, a hybrid attention module enhances the features with key information.

[0038] The semantic feature extraction module is one of the core innovations of the generator, and its structure is as follows: Figure 4 As shown, this module first processes the deep features output by the encoder through a semantic preliminary extraction unit (composed of two 3×3 convolutional layers). Then, its output is simultaneously fed into two parallel branches: a downsampling high-frequency feature extraction branch and a full-frequency feature extraction branch. The downsampling high-frequency feature extraction branch first performs downsampling, then utilizes its built-in high-frequency feature spatial attention module to focus on and enhance high-frequency details (such as edges and textures) in the image. The full-frequency feature extraction branch utilizes its built-in multi-scale dilated convolutional spatial attention module to extract and enhance key deep semantic information under different receptive fields. Finally, in the feature fusion unit, the module's original input (via skip connections), the upsampled high-frequency branch output features, and the full-frequency branch output features are concatenated and fused, achieving multi-scale, comprehensive fusion of shallow details, high-frequency information, and deep semantics.

[0039] The decoder is used for feature upsampling and image reconstruction. For example... Figure 2 As shown, it comprises three cascaded upsampling feature extraction sections. Each section contains a 4×4 deconvolutional layer (used to double the feature map size), a batch normalization layer, a LeakyReLU activation function layer, and the aforementioned progressive feature extraction module. After three upsampling layers, the feature map size is restored to the input image size. Finally, an image generation section, consisting of a 3×3 convolutional layer, a batch normalization layer, and a Tanh activation function layer, outputs the final dehazed (or hazy) image.

[0040] The structure of the discriminator is as follows Figure 5 As shown, unlike traditional discriminators that output a single probability of true or false data, this algorithm aims to output a local feature map to improve the ability to discriminate local details in an image. It first uses a 1×1 convolutional layer to increase the channel dimension, then uses four parallel convolutional layers (1×1, 3×3, 5×5, and 7×7) to extract shallow features at different scales. After concatenating these multi-scale features, they are sequentially passed through a 3×3 convolutional layer and three 4×4 convolutional layers with a stride of 2 for feature fusion and downsampling. Finally, a 1×1 convolutional layer outputs a feature map containing rich local information.

[0041] During training, a cycle-consistency training strategy is employed. Through two training cycles, and by calculating adversarial loss and cycle-consistency loss, the training of the network is jointly constrained, enabling the first generator to learn the complex mapping relationship from foggy images to high-quality fog-free images, while preserving the details of the transmission lines well.

[0042] Table 1 shows the quantitative comparison results between the method of this invention and other advanced dehazing algorithms. The method of this invention achieves the best values ​​in PSNR, SSIM and MSE, which proves its effectiveness.

[0043]

[0044] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0045] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0046] Figure 6 This is a schematic diagram of a defogging device for remote sensing images of power transmission lines provided in an embodiment of this application. Figure 6 As shown, the dehazing device for remote sensing images of the transmission line includes: The training dataset acquisition module 601 is configured to acquire a training dataset, which includes a remote sensing image sample set of transmission lines with fog and a remote sensing image sample set of transmission lines without fog. An unsupervised generative adversarial network construction module 602 is configured to construct an unsupervised generative adversarial network, which includes a first generator and a second generator with the same structure, and a first discriminator and a second discriminator with the same structure. Training module 603 is configured to perform recurrent adversarial training on an unsupervised generative adversarial network. The training process includes: First training loop: Input the remote sensing image of the foggy transmission line into the first generator to obtain the first generated fog-free image, and input the first generated fog-free image into the first discriminator for authenticity judgment; at the same time, input the first generated fog-free image into the second generator to obtain the first reconstructed foggy image; Second training loop: Input the fog-free transmission line remote sensing image into the second generator to obtain the second generated foggy image, and input the second generated foggy image into the second discriminator for authenticity judgment; at the same time, input the second generated foggy image into the first generator to obtain the second reconstructed fog-free image; The dehazing model acquisition module 604 is configured to calculate the cycle consistency loss and adversarial loss based on the first generated hazy image, the first reconstructed hazy image, the second generated hazy image, the second reconstructed hazy image, and the discrimination results of the first discriminator and the second discriminator, so as to update the network parameters of the first generator, the second generator, the first discriminator, and the second discriminator; iteratively execute until the network model converges to obtain a fully trained dehazing model; The target fog-free transmission line remote sensing image output module 605 is configured to input the foggy transmission line remote sensing image to be defogged into the first generator in the fully trained defogging model, and output the final target fog-free transmission line remote sensing image.

[0047] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0048] Figure 7 This is a schematic diagram of the electronic device 7 provided in an embodiment of this application. Figure 7 As shown, the electronic device 7 of this embodiment includes a processor 701, a memory 702, and a computer program 703 stored in the memory 702 and executable on the processor 701. When the processor 701 executes the computer program 703, it implements the steps in the various method embodiments described above. Alternatively, when the processor 701 executes the computer program 703, it implements the functions of each module / unit in the various device embodiments described above.

[0049] Electronic device 7 can be a desktop computer, laptop, handheld computer, cloud server, or other electronic device. Electronic device 7 may include, but is not limited to, processor 701 and memory 702. Those skilled in the art will understand that... Figure 7 This is merely an example of electronic device 7 and does not constitute a limitation on electronic device 7. It may include more or fewer components than shown, or different components.

[0050] Processor 701 may be a central processing unit (CPU) It can also be other general-purpose processors or digital signal processors (DSPs). Application-Specific Integrated Circuits (ASICs) Field-Programmable Gate Arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0051] The memory 702 can be an internal storage unit of the electronic device 7, such as a hard disk or memory of the electronic device 7. The memory 702 can also be an external storage device of the electronic device 7, such as a plug-in hard disk or smart memory card equipped on the electronic device 7. This includes Secure Digital (SD) cards, FlashCards, etc. The memory 702 can also include both internal storage units and external storage devices of the electronic device 7. The memory 702 is used to store computer programs and other programs and data required by the electronic device.

[0052] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0053] If integrated modules / units are implemented as software functional units and sold or used as independent products, they can be stored in a readable storage medium (e.g., a computer-readable storage medium). Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program may include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. A computer-readable storage medium may include: any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), etc. Random Access Memory (RAM) (including) electrical carrier signals, telecommunication signals, and software distribution media, etc.

[0054] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method for dehazing remote sensing images of power transmission lines, characterized in that, The method includes: Obtain a training dataset, which includes a remote sensing image sample set of transmission lines in fog and a remote sensing image sample set of transmission lines without fog; Construct an unsupervised generative adversarial network, wherein the unsupervised generative adversarial network includes a first generator and a second generator with the same structure, and a first discriminator and a second discriminator with the same structure; Both the first generator and the second generator include an encoder, a semantic feature extraction module, and a decoder connected in sequence; The encoder is configured to downsample the input image and extract multi-scale features. The encoder includes a primary feature extraction section and three cascaded high-level feature extraction sections. The primary feature extraction section includes a 3×3 convolutional layer, a batch normalization layer, and a LeakyReLU activation function layer connected in sequence. Each high-level feature extraction section includes a progressive feature extraction module, a 4×4 convolutional layer with a stride of 2, a batch normalization layer, and a LeakyReLU activation function layer connected in sequence. The progressive feature extraction module includes a first convolutional branch, a feature fusion branch, and a feature fusion branch connected in sequence. The system comprises a node, a second convolutional branch, and a hybrid attention module. The first convolutional branch includes parallel 7×7 and 5×5 convolutional layers for extracting global information from the input features. The feature fusion node is used to element-wise add the output of the first convolutional branch to the input of the progressive feature extraction module. The second convolutional branch includes sequentially connected 3×3 and 1×1 convolutional layers for extracting detailed information from the fused features. The hybrid attention module, connected after the second convolutional branch, is used to enhance key information in the processed feature map. The semantic feature extraction module is configured to perform semantic enhancement and feature fusion on the features output by the encoder. The semantic feature extraction module also integrates a high-frequency feature spatial attention module and a multi-scale dilated convolutional spatial attention module. The semantic feature extraction module includes a preliminary semantic extraction unit, parallel-connected downsampling high-frequency feature extraction branches and full-frequency feature extraction branches, and a feature fusion unit. The preliminary semantic extraction unit is used to perform preliminary semantic extraction on the input features. The preliminary semantic extraction unit includes two sequentially connected 3×3 convolutional layers, each followed by a batch normalization layer and a LeakyReLU activation function layer. The high-frequency feature extraction branch receives the output of the initial semantic extraction unit and processes it through the included high-frequency feature spatial attention module to extract and enhance high-frequency feature information. The full-frequency feature extraction branch receives the output of the initial semantic extraction unit and processes it through the included multi-scale dilated convolution spatial attention module to extract multi-scale deep semantic information. The feature fusion unit concatenates and fuses the input features of the semantic feature extraction module, the upsampled features of the output features of the downsampled high-frequency feature extraction branch, and the output features of the full-frequency feature extraction branch. The decoder is configured to upsample the features output by the semantic feature extraction module to reconstruct the output image; the decoder includes three cascaded upsampled feature extraction parts and an image generation part; each upsampled feature extraction part includes a 4×4 deconvolution layer, a batch normalization layer, a LeakyReLU activation function layer and a progressive feature extraction module connected in sequence; the image generation part includes a 3×3 convolution layer, a batch normalization layer and a Tanh activation function layer connected in sequence, used to generate an output image with the same size as the input image; Both the first and second discriminators are configured to extract local feature maps of the input image for authenticity determination. Each discriminator includes: a multi-scale shallow feature extraction module, comprising a 1×1 convolutional layer and a LeakyReLU activation function layer for channel upscaling of the input image, and parallel 1×1, 3×3, 5×5, and 7×7 convolutional layers, each followed by a batch normalization layer and a LeakyReLU activation function layer, to extract multi-scale shallow features; a feature fusion and downsampling module, used to concatenate the multi-scale shallow features and then pass them sequentially through a 3×3 convolutional layer and three 4×4 convolutional layers with a stride of 2 for equalization, extraction, and downsampling to obtain a small-scale feature map; and an output layer, including a 1×1 convolutional layer, used to adjust the channels of the small-scale feature map and output a local feature map. The unsupervised generative adversarial network is subjected to cyclic adversarial training, the training process of which includes: First training loop: Input the remote sensing image of the foggy transmission line into the first generator to obtain the first generated fog-free image, and input the first generated fog-free image into the first discriminator for authenticity judgment; at the same time, input the first generated fog-free image into the second generator to obtain the first reconstructed foggy image; Second training cycle: Input the fog-free transmission line remote sensing image into the second generator to obtain a second generated foggy image, and input the second generated foggy image into the second discriminator for authenticity judgment; at the same time, input the second generated foggy image into the first generator to obtain a second reconstructed fog-free image; Based on the first generated hazy image, the first reconstructed hazy image, the second generated hazy image, the second reconstructed hazy image, and the discrimination results of the first discriminator and the second discriminator, the cycle consistency loss and adversarial loss are calculated to update the network parameters of the first generator, the second generator, the first discriminator, and the second discriminator; the process is iteratively executed until the network model converges, and a fully trained dehazing model is obtained. The remote sensing image of the foggy transmission line to be defogged is input into the first generator in the fully trained defogging model, and the final target fog-free remote sensing image of the transmission line is output.

2. A device for dehazing remote sensing images of power transmission lines, characterized in that, The device is used to perform the dehazing method for remote sensing images of transmission lines as described in claim 1, including: The training dataset acquisition module is used to acquire the training dataset, which includes a remote sensing image sample set of transmission lines with fog and a remote sensing image sample set of transmission lines without fog. An unsupervised generative adversarial network construction module is used to construct an unsupervised generative adversarial network, wherein the unsupervised generative adversarial network includes a first generator and a second generator with the same structure, and a first discriminator and a second discriminator with the same structure; The training module is used to perform cyclic adversarial training on the unsupervised generative adversarial network. The training process includes: First training loop: Input the remote sensing image of the foggy transmission line into the first generator to obtain the first generated fog-free image, and input the first generated fog-free image into the first discriminator for authenticity judgment; at the same time, input the first generated fog-free image into the second generator to obtain the first reconstructed foggy image; Second training cycle: Input the fog-free transmission line remote sensing image into the second generator to obtain a second generated foggy image, and input the second generated foggy image into the second discriminator for authenticity judgment; at the same time, input the second generated foggy image into the first generator to obtain a second reconstructed fog-free image; The dehazing model acquisition module is used to calculate the cycle consistency loss and adversarial loss based on the first generated hazy image, the first reconstructed hazy image, the second generated hazy image, the second reconstructed hazy image, and the discrimination results of the first discriminator and the second discriminator, so as to update the network parameters of the first generator, the second generator, the first discriminator, and the second discriminator; iteratively executes until the network model converges to obtain a fully trained dehazing model; The target fog-free transmission line remote sensing image output module is used to input the foggy transmission line remote sensing image to be defogged into the first generator in the fully trained defogging model, and output the final target fog-free transmission line remote sensing image.

3. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in claim 1.

4. A readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in claim 1.

Images