Power image denoising method and system based on attention mechanism

Abstract

The invention provides a power image denoising method and system based on an attention mechanism. The method comprises the following steps: constructing a power image denoising model based on the attention mechanism; training the electric power image denoising model by using the preprocessed data set; and processing the acquired power image data by using the trained power image denoising model to obtain a denoised power image. According to the method, on the basis of the convolutional neural network, the attention mechanism and the expansion convolution technology are combined, so that the denoising performance is improved, the model can more effectively capture feature information of different levels in the power image, and the feature information comprises fine local details and also covers the global context. The attention mechanism enables the network to dynamically adjust the attention degree of key features, suppress noise and reserve important information, thereby reducing the risk of misjudgment. And the receptive field is expanded by introducing the expansion convolution, so that the network can obtain wider context information, and the processing capability on a complex scene is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of power image processing, and particularly relates to a power image denoising method and system based on an attention mechanism. Background Art

[0002] The statements in this part merely provide background technical information related to the present invention and do not necessarily constitute prior art.

[0003] In the monitoring and maintenance of power systems, power images play a crucial role in ensuring power safety and equipment operation. However, images are usually interfered by various noises, such as light changes, sensor noises or environmental factors, resulting in a decline in image quality and affecting the judgment of equipment status and subsequent processing and analysis.

[0004] However, traditional denoising methods, although able to reduce the noise in images to a certain extent, mainly focus on the processing of local features and have many limitations. Mean filtering removes noise by calculating the average value of the neighboring pixels around each pixel in the image. However, since mean filtering assigns the same weight to all pixels, when processing the edges and details in the image, these important features will be blurred. In power images, many key information is usually located in the edge areas of the image. Mean filtering may lead to the loss of these details, thus affecting the accurate judgment of equipment status. Median filtering removes noise by replacing the value of each pixel with the median of the neighboring pixel values. This method performs well in dealing with salt-and-pepper noise and can effectively retain edge information, avoiding the blurring problem caused by mean filtering. However, median filtering also fails to effectively capture the global information in power images, so that in dealing with complex scenes, it may not be able to provide sufficient context information, resulting in the inability to effectively distinguish noise and useful signals.

[0005] In addition, the computational complexity of traditional denoising methods is relatively low and is suitable for scenarios with low real-time processing requirements. With the continuous increase in the number and types of power equipment and the diversification of monitoring requirements, traditional denoising methods cannot effectively adapt to the characteristics of new power images, resulting in poor denoising effects. Summary of the Invention

[0006] To solve the above problems, the present invention proposes a power image denoising method and system based on an attention mechanism. The present invention is based on a convolutional neural network (CNN), combines the attention mechanism and dilated convolution technology to improve the denoising performance. Through this design, the model can more effectively capture feature information at different levels in power images, including both subtle local details and global context. The attention mechanism enables the network to dynamically adjust the degree of attention to key features, suppress noise and retain important information, thereby reducing the risk of misjudgment. The introduction of dilated convolution expands the receptive field, enabling the network to obtain more extensive context information and enhancing the processing ability for complex scenarios.

[0007] According to some embodiments, the present invention adopts the following technical solutions:

[0008] A power image denoising method based on an attention mechanism, comprising the following steps:

[0009] Obtain historical power image data to form a data set, and preprocess the data set;

[0010] Construct a power image denoising model based on an attention mechanism;

[0011] Use the preprocessed data set to train the power image denoising model;

[0012] Use the trained power image denoising model to process the obtained power image data to obtain a denoised power image;

[0013] The power image denoising model includes an input layer, a feature extraction module, an attention mechanism module, a feature fusion module, and an output layer connected in sequence. The input layer is used to receive a power image, and the feature extraction module is used to extract features from the input power image and perform feature normalization. The feature extraction module includes multiple convolutional layers, and at least some convolutional layers are added with dilated convolutional layers to expand the receptive field, and some dilated convolutional layers introduce holes in the convolutional kernel; the attention mechanism module receives the processed features and dynamically adjusts the degree of attention to each feature channel, and enhances important features through a weighting mechanism; the feature fusion module is used to fuse the features processed by the attention mechanism module and generate a denoised power image through the output layer.

[0014] As an alternative implementation, the preprocessing process includes size adjustment and normalization processing.

[0015] As an alternative implementation, the preprocessing process further includes data augmentation, and the data augmentation methods include several of the following:

[0016] Randomly rotate the image to increase image samples at different angles;

[0017] Horizontally or vertically flip the image to increase sample diversity;

[0018] Randomly crop partial regions of the image to enhance the learning of local features;

[0019] Randomly scale the image within a certain range to simulate shooting effects at different distances;

[0020] Randomly adjust brightness, contrast, and saturation to increase the robustness to illumination changes;

[0021] Add random noise to the clear image to generate a new noisy image.

[0022] As an alternative implementation, the feature extraction module includes a first convolutional layer and multiple dilated convolutional layers. The first convolutional layer is used for preliminary feature extraction and adds zero-padding of 1 pixel around the input image to maintain the spatial size of the feature map;

[0023] Between every two layers of the dilated convolutional layers, the dilation rate doubles.

[0024] Furthermore, behind each convolutional layer, there is a batch normalization layer and a ReLU activation layer.

[0025] As an alternative implementation, the attention mechanism module includes two convolutional layers and a Sigmoid activation function. The first convolutional layer is used to compress the feature channels to 1 / 8 of the original size to reduce the dimension of the feature map. The second convolutional layer maps the compressed feature back to the original number of channels to ensure the retention of information. The Sigmoid activation function is used to map the output to the range of (0,1) to generate the attention weights for each channel.

[0026] As an alternative implementation, the attention mechanism module performs element-wise multiplication on the feature map and the attention weight map.

[0027] As an alternative implementation, after the attention mechanism module, there is also a convolutional layer used to map the feature map back to the original number of input channels to ensure consistent output format.

[0028] As an alternative implementation, during the training process, the mean squared error is used as the loss function to measure the difference between the denoised image and the clear image.

[0029] As an alternative implementation, during the training process, the Adam optimizer is used for parameter update. The specific process includes: setting the training stage to multiple epochs. In each epoch, the model is set to the training mode, and a variable is initialized to record the total loss of the current epoch. When traversing the entire dataset, each batch of noisy images and clear images is processed. At the beginning of each batch, the gradients of the optimizer are cleared. The noisy images are input into the model for forward propagation to calculate the loss between the output image and the clear image. Then, backpropagation is performed to calculate the gradients and update the model parameters, and at the same time, the loss of the current batch is accumulated into the total loss variable.

[0030] A power image denoising system based on an attention mechanism, comprising:

[0031] A data acquisition module, configured to acquire historical power image data, form a dataset, and preprocess the dataset;

[0032] A model construction module, configured to construct a power image denoising model based on an attention mechanism;

[0033] A model training module, configured to use the preprocessed dataset to train the power image denoising model;

[0034] A power image denoising module, configured to use the trained power image denoising model to process the acquired power image data to obtain a denoised power image;

[0035] The power image denoising model includes an input layer, a feature extraction module, an attention mechanism module, a feature fusion module, and an output layer connected in sequence. The input layer is used to receive power images. The feature extraction module is used to extract features from the input power images and perform feature normalization. The feature extraction module includes multiple convolutional layers, and at least some convolutional layers are added with dilated convolutional layers to expand the receptive field, and some dilated convolutional layers introduce holes in the convolutional kernel. The attention mechanism module receives the processed features and dynamically adjusts the attention degree to each feature channel, and enhances important features through a weighting mechanism. The feature fusion module is used to fuse the features processed by the attention mechanism module and generate a denoised power image through the output layer.

[0036] A computer-readable storage medium for storing computer instructions, which when executed by a processor, complete the steps in the above method.

[0037] An electronic device, including a memory and a processor, and computer instructions stored on the memory and running on the processor, which when executed by the processor, complete the steps in the above method.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0039] The present invention innovatively proposes a power image denoising method based on the attention mechanism. By combining deep learning with the attention mechanism and dilated convolution, the denoising effect of power images is improved. Compared with traditional mean filtering and median filtering, it can more effectively remove noise and retain key details; it not only improves the monitoring efficiency, but also effectively reduces the equipment maintenance cost, providing strong support for the safe and stable operation of the power system.

[0040] In order to make the above objects, features, and advantages of the present invention more obvious and understandable, the following specific preferred embodiments are given, and in conjunction with the accompanying drawings, the detailed description is as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings forming a part of this specification are used to provide a further understanding of the present invention. The schematic embodiments and descriptions thereof of the present invention are used to explain the present invention and do not constitute an improper limitation of the present invention.

[0042] Figure 1 It is a schematic diagram of the image denoising model framework of an embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0043] The present invention will be further described below in conjunction with the drawings and embodiments.

[0044] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0045] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0046] In the case of no conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0047] Embodiment 1

[0048] A power image denoising method based on the attention mechanism includes the following steps:

[0049] Obtain historical power image data, form a data set, and preprocess the data set;

[0050] Construct a power image denoising model based on the attention mechanism;

[0051] Use the preprocessed dataset to train the power image denoising model;

[0052] Use the trained power image denoising model to process the real-time acquired power image data to obtain the denoised power image.

[0053] The following is a detailed introduction to each step:

[0054] First is data collection. Collect and organize the power image dataset, including noisy images and corresponding clear images. Preprocess the images, such as resizing and normalizing, to adapt to the model input. To enhance the generalization ability of the model and reduce overfitting, data augmentation is performed.

[0055] In this embodiment, the following method is used for image augmentation.

[0056] Rotation: Randomly rotate the image (e.g., -15° to 15°) to increase image samples at different angles.

[0057] Flipping: Horizontally or vertically flip the image to increase sample diversity.

[0058] Cropping: Randomly crop a part of the image to increase the learning of local features.

[0059] Scaling: Randomly scale the image within a certain range to simulate the shooting effects at different distances.

[0060] Color transformation: Randomly adjust the brightness, contrast, and saturation to increase the robustness to light changes.

[0061] Noise addition: Add random noise (such as Gaussian noise) to the clear image to generate new noisy images and enhance the model's adaptability to noise

[0062] After data preparation is completed, model construction is carried out. Use PyTorch to implement the power image denoising network based on the attention mechanism.

[0063] The power image denoising network of this embodiment is as Figure 1 shown, mainly composed of several key modules to achieve efficient denoising effects.

[0064] (1) Input layer: The input layer receives the original noisy power image. The image format is RGB, and the size is 256x256 pixels. This module ensures that the data can smoothly enter the subsequent processing stage.

[0065] (2) Feature extraction module: The feature extraction module contains multiple convolutional layers, each followed by Batch Normalization and the ReLU activation function. The convolutional layers extract important information helpful for denoising by learning local features in the image. Batch Normalization helps accelerate the training speed and improve the model stability, while the ReLU activation function introduces non-linearity, enabling the model to capture more complex features.

[0066] (3) Dilated convolutional layer: The dilated convolutional layer is used to expand the receptive field, so as to capture global features in the power image and provide richer context information. This feature is crucial for dealing with complex backgrounds and identifying important device states.

[0067] (4) Attention mechanism module: The attention mechanism module dynamically adjusts the attention degree to each feature channel and enhances important features through a weighting mechanism. This design can effectively suppress noise and highlight key features, ensuring that the denoised image retains the state information of the device.

[0068] (5) Output layer: The output layer generates the denoised power image, and its number of channels is the same as that of the input image, ensuring the consistency of the output format. The design of the overall architecture enables this network to efficiently process power images, improve the image quality, and provide reliable support for the monitoring of the power system.

[0069] The process of building the model includes:

[0070] First is the input layer, which is responsible for receiving the original noisy image as input. The input image is in RGB format, containing three color channels (red, green, and blue), and its pixel values are usually between 0 and 255. The input layer provides the original image data for the subsequent feature extraction, attention mechanism, and feature fusion modules, laying the foundation for the entire denoising process. Then, the feature extraction module is mainly responsible for extracting useful feature information from the input noisy image to capture textures and details at different levels.

[0071] In this embodiment, the feature fusion module adopts a multi-scale fusion design method. It uses convolutional layers with multiple different receptive fields to extract multi-scale information, extracts features in a parallel convolutional branch manner, and realizes the fusion by pointwise weighting. The weights are determined by the network through adaptive learning. Feature fusion helps the model distinguish image signals and noise and reduces the impact of noise on image reconstruction and fuses the features of the image. First, the image undergoes preliminary feature extraction through a conventional convolutional layer. The input has 3 channels (RGB image), the convolutional layer outputs 64 channels, uses a 3x3 convolutional kernel, and adds 1-pixel zero-padding around the input image to maintain the spatial size of the feature map. Then, a batch normalization layer is used to normalize the features, improving the training stability and generalization performance of the model.

[0072] Starting from the second convolutional layer, a series of dilated convolutional layers are gradually introduced. Dilated convolutions can expand the receptive field, enabling the model to better capture the global information of the noise. Between every two layers, the dilation rate doubles, gradually increasing the receptive field. After each convolutional layer, a batch normalization layer and a ReLU activation layer follow. Traditional convolutional layers mainly capture local information, while image noise usually has global characteristics. To address this limitation, dilated convolutions are adopted, which expand the receptive field by introducing holes in the convolutional kernel, enabling the model to perceive more extensive regional information such as power environmental noise in power images, Gaussian noise introduced by image acquisition devices, and artificial marking errors or interferences (such as bright spots caused by arcs), and considering the global context of the image to distinguish noise from useful information. Feature maps are output at the intermediate layers of the convolutional network, and activation intensities are used to determine which regions contain key feature information. Response values are directly normalized or weighted on the feature map. Feature information with high responses represents possible important features, and finally, their weights are amplified. After each dilated convolutional layer, a batch normalization layer and a ReLU activation layer are also used. This combination stabilizes the training process, reduces internal covariate shift, and enhances the network's ability to represent non-linearity, enabling the model to more effectively capture complex noise features.

[0073] Before the last convolutional layer, an attention mechanism module is added. This module consists of two convolutional layers and a Sigmoid activation function. First, the initial convolutional layer compresses the feature channels to 1 / 8 of the original size, effectively reducing the dimension of the feature map, which helps extract more representative features. The next convolutional layer maps the compressed feature back to the original number of channels to ensure information retention. Finally, the Sigmoid activation function maps the output to the range (0, 1) to generate the attention weights for each channel. By element-wise multiplying the feature map and the attention weight map, the model can emphasize important features while suppressing unimportant features. This mechanism enhances the model's ability to focus on key information. After the attention mechanism, the last convolutional layer maps the feature map back to the original number of input channels to ensure consistent output format.

[0074] After the model is built, model training is carried out. The dataset images are stored in a specified folder for subsequent processing. PyTorch's DataLoader is used to read these images, and batch processing parameters are set to improve training efficiency. After loading the data, preprocessing is performed on the images. All images are resized to a unified size (256x256 pixels) to ensure the consistency of the input data. At the same time, the images are converted to tensor format to meet the input requirements of the model and prepare for training.

[0075] In this embodiment, the mean squared error (MSE) is used as the loss function to measure the difference between the denoised image and the clear image. And the Adam optimizer is used for parameter update to accelerate the convergence of the model. The model training stage is set to 20 epochs. In each epoch, the model is set to the training mode, and a variable is initialized to record the total loss of the current epoch. When traversing the entire training set, each batch of noisy images and clear images is processed. At the beginning of each batch, the gradients of the optimizer are cleared, the noisy images are input into the model for forward propagation, and the loss between the output image and the clear image is calculated. Then, backpropagation is performed to calculate the gradients and update the model parameters, and the loss of the current batch is accumulated into the total loss variable. At the end of each epoch, the average loss value of the current epoch is printed to monitor the training progress.

[0076] After completing all training epochs, the trained model parameters are saved using the torch.save() method for subsequent use and evaluation. This step ensures that the model can perform power image denoising in practical applications.

[0077] Model evaluation: Set the trained model to the evaluation mode. This can be achieved through model.eval(), which disables the training features of dropout and batch normalization to ensure the stability of the evaluation results. During the evaluation process, metrics for evaluation are defined, such as peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM). These two metrics can effectively measure the quality difference between the denoised image and the clear image. Traverse all batches of the validation set and perform the following steps:

[0078] Perform forward propagation on each batch of noisy images to obtain the denoised images output by the model.

[0079] Calculate the PSNR and SSIM values between the output image and the corresponding clear image, and accumulate the results.

[0080] After processing all batches, calculate and record the average PSNR and SSIM values for overall evaluation.

[0081] After the evaluation is completed, print the average PSNR and SSIM values of the model to help judge the performance of the model in the power image denoising task.

[0082] Embodiment 2

[0083] A power image denoising system based on an attention mechanism, comprising:

[0084] A data acquisition module configured to acquire historical power image data, form a data set, and preprocess the data set;

[0085] A model construction module, configured to construct a power image denoising model based on an attention mechanism;

[0086] A model training module, configured to train the power image denoising model using the preprocessed data set;

[0087] A power image denoising module, configured to process the acquired power image data using the trained power image denoising model to obtain a denoised power image;

[0088] The power image denoising model includes an input layer, a feature extraction module, an attention mechanism module, a feature fusion module, and an output layer connected in sequence. The input layer is used to receive a power image. The feature extraction module is used to extract features from the input power image and perform feature normalization. The feature extraction module includes a plurality of convolutional layers, and at least some of the convolutional layers are added with dilated convolutional layers to expand the receptive field, and some of the dilated convolutional layers introduce holes in the convolutional kernel. The attention mechanism module receives the processed features and dynamically adjusts the attention degree to each feature channel, and enhances important features through a weighting mechanism. The feature fusion module is used to fuse the features processed by the attention mechanism module and generate a denoised power image through the output layer.

[0089] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0090] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, and the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the functions specified in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0091] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to work in a particular manner, such that the instructions stored in the computer-readable memory produce a manufacture including an instruction means that implements the function specified in one process or a plurality of processes and / or one block or a plurality of blocks of the flowchart. Figure 1 one process or a plurality of processes and / or Figure 1 one block or a plurality of blocks.

[0092] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus, so as to perform a series of operation steps on the computer or other programmable apparatus to generate a computer-implemented process, and thus the instructions executed on the computer or other programmable apparatus provide steps for implementing the function specified in one process or a plurality of processes and / or one block or a plurality of blocks of the flowchart. Figure 1 one process or a plurality of processes and / or Figure 1 one block or a plurality of blocks.

[0093] The foregoing is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made by those skilled in the art without creative efforts within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A power image denoising method based on an attention mechanism, characterized in that, It includes the following steps: Obtain historical power image data, form a data set, and preprocess the data set; Construct a power image denoising model based on the attention mechanism; Use the preprocessed data set to train the power image denoising model; Use the trained power image denoising model to process the obtained power image data to obtain a denoised power image; The power image denoising model includes an input layer, a feature extraction module, an attention mechanism module, a feature fusion module, and an output layer connected in sequence. The input layer is used to receive a power image. The feature extraction module is used to extract features from the input power image and perform feature normalization. The feature extraction module includes multiple convolutional layers, and at least some convolutional layers are added with dilated convolutional layers to expand the receptive field, and some dilated convolutional layers introduce holes in the convolutional kernel. The attention mechanism module receives the processed features and dynamically adjusts the attention degree to each feature channel, and enhances important features through a weighting mechanism. The feature fusion module is used to fuse the features processed by the attention mechanism module and generate a denoised power image through the output layer.

2. The method for denoising power images based on the attention mechanism according to claim 1, wherein, The preprocessing process includes size adjustment and normalization processing.

3. The power image denoising method based on the attention mechanism according to claim 1, characterized in that, The preprocessing process further includes data augmentation, and the data augmentation methods include several of the following: Randomly rotate the image to increase image samples at different angles; Horizontally or vertically flip the image to increase sample diversity; Randomly crop a part of the image area to increase the learning of local features; Randomly scale the image within a certain range to simulate shooting effects at different distances; Randomly adjust the brightness, contrast, and saturation to increase the robustness to light changes; Add random noise to the clear image to generate a new noisy image.

4. A power image denoising method based on an attention mechanism according to claim 1, characterized in that, The feature extraction module includes a first convolutional layer and multiple dilated convolutional layers. The first convolutional layer is used for preliminary feature extraction and adds 1-pixel zero padding around the input image to maintain the spatial size of the feature map; Between every two layers of the dilated convolutional layers, the dilation rate is doubled.

5. The method for denoising power images based on the attention mechanism according to claim 4, characterized in that, Behind each convolutional layer, there is a batch normalization layer and a ReLU activation layer.

6. The power image denoising method based on the attention mechanism according to claim 1, characterized in that, The attention mechanism module includes two convolutional layers and a Sigmoid activation function. The first convolutional layer is used to compress the feature channels to 1 / 8 of the original size to reduce the dimension of the feature map. The second convolutional layer maps the compressed feature back to the original number of channels to ensure the retention of information. The Sigmoid activation function is used to map the output to the range of (0,1) to generate the attention weight for each channel; The attention mechanism module performs element-wise multiplication on the feature map and the attention weight map.

7. A power image denoising method based on an attention mechanism according to claim 1 or 6, characterized in that, After the attention mechanism module, there is also a convolutional layer, which is used to map the feature map back to the original input number of channels to ensure consistent output format.

8. The power image denoising method based on the attention mechanism according to claim 1, characterized in that During the training process, the mean squared error is used as the loss function to measure the difference between the denoised image and the clear image.

9. The power image denoising method based on the attention mechanism according to claim 1, characterized in that During the training process, the Adam optimizer is used for parameter updates. The specific process includes: setting the training phase to multiple epochs. In each epoch, the model is set to the training mode, and a variable is initialized to record the total loss of the current epoch. When traversing the entire dataset, each batch of noisy images and clear images is processed. At the beginning of each batch, the gradients of the optimizer are cleared. The noisy images are input into the model for forward propagation to calculate the loss between the output image and the clear image. Then, backpropagation is performed to calculate the gradients and update the model parameters, and at the same time, the loss of the current batch is accumulated into the total loss variable.

10. A power image denoising system based on an attention mechanism, characterized in that, Including: A data acquisition module configured to acquire historical power image data, form a dataset, and preprocess the dataset; A model construction module configured to construct a power image denoising model based on an attention mechanism; A model training module configured to train the power image denoising model using the preprocessed dataset; A power image denoising module configured to process the acquired power image data using the trained power image denoising model to obtain a denoised power image; The power image denoising model includes an input layer, a feature extraction module, an attention mechanism module, a feature fusion module, and an output layer connected in sequence. The input layer is used to receive power images. The feature extraction module is used to extract features from the input power images and perform feature normalization. The feature extraction module includes multiple convolutional layers, and at least some of the convolutional layers are added with dilated convolutional layers to expand the receptive field, and some of the dilated convolutional layers introduce holes in the convolutional kernel. The attention mechanism module receives the processed features and dynamically adjusts the attention degree to each feature channel, and enhances important features through a weighting mechanism. The feature fusion module is used to fuse the features processed by the attention mechanism module and generate a denoised power image through the output layer.

Images