Method and device for denoising transmission tower component image

Abstract

The application provides a power transmission tower component image denoising method and device, and belongs to the technical field of image denoising. The method comprises the following steps: obtaining a power transmission tower component image to be denoised; inputting the power transmission tower component image to be denoised into a trained image denoising model to obtain a denoised power transmission tower component image; wherein the image denoising model is an HR-Net network, and each DACNN module in the HR-Net network is improved by a double attention mechanism. The application adopts the HR-Net network to add the double attention mechanism as an image model, utilizes the high-resolution characteristics of the HR-Net network, and fully processes details and local regions in the power transmission tower component image based on the noise characteristics of the power transmission tower component image, so that the power transmission tower component image can be accurately denoised, and better denoising effect can be obtained.

Description

Technical Field

[0001] This invention relates to the field of image noise reduction technology, and in particular to a method and apparatus for noise reduction of images of power transmission tower components. Background Technology

[0002] Images of transmission tower components are often noisy due to outdoor environments and lighting conditions. Precise denoising of these specific images can improve image quality, helping engineers and technicians better analyze and assess the condition of transmission towers. Transmission towers are also a crucial component of the power grid system; accurate detection and diagnosis of their condition are vital for the operation and safety of the power system. Denoising images of transmission tower components can help identify and locate anomalies in a timely manner, allowing for proactive maintenance or replacement measures to ensure the normal operation and safety of the power grid.

[0003] Currently, methods for denoising and enhancing images of power transmission tower components are mainly based on traditional signal processing techniques and machine learning models. Traditional signal processing techniques typically rely on filters and image processing algorithms to handle noise. However, in complex images of power transmission tower components, noise often exhibits diverse and nonlinear characteristics, making traditional methods inadequate for such complexities and resulting in poor denoising performance. In recent years, deep learning models, such as Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs), have become increasingly mature in image processing. For pixel-level image tasks such as image segmentation (semantic segmentation, instance segmentation) and image generation (AI painting, AI redrawing, image denoising), the U-Net network structure is commonly used and has made significant progress in image denoising and modeling. However, these methods often focus only on feature extraction of the overall image, which has limitations when processing complex images of power transmission tower components. Details and local regions in the U-Net structure are not adequately processed in the shallow feature layers, leading to unsatisfactory denoising results. Summary of the Invention

[0004] This invention provides a method and apparatus for denoising images of power transmission tower components to address the problem of improving the denoising effect of power transmission tower images.

[0005] In a first aspect, embodiments of the present invention provide a method for denoising images of transmission tower components, comprising:

[0006] Acquire images of the transmission tower components to be denoised;

[0007] The image of the transmission tower component to be denoised is input into the trained image denoising model to obtain the denoised image of the transmission tower component; the image denoising model is the HR-Net network, and each DACNN module in the HR-Net network has been improved with a dual attention mechanism.

[0008] In one possible implementation, each DACNN module includes a convolutional module, a spatial attention module, a channel attention module, and a batch normalization module that are sequentially connected.

[0009] The spatial attention module includes a max pooling layer, an average pooling layer, and a convolutional layer. The max pooling layer and the average pooling layer are connected in parallel and then connected in series with the convolutional layer.

[0010] The channel attention module consists of an average pooling layer, a first fully connected layer, and a second fully connected layer connected in series.

[0011] In one possible implementation, the resolution and number of channels of the four high-resolution feature layers in the HR-Net network are 1 / 2 and C, 1 / 4 and 2C, 1 / 8 and 3C, and 1 / 16 and 4C, respectively.

[0012] In one possible implementation, before inputting the image of the transmission tower component to be denoised into the trained image denoising model, the following is also included:

[0013] Obtain the initial image denoising model;

[0014] The initial batch size, initial learning rate, number of iterations, learning rate decay strategy, and loss function of the image denoising model are set, and the loss function is optimized based on the training sample set to obtain the trained image denoising model. The training sample set includes multiple training samples, each of which is an image of a transmission tower component with noise added, and the corresponding image of a transmission tower component without noise added is the sample label.

[0015] In one possible implementation, before optimizing the loss function based on the training sample set, the following is also included:

[0016] Acquire multiple images of transmission tower components;

[0017] For each image of a transmission tower component, speckle noise is added to the image to obtain a noisy transmission tower component image. The noisy transmission tower component image is then used as a training sample and added to the training sample set as a sample label.

[0018] In one possible implementation, the formula for adding speckle noise to the image of the transmission tower component is:

[0019]

[0020] Where n is the noise signal, y is the output signal after noise pollution, and x is the input signal, and x and n are independent of each other.

[0021] In one possible implementation, speckle noise is added to the image of the transmission tower component, including:

[0022] Based on the component categories contained in the image of the transmission tower component, the noise signal corresponding to the image of the transmission tower component is determined in different levels of noise signals, and speckle noise is added to the image of the transmission tower component based on the noise signal.

[0023] In one possible implementation, speckle noise is added to the image of the transmission tower component, including:

[0024] Based on the region where the component is located in the image of the transmission tower component, the noise region corresponding to the image of the transmission tower component is determined, and speckle noise is added to the image of the transmission tower component based on the noise region.

[0025] Secondly, embodiments of the present invention provide a noise reduction device for images of transmission tower components, comprising:

[0026] The acquisition module is used to acquire images of the transmission tower components to be denoised;

[0027] The denoising module is used to input the image of the transmission tower component to be denoised into the trained image denoising model to obtain the denoised image of the transmission tower component; the image denoising model is the HR-Net network, and each DACNN module in the HR-Net network has been improved with a dual attention mechanism.

[0028] This invention provides a method and apparatus for denoising images of power transmission tower components. It employs an HR-Net network with a dual attention mechanism as the image model. By utilizing the high-resolution characteristics of the HR-Net network and based on the noise characteristics of the power transmission tower component images, it fully processes the details and local regions in the images, enabling precise denoising of the power transmission tower component images and achieving better denoising results. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention, 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0030] Figure 1This is a flowchart illustrating the implementation of the noise reduction method for images of power transmission tower components provided in this embodiment of the invention.

[0031] Figure 2 This is a schematic diagram of the structure of the DACNN module provided in an embodiment of the present invention;

[0032] Figure 3 This is a schematic diagram of the spatial attention module provided in an embodiment of the present invention;

[0033] Figure 4 This is a schematic diagram of the channel attention module provided in an embodiment of the present invention;

[0034] Figure 5 This is a schematic diagram of the HR-Net network structure provided in an embodiment of the present invention;

[0035] Figure 6 These are images of noisy transmission tower components provided in embodiments of the present invention;

[0036] Figure 7 This is a denoised image of a power transmission tower component provided in an embodiment of the present invention;

[0037] Figure 8 This is a structural schematic diagram of the noise reduction device for images of power transmission tower components provided in an embodiment of the present invention. Detailed Implementation

[0038] 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 the invention. However, those skilled in the art will understand that the invention can be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0039] To make the objectives, technical solutions, and advantages of the present invention clearer, specific embodiments will be described below in conjunction with the accompanying drawings.

[0040] Figure 1 The flowchart illustrating the noise reduction method for images of power transmission tower components provided in this embodiment of the invention is described in detail below:

[0041] Step 101: Obtain images of the transmission tower components to be denoised.

[0042] In this embodiment, the images of transmission tower components are noisy due to factors such as outdoor environment and lighting conditions. This noise manifests as stripes, spots, and blurring. General images, including natural landscapes and portraits, exhibit different levels of noise under different environments and shooting conditions, such as Gaussian noise and salt-and-pepper noise. Therefore, denoising images of transmission tower components differs from denoising general images in terms of processing objects, application scenarios, achieved effects, and noise types, and cannot be generalized. For images of transmission tower components in these specific scenarios, precise denoising processing can improve image quality, helping engineers and technicians better analyze and evaluate the condition of transmission towers. Transmission towers are also a particularly important component of the power grid system; accurate detection and diagnosis of their condition are crucial for the operation and safety of the power system. Denoising images of transmission tower components can help to promptly detect and locate abnormal conditions, allowing for early maintenance or replacement measures to ensure the normal operation and safety of the power grid.

[0043] Currently, methods for denoising and enhancing images of transmission tower components are mainly based on traditional signal processing techniques and machine learning models. These methods have certain limitations when processing complex images of transmission tower components, such as poor denoising performance and low modeling accuracy.

[0044] Traditional signal processing techniques typically rely on filters and image processing algorithms to handle noise. However, in complex images of power transmission tower components, noise often exhibits diverse and nonlinear characteristics, making traditional methods inadequate for such complexities and resulting in poor denoising performance. In recent years, deep learning models, such as Convolutional Neural Networks (CNNs) and Generative Adversarial Networks (GANs), have become increasingly sophisticated for image processing tasks. The U-Net network structure is commonly used for pixel-level image tasks such as image segmentation (semantic segmentation, instance segmentation) and image generation (AI painting, AI redrawing, image denoising), achieving significant progress in image denoising and modeling. However, these methods often focus only on feature extraction from the overall image, neglecting to adequately process details and local regions in the shallow feature layers of the U-Net structure, leading to unsatisfactory denoising results.

[0045] Step 102: Input the image of the transmission tower component to be denoised into the trained image denoising model to obtain the denoised image of the transmission tower component; wherein, the image denoising model is the HR-Net network, and each DACNN module in the HR-Net network has been improved with a dual attention mechanism.

[0046] In this embodiment, this patent relates to a convolutional neural network for image denoising of transmission tower components using an improved dual-attention mechanism. This method utilizes a dual-attention mechanism model to process and analyze images of transmission tower components, achieving denoising enhancement and modeling. Based on the characteristics of the transmission tower component images, this method focuses on important details and local regions through an attention mechanism, thereby improving denoising effectiveness and modeling accuracy. Furthermore, this method can achieve different levels of denoising and modeling effects by adjusting the parameters of the attention mechanism according to actual needs.

[0047] To address the aforementioned issues, this patent proposes an improved convolutional neural network for image denoising of power transmission tower components using a dual-attention mechanism. Specifically, the convolutional neural network employs an improved HR-Net (High-Resolution Network). The primary goal of the HR-Net network is to handle high-resolution image tasks and multi-scale image tasks, providing more accurate predictions while maintaining a smaller number of network parameters. HR-Net utilizes the idea of ​​multi-resolution information flowing through multiple branch networks to process input images of different resolutions in parallel, preserving and transmitting high-level features while sharing low-level features.

[0048] Secondly, this method employs a dual-attention mechanism, focusing on both global and local information of the image. The global attention mechanism allows the model to capture the overall structure and key features of the image, while the local attention mechanism focuses on details and local regions, extracting more accurate noise and modeling information. This dual-attention mechanism better balances global and local information, improving both noise reduction and modeling accuracy. Furthermore, training and optimizing the dual-attention model requires collecting and annotating a large amount of image data of transmission tower components. The annotation process typically requires professional personnel, including labeling noisy regions and adjusting modeling parameters. The quality of these datasets and annotations is crucial for training and optimizing the model.

[0049] As can be seen from the above background technology introduction, the convolutional neural network for image denoising of transmission tower components with dual attention mechanism can effectively overcome the limitations of traditional methods and other machine learning models, and provide more accurate and reliable denoising and modeling results, which is of great significance to the operation and maintenance of power systems.

[0050] This invention employs the HR-Net network with a dual attention mechanism as an image denoising model. Leveraging the high-resolution characteristics of the HR-Net network and based on the noise characteristics of transmission tower component images, it fully processes the details and local regions in the transmission tower component images, enabling precise denoising of the transmission tower component images and achieving better denoising results.

[0051] In one possible implementation, each DACNN module includes a convolutional module, a spatial attention module, a channel attention module, and a batch normalization module that are sequentially connected.

[0052] The spatial attention module includes a max pooling layer, an average pooling layer, and a convolutional layer. The max pooling layer and the average pooling layer are connected in parallel and then connected in series with the convolutional layer.

[0053] The channel attention module consists of an average pooling layer, a first fully connected layer, and a second fully connected layer connected in series.

[0054] In this embodiment, as Figure 2 , 3 As shown in Figure 4, the image denoising model includes an HR-Net network with total convolutional operations, a ReLU activation function, and a dual attention mechanism module. The dual attention mechanism module consists of a spatial attention module and a channel attention module. The input to the HR-Net network is a convolutional layer using the ReLU activation function, and the output is a convolutional layer using the sigmoid activation function. The intermediate layers include multiple DACNN modules. Setting the feature layers to resolutions of 1 / 2, 1 / 4, 1 / 8, and 1 / 16 with channel counts of C, 2C, 3C, and 4C respectively (C=24, compared to C=48 in the original HRNet) increases the number of model parameters; therefore, the number of feature layer channels is reduced.

[0055] The overall process of using the image denoising model provided by this invention for denoising transmission tower images can include:

[0056] S1 constructs the training set;

[0057] The operating principle of the S2 image denoising model;

[0058] S3 trains and tests the initial image denoising model;

[0059] S4 uses a trained image denoising model for denoising.

[0060] The image denoising model in S2 mainly processes noisy images of transmission tower components by following these steps:

[0061] S21 first feeds the noisy image of the transmission tower components into a convolutional layer for convolution, and then uses the ReLU activation function to perform a non-linear transformation on the convolution result. The convolution result is then fed into a two-layer attention neural network (DACNN) for higher-level feature extraction. The output feature F (of shape C*H*W) of the convolutional layer is used as the input of this module. First, max pooling and average pooling are applied in the channel dimension to obtain two different feature descriptions of shape 1*H*W. Then, concatenate the two feature descriptions, resulting in a concatenated data of shape 2*H*W. The concatenated result is then fed into the convolutional layer for convolution, and the Sigmoid function is used to compress the convolution result to between 0 and 1. Spatial feature weighting is applied to the spatial principal axis and then input into the channel attention module. Each convolution operation is followed by a DACNN operation. The ReLU activation function formula is as follows:

[0062] ReLU(x) = max(x,0)

[0063] The S22 spatial attention module includes two pooling layers and a spatial principal axis. The two pooling layers are a max pooling layer and an average pooling layer. The feature maps are concatenated into a large feature map after max pooling and average pooling operations. This large feature map is then convolved by a convolutional layer and subsequently weighted spatially by a sigmoid activation function. The spatial principal axis is then weighted and input into the channel attention module. The sigmoid activation function formula is as follows:

[0064]

[0065] Spatial Attention Module F sa The calculation formula is as follows:

[0066] F sa (x)=σ(f i *Cat(x avg ,x max ))

[0067] In the formula, σ is the sigmoid activation function, and f i For convolution operations with kernel size i, Cat represents the concatenation operation, and x represents the concatenation operation. avg and x max These are the feature maps after average pooling and after max pooling, respectively.

[0068] The S23 channel attention module includes an average pooling layer, a multilayer perceptron, and a softmax activation function. The multilayer perceptron consists of two linear layers, ultimately producing attention weights for the main branches and non-attention weights for a pair of residual connections within the main branches. The input to this module is the output of a convolutional layer, a C*H*W 3D dataset, where C represents the number of channels (i.e., the number of convolutional kernels), and H and W represent the height and depth of the input image, respectively. Global Max Pooling and Global Average Pooling operations are performed on the input data. The number of neurons in the first Fully Connection is C / R (where R represents the compression ratio), and the number of neurons in the second Fully Connection is C. This ultimately produces attention weights for the main branches and non-attention weights for the main branches, constructing better subsequent features. The softmax activation function is calculated as follows:

[0069]

[0070] Inter-channel attention module F ca The calculation formula is as follows:

[0071] F ca (x)=δ[W2(W1x avg +b1)+b2]

[0072] In the formula, δ is the Softmax activation function, W1 and W2 are the weights of the multilayer perceptron, b1 and b2 are the biases of the multilayer perceptron, and x is the weight of the multilayer perceptron. avg This is the feature map after average pooling.

[0073] In one possible implementation, the resolution and number of channels of the four high-resolution feature layers in the HR-Net network are 1 / 2 and C, 1 / 4 and 2C, 1 / 8 and 3C, and 1 / 16 and 4C, respectively.

[0074] In this embodiment, the original maximum 1 / 4 high-resolution feature layer in the HR-Net network is modified to a 1 / 2 high-resolution feature layer, thus focusing more on image texture. This increases the number of model parameters, therefore the number of feature layer channels is reduced. The modified resolutions of 1 / 2, 1 / 4, 1 / 8, and 1 / 16 have C, 2C, 3C, and 4C channels respectively, better handling more details. Multiple parallel branch networks process input images of different resolutions, allowing the network to simultaneously capture low-level and high-level features, sharing and fusing them to increase the network's receptive field and improve network performance. Secondly, an end-to-end model is constructed using an improved convolutional neural network. This model mainly includes convolution operations, channel attention, spatial attention, and batch normalization. Convolutional neural networks mainly consist of convolution and pooling. Pooling operations can lead to the loss of some image information. To maximize the preservation of complete information in the image, this invention uses fully convolutional operations in the construction of the denoising model. Simultaneously, spatial attention and channel attention mechanisms are introduced to improve the existing convolutional neural network model, further enhancing the model's feature extraction capabilities. With these improvements, the model is able to better capture the feature information of images and achieve better results in noise reduction tasks.

[0075] To prevent the loss of features in the final extraction, this invention adopts the idea of ​​the nonlocal-means algorithm in the spatial attention mechanism. This algorithm can balance the global and local information of the image. By selecting an area of ​​appropriate size, it can extract global information without significantly increasing the training parameters of the model.

[0076] This invention introduces a channel attention mechanism to dynamically adjust the weight coefficients of each channel. The channel attention module includes an average pooling layer, a multilayer perceptron, and a softmax activation function. The multilayer perceptron includes two linear layers. Ultimately, it produces attention weights for the main branch and non-attention weights in the main branch, so that the model can extract better features.

[0077] The convolutional neural network for image denoising of transmission tower components provided by this invention with a dual attention mechanism has the following beneficial effects:

[0078] 1. An improved HR-Net network is applied to image denoising of power transmission tower components. By processing input images in a multi-resolution manner, it can more effectively capture texture details and image structure at different scales, remove unnecessary detail information and noise, thereby improving image quality. The improved HR-Net network structure optimizes the way information is transmitted between branch networks, improves the utilization rate of network parameters, and enables the processing of a large number of power transmission tower component images in a shorter time, thereby improving data processing efficiency and accuracy.

[0079] 2. The application of a dual attention mechanism to image denoising of transmission tower components is an innovation of this invention. This method combines an attention mechanism with image denoising techniques, achieving significant results in processing images of transmission tower components. By introducing a dual attention mechanism, this method can focus on key image regions and better capture the spatial and channel correlations of the image, thereby reducing image noise while preserving more detailed information. This innovation effectively improves the accuracy and stability of the image denoising algorithm for transmission tower components.

[0080] 3. In the spatial attention mechanism, this invention employs the idea of ​​the NonLocal-means algorithm. This algorithm can balance global and local information in an image by selecting a region of appropriate size to extract global information without significantly increasing the model's training parameters. This invention uses a similar approach in the spatial attention mechanism to ensure the full extraction of overall features. By introducing this mechanism, global and local feature information can be better balanced, thereby improving the model's performance and robustness while maintaining its efficiency and trainability. This improvement helps to increase the model's ability to perceive complex image scenes and enhances the quality of noise reduction results.

[0081] 4. In this invention, a channel attention mechanism is introduced to dynamically adjust the weight coefficients of each channel, enabling the model to extract high-quality features from the image more accurately. The main purpose of the channel attention mechanism is to adaptively assign feature weights to each channel. Through the learned attention weights, the network focuses more on important feature channels in a specific task, while ignoring channels that are irrelevant to the task. The channel attention mechanism dynamically calculates the weight coefficients of each channel based on the content of the input image. These weight coefficients are applied to the feature map of each channel to amplify useful information and suppress noise and irrelevant features.

[0082] In one possible implementation, before inputting the image of the transmission tower component to be denoised into the trained image denoising model, the following is also included:

[0083] Obtain the initial image denoising model;

[0084] The initial batch size, initial learning rate, number of iterations, learning rate decay strategy, and loss function of the image denoising model are set, and the loss function is optimized based on the training sample set to obtain the trained image denoising model. The training sample set includes multiple training samples, each of which is an image of a transmission tower component with noise added, and the corresponding image of a transmission tower component without noise added is the sample label.

[0085] In this embodiment, the specific steps of step S3 are as follows:

[0086] S3 uses preprocessed noisy images of transmission tower components to train a denoising network model, and uses a loss function to constrain the model, thereby obtaining a network model capable of denoising images of transmission tower components. This step mainly includes the following steps:

[0087] S31 sets the batch size, initial learning rate, number of iterations, learning rate decay strategy, and loss function of the denoising network model, and optimizes the loss function, where the loss function is the mean squared error, and its mathematical expression is:

[0088]

[0089] In the formula: Θ represents the network model parameters, R(y) i ;Θ) represents the image learned by the network, y i For noise images of transmission tower components, x i This represents a clean image, and N is the number of training samples.

[0090] To simulate noise in real-world scenarios, S32 adds different levels of noise and uses the obtained image denoising model to denoise the test set, resulting in a denoised image.

[0091] S4 evaluates the denoised transmission tower component image and the noisy transmission tower component image using PNSR and SSIM values. In the field of image denoising, Peak Signal-to-Noise Ratio (PSNR, R0) is the most important factor. PSNR ) and structural similarity (SSIM, f SSIM This is often used as an evaluation metric for quantifying noise reduction effectiveness. This step mainly includes the following steps:

[0092] S41 PNSR measures the distortion between the denoised image and the original image by calculating the error between corresponding pixels. The calculation formula is as follows:

[0093]

[0094] in, In the formula: α MSE E represents the mean square error; max represents the maximum pixel in the entire image; m and n represent the width and height of the image, respectively; I represents the image after denoising; K represents the image before denoising, which is obtained by summing the squares of the differences between corresponding pixels in the two images.

[0095] S42 SSIM measures the similarity between the denoised image and the original image in terms of brightness, contrast, and structure. The calculation formula is as follows:

[0096]

[0097] Where x is the natural image; y is the image after model denoising; μ x and μ y σ represents the mean gray level of x and y, respectively; x and σ y σ represents the standard deviation of the gray values ​​of x and y. xy Let C1 and C2 represent the covariance of x and y; C1 and C2 are both constants.

[0098] In one possible implementation, before optimizing the loss function based on the training sample set, the following is also included:

[0099] Acquire multiple images of transmission tower components;

[0100] For each image of a transmission tower component, speckle noise is added to the image to obtain a noisy transmission tower component image. The noisy transmission tower component image is then used as a training sample and added to the training sample set as a sample label.

[0101] In this embodiment, the specific steps of step S1 are as follows:

[0102] S1 adds noise to the input training images of transmission tower components to obtain a noisy image. This step includes the following sub-steps:

[0103] S11 Select training samples from the training transmission tower component image dataset as the original training set. The images in the training samples are all noise-free images of the same size.

[0104] S12 constructs a noisy image by adding speckle noise to images of transmission tower components based on the speckle noise model proposed by Loupas. The specific formula of the speckle noise model is as follows:

[0105]

[0106] In the formula, n is the noise signal, y is the output signal after noise pollution, and x is the input signal, and x and n are independent of each other.

[0107] S13 performs augmentation operations on the extracted image patches, that is, rotates and scales the images respectively, with rotation angles of 0, 90, 180, and 270 degrees, and scaling ratios of 0.7, 0.8, 0.9, and 1 times, which are used as the training dataset.

[0108] In one possible implementation, the formula for adding speckle noise to the image of the transmission tower component is:

[0109]

[0110] Where n is the noise signal, y is the output signal after noise pollution, and x is the input signal, and x and n are independent of each other.

[0111] In this embodiment, the images of transmission tower components are noisy due to factors such as outdoor environment and lighting conditions. The noise exists in the form of stripes, spots, and blurring. The spot noise model is more consistent with the noise characteristics in the images of transmission tower components.

[0112] In one possible implementation, speckle noise is added to the image of the transmission tower component, including:

[0113] Based on the component categories contained in the image of the transmission tower component, the noise signal corresponding to the image of the transmission tower component is determined in different levels of noise signals, and speckle noise is added to the image of the transmission tower component based on the noise signal.

[0114] In this embodiment, the severity of noise and the difficulty of noise reduction in the image of the transmission tower component vary depending on the type of transmission tower component. Therefore, the noise level of each type of transmission tower component can be set according to the specific situation of each type of transmission tower component, and the corresponding noise signal can be selected from different levels of noise signals to add noise to the image in order to achieve a better noise reduction effect.

[0115] In one possible implementation, speckle noise is added to the image of the transmission tower component, including:

[0116] Based on the region where the component is located in the image of the transmission tower component, the noise region corresponding to the image of the transmission tower component is determined, and speckle noise is added to the image of the transmission tower component based on the noise region.

[0117] In this embodiment, the images of power transmission tower components typically contain both the component parts and the environment. To denoise the images of power transmission tower components, it is necessary to obtain clear images of the component parts. However, the noise in the component parts and the noise in the environment have different characteristics. Therefore, speckle noise can be added only to the image parts to improve the denoising effect of the model.

[0118] In one specific embodiment, after training the image denoising model provided by the present invention, 10 images were selected for testing based on two evaluation metrics, PSNR and SSIM. The test results are shown in Table 1.

[0119] Table 1

[0120]

[0121] As can be seen from Table 1, the evaluation results show that the PSNR of the noise-removed transmission tower component image is significantly improved compared with the traditional CNN, and the SSIM value is closer to 1. The convolutional neural network for noise reduction of transmission tower component images provided by the dual attention mechanism of this invention has a much better noise reduction performance than CNN.

[0122] Based on this model Figure 6 The image of the transmission tower component shown is subjected to noise reduction processing to obtain... Figure 7 The image shown demonstrates that this method effectively removes noise and better preserves the edge information of transmission tower components.

[0123] 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 the present invention.

[0124] The following are device embodiments of the present invention. For details not described in detail, please refer to the corresponding method embodiments described above.

[0125] Figure 8 A schematic diagram of a noise reduction device for images of transmission tower components provided in an embodiment of the present invention is shown. For ease of explanation, only the parts related to the embodiment of the present invention are shown, and are described in detail below:

[0126] like Figure 8 As shown, the noise reduction device 8 for the image of transmission tower components includes:

[0127] Acquisition module 81 is used to acquire images of the transmission tower components to be denoised;

[0128] The denoising module 82 is used to input the image of the transmission tower component to be denoised into the trained image denoising model to obtain the denoised image of the transmission tower component; wherein, the image denoising model is an HR-Net network, and each DACNN module in the HR-Net network has been improved with a dual attention mechanism.

[0129] In one possible implementation, each DACNN module includes a convolutional module, a spatial attention module, a channel attention module, and a batch normalization module that are sequentially connected.

[0130] The spatial attention module includes a max pooling layer, an average pooling layer, and a convolutional layer. The max pooling layer and the average pooling layer are connected in parallel and then connected in series with the convolutional layer.

[0131] The channel attention module consists of an average pooling layer, a first fully connected layer, and a second fully connected layer connected in series.

[0132] In one possible implementation, the resolution and number of channels of the four high-resolution feature layers in the HR-Net network are 1 / 2 and C, 1 / 4 and 2C, 1 / 8 and 3C, and 1 / 16 and 4C, respectively.

[0133] In one possible implementation, the device 8 further includes a training module for:

[0134] Before inputting the images of the transmission tower components to be denoised into the trained image denoising model, an initial image denoising model is obtained;

[0135] The initial batch size, initial learning rate, number of iterations, learning rate decay strategy, and loss function of the image denoising model are set, and the loss function is optimized based on the training sample set to obtain the trained image denoising model. The training sample set includes multiple training samples, each of which is an image of a transmission tower component with noise added, and the corresponding image of a transmission tower component without noise added is the sample label.

[0136] In one possible implementation, the training module is also used for:

[0137] Before optimizing the loss function based on the training sample set, multiple images of transmission tower components are acquired;

[0138] For each image of a transmission tower component, speckle noise is added to the image to obtain a noisy transmission tower component image. The noisy transmission tower component image is then used as a training sample and added to the training sample set as a sample label.

[0139] In one possible implementation, the formula for adding speckle noise to the image of the transmission tower component is:

[0140]

[0141] Where n is the noise signal, y is the output signal after noise pollution, and x is the input signal, and x and n are independent of each other.

[0142] In one possible implementation, the training module is specifically used for:

[0143] Based on the component categories contained in the image of the transmission tower component, the noise signal corresponding to the image of the transmission tower component is determined in different levels of noise signals, and speckle noise is added to the image of the transmission tower component based on the noise signal.

[0144] In one possible implementation, the training module is specifically used for:

[0145] Based on the region where the component is located in the image of the transmission tower component, the noise region corresponding to the image of the transmission tower component is determined, and speckle noise is added to the image of the transmission tower component based on the noise region.

[0146] This invention employs the HR-Net network with a dual attention mechanism as an image model. Leveraging the high-resolution characteristics of the HR-Net network and based on the noise characteristics of transmission tower component images, it fully processes the details and local regions in the transmission tower component images, enabling precise denoising and achieving better noise reduction results.

[0147] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0148] Those skilled in the art will recognize that the templates, units, and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0149] If the module / unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the above-described methods for noise reduction of images of various transmission tower components. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the 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, a random access memory, an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0150] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention 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 the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A method for denoising images of transmission tower components, characterized in that, include: Acquire images of the transmission tower components to be denoised; The image of the transmission tower component to be denoised is input into a trained image denoising model to obtain a denoised image of the transmission tower component; wherein, the image denoising model is an HR-Net network, and each DACNN module in the HR-Net network has been improved with a dual attention mechanism; The image denoising model is trained based on a training sample set, which is constructed as follows: Acquire multiple images of transmission tower components; For each image of a transmission tower component, speckle noise is added to the image to obtain a noisy transmission tower component image. The noisy transmission tower component image is then used as a training sample and added to the training sample set as a sample label. The addition of speckle noise to the image of the transmission tower component includes: Based on the component categories contained in the image of the transmission tower component, the noise signal corresponding to the image of the transmission tower component is determined from different levels of noise signals, and speckle noise is added to the image of the transmission tower component based on the noise signal.

2. The noise reduction method for images of transmission tower components according to claim 1, characterized in that, Each DACNN module includes a convolutional module, a spatial attention module, a channel attention module, and a batch normalization module, which are connected in sequence. The spatial attention module includes a max pooling layer, an average pooling layer, and a convolutional layer. The max pooling layer and the average pooling layer are connected in parallel and then connected in series with the convolutional layer. The channel attention module includes an average pooling layer, a first fully connected layer, and a second fully connected layer connected in series.

3. The noise reduction method for images of transmission tower components according to claim 1, characterized in that, The resolutions and number of channels of the four high-resolution feature layers in the HR-Net network are 1 / 2 and C, 1 / 4 and 2C, 1 / 8 and 3C, and 1 / 16 and 4C, respectively.

4. The noise reduction method for images of transmission tower components according to claim 1, characterized in that, Before inputting the image of the transmission tower component to be denoised into the trained image denoising model, the method further includes: Obtain the initial image denoising model; The initial image denoising model is configured with batch size, initial learning rate, number of iterations, learning rate decay strategy, and loss function. The loss function is then optimized based on the training sample set to obtain the trained image denoising model. The training sample set includes multiple training samples, each of which is an image of a transmission tower component with added noise, and the corresponding image of the transmission tower component without added noise serves as the sample label.

5. The noise reduction method for images of transmission tower components according to claim 1, characterized in that, The formula for adding speckle noise to the image of this transmission tower component is: Where n is the noise signal, y is the output signal after noise pollution, and x is the input signal, and x and n are independent of each other.

6. The noise reduction method for images of transmission tower components according to claim 1, characterized in that, The addition of speckle noise to the image of the transmission tower component includes: Based on the region where the component is located in the image of the transmission tower component, the noise region corresponding to the image of the transmission tower component is determined, and speckle noise is added to the image of the transmission tower component based on the noise region.

7. A noise reduction device for images of transmission tower components, characterized in that, include: The acquisition module is used to acquire images of the transmission tower components to be denoised; The denoising module is used to input the image of the transmission tower component to be denoised into a trained image denoising model to obtain a denoised image of the transmission tower component; wherein, the image denoising model is an HR-Net network, and each DACNN module in the HR-Net network has been improved with a dual attention mechanism; The image denoising model is trained based on a training sample set; The training module is used for: Acquire multiple images of transmission tower components; For each image of a transmission tower component, speckle noise is added to the image to obtain a noisy transmission tower component image. The noisy transmission tower component image is then used as a training sample and added to the training sample set as a sample label. The training module is specifically used for: Based on the component categories contained in the image of the transmission tower component, the noise signal corresponding to the image of the transmission tower component is determined from different levels of noise signals, and speckle noise is added to the image of the transmission tower component based on the noise signal.

8. The noise reduction device for images of transmission tower components according to claim 7, characterized in that, Each DACNN module includes a convolutional module, a spatial attention module, a channel attention module, and a batch normalization module, which are connected in sequence. The spatial attention module includes a max pooling layer, an average pooling layer, and a convolutional layer. The max pooling layer and the average pooling layer are connected in parallel and then connected in series with the convolutional layer. The channel attention module includes an average pooling layer, a first fully connected layer, and a second fully connected layer connected in series.

Images