Power transmission line icing detection method based on SAF-MSNet

Abstract

The invention discloses a power transmission line icing detection method based on SAF-MSNet. According to the method, the feature extraction capability of the network is enhanced through a dynamic feature adaptive convolution module, a multi-scale feature regulation and control mechanism is introduced to capture irregular icing features, and accurate positioning is realized through a multi-scale target detection layer, so that the recognition precision is remarkably improved. The ice coating degree of the wire can be rapidly and accurately detected, and the problem that the power supply stability and reliability of a power system are poor in complex weather is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a method for detecting ice coating on a power transmission line based on SAF-MSNet. Background Art

[0002] Conductor ice is a common natural phenomenon in transmission line facilities in cold and humid environments. It not only poses a great threat to the safe operation of the power grid, but also has a significant impact on the stability and economy of the power system. Ice will increase the mechanical load on the conductors and towers, which may cause serious accidents such as conductor breakage and tower collapse. In extreme cases, it may even cause large-scale power outages. Therefore, quickly and accurately detecting the degree of conductor ice is crucial to ensure the safe operation of power facilities and reduce maintenance costs.

[0003] Due to the development of computer vision image processing technology, an algorithm for icing detection of power transmission lines based on SAF-MSNet has emerged in recent years. This algorithm first obtains the conductor icing data set through an image acquisition device. By training a large amount of data, it can automatically extract features related to icing from the conductor image, and perform classification and detection. This data-driven method does not rely on complex physical models, which not only improves the real-time performance of detection, but also shows great potential in saving manpower and reducing costs. Summary of the invention

[0004] The purpose of the present invention is to provide a transmission line icing detection method based on SAF-MSNet, which can quickly and accurately detect the degree of conductor icing and solve the problem of poor power supply stability and reliability of the power system under complex weather conditions.

[0005] The technical solution of the present invention is a transmission line icing detection method based on SAF-MSNet, which is specifically implemented according to the following steps:

[0006] Step 1: Collect ice-covered images of transmission lines, pre-process the collected ice-covered images of transmission lines, annotate them and divide them into a training set, a verification set and a test set;

[0007] Step 2: Build a transmission line icing detection network model based on SAF-MSNet;

[0008] Step 3: Input the images in the training set into the network model built in step 2 for training;

[0009] Step 4: Input the images in the test set into the network model trained in step 3 for detection, and obtain the transmission line icing detection result map.

[0010] Furthermore, step 1 is specifically implemented according to the following steps:

[0011] Step 1.1, collect a large number of images through tower cameras and inspection drones, including images of ice and snow covering various transmission lines. Through close-up and close-up images, keep the number of images of ice and snow covering each type of transmission line consistent, and obtain transmission line ice-covered images;

[0012] Step 1.2, performing data enhancement and expansion on the transmission line ice-covered image obtained in step 1.1, including mirroring, inversion, contrast, and brightness processing, so as to obtain an expanded transmission line ice-covered image;

[0013] Step 1.3: The transmission line ice-covered image obtained in step 1.1 and the expanded transmission line ice-covered image obtained in step 1.2 are taken together as a transmission line ice-covered image set, and divided into a training set, a validation set, and a test set according to a ratio of 8:1:1.

[0014] Furthermore, step 2 is specifically implemented according to the following steps:

[0015] Step 2.1, using the C2f architecture (Cascaded Fusion), combining the feature information of the lower and higher layers between multiple network layers to perform feature fusion step by step, and obtain the output feature map;

[0016] Step 2.2: The feature map obtained in step 2.1 is passed through a dynamic snake convolution layer to generate the weight of the convolution kernel. According to the weight output by the feature network, a snake convolution kernel is generated and the convolution kernel is assigned a value.

[0017] The generated snake convolution is moved on the input image or feature map in a sliding window manner to cover every target area. For each area, the weight of the convolution kernel is dot-producted with the input feature image to generate a new feature map.

[0018] The 3×3 snake-shaped convolution kernel K with expansion is expressed as:

[0019] K={(x-1,y-1),(x-1,y),…,(x+1,y+1)}Equation 1

[0020] Formula 1 needs to introduce deformation offset, adopt an iterative strategy, select key positions for each target in turn, and adjust the offset to ensure that the convolution kernel conforms to the linear morphological structure;

[0021]

[0022] And Equation 2 becomes:

[0023]

[0024] Since the offset is usually fractional, bilinear interpolation is implemented as:

[0025] K=∑ K' B(K',K)K' Formula 4

[0026] Where K represents the fractional position of Equation 2 and Equation 3, K' enumerates all integer spatial positions, and B is the bilinear interpolation kernel, which is divided into two one-dimensional kernels:

[0027] B(K,K')=b(K x , K' x )b(K y , K' y ) Formula 5

[0028] Due to the two-dimensional (x, y) changes, the dynamic feature adaptive convolution covers a 9×9 range during the deformation process;

[0029] Step 2.3: Send the new feature map obtained in step 2.2 to the spatial pyramid pooling layer for multi-scale feature aggregation. The Adaptive Multi-Scale Feature Enhancement Mechanism is integrated to extract the horizontal and vertical features of the input feature map and generate a preliminary attention map to help the model focus on the important parts of the image. After obtaining the preliminary attention map, the multi-scale adaptive feature enhancement mechanism uses spatial dilated convolutions with different dilation rates to further extract features. By operating in the horizontal and vertical directions respectively, the image features are processed more carefully and the model's understanding of the spatial relationship in the image is enhanced. After a series of convolution operations, the multi-scale adaptive feature enhancement mechanism fuses the features obtained through the last convolution layer to generate the final attention map A. C ; This attention map will perform element-wise multiplication with the original input feature map, so that each element in the original input feature map will be weighted according to the value of the attention map, highlighting important features and suppressing unimportant features;

[0030] The formula is as follows:

[0031]

[0032]

[0033] A C =W 1×1 *Z C Formula 8

[0034] Where d is the expansion rate, represents the output of a depthwise convolution with a kernel size of (2d-1)×1, which captures local spatial information and compensates for the grid effect of the subsequent depthwise dilated convolution; the size of the convolution kernel is *and Respectively represent convolution and Hadamard product; Z C It is done by combining the convolution kernel W of size k×k with the input feature map F C The output of the deep convolution obtained by convolution, the k in the equation represents the maximum receptive field of the kernel W;

[0035] Step 2.4: Send the feature map obtained in step 2.3 to the positioning and recognition target detection layer, which is mainly composed of two modules: the first module is the convolution layer, which combines group normalization and activation function; group normalization divides the feature map into multiple groups and normalizes them separately; the second module involves creating objects for generating mask prototypes; through the mask prototype, the convolution layer can process the input feature map to obtain mask coefficients, and the mask coefficients and mask prototypes work together to determine the shape characteristics of the ice-covered area, thereby guiding the model to generate a mask for each area instance; the mask coefficients of each detection box are spliced ​​into a mask coefficient matrix, so that the model can return the output, mask coefficients and mask prototypes;

[0036] Step 2.5, combining the characteristics of Wise-IoU and inner-IoU, the Wise-Inner-IoU method is used to evaluate the similarity between the predicted box and the real box to improve the recognition accuracy. The calculation formula of Wise-inner-IoU is:

[0037] L Inner-WIoU =L WIoU +IoU-IoU inner Formula 9

[0038] Wise-inner-IoU uses smoothing in its calculation to improve robustness in the case of slight offsets of the bounding box; adjusts the calculation method of the overlapping area of ​​the bounding box to reduce sensitivity to extreme cases, and assigns different weights to the IoU of different areas; and uses Wise-inner-IoU as the loss function in model training;

[0039] After the above steps, the SAF-MSNet network model is obtained.

[0040] Furthermore, step 3 is specifically implemented according to the following steps:

[0041] Step 3.1: Input the training set images divided in step 2 into the SAF-MSNet network model for training, set batch_size to 4, and perform 300 rounds of stochastic gradient descent SGD training to obtain the training results;

[0042] Step 3.2: Use the validation set to test the SAF-MSNet network model, obtain the validation results, and analyze the differences between the training results and the validation results;

[0043] Step 3.3: Fine-tune the hyperparameters of the SAF-MSNet network on the validation set, fine-tune the learning rate to 0.002, set the batch_size to 4, and obtain the optimized SAF-MSNet network model.

[0044] Furthermore, step 4 is specifically implemented according to the following steps:

[0045] The test set is input into the optimized SAF-MSNet network model obtained in step 3.3 for optimization. The specific process is as follows: SAF-MSNet model hyperparameter setting: input image pixel size 640*640, training 300 rounds, batch_size 4, num_workers8, Adam optimizer, attenuation weight coefficient 5*10-4, initial learning rate 1*10-5, when testing the test set, the IoU threshold is set to 0.5 for experimentation, and when verifying, the learning rate is fine-tuned to 0.002.

[0046] The beneficial effect of the present invention is that conductor icing is a common phenomenon in transmission lines under severe weather, which has a significant impact on the stability and reliability of power supply. In view of the problems of complex background, irregular shape and overlap in the icing area, the recognition accuracy is low. The present invention proposes a transmission line icing detection method based on SAF-MSNet. This method enhances the feature extraction capability of the network through a dynamic feature adaptive convolution module, and introduces a multi-scale feature regulation mechanism to capture irregular icing features, and achieves precise positioning through a multi-scale target detection layer, thereby significantly improving the recognition accuracy. In addition, the Wise-Inner-IoU method is used to evaluate the similarity between the predicted box and the true box to improve the recognition accuracy. Experimental results show that the recognition accuracy of this method reaches 90.5%, which is better than the existing baseline model, and shows important application value in the field of conductor icing recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a schematic diagram of the algorithm in the power transmission line icing detection method based on SAF-MSNet of the present invention;

[0048] Figure 2 It is a schematic diagram of a dynamic feature adaptive convolution module in the SAF-MSNet transmission line icing detection method of the present invention;

[0049] Figure 3 It is a principle diagram of the multi-scale adaptive feature enhancement mechanism in the power transmission line icing detection method based on SAF-MSNet of the present invention;

[0050] Figure 4 This is a schematic diagram of the target detection layer for positioning and identifying in the power transmission line icing detection method based on SAF-MSNet of the present invention. DETAILED DESCRIPTION

[0051] like Figure 1 As shown, this embodiment provides a transmission line icing detection method based on SAF-MSNet, which is specifically implemented according to the following steps:

[0052] Step 1: Collect ice-covered images of transmission lines, pre-process the collected ice-covered images of transmission lines, annotate them and divide them into a training set, a verification set and a test set;

[0053] Step 2: Build a transmission line icing detection network model based on SAF-MSNet;

[0054] Step 3: Input the images in the training set into the network model built in step 2 for training;

[0055] Step 4: Input the images in the test set into the network model trained in step 3 for detection, and obtain the transmission line icing detection result map.

[0056] Step 1 is specifically implemented as follows:

[0057] Step 1.1, collect a large number of images through tower cameras and inspection drones, including images of ice and snow covering various transmission lines. Through close-up and close-up images, keep the number of images of ice and snow covering each type of transmission line almost the same, and obtain a dataset of ice and snow covering the conductors;

[0058] Step 1.2, performing data enhancement and expansion on the image obtained in step 1.1, including mirroring, inversion, contrast, and brightness processing, so as to obtain an expanded transmission line ice-covered image;

[0059] Step 1.3: The transmission line icing images obtained in step 1.1 and the expanded images obtained in step 1.2 are used together as a transmission line icing image library, and divided into a training set, a validation set, and a test set according to a ratio of 8:1:1.

[0060] Step 2 is implemented as follows:

[0061] Step 2.1, according to the formation of C2f (Cascaded Fusion), the model's expressiveness is enhanced by fusing features of different levels step by step, and features are fused step by step between multiple network layers, combining the feature information of low and high layers to obtain the output feature map; this structure can make full use of the detailed features of the low layer and the semantic features of the high layer to enhance the expressiveness of the model. The C2f architecture can effectively process multi-scale inputs, allowing the model to pay attention to different scales of the object at the same time. By gradually fusing features of different scales, C2f improves the robustness of the model to size changes and morphological changes.

[0062] Step 2.2: The feature map obtained in step 2.1 is passed through a dynamic snake convolution layer. The convolution layer improves the high-dimensional features of the input map. The module is used to generate the weight of the convolution kernel, which changes with the input data. According to the weight output by the feature network, a snake convolution kernel is generated and assigned to the convolution kernel. The generated snake convolution is moved on the input image or feature map in a sliding window manner to cover each target area. For each area, the weight of the convolution kernel is dot-producted with the input feature image to generate a new feature map. Since the transmission line is easily deformed after being covered with ice, the difficulty of detection is increased. The retractable 3×3 kernel K is expressed as:

[0063] K={(x-1,y-1),(x-1,y),…,(x+1,y+1)} Formula 1

[0064] In order to make the convolution kernel more flexible to focus on the complex geometric features of the target, we introduced deformation offset. However, if the model is allowed to freely learn deformation offset, the perception field may deviate from the target. Therefore, we adopt an iterative strategy to select key positions for each target in turn to ensure the continuity of attention and avoid excessive diffusion of the perception field due to excessive offset. In dynamic feature adaptive convolution, the standard convolution kernel is straightened on the x-axis and y-axis. Taking the x-axis as an example, assuming the convolution kernel size is 9, each grid position is represented as K i±c =(x i±c ,y i±c ), where c = {0, 1, 2, 3, 4} represents the horizontal distance to the center grid. i±c The selection is cumulative. Starting from the center position K i , the position away from the center grid depends on the previous position: K i+1 The offset is increased by Δ = {δ|δ∈[-1,1]}. Therefore, the offset needs to be adjusted to ensure that the convolution kernel conforms to the linear morphology structure.

[0065]

[0066] And Equation 2 becomes in the y-axis direction:

[0067]

[0068] Since the offset is usually fractional, bilinear interpolation is implemented as:

[0069] K=∑ K' B(K',K)K' Formula 4

[0070] where K represents the fractional position in Equation 2 and Equation 3, K' enumerates all integer spatial positions, and B is the bilinear interpolation kernel, which is split into two one-dimensional kernels:

[0071] B(K,K')=b(K x , K' x )b(K y , K' y ) Formula 5

[0072] Due to the two-dimensional (x, y) changes, our dynamic feature adaptive convolution covers a 9×9 range during the deformation process. Figure 2 As shown, the dynamic feature adaptive convolution is designed to better adapt to and thus better perceive the local characteristics of the wire structure after ice coating.

[0073] Step 2.3: Send the new feature map obtained in step 2.2 to the spatial pyramid pooling layer for aggregating features at multiple scales. The Adaptive Multi-Scale Feature Enhancement Mechanism is incorporated to extract the horizontal and vertical features of the input feature map. Figure 3 As shown in Figure 1, this step is to generate a preliminary attention map to help the model focus on the important parts of the image. After obtaining the preliminary attention map, the multi-scale adaptive feature enhancement mechanism uses spatial dilated convolutions with different dilation rates to further extract features. These convolutional layers can cover a larger receptive field and capture a wider range of contextual information without increasing computational costs. By operating in the horizontal and vertical directions respectively, they can process image features more carefully and enhance the model's understanding of spatial relationships in the image. After a series of convolution operations, the multi-scale adaptive feature enhancement mechanism fuses the features obtained through the last convolutional layer to generate the final attention map A. C This attention map is multiplied element-wise with the original input feature map, i.e., the attention mechanism is applied. In this way, each element in the original feature map is weighted according to the value of the attention map, highlighting important features and suppressing unimportant features.

[0074] Through the fast feature pyramid, features of different scales are captured, and multiple pooling operations are used to capture large, medium and small scale features, so that the model can capture large-scale and small ice-covered area features at the same time, thereby improving the robustness and adaptability of the model under different conditions. The formula is as follows:

[0075]

[0076]

[0077] A C =W 1×1 *Z C Formula 8

[0078]

[0079] Where d is the expansion rate. represents the output of a depthwise convolution with a kernel size of (2d-1)×1, which captures local spatial information and compensates for the grid effect of the subsequent depthwise dilated convolution. The size of the convolution kernel is *and denote convolution and Hadamard product respectively. Z C is the output of the depthwise convolution obtained by convolving a convolution kernel W of size k×k with the input feature map F. Each channel C in F is convolved with its corresponding channel in the following kernel W. The k1 in the equation represents the maximum receptive field of the kernel W.

[0080] Step 2.4: Send the feature map obtained in step 2.3 to the positioning and recognition target detection layer, such as Figure 4 As shown in the figure, the detection layer mainly consists of two modules. The first module is a convolutional layer that combines group normalization and activation functions. Group normalization divides the feature map into multiple groups and normalizes them separately, which is more robust to feature maps of small batch sizes and helps to enhance the expressiveness of the model. The second module involves creating objects for generating mask prototypes. Through the mask prototype, the convolutional layer can process the input feature map to obtain the mask coefficients. The coefficients and prototypes work together to determine the shape characteristics of the ice-covered area, thereby guiding the model to generate a mask for each area instance. The mask coefficients of each detection box are concatenated into a mask coefficient matrix, so that the model can return the output, mask coefficients, and mask prototypes. This module provides pixel-level segmentation results. By generating masks and mask coefficients, the model can accurately segment different object instances in the image and accurately predict the location and shape of the target. This improves the positioning accuracy of the model against the background and can effectively separate and identify overlapping ice-covered areas to achieve the best results.

[0081] Step 2.5: Using Wise-inner-IoU, we combine the characteristics of Wise-IoU and inner-IoU. The calculation formula of Wise-inner-IoU is:

[0082]

[0083] Wise-inner-IoU uses smoothing in its calculations to improve robustness in the case of slight offsets of the bounding box. Adjusts the calculation method of the overlapping area of ​​the bounding box to reduce sensitivity to extreme cases. Gives different weights to the IoU of different areas. Use Wise-inner-IoU as the loss function in model training, replacing the traditional IoU loss and other bounding box regression losses. In practical applications, Wise-inner-IoU can significantly improve the performance of target detection models and improve the detection accuracy of small targets. Improve the accuracy of bounding box prediction in target-dense scenes and improve the robustness to non-rectangular or complex-shaped targets.

[0084] After the above steps, the SAF-MSNet network model is obtained.

[0085] Step 3 is implemented as follows:

[0086] Step 3.1: Input the training set images divided in step 2 into the SAF-MSNet network model for training, set batch_size to 4, and perform 300 rounds of stochastic gradient descent SGD training to obtain the training results;

[0087] Step 3.2: Use the validation set to test the SAF-MSNet network model, obtain the validation results, and analyze the differences between the training results and the validation results;

[0088] Step 3.3: Fine-tune the hyperparameters of the SAF-MSNet network on the validation set, fine-tune the learning rate to 0.002, increase the learning rate to obtain better robustness, set batch_size to 4, and obtain the optimized SAF-MSNet network model.

[0089] Step 4 is implemented according to the following steps:

[0090] The test set is input into the optimized SAF-MSNet network model obtained in step 3.3 for optimization. The specific process is as follows: SAF-MSNet model hyperparameter setting: input image pixel size 640*640, training 300 rounds, batch_size 4, num_workers8, Adam optimizer, attenuation weight coefficient 5*10-4 initial learning rate 1*10-5, when testing the test set, the IoU threshold is set to 0.5 for experimentation, and when verifying, the learning rate is fine-tuned to 0.002.

Claims

1. A transmission line icing detection method based on SAF-MSNet, characterized in that: Follow the steps below to implement it: Step 1: Collect ice-covered images of transmission lines, pre-process the collected ice-covered images of transmission lines, annotate them and divide them into a training set, a verification set and a test set; Step 2: Build a transmission line icing detection network model based on SAF-MSNet; Step 3: Input the images in the training set into the network model built in step 2 for training; Step 4: Input the images in the test set into the network model trained in step 3 for detection, and obtain the transmission line icing detection result map.

2. The method for detecting ice coating on power transmission lines based on SAF-MSNet according to claim 1, characterized in that: The step 1 is specifically implemented according to the following steps: Step 1.1, collect a large number of images through tower cameras and inspection drones, including images of ice and snow covering various transmission lines. Through close-up and close-up images, keep the number of images of ice and snow covering each type of transmission line consistent, and obtain transmission line ice-covered images; Step 1.2, performing data enhancement and expansion on the transmission line ice-covered image obtained in step 1.1, including mirroring, inversion, contrast, and brightness processing, so as to obtain an expanded transmission line ice-covered image; Step 1.3: The transmission line ice-covered image obtained in step 1.1 and the expanded transmission line ice-covered image obtained in step 1.2 are taken together as a transmission line ice-covered image set, and divided into a training set, a validation set, and a test set according to a ratio of 8:1:

1.

3. The method for detecting ice coating on power transmission lines based on SAF-MSNet according to claim 1, characterized in that: The step 2 is specifically implemented according to the following steps: Step 2.1, using the C2f architecture (Cascaded Fusion), combining the feature information of the lower and higher layers between multiple network layers to perform feature fusion step by step, and obtain the output feature map; Step 2.2: The feature map obtained in step 2.1 is passed through a dynamic snake convolution layer to generate the weight of the convolution kernel. According to the weight output by the feature network, a snake convolution kernel is generated and the convolution kernel is assigned a value. The generated snake convolution is moved on the input image or feature map in a sliding window manner to cover every target area. For each area, the weight of the convolution kernel is dot-producted with the input feature image to generate a new feature map. The 3×3 snake-shaped convolution kernel K with expansion is expressed as: K={(x-1,y-1),(x-1,y)…(x+1,y+1)} Formula 1 Formula 1 needs to introduce deformation offset, adopt an iterative strategy, select key positions for each target in turn, and adjust the offset to ensure that the convolution kernel conforms to the linear morphological structure; And Equation 2 becomes: Since the offset is usually fractional, bilinear interpolation is implemented as: K=∑ K' B(K',K)K' Formula 4 Where K represents the fractional position of Equation 2 and Equation 3, K' enumerates all integer spatial positions, and B is the bilinear interpolation kernel, which is divided into two one-dimensional kernels: B(K,K')=b(K x , K' x )b(K y , K' y ) Formula 5 Due to the two-dimensional (x, y) changes, the dynamic feature adaptive convolution covers a 9×9 range during the deformation process; Step 2.3: Send the new feature map obtained in step 2.2 to the spatial pyramid pooling layer for multi-scale feature aggregation. The Adaptive Multi-Scale Feature Enhancement Mechanism is integrated to extract the horizontal and vertical features of the input feature map and generate a preliminary attention map to help the model focus on the important parts of the image. After obtaining the preliminary attention map, the multi-scale adaptive feature enhancement mechanism uses spatial dilated convolutions with different dilation rates to further extract features. By operating in the horizontal and vertical directions respectively, the image features are processed more carefully and the model's understanding of the spatial relationship in the image is enhanced. After a series of convolution operations, the multi-scale adaptive feature enhancement mechanism fuses the features obtained through the last convolution layer to generate the final attention map A. C ; This attention map will perform element-wise multiplication with the original input feature map, so that each element in the original input feature map will be weighted according to the value of the attention map, highlighting important features and suppressing unimportant features; The formula is as follows: Where d is the expansion rate, represents the output of a depthwise convolution with a kernel size of (2d-1)×1, which captures local spatial information and compensates for the grid effect of the subsequent depthwise dilated convolution; the size of the convolution kernel is *and Respectively represent convolution and Hadamard product; Z C It is done by combining the convolution kernel W of size k×k with the input feature map F C The output of the deep convolution obtained by convolution, the k in the equation represents the maximum receptive field of the kernel W; Step 2.4: Send the feature map obtained in step 2.3 to the positioning and recognition target detection layer, which consists of two modules: the first module is a convolutional layer that combines group normalization and activation function; group normalization divides the feature map into multiple groups and normalizes them separately; the second module involves creating objects for generating mask prototypes; through the mask prototype, the convolutional layer can process the input feature map to obtain mask coefficients, and the mask coefficients and mask prototypes work together to determine the shape characteristics of the ice-covered area, thereby guiding the model to generate a mask for each area instance; the mask coefficients of each detection box are spliced ​​into a mask coefficient matrix, so that the model can return the output, mask coefficients and mask prototypes; Step 2.5, combining the characteristics of Wise-IoU and inner-IoU, the Wise-Inner-IoU method is used to evaluate the similarity between the predicted box and the real box to improve the recognition accuracy. The calculation formula of Wise-inner-IoU is: L Inner-WIoU =L WIoU +IoU-IoU inner formula 9 Wise-inner-IoU uses smoothing in its calculation to improve robustness in the case of slight offsets of the bounding box; adjusts the calculation method of the overlapping area of ​​the bounding box to reduce sensitivity to extreme cases, and assigns different weights to the IoU of different areas; and uses Wise-inner-IoU as the loss function in model training; After the above steps, the SAF-MSNet network model is obtained.

4. The method for detecting ice coating on power transmission lines based on SAF-MSNet according to claim 1, characterized in that: The step 3 is specifically implemented according to the following steps: Step 3.1: Input the training set images divided in step 2 into the SAF-MSNet network model for training, set batch_size to 4, and perform 300 rounds of stochastic gradient descent SGD training to obtain the training results; Step 3.2: Use the validation set to test the SAF-MSNet network model, obtain the validation results, and analyze the differences between the training results and the validation results; Step 3.3: Fine-tune the hyperparameters of the SAF-MSNet network on the validation set, fine-tune the learning rate to 0.002, set the batch_size to 4, and obtain the optimized SAF-MSNet network model.

5. The method for detecting ice coating on power transmission lines based on SAF-MSNet according to claim 1, characterized in that: The step 4 is specifically implemented according to the following steps: The test set is input into the optimized SAF-MSNet network model obtained in step 3.3 for optimization. The specific process is as follows: SAF-MSNet model hyperparameter setting: input image pixel size 640*640, training 300 rounds, batch_size 4, num_workers8, Adam optimizer, attenuation weight coefficient 5*10-4, initial learning rate 1*10-5, when testing the test set, the IoU threshold is set to 0.5 for experimentation, and when verifying, the learning rate is fine-tuned to 0.002.

Images