Line icing detection method and device, equipment, storage medium and program product

Abstract

The embodiment of the invention provides a line icing detection method and device, equipment, a storage medium and a program product. The method comprises the following steps: acquiring a line icing data set; constructing an icing detection model comprising at least one multi-dynamic average pooling edge enhancement module and at least one multi-path dynamic deformation convolution attention module; based on the line icing data set and a preset loss function, training an icing detection model to obtain a trained icing detection model; and based on the trained icing detection model and the line image of the to-be-detected line, determining whether the to-be-detected line is iced or not. According to the method, whether the power transmission line is iced or not can be accurately detected.

Description

Technical Field

[0001] The present application relates to the field of electrical technology, and in particular to a line icing detection method, device, equipment, storage medium and program product. Background Art

[0002] Icing on transmission lines is a major threat to the safe operation of power systems. In low-temperature, high-humidity environments, ice on conductors and insulators significantly increases the mechanical load on the lines, leading to conductor galloping, insulator flashover, and even line breakage and tower collapse, seriously threatening grid stability and power supply reliability.

[0003] In existing technologies, deep learning technology can be used to realize automatic identification of ice cover based on target detection and semantic segmentation models based on convolutional neural networks.

[0004] However, existing deep learning technology has obvious limitations in transmission line scenarios and has poor detection accuracy for ice-covered lines. Summary of the Invention

[0005] The embodiments of the present application provide a line icing detection method, device, equipment, storage medium and program product. By constructing an icing detection model, it is possible to accurately detect whether a transmission line is iced in a transmission line application scenario.

[0006] In a first aspect, an embodiment of the present application provides a line icing detection method, comprising:

[0007] Acquire a line ice-covered data set, wherein the line ice-covered data set includes ice-covered line images and non-ice-covered line images;

[0008] Constructing an ice detection model, the ice detection model includes at least one multi-dynamic average pooling edge enhancement module and at least one multi-channel dynamic deformation convolution attention module, the multi-dynamic average pooling edge enhancement module is used to adaptively enhance the ice edge area in the image and dynamically suppress the background interference of the image, and the multi-channel dynamic deformation convolution attention module is used to identify lighting changes, occluded objects and interfering background in the image;

[0009] Based on the line icing dataset and a preset loss function, the icing detection model is trained to obtain a trained icing detection model;

[0010] Based on the trained ice detection model and the line image of the line to be detected, it is determined whether the line to be detected is iced.

[0011] In a second aspect, an embodiment of the present application provides a line icing detection device, comprising:

[0012] An acquisition module, configured to acquire a line ice-covered data set, wherein the line ice-covered data set includes ice-covered line images and non-ice-covered line images;

[0013] A model construction module for constructing an ice detection model, the ice detection model including at least one multi-dynamic average pooling edge enhancement module and at least one multi-channel dynamic deformation convolution attention module, the multi-dynamic average pooling edge enhancement module is used to adaptively enhance the ice edge area in the image and dynamically suppress the background interference of the image, and the multi-channel dynamic deformation convolution attention module is used to identify lighting changes, occluded targets and interfering background in the image;

[0014] A model training module is used to train the ice detection model based on the line ice data set and a preset loss function to obtain a trained ice detection model;

[0015] The detection module is used to determine whether the line to be detected is covered with ice based on the trained ice detection model and the line image of the line to be detected.

[0016] In a third aspect, an embodiment of the present application provides an electronic device, comprising: a memory, a processor;

[0017] The memory stores computer-executable instructions;

[0018] The processor executes the computer-executable instructions stored in the memory, so that the processor executes the above first aspect and / or various possible implementations of the first aspect.

[0019] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, in which computer-executable instructions are stored. When the computer-executable instructions are executed by a processor, they are used to implement the first aspect above and / or various possible implementation methods of the first aspect.

[0020] In a fifth aspect, an embodiment of the present application provides a computer program product, including a computer program, which, when executed by a processor, implements the above first aspect and / or various possible implementation methods of the first aspect.

[0021] The line icing detection method, apparatus, equipment, storage medium and program product provided in the embodiments of the present application construct multiple dynamic average pooling edge enhancement modules and multiple dynamic deformation convolution attention modules in the icing detection model, adaptively enhance the ice edge area in the image through the multiple dynamic average pooling edge enhancement modules, and dynamically suppress the background interference of the image. In addition, the multiple dynamic deformation convolution attention modules are used to identify lighting changes, occluded targets and interfering backgrounds in the image, thereby improving the model's adaptability to iced line detection in transmission line scenarios and achieving efficient, accurate and automated identification of icing conditions on transmission lines. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0023] Figure 1 Schematic diagram of the scenario for ice-covered line identification provided by this application;

[0024] Figure 2 A flow chart of line icing detection provided for this application;

[0025] Figure 3 A schematic diagram of the structural framework of a line icing detection model provided in an embodiment of the present application;

[0026] Figure 4 This is a diagram showing the overall architecture of the transmission line icing model provided in an embodiment of the present application;

[0027] Figure 5 This is a structural diagram of the multi-dynamic average pooling edge enhancement module provided in an embodiment of the present application;

[0028] Figure 6 This is a diagram of the architecture of the multi-channel dynamic deformation convolution attention MDC-Attention module provided in the embodiment of the present application;

[0029] Figure 7 A schematic flow chart of a line icing detection method according to another embodiment of the present application;

[0030] Figure 8 This is a schematic diagram of the structure of the line icing detection device provided in this application;

[0031] Figure 9 This is a schematic diagram of the structure of the electronic device provided in this application.

[0032] The above drawings illustrate specific embodiments of the present application, which will be described in more detail below. These drawings and the textual description are not intended to limit the scope of the present application in any way, but rather to illustrate the concepts of the present application to those skilled in the art by reference to specific embodiments. DETAILED DESCRIPTION

[0033] Exemplary embodiments will be described in detail herein, with examples illustrated in the accompanying drawings. In the following description, when referring to the drawings, identical numerals in different figures represent identical or similar elements, unless otherwise indicated. The embodiments described in the following exemplary embodiments are not intended to represent all embodiments consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the present application, as detailed in the appended claims.

[0034] Icing on transmission lines is a major threat to the safe operation of power systems, seriously threatening grid stability and power supply reliability. Traditional ice detection methods mainly include manual inspections, sensor monitoring, and image analysis technology. Manual inspections rely on experience and judgment, which is inefficient, costly, and difficult to cover complex terrain. Sensor monitoring (such as tension sensors and meteorological sensors) can provide real-time data feedback, but it has problems such as high deployment costs, difficult maintenance, and susceptibility to environmental interference. It can only reflect the status of local lines. In recent years, computer vision technology based on visible light or infrared images has gradually been applied to ice detection. However, it mostly relies on manually designed features (such as edge extraction and color threshold segmentation), is sensitive to lighting changes and background interference, and has difficulty distinguishing ice from other attachments (such as snow and dirt), resulting in high false detection and missed detection rates. Furthermore, with the development of deep learning technology, convolutional neural network-based object detection and semantic segmentation models are gradually being applied to automated ice recognition. However, deep learning technology has significant limitations in transmission line scenarios. On the one hand, ice has diverse forms and irregular distribution, and existing models are insufficiently capable of extracting the features of small ice targets. On the other hand, complex backgrounds (such as mountainous areas and vegetation obstruction) and extreme weather conditions (such as haze and snow) lead to image quality degradation, significantly reducing model generalization performance. Furthermore, existing deep learning technologies often rely on a single data source (such as static images) and lack the integration and analysis of time series information and multimodal data (visible light, infrared, and meteorological parameters), making it difficult to accurately predict the dynamic growth of ice cover.

[0035] To address the above problems, the present application provides a line icing detection solution, which constructs multiple dynamic average pooling edge enhancement modules and multi-channel dynamic deformation convolution attention modules in the model, adaptively enhances the ice edge area in the image through the multiple dynamic average pooling edge enhancement modules, and dynamically suppresses the background interference of the image. In addition, the multi-channel dynamic deformation convolution attention module is used to identify illumination changes, occluded targets and interfering background in the image, which can improve the model's adaptability to iced line detection in the transmission line scenario, and realize efficient, accurate and automatic identification of icing conditions on transmission lines.

[0036] Figure 1 The schematic diagram of the ice-covered line identification scenario provided by this application is as follows: Figure 1 As shown, an image 10 of a certain section of a transmission line can be collected by an image acquisition device and input into a trained ice detection model 11 . The ice detection model performs ice detection on the image to obtain an output image 12 .

[0037] Among them, for the output image 12, if the ice detection model recognizes that there is ice in the transmission line, the ice detection model can predict the ice by the prediction box ( Figure 1 The ice-covered lines are selected (the dotted box in the figure) and the corresponding ice-covered probability is marked near the prediction box.

[0038] The following specific embodiments describe in detail the technical solution of the present application and how the technical solution of the present application solves the above-mentioned technical problems. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0039] Figure 2 This is a flow chart of line ice detection provided by this application. The method can be executed in an electronic device. For example, the electronic device can include a display interface for displaying images. The electronic device also has an input and output interface, and the line image of the line to be detected can be received through the input interface. Figure 2 As shown, the method includes:

[0040] S210: Obtain a line ice cover dataset.

[0041] The line icing dataset includes ice-covered line images and non-ice-covered line images.

[0042] S220: Construct an ice detection model.

[0043] Among them, the ice detection model includes at least one multi-dynamic average pooling edge enhancement module and at least one multi-channel dynamic deformation convolution attention module. The multi-dynamic average pooling edge enhancement module is used to adaptively enhance the ice edge area in the image and dynamically suppress the background interference of the image. The multi-channel dynamic deformation convolution attention module is used to identify the lighting changes, occluded targets and interfering background in the image.

[0044] S230: Based on the line icing dataset and a preset loss function, an icing detection model is trained to obtain a trained icing detection model.

[0045] S240: Determine whether ice is present on the line to be inspected based on the trained ice detection model and the line image of the line to be inspected.

[0046] The method provided in the embodiment of the present application constructs an ice detection model configured with multiple dynamic average pooling edge enhancement modules and multiple dynamic deformation convolution attention modules. On the one hand, the multiple dynamic average pooling edge enhancement modules can effectively enhance the anti-interference ability of ice edge features under extreme weather conditions; on the other hand, the multiple dynamic deformation convolution attention modules can improve the deformation feature extraction accuracy of small target ice and occluded areas, thereby achieving a comprehensive improvement in ice detection accuracy and positioning stability in complex environments.

[0047] For the above step S210:

[0048] First, we can construct an original transmission line icing dataset. For example, we can use a camera (including but not limited to drones, fixed-position high-definition cameras, etc.) to photograph the area where the transmission lines are located and collect images of suspected icing on the transmission lines.

[0049] Then, we annotated the suspected iced transmission line images in the original transmission line icing dataset. This involves using annotation tools to select the suspected iced locations on the transmission lines, determining the coordinates of the top-left and bottom-right vertices of the box, and marking the suspected iced locations with bounding boxes. Each bounding box is then assigned a classification label to indicate the iced status of the transmission line within the box. There are two types of classification labels: one indicating that the transmission line is iced, and the other indicating that the transmission line is not iced.

[0050] Finally, the original transmission line ice image dataset is preprocessed to obtain a preprocessed transmission line ice image dataset, and the preprocessed transmission line ice dataset is divided to facilitate the subsequent training of the transmission line ice detection model based on deep learning. The specific process is as follows: (1) All suspected ice images of the transmission line in the original transmission line ice dataset are adjusted to a uniform size; (2) When the size of the original dataset is limited, each suspected ice image of the transmission line can be enhanced by using data enhancement techniques including but not limited to geometric transformation, color gamut transformation, clarity transformation, noise injection, local erasure, etc. to obtain multiple enhanced suspected ice images of the transmission line and add them to the dataset to ensure that there are sufficient samples in the training, verification and testing stages, thereby enhancing the robustness of the transmission line ice detection model and reducing the model's sensitivity to subtle changes in the suspected ice images of the transmission line. (3) The enhanced transmission line ice dataset is scientifically divided into a training set, a verification set and a test set according to a ratio of 6:2:2 to facilitate the subsequent training of the ice detection model.

[0051] For the above step S220, a transmission line icing detection model based on deep learning is constructed, which can be used to identify suspected icing images of the transmission line, and output the type of icing on the transmission line (specifically divided into two categories: one is that there is icing on the transmission line, and the other is that there is no icing on the transmission line) and the location of the icing.

[0052] Exemplarily, when the Multi-path Dynamic Average Pooling Edge Enhancement Module (MDAPM) in the ice detection model performs image processing, it can specifically include the following processes: first, dynamic average pooling operations are performed on the input feature maps respectively; it uses differentiated multi-path dilation convolution circuits (for example, the dilation rates are 3, 5, and 7 respectively) to obtain multi-scale edge enhancement feature maps, and then generates an attention weight map through channel dimension feature fusion, and then performs adaptive weighted fusion with the input feature map, and finally, adjusts the channel dimension output through 1×1 convolution, thereby achieving adaptive enhancement of the ice edge area in the image and dynamic suppression of the background interference of the image.

[0053] For example, the Multi-path Dynamic Deformable Convolutional Attention Module (MDC-Attention) in the ice detection model can specifically include the following steps when performing image processing: first, the multi-dimensional deformation features of the ice-covered area are captured through parallel multi-scale dynamic deformable convolution paths (for example, 5×5, 7×7, and 9×9 kernels); second, a dynamic step size calculation mechanism based on adaptive convolution kernel size is used to achieve robust perception of ice edge deformation, local occlusion, and ice layer attached to wires; then, channel fusion and spatial compression techniques (for example, global average pooling + adaptive one-dimensional convolution kernel) are used to generate channel-spatial joint attention weights, combined with sigmoid activation and feature reweighting mechanisms to adaptively focus on key ice-covered areas; finally, an adaptive one-dimensional convolution kernel size formula is introduced (dynamically determined by the number of channels c and hyperparameters γ and b) to improve the model's ability to differentiate between different channel features, thereby identifying illumination changes, occluded targets, and interfering backgrounds in the image, providing technical support for ice thickness estimation and safety warning in complex natural environments.

[0054] For the above-mentioned step S230, the loss function may refer to the intersection over union (IoU) loss function used for target detection and image segmentation tasks, which optimizes the model by measuring the degree of overlap between the predicted bounding box (or mask) and the true annotation box. The core is the intersection over union, which is the intersection area of ​​the predicted bounding box (prediction area) and the true annotation box (true area) divided by the union area. The IoU loss function is defined as 1-IoU, with a value range of [0,1]. When the predicted area completely overlaps with the true area, the loss is 0, and when there is no overlap at all, it is 1. Compared with the traditional L1 / L2 loss, the IoU loss function directly optimizes the core evaluation indicators of the detection task, and is scale-invariant to changes in the scale and position of the bounding box.

[0055] Furthermore, considering that the IoU loss function cannot provide a gradient when the prediction has no overlap with the real box, and cannot distinguish the overlapping differences in different directions. For this reason, the IoU loss function can be improved. Specifically, in some embodiments, the loss function can also be the FBIoU loss function, which integrates multi-dimensional geometric constraints on the basis of the IoU loss function: ① Strengthen the spatial alignment ability of the frame through the Euclidean distance penalty term of the center point to solve the positioning deviation caused by the center offset of small targets (such as thin ice layers); ② Introduce the aspect ratio inverse tangent correction factor, and use the angle difference to quantify the shape similarity between the predicted frame and the real frame, effectively responding to the regression requirements of ice-covered targets with variable width and height ratios (such as ice on slender wires); ③ Adaptively balance the overlap and shape optimization through a dynamic weight mechanism to improve the sensitivity of the bounding box to blurred edges in complex backgrounds and occlusion scenes.

[0056] The design of the FBIoU loss function combines optimization of position, shape, and overlap, significantly improving the model's positioning accuracy for diverse ice area sizes, low contrast, and occlusion. This addresses challenges such as variable target size, complex backgrounds, and blurred edges in transmission line ice detection, and improves the model's adaptability to these scenarios.

[0057] During the training process of the icing detection model, it is necessary to train and update the parameters of each layer of the transmission line icing detection model. Specifically, all neural network parameters are initialized and hyperparameters related to the transmission line icing detection model are set. These hyperparameters include, but are not limited to, the number of training rounds, batch size, optimizer selection, learning rate, weight initialization method, and dropout ratio. After initializing the parameters, the training and validation data sets are divided into multiple batches. Each batch of training data is fed into the transmission line icing detection model for training. The difference between the predicted box and the ground-truth box is calculated using the FBIoU loss function to obtain the training loss value T for each batch. After completing a round of training on all batches of the training data, the validation set is fed into the transmission line icing detection model batch by batch, obtaining the corresponding batch loss value, Batch T. The validation set loss value is primarily used to monitor whether the transmission line icing model is overfitting and to adjust the training strategy, such as early termination of training or adjusting the learning rate. During training and validation, the FBIoU loss function is also used to calculate the T and batch T values ​​for the corresponding batches. Automatic learning and parameter adjustment are then performed based on T and batch T. Training of the transmission line ice detection model is completed when the batch T value converges after one or more training rounds.

[0058] Here, batch L refers to the batch training loss, specifically the degree of discrepancy between the model's predictions and the true labels for all samples in a training batch. This discrepancy is typically measured using a loss function, which calculates the error between the predicted values ​​and the true values.

[0059] In this embodiment, with respect to step S240 above, after model training is complete, the trained line icing detection model is applied to identify and analyze the current transmission line data. For example, the coordinates of the center point of the final prediction box on the final output feature map are (x4, y4), the width of the prediction box is w4, and the height is h4. The corresponding probability value near the final prediction box indicates the probability of ice being present at that location. When this probability exceeds a preset threshold, it is determined that ice has been detected on the transmission line.

[0060] The preset threshold value can be set and adjusted according to actual conditions. For example, the preset threshold value is 0.5.

[0061] Figure 3 A schematic diagram of the structural framework of the line icing detection model provided in the embodiment of the present application is shown in FIG. Figure 3 As shown, it includes a first processing unit 31 , a second processing unit 32 , a third processing unit 33 , a first detection head 34 , a second detection head 35 , a third detection head 36 and a merging output module 37 .

[0062] Among them, the first processing unit 31 and the second processing unit 32 include multiple dynamic average pooling edge enhancement modules, and the third processing unit 33 includes multiple dynamic deformation convolution attention modules.

[0063] The first processing unit 31 is used to process the input image of the ice detection model to obtain at least three ice shallow layer feature maps. The second processing unit 32 is used to process the at least three ice shallow layer feature maps to obtain at least three ice middle layer feature maps.

[0064] The third processing unit 33 is used to process at least three ice-covered middle layer characteristic maps to obtain at least three ice-covered deep layer characteristic maps.

[0065] Each detection head is used to obtain at least one deep ice feature map for regression, convolution, and classification. The merge output module is used to merge the images output by each detection head to obtain a target image as the output feature map. The output feature map is used to determine whether the track in the input image is covered with ice.

[0066] Specifically, Figure 4 The overall architecture diagram of the transmission line icing model provided in the embodiment of the present application is as follows: Figure 4 As shown, the first processing unit 31 includes four multi-dynamic average pooling edge enhancement modules (i.e. Figure 4 The first MDPAM module, the second MDPAM module, the third MDPAM module and the fourth MDPAM module in the first processing unit 31 are as follows:

[0067] (1) Taking any suspected ice-covered image F1 of the transmission line as the input feature map of the transmission line ice-covered detection model, the first and second combined layers are used to perform convolution, batch normalization and activation function operations on F1 to obtain the first ice-covered shallow layer feature map F2;

[0068] (2) Apply the second second combination layer to perform convolution, batch normalization and activation function operations on F2 to obtain the second ice shallow layer feature map F3;

[0069] (3) Apply the first MDAPM module to process F3 to obtain the third ice shallow layer feature map F4;

[0070] (4) Apply the first combined layer to perform convolution, batch normalization and activation function operations on F4 to obtain the fourth ice shallow layer feature map F5;

[0071] (5) Input F5 into the second MDAPM module for processing to obtain the fifth ice shallow layer feature map F6;

[0072] (6) Apply the first combination layer to perform convolution, batch normalization and activation function operations on F6 to obtain the sixth ice shallow layer feature map F7;

[0073] (7) Input F7 into the third MDAPM module for processing to obtain the seventh ice-covered shallow layer feature map F8;

[0074] (8) Apply the first combination layer to perform convolution, batch normalization and activation function operations on F8 to obtain the eighth ice shallow layer feature map F9;

[0075] (9) Input F9 into the fourth MDAPM module for processing to obtain the ninth ice-covered shallow layer characteristic map F10.

[0076] Among them, the above-mentioned first combination layer and second combination layer can refer to the Conv_BN_ReLU module. In addition, the third combination layer, the fourth combination layer and other combination layers mentioned later can all refer to the Conv_BN_ReLU module. Specifically, the Conv_BN_ReLU module is an existing standard neural network unit composed of a convolution kernel size of 3×3 and a step size of 2, a batch normalization (BN) layer and a ReLU activation function layer connected in series. Its workflow can be divided into three stages: first, the convolution layer extracts the underlying features (such as edges, textures and geometric shapes) in the image through local perception and weight sharing mechanisms; then, the batch normalization (BN) layer standardizes the feature data, and by adjusting the mean and variance of the data distribution, it effectively improves the model training convergence speed and alleviates the gradient anomaly problem; finally, the ReLU activation function performs a nonlinear transformation on the feature map, suppressing negative responses while retaining positive features, thereby enhancing the model's ability to express complex feature patterns.

[0077] Further, Figure 5 The structure diagram of the multi-dynamic average pooling edge enhancement module provided in the embodiment of this application is as follows Figure 5 As shown, the multi-dynamic average pooling edge enhancement module includes a shallow expansion pooling feature extraction route La, a medium expansion pooling feature extraction route Lb and a deep expansion pooling feature extraction route Lc.

[0078] Figure 5 The multi-dynamic average pooling edge enhancement module in can refer to Figure 4 The structure and operation process of any MDAPM module in are the same. For example, assume Figure 5The multi-dynamic average pooling edge enhancement module in Figure 4 The first MDAPM module in Figure 5 The operation and construction process of the multi-dynamic average pooling edge enhancement module in is described. Figure 5 As shown, the original Figure 4 The second shallow ice feature map F3 in Figure 5 The input feature map of the multi-dynamic average pooling edge enhancement module. For the convenience of description, the second ice-covered shallow layer feature map F3 is recorded as M0. The above-mentioned shallow expansion pooling feature extraction route La, medium expansion pooling feature extraction route Lb and deep expansion pooling feature extraction route Lc will all process M0, as follows:

[0079] In the shallow expansion pooling feature extraction route La, first, the dynamic average pooling block is used to perform dynamic pooling operation on M0, where for M0 (assuming the size of M0 H I *W I Pixel, H I is the height of M0, W I The first multi-dynamic average pooling edge enhancement map M1 is obtained by the dynamic pooling operation of the dynamic average pooling block (the size of M1 is H0*W0 pixels, H0 is the height of M1, and W0 is the width of M1). Then, a dilated convolution with a dilation rate of 3 is performed on M1 to obtain the second multi-dynamic average pooling edge enhancement map M2.

[0080] In the dilated pooling feature extraction route Lb, first, a dynamic pooling operation is performed on M0 using the dynamic average pooling block. The dynamic pooling operation of the dynamic average pooling block on M0 obtains a third multi-dynamic average pooling edge enhancement map M3. The dynamic pooling operation process is consistent with the processing route La. Then, a dilated convolution with a dilation rate of 5 is performed on M3 to obtain a fourth multi-dynamic average pooling edge enhancement map M4.

[0081] In the deep dilated pooling feature extraction route Lc, a dynamic pooling operation is performed on M0 using the dynamic average pooling block. The dynamic pooling operation of the dynamic average pooling block on M0 results in a fifth multi-dynamic average pooling edge enhancement map M5. The dynamic pooling operation process is consistent with the processing route La. Then, a dilated convolution with a dilation rate of 7 is performed on M5 to obtain a sixth multi-dynamic average pooling edge enhancement map M6.

[0082] In some embodiments, performing the dynamic pooling operation is specifically:

[0083]

[0084] In the above formula, M0 is the input feature map, M1 is the image after the dynamic pooling operation of M0, and Mi，j Represents the element value i∈[0,H0],j∈[0,W0],m with coordinates (i, j) in M1 p,q Represents the element value with coordinates (p, q) in M0, p∈[0,H1],q∈[0,W1],H t Denotes the height of the dynamic average pooling block of size t, W t Denotes the width of the dynamic average pooling block of size t, H s Indicates the sliding step size s of the dynamic average pooling block in the vertical direction, and the distance the window moves vertically each time, W s Represents the sliding step size s of the dynamic average pooling block in the horizontal direction, and the distance the window moves horizontally each time.

[0085] In the above formula, for example, for the shallow dilation pooling feature extraction route La, the sizes of t and s are both set to 32. For the medium dilation pooling feature extraction route Lb, the sizes of t and s are both set to 32. For the deep dilation pooling feature extraction route Lc, the sizes of t and s are both set to 32.

[0086] Continue to refer to the above Figure 5 After obtaining M2, M4 and M6, M2, M4 and M6 can be fused in the channel dimension through the first fusion layer to obtain the seventh multi-dynamic average pooling edge enhancement map M7. Then, the seventh multi-dynamic average pooling edge enhancement map M7 is further subjected to the Sigmoid activation function to obtain the weight ω of each channel, and then the weight ω of each channel is obtained through Figure 5 The first weighted layer in , the weight ω is multiplied by M0 element by element to obtain the eighth multi-dynamic average pooling edge enhancement map M8. Finally, the convolution layer (such as a normal convolution of size 1×1) is used to adjust the number of channels of M8 to obtain the ninth multi-dynamic average pooling edge enhancement map M9, that is, the original Figure 4 Output feature map of the first MDAPM module in .

[0087] In this embodiment, the multi-dynamic average pooling edge enhancement module uses a multi-path dynamic average pooling and dilated convolution collaborative enhancement method, through three-way parallel processing La, Lb, and Lc (dynamic pooling window parameters t, s and dilation rates 3, 5, and 7 are set respectively), and at the same time uses the dynamic average pooling block. Through the adaptive combination of adjustable window size t and sliding step size s, key edge information is retained and illumination changes and noise interference are suppressed during the pooling process. The weight map is generated by channel cascade of multi-path output features (M2, M4, M6), combined with the element-by-element product of the Sigmoid activation function and the original feature map (M0), to achieve adaptive enhancement of the ice edge area and dynamic suppression of background interference.

[0088] In some embodiments, continue to refer to the above Figure 4The second processing unit 32 includes two multi-dynamic average pooling edge enhancement modules (i.e. Figure 4 The fifth MDPAM module and the sixth MDPAM module in the spatial pyramid pooling fast (SPPF) module, two upsampling modules (i.e. Figure 4 The first upsampling module and the second upsampling module in the ) and two fusion layers (i.e. Figure 4 The second and third fusion layers in ).

[0089] The fifth MDPAM module, as the first multi-dynamic average pooling edge enhancement module, processes the ninth ice shallow layer feature map F10 output by the fourth MDPAM module in the first processing unit to obtain the first ice middle layer feature map F11. The sixth MDPAM module serves as the next multi-modal average pooling edge enhancement module. The operation and construction process of the second processing unit 32 is as follows:

[0090] (1) applying the fifth MDAPM module to process the ninth ice shallow layer feature map F10 output by the fourth MDPAM module in the first processing unit to obtain the first ice middle layer feature map F11;

[0091] (2) Apply SPPF to process F11 and obtain the second ice-covered middle layer feature map F12;

[0092] (3) performing an upsampling operation through the first upsampling module to obtain a third ice-covered middle layer feature map F13;

[0093] (4) After the second fusion layer, the seventh ice shallow layer feature map F8 and F13 output by the third MDAPM module in the original first processing module 31 are fused in the channel dimension to obtain the fourth ice middle layer feature map F14;

[0094] (5) Apply the sixth MDAPM module to process F14 to obtain the fifth ice-covered middle layer characteristic map F15;

[0095] (6) performing an upsampling operation on F15 through a second upsampling module to obtain a sixth ice-covered middle layer feature map F16;

[0096] (7) After the third fusion layer, the fifth ice shallow layer feature map F6 and F16 output by the second MDAPM module in the original first processing module 31 are fused in the channel dimension to obtain the seventh ice middle layer feature map F17, which is input to the third processing unit 33.

[0097] In this embodiment, SPPF is an optimization module introduced in YOLOv8. SPPF uses a fast spatial pyramid pooling method to fuse global information of different scales to improve the performance of target detection, while reducing computational redundancy, achieving higher efficiency and lower floating-point operations.

[0098] In an embodiment of the present application, through a multi-dynamic average pooling edge enhancement module, using shallow, medium and deep three-layer expansion pooling feature extraction routes (expansion rates are 3, 5 and 7 respectively), and combining dynamic average pooling with channel fusion mechanism, the anti-interference ability of ice edge features under extreme weather conditions can be effectively enhanced, thereby improving the detection accuracy of the ice detection model.

[0099] In other embodiments, continue to refer to the above Figure 4 The third processing unit 33 specifically includes three multi-channel dynamic deformation convolution attention modules (i.e. Figure 4 The operation and construction process of the third processing unit 33 is as follows:

[0100] (1) Input the seventh ice-covered middle layer feature map F17 into the first MDC-Attention module to obtain the first ice-covered deep layer feature map F18;

[0101] (2) Apply the third combination layer to perform convolution, batch normalization and activation function operations on F18 to obtain the second ice cover deep layer feature map F19;

[0102] (3) Using the fourth fusion layer, F19 is fused with the fifth ice middle layer feature map F15 output by the sixth MDAPM module of the original second processing unit 32 in the channel dimension to obtain the third ice deep layer feature map F20;

[0103] (4) Apply the second MDC-Attention module to process F20 to obtain the fourth ice deep layer feature map F21;

[0104] (5) Apply the fourth combination layer to perform convolution, batch normalization and activation function operations on F21 to obtain the fifth ice deep layer feature map F22;

[0105] (6) Using the fifth fusion layer, the second ice-covered middle layer feature maps F12 and F22 output by the SPPF module of the original second processing unit 32 are fused in the channel dimension to obtain the sixth ice-covered deep layer feature map F23;

[0106] (7) Apply the third MDC-Attention module to process F23 and obtain the seventh ice deep layer feature map F24.

[0107] Figure 6 This is a diagram of the architecture of the multi-channel dynamic deformation convolution attention MDC-Attention module provided in the embodiment of the present application. The multi-channel dynamic deformation convolution attention module MDC-Attention can Figure 4 Any MDC-Attention module in Figure 6 As shown in FIG, the MDC-Attention module includes the first dynamic deformation convolution, the second dynamic deformation convolution and the third dynamic deformation convolution, the sixth fusion layer, the fifth combination layer and the second weighted layer. Figure 4 The seventh ice-covered mid-layer feature map F17 is Figure 6 The input feature map of the MDC-Attention module in

[15] is shown in Figure 1. For ease of description, F17 is denoted as Q0. The operation and construction process of the MDC-Attention module is as follows:

[0108] (1) Convolution operation is performed on Q0 using the first dynamic deformation convolution, the second dynamic deformation convolution and the third dynamic deformation convolution to obtain the first dynamic deformation convolution attention map Q1, the second dynamic deformation convolution attention map Q2 and the third dynamic deformation convolution attention map Q3 respectively;

[0109] (2) Using the sixth fusion layer, Q1, Q2, and Q3 are fused in the channel dimension to obtain the fourth dynamic deformation convolution attention map Q4;

[0110] (3) Using the global average pooling in the fifth combination layer, the spatial feature compression of the fourth dynamic deformation convolution attention map Q4 is performed, that is, the pooling operation is used in the spatial dimension to obtain a one-dimensional vector of length C;

[0111] (4) Using the convolutional layer in the fifth combined layer (i.e. Figure 6 The one-dimensional vector obtained by Convld k*k) in the fifth combination layer is subjected to one-dimensional convolution and the weight ω of each channel is obtained by the Sigmoid activation function in the fifth combination layer;

[0112] (5) Using the second valence full layer, the weight ω is element-wise multiplied with Q4 to obtain the fifth dynamic deformation convolution attention map Q5, which is used as the output image of the MDC-Attention module.

[0113] In some embodiments, the sizes of the first, second, and third dynamic deformable convolutions are 5×5, 7×7, and 9×9, respectively. The principle of the first, second, and third dynamic deformable convolutions is to obtain the initial sampling coordinates based on the size of the convolution kernel and dynamically determine the step size B of the dynamic deformable convolution. The specific formula is as follows:

[0114]

[0115] In the above formula, t represents the size of the convolution kernel, d represents the result of rounding down the square root of t; row represents the result of rounding down the value of t divided by d; B is the step size of the dynamic deformation convolution during the convolution process; [j] odd represents the nearest odd number of j, c represents the number of channels of the output feature map, and β and θ are both hyperparameters. For example, β and θ are set to 2 and 1 respectively.

[0116] In the above formula, d, row, and M represent the initial sampling coordinates calculated based on the value of t. First, a tensor is generated based on the values ​​of d and row, which serve as the standard convolution kernel sampling coordinates. This tensor is then flattened into a one-dimensional vector. The M value defines the incremental step length, and all tensors are sequentially flattened into one-dimensional vectors. The irregular convolution kernel sampling coordinates are then determined, and the complete sampling coordinates are finally output.

[0117] In some embodiments, the process of fusing Q1, Q2, and Q3 in the channel dimension using the sixth fusion layer to obtain the fourth dynamically deformable convolutional attention map Q4 can be specifically implemented by the following formula:

[0118] Q5=ω⊙Q4=S(C1D k (GlobalAvgPool(x)))⊙Q4

[0119]

[0120] Q5 is the image output by the multi-channel dynamic deformation convolution attention module, Q4 is the fourth dynamic deformation convolution attention map, GlobalAvgPool represents the global average pooling; C1D k represents a one-dimensional convolution with a kernel size of k; S represents the activation function; ⊙ represents the element-wise product; [t] odd represents the nearest odd integer to t, where t represents the convolution kernel size, c represents the number of channels of the image output by the multi-way dynamic deformation convolution attention module, and γ and b are hyperparameters. For example, γ and b are set to 2 and 1, respectively.

[0121] In an embodiment of the present application, an MDC-Attention module based on a multi-path dynamic deformation convolution attention mechanism is constructed, and ice deformation features are extracted in parallel through three groups of dynamic deformation convolutions of 5×5, 7×7, and 9×9. Combined with the calculation formula of the dynamic step size B (based on the convolution kernel size t, the number of channels c and the hyperparameters β and θ to generate irregular sampling coordinates), differentiated feature modeling of ice edge deformation, wire occlusion area and background interference is achieved. Combined with global average pooling and Sigmoid activation function, a channel-space attention weight map is generated to achieve dynamic focusing and noise suppression of key areas of the feature map. In addition, through the joint optimization mechanism of multi-path dynamic deformation convolution feature fusion (channel cascade of Q1, Q2, Q3), spatial compression and attention re-addition, the sampling offset and feature loss problems caused by fixed convolution kernels in traditional methods are solved, and the characterization capability and detection robustness of ice deformation features under complex backgrounds are significantly improved. In addition, by using the MDC-Attention module, through the sampling coordinate generation technology of dynamic deformation convolution, the multi-scale deformation feature fusion method and the adaptive attention guidance mechanism, high-precision distinction between illumination changes, target occlusion and wire / tower background interference is achieved in transmission line icing detection, providing support for ice thickness estimation and safety warning in complex natural environments.

[0122] After introducing the above Figure 4 After the first processing unit 31, the second processing unit 33 and the third processing unit 33, the following continues Figure 4 The detection head and merge output module in the . The specific operation process is as follows:

[0123] (1) The first detection head can perform regression convolution operation and classification convolution operation on F18 to obtain the eighth ice cover deep layer feature map F25.

[0124] Among them, the role of applying the regression convolution operation is to determine the geometric parameters of the prediction box covering the transmission line area on F25. The prediction box geometric parameters include the coordinates of the center point of the prediction box (x1, y1), the width w1 and the height h1 of the prediction box; the role of applying the classification convolution operation is to calculate the probability value of the existence of ice-covered targets in the prediction box area in F18 and mark the probability value near the prediction box.

[0125] The classification convolution establishes a category-sensitive feature response in the channel dimension by analyzing the spatial correlation of feature maps. In this embodiment, the classification convolution uses a 1×1 convolution kernel to achieve dimensionality reduction of high-dimensional features. It combines this with a Sigmoid function to output independent probabilities for multiple categories, rather than the traditional Softmax mutually exclusive classification model. This structure uses an attention mechanism to strengthen the weights of key channels, targeting the unique texture characteristics of ice-covered transmission lines, effectively improving the ability to filter out category false detections.

[0126] The regression convolution in the YOLOv8 detection head uses multi-level convolution kernels for feature encoding, converting the geometric representation of ice-covered transmission lines into numerical outputs. This regressive convolution employs channel dimension parameter reorganization techniques to generate a four-dimensional vector containing the target center point (normalized coordinates based on grid cells), width, and height (scale coefficients relative to the original image size). These values ​​are constrained to the (0, 1) interval using a sigmoid function, establishing a differentiable mapping relationship with the image space coordinate system, providing the geometric transformation foundation for subsequent decoder decoding.

[0127] The detection head, serving as the final decision layer, simultaneously analyzes the spatial semantics of feature maps through a parallel dual-branch architecture (regression / classification). Its output tensor interweaves multiple candidate box parameters for each grid cell in the channel dimension, including five-dimensional geometric parameters (x, y, w, h, confidence) and an N-dimensional class probability vector. This multi-scale prediction mechanism achieves dense coverage prediction for ice-covered objects of varying sizes through hierarchical feature fusion.

[0128] (2) The second detection head of the transmission line icing detection model is used to perform regression convolution operations and classification convolution operations on F21 to obtain the ninth icing deep feature map F26.

[0129] Among them, the role of applying the regression convolution operation is to determine the geometric parameters of the prediction box covering the transmission line area on F26. The prediction box geometric parameters include the coordinates of the center point of the prediction box (x2, y2), the width w2 and the height h2 of the prediction box; the role of applying the classification convolution operation is to calculate the probability value of the existence of ice-covered targets in the prediction box area in F21 and mark the probability value near the prediction box.

[0130] (3) The third detection head of the transmission line icing detection model is used to perform regression convolution operations and classification convolution operations on F24 to obtain the tenth icing deep feature map F27.

[0131] Among them, the role of applying the regression convolution operation is to determine the geometric parameters of the prediction box covering the transmission line area on F26. The prediction box geometric parameters include the coordinates of the center point of the prediction box (x3, y3), the width w3 and the height h3 of the prediction box. The role of applying the classification convolution operation is to calculate the probability value of the existence of ice-covered targets in the prediction box area in F24 and mark the probability value near the prediction box.

[0132] (4) Using the merge output module, F25, F26 and F27 are merged, and each prediction box on F25, F26 and F27 is selected. The selection operation uses the non-maximum suppression algorithm to screen out the prediction box with the highest probability value and no significant overlap based on the probability value of the prediction box and the intersection-over-union ratio threshold. The final prediction box with the highest probability value and no significant overlap is then marked on the merged result of F25, F26 and F27 to obtain the eleventh ice cover deep layer feature map F28.

[0133] Among them, the Non-Maximum Suppression (NMS) algorithm is a post-processing technique for target detection, mainly used to eliminate redundant prediction boxes that overlap a large number of times around the same target. Its core idea is: for all prediction boxes of the same category, sort them from high to low by confidence, use the box with the highest confidence as the benchmark, and calculate its intersection over union (IoU, a measure of the degree of overlap between two boxes) with the remaining boxes. If the IoU of a box and the benchmark box exceeds a preset threshold, it is considered that they are detecting the same target, thereby suppressing redundant boxes with lower confidence. This process is iterated round by round, and the box with the highest confidence among the remaining boxes is selected in turn, and its overlapping boxes are suppressed until all boxes are processed. NMS can effectively solve the problem of the detection model repeatedly predicting the same target multiple times, ensuring that the output results are concise and accurate. It is an indispensable key step in target detection algorithms such as YOLO.

[0134] Bounding box prediction involves associating a preset number of anchor box templates with each feature map grid cell. The regression branch then predicts coordinate corrections for each anchor box: the center offset Δx / Δy is constrained by a sigmoid to achieve precise positioning within the grid, and the width and height scaling factors are converted using an exponential function to maintain the numerical stability of the size prediction. This mechanism significantly improves the positioning accuracy of small insulating clothing targets by decoupling coordinate prediction from prior knowledge of anchor boxes.

[0135] (5) The coordinates of the center point of the prediction box in F28 are (x4, y4), the width of the prediction box is w4, and the height is h4. The probability value corresponding to the vicinity of the final prediction box indicates the probability of ice coverage at that location. When the probability exceeds the preset threshold, it is determined that the transmission line is covered with ice.

[0136] In addition, as mentioned above, the preset loss function can be the FBIoU loss function. In some embodiments, the FBIoU loss function can be used to optimize the position, size, confidence and accuracy of the category prediction of the bounding box. By using the FBIoU loss function to accurately calculate the difference between the predicted box and the true box, the training of the ice detection model is finally realized.

[0137] Among them, the FBIoU loss function is:

[0138]

[0139] In the above formula, B pred Indicates the prediction box of ice-covered roads in the output image of the ice detection model during training, B gt represents the real box, IoU represents the intersection-over-union ratio between the predicted box and the real box; v is the correction factor for the difference in the inverse tangent of the aspect ratio; w G and h G Represents the width and height of the real box respectively; w prod and h prod Represent the width and height of the prediction box respectively; b and b gt They represent the center points of the predicted box and the true box respectively; ρ(·) represents the Euclidean distance between two pixels.

[0140] The confidence prediction value reflects the probability of a valid target within the current anchor box. After element-by-element multiplication with the classification probability matrix, the final detection confidence is generated. By setting confidence and category prediction, the task of determining target presence is decoupled from the fine-grained classification task, effectively alleviating the problem of misidentification of ice-covered transmission lines and backgrounds in power transmission line scenarios.

[0141] The preset thresholds refer to the classification confidence threshold and the target intersection-over-union (IoU) threshold. Specifically, the classification confidence threshold (such as 0.25 or 0.5) is used to filter low-confidence prediction boxes: a prediction box is considered valid only when the ice category prediction score at a certain location (for example, the probability value corresponding to the "ice cover presence" category) is higher than the threshold; and the IoU threshold (such as 0.45) is used in the non-maximum suppression stage. When the IoU of two prediction boxes exceeds this threshold, only the prediction box with higher confidence is retained. These two thresholds need to be adjusted through the experimental validation set to balance detection accuracy (reducing missed detections) and false alarm rate (reducing false detections).

[0142] In an embodiment of the present application, the bounding box regression error problem caused by variable target size, complex background, and blurred ice edges in transmission line ice detection is addressed by employing a loss function, FBIoU. Compared to the traditional intersection-over-union (IoU) function, the FBIoU function strengthens the spatial alignment of the predicted box and the ground-truth box by introducing the Euclidean distance of the center points. A correction factor v, based on the inverse tangent difference of the aspect ratio, is designed to quantify the shape similarity between the predicted box and the ground-truth box. By jointly optimizing IoU, center offset, aspect ratio matching, and a dynamic adjustment mechanism, FBIoU significantly improves the bounding box regression accuracy of ice-covered areas under complex backgrounds, addressing the traditional loss function's insensitivity to box positioning in thin ice layers, small targets, and occluded scenes.

[0143] Figure 7 A schematic flow chart of a line icing detection method according to another embodiment of the present application is shown in FIG. Figure 7 As shown, it includes the following steps:

[0144] S710, constructing an original transmission line ice coverage dataset;

[0145] S720: Preprocess the original data set of transmission line icing to obtain a processed data set of transmission line icing, and divide the data set of transmission line icing;

[0146] S730: Construct a transmission line icing detection model, insert the transmission line image into the model, and output a determination result of the transmission line icing condition;

[0147] S740: Training a transmission line icing detection model based on the transmission line icing dataset to obtain a trained transmission line icing detection model;

[0148] S750: Apply the trained transmission line icing detection model to determine the current icing situation on the transmission line.

[0149] Regarding the above steps S710 to S750, the specific process is described below using distance:

[0150] (1) All suspected ice-covered transmission line images in the original transmission line ice-covered dataset are resized to a uniform size of H × W × C, where H is the height of the feature map, W is the width of the feature map, and C is the number of channels in the feature map. For example, H is 512, W is 512, and C is 3.

[0151] (2) The suspected ice-covered image F1 of the transmission line is used as the input of the ice-covered detection model of the transmission line. The size of F1 is 512×512×3. First, the Conv_BN_ReLU module is applied to process F1 to obtain the first shallow ice feature map F2. The size of F2 is 512×512×64. Then, the Conv_BN_ReLU module is applied to perform convolution, batch normalization and activation function operations on F2 to obtain the second shallow ice feature map F3. The size of F3 is 128×128×128. After that, the first MDAPM module is applied to process F3 to obtain the third shallow ice feature map F4. The size of F4 is 128×128×256. Secondly, the Conv_BN_ReLU module is applied to process F5 to obtain the fourth shallow ice feature map F5. The size of F5 is 64×64×512. F5 is input to the second MDAPM module for processing to obtain the fifth shallow ice feature map F5. The sixth shallow ice feature map F7 is obtained by convolution, batch normalization and activation function operation on F6, and the size of F7 is 32×32×512. F7 is input to the third MDAPM module for processing to obtain the seventh shallow ice feature map F8, and the size of F8 is 32×32×512. Finally, the Conv_BN_ReLU module is applied to F8 for convolution, batch normalization and activation function operation to obtain the eighth shallow ice feature map F9, and the size of F9 is 16×16×512. F9 is input to the fourth MDAPM module for processing to obtain the ninth shallow ice feature map F10, and the size of F10 is 16×16×512.

[0152] (2) Apply the fifth MDAPM module to process F10 to obtain the first ice-covered middle layer feature map F11, the size of which is 16×16×512; secondly, apply SPPF to process F11 to obtain the second ice-covered middle layer feature map F12, the size of which is 16×16×512, and then perform upsampling operation (i.e. Figure 2 The upsample operation in the icing module is performed to obtain the third ice-covered middle layer feature map F13, and the size of F13 is 32×32×512; then, F8 and F13 are fused in the channel dimension to obtain the fourth ice-covered middle layer feature map F14, and the size of F14 is 32×32×1024; then, the sixth MDAPM module is applied to process F14 to obtain the fifth ice-covered middle layer feature map F15, and the size of F15 is 32×32×1024, and F15 is upsampled to obtain the sixth ice-covered middle layer feature map F16, and the size of F16 is 64×64×512; finally, F6 and F16 are fused in the channel dimension to obtain the seventh ice-covered middle layer feature map F17, and the size of F17 is 64×64×1024.

[0153] (3) Input F17 into the first MDC-Attention module to obtain the first ice-covered deep feature map F18, the size of F18 is 64×64×512; secondly, apply the Conv_BN_ReLU module to perform convolution, batch normalization and activation function operations on F18 to obtain the second ice-covered deep feature map F19, the size of F19 is 32×32×512, and F19 is fused with F15 in the channel dimension to obtain the third ice-covered deep feature map F20, the size of F20 is 32×32×1536; then, apply the second MDC-Attention module to process F20, and obtain To the fourth ice-covered deep feature map F21, the size of F21 is 32×32×512, and the Conv_BN_ReLU module is applied to F21 for convolution, batch normalization and activation function operations to obtain the fifth ice-covered deep feature map F22, the size of F22 is 16×16×512; then, F12 and F22 are fused in the channel dimension to obtain the sixth ice-covered deep feature map F23, the size of F23 is 16×16×1024; finally, the third MDC-Attention module is applied to process F23 to obtain the seventh ice-covered deep feature map F24, the size of F24 is 16×16×512.

[0154] (4) Using the first detection head to perform regression convolution and classification convolution operations on F18, the eighth ice cover deep layer feature map F25 is obtained; using the second detection head to perform regression convolution and classification convolution operations on F21, the ninth ice cover deep layer feature map F26 is obtained; using the third detection head to perform regression convolution and classification convolution operations on F24, the tenth ice cover deep layer feature map F27 is obtained. Merge F25, F26, and F27, and merge F25, F26, and F27, and perform a selection operation on each prediction box on F25, F26, and F27. The selection operation is to use the non-maximum suppression algorithm NMS to filter out the prediction box with the highest probability value and no significant overlap, and then mark the final prediction box with the highest probability value and no significant overlap on the merged result of F25, F26, and F27 to obtain the eleventh ice cover deep layer feature map F28.

[0155] (5) The coordinates of the center point of the final prediction box in F28 are (x4, y4), the width of the prediction box is w4, and the height is h4. The corresponding probability value near the final prediction box is the probability that ice is present at that location. The preset threshold is set to 0.5. When the probability exceeds 0.5, it is determined that ice is present on the transmission line.

[0156] Figure 8 The schematic diagram of the line ice detection device provided in this application is as follows: Figure 8As shown, the line icing detection device 80 provided in this embodiment includes: an acquisition module 810 , a model construction module 820 , a model training module 830 and a detection module 840 .

[0157] The acquisition module 810 is used to acquire a line ice dataset, which includes ice-covered line images and non-ice-covered line images. The model construction module 820 is used to construct an ice detection model.

[0158] Among them, the ice detection model includes at least one multi-dynamic average pooling edge enhancement module and at least one multi-channel dynamic deformation convolution attention module. The multi-dynamic average pooling edge enhancement module is used to adaptively enhance the ice edge area in the image and dynamically suppress the background interference of the image. The multi-channel dynamic deformation convolution attention module is used to identify lighting changes, occluded targets and interfering backgrounds in the image;

[0159] Model training module 830 is used to train an ice detection model based on the line ice dataset and a preset loss function, thereby obtaining a trained ice detection model. Detection module 840 is used to determine whether ice is present on the line to be inspected based on the trained ice detection model and an image of the line to be inspected.

[0160] In a possible implementation, the model building module is specifically used to:

[0161] Based on the multi-dynamic average pooling edge enhancement module, a first processing unit and a second processing unit are constructed. The first processing unit is used to process the input image of the ice detection model to obtain at least three shallow ice feature maps. The second processing unit is used to process the at least three shallow ice feature maps to obtain at least three mid-ice feature maps.

[0162] Based on the multi-channel dynamic deformation convolution attention module, a third processing unit is constructed, and the third processing unit is used to process at least three ice cover middle layer feature maps to obtain at least three ice cover deep layer feature maps;

[0163] Construct a first detection head, a second detection head, a third detection head and a merging output module. Each detection head is used to obtain at least one deep feature map of ice cover for regression convolution and classification operations. The merging output module is used to merge the images output by each detection head to obtain a target image. The target image is used to determine whether there is ice cover on the line in the input image.

[0164] In a possible implementation, the model building module is specifically used to:

[0165] Construct at least two multi-dynamic average pooling edge enhancement modules;

[0166] Constructing at least one first combination layer, the first combination layer is used to connect the previous multi-dynamic average pooling edge enhancement module and the next multi-dynamic average pooling edge enhancement module, and the first combination layer is used to perform convolution, normalization and activation function operations;

[0167] Constructing at least one second combination layer and connecting it to the first multi-dynamic average pooling edge enhancement module, the second combination layer is used to perform convolution, normalization and activation function operations;

[0168] Connect the last multi-dynamic average pooling edge enhancement module to the second processing unit.

[0169] In a possible implementation, the model building module is specifically used to:

[0170] Constructing a shallow expansion pooling feature extraction route, a medium expansion pooling feature extraction route, and a deep expansion pooling feature extraction route, wherein the shallow expansion pooling feature extraction route, the medium expansion pooling feature extraction route, and the deep expansion pooling feature extraction route each include at least one dynamic average pooling block and at least one dilated convolution, the dynamic average pooling block is used to perform a dynamic pooling operation on the input feature map of the multiple dynamic average pooling edge enhancement module, and the dilated convolution is used to perform a convolution operation on the image after the dynamic pooling operation;

[0171] Constructing a first fusion layer, which is used to fuse the images output by the shallow expansion pooling feature extraction route, the medium expansion pooling feature extraction route, and the deep expansion pooling feature extraction route in the channel dimension;

[0172] Constructing a first weighted layer, which is used to weight the input feature map according to the weights of each channel in the image fused by the first fusion layer;

[0173] Construct a convolutional layer, which is used to adjust the number of channels of the weighted input feature map to obtain the output feature map of the multi-dynamic average pooling edge enhancement module.

[0174] In one possible implementation, the model building module is specifically used to perform dynamic pooling operations using the following formula:

[0175]

[0176] In the above formula, M0 is the input feature map, M1 is the image after the dynamic pooling operation of M0, and M i,j Represents the element value with coordinates (i, j) in M1, i∈[0,H O ],j∈[0,W O ],m p,q Represents the element value with coordinates (p, q) in M0, p∈[0,H I ],q∈[0,W I ], Ht Denotes the height of the dynamic average pooling block of size t, W t Denotes the width of the dynamic average pooling block of size t, H s Indicates the sliding step size s of the dynamic average pooling block in the vertical direction, the distance the window moves vertically each time, W s Represents the sliding step size s of the dynamic average pooling block in the horizontal direction, and the distance the window moves horizontally each time.

[0177] In a possible implementation, the model building module is specifically used to:

[0178] Constructing a first multi-dynamic average pooling edge enhancement module and a spatial pyramid pooling module. The first multi-dynamic average pooling edge enhancement module is used to process the shallow ice feature map output by the first processing unit to obtain a first middle ice feature map. The spatial pyramid pooling module is used to pool the first middle ice feature map to obtain a second middle ice feature map.

[0179] Construct a first upsampling module and a second fusion layer. The first upsampling module is used to upsample the second ice-covered middle layer feature map to obtain a third ice-covered middle layer feature map. The second fusion layer is used to obtain the ice-covered shallow layer feature map from the first processing unit and fuse it with the third ice-covered middle layer feature map in the channel dimension to obtain a fourth ice-covered middle layer feature map.

[0180] Constructing a next polymorphic average pooling edge enhancement module connected to the second fusion layer and a second upsampling module connected to the next polymorphic average pooling edge enhancement module, the next multi-dynamic average pooling edge enhancement module is used to process the fourth ice-covered middle layer feature map to obtain a fifth ice-covered middle layer feature map, and the second upsampling module is used to upsample the fifth ice-covered middle layer feature map to obtain a sixth ice-covered middle layer feature map;

[0181] Construct a third fusion layer, which is used to obtain the shallow ice layer feature map from the first processing unit, fuse it with the sixth middle ice layer feature map in the channel dimension, obtain the seventh middle ice layer feature map, and input it into the third processing unit.

[0182] In a possible implementation, the model building module is specifically used to:

[0183] Constructing a first multi-channel dynamic deformation convolution attention module, the first multi-channel dynamic deformation convolution attention module is used to process the seventh ice-covered middle layer feature map to obtain a first ice-covered deep layer feature map, and input it into the first detection head;

[0184] Construct a third combination layer, a fourth fusion layer, and a second multi-channel dynamic deformation convolution attention module. The third combination layer is used to perform convolution, normalization, and activation function operations on the first ice-covered deep feature map to obtain a second ice-covered deep feature map. The fourth fusion layer is used to fuse the fifth ice-covered middle layer feature map with the second ice-covered deep feature map on the channel to obtain a third ice-covered deep feature map. The second multi-channel dynamic deformation convolution attention module is used to generate a fourth ice-covered deep feature map based on the third ice-covered deep feature map and output it to the second detection head.

[0185] Construct the fourth combination layer, the fifth fusion layer and the third multi-channel dynamic deformation convolution attention module. The fourth combination layer is used to perform convolution, normalization and activation function operations on the fourth ice-covered deep feature map to obtain the fifth ice-covered deep feature map. The fifth fusion layer is used to fuse the second ice-covered middle layer feature map and the fifth ice-covered deep feature map in the channel dimension to obtain the sixth ice-covered depth feature map. The third multi-channel dynamic deformation convolution attention module is used to generate the seventh ice-covered deep feature map based on the sixth ice-covered depth feature map and output it to the third detection head.

[0186] In one possible implementation, the multi-channel dynamic deformable convolution attention module includes a first dynamic deformable convolution, a second dynamic deformable convolution, a third dynamic deformable convolution, a sixth fusion layer, a fifth combination layer, and a second weighted layer; the model construction module is specifically used to:

[0187] The feature maps input to the multi-way dynamic deformable convolution attention module are convolved by the first dynamic deformable convolution, the second dynamic deformable convolution, and the third dynamic deformable convolution respectively to obtain the first dynamic deformable convolution attention map, the third dynamic deformable convolution attention map, and the third dynamic deformable convolution attention map. The convolution sizes of different dynamic deformable convolutions are different.

[0188] The first dynamic deformable convolutional attention map, the third dynamic deformable convolutional attention map, and the third dynamic deformable convolutional attention map are fused in the channel dimension through the sixth fusion layer to obtain the fourth dynamic deformable convolutional attention map;

[0189] The fourth dynamic deformation convolution attention map is sequentially subjected to global average pooling, one-dimensional convolution, and activation function operations through the fifth combination layer to obtain the weights of each channel;

[0190] The fourth dynamic deformation convolution attention map is weighted by the weight of each channel through the second weighted layer to obtain the image output by the multi-channel dynamic deformation convolution attention module;

[0191] According to the image output by the multi-channel dynamic deformation convolution attention module, the illumination changes, occluded objects and interfering background in the image are identified.

[0192] In a possible implementation, the model building module is specifically used to:

[0193] Q5=ω⊙Q4=S(C1D k (GlobalAvgPool(x)))⊙Q4

[0194]

[0195] Q5 is the image output by the multi-channel dynamic deformation convolution attention module, Q4 is the fourth dynamic deformation convolution attention map, GlobalAvgPool represents the global average pooling; C1D k represents a one-dimensional convolution with a kernel size of k; S represents the activation function; ⊙ represents the element-wise product; [t] odd Represents the nearest odd number to t, t represents the convolution kernel size, c represents the number of channels of the image output by the multi-way dynamic deformation convolution attention module, and γ and b are hyperparameters.

[0196] In a possible implementation, the method further includes a step length determination module, configured to:

[0197]

[0198] In the above formula, t represents the size of the convolution kernel, d represents the result of rounding down the square root of t; row represents the result of rounding down the value of t divided by d; B is the step size of the dynamic deformation convolution during the convolution process; [j] odd represents the nearest odd number to j, c represents the number of channels of the output feature map, and β and θ are both hyperparameters.

[0199] In one possible implementation, the preset loss function is:

[0200]

[0201] In the above formula, B pred Indicates the prediction box of ice-covered roads in the output image of the ice detection model during training, B gt represents the real box, IoU represents the intersection-over-union ratio between the predicted box and the real box; v is the correction factor for the difference in the inverse tangent of the aspect ratio; w G and h G Represents the width and height of the real box respectively; w prod and h prod Represent the width and height of the prediction box respectively; b and b gt They represent the center points of the predicted box and the true box respectively; ρ(·) represents the Euclidean distance between two pixels.

[0202] The line icing detection device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effects are similar, and are not described in detail in this embodiment.

[0203] Figure 9 This is a schematic diagram of the structure of the electronic device provided in this application. Figure 9 As shown, the electronic device 90 provided in this embodiment includes: at least one processor 901 and a memory 902. Optionally, the device 90 further includes a communication component 903. The processor 901, the memory 902 and the communication component 903 are connected via a bus.

[0204] During the specific implementation process, at least one processor 901 executes the computer-executable instructions stored in the memory 902, so that the at least one processor 901 performs the above method.

[0205] The specific implementation process of the processor 901 can be found in the above method embodiment. Its implementation principle and technical effects are similar and will not be repeated here in this embodiment.

[0206] In the above embodiments, it should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASICs), etc. A general-purpose processor may be a microprocessor or any conventional processor. The steps of the method disclosed in the present invention may be directly executed by a hardware processor or by a combination of hardware and software modules in the processor.

[0207] The memory may include a high-speed memory (Random Access Memory, RAM), and may also include a non-volatile memory (NVM), such as at least one disk memory.

[0208] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus. Buses can be classified as address buses, data buses, and control buses. For ease of illustration, the buses in the drawings of this application are not limited to just one bus or just one type of bus.

[0209] The present application also provides a computer program product, including a computer program, which implements the above method when executed by a processor.

[0210] The present application also provides a computer-readable storage medium, in which computer-executable instructions are stored. When a processor executes the computer-executable instructions, the above method is implemented.

[0211] The above-mentioned readable storage medium can be implemented by any type of volatile or non-volatile memory device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. The readable storage medium can be any available medium that can be accessed by a general-purpose or special-purpose computer.

[0212] An exemplary readable storage medium is coupled to a processor so that the processor can read information from the readable storage medium and write information to the readable storage medium. Of course, the readable storage medium can also be an integral part of the processor. The processor and the readable storage medium can be located in an application specific integrated circuit (ASIC). Of course, the processor and the readable storage medium can also exist in a device as discrete components.

[0213] The division of units is merely a logical functional division; actual implementations may employ alternative divisions, such as combining or integrating multiple units or components into another system, or omitting or disabling certain features. Furthermore, any direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between devices or units, either through an interface, electrical, mechanical, or other means.

[0214] Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0215] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0216] If the function is implemented in the form of 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, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program code.

[0217] Those skilled in the art will appreciate that all or part of the steps in the above-described method embodiments can be implemented using hardware associated with program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0218] Finally, it should be noted that those skilled in the art will readily identify other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. The present invention is intended to cover any variations, uses, or adaptations of the present invention that follow the general principles of the present invention and include common knowledge or customary techniques in the art not disclosed herein. The present invention is not limited to the precise structure described above and illustrated in the accompanying drawings, and various modifications and variations may be made without departing from the scope thereof. The scope of the present invention is limited solely by the appended claims.

Claims

1. A line icing detection method, characterized in that: include: Acquire a line ice-covered data set, wherein the line ice-covered data set includes ice-covered line images and non-ice-covered line images; Constructing an ice detection model, the ice detection model includes at least one multi-dynamic average pooling edge enhancement module and at least one multi-channel dynamic deformation convolution attention module, the multi-dynamic average pooling edge enhancement module is used to adaptively enhance the ice edge area in the image and dynamically suppress the background interference of the image, and the multi-channel dynamic deformation convolution attention module is used to identify lighting changes, occluded objects and interfering background in the image; Based on the line icing dataset and a preset loss function, the icing detection model is trained to obtain a trained icing detection model; Based on the trained ice detection model and the line image of the line to be detected, it is determined whether the line to be detected is iced.

2. The method according to claim 1, characterized in that The constructing of the ice detection model includes: Based on the multi-dynamic average pooling edge enhancement module, a first processing unit and a second processing unit are constructed, wherein the first processing unit is used to process the input image of the ice detection model to obtain at least three shallow ice feature maps, and the second processing unit is used to process the at least three shallow ice feature maps to obtain at least three mid-ice feature maps; Based on the multi-channel dynamic deformation convolution attention module, a third processing unit is constructed, wherein the third processing unit is used to process the at least three ice cover middle layer feature maps to obtain at least three ice cover deep layer feature maps; Construct a first detection head, a second detection head, a third detection head and a merging output module. Each detection head is used to obtain at least one deep feature map of ice cover for regression convolution and classification operations. The merging output module is used to merge the images output by each detection head to obtain a target image. The target image is used to determine whether there is ice cover on the line in the input image.

3. The method according to claim 2, characterized in that The first processing unit is constructed based on the multi-dynamic average pooling edge enhancement module, including: Construct at least two multi-dynamic average pooling edge enhancement modules; Constructing at least one first combination layer, where the first combination layer is used to connect a previous multi-dynamic average pooling edge enhancement module and a next multi-dynamic average pooling edge enhancement module, and the first combination layer is used to perform convolution, normalization, and activation function operations; Constructing at least one second combination layer and connecting it to the first multi-dynamic average pooling edge enhancement module, wherein the second combination layer is used to perform convolution, normalization, and activation function operations; Connect the last multi-dynamic average pooling edge enhancement module to the second processing unit.

4. The method according to claim 3, characterized in that Construct a multi-dynamic average pooling edge enhancement module, including: Constructing a shallow expansion pooling feature extraction route, a medium expansion pooling feature extraction route, and a deep expansion pooling feature extraction route, wherein the shallow expansion pooling feature extraction route, the medium expansion pooling feature extraction route, and the deep expansion pooling feature extraction route each include at least one dynamic average pooling block and at least one dilated convolution, the dynamic average pooling block is used to perform a dynamic pooling operation on the input feature map of the multiple dynamic average pooling edge enhancement module, and the dilated convolution is used to perform a convolution operation on the image after the dynamic pooling operation; Constructing a first fusion layer, where the first fusion layer is used to fuse the images output by the shallow dilation pooling feature extraction route, the medium dilation pooling feature extraction route, and the deep dilation pooling feature extraction route in the channel dimension; Constructing a first weighted layer, wherein the first weighted layer is used to weight the input feature map according to the weights of each channel in the image fused by the first fusion layer; Construct a convolution layer, which is used to adjust the number of channels of the weighted input feature map to obtain the output feature map of the multi-dynamic average pooling edge enhancement module.

5. The method according to claim 4, characterized in that The performing a dynamic pooling operation on the input feature map of the multi-dynamic average pooling edge enhancement module includes: In the above formula, M0 is the input feature map, M1 is the image after the dynamic pooling operation of M0, and M i，j Represents the element value i∈[0,H0],j∈[0,W0],m with coordinates (i, j) in M1 p,q Represents the element value with coordinates (p, q) in M0, p∈[0,H1],q∈[0,W1],H t Denotes the height of the dynamic average pooling block of size t, W t Denotes the width of the dynamic average pooling block of size t, H s Indicates the sliding step size s of the dynamic average pooling block in the vertical direction, and the distance the window moves vertically each time, W s Represents the sliding step size s of the dynamic average pooling block in the horizontal direction, and the distance the window moves horizontally each time.

6. The method according to claim 2, characterized in that Based on the multi-dynamic average pooling edge enhancement module, a second processing unit is constructed, including: Constructing a first multi-dynamic average pooling edge enhancement module and a spatial pyramid pooling module, wherein the first multi-dynamic average pooling edge enhancement module is used to process the ice cover shallow layer feature map output by the first processing unit to obtain a first ice cover middle layer feature map, and the spatial pyramid pooling module is used to pool the first ice cover middle layer feature map to obtain a second ice cover middle layer feature map; Constructing a first upsampling module and a second fusion layer, wherein the first upsampling module is used to upsample the second ice-covered middle layer feature map to obtain a third ice-covered middle layer feature map, and the second fusion layer is used to obtain the ice-covered shallow layer feature map from the first processing unit and fuse it with the third ice-covered middle layer feature map in the channel dimension to obtain a fourth ice-covered middle layer feature map; Constructing a next polymorphic average pooling edge enhancement module connected to the second fusion layer and a second upsampling module connected to the next polymorphic average pooling edge enhancement module, the next multi-dynamic average pooling edge enhancement module is used to process the fourth ice-covered middle layer feature map to obtain a fifth ice-covered middle layer feature map, and the second upsampling module is used to upsample the fifth ice-covered middle layer feature map to obtain a sixth ice-covered middle layer feature map; Construct a third fusion layer, which is used to obtain the shallow ice layer feature map from the first processing unit, fuse it with the sixth middle ice layer feature map in the channel dimension, obtain the seventh middle ice layer feature map, and input it into the third processing unit.

7. The method according to claim 2, characterized in that Based on the multi-dynamic average pooling edge enhancement module, a third processing unit is constructed, including Constructing a first multi-channel dynamic deformation convolutional attention module, wherein the first multi-channel dynamic deformation convolutional attention module is used to process the seventh ice-covered middle layer feature map to obtain a first ice-covered deep layer feature map, and input the first ice-covered deep layer feature map to the first detection head; Constructing a third combination layer, a fourth fusion layer, and a second multi-channel dynamic deformation convolution attention module, wherein the third combination layer is used to perform convolution, normalization, and activation function operations on the first ice-covered deep feature map to obtain a second ice-covered deep feature map, and the fourth fusion layer is used to fuse the fifth ice-covered middle layer feature map with the second ice-covered deep feature map on the channel to obtain a third ice-covered deep feature map, and the second multi-channel dynamic deformation convolution attention module is used to generate a fourth ice-covered deep feature map based on the third ice-covered deep feature map, and output it to the second detection head; Construct a fourth combination layer, a fifth fusion layer and a third multi-channel dynamic deformation convolution attention module, wherein the fourth combination layer is used to perform convolution, normalization and activation function operations on the fourth ice-covered deep feature map to obtain the fifth ice-covered deep feature map, and the fifth fusion layer is used to fuse the second ice-covered middle layer feature map and the fifth ice-covered deep feature map in the channel dimension to obtain the sixth ice-covered depth feature map, and the third multi-channel dynamic deformation convolution attention module is used to generate a seventh ice-covered deep feature map based on the sixth ice-covered depth feature map, and output it to the third detection head.

8. The method according to claim 1 or 7, characterized in that The multi-channel dynamic deformation convolution attention module includes a first dynamic deformation convolution, a second dynamic deformation convolution and a third dynamic deformation convolution, a sixth fusion layer, a fifth combination layer and a second weighted layer; The identification of illumination changes, occluded objects, and interfering backgrounds in an image includes: Performing convolution operations on the feature maps input to the multi-way dynamic deformable convolution attention module through the first dynamic deformable convolution, the second dynamic deformable convolution, and the third dynamic deformable convolution, respectively, to obtain a first dynamic deformable convolution attention map, a third dynamic deformable convolution attention map, and a third dynamic deformable convolution attention map, where the convolution sizes of different dynamic deformable convolutions are different; fusing the first dynamic deformable convolutional attention map, the third dynamic deformable convolutional attention map, and the third dynamic deformable convolutional attention map in the channel dimension through the sixth fusion layer to obtain a fourth dynamic deformable convolutional attention map; Performing global average pooling, one-dimensional convolution, and activation function operations on the fourth dynamically deformed convolutional attention map in sequence through the fifth combination layer to obtain the weights of each channel; weighting the fourth dynamic deformation convolution attention map by using the weights of each channel through the second weighted layer to obtain an image output by the multi-channel dynamic deformation convolution attention module; According to the image output by the multi-channel dynamic deformation convolution attention module, the illumination changes, occluded targets and interfering background in the image are identified.

9. The method according to claim 8, characterized in that The method of weighting the fourth dynamic deformation convolution attention map by using the weights of each channel to obtain the image output by the multi-channel dynamic deformation convolution attention module includes: Q5=ω⊙Q4=S(C1D k (GlobalAvgPool(x)))⊙Q4 Q5 is the image output by the multi-channel dynamic deformation convolution attention module, Q4 is the fourth dynamic deformation convolution attention map, GlobalAvgPool represents the global average pooling; C1D k represents a one-dimensional convolution with a kernel size of k; S represents the activation function; ⊙ represents the element-wise product, [t] odd Represents the nearest odd number to t, t represents the convolution kernel size, c represents the number of channels of the image output by the multi-way dynamic deformation convolution attention module, and γ and b are hyperparameters.

10. The method according to claim 8, characterized in that The method further includes determining the step sizes of the first dynamic deformable convolution, the second dynamic deformable convolution, and the third dynamic deformable convolution by the following formula: In the above formula, t represents the size of the convolution kernel, d represents the result of rounding down the square root of t; row represents the result of rounding down the value of t divided by d, and B is the step size of the dynamic deformation convolution during the convolution process; [j] odd represents the nearest odd number to j, c represents the number of channels of the output feature map, and b and θ are both hyperparameters.

11. The method according to any one of claims 1 to 7, 9 and 10, characterized in that: The preset loss function is: In the above formula, B pred Indicates the prediction box of ice-covered lines in the output image of the ice detection model during training, B gt represents the real box, IoU represents the intersection-over-union ratio between the predicted box and the real box; V is the correction factor for the difference in the inverse tangent of the aspect ratio; W G and h G Represents the width and height of the real box respectively; W prod and H prod Represent the width and height of the prediction box respectively; b and b gt Represent the center points of the predicted box and the real box respectively; p() represents the Euclidean distance between two pixels.

12. A line icing detection device, characterized in that: include: An acquisition module, configured to acquire a line ice-covered data set, wherein the line ice-covered data set includes ice-covered line images and non-ice-covered line images; A model construction module for constructing an ice detection model, the ice detection model including at least one multi-dynamic average pooling edge enhancement module and at least one multi-channel dynamic deformation convolution attention module, the multi-dynamic average pooling edge enhancement module is used to adaptively enhance the ice edge area in the image and dynamically suppress the background interference of the image, and the multi-channel dynamic deformation convolution attention module is used to identify lighting changes, occluded targets and interfering background in the image; A model training module is used to train the ice detection model based on the line ice data set and a preset loss function to obtain a trained ice detection model; The detection module is used to determine whether the line to be detected is covered with ice based on the trained ice detection model and the line image of the line to be detected.

13. An electronic device, characterized in that: include: Memory, processor; The memory stores computer-executable instructions; The processor executes the computer-executable instructions stored in the memory, so that the processor performs the method according to any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer-executable instructions, which are used to implement the method according to any one of claims 1 to 11 when executed by a processor.

15. A computer program product, characterized in that The invention comprises a computer program, which implements the method according to any one of claims 1 to 11 when being executed by a processor.

Images