An icing image segmentation method and device, electronic equipment and storage medium

Abstract

Embodiments of the present application disclose an icing image segmentation method and device, electronic equipment and storage medium, the method comprising: when the initial segmentation model established in advance does not meet the preset condition, obtaining the icing image set of the power transmission line; based on the hollow convolution unit and attention unit of the feature extraction module, the feature of the icing image in the icing image set is extracted, and the first feature is obtained; according to the first feature and the up-sampling module, the target feature of the icing image is obtained; based on the target feature and the preset loss function, the initial segmentation model is optimized until the initial segmentation model meets the icing image segmentation preset condition, and the target segmentation model is obtained. The target segmentation model of the present application can effectively segment the icing image of the power transmission line.

Description

TECHNICAL FIELD

[0001] The embodiment of the present application relates to the technical field of power transmission line safety control, and particularly relates to an icing image segmentation method and device, electronic equipment and storage medium. BACKGROUND

[0002] The power transmission line is easily affected by ice and snow covering in winter. The icing not only increases the weight of the power transmission line, resulting in increased mechanical load, but also can cause conductor jumping, wire breaking and other safety accidents, which seriously affects the safe operation of the power system. Therefore, real-time monitoring and timely cleaning of the icing on the power transmission line are important tasks to ensure the safety of the power system.

[0003] In recent years, deep learning technology has made significant progress in the field of computer vision, especially in image classification, target detection and image segmentation tasks. The automatic icing detection and segmentation method based on deep learning can train the model through large-scale data, automatically extract icing features, adapt to complex and variable environments, and significantly improve the accuracy and robustness of power transmission line icing image segmentation. However, the current deep learning-based icing image segmentation method has too many deep learning network parameters, resulting in long model training time and low running efficiency, which reduces the efficiency of power transmission line icing image segmentation. SUMMARY

[0004] The embodiment of the present application provides an icing image segmentation method, device, electronic equipment and storage medium, which can quickly and accurately obtain a model capable of segmenting a power transmission line icing image, improve the efficiency of model training, further improve the efficiency of power transmission line icing image segmentation, and facilitate subsequent discrimination operation according to the icing condition.

[0005] In a first aspect, the embodiment of the present application provides an icing image segmentation method, comprising:

[0006] When the pre-established initial segmentation model does not meet the preset condition, an icing image set of the power transmission line is obtained; the initial segmentation model comprises a feature extraction module and an up-sampling module;

[0007] The feature extraction module is based on the hollow convolution unit and the attention unit to extract features of the icing image in the icing image set, and obtain a first feature;

[0008] According to the first feature and the up-sampling module, a target feature of the icing image is obtained;

[0009] Based on the target feature and the pre-set loss function, the initial segmentation model is optimized until the initial segmentation model meets the preset condition, and a target segmentation model is obtained, which is used for image segmentation of the power transmission line icing image.

[0010] In a second aspect, an embodiment of the present application provides an icing image segmentation device, the device comprising:

[0011] a data acquisition module configured to acquire an icing image set of a power transmission line when a pre-established initial segmentation model does not meet a preset condition, wherein the initial segmentation model comprises a feature extraction module and an up-sampling module;

[0012] a first determination module configured to perform feature extraction on an icing image in the icing image set based on a cavity convolution unit and an attention unit of the feature extraction module to obtain first features;

[0013] a second determination module configured to obtain target features of the icing image according to the first features and the up-sampling module;

[0014] a model optimization module configured to optimize the initial segmentation model based on the target features and a pre-set loss function until the initial segmentation model meets the preset condition to obtain a target segmentation model, wherein the target segmentation model is used for image segmentation on an icing image of a power transmission line.

[0015] In a third aspect, an embodiment of the present application further provides an electronic device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements a kind of icing image segmentation method as any of the embodiments of the present application when executing the program.

[0016] In a fourth aspect, an embodiment of the present application further provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement a kind of icing image segmentation method as any of the embodiments of the present application.

[0017] In the embodiments of the present application, the cavity convolution is used instead of the traditional convolution, which expands the receptive field without reducing the resolution, and the attention unit is combined on the basis of the cavity convolution unit, so that the target segmentation model can fuse icing feature information of multiple scales, improve the icing image segmentation quality of the power transmission line, and suppress interference factors. The up-sampling module saves the parameter calculation amount of the target segmentation model, improves the efficiency of model training, and enables the target segmentation model to effectively segment the icing image of the power transmission line. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor.

[0019] Figure 1 A first flow chart of an icing image segmentation method provided by an embodiment of the present application;

[0020] Figure 2 A schematic diagram of a feature extraction module provided by an embodiment of the present application;

[0021] Figure 3 A model structure diagram of an initial segmentation model provided by an embodiment of the present application;

[0022] Figure 4 A second flow chart of an icing image segmentation method provided by an embodiment of the present application;

[0023] Figure 5 A schematic diagram of a cavity convolution unit processing input features provided by an embodiment of the present application;

[0024] Figure 6 A schematic diagram of a pooling unit processing input features provided by an embodiment of the present application;

[0025] Figure 7 A schematic diagram of an attention unit processing input features provided by an embodiment of the present application;

[0026] Figure 8 A third flow chart of an icing image segmentation method provided by an embodiment of the present application;

[0027] Figure 9 A structural schematic diagram of an icing image segmentation device provided by an embodiment of the present application;

[0028] Figure 10 A structural schematic diagram of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0029] The present application will be further described below in conjunction with the accompanying drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application. In addition, it should be noted that, for the convenience of description, only the parts related to the present application are shown in the drawings, but not all the structures.

[0030] Figure 1A first flowchart of an icing image segmentation method provided by an embodiment of the present application is shown in FIG. 1. The method of the embodiment of the present application can quickly and accurately obtain a model capable of segmenting an icing image of a power transmission line, improve the efficiency of model training, further improve the efficiency of icing image segmentation of the power transmission line, and facilitate subsequent discrimination operations according to the icing condition. The method can be executed by an icing image segmentation device provided by an embodiment of the present application. The device can be implemented in the form of software and / or hardware. The following embodiments will be described by taking an example in which the device is integrated in an electronic device. The electronic device can be a computer device or a server, etc. Referring to FIG. 1, the method can include the following steps. Figure 1 The method can specifically include the following steps:

[0031] In step 101, when a pre-established initial segmentation model does not meet a preset condition, an icing image set of a power transmission line is obtained.

[0032] The initial segmentation model includes a feature extraction module and an up-sampling module. The initial segmentation model is a pre-established model that has not been trained. The feature extraction module is used for feature extraction and feature refinement of the icing image of the power transmission line. The up-sampling module is used for up-sampling of a feature vector. The preset condition can be that a pre-set loss function value reaches convergence or the number of iterations reaches a pre-set number.

[0033] Specifically, when the initial segmentation model does not meet the preset condition, it indicates that the initial segmentation model cannot accurately segment the icing image, and the initial segmentation model needs to be trained and optimized according to the icing image set of the power transmission line. The server can obtain the icing image set of the power transmission line in the historical period from the power system through a pre-set data interface, or directly receive the icing image set of the power transmission line uploaded by the user.

[0034] In step 102, the feature extraction module is used to extract features of the icing image in the icing image set based on a hollow convolution unit and an attention unit of the feature extraction module, to obtain first features.

[0035] The feature extraction module includes a hollow convolution unit, an attention unit and a pooling unit. The hollow convolution unit is composed of a hollow convolution layer. The hollow convolution expands its receptive field by adding spaces (zeros) in the convolution kernel. Figure 2 A schematic diagram of the feature extraction module provided by an embodiment of the present application is shown in FIG. 2. Figure 2As shown in the figure, the feature extraction module includes a first dilated convolution unit, a second dilated convolution unit, a pooling unit and an attention unit. The dilated rate of the first dilated convolution unit is 2, and the dilated rate of the second dilated convolution unit is 3. The pooling unit is used to increase the network depth and width of the initial segmentation model, thereby obtaining icing features of multiple perspectives and multiple scales. The attention unit is used to implement the attention mechanism, so that the initial segmentation model increases the weight value of important feature maps or feature channels and reduces the weight value of unimportant features, thereby improving the performance of the initial segmentation model.

[0036] In an optional embodiment, after obtaining the icing image set, the icing images in the icing image set are input into the initial segmentation model. After receiving the icing images, the initial segmentation model can perform convolution and batch normalization on the icing images to reduce the internal covariate shift. The icing images after batch normalization are activated according to a pre-set activation function (such as a ReLU activation function) to obtain first output features. The first output features are input into the feature extraction module, and the first output features are sequentially subjected to dilated convolution with a dilated rate of 2 and dilated convolution with a dilated rate of 3 through the dilated convolution units of the feature extraction module to obtain second output features. The second output features are input into the pooling unit to obtain third output features of multiple perspectives and multiple scales. The third output features are input into the attention unit to adjust the weight values of the third output features through the attention unit to obtain first features.

[0037] Step 103, obtaining target features of the icing image by the initial segmentation model according to the first features and the icing image.

[0038] The upsampling module is used to gradually restore the resolution of the icing feature map to the size of the original input transmission line icing image. The target features are the features obtained by image segmentation of the icing image by the initial segmentation model. Specifically, according to the positions of the pixels in the icing image, the first features are dynamically weighted by the upsampling module using transposed convolution and bilinear interpolation. In the upsampling process, the closer the pixel position is to the edge of the icing image, the greater the weight value of the bilinear interpolation, and the closer the pixel position is to the center of the icing image, the greater the weight value of the transposed convolution.

[0039] In this scheme, the initial segmentation model further includes a feature extraction sub-module composed of a dilated convolution unit and an attention unit in series. The dilated rate of the dilated convolution in the feature extraction sub-module can be 2. In an optional embodiment, the initial segmentation model can be composed of multiple convolution layers, multiple feature extraction modules, multiple upsampling modules, multiple dilated convolution units, multiple attention units and multiple convolution units. Figure 3 The model structure diagram of the initial segmentation model provided by the embodiment of the present application is shown in the figure. Figure 3 As shown in the figure,Figure 3 The "+" in the formula indicates channel dimension splicing. The icing image is input into the initial segmentation model. The convolution unit 1 of the initial segmentation model uses a convolution kernel with a size of 3*3, and the step is set to 1. The input icing image is subjected to a convolution operation, and the convolved icing image is subjected to batch normalization processing. Then, the icing image is processed through a nonlinear activation to obtain F1. F1 is subjected to a convolution operation through the convolution unit 2 using a convolution kernel with a size of 7*7. The convolved F1 is input into the feature extraction module 1 to obtain F2. The input icing image is subjected to a convolution operation through the convolution unit 3 using a convolution kernel with a size of 3*3. The convolved F2 is input into the feature extraction module 2 to obtain F3. The input feature is subjected to a convolution operation through the convolution unit 4 using a convolution kernel with a size of 3*3. The convolved F3 is input into the empty convolution unit to obtain F4. F4 is input into the upsampling module. The features output by the upsampling module are channel dimension spliced with F3. The spliced vector is input into the feature extraction submodule 1 to obtain F5. F5 is input into the upsampling module 2 for upsampling. F2 and the upsampled F5 are channel dimension spliced. The spliced vector is input into the feature extraction submodule 2 to obtain F6. F6 is input into the upsampling module 3 for upsampling. The upsampled F6 and F1 are channel dimension spliced. The spliced features are input into the attention unit for weight adjustment. The features with adjusted weights are input into the convolution unit 5. The input features are subjected to a convolution operation through the convolution unit 5 using a convolution kernel with a size of 3*3 to obtain F7. F7, F5, F6 and F4 are channel dimension spliced. The spliced features are input into the convolution unit 6. The input features are subjected to a convolution operation through the convolution unit 6 using a convolution kernel with a size of 3*3 to obtain the target feature (the image segmentation result corresponding to the icing image).

[0040] In step 104, the initial segmentation model is optimized based on the target feature and the pre-set loss function until the initial segmentation model meets the pre-set condition to obtain the target segmentation model.

[0041] The target segmentation model is used for image segmentation of the icing image of the power transmission line. The preset condition can be that the value of the loss function reaches convergence. When the initial segmentation model does not meet the preset condition, an icing image is selected from the icing image set as a current image, and the current image is input to the initial segmentation model. After the initial segmentation model receives the current image, the initial segmentation model processes the current image and outputs a current segmentation result (a target feature corresponding to the current icing image) corresponding to the current image. The target feature corresponding to the current icing image and a label feature (a standard segmentation result corresponding to the current icing image) corresponding to the current icing image determined in advance are input to the loss function to obtain a loss function value. After the loss function value is obtained, it is determined whether the loss function value reaches the preset condition. If the loss function value reaches the preset condition, the initial segmentation model is determined as the target segmentation model. If the loss function value does not reach the preset condition, the model parameters of the initial segmentation model are adjusted according to the loss function value, and the next icing image is selected from the icing image set as the current image. The step of inputting the current image to the initial segmentation model is repeatedly performed, and the model parameters of the initial segmentation model are continuously adjusted until the initial segmentation model meets the preset condition, and the target segmentation model is obtained.

[0042] The technical scheme of the embodiment, when the pre-established initial segmentation model does not meet the preset condition, an icing image set of a power transmission line is obtained. The initial segmentation model includes a feature extraction module and an upsampling module. The feature extraction module is based on a cavity convolution unit and an attention unit to extract features of the icing images in the icing image set to obtain first features. The target feature of the icing image is obtained according to the first features and the upsampling module. The initial segmentation model is optimized based on the target feature and the pre-set loss function until the initial segmentation model meets the preset condition, and the target segmentation model is obtained. The target segmentation model is used for image segmentation of the icing image of the power transmission line. The technical scheme of the embodiment uses a cavity convolution instead of a traditional convolution, expands the receptive field without reducing the resolution, and combines the attention unit based on the cavity convolution unit, so that the target segmentation model can fuse icing feature information of multiple scales, improve the icing image segmentation quality of the power transmission line, and suppress interference factors. The upsampling module can save the parameter calculation amount of the target segmentation model, so that the target segmentation model can effectively segment the icing image of the power transmission line.

[0043] Figure 4 A second flowchart of an icing image segmentation method provided by the embodiment of the application is shown in the following table. Figure 4 The method can include the following steps:

[0044] Step 401, when a pre-established initial segmentation model does not meet a preset condition, an icing image set of a power transmission line is obtained.

[0045] The initial segmentation model comprises a feature extraction module and an up-sampling module.

[0046] At step 402, for each icing image in the set of icing images, the current icing image is input into the initial segmentation model to obtain first output features of the current icing image output by a convolution unit of the initial segmentation model.

[0047] The convolution unit comprises a convolution layer, a normalization layer and an activation layer, and the convolution kernel of the convolution layer can be 3x3. In an optional implementation, after obtaining the set of current icing images, the current icing image is input into the initial segmentation model. After receiving the current icing image, the initial segmentation model performs a convolution operation on the input current icing image using a convolution kernel with a size of 3x3 and a step size of 1 through the convolution unit. The current icing image after convolution is batch normalized through the normalization layer. The current icing image after normalization is nonlinearly activated using a nonlinear activation function through the activation layer to obtain the first output features of the current icing image.

[0048] At step 403, the first output features are processed by a dilated convolution unit to obtain second output features.

[0049] The dilated convolution unit comprises a first convolution layer and a second convolution layer. Dilated convolution expands its receptive field by adding spaces (zeros) in the convolution kernel. In this scheme, the first output features are processed by the dilated convolution unit to obtain the second output features, which comprises: performing dimension reduction convolution and dilated convolution on the first output features through the first convolution layer of the dilated convolution unit to obtain first dilated features; performing dimension reduction convolution on the first output features through the second convolution layer of the dilated convolution unit to obtain second dilated features; element-wise adding the first dilated features and the second dilated features to obtain candidate second output features, and obtaining the second output features based on the candidate second output features.

[0050] Figure 5 A schematic diagram of the dilated convolution unit provided by the embodiment of the present application processing input features is shown in FIG. 3. Figure 5 As shown in FIG. 3, the dilated convolution unit comprises a first convolution layer and a second convolution layer. Figure 5Add represents element-wise addition, after the input feature (the first output feature) is obtained, the empty convolution unit uses a convolution kernel with a size of 1*1 to perform convolution on the input feature through the convolution layer in the first convolution layer. The input feature after convolution is batch normalized through the normalization layer. The input feature after normalization is nonlinearly activated through the activation layer using a nonlinear activation function. The activated input feature is dilated convolution with a dilation rate of 2 or 3 through the dilated convolution layer, and the input feature after dilated convolution is obtained. The input feature after dilated convolution is batch normalized and nonlinearly activated. The input feature after nonlinear activation is convolved using a convolution kernel with a size of 1*1 through the convolution layer, and the first dilated feature is obtained. At the same time, the input feature is convolved using a convolution kernel with a size of 1*1 through the convolution layer of the second convolution layer of the dilated convolution unit, and the input feature after convolution is batch normalized to obtain the second dilated feature. After obtaining the first dilated feature and the second dilated feature, the first dilated feature and the second dilated feature are added element by element to obtain the candidate output feature. After obtaining the candidate output feature, the candidate output feature is nonlinearly activated using a nonlinear activation function to obtain the output feature (the second output feature).

[0051] For example, BN represents batch normalization, ReLU represents nonlinear activation, Intput represents the input feature of the dilated convolution unit, Output represents the output feature of the dilated convolution unit, do1 represents the first dilated feature, do2 represents the second dilated feature, Add represents element-wise addition, conv d represents dilated convolution, conv represents convolution, n represents dilation rate, and n is equal to 2 or 3. Further, the calculation process of the output feature of the dilated convolution unit can be represented by the following formula:

[0052]

[0053] In the above steps, the dilated convolution is used instead of the traditional convolution, which expands the receptive field without reducing the resolution, so that the network can fuse ice coating feature information of multiple scales.

[0054] Step 404, performing feature extraction on the second output feature through the pooling unit to obtain a third output feature.

[0055] The pooling unit is used to increase the network depth and width of the initial segmentation model, so as to obtain icing features of multiple perspectives and multiple scales. The pooling unit comprises a first convolutional layer, a second convolutional layer, a third convolutional layer and a maximum pooling layer. In the scheme, the second output feature is extracted by the pooling unit to obtain the third output feature, including: the second output feature is extracted by the first convolutional layer, the second convolutional layer and the third convolutional layer to obtain the first convolutional feature, the second convolutional feature and the third convolutional feature; the second output feature is fused by the maximum pooling layer to obtain the fourth pooling feature; and the first convolutional feature, the second convolutional feature, the third convolutional feature and the fourth pooling feature are spliced to obtain the third output feature.

[0056] Figure 6 The pooling unit provided by the embodiment of the application processes the input feature. As shown in Figure 6 Figure 6 The "+" in the formula represents channel dimension splicing. After receiving the input feature (the second output feature), the pooling unit uses a convolution kernel with a size of 1*1 in the first convolutional layer to perform convolution operation on the input feature, to obtain the first convolutional feature. The second convolutional layer uses a convolution kernel with a size of 1*1 and a convolution kernel with a size of 3*3 to perform convolution operation on the input feature, to obtain the second convolutional feature. The third convolutional layer uses a convolution kernel with a size of 1*1 and two convolution kernels with a size of 3*3 to perform convolution operation on the input feature, to obtain the third convolutional feature. The maximum pooling layer performs maximum pooling with a pooling window size of 3*3 on the input feature, and then uses a convolution kernel with a size of 1*1 to perform convolution on the maximum pooled input feature, to obtain the fourth pooling feature. Further, the first convolutional feature, the second convolutional feature, the third convolutional feature and the fourth pooling feature are spliced in the channel dimension, to obtain the output feature (the third output feature).

[0057] ​Exemplarily, Input represents the input feature of the pooling unit, and po1 represents the first convolutional feature. The channel number of Input is reduced through the first convolutional layer to reduce the calculation complexity, and the output of the first convolutional layer po1 is obtained. po2 represents the second convolutional feature. In the second convolutional layer, Input is first reduced in channel number through a 1*1 convolutional kernel, and then 3*3 convolutional kernel is used for ice feature extraction, and the output po2 of the second convolutional layer is obtained. po3 represents the third convolutional feature. In the third convolutional layer, Input is first reduced in channel number through a 1*1 convolutional kernel, and then two consecutive 3*3 convolutional kernels are used for deeper ice feature extraction, and the output po3 of the third convolutional layer is obtained. In the present scheme, the convolutional layers corresponding to the two 3*3 convolutional kernels have the same receptive field as the convolutional layer corresponding to the 5*5 convolutional kernel, but have fewer calculation parameters, which can reduce the network parameter amount. po4 represents the fourth pooling feature. In the maximum pooling layer, Input is first maximum-pooled (pooling window is 3*3), and then a 1*1 convolutional kernel is used to fuse the ice features, and po4 is obtained as the output of the maximum pooling layer. The channel dimension of po1, po2, po3 and po4 is spliced to obtain Output. Therefore, the calculation formula of Output is as follows:

[0058]

[0059] conv represents convolution, and Concat represents channel dimension splicing. maxpooling represents maximum pooling.

[0060] In the above steps, through the pooling unit, the performance of the target segmentation model can be improved, the parameter amount is reduced, and the depth and width of the network are increased, so that multiple perspective and multi-scale ice features are obtained.

[0061] In step 405, the third output feature is adjusted in weight according to the attention unit to obtain the first feature.

[0062] The attention unit is used to implement the attention mechanism, so that the initial segmentation model increases the weight of important feature maps or feature channels and reduces the weight of unimportant features, thereby improving the performance of ice image segmentation. The attention unit includes a convolutional layer and a global pooling layer. In the present scheme, the third output feature is adjusted in weight according to the attention unit to obtain the first feature, which includes: performing dimension reduction convolution on the third output feature through the convolutional layer of the attention unit, and performing spatial weighting on the third output feature and the third output feature after dimension reduction convolution to obtain the first attention feature; performing global pooling processing on the third output unit through the global pooling layer of the attention unit, and performing channel weighting on the third output feature and the third output feature after global pooling processing to obtain the second attention feature; and element-level adding the first attention feature and the second attention feature to obtain the first feature.

[0063] Figure 7 The schematic diagram of the attention unit processing the input feature is provided for the embodiment of the present application. As shown in the figure, Figure 7 Figure 7 The "X" in the figure represents weighting. After receiving the input feature (the third output feature), the attention unit uses a convolution kernel with a size of 1x1 to perform convolution on the input feature through a convolution layer, uses an activation function to activate the input feature after convolution, and obtains the activated input feature (the third output feature after dimension reduction convolution). The input feature and the activated input feature are spatially weighted to obtain the first attention feature. At the same time, the input feature is processed by a global pooling layer for global average pooling, two convolution kernels with a size of 1x1 are used to perform convolution on the input feature after global average pooling, an activation function is used to activate the input feature after convolution, and the activated input feature (the third output feature after global pooling processing) is obtained. Further, the input feature and the activated input feature are channel weighted to obtain the second attention feature. The first attention feature and the second attention feature are added element by element to obtain the output feature (the first feature).

[0064] For example, Input represents the input feature of the attention unit, and conv represents convolution using a 1x1 convolution kernel. The dimension of Input is CxHxW. First, Input is conv(1x1) to obtain an attention feature S1 with a dimension of 1xHxW, then S1 is activated by a Sigmoid function to obtain an activated attention feature S2 with a dimension of 1xHxW, and finally the spatial weighting of S2 and Input is calculated by Spatial-wise (spatial weighting algorithm) to obtain the first attention feature S. On the other hand, the global average pooling of Input with a dimension of CxHxW is performed to obtain an attention feature C1 with a dimension of Cx1x1, then C1 is conv(1x1) to obtain a feature C2 with a dimension of C3 with a dimension of Cx1x1 is obtained by performing conv(1x1) on C2 and activating it using a Sigmoid function, and the channel weighting of C3 and Input is performed by Channel-wise (channel weighting algorithm) to obtain the second attention feature C. After Add of S and C, the Output is obtained. gap is global average pooling. Further, the calculation process of the Output of the attention unit can be described by the following formula:

[0065]

[0066] ​In one alternative implementation, for a 1×H×W spatial weight map S2 and an Input of C×H×W, the 1×H×W spatial weight map S2 is first broadcast spatially-wise to each channel, generating a C×H×W S3. Each channel uses the same H×W attention map, i.e., H×W is copied C times along the channel. Then, S3 is multiplied element-wise with the Input to obtain S:

[0067] S c,i,j =s c,i,j ×I c,i,j

[0068] Among them, S c,i,j Let s be the value in the i-th row and j-th column of the c-th channel of S. c,i,j Let I be the value in the i-th row and j-th column of the c-th channel of S3. c,i,j Let C3 be the value in the i-th row and j-th column of the c-th channel of the Input. For an Input with dimensions C×1×1 and C×H×W, Channel-wise treats the C×1×1 C3 as the weight of each channel c, and then assigns the 1×1 weight W to each channel c. c Broadcast to every position in H×W, resulting in C:

[0069] C c,i,j =W c ×I c,i,j

[0070] Among them, C c,i,j Let I be the value in the i-th row and j-th column of the c-th channel of C. c,i,j This represents the value in the i-th row and j-th column of the c-th channel of the Input. Attention units allow the network to increase the weights of important feature maps or feature channels and decrease the weights of unimportant features, thereby improving the performance of the object segmentation model.

[0071] Step 406: For the first feature of each icing image, determine the input features of the upsampling module based on the first feature of the current icing image, the feature extraction module, and the dilated convolutional unit of the feature extraction module.

[0072] Specifically, after obtaining the first feature, it is input into the feature extraction module, and then a 1×1 convolution kernel is used to convolve the first feature output by the feature extraction module. The convolved first feature is then input into the feature extraction module again, and a 1×1 convolution kernel is used to convolve the first feature output by the feature extraction module. The convolved first feature is then input into a dilated convolution unit to obtain the input feature of the upsampling module output by the dilated convolution unit.

[0073] like Figure 3 As shown, Figure 3F2 is a first feature, F2 is input into a feature extraction module after a Conv(3x3), step=2 operation, the icing features of F2 are extracted and refined to obtain F3. The resolution of F3 is half of F2. F3 is input into a dilated convolution unit with a dilated rate of 2 after a Conv(3x3), step=2 operation, the icing features of F3 are extracted and refined to obtain the input feature F4 of the upsampling module, and the resolution of F4 is half of F3.

[0074] In step 407, the input feature of the upsampling module is transposed and dynamically weighted by bilinear interpolation based on the pixel position of the current icing image to obtain a second feature.

[0075] The pixel position of the current icing image is the position of the pixel in the upsampling process of the current icing image. The bilinear interpolation uses the four real existing pixel values around the target point in the original image to jointly determine a pixel value in the target image, and the core idea is to perform linear interpolation in two directions respectively. The transposed convolution is to obtain a transposed matrix to fill the feature map with low resolution, so as to realize the expansion of the resolution of the feature map.

[0076] In an optional embodiment, after obtaining the input feature of the upsampling module, if the pixel position of the current icing image is closer to the image edge, the weight of the bilinear interpolation is made larger. If the pixel position of the current icing image is closer to the image center, the weight of the transposed convolution is made larger, and the balance of the two is used to reduce the loss of details and the calculation overhead. The specific weight adjustment equation is as follows:

[0077]

[0078] Wherein, W and H represent the height and width of the input feature of the upsampling module, and UPres(i,j) represents the pixel value corresponding to the coordinate (i,j) after upsampling. tc(i,j) represents the pixel value corresponding to the coordinate (i,j) after transposed convolution, and the calculation formula of tc(i,j) is as follows:

[0079]

[0080] Wherein, T is a deconvolution kernel, the size of T is KxK, S is a step, Downward rounding. bl(i,j) represents the pixel value corresponding to the coordinate (i,j) after bilinear interpolation. The calculation formula of bl(i,j) is as follows:

[0081]

[0082] Wherein, S h , Sw As a scaling factor, H in W in H represents the height and width of the input feature map, respectively. out and W out These represent the height and width of the output feature map, respectively. This indicates rounding down. This indicates rounding up. The upsampling module can dynamically and adaptively adjust the weights of the two upsampling methods based on the position of the icing feature map. This can alleviate the problem of blurred edge details caused by traditional deconvolution upsampling methods, and also save parameter computation to the maximum extent, thereby improving the upsampling efficiency of the target segmentation model.

[0083] Step 408: Determine the target features based on the second feature, the first feature, the feature extraction module, the upsampling module, and the convolutional units of the initial segmentation model.

[0084] Specifically, after obtaining the second feature output by the upsampling module, the second feature and the first feature are further processed through the feature extraction submodule, the upsampling module, the convolution unit, the attention unit, and the feature extraction submodule to obtain the target feature.

[0085] For example, such as Figure 3 As shown, F4 is the input feature of the upsampling module. After being upsampled by the upsampling module, F4 is concatenated with F3 along the channel dimension. The concatenated feature is then input into the feature extraction submodule to obtain F5, whose resolution is twice that of F4. F5 is then upsampled by the upsampling module and concatenated with F2 along the channel dimension. The concatenated feature is then input into the feature extraction submodule to obtain F6, whose resolution is twice that of F5. F6 is then upsampled by the upsampling module and concatenated with F1 along the channel dimension. The concatenated feature is then input into the attention unit. The result of the attention unit is processed by a Conv(3x3) convolution operation to obtain F7, whose resolution is twice that of F6. Concatenating the channel dimensions of F7, F5, F6, and F4, and then concatenating the concatenated features by a Conv(3x3) convolution operation, yields the target feature, which is the image segmentation result corresponding to the current icing image.

[0086] Step 409: Optimize the initial segmentation model based on the target features and the pre-set loss function until the initial segmentation model meets the preset conditions, and obtain the target segmentation model.

[0087] The target segmentation model is used for image segmentation of icy transmission line images.

[0088] In the technical scheme of the embodiment, when the pre-established initial segmentation model does not meet the preset condition, an icing image set of a power transmission line is acquired; the initial segmentation model comprises a feature extraction module and an up-sampling module; for each icing image in the icing image set, the current icing image is input into the initial segmentation model to obtain first output features of the current icing image output by a convolution unit of the initial segmentation model. The first output features are processed by a hollow convolution unit to obtain second output features. The second output features are processed by a pooling unit to obtain third output features. The third output features are adjusted in weight by an attention unit to obtain first features. For the first features of each icing image, input features of the up-sampling module are determined based on the first features of the current icing image, the feature extraction module and the hollow convolution unit of the feature extraction module. The input features of the up-sampling module are processed by the up-sampling module based on pixel positions of the current icing image to obtain second features through transposed convolution and dynamic weighting of bilinear interpolation. Target features are determined based on the second features, the first features, the feature extraction module, the up-sampling module and the convolution unit of the initial segmentation model. The initial segmentation model is optimized based on the target features and a pre-set loss function until the initial segmentation model meets the preset condition, and a target segmentation model is obtained, which is used for image segmentation of the icing image of the power transmission line. In the technical scheme of the embodiment, the hollow convolution unit can expand the receptive field without reducing the resolution, so that the network can fuse icing feature information of multiple scales. The pooling unit can reduce the parameter amount while increasing the depth and width of the network to obtain icing features of multiple perspectives and multiple scales. The attention unit can realize the attention mechanism, so that the network can increase the weight of important feature maps or feature channels and reduce the weight of unimportant features, thereby improving the performance of the icing image segmentation. The up-sampling module can not only alleviate the problem of relatively blurred partial edge details caused by the traditional deconvolution up-sampling method, but also maximize the parameter calculation amount and improve the up-sampling efficiency of the target segmentation model.

[0089] Figure 8 A third flowchart of an icing image segmentation method provided by the embodiment of the application is shown in FIG. 8. The method can include the following steps: Figure 8

[0090] Step 801: Acquire a to-be-processed icing image of a power transmission line.

[0091] The to-be-processed icing image is an image that needs to be subjected to image segmentation. In actual application, when the icing condition of the power transmission line needs to be analyzed, a user can upload the to-be-processed icing image of the power transmission line to a server, so that the server acquires the to-be-processed icing image of the power transmission line.

[0092] ​In step 802, the icing image to be processed is input into the target segmentation model, and the target segmentation model is used to perform image segmentation on the icing image to be processed to obtain a segmentation result image of the icing image to be processed.

[0093] The target segmentation model is used to perform image segmentation on the icing image of the power transmission line. After obtaining the icing image to be processed, the icing image to be processed is input into the target segmentation model. The target segmentation model can perform image segmentation on the icing image to be processed to obtain a segmentation result image of the icing image to be processed.

[0094] The model structure of the target segmentation model is shown in Figure 3 As an example, assuming that the dimension of the icing image to be processed is 1x1024x1024 (channelxwidthxheight). The icing image to be processed is input into the target segmentation model, and first processed by the convolution unit (Conv(3x3)_BN_ReLU) to obtain F1 with a dimension of 16x1024x1024. F1 is subjected to a downsampling operation of Conv(7x7) with a step size of 2, and then input into a feature extraction module to obtain F2 with a dimension of 32x512x512. F2 is subjected to a downsampling operation of Conv(3x3) with a step size of 2, and then input into the feature extraction module to extract and refine the icing features of F2 to obtain F3 with a dimension of 48x256x256. F3 is subjected to a downsampling operation of Conv(3x3) with a step size of 2, and then input into a dilated convolution unit with a dilated rate of 2 to extract and refine the icing features of F3 to obtain F4 with a dimension of 64x128x128. F4 is upsampled by an upsampling module, and then subjected to a channel dimension splicing operation with F3. The feature map after the channel dimension splicing operation is input into a feature extraction submodule to obtain F5 with a dimension of 56x256x256. F5 is upsampled by an upsampling module, and then subjected to a channel dimension splicing operation with F2. The feature after the channel dimension splicing operation is input into a feature extraction submodule to obtain F6 with a dimension of 44x512x512. F6 is upsampled by an upsampling module, and then subjected to a channel dimension splicing operation with F1. The feature after the channel dimension splicing operation is input into an attention unit. The result processed by the attention unit is subjected to a convolution operation of Conv(3x3) with a step size of 1 to obtain F7 with a dimension of 44x1024x1024. The obtained F4, F5, F6 and F7 of different scales are respectively restored to a resolution of 1024x1024 by upsampling operations, and then spliced in the channel dimension. The feature with a dimension of (44+44+56+64)x1024x1024 after the splicing operation is subjected to a convolution operation of Conv(3x3) with a step size of 1, and then converted into a single channel. Finally, the power transmission line icing image segmentation result image with a final dimension of 1x1024x1024 is obtained by a Sigmoid activation function.

[0095] In step 803, the segmentation result image is displayed to the user, so that the user determines the icing condition of the power transmission line based on the segmentation result image.

[0096] Specifically, after obtaining the segmentation result image, the server can display the segmentation result image to the user through a display screen. Alternatively, the segmentation result image is sent to a terminal device such as a mobile phone or a tablet of the user, so that the user views the segmentation result image through the terminal device and determines the icing condition of the power transmission line based on the segmentation result image.

[0097] In the technical scheme of the embodiment, the icing image to be processed of the power transmission line is obtained, the icing image to be processed is input into a target segmentation model, the icing image to be processed is subjected to image segmentation through the target segmentation model, and a segmentation result image of the icing image to be processed is obtained. The segmentation result image is displayed to the user, so that the user determines the icing condition of the power transmission line based on the segmentation result image. The technical scheme of the embodiment can use the target segmentation model to provide the user with an accurate segmentation result image, so that the user can effectively understand the icing condition of the power transmission line according to the segmentation result image, and the user experience is improved.

[0098] Figure 9 A structural schematic diagram of an icing image segmentation device provided by the embodiment of the present application is provided, and the device is suitable for executing the icing image segmentation method provided by the embodiment of the present application. As shown in the figure, the device can specifically include: Figure 9

[0099] The data acquisition module 901 is configured to acquire a set of icing images of the power transmission line when a pre-established initial segmentation model does not meet a preset condition. The initial segmentation model includes a feature extraction module and an up-sampling module.

[0100] The first determination module 902 is configured to perform feature extraction on the icing images in the set of icing images based on a cavity convolution unit and an attention unit of the feature extraction module, and obtain first features.

[0101] The second determination module 903 is configured to obtain target features of the icing images according to the first features and the up-sampling module.

[0102] The model optimization module 904 is configured to optimize the initial segmentation model based on the target features and a pre-set loss function until the initial segmentation model meets the preset condition, so as to obtain a target segmentation model. The target segmentation model is used for image segmentation on the icing images of the power transmission line.

[0103] Optionally, the first determination module 902 is specifically configured to: for each icing image in the set of icing images, input a current icing image into the initial segmentation model, so as to obtain first output features of the current icing image output by a convolution unit of the initial segmentation model.​

[0104] perform feature extraction on the second output feature to obtain a third output feature; and adjust a weight of the third output feature according to the attention unit to obtain the first feature.

[0105] Optionally, the first determining module 902 is further configured to perform dimension reduction convolution and dilated convolution on the first output feature through a first convolution layer of the dilated convolution unit to obtain a first dilated feature.

[0106] perform dimension reduction convolution on the first output feature through a second convolution layer of the dilated convolution unit to obtain a second dilated feature.

[0107] add the first dilated feature and the second dilated feature at an element level to obtain a candidate second output feature, and obtain the second output feature based on the candidate second output feature.

[0108] Optionally, the first determining module 902 is further configured to perform feature extraction on the second output feature through the first convolution layer, the second convolution layer and the third convolution layer respectively to obtain a first convolution feature, a second convolution feature and a third convolution feature.

[0109] perform feature fusion on the second output feature through the max-pooling layer to obtain a fourth pooling feature; and perform feature splicing on the first convolution feature, the second convolution feature, the third convolution feature and the fourth pooling feature to obtain the third output feature.

[0110] Optionally, the first determining module 902 is further configured to perform dimension reduction convolution on the third output feature through a convolution layer of the attention unit, and perform spatial weighting on the third output feature and the third output feature after the dimension reduction convolution to obtain a first attention feature.

[0111] perform global pooling processing on the third output unit through a global pooling layer of the attention unit, and perform channel weighting on the third output feature and the third output feature after the global pooling processing to obtain a second attention feature.

[0112] add the first attention feature and the second attention feature at an element level to obtain the first feature.

[0113] Optionally, the second determining module 902 is specifically configured to, for the first feature of each icing image, determine an input feature of the up-sampling module based on the first feature of the current icing image, the feature extraction module and the dilated convolution unit of the feature extraction module.

[0114] By the upsampling module, based on the pixel position of the current ice-coated image, the input feature of the upsampling module is transposed convolution and dynamic weighting of bilinear interpolation, to obtain a second feature;

[0115] Based on the second feature, the first feature, the feature extraction module, the convolution unit of the upsampling module and the initial segmentation model, the target feature is determined.

[0116] Optionally, the model optimization module 904 is specifically configured to: acquire an ice-coated image to be processed of a power transmission line;

[0117] The ice-coated image to be processed is input into the target segmentation model, and the ice-coated image to be processed is subjected to image segmentation by the target segmentation model to obtain a segmentation result image of the ice-coated image to be processed;

[0118] The segmentation result image is displayed to a user, so that the user determines the ice-coating condition of the power transmission line based on the segmentation result image.

[0119] The ice-coated image segmentation device provided in the embodiment of the present application can execute the ice-coated image segmentation method provided in any embodiment of the present application, and has the corresponding function modules and beneficial effects of the execution method. The contents not described in detail in the present embodiment can be referred to the description in any method embodiment of the present application.

[0120] Figure 10 The structure schematic diagram of an electronic device provided in the embodiment of the present application is shown in Figure 10 , Figure 10 The electronic device 12 shown is only an example, and should not bring any limitation to the function and use range of the embodiments of the present application. For example, Figure 10As shown, the electronic device 12 is in the form of a general-purpose computing device. The components of the electronic device 12 can include, but are not limited to, one or more processors or processing units 16, a system memory 28, and a bus 18 that connects the various system components, including the system memory 28 and the processing unit 16. The bus 18 represents one or more of any of several bus structures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration bus, a processor or local bus using any of a variety of bus architectures, etc. By way of example, these architectures include, but are not limited to, an Industry Standard Architecture (ISA) bus, a Micro Channel Architecture (MAC) bus, an Enhanced ISA bus, a Video Electronics Standards Association (VESA) local bus, and a Peripheral Component Interconnect (PCI) bus. The electronic device 12 typically includes a variety of computer system readable media. Such media can be any available media that is accessible by the electronic device 12 and includes both volatile and non-volatile media, removable and non-removable media. The system memory 28 can include computer system readable media in the form of volatile memory, such as random access memory (RAM) 30 and / or cache memory 32. The electronic device 12 can further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, a storage system 34 can be provided for reading from and writing to non-removable, non-volatile magnetic media (e.g., a "hard drive") Figure 10 (not shown), commonly referred to as a "hard disk drive"). Although not specifically shown, the electronic device 12 can further include other removable / non-removable, volatile / non-volatile computer system storage media including, but not limited to, a magnetic floppy disk drive (e.g., to read from and / or write to a removable magnetic floppy diskette) and an optical disk drive (e.g., to read from and / or write to a removable optical disk (e.g., a CD-ROM)). Figure 10A disk drive, a floppy disk drive, and a CD-ROM drive can be provided as shown in FIG. 1, for reading from or writing to a removable nonvolatile magnetic media (e.g., a "floppy disk"), and to a removable nonvolatile optical media (e.g., CD-ROM, DVD-ROM, or other optical media). Each of the drives can be connected to the system bus 18 through a respective one or more data media interfaces. The memory 28 can include at least one program product having a set (e.g., at least one) of program modules that are configured to carry out the functions of embodiments of the application. Program / utility 40, having a set (at least one) of program modules 42, can be stored in memory 28 by way of example, and includes an operating system, one or more application programs, other program modules, and program data, each or some combination thereof, which may Figure 10 Other hardware and / or software modules can be used in conjunction with the electronic device 12, including but not limited to microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data archival storage systems, etc. The processing unit 16 executes the program code accessed from the system memory 28, in order to carry out various functions and data processing, such as implementing an ice-coated image segmentation method provided by embodiments of the present application: when a pre-established initial segmentation model does not meet a preset condition, obtaining an ice-coated image set of a power transmission line; the initial segmentation model includes a feature extraction module and an up-sampling module; performing feature extraction on an ice-coated image in the ice-coated image set based on a cavity convolution unit and an attention unit of the feature extraction module, to obtain a first feature; obtaining a target feature of the ice-coated image according to the first feature and the up-sampling module; optimizing the initial segmentation model based on the target feature and a pre-set loss function, until the initial segmentation model meets the preset condition, to obtain a target segmentation model, the target segmentation model being used for image segmentation on an ice-coated image of a power transmission line.

[0121] The embodiment of the present application provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the ice-coated image segmentation method provided by all the embodiments of the present application. The computer readable medium can be a computer readable signal medium or a computer readable storage medium. The computer readable storage medium can be, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor electronic device, device or component, or any combination thereof. More specific examples (non-exhaustive list) of the computer readable storage medium include: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this document, the computer readable storage medium can be any tangible medium that contains or stores a program that can be used by or in conjunction with an instruction execution electronic device, device or component. The computer readable signal medium can include a data signal carried in a baseband or as a part of a carrier wave, which carries computer readable program code. Such a propagated data signal can take various forms, including but not limited to an electromagnetic signal, an optical signal or any suitable combination thereof. The computer readable signal medium can also be any computer readable medium other than the computer readable storage medium, which can send, propagate or transmit a program for use by or in conjunction with an instruction execution electronic device, device or component. The program code contained in the computer readable medium can be transmitted by any suitable medium, including but not limited to wireless, wire, optical cable, RF, etc., or any suitable combination thereof. Computer program code for performing the operations of the present application can be written in one or more programming languages or combinations thereof, including object oriented programming languages such as Java, Smalltalk, C++, and conventional procedural programming languages such as "C" programming language or similar programming languages. The program code can be executed entirely on a user computer, partially on a user computer, as a separate software package, partially on a user computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer can be connected to the user computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computer (for example, connected to the Internet through an Internet service provider).

[0122] Note that the above merely describes preferred embodiments of the present application and the principles of the technology applied. Those skilled in the art will understand that the present application is not limited to the specific embodiments described herein, and that various obvious changes, modifications and substitutions can be made thereto without departing from the scope of the present application. Therefore, although the present application has been described in detail through the above embodiments, the present application is not limited to the above embodiments, and can include more other equivalent embodiments without departing from the concept of the present application, and the scope of the present application is determined by the scope of the claims.

Claims

1. A method for segmenting icy images, characterized in that, The method includes: When the pre-established initial segmentation model does not meet the preset conditions, an icing image set of the transmission line is obtained; the initial segmentation model includes a feature extraction module and an upsampling module, the feature extraction module includes: dilated convolutional units, attention units and pooling units; the pooling unit includes: a first convolutional layer, a second convolutional layer, a third convolutional layer and a max pooling layer; For each icing image in the icing image set, the current icing image is input into the initial segmentation model to obtain the first output feature of the current icing image output by the convolutional unit of the initial segmentation model; The first output feature is obtained by performing dimensionality reduction convolution and dilated convolution on the first convolutional layer of the dilated convolutional unit; The first output feature is subjected to dimensionality reduction convolution through the second convolutional layer of the dilated convolutional unit to obtain the second dilated feature; The first hole feature and the second hole feature are added element-wise to obtain a candidate second output feature, and the second output feature is obtained based on the candidate second output feature; The third output feature is obtained by extracting features from the second output feature through the pooling unit, including: extracting features from the second output feature through the first convolutional layer, the second convolutional layer, and the third convolutional layer respectively to obtain a first convolutional feature, a second convolutional feature, and a third convolutional feature; fusing features from the second output feature through the max pooling layer to obtain a fourth pooling feature; and concatenating the first convolutional feature, the second convolutional feature, the third convolutional feature, and the fourth pooling feature to obtain the third output feature. The third output feature is subjected to dimensionality reduction convolution through the convolutional layer of the attention unit, and the third output feature and the dimensionality reduction convolutioned third output feature are spatially weighted to obtain the first attention feature; The third output unit is globally pooled through the global pooling layer of the attention unit, and the third output feature and the third output feature after global pooling are channel-weighted to obtain the second attention feature. The first attention feature and the second attention feature are added element-wise to obtain the first feature; For each icing image, the input features of the upsampling module are determined based on the first feature of the current icing image, the feature extraction module, and the dilated convolution unit of the feature extraction module. This includes: after obtaining the first feature, inputting the first feature into the feature extraction module, then convolving the first feature output by the feature extraction module with a 1×1 convolution kernel, inputting the convolved first feature back into the feature extraction module, then convolving the first feature output by the feature extraction module with a 1×1 convolution kernel, and finally inputting the convolved first feature into the dilated convolution unit to obtain the input features of the upsampling module output by the dilated convolution unit. The upsampling module dynamically weights the input features of the current iced image using transposed convolution and bilinear interpolation based on the pixel positions, yielding a second feature. Specifically, the closer the pixel position is to the edge of the iced image, the greater the weight of the bilinear interpolation; conversely, the closer the pixel position is to the center of the iced image, the greater the weight of the transposed convolution. The dynamic weighting adjustment equation is as follows: ； in, and These represent the height and width of the input features of the upsampling module, respectively. Represents the coordinates after upsampling The corresponding pixel value, Represents the coordinates after transpose convolution The corresponding pixel value, Represents the coordinates after bilinear interpolation The corresponding pixel value; The target feature is determined based on the second feature, the first feature, the feature extraction module, the upsampling module, and the convolutional unit of the initial segmentation model; The initial segmentation model is optimized based on the target features and a pre-defined loss function until the initial segmentation model meets the preset conditions, thereby obtaining the target segmentation model, which is used for image segmentation of icing images of transmission lines.

2. The method according to claim 1, characterized in that, The method further includes: Obtain the icing image of the transmission line to be processed; The ice-covered image to be processed is input into the target segmentation model, and the target segmentation model is used to segment the ice-covered image to obtain the segmentation result image of the ice-covered image to be processed. The segmentation result image is displayed to the user so that the user can determine the icing status of the transmission line based on the segmentation result image.

3. An icing image segmentation device, characterized in that, include: The data acquisition module is used to acquire a set of icing images of the transmission line when the pre-established initial segmentation model does not meet the preset conditions. The initial segmentation model It includes a feature extraction module and an upsampling module. The feature extraction module includes: a dilated convolutional unit, an attention unit, and a pooling unit. The pooling unit includes: a first convolutional layer, a second convolutional layer, a third convolutional layer, and a max pooling layer. The first determining module is configured to, for each icing image in the icing image set, input the current icing image into the initial segmentation model to obtain a first output feature of the current icing image output by the convolutional unit of the initial segmentation model; perform dimensionality reduction convolution and dilated convolution on the first output feature through the first convolutional layer of the dilated convolutional unit to obtain a first dilated feature; perform dimensionality reduction convolution on the first output feature through the second convolutional layer of the dilated convolutional unit to obtain a second dilated feature; add the first dilated feature and the second dilated feature element-wise to obtain a candidate second output feature, and obtain the second output feature based on the candidate second output feature; and extract features from the second output feature through the pooling unit to obtain a third output feature, including: processing the second output feature through the first convolutional layer, the second convolutional layer, and the third convolutional layer respectively. Feature extraction is performed to obtain a first convolutional feature, a second convolutional feature, and a third convolutional feature. The second output feature is then fused using the max pooling layer to obtain a fourth pooling feature. The first, second, third, and fourth convolutional features are concatenated to obtain a third output feature. The third output feature is then subjected to dimensionality-reduced convolution using the convolutional layer of the attention unit, and spatial weighting is applied to both the third output feature and the dimensionality-reduced convolutional feature to obtain a first attention feature. The third output feature is then subjected to global pooling using the global pooling layer of the attention unit, and channel weighting is applied to both the third output feature and the globally pooled feature to obtain a second attention feature. Finally, the first and second attention features are element-wise added to obtain a first feature. The second determining module is used to determine the input features of the upsampling module based on the first feature of the current icing image, the feature extraction module, and the dilated convolution unit of the feature extraction module, for each icing image. This includes: after obtaining the first feature, inputting the first feature to the feature extraction module, then convolving the first feature output by the feature extraction module with a 1×1 convolution kernel, inputting the convolved first feature back to the feature extraction module, then convolving the first feature output by the feature extraction module with a 1×1 convolution kernel, and inputting the convolved first feature into the dilated convolution unit to obtain the input features of the upsampling module output by the dilated convolution unit; through the upsampling module, based on the pixel position of the current icing image, dynamically weighting the input features of the upsampling module using transposed convolution and bilinear interpolation to obtain the second feature; wherein, the closer the pixel position is to the edge of the icing image, the larger the weight of bilinear interpolation; the closer the pixel position is to the center of the icing image, the larger the weight of transposed convolution; the weight adjustment equation for the dynamic weighting is: ； in, and These represent the height and width of the input features of the upsampling module, respectively. Represents the coordinates after upsampling The corresponding pixel value, Represents the coordinates after transpose convolution The corresponding pixel value, Represents the coordinates after bilinear interpolation The corresponding pixel value; the target feature is determined based on the second feature, the first feature, the feature extraction module, the upsampling module, and the convolutional unit of the initial segmentation model; The model optimization module is used to optimize the initial segmentation model based on the target features and a pre-set loss function until the initial segmentation model meets the preset conditions to obtain the target segmentation model, which is used to segment images of icing on transmission lines.

4. An electronic device, the electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements an ice-covered image segmentation method as described in any one of claims 1 to 2.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements an ice-covered image segmentation method as described in any one of claims 1 to 2.

Images