Analytical method for design image of power transmission tower

Abstract

The invention discloses an analysis method for a design image of a power transmission tower. The analysis content comprises information extraction of shape, parameterization, layout and the like of a design structure. Firstly, an original design drawing is converted into a clean topological mask drawing through a semantic segmentation network, and the shape of a design structure is obtained; then, through a vector extraction network, extracting a skeleton center vector and endpoint information from the topological mask map; then, numbers and symbols on the image and in the table are recognized and extracted, corresponding numbers beside the mask image are numbered, the same numbers are found in the table, the actual size corresponding to the numbers is obtained, and parameterization assignment of the actual size of the shape vector is achieved; and finally, constructing a visual angle classification network, analyzing the image layout of each shape, and obtaining the visual angle orientation of the three-dimensional model corresponding to the planar graph. The method lays a foundation for realizing automatic three-dimensional modeling of the drawing, and can be expanded to be used for intelligently verifying, improving and checking the design drawing of the iron tower and the like.

Description

Technical Field

[0001] The present invention relates to the fields of power grid, image processing, computer vision and artificial intelligence technology, and in particular to a method for analyzing transmission tower design images. Background Art

[0002] Smart grids are the driving force behind the evolution of today's power systems. Digitalization, visualization, intelligence, and information interconnection have become design goals throughout every phase of the grid's lifecycle. In this digital era, there's a growing demand for automated analysis of paper design drawings for transmission towers.

[0003] During the digitization phase of transmission tower paper drawings, automated analysis of transmission tower design drawings is crucial to improve efficiency and accuracy. The automatic conversion of 2D design drawings into 3D solid models has already begun to be applied in current digital construction processes.

[0004] However, existing image processing tools cannot accurately meet the requirements of three-dimensional solid model accuracy. Further improving the accuracy of extracting the skeleton center vector of the transmission tower drawing and further obtaining the viewing angle orientation of the drawing to improve the accuracy and consistency of the data can not only save a lot of manpower and time costs, but also improve the accuracy and consistency of the data, and promote the digital construction of the power system to a new level. Summary of the Invention

[0005] The problem to be solved by the present invention is to provide a method for analyzing transmission tower design images, which can automatically obtain the skeleton center vector and viewing angle of the transmission tower drawing without relying on manual operation by professionals, thereby further improving the accuracy of the three-dimensional solid model.

[0006] The present invention adopts the following technical solution: a method for analyzing a transmission tower design image, comprising the following steps:

[0007] Step 1: Using a shape segmentation network, the input transmission line tower design original image is converted into a topological mask graphic;

[0008] Step 2: extracting the skeleton centerline vector and endpoint information of the transmission line tower member from the topological mask graphic output in step 1 through a vector extraction network to form a contour vector set of the shape;

[0009] Step 3: Identify identifiers in the original image of the transmission line tower design using an image processing method, and identify text on the original image, including numbers and letters next to the graphic shape vectors, and numbers and text in the component list. Use the numerical numbers next to the shape vectors to find the corresponding numerical numbers in the component list. Determine the actual size corresponding to each shape vector in a table, and assign the actual size to the shape vector to achieve parameterized assignment of each vector.

[0010] Step 4: Combining the contour vector set and the identified digital number, a view classification network is constructed to achieve view classification of the plane projection of the local structure of each vector;

[0011] Step 5: Save the obtained contour vector set, the parameterized value of each vector, and the perspective classification results of the plane projection of the local structure of each vector accordingly to complete the content analysis of the design drawing.

[0012] Preferably, the design image input in step 1 is a two-dimensional shape structure diagram of a plane transmission line tower actually used in a power grid, which includes a steel frame structure diagram, steel bar numbers, and component list information, and is stored in JPG, PNG, or TIF format.

[0013] Preferably, the shape segmentation network in step 1 is based on the first encoder-decoder framework. The original image of the transmission line tower design is first encoded by the encoder E to encode the feature vector of the input design image, and then the decoder SD is used to obtain the binary mask of the geometric shape. Finally, the binary cross entropy is used to calculate the loss L. Seg ;

[0014] The encoder E consists of a 7*7 convolutional layer and two ResNet modules, and the decoder SD consists of two upsampling module layers and a 7*7 deconvolution layer.

[0015] Preferably, the ResNet module includes 6 residual blocks, each of which uses 1×1, 3×3, and 1×1 convolutions to extract image features, and adds a BatchNorm layer and ReLU after each convolution operation. At the same time, a skip layer connection is added to add the input features and the features extracted by convolution as the final output result, which is used to solve the problem of vanishing model gradients when the network depth increases;

[0016] The calculation process of the residual block is as follows:

[0017] y=F(x,{W i})+x,

[0018] Among them, x and y are the input and output of each layer respectively, and the function F(x,{W i}) is a learnable residual mapping, which is expressed as follows:

[0019] F=σ(BN(W3σ(BN(W2σ(BN(W1σ)))))),

[0020] Among them, W1, W2, and W3 represent 1×1, 3×3, and 1×1 convolution parameter matrices respectively, BN(﹒) represents the BatchNorm normalization operation, and σ(﹒) represents the activation function.

[0021] Preferably, the upsampling module uses transposed convolution to upsample the input feature map to twice the resolution, and outputs a feature map with the same resolution as the original image, and the depth of the feature map is the number of classification categories;

[0022] The calculation process of the upsampling module is as follows:

[0023] y′=σ(BN(W4x′)),

[0024] Where x′ and y′ represent the input and output of each upsampling module, respectively, and W4 represents the transposed convolution parameter matrix;

[0025] The shape segmentation network uses binary cross entropy to calculate the loss L Seg , the formula is as follows:

[0026]

[0027] in, is the generated mask and m is the ground-truth segmentation mask.

[0028] Preferably, the vector extraction network in step 2 is specifically performed as follows:

[0029] Step 2.1, extracting the mask image of the transmission line tower truss through the geometric shape extraction module;

[0030] Step 2.2, using the second encoder-decoder structure, respectively predicting the node heat map and the skeleton line heat map of the transmission line tower truss;

[0031] Step 2.3: construct a cross-fusion module to establish the correct connection between the transmission line tower truss nodes, extract the skeleton centerline vector and endpoint information of the transmission line tower rods, and form a contour vector set of the shape;

[0032] The skeleton centerline vector is a connecting line segment composed of two end points, which is used to express the spatial position of the transmission line tower truss nodes and the mutual connection relationship between the nodes.

[0033] Preferably, the second encoder-decoder structure first passes the mask image of the transmission line tower truss through the encoder E′ for feature encoding, and then passes it through the decoder SD′ to obtain the binary value of each point;

[0034] The encoder E′ downsamples the features extracted from the convolutional layer through three pyramid residual modules, and generates feature values ​​of dimension W×H×C after a maximum pooling operation and five stacked hourglass modules composed of PRMs, forming a grid composed of high-level descriptors;

[0035] After the grid passes through the decoder SD′ composed of two fully convolutional neural networks, the binary value of each point is estimated on the grid map generated by the decoder. It is 1 at the node or center skeleton line and 0 otherwise. The node heat map and skeleton line heat map with a resolution of H×W are output to show the probability of the pixel point belonging to the node or skeleton line. The distribution probability of the point belonging to the node or skeleton line is different at different pixels. The location of the pixel point with the highest probability is selected to describe the node position and skeleton line structure of the mask shape of the transmission line tower truss.

[0036] The vector extraction network regards the node position or skeleton line structure prediction as a classification problem and uses the average binary cross entropy as the loss function. The formula is as follows:

[0037]

[0038] Among them, y i is a binary label 0 or 1, p(y i ) is the output belonging to y i The probability of the label, N represents the total number of samples.

[0039] Preferably, the cross fusion module fuses the node heat map and the skeleton line heat map information in the following manner:

[0040] Connect all predicted nodes and construct a list, and generate a candidate line list after uniform sampling;

[0041] Align the candidate line list with the skeleton line heatmap features, rasterize the lines in the candidate line list onto the skeleton line heatmap, and use the heatmap value as the confidence of the pixel skeleton line for screening and filtering;

[0042] The confidence of each candidate line is defined as the average pixel density along the candidate line. By setting a threshold, the correct skeleton line is obtained.

[0043] Preferably, the perspective classification in step 4 is to classify the perspectives of each plane in the transmission line tower design image into: front view perspective, side view perspective, and top view perspective, including the following sub-steps:

[0044] Step 4.1: embed the identifier corresponding to the contour vector set into the topological mask pattern obtained in step 1 through RGB pixels, so that the topological mask pattern carries the position information of the associated number, and use the result as input;

[0045] Step 4.2: Using ResNet50 as the backbone network, an attention module is connected in series to the residual module of each layer of the backbone network, so that the backbone network can focus on the local feature information of the identifier.

[0046] Preferably, the attention module adopts CBAM, including channel attention and spatial attention;

[0047] The channel attention adopts average pooling and maximum pooling to aggregate the channel information of the feature map F obtained by the residual module, generates two different spatial context descriptors, obtains the fused feature vector through a shared multi-layer perceptron, and inputs it into the sigmoid function σ to output the channel attention feature F CR ;

[0048] Channel Attention M C (F), calculated as follows:

[0049] M C (F)=σ[MLP(GAP(F))+MLP(GMP(F))],

[0050] Among them, MLP stands for multi-layer perceptron, GAP stands for average pooling, and GMP stands for maximum pooling;

[0051] The spatial attention performs GAP and GMP processing on the feature map F respectively, connects the results together, and highlights the key information area; then uses a 7×7 convolution layer f to perform convolution operation on the features, and inputs the convolution output features into the sigmoid function σ;

[0052] Spatial Attention M S (F), calculated as follows:

[0053]

[0054] Among them, the symbol Represents the concatenation operator;

[0055] For a given input feature vector CBAM is based on one-dimensional channel attention and two-dimensional spatial attention Output the final attention feature:

[0056]

[0057] Among them, the symbol Represents the multiplication operation, F CR is the channel attention feature, F output is the output attention feature.

[0058] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0059] 1. Compared with traditional manual drawing or existing basic drawing software, the method for analyzing transmission tower design images of the present invention can automatically identify and process complex transmission tower structures in design transmission tower drawings, laying the foundation for realizing automatic three-dimensional modeling of drawings. It can also quickly generate high-quality skeleton vectors and viewing angles without requiring a large amount of manpower, saving time and costs, and improving the level of intelligence and work efficiency.

[0060] 2. The skeleton vectors and viewing angles generated by the analytical method for the transmission tower design image of the present invention can be directly applied to various engineering and academic research fields, such as terrain analysis, structural simulation, planning and design, etc., providing a good foundation for subsequent work (such as three-dimensional modeling). It can also be expanded to intelligent verification, improvement, and inspection of tower design drawings, etc., has broad application prospects and value, and is of great significance to the comprehensive promotion of the digital transformation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 A shape segmentation network structure diagram of an analytical method for designing an image of a transmission tower according to the present invention;

[0062] Figure 2 A vector extraction network structure diagram of an analytical method for designing an image of a transmission tower according to the present invention;

[0063] Figure 3 A vector flow chart for extracting a transmission tower drawing according to an embodiment of the present invention;

[0064] Figure 4 A perspective classification network structure diagram of the analytical method for designing images of transmission towers according to the present invention;

[0065] Figure 5 A diagram showing the relationship between the plane viewing angle and its contour vector identifier of the present invention;

[0066] Figure 6 This is the network structure diagram of the attention module CBAM of the analysis method for the transmission tower design image of the present invention. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the application are further elaborated in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in the present invention. All non-innovative embodiments of other researchers in this field on this embodiment fall within the scope of protection of the present invention. At the same time, the step numbers in the embodiments of the present invention are only set for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0068] The present invention provides a method for analyzing a transmission tower design image, comprising the following steps:

[0069] Step 1: Convert the input transmission line tower design image into a topological mask graphic through a shape segmentation network;

[0070] Step 2: extracting the skeleton centerline vector and endpoint information of the transmission line tower member from the topological mask graphic output in step 1 through a vector extraction network to form a contour vector set of the shape;

[0071] Step 3: Identify identifiers in the original image of the transmission line tower design using an image processing method, and identify text on the design image, including numbers and letters next to the graphic shapes, and numbers and text in the component list. Use the numerical numbers next to the shape vectors to find the corresponding numerical numbers in the table, thereby determining the actual size (in meters) corresponding to the shape vector in the table, and assigning the size to the shape vector to achieve vector parameterization assignment.

[0072] Step 4: Combining the contour vector set and the identified number, a view classification network is constructed to classify the projection view of the plane where each local structure is located;

[0073] Step 5: Save the shape outline vector set determined in the above steps, the parameterized value assignment of each vector, and the local structure perspective classification result including each vector accordingly to complete the content analysis of the design drawing.

[0074] In one embodiment of the present invention, a method for analyzing transmission tower design drawings specifically includes the following steps:

[0075] 1. A shape segmentation network is used to convert the original input image of the transmission line tower design into a clean topological mask.

[0076] In this embodiment, the design drawings used are two-dimensional design drawings of transmission line towers actually used in the power grid, which are stored in JPG, PNG, or TIF formats and include steel frame structure drawings, steel bar numbers, and component details.

[0077] Shape segmentation networks, such as Figure 1 As shown, based on the first encoder (E)-decoder (SD) framework, for a truss design drawing, the network first encodes the feature vector of the input design image through the encoder E, and then obtains the binary mask of the geometric shape through the decoder SD.

[0078] In this embodiment, the encoder E is composed of a 7*7 convolutional layer and two ResNet modules; the decoder SD is composed of two upsampling module layers and a 7*7 deconvolution layer;

[0079] Specifically, the ResNet module consists of six residual blocks, each of which sequentially extracts image features using 1×1, 3×3, and 1×1 convolutions. A BatchNorm layer and ReLU are added after each convolution operation to ensure better model convergence. Furthermore, skip-layer connections are added to sum the input features with the features extracted by the convolution as the final output, addressing the problem of vanishing gradients as the network depth increases.

[0080] The calculation process of the residual block can be expressed as:

[0081] y=F(x,{W i})+x,

[0082] Among them, x and y are the input and output of each layer respectively, and the function F(x,{W i}) represents a learnable residual mapping, which is expressed as:

[0083] F=σ(BN(W3σ(BN(W2σ(BN(W1σ)))))),

[0084] Among them, W1, W2, and W3 represent 1×1, 3×3, and 1×1 convolution parameter matrices respectively, BN(﹒) represents the BatchNorm normalization operation, and σ(﹒) represents the activation function, which can be ReLU.

[0085] Specifically, the upsampling module uses transposed convolution to upsample the input feature map to twice the resolution, and finally outputs a feature map with the same resolution as the original image. The depth of the feature map is the number of classification categories. The calculation process of the upsampling module can be expressed as follows:

[0086] y=σ(BN(W4x)),

[0087] Where x and y are the input and output of each upsampling module, respectively, and W4 represents the transposed convolution parameter matrix. Reverse convolution is the process of reversing the forward and backward propagation of convolution.

[0088] In this embodiment, the shape segmentation network module uses binary cross entropy to calculate the loss:

[0089]

[0090] in, is the generated mask and m is the ground-truth segmentation mask.

[0091] 2. Through a vector extraction network, the skeleton center vector of the high-voltage tower rod is extracted from the topological mask image.

[0092] The skeleton centerline vector is a connecting line segment composed of two end points, which expresses the spatial position of the truss nodes and their mutual connection relationship.

[0093] Vector extraction network, such as Figure 2 As shown, the working steps are as follows:

[0094] First, the mask image of the transmission line tower truss is extracted through the geometric shape extraction module;

[0095] Then, the node heat map and skeleton line heat map of the transmission line tower truss are predicted respectively through the second encoder-decoder structure;

[0096] Finally, a cross-fusion module is constructed to suggest the correct connections between the truss nodes of the transmission line tower and extract the skeleton center vectors of the transmission line tower members.

[0097] In this embodiment, in the initial stage of the vector extraction network, the mask image generated by the geometric shape extraction module is first feature encoded by the encoder E′, and the encoder E′ downsamples the features extracted from the convolutional layer through three pyramid residual modules (PRMs). After the maximum pooling operation and five stacked hourglass modules composed of PRMs, feature values ​​with a dimension of W×H×C are generated to form a grid composed of high-level descriptors.

[0098] Then, the grid passes through the decoder SD′ composed of two fully convolutional neural networks (FCNs), and estimates the binary value of each point on the grid map generated by the decoder (1 on the node or center skeleton line, otherwise 0), thereby outputting a node heat map and a skeleton line heat map with a resolution of H×W, thereby showing the probability of a pixel point belonging to a node or skeleton line.

[0099] Next, since the distribution probabilities of points belonging to nodes or skeleton lines are different at different pixels, the location of the pixel with the highest probability is selected to describe the node position and skeleton structure of the mask shape of the transmission line tower truss, respectively.

[0100] Here, node or skeleton line prediction can be regarded as a classification problem, so the average binary cross entropy is used as the loss function:

[0101]

[0102] Among them, y i is a binary label 0 or 1, p(y i ) is the output belonging to y i The probability of the label, N represents the total number of samples.

[0103] In this embodiment, the cross-fusion module mainly fuses the information of the node heat map and the skeleton line heat map. The design idea of ​​this module is: first, all the predicted nodes are connected and a list is constructed, and a candidate line list is generated after uniform sampling, which is aligned with the skeleton line heat map features, that is, the lines in the candidate line list are rasterized onto the skeleton line heat map, and the heat map value is used as the confidence that the pixel is a skeleton line, so as to perform screening and filtering.

[0104] It should be noted that this embodiment defines the confidence of each candidate line as the average pixel density along the candidate line. By setting a simple threshold, the correct skeleton line can be obtained. This embodiment extracts the skeleton center vector of the high-voltage tower rod from the topological mask image. The result is as follows: Figure 3 shown.

[0105] 3. Recognize the numbers, letters and other text on the design drawing as well as the text in the table, identify the identification symbols such as numbers through image processing methods, and obtain a set of contour vectors.

[0106] 4. For the mask image and its outer contour vector corresponding identifier, the viewing angle of each plane is obtained through the classification network, such as Figure 4 shown.

[0107] Among them, the viewing angle is to divide the viewing angles of each plane in the drawing into three categories: front view, side view, and top view.

[0108] The viewpoint recognition module of this embodiment first embeds the identifier corresponding to the contour vector set into the topological mask pattern obtained in step 1 through RGB pixels, so that the mask pattern carries the position information of the associated number. The result is used as input. ResNet50 is used as the backbone network, and an attention module is connected in series to the residual module of each layer, so that the backbone network can better focus on the local feature information of the identifier.

[0109] The input of the network is a planar structure extracted from the drawing, including the mask of its geometric shape and the identifier corresponding to the outer contour vector, such as Figure 5 shown.

[0110] Preferably, the residual module of each layer of the ResNet50 neural network is connected in series with an attention module to strengthen the network's attention to the contour vector identifier. The attention module uses CBAM, such as Figure 6 As shown in Figure 3, it includes two parts: channel attention (C-Attention) and spatial attention (S-Attention).

[0111] The attention module first uses Global Average Pooling (GAP) and Global Max Pooling (GMP) to aggregate the channel information of the feature map F to generate two different spatial context descriptors.

[0112] Subsequently, the two descriptors are passed through a shared multi-layer perceptron to obtain a fused feature vector. Finally, the feature vector is input into the sigmoid function (σ) to output a channel attention feature F CR .

[0113] The calculation formula of channel attention is as follows:

[0114] M C (F)=σ[MLP(GAP(F))+MLP(GMP(F))].

[0115] Spatial attention focuses on finding the meaningful content in a given image. First, it performs GAP and GMP on the feature maps F separately, then concatenates them together to effectively highlight the key information areas. Next, a 7×7 convolutional layer is used to convolve these features. Finally, the output features of the convolution are input into the sigmoid function (σ).

[0116] The calculation formula of spatial attention is as follows:

[0117]

[0118] Among them, the symbol Represents the concatenation operator.

[0119] For a given input feature vector CBAM is based on the above one-dimensional channel attention and two-dimensional spatial attention Get the final output features:

[0120]

[0121] Among them, the symbol Represents the multiplication operation, F CR is the channel attention feature, F output is the output attention feature.

[0122] Furthermore, in this embodiment, the shape outline vector set determined in the above steps, the parameterized value assignment of each vector, and the local structure perspective classification result including each vector are correspondingly saved to complete the content analysis of the design drawing.

[0123] It should be noted that the deep learning network structure provided by the present invention can also have multiple variations and improvements. Without inventive efforts, those skilled in the art can also adopt different drawing data, different network structures, different algorithm variations, etc. in the above manner, which are all included in the scope of patent protection of the present invention.

[0124] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for analyzing a transmission tower design image, characterized in that: The steps include: Step 1: Convert the input transmission line tower design original image into a topological mask graphic through a shape segmentation network; Step 2: extracting the skeleton centerline vector and endpoint information of the transmission line tower rod from the topological mask graph output in step 1 through a vector extraction network to form a contour vector set of the shape; Step 3, identifying identifiers in the original image of the transmission line tower design by an image processing method, and identifying text on the original image of the design, including: numbers and letters next to the graphic shape vector, and numbers and text in the component list, finding the corresponding number in the component list through the number next to the shape vector, determining the actual size corresponding to each shape vector in the table, assigning the actual size to the shape vector, and realizing parameterized assignment of each vector; Step 4: Combining the contour vector set and the identified digital number, a view classification network is constructed to achieve view classification of the plane projection of the local structure of each vector; Step 5: Save the obtained contour vector set, the parameterized value of each vector, and the perspective classification results of the plane projection of the local structure of each vector accordingly to complete the content analysis of the design drawing.

2. The method for analyzing the transmission tower design image according to claim 1, characterized in that: The design image input in step 1 is a two-dimensional shape structure diagram of a plane transmission line tower actually used in the power grid, which includes a steel frame structure diagram, steel bar numbers, and component list information, and is stored in JPG, PNG, and TIF formats.

3. The method for analyzing the design image of a transmission tower according to claim 2, characterized in that: The shape segmentation network described in step 1 is based on the first encoder-decoder framework. The original image of the transmission line tower design first passes through the encoder E to encode the feature vector of the input design image, and then passes through the decoder SD to obtain the binary mask of the geometric shape. Finally, the binary cross entropy is used to calculate the loss L Seg ; The encoder E consists of a 7*7 convolutional layer and two ResNet modules, and the decoder SD consists of two upsampling module layers and a 7*7 deconvolutional layer.

4. The method for analyzing the transmission tower design image according to claim 3, characterized in that: The ResNet module includes 6 residual blocks, each of which uses 1×1, 3×3, and 1×1 convolutions to extract image features in turn, and adds a BatchNorm layer and ReLU after each convolution operation. At the same time, a skip layer connection is added to add the input features to the features extracted by the convolution as the final output result, which is used to solve the problem of model gradient disappearance when the network depth increases; The calculation process of the residual block is as follows: y=F(x,{W i })+x, Among them, x and y are the input and output of each layer respectively, and the function F(x,{W i }) is a learnable residual mapping, expressed as follows: F=σ(BN(W3σ(BN(W2σ(BN(W1σ)))))), Among them, W1, W2, and W3 represent 1×1, 3×3, and 1×1 convolution parameter matrices respectively, BN(﹒) represents the BatchNorm normalization operation, and σ(﹒) represents the activation function.

5. The method for analyzing the transmission tower design image according to claim 3, characterized in that: The upsampling module uses transposed convolution to upsample the input feature map to twice the resolution, and outputs a feature map with the same resolution as the original image, and the depth of the feature map is the number of classification categories; The calculation process of the upsampling module is as follows: y′=σ(BN(W4x′)), Where x′ and y′ represent the input and output of each upsampling module, respectively, and W4 represents the transposed convolution parameter matrix; The shape segmentation network uses binary cross entropy to calculate the loss L Seg , the formula is as follows: in, is the generated mask and m is the ground-truth segmentation mask.

6. The method for analyzing a transmission tower design image according to claim 1, characterized in that: The vector extraction network described in step 2 is specifically described as follows: Step 2.1, extracting the mask image of the transmission line tower truss through the geometric shape extraction module; Step 2.2, predicting the node heat map and the skeleton line heat map of the transmission line tower truss respectively through the second encoder-decoder structure; Step 2.3, construct a cross-fusion module, establish the correct connection between the nodes of the transmission line tower truss, extract the skeleton centerline vector and endpoint information of the transmission line tower rod, and form a contour vector set of the shape; The skeleton centerline vector is a connecting line segment composed of two end points, which is used to express the spatial position of the transmission line tower truss nodes and the mutual connection relationship between the nodes.

7. The method for analyzing transmission tower design images according to claim 6, characterized in that: The second encoder-decoder structure first passes the mask image of the transmission line tower truss through the encoder E′ for feature encoding, and then passes the decoder SD′ to obtain the binary value of each point; The encoder E′ downsamples the features extracted from the convolutional layer through three pyramid residual modules, generates feature values ​​of dimension W×H×C through a maximum pooling operation and five stacked hourglass modules composed of PRMs, and forms a grid composed of high-level descriptors; After the grid passes through the decoder SD′ composed of two fully convolutional neural networks, the binary value of each point is estimated on the grid map generated by the decoder, which is 1 at the node or the center skeleton line, otherwise it is 0, and the node heat map and skeleton line heat map with a resolution of H×W are output, thereby showing the probability of the pixel point belonging to the node or the skeleton line. The distribution probability of the points belonging to the node or the skeleton line is different in different pixels. The position of the pixel point with the largest probability is selected to describe the node position and skeleton line structure of the mask shape of the transmission line tower truss; The vector extraction network regards the node position or skeleton line structure prediction as a classification problem and uses the average binary cross entropy as the loss function. The formula is as follows: Among them, y i is a binary label 0 or 1, p(y i ) is the output belonging to y i The probability of the label, N represents the total number of samples.

8. The method for analyzing transmission tower design images according to claim 6, characterized in that: The cross fusion module fuses the node heat map and the skeleton line heat map information in the following way: Connect all predicted nodes and construct a list, and generate a candidate line list after uniform sampling; Align the candidate line list with the skeleton line heat map features, rasterize the lines in the candidate line list onto the skeleton line heat map, and use the heat map value as the confidence of the pixel skeleton line for screening and filtering; The confidence of each candidate line is defined as the average pixel density along the candidate line, and the correct skeleton line is obtained by setting a threshold.

9. The method for analyzing transmission tower design images according to claim 1, characterized in that: The perspective classification described in step 4 is to classify the perspectives of each plane in the transmission line tower design image into: front view perspective, side view perspective, and top view perspective, including the following sub-steps: Step 4.1, embed the identifier corresponding to the contour vector set into the topological mask graphic obtained in step 1 through RGB pixels, so that the topological mask graphic carries the position information of the associated number, and use the obtained result as input; Step 4.2: Using ResNet50 as the backbone network, an attention module is connected in series to the residual module of each layer of the backbone network, so that the backbone network can focus on the local feature information of the identifier.

10. The method for analyzing transmission tower design images according to claim 7, characterized in that: In step 4.2, the attention module adopts CBAM, including channel attention and spatial attention; The channel attention uses average pooling and maximum pooling to aggregate the channel information of the feature map F obtained by the residual module, generate two different spatial context descriptors, obtain a fused feature vector through a shared multi-layer perceptron, and input it into the sigmoid function σ to output the channel attention feature F CR ; Channel Attention M C (F), calculated as follows: M C (F)=σ[MLP(GAP(F))+MLP(GMP(F))], Among them, MLP stands for multi-layer perceptron, GAP stands for average pooling, and GMP stands for maximum pooling; The spatial attention performs GAP and GMP processing on the feature map F respectively, connects the results together, and highlights the key information area; then uses a 7×7 convolution layer f to perform convolution operation on the features, and inputs the output features of the convolution into the sigmoid function σ; Spatial AttentionM S (F), calculated as follows: Among them, the symbol Represents the concatenation operator; For a given input feature vector CBAM is based on one-dimensional channel attention and 2D spatial attention Output the final attention feature: Among them, the symbol Represents the multiplication operation, F CR is the channel attention feature, F output is the output attention feature.

Images