Power grid infrastructure project scene ground feature intelligent identification method and system based on AI technology

Abstract

The invention relates to the technical field of scene intelligent identification, and particularly discloses a power grid infrastructure project scene ground feature intelligent identification method and system based on an AI technology, and the method mainly comprises the steps: obtaining a remote sensing image of a power grid infrastructure project scene; according to the method, a convolutional neural network model is adopted to identify the surface feature type of a remote sensing image, in an initial stage, the convolutional neural network model activates a preset number of convolutional layers and pooling layers, and if there is an unknown surface feature type, the number and depth of the convolutional layers and pooling layers are further activated to extract more surface feature features. The depth and the number of extracted features can be dynamically adjusted according to the ground feature condition in the image, so that the method can better adapt to different types of remote sensing images, ground feature type identification and coordinate positioning tasks are separated, the calculation complexity is effectively reduced, the efficiency and the speed of the model are improved, various ground feature types are better identified, and the accuracy of the model is improved. And remarkable economic and social benefits are brought.

Description

Technical Field

[0001] The present application relates to the technical field of scene intelligent recognition, and in particular to a method for intelligent recognition of objects in power grid infrastructure project scenes based on AI technology. Background Art

[0002] Power grid infrastructure projects are projects that plan, design, build and transform power systems to meet power supply needs. Using static satellite remote sensing data for power grid infrastructure planning helps provide detailed basic information such as topography, land use and vegetation coverage, supports preliminary facility location and route selection, reduces decision makers' blindness to environmental and geographical conditions, and thus improves the accuracy and reliability of planning.

[0003] Although existing technologies have made certain progress in ground object recognition in power grid infrastructure project scenarios, the static model structure cannot flexibly increase the network depth when facing unknown ground object types, which limits the adaptability and generalization ability of the model. It also shows limitations in responding to real-time changes and dynamic demands in power grid infrastructure planning, and cannot effectively cope with the diversity and variability of ground object types. Summary of the invention

[0004] This application provides an AI-based method and system for intelligently identifying objects in power grid infrastructure projects, solving the problem that static model structures in the prior art cannot flexibly increase network depth when facing unknown object types. It can dynamically adjust the depth and number of extracted features according to the objects in the image, thereby better adapting to different types of remote sensing images. It reduces resource consumption when processing common objects, not only saving storage and processing resources, but also maintaining high efficiency and accuracy in complex image processing.

[0005] The present application provides a method for intelligently identifying ground objects in power grid infrastructure project scenes based on AI technology, which includes the following steps: acquiring a remote sensing image of a power grid infrastructure project scene; using a convolutional neural network model to identify the ground object type of the remote sensing image, including: the convolutional neural network model uses stacked convolutional layers and pooling layers, in the initial stage, the convolutional neural network model activates a preset number of convolutional layers and pooling layers, after extracting ground object features of the remote sensing image, the extracted ground object features are transferred to the connection layer to obtain a ground object type probability value; judging whether the ground object type probability value is greater than a preset judgment probability, and identifying the ground object types with probability values ​​greater than and less than the judgment probability as known ground object types and unknown ground object types respectively; if there is an unknown ground object type, further activating the number and depth of the convolutional layers and the pooling layers to extract more ground object features, and identifying the ground object type probability value of the unknown ground object type based on the ground object features, until a feature map of all known ground object types is obtained; outputting a result, the result including a remote sensing image marked with a known ground object type.

[0006] By increasing the number and depth of convolutional layers and pooling layers, and optimizing the feature extraction strategy, the depth and number of extracted features can be dynamically adjusted according to the ground objects in the image. When processing unknown ground object types, the network structure can be dynamically adjusted to better identify various ground object types, providing a more accurate and efficient ground object identification solution for power grid infrastructure projects.

[0007] Furthermore, the output remote sensing image also has coordinates of known land object types.

[0008] Furthermore, the convolutional neural network model also includes an additional fully connected layer, the feature map of the known land feature type is converted into a one-dimensional vector through a shape transformation matrix, the one-dimensional vector is input into the fully connected layer, a linear transformation is performed through a weight matrix to obtain an intermediate output, an activation function is applied to the intermediate output, and the activated output is used as the input of the output layer to obtain the coordinates of the known land feature type.

[0009] Furthermore, the number of activation layers of the convolution layer and the pooling layer when identifying all specified objects is obtained, and the number of activation layers is used as the number of stacked layers of the convolution layer and the pooling layer.

[0010] Furthermore, further activating the number and depth of convolutional layers and pooling layers specifically includes: obtaining the maximum value of the probability value of the ground object type of all unknown ground object types, dividing the maximum value of the ground object type probability value of all unknown ground object types into a specified number of clusters according to the numerical value; and determining the number of times to activate the convolutional layer and the pooling layer based on the number of clusters.

[0011] Furthermore, a clustering algorithm based on data density is used to determine the number of clusters.

[0012] Furthermore, determining the number of clusters specifically includes: taking the maximum value of the feature type probability value of the unknown feature type as the data point, defining parameters ε and MinPts, ε represents the distance threshold of the neighborhood, and MinPts represents the minimum number of data points in the ε neighborhood; visiting each data point, randomly selecting an unvisited data point each time, calculating the number of data points in the ε neighborhood of the point, and if it is greater than or equal to MinPts, marking the point as a core point; starting from the core point, adding its density-reachable points to the same cluster in a density-reachable manner, and adding points that do not belong to any cluster to the nearest cluster.

[0013] Furthermore, each cluster corresponds to the number of activation layers n i , n i For: n i =a i ·(Ls); Where n i represents the number of activation layers corresponding to the i-th cluster, a irepresents the length ratio of the i-th cluster, L represents the number of stacked convolutional layers and pooling layers, s represents the number of convolutional layers and pooling layers activated in the initial stage, and A represents the number of clusters.

[0014] Furthermore, the ground feature features include edge features, texture features, shape features, spectral features, spatial distribution features and context features.

[0015] In the second aspect, the present application provides an AI-based intelligent ground object recognition system for power grid infrastructure project scenes, which adopts the AI-based intelligent ground object recognition method for power grid infrastructure project scenes as described in the first aspect, and includes: an image acquisition module and a ground object recognition module. The image acquisition module is used to acquire remote sensing images of power grid infrastructure engineering scenes; the object recognition module is used to use a convolutional neural network model to identify the object type of the remote sensing image, including: the convolutional neural network model uses stacked convolutional layers and pooling layers. In the initial stage, the convolutional neural network model activates a preset number of convolutional layers and pooling layers. After extracting the object features of the remote sensing image, the extracted object features are transferred to the connection layer to obtain the object type probability value; it is determined whether the object type probability value is greater than the preset judgment probability, and the object types with probability values ​​greater than and less than the judgment probability are respectively identified as known object types and unknown object types. If an unknown object type exists, the number and depth of the convolutional layers and pooling layers are further activated to extract more object features, and the object type probability value of the unknown object type is identified based on the object features until a feature map of all known object types is obtained; and the result is output, and the result includes a remote sensing image marked with a known object type.

[0016] In the third aspect, the present application provides an AI-based device for intelligently identifying ground objects in power grid infrastructure project scenes, including a memory and a processor; the memory is used to store computer programs; the processor is used to implement the steps of the AI-based method for intelligently identifying ground objects in power grid infrastructure project scenes as in the first aspect when executing the computer program.

[0017] The technical solution provided by this application has at least the following technical effects or advantages:

[0018] By increasing the number and depth of convolutional layers and pooling layers, and optimizing the feature extraction strategy, the problem in the existing technology that the static model structure cannot flexibly increase the network depth when facing unknown land object types is effectively solved. The depth and number of extracted features can be dynamically adjusted according to the land object conditions in the image, so as to better adapt to different types of remote sensing images, separate the tasks of land object type identification and coordinate positioning, effectively reduce the computational complexity, and improve the efficiency and speed of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1This is a flow chart of the method for intelligently identifying objects in power grid infrastructure project scenes based on AI technology in this application;

[0020] Figure 2 For this application Figure 1 Flowchart of the number of further activations of convolutional and pooling layers in . DETAILED DESCRIPTION

[0021] The present application can solve the defect in the prior art that the static model structure cannot flexibly increase the network depth when facing unknown land object types. It can dynamically adjust the depth and number of extracted features according to the land object conditions in the image. When processing unknown land object types, it can dynamically adjust the network structure and separate the tasks of land object type identification and coordinate positioning, effectively reducing the computational complexity and improving the efficiency and speed of the model, thereby better identifying various land object types and bringing significant economic and social benefits.

[0022] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0023] Embodiment 1

[0024] A method for intelligently identifying objects in power grid infrastructure engineering scenes based on AI technology, comprising the following steps:

[0025] S100. Acquire remote sensing images of power grid infrastructure project scenes.

[0026] The acquisition and analysis of remote sensing images play a key role in power grid infrastructure projects. Remote sensing images provide detailed geographic information through different acquisition methods, such as satellite remote sensing, aerial remote sensing, and ground measurement. These images provide digitized multi-band data, including visible light, infrared, and radar bands, as well as corresponding geospatial coordinate information. Remote sensing images should have appropriate resolution and sensor information, which provides a rich source of information for the convolutional neural network (CNN) model, enabling the model to accurately identify and classify different types of objects, thereby helping decision makers understand and manage the status and environmental impact of power grid infrastructure.

[0027] S200. Using convolutional neural network model to identify the type of land objects in remote sensing images.

[0028] Land object features refer to visual features extracted from remote sensing images, including edge features, texture features, shape features, spectral features, spatial distribution features, and contextual features. Texture features include the wall texture of buildings and the road surface texture, shape features include the outline of buildings and the shape of roads, spectral features include the characteristics of different land object types in visible light, infrared and other bands, spatial distribution features include the dense distribution of buildings and the distribution pattern of water bodies, and contextual features include the relative positions and interactions between different land object types.

[0029] Construct a convolutional neural network model. The input data is high-resolution, multi-band remote sensing image data. The label data is the ground object type label of each image. Use the convolution kernel to perform convolution operation on the input data and calculate the output feature map of the convolution layer. For the kth convolution kernel, the convolution calculation is: In the formula, F represents the size of the convolution kernel, S represents the stride, and b k Represents bias; the output of the convolutional layer is transformed nonlinearly, and the ReLU activation function ReLU is applied. The activation function is calculated as: A i,j,k =max(0,Z i,j,k ); Perform pooling operation on the feature map to reduce the size of the feature map and reduce the computational complexity. For maximum pooling, the pooling calculation is: In the formula, F represents the size of the pooling window; the output of the pooling layer is flattened into a one-dimensional vector, the flattened feature map is matrix multiplied with the weight matrix of the fully connected layer, a bias is added, and an activation function is applied. The full connection is calculated as: O = ReLU (W · X + b), where W represents the weight matrix and b represents the bias; the output of the fully connected layer is converted into a probability distribution, representing the predicted probability of each category, and the Softmax function is calculated as: In the formula, O k Represents the output value of the kth category, K represents the total number of categories; Calculate the loss: the difference between the prediction of the evaluation model and the actual label, the cross entropy loss function is calculated as: Where Y i,k represents the true label of the i-th sample, P i,k Represents the probability predicted by the model; calculates the gradient of the loss function with respect to each parameter (weight, bias), and updates the parameters using an optimization algorithm (such as gradient descent); selects an optimization algorithm: such as stochastic gradient descent (SGD), Adam, etc., and adjusts the learning rate and parameters according to the gradient update rule.

[0030] S210. The convolutional neural network model uses stacked convolutional layers and pooling layers. In the initial stage, the convolutional neural network model activates a preset number of convolutional layers and pooling layers. After extracting the ground object features of the remote sensing image, the extracted ground object features are passed to the connection layer to obtain the ground object type probability value.

[0031] The convolution layer is the core component for extracting features from remote sensing images. Each convolution layer consists of multiple convolution kernels. Each convolution kernel slides on the image and calculates the dot product with its overlapping area to produce a feature map. The output of each convolution layer is a set of feature maps. Each feature map contains multiple features detected in the input image. As the depth of the network increases, the complexity and abstractness of the convolution kernel also increase, and more advanced and abstract features can be extracted.

[0032] In the initial stage, a preset number of convolutional layers and pooling layers are activated to extract features of land objects. These convolutional layers and pooling layers extract features in the image through filters, such as edge features, texture features, shape features, spectral features, spatial distribution features, and context features, and downsample through the pooling layer to retain the main features. The features of land objects extracted by the convolutional layers and pooling layers are flattened and passed to the connection layer, which is responsible for mapping the extracted features to the probability values ​​of land object types. In the connection layer, the activation function will perform nonlinear transformation on the features to increase the network's expression ability and learn nonlinear features. These nonlinearly transformed features will be used for probability prediction of land object types. The features output by the connection layer are processed by the softmax function to obtain the probability values ​​of each land object type.

[0033] By stacking multiple layers of convolution and pooling layers, the model is able to extract features of objects in remote sensing images at different levels of abstraction. This hierarchical structure allows the model to gradually capture features from simple to complex, such as edges, textures to more advanced structures and shapes of objects. The open preset number of layers can be adjusted according to the complexity of specific tasks and data. In the initial stage, the model can use fewer convolution and pooling layers to quickly extract basic features for preliminary prediction of object types.

[0034] S220. Determine whether the probability value of the feature type is greater than the preset judgment probability, and identify the feature types with probability values ​​greater than and less than the judgment probability as known feature types and unknown feature types respectively. If an unknown feature type exists, further activate the number and depth of the convolutional layer and the pooling layer to extract more feature features, and identify the feature type probability value of the unknown feature type based on the feature features, until the feature maps of all known feature types are obtained.

[0035] For each feature type probability value, determine whether it is greater than the preset judgment probability. If the probability value is greater than the judgment probability, the feature type is identified as a known feature type. If the probability value is less than the judgment probability, the feature type is identified as an unknown feature type. If an unknown feature type exists, the number and depth of the convolutional layer and the pooling layer need to be further activated according to the existing unknown feature types to extract more feature features. This can be achieved by activating more convolutional layers and pooling layers. The adjusted convolutional neural network extracts more feature features and then passes them to the connection layer again. This iterative judgment is used to obtain the feature type probability value of the unknown feature type until the feature map of all known feature types is obtained.

[0036] The types of features with probability values ​​greater than and less than the judgment probability are respectively identified as known feature types and unknown feature types, so that some features that are difficult to accurately identify can be identified. For the processing of unknown feature types, further activating the number and depth of convolutional layers and pooling layers can help the system to extract more feature features in a targeted manner, which helps to make up for the deficiencies that may exist in the initial feature extraction stage. In addition, the system can also continuously optimize and improve the recognition results of unknown feature types during the recognition process, thereby improving the overall recognition accuracy.

[0037] S300. Output results, including remote sensing images marked with known ground object types.

[0038] By using an appropriate amount of convolutional layers and pooling layers to identify the type of land objects in the initial stage, the model is allowed to dynamically increase the network depth when encountering unknown land object types to extract richer features while maintaining rapid recognition of known land object types, thereby improving overall adaptability and recognition accuracy, and being able to dynamically adjust the depth and number of extracted features according to the land object conditions in the image, thereby better adapting to different types of remote sensing images. Reducing resource consumption when processing common land objects not only saves storage and processing resources, but also maintains efficiency and accuracy in complex image processing.

[0039] The output remote sensing image also has the coordinates of known land feature types.

[0040] The coordinate information of known land feature types combined with data such as terrain and land use can help accurately assess terrain and resource distribution, optimize line layout, reduce environmental impact, ensure safe and stable operation of the power grid, and reduce construction costs.

[0041] S230. The convolutional neural network model also includes an additional fully connected layer. The feature map of the known land object type is converted into a one-dimensional vector through a shape transformation matrix, and the one-dimensional vector is input into the fully connected layer. A linear transformation is performed through the weight matrix to obtain an intermediate output. An activation function is applied to the intermediate output, and the activated output is used as the input of the output layer to obtain the coordinates of the known land object type.

[0042] When training this additional fully connected layer, the weight matrix W and bias vector b are randomly initialized. During the forward propagation, the feature maps of all known object types are converted into one-dimensional vectors: flatten(X k )=X k ⊙reshape(I), where X k represents the feature map of all known feature types, reshape(I) represents a shape transformation matrix, which is used to convert the feature map of all known feature types into a one-dimensional vector and calculate the intermediate output: intermediate_output = W·flatten(X k )+b, apply an activation function such as ReLU to the intermediate output, and use the mean square error (MSE) as the loss function during back propagation. Calculate the gradient of the loss function and use gradient descent or other optimization algorithms to update weights and biases until the maximum number of iterations is reached or the loss function converges.

[0043] After obtaining the feature maps of all known types of objects, they are converted into one-dimensional vectors through a shape transformation matrix reshape(I), and the one-dimensional vectors are mapped to the predicted values ​​of the object coordinates. The coordinates are predicted through two neurons to obtain the output: output = [output x output y ], where output x and output y The coordinates obtained through the activation function realize the precise positioning of the known type of land object.

[0044] In conventional CNNs, identifying object types and coordinates is usually handled as a joint task. This means that the CNN architecture includes convolution and pooling layers to extract rich features from remote sensing images, and then performs classification (identifying object types) and regression (obtaining object coordinates) through fully connected layers. Therefore, in this case, multiple convolution and pooling operations are indeed required throughout the network in order to fully extract and utilize the information in the image to complete these two tasks at the same time.

[0045] In S210, the convolutional neural network mainly extracts the ground object features of the remote sensing image through stacked convolutional layers and pooling layers, and S220 classifies the extracted features, determines the ground object type, and determines the known ground object type and the unknown ground object type. These two steps are mainly concentrated on the feature extraction and ground object type judgment process of the convolutional neural network model, focusing on the extraction and classification of ground object features in the remote sensing image, but steps S210 and S220 provide the ground object features and classification information for S230, and S230 converts the feature map of the known ground object type into a one-dimensional vector, and inputs it into the fully connected layer for processing, and finally obtains the specific position of the known ground object type in the image. This embodiment scheme adopts a simplified processing flow by separating the tasks of ground object type identification and coordinate positioning. First, multiple convolution and pooling operations are used to extract key features for ground object type identification. This step effectively reduces the computational complexity. When performing coordinate positioning, S230 avoids repeated convolution and pooling, and directly processes on the extracted features, thereby further improving the efficiency and speed of the model.

[0046] Get the number of activation layers of the convolution layer and the pooling layer when identifying all specified objects, and use the number of activation layers as the number of stacked layers of the convolution layer and the pooling layer.

[0047] Regarding the number of stacked convolutional and pooling layers, the types of objects that need to be identified can be selected as specified objects during the model training stage. The number of activation layers of the convolutional and pooling layers when identifying all specified objects can be used as the number of stacked convolutional and pooling layers. If there are difficult-to-identify types of objects among the types of objects that need to be identified, the difficult-to-identify types of objects can be removed from the specified known object types to ensure that the model can identify all specified known object types as much as possible.

[0048] In S220, further activating the number and depth of convolutional layers and pooling layers specifically includes: obtaining the maximum value of the probability value of the ground object type of all unknown ground object types, dividing the maximum value of the ground object type probability value of all unknown ground object types into a specified number of clusters according to the numerical value; and determining the number of times to activate the convolutional layer and the pooling layer based on the number of clusters.

[0049] The probability value reflects the model's confidence in different types of objects, and the number of clusters determines the number of layers to be activated in the convolutional neural network next. Each cluster can be associated with a set of convolutional and pooling layer operations in the network to more effectively extract and identify similar types of objects. The number and depth of convolutional and pooling layers are dynamically determined based on actual data, so that the model can be optimized based on the characteristics and complexity of each type of object. This dynamic adaptability helps improve the generalization and accuracy of the model, and can also reduce the computational load while maintaining recognition speed, thereby improving overall operating efficiency.

[0050] A clustering algorithm based on data density is used to determine the number of clusters, specifically including: taking the maximum value of the feature type probability value of the unknown feature type as the data point, defining parameters ε and MinPts, ε represents the distance threshold of the neighborhood, and MinPts represents the minimum number of data points in the ε neighborhood; visiting each data point, randomly selecting an unvisited data point each time, calculating the number of data points in the ε neighborhood of the point, and if it is greater than or equal to MinPts, marking the point as a core point; starting from the core point, adding its density-reachable points to the same cluster in a density-reachable manner, and adding points that do not belong to any cluster to the nearest cluster.

[0051] The selection of the two parameters ε and MinPts will affect the clustering results and can be adjusted according to specific problems and data characteristics. The core points represent areas with higher data density. Points densely connected to the core points are classified into the same category to form clusters. The DBSCAN algorithm does not require the pre-specification of the number of clusters, but determines the number of clusters by the density of the data. Therefore, it has good applicability for determining the number of clusters of unknown landform types.

[0052] The number of activation layers n corresponding to clustering i For: n i =a i ·(Ls); Where n i represents the number of activation layers corresponding to the i-th cluster, a i represents the length ratio of the i-th cluster, L represents the number of stacked convolutional layers and pooling layers, s represents the number of convolutional layers and pooling layers activated in the initial stage, and A represents the number of clusters.

[0053] Since the length ratios of different types of objects are different, the number of layers activated each time can be dynamically determined and adjusted according to the complexity of the category, which can make the network more targeted and adaptable when dealing with different types of objects, thereby improving the accuracy of classification. For categories with longer lengths, more layers are allocated for feature extraction and recognition; for categories with shorter lengths, the number of activated layers can be reduced, and the resources of the network can be allocated more reasonably, avoiding the use of the same number of layers for all categories, thereby reducing the risk of overfitting. For categories with shorter lengths, reducing the number of activated layers can save computing resources and improve the operating efficiency of the network.

[0054] Embodiment 2:

[0055] This embodiment provides an AI-based intelligent ground object recognition system for power grid infrastructure project scenes, which adopts the AI-based intelligent ground object recognition method for power grid infrastructure project scenes as in Embodiment 1, and includes: an image acquisition module and a ground object recognition module.

[0056] The image acquisition module is used to acquire remote sensing images of power grid infrastructure engineering scenes; the land object identification module is used to use a convolutional neural network model to identify the land object type of the remote sensing image, including: the convolutional neural network model uses stacked convolutional layers and pooling layers. In the initial stage, the convolutional neural network model activates a preset number of convolutional layers and pooling layers. After extracting the land object features of the remote sensing image, the extracted land object features are transferred to the connection layer to obtain the land object type probability value; it is determined whether the land object type probability value is greater than the preset judgment probability, and the land object types with probability values ​​greater than and less than the judgment probability are respectively identified as known land object types and unknown land object types. If an unknown land object type exists, the number and depth of the convolutional layers and pooling layers are further activated to extract more land object features, and the land object type probability value of the unknown land object type is identified based on the land object features until a feature map of all known land object types is obtained; the output result includes a remote sensing image marked with a known land object type.

[0057] This embodiment has all the advantages of the method for intelligently identifying ground objects in power grid infrastructure project scenes based on AI technology in Embodiment 1, and can automatically implement all steps of the method for intelligently identifying ground objects in power grid infrastructure project scenes based on AI technology.

[0058] Embodiment three:

[0059] The present embodiment provides an AI-based device for intelligently identifying objects in power grid infrastructure project scenes, including a memory and a processor, wherein the memory is used to store computer programs; and the processor is used to implement the steps of the AI-based method for intelligently identifying objects in power grid infrastructure project scenes as described in Embodiment 1 when executing the computer program.

[0060] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0061] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1The steps of the functions specified in a box or multiple boxes. Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they know the basic creative concept. Therefore, the attached claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

Claims

1. An intelligent method for identifying objects in power grid infrastructure projects based on AI technology, characterized by: Obtain remote sensing images of power grid infrastructure project scenes; The convolutional neural network model is used to identify the types of objects in remote sensing images, including: The convolutional neural network model uses stacked convolutional layers and pooling layers. In the initial stage, the convolutional neural network model activates a preset number of convolutional layers and pooling layers. After extracting the ground feature of the remote sensing image, the extracted ground feature is transferred to the connection layer to obtain the ground feature type probability value; Determine whether the feature type probability value is greater than a preset determination probability, and identify the feature types with probability values ​​greater than and less than the determination probability as known feature types and unknown feature types, respectively. If an unknown feature type exists, further activate the number and depth of the convolution layer and the pooling layer to extract more feature features, and identify the feature type probability value of the unknown feature type based on the feature features, until feature maps of all known feature types are obtained; Output results, including remote sensing images with known object types marked.

2. The method for intelligently identifying ground objects in power grid infrastructure projects based on AI technology according to claim 1, characterized in that: The output remote sensing image also has the coordinates of known land feature types.

3. The method for intelligently identifying ground objects in power grid infrastructure projects based on AI technology as claimed in claim 2, characterized in that: The convolutional neural network model also includes an additional fully connected layer. The feature map of the known land feature type is converted into a one-dimensional vector through a shape transformation matrix, the one-dimensional vector is input into the fully connected layer, a linear transformation is performed through a weight matrix to obtain an intermediate output, an activation function is applied to the intermediate output, and the activated output is used as the input of the output layer to obtain the coordinates of the known land feature type.

4. The method for intelligently identifying ground objects in power grid infrastructure projects based on AI technology according to claim 1, characterized in that: Get the number of activation layers of the convolution layer and the pooling layer when identifying all specified objects, and use the number of activation layers as the number of stacked layers of the convolution layer and the pooling layer.

5. The method for intelligently identifying ground objects in power grid infrastructure project scenes based on AI technology as claimed in claim 4, characterized in that: The number and depth of further activation convolutional and pooling layers include: Obtain the maximum value of the feature type probability value of all unknown feature types, and divide the maximum value of the feature type probability value of all unknown feature types into a specified number of clusters according to the value size; The number of times to activate the convolutional layers and the pooling layers is determined based on the number of clusters.

6. The method for intelligently identifying ground objects in power grid infrastructure project scenes based on AI technology as claimed in claim 5, characterized in that: The number of clusters is determined based on the data density clustering algorithm.

7. The method for intelligently identifying ground objects in power grid infrastructure projects based on AI technology according to claim 6, characterized in that: Determining the number of clusters specifically includes: The maximum value of the feature type probability value of the unknown feature type is taken as the data point, and the parameters ε and MinPts are defined. ε represents the distance threshold of the neighborhood, and MinPts represents the minimum number of data points in the ε neighborhood. Visit each data point, randomly select an unvisited data point each time, calculate the number of data points in the ε neighborhood of the point, and if it is greater than or equal to MinPts, mark the point as a core point; starting from the core point, add its density-reachable points to the same cluster in a density-reachable manner, and add points that do not belong to any cluster to the nearest cluster.

8. The method for intelligently identifying ground objects in power grid infrastructure project scenes based on AI technology as claimed in claim 5, characterized in that: Each cluster corresponds to the number of activation layers n i , n i for: n i =a i ·(L-s) Where n i represents the number of activation layers corresponding to the i-th cluster, a i represents the length ratio of the i-th cluster, L represents the number of stacked convolutional layers and pooling layers, s represents the number of convolutional layers and pooling layers activated in the initial stage, and A represents the number of clusters.

9. The method for intelligently identifying ground objects in power grid infrastructure project scenes based on AI technology according to claim 1, characterized in that: The ground feature features include edge features, texture features, shape features, spectral features, spatial distribution features and context features.

10. The intelligent ground object recognition system for power grid infrastructure projects based on AI technology is characterized by: include: An image acquisition module, which is used to acquire remote sensing images of power grid infrastructure project scenes; The ground object recognition module is used to identify the ground object type in remote sensing images using a convolutional neural network model, including: The convolutional neural network model uses stacked convolutional layers and pooling layers. In the initial stage, the convolutional neural network model activates a preset number of convolutional layers and pooling layers. After extracting the ground feature of the remote sensing image, the extracted ground feature is transferred to the connection layer to obtain the ground feature type probability value; Determine whether the feature type probability value is greater than a preset determination probability, and identify the feature types with probability values ​​greater than and less than the determination probability as known feature types and unknown feature types, respectively. If an unknown feature type exists, further activate the number and depth of the convolution layer and the pooling layer to extract more feature features, and identify the feature type probability value of the unknown feature type based on the feature features, until feature maps of all known feature types are obtained; Output results, including remote sensing images with known object types marked.

Images