Synthetic aperture radar image power tower detection method and system based on deep learning

Abstract

The application provides a deep learning synthetic aperture radar image power tower detection method and system, comprising: acquiring a to-be-recognized synthetic aperture radar image of a power tower; inputting the to-be-recognized synthetic aperture radar image into a trained power tower detection model for inference to obtain a power tower detection result in the to-be-recognized synthetic aperture radar image; wherein the power tower detection model first trains a detection network based on generalized convolution attention using historical synthetic aperture radar images to obtain an initial detection model, and then trains the initial detection model again using knowledge distillation. The power tower detection model can realize the process of quickly extracting the synthetic aperture radar image of the power tower in the candidate region, so that the time consumption of the synthetic aperture radar image of the power tower in the candidate region extraction process is shortened.

Description

Technical Field

[0001] This invention belongs to the field of power facility technology, specifically relating to a method and system for detecting power poles using synthetic aperture radar images based on deep learning. Background Technology

[0002] With the continuous growth of electricity demand and increasing requirements for power system reliability, the effective monitoring and maintenance of power infrastructure has become increasingly important. Among power infrastructure, the inspection of high-voltage transmission towers is particularly critical. These towers are typically 30 to 60 meters high, constructed of steel, and possess significant metallic characteristics. However, power poles are widely distributed in complex environments such as suburbs and rural areas.

[0003] Existing methods for detecting power poles based on high-resolution synthetic aperture radar (SAR) images overcome the limitations of traditional optical images by leveraging the advantages of SAR images for all-weather, all-time ground observation, thus enabling routine monitoring of power poles.

[0004] However, in high-resolution synthetic aperture radar (SAR) images, the identification of power poles is easily affected by clutter interference and local differences in incident angle, posing significant challenges to the tower detection task. This results in a prolonged time consumption during candidate region extraction in power pole detection images. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, in a first aspect, this invention proposes a deep learning-based method for detecting power poles in synthetic aperture radar images, comprising:

[0006] Acquire synthetic aperture radar images of power poles to be identified;

[0007] The synthetic aperture radar image to be identified is input into the trained power pole detection model for inference to obtain the power pole detection result in the synthetic aperture radar image to be identified.

[0008] The power pole detection model first uses historical synthetic aperture radar images to train a detection network based on generalized convolutional attention to obtain an initial detection model, and then uses knowledge distillation to train the initial detection model again.

[0009] Preferably, the training process of the power pole detection model includes:

[0010] Historical synthetic aperture radar (SAR) images of power poles are acquired and used as input for training data.

[0011] The historical power pole labels in the historical synthetic aperture radar images are used as the output of the training data.

[0012] Using the input and output of the training data, a detection network based on generalized convolutional attention is trained to obtain an initial detection model;

[0013] The initial detection model is used as the teacher network; the number of feature channels in the initial detection model is reduced, and the reduced initial detection model is used as the student network.

[0014] Using knowledge distillation, the teacher network guides the student network in training, and the trained student network is then identified as the power pole detection model.

[0015] Preferably, after acquiring the synthetic aperture radar image of the power pole to be identified, the process includes:

[0016] The synthetic aperture radar image to be identified is preprocessed;

[0017] The preprocessing includes: metadata tagging, radiometric calibration, multi-view processing, and geocoding.

[0018] Preferably, the step of training a detection network based on generalized convolutional attention using the input and output of the training data to obtain an initial detection model includes:

[0019] The input of the training data is predicted using a detection network based on generalized convolutional attention to obtain a predicted output;

[0020] The loss function value is calculated based on the output of the training data and the predicted output;

[0021] If the iteration termination condition is reached, then the detection network based on generalized convolutional attention corresponding to the model parameters at the time the iteration termination condition is reached is determined as the initial detection model; the iteration termination condition includes the loss function value reaching a preset loss threshold and / or the current iteration number reaching a preset iteration number;

[0022] If the iteration termination condition is not met, the model parameters of the detection network based on generalized convolutional attention are adjusted according to the loss function value, and the detection network based on generalized convolutional attention with adjusted model parameters is used to re-predict the input of the training data until the iteration termination condition is met, thus obtaining the initial detection model.

[0023] Preferably, the detection network based on generalized convolutional attention has multiple levels, and each level has a generalized convolutional attention module;

[0024] The initial detection model is obtained by training a detection network based on generalized convolutional attention using the input and output of the training data, including:

[0025] The training data is used as the input to the detection network based on generalized convolutional attention; the first level among the multiple levels is taken as the current level, the output feature tensor of the generalized convolutional attention module in the current level is taken as the initial feature tensor, the initial feature tensor is passed to the generalized convolutional attention module of the next level, and then the size of the feature tensor output by the generalized convolutional attention module of the next level is enlarged to form an enlarged feature tensor;

[0026] The next level is taken as the new current level, and the expanded feature tensor is taken as the feature tensor of the generalized convolutional attention module of the new current level. It is then passed to the generalized convolutional attention module of the next level of the new current level until the feature tensor output by the generalized convolutional attention module of the last level among the multiple levels is obtained, forming a feature tensor that is gradually expanded according to the multiple levels.

[0027] The gradually expanding feature tensor is used as the prediction output. Based on the output of the training data and the prediction output, a detection network based on generalized convolutional attention is trained to obtain an initial detection model.

[0028] Preferably, the step of taking the next level as the new current level, and taking the expanded feature tensor as the feature tensor of the generalized convolutional attention module of the new current level, and passing it to the generalized convolutional attention module of the next level of the new current level, until the feature tensor output by the generalized convolutional attention module of the last level among the multiple levels is obtained, forming a feature tensor that gradually expands according to the multiple levels, includes:

[0029] The initial feature tensor is passed to the next level of generalized convolutional attention module, and the initial feature tensor is processed by the next level of generalized convolutional attention module to obtain the processed initial feature tensor.

[0030] The initial feature tensor and the processed initial feature tensor are divided into multiple partial tensors according to the same segmentation rules;

[0031] Each part of the initial feature tensor is concatenated with the corresponding part of the processed initial feature tensor to form multiple sub-feature tensors;

[0032] Multiple sub-feature tensors are concatenated; the concatenation result is used as the enlarged feature tensor output by the next level of generalized convolutional attention module.

[0033] Preferably, the enlarged feature tensor satisfies the following formula:

[0034] Out=X·MLP(flatten(X))·σ·(ConvB(X))

[0035] In the above formula, MLP represents the next level of generalized convolutional attention module, ConvB() represents the single-channel weight matrix generated by performing convolution operation on the convolutional blocks in the generalized convolutional attention module; σ represents the nonlinear function used for activation; flatten() represents the concatenation of multiple sub-feature tensors; X represents multiple sub-feature tensors; and Out represents the expanded feature tensor.

[0036] Preferably, the step of employing knowledge distillation, using the teacher network to guide the student network in training, and determining the trained student network as the power pole detection model, includes:

[0037] The training data is input into the teacher network to obtain the intermediate features output by the teacher network; the training data is input into the student network to obtain the intermediate features output by the student network.

[0038] Calculate the distance metric between the intermediate features of the student network and the intermediate features of the teacher network;

[0039] The distance between the intermediate features of the student network and the intermediate features of the teacher network is determined as the distillation loss function value;

[0040] If the distillation loss function value reaches the preset distillation loss threshold, then the student network corresponding to the model parameters when the preset distillation loss threshold is determined as the trained student network.

[0041] If the distillation loss function value does not reach the preset distillation loss threshold, then the model parameters of the student network are adjusted according to the distillation loss function value, and the intermediate features re-output by the student network after adjusting the model parameters are used to iterate repeatedly until the distillation loss function value reaches the preset distillation loss threshold, thus obtaining the trained student network.

[0042] Extract the feature dimensions from the fully convolutional adaptive layers of the teacher network and the trained student network, respectively;

[0043] Based on the feature dimensions in the fully convolutional adaptive layers of the teacher network and the trained student network, the feature dimensions of the trained student network and the teacher network are aligned; the student network after feature dimension alignment is used as the power pole detection model.

[0044] Preferably, the expression for the distillation loss function is as follows:

[0045]

[0046] In the above formula, loss dis N represents the distillation loss value of the distillation loss function. p f is the size of the feature. cov W represents the total number of feature dimensions used to calculate the distillation loss; H represents the width of the feature map output in the fully convolutional adaptive layer; c represents the height of the feature map output in the fully convolutional adaptive layer; c represents the number of the c-th feature channel; C represents the total number of feature channels; S represents the intermediate features of the student network; t represents the intermediate features of the teacher network; and i and j represent the two-dimensional spatial coordinates of the feature map.

[0047] Secondly, this application also proposes a deep learning-based synthetic aperture radar image power pole tower detection system, comprising:

[0048] The acquisition module is used to acquire synthetic aperture radar images of power poles to be identified;

[0049] The power pole detection result acquisition module is used to input the synthetic aperture radar image to be identified into the trained power pole detection model for inference, and obtain the power pole detection result in the synthetic aperture radar image to be identified.

[0050] The power pole detection model first uses historical synthetic aperture radar images to train a detection network based on generalized convolutional attention to obtain an initial detection model, and then uses knowledge distillation to train the initial detection model again.

[0051] Thirdly, this application also proposes an electronic device, comprising: at least one processor and a memory; wherein the memory and the processor are connected via a bus;

[0052] The memory is used to store one or more programs;

[0053] When the one or more programs are executed by the at least one processor, the deep learning-based synthetic aperture radar image power pole detection method is implemented.

[0054] Fourthly, this application also proposes a readable storage medium having an executable program stored thereon, which, when executed, implements the deep learning-based synthetic aperture radar image power pole tower detection method.

[0055] Compared with the closest prior art, the present invention application has the following beneficial effects:

[0056] This invention discloses a method and system for detecting power poles in synthetic aperture radar (SAR) images using deep learning. The method includes: acquiring a SAR image of a power pole to be identified; inputting the SAR image into a trained power pole detection model for inference to obtain a power pole detection result in the SAR image; wherein the power pole detection model first trains an initial detection model using historical SAR images on a detection network based on generalized convolutional attention, and then retrains the initial detection model using knowledge distillation. Because the initial detection model is trained using a detection network based on generalized convolutional attention and knowledge distillation, a power pole detection model is obtained. This model enables rapid extraction of candidate regions from SAR images of power poles, thus reducing the time required for candidate region extraction. Attached Figure Description

[0057] Figure 1 A flowchart of a deep learning-based synthetic aperture radar image power pole tower detection method provided in this invention application;

[0058] Figure 2 The structure of the detection network based on generalized convolutional attention provided in this invention application;

[0059] Figure 3 The structure of the generalized convolutional attention module provided in this invention application;

[0060] Figure 4 This is a training diagram of the teacher network on the student network provided in this invention application;

[0061] Figure 5 A flowchart for verifying the accuracy of the power pole detection model provided for the invention application;

[0062] Figure 6 This is a schematic diagram of the SAR image edge overlap strategy provided in this invention application;

[0063] Figure 7 A graph showing the relationship between training loss and number of iterations for a detection network based on generalized convolutional attention, as provided in this invention application;

[0064] Figure 8 The results of the power pole detection model provided in this invention application identifying power poles in different scenarios;

[0065] Figure 9 A flowchart of a deep learning-based synthetic aperture radar image power pole detection system provided in this invention application;

[0066] Figure 10This is a schematic diagram of the operation of an electronic device provided in this invention application. Detailed Implementation

[0067] The specific embodiments of this invention will be further described in detail below with reference to the accompanying drawings.

[0068] Example 1:

[0069] like Figure 1 As shown, this invention application proposes a deep learning-based method for detecting power poles in synthetic aperture radar images, comprising the following steps:

[0070] Step 1: Acquire synthetic aperture radar images of the power poles to be identified;

[0071] Step 2: Input the synthetic aperture radar image to be identified into the trained power pole detection model for inference to obtain the power pole detection result in the synthetic aperture radar image to be identified; wherein, the power pole detection model first uses historical synthetic aperture radar images to train a detection network based on generalized convolutional attention to obtain an initial detection model, and then uses knowledge distillation to train the initial detection model again.

[0072] Step 1 above, after acquiring the synthetic aperture radar image of the power pole to be identified, may include:

[0073] The synthetic aperture radar image to be identified is preprocessed;

[0074] The preprocessing includes: metadata tagging, radiometric calibration, multi-view processing, and geocoding.

[0075] The metadata tagging process is as follows: obtain the parameter information of the synthetic aperture radar image to be identified from the header file of the image to be identified. The parameter information includes the geographic range of the image, azimuth resolution, and range resolution. Based on the parameter information, the synthetic aperture radar image to be identified is tagged with metadata according to the parameter information.

[0076] The radiometric calibration process described above is as follows: The backscattering intensity information of the synthetic aperture radar (SAR) image to be identified is calibrated using parameters such as scattering area and antenna gain to obtain scattering coefficients; then, the scattering coefficients are normalized to a unified standard to eliminate the influence of the observation geometry of the SAR image data (such as different sensors or receiving modes) on the results, thereby facilitating comparison and analysis; based on the normalization results, the SAR image to be identified is radiometrically calibrated.

[0077] The multi-view processing procedure is as follows: acquire the range resolution, range view count, azimuth resolution, and azimuth view count of the synthetic aperture radar image to be identified; customize the range view count and azimuth view count so that the product of the azimuth resolution and the azimuth view count, and the product of the range resolution and the range view count, are close to preset values ​​(for example, the preset value can be 1m); mark the azimuth view counts that are close to the preset values ​​in the synthetic aperture radar image to be identified;

[0078] The geocoding process is as follows: obtain the slant range coordinate system in the synthetic aperture radar image to be identified, convert the slant range coordinate system in the synthetic aperture radar image data into a geographic coordinate system using a digital elevation model (DEM), and geocode the synthetic aperture radar image to be identified based on the geographic coordinate system.

[0079] The preprocessing of the synthetic aperture radar image to be identified mainly involves obtaining a backscattering coefficient map with geocoded coordinates.

[0080] In step 2 above, the training process of the power pole detection model may include:

[0081] Step 2.1: Obtain historical synthetic aperture radar images of power poles and use these historical synthetic aperture radar images as input for training data;

[0082] Step 2.2: Use the historical power pole labels in the historical synthetic aperture radar images as the output of training data;

[0083] Step 2.3: Using the input and output of the training data, train the detection network based on generalized convolutional attention to obtain the initial detection model;

[0084] Step 2.4: Use the initial detection model as the teacher network; reduce the number of feature channels in the initial detection model, and use the reduced initial detection model as the student network;

[0085] Step 2.5: Using knowledge distillation, the teacher network guides the student network in training, and the trained student network is determined as the power pole detection model.

[0086] In step 2.2 above, the historical power pole labels in the historical synthetic aperture radar images all indicate the actual location information of the power poles.

[0087] In step 2.3 above: first, the parameters of the detection network based on generalized convolutional attention are randomly initialized; then, the training data is input into the initialized detection network based on generalized convolutional attention in batches.

[0088] In step 2.3, after randomly initializing the parameters of the detection network based on generalized convolutional attention, training the detection network based on generalized convolutional attention using the input and output of the training data to obtain an initial detection model may include:

[0089] Step 2.3.1: Use a detection network based on generalized convolutional attention to predict the input of the training data to obtain the predicted output;

[0090] Step 2.3.2: Calculate the loss function value based on the output of the training data and the predicted output;

[0091] Step 2.3.3: If the iteration termination condition is met, then the detection network based on generalized convolutional attention corresponding to the model parameters at the time the iteration termination condition is met is determined as the initial detection model; the iteration termination condition includes the loss function value reaching a preset loss threshold and / or the current iteration number reaching a preset iteration number.

[0092] Step 2.3.4: If the iteration termination condition has not been met, then the model parameters of the detection network based on generalized convolutional attention are adjusted according to the loss function value, and the detection network based on generalized convolutional attention with the adjusted model parameters is used to re-predict the input of the training data until the iteration termination condition is met, and the initial detection model is obtained.

[0093] In step 2.3.2 above, the output of the training data is the actual location information of the power poles. This mainly involves comparing the predicted output with the actual location information of the power poles; and calculating the loss function value based on the comparison result.

[0094] When calculating the loss function value, this loss function enhances the sensitivity to geometric differences between the predicted bounding box and the true bounding box of the power pole's actual location information by introducing a penalty term for bounding box center offset and aspect ratio difference. Building upon this, a focusing concept is further introduced, assigning different weights to different training data, focusing on training data that is difficult to predict, in order to alleviate the problem of imbalanced training data. This loss function can provide larger gradients in regions with low IOU (Intersection over Union) values, representing an improvement over traditional loss functions, thereby improving the gradient vanishing problem caused by unpredictable imbalanced training data during training. See the following formula for details:

[0095]

[0096]

[0097]

[0098] loss=λ1×cls loss (p t )+λ2×box loss +λ3×dfl loss (S i ,S i+1 )

[0099] In the above formula, cls loss For classification loss, cls loss (p t p is the cross-entropy loss used to measure the classification error of the training data. t For the prediction tensor of the training data, log(p) t The box represents the logarithmic probability that the training data is correctly classified. loss The bounding box loss is D2, where D is the distance between the center points of the predicted box and the ground truth box. C α2 is the diagonal distance of the minimum bounding rectangle C; α2 is the weight coefficient for adjusting the geometric penalty term; v is the penalty term caused by the difference in height ratio. Where π is the mathematical constant pi, and w gt h is the width of the target bounding box. gt w is the height of the target bounding box, h is the width of the predicted bounding box, and dfl is the height of the predicted bounding box. loss For distributed loss, dfl loss (S i ,S i+1 ) is the loss function that measures the difference between the predicted bounding box distribution and the true location; y represents the true value; S i It is the probability of being closest to the left of the target position y in the predicted distribution; S i+1 α1, α3, α4, γ1, γ2, γ3, λ1, λ2, and λ3 are all hyperparameters, and loss is the loss function value.

[0100] In step 2.3.4 above: if the iteration termination condition has not been met, the parameters of the detection network based on generalized convolutional attention can be updated using the backpropagation algorithm according to the loss function value;

[0101] In steps 2.1-2.3 above, during training, a validation set of the detection network based on generalized convolutional attention can be used to monitor the training process and prevent overfitting. If the loss function value continues to decrease during training, it means that the detection network based on generalized convolutional attention is gradually converging and its performance is improving. The test set is used to evaluate the final performance of the initial detection model. The test set is not used during training to prevent the initial detection model from generating false generalizations by fitting the test set data.

[0102] like Figure 2 As shown, the detection network based on generalized convolutional attention introduced in this invention is a deep convolutional neural network, including: (1) an input module responsible for receiving input data; (2) a backbone network based on deep convolutional layers for extracting preliminary features; (3) a neck network based on the GCAL (Global Convolution Attention Layer) module for extracting deep information through the generalized convolutional attention layer; (4) a multi-level detection head, including: detection head-1, detection head-2, and detection head-3, for detecting the extracted features; and (5) a loss function for calculating the loss function value of the output result. During the training phase, a randomly initialized detection network based on generalized convolutional attention is used for training. The training set is input into the detection network based on generalized convolutional attention in batches. Feature information is extracted through the backbone network and the GCAL module. The multi-level detection head outputs the results. The loss function is calculated based on the output results and manually labeled information. The network parameters are updated using the backpropagation algorithm. The model is iterated repeatedly until it converges. After reaching the predetermined stopping condition, the model is saved and the training ends. Where C1-C7 represent the convolutional block features generated sequentially by the network, the blue arrows represent feature pooling, and the red arrows represent feature upsampling. Because detection networks based on generalized convolutional attention output a large number of predicted boxes, but there are actually only a few targets, non-maximum suppression is needed to eliminate redundant candidate boxes and retain only the best candidate boxes, thereby improving the accuracy of detection networks based on generalized convolutional attention.

[0103] The detection network based on generalized convolutional attention (GCAL) performs deeper analysis on the initial features extracted by the backbone network through multi-path feature aggregation, making it suitable for subsequent detection tasks. Each GCAL module in the network employs an attention-based multi-path feature aggregation method, enabling the network to maintain relatively low computational complexity while achieving powerful performance and efficiency. Although each GCAL module includes feature segmentation, feature connection, convolution, normalization, and activation functions, the dimension and size of the processed feature maps vary at different levels. Therefore, it is necessary to consider how to improve the stability of the initial detection model after training and its ability to recognize synthetic aperture radar (SAR) images.

[0104] In step 2.3 above, the detection network based on generalized convolutional attention has multiple levels, and each level has a generalized convolutional attention module;

[0105] To improve the stability of the initial detection model after training and the ability to recognize synthetic aperture radar images, the initial detection model is obtained by training a detection network based on generalized convolutional attention using the input and output of the training data. This can include:

[0106] Step a: The input of the training data is used as the input of the detection network based on generalized convolutional attention; the first level among the multiple levels is used as the current level, the output feature tensor of the generalized convolutional attention module in the current level is used as the initial feature tensor, the initial feature tensor is passed to the generalized convolutional attention module of the next level, and then the size of the feature tensor output by the generalized convolutional attention module of the next level is enlarged to form an enlarged feature tensor;

[0107] Step b: Take the next level as the new current level, and take the expanded feature tensor as the feature tensor of the generalized convolutional attention module of the new current level, and pass it to the generalized convolutional attention module of the next level of the new current level, until the feature tensor output by the generalized convolutional attention module of the last level among the multiple levels is obtained, forming a feature tensor that gradually expands according to the multiple levels.

[0108] Step c: Using the gradually expanding feature tensor as the prediction output, and based on the output of the training data and the prediction output, train the detection network based on generalized convolutional attention to obtain the initial detection model.

[0109] Taking steps a, b, and c as examples, training data of size 512×512×1 pixels is input into a detection network based on generalized convolutional attention. The backbone of the detection network based on generalized convolutional attention outputs a three-dimensional feature tensor, with feature tensors of 16×16×1024, 32×32×1024, and 64×64×512, respectively. Each GCAL module processes a different dimension of feature tensor and passes it to the next level. Compared to the previous level, the size of the feature tensor output by the generalized convolutional attention module at each level is doubled.

[0110] In step b above, taking the next level as the new current level, and using the expanded feature tensor as the feature tensor of the generalized convolutional attention module of the new current level, and passing it to the generalized convolutional attention module of the next level of the new current level, until the feature tensor output by the generalized convolutional attention module of the last level among the multiple levels is obtained, forming a feature tensor that gradually expands according to the multiple levels, may include:

[0111] Step b1: Pass the initial feature tensor to the next level of generalized convolutional attention module, and use the next level of generalized convolutional attention module to process the initial feature tensor to obtain the processed initial feature tensor.

[0112] Step b2: According to the same segmentation rules, the initial feature tensor and the processed initial feature tensor are segmented into multiple partial tensors;

[0113] Step b3: Concatenate each part tensor of the initial feature tensor with the corresponding part tensor in the processed initial feature tensor to form multiple sub-feature tensors;

[0114] Step b4: Concatenate the multiple sub-feature tensors; use the concatenation result as the enlarged feature tensor output by the generalized convolutional attention module of the next level.

[0115] like Figure 3 The diagram illustrates the structure of a generalized convolutional attention module, consisting of multiple parallel convolutions and multiple levels, which represent multiple layers. The gradually expanding feature tensors create smooth non-linearity. Specifically, after the convolutional operation, normalization and activation functions are typically applied. This normalizes the mini-batch inputs of each layer and introduces a smooth non-linear transition, accelerating training and improving the stability of the initial detection model. The specific process is as follows:

[0116]

[0117] In the above formula, Feature i Represents the i-th feature tensor; CovBlock1 represents the first convolutional block in the generalized convolutional attention module; the feature tensor input to the generalized convolutional attention module is divided into two parts: Chunk 1 / 2 and Chunk 2 / 2 CovBlock i Feature represents the i-th convolutional block in the generalized convolutional attention module. i-1 Let i represent the (i-1)th feature tensor; i represents the number of feature tensors.

[0118] X = ΠFeature i

[0119] In the above formula, Π represents concatenating the i sub-feature tensors, and X represents the concatenation result;

[0120] In step b4 above, starting from the second convolutional block, the current convolutional block processes the feature tensor output by the previous convolutional block sequentially. After all convolutional blocks have finished processing, all sub-feature tensors are concatenated. Finally, the concatenated total tensor is input into the detection network based on generalized convolutional attention to form an enlarged feature tensor, which satisfies the following formula:

[0121] Out=X·MLP(flatten(X))·σ·(ConvB(X))

[0122] In the above formula, MLP represents the next level of generalized convolutional attention module, ConvB() represents the single-channel weight matrix generated by performing convolution operation on the convolutional blocks in the generalized convolutional attention module; σ represents the nonlinear function used for activation; flatten() represents the concatenation of multiple sub-feature tensors; X represents multiple sub-feature tensors; and Out represents the expanded feature tensor.

[0123] In step 2.4 above: when reducing the number of feature channels of the initial detection model, it can be reduced by a factor of several to form a network with multiple channel reductions (for example, it can be reduced by a factor of 2).

[0124] In step 2.5 above, to improve the versatility of the power pole detection model, the method of knowledge distillation is used to train the student network using the teacher network, and the trained student network is then identified as the power pole detection model. This may include:

[0125] Step 2.5.1: Input the training data into the teacher network and obtain the intermediate features output by the teacher network; input the training data into the student network and obtain the intermediate features output by the student network.

[0126] Step 2.5.2: Calculate the distance metric between the intermediate features of the student network and the intermediate features of the teacher network; determine the distance metric between the intermediate features of the student network and the intermediate features of the teacher network as the distillation loss function value;

[0127] Step 2.5.3: If the distillation loss function value reaches the preset distillation loss threshold, then the student network corresponding to the model parameters when the preset distillation loss threshold is determined as the trained student network.

[0128] Step 2.5.4: If the distillation loss function value does not reach the preset distillation loss threshold, then the model parameters of the student network are adjusted according to the distillation loss function value, and the intermediate features re-output by the student network after adjusting the model parameters are used to iterate repeatedly until the distillation loss function value reaches the preset distillation loss threshold, and the trained student network is obtained.

[0129] Step 2.5.5: Extract the feature dimensions from the fully convolutional adaptive layers of the teacher network and the trained student network, respectively;

[0130] Step 2.5.6: Based on the feature dimensions in the fully convolutional adaptive layers of the teacher network and the trained student network, align the feature dimensions of the trained student network and the teacher network; use the student network with aligned feature dimensions as the power pole detection model.

[0131] like Figure 4 As shown, in the knowledge distillation process, the features of the student network are processed through a fully convolutional adaptive layer to calculate the intermediate features of the student network. The functions of the fully convolutional adaptive layer include: (1) when the feature dimensions of the student network are different from those of the teacher network, the feature dimensions in the fully convolutional adaptive layers of the student network and the teacher network can be aligned to facilitate the calculation of distance metrics. (2) when the features of the student network and the teacher network are compatible, to prevent the features of the student network from overfitting the features of the teacher network, a distillation loss function is set during the training process. When adjusting the model parameters of the student network according to the value of the distillation loss function, the parameters in the student network can be adjusted by gradient descent. The main implementation is to use knowledge distillation to train the student network using the teacher network, thereby obtaining a high-precision and lightweight student network. The lightweight student network has versatility. The distillation loss function is used to measure the similarity between the intermediate features of the student network and the teacher network. It can also be combined with the distillation loss function of object detection (such as classification loss or bounding box loss to form the total loss function, which is used as the distillation loss function). In this way, the student network not only needs to pay attention to the error between its own prediction results and the true labels during the learning process, but also needs to consider learning the output features of the teacher network, thereby improving its own feature expression ability.

[0132] The expression for the distillation loss function mentioned above is as follows:

[0133]

[0134] In the above formula, loss dis N represents the distillation loss value of the distillation loss function. p f is the size of the feature. cov W represents the total number of feature dimensions used to calculate the distillation loss; H represents the width of the feature map output in the fully convolutional adaptive layer; c represents the height of the feature map output in the fully convolutional adaptive layer; c represents the number of the c-th feature channel; C represents the total number of feature channels; S represents the intermediate features of the student network; t represents the intermediate features of the teacher network; and i and j represent the two-dimensional spatial coordinates of the feature map.

[0135] In deep learning, while complex teacher networks typically possess stronger recognition capabilities and higher accuracy, they also come with higher computational and storage costs. Therefore, their application in resource-constrained scenarios (such as mobile devices and embedded systems) is limited. Knowledge distillation is a method that allows student networks (smaller models) to learn from the knowledge acquired by a teacher network (larger model). The student network can significantly reduce the complexity of the teacher network while maintaining its performance. The formula for the distillation process is as follows:

[0136] L=loss+λ×loss dis

[0137] In the above formula, L is the total loss, representing the final loss of the student network, used to simultaneously consider the detection training loss and distillation loss; loss is the detection training loss function, representing the loss of the student network on the object detection task, which usually includes classification loss and localization loss; λ is the distillation loss weight balancing factor, used to adjust the influence of distillation loss in the total loss, ensuring that the contributions of classification loss and localization loss are balanced during training.

[0138] After the power pole detection model is completed, such as Figure 5 As shown, in order to verify the accuracy of the power pole detection model of the present invention, the power pole detection model was verified as follows:

[0139] 1) First, acquire the synthetic aperture radar image of the power pole to be identified as follows:

[0140] Based on SAR images downloaded from the Resource Satellite Application Center (i.e., the original images), the SAR images are in spotlight mode with single polarization, an image resolution of approximately 1 meter, and a pixel size of approximately 20000×12000. Eight SAR images were manually labeled, and then 1005 synthetic aperture radar (SAR) images of 512×512 pixels each were extracted from these images to identify the power poles. The distribution of the sampled SAR images is shown in Table 1 below.

[0141] Table 1 Data sampling locations

[0142]

[0143] like Figure 6 As shown, an edge overlap strategy was adopted in the SAR image cropping process, that is, overlapping cropping of the SAR image was performed. The size of each cropping image box is 512 pixels, and the step size for each movement is 410 pixels, so the overlap ratio of adjacent cropping images is about 20%.

[0144] Preprocessed synthetic aperture radar (SAR) images to be identified were randomly selected as samples. Approximately 80% of these samples were used as training data, forming the training set; the remaining approximately 20% were divided into validation and test sets. The sample distribution is shown in Table 2 below.

[0145] Table 2 Pole Tower Dataset

[0146] training set Validation set test set Sample size 783 111 111

[0147] The samples are then input into the power pole detection model, as follows:

[0148] Initialize all weights, biases, and batch normalization scaling factors for the power pole detection model. Set the initial learning rate of the power pole detection model to 0.01 and the sample input batch size to 8. Perform the following calculations using the Stochastic Gradient Descent (SGD) method:

[0149] lr=lr base -k*epoch

[0150]

[0151] Where lr is the learning rate of the power pole detection model, lr base γ is the initial learning rate, k is the decay coefficient of the power pole detection model, epoch is the number of training rounds, representing the number of times the model has traversed the entire training set; γ is the momentum coefficient of the power pole detection model, and mo is the initial learning rate. t Let t be the momentum term of the current iteration step, and mo be the current iteration step number. t-1 Let θ be the momentum term from the previous iteration, and θ be the parameter of the power pole detection model. (t) For the parameters before the update, θ (t+1) Here are the updated parameters, and J is the loss function. It is the gradient of the loss function J with respect to the parameter θ, x (i) Let y be the input feature of the i-th sample. (i) Let be the true label of the i-th sample.

[0152] The training set is input into the power pole detection model to obtain the predicted output value of the training set; the predicted output value is compared with the actual location information of the power pole to obtain the precision, recall and average precision of the calculation results.

[0153] (1.1) To evaluate the power pole detection model, it is first necessary to provide a cross-union ratio (CUP) as a metric for the accuracy of the measurement bounding box of the power pole detection model:

[0154]

[0155] Where IOU is the intersection-union ratio, which is used to observe the overlap rate between the predicted output value of the power pole detection model and the actual location of the power pole. Pre is the predicted output value of the model, GT is the actual location information of the power pole, ∩ is the intersection, and ∪ is the union.

[0156] (1.2) Based on the calculation of the IOU value, by setting a threshold, the number of correctly detected (TP), the number of false alarms (FP), and the number of missed detections (FN) in the synthetic aperture radar image of the power pole to be identified can be obtained. Then, the precision and recall of the power pole detection model can be obtained.

[0157]

[0158] By adjusting the confidence threshold of the power pole detection model output, the precision-recall curves can be derived from the above formula. The area formed by the precision-recall curve on the XY coordinate axis is AP (Average Precision). If the IOU threshold is set to α, then the AP corresponding to this confidence threshold is represented as AP@α.

[0159] Furthermore, the F1 (F-measure) score, as the harmonic mean of precision and recall, can be used to evaluate the performance of the power pole detection model. The specific calculation formula is as follows:

[0160]

[0161] Furthermore, FPS (Frames Per Second) is a metric for evaluating the real-time detection capability of power pole detection models, representing the number of image frames processed by the model per second. A higher FPS indicates a faster processing speed, making it suitable for real-time applications. FPS can be calculated using the following formula:

[0162]

[0163] 2) An example of the training rounds for the detection network based on generalized convolutional attention in this invention is as follows:

[0164] During the training of a detection network based on generalized convolutional attention, the relationship between the training loss and the number of iterations is as follows: Figure 7 As shown. From Figure 7 As can be seen, the detection network based on generalized convolutional attention converges rapidly. In the first 10 epochs, both the training and validation loss functions decrease rapidly; in subsequent epochs, the loss functions gradually converge, eventually leveling off during training and slowly increasing during validation. After validation on the test set, the result of the 88th epoch was selected as the power pole detection model.

[0165] To comprehensively evaluate the effectiveness of the power pole detection model of the present invention, it was compared with existing models according to the index type as follows. The models compared included Faster R-CNN (Faster Region-based Convolutional Neural Network), RTMDet (Real-Time Multi-Object Detection Network), and YOLOX (YouOnly Look Once Extended, a regression-based object detection algorithm). The comparison results are shown in Table 3 below.

[0166] Table 3 Comparison Results

[0167]

[0168] As shown in Table 3, the accuracy of the power pole detection model proposed in this invention is superior to that of the other models in the comparative examples. This is likely because the power pole detection model of this invention adds a cross-union ratio (CUI) metric to measure the accuracy of the measurement bounding box, resulting in an F1 score of 90.09% and an AP@α score of 94.26. Compared to the YOLOX algorithm in the comparative examples, the power pole detection model improves AP@α and F1 scores by 7.16% and 1.99%, respectively. Furthermore, it can be seen that the precision and recall of the power pole detection model of this invention are relatively balanced, with no significant missed detections or false positives.

[0169] like Figure 8 This invention demonstrates the ability of a power pole detection model to identify power poles in different scenarios. The red boxes indicate power poles, and the numbers represent confidence levels. Figure 8 (a, b, c) are detection images of a plain area. Due to different imaging incident angles, the sizes of the power towers in the SAR images also vary. The algorithm of this invention is for... Figure 8 The isolated power tower identification in (a) did not produce false alarms. Figure 8 No dense power tower areas in (b) and (c) were missed during inspection. Figure 8 (d, e, f) show the detection results of power towers in urban and mountainous areas. Due to the high backscattering coefficient of buildings and the shrinking imaging of hillsides, power towers in these areas are more difficult to identify than in plains. However, the power tower detection model of this invention did not generate false alarms in built-up areas and did not miss any detections on hillsides. Therefore, it can be seen that the power tower detection model can accurately identify power towers in different regions and has strong robustness.

[0170] Furthermore, the improved generalized convolutional attention module and the introduction of knowledge distillation in this invention are analyzed and verified; specifically as follows:

[0171] Table 4 below lists the performance of the initial detection model with the improved generalized convolutional attention module (Experiment 2) and the model before improvement (Experiment 1). The initial detection model with the improved generalized convolutional attention module improved the F1 score by 1.41% and AP@α by 1.09%. After adding the distillation loss function (Experiment 3), the recall and AP@α of the initial detection model were significantly improved, indicating that the improved initial detection model (Experiment 4) performed better in recognizing difficult targets and regressing bounding boxes. Although these improvements improved the recognition accuracy of the initial detection model, they also led to a decrease in the efficiency of the initial detection model. After the improvement, the FPS of the initial detection model decreased by about 9.14%, which was 13.47% slower than the YOLOX model. To alleviate the need for high accuracy and high efficiency in the network, this invention introduces a knowledge distillation method. The input image size of the student model (Experiment 5) after distillation is still 512×512 pixels, but the convolutional feature dimension is half that of the teacher network (Experiment 4). The student network achieved a speed of 84.7 FPS, enabling real-time detection without compromising detection efficiency while maintaining algorithm accuracy. Comprehensive analysis shows that the method proposed in this invention significantly improves performance compared to current mainstream algorithms.

[0172] Table 4

[0173]

[0174] The above verification confirms that the power pole detection model of this invention achieves high-precision power pole detection and can adapt to power pole identification in different scenarios. Compared with mainstream target detection networks, the innovations of this power pole detection model are: 1. It integrates a detection network based on generalized convolutional attention. It uses features from the GCAL module to extract deeper features more suitable for power pole detection. 2. It proposes a loss function based on IOU focusing. It assigns greater loss weights to targets that are difficult to identify, alleviating the problems of imbalanced positive and negative detection samples and missed detection of difficult targets. 3. It simplifies the complex teacher network using knowledge distillation. It improves detection efficiency while maintaining accuracy. Experiments show that the power pole detection model of this invention can effectively identify power poles in various scenarios, with F1 scores and AP@α both exceeding 90%; the student network trained through knowledge distillation achieves a 33% improvement in detection speed while maintaining high accuracy.

[0175] In summary, the main improvements of the power pole detection model proposed in this invention are as follows: (1) To address the problem of reduced power pole identifiability due to coherent noise in the SAR image background, a detection network based on generalized convolutional attention is proposed. This network extracts features at different scales, enhancing the initial detection model's understanding of information at different levels, making it easier for the trained power pole detection model to learn the detection results. (2) To address the imbalance between background and target power poles in the synthetic aperture radar image of the target power pole and the difficulty in identifying small power poles, a loss function is introduced to improve the accuracy and robustness of the initial detection model in detecting imbalanced and difficult targets. (3) To address the heavy resource burden of high-performance networks, knowledge distillation technology is used to compress the high-precision model into a smaller, lighter model. While maintaining accuracy, the model's real-time processing capability is improved, enabling the trained power pole detection model to be deployed on various power pole detection platforms.

[0176] Those skilled in the art should know that:

[0177] In existing deep learning-based synthetic aperture radar (SAR) image-based power pole detection methods and systems, researchers have proposed a constant false alarm rate (CFAR) method based on Bayesian theory to address noise and clutter interference issues, enabling radar signal detection in Gaussian background noise. While this method is currently the most widely and thoroughly studied, its accuracy is relatively low and its generalization ability is limited. To improve detection accuracy, researchers have introduced a detection algorithm incorporating deep learning. This algorithm processes SAR images through deep neural networks, directly using the original pixels of the image as input, reducing data processing requirements and avoiding complex feature engineering.

[0178] To ensure the effectiveness of power pole detection, researchers have used deep learning networks to model specific target detection problems. Deep learning-based target detection models mainly include: a two-stage detection model using R-CNN (Faster Region-based Convolutional Neural Network) and a one-stage model using YOLO (You Only Look Once). The two-stage algorithm, based on candidate regions, transforms the target detection problem into a classification problem, extracting target features through convolutional neural networks for target classification and localization. This method enhances the accuracy of target detection, but is time-consuming in the candidate region extraction process. The one-stage algorithm, based on regression, combines feature extraction and target detection processes, directly outputting coordinates, category, and confidence information. It has advantages such as simple structure, fast processing speed, and high accuracy.

[0179] To meet real-time requirements, existing technologies have improved the one-stage algorithm to ensure accuracy while maintaining high speed. Detection speed and efficiency have been improved by optimizing the network structure and reducing computational load. Liang Yiqing et al., while modifying the loss function, proposed the SAR-YOLO-960 (Synthetic Aperture Radar-You Only Look Once-960, a fast target detection model based on synthetic aperture radar data) algorithm based on the characteristics of SAR image targets. This algorithm achieved high-performance detection on high-resolution SAR image datasets, with a detection speed of 32.8 frames / second and precision and recall both exceeding 90%. Yang Zhi et al. proposed a two-stage target detection algorithm combining YOLOv2 (You Only Look Once version 2, the second version of the fast target detection model) and the VGG (Visual Geometry Group) model. YOLO is used for initial detection, and then VGG (Visual Geometry Group) is used to eliminate false positives. The algorithm was validated in COSMO (Constellation of Small Satellites for Mediterranean Basin Observation) images, and the results showed that the method can accurately identify power tower targets in SAR images. Ye Shufen et al. proposed a lightweight network structure and achieved high recognition rate and real-time performance on the Jetson Nano embedded platform. Although the above methods have achieved good results in power tower detection, there are still some problems in practical applications. (1) There are certain differences in the shape, height and structural features of power towers. Due to the change in the incident angle of SAR imaging, power towers in different locations appear differently in the image. This requires complex deep neural networks to achieve high-precision detection. (2) Compared with power towers in farmland and open areas, power towers in forests and mountains are blurred in shape in SAR images and are easily affected by clutter, which increases the difficulty of identification. (3) Lightweight algorithm and low resource requirements are important indicators for model deployment. High-performance detection networks usually require a lot of computing resources. Therefore, the current challenge is how to significantly reduce computational burden and energy consumption while maintaining high performance, and improve system efficiency and reliability.

[0180] Therefore, traditional manual inspection methods are not only time-consuming and labor-intensive, but also pose safety risks. Compared with traditional technologies, the method of this invention can quickly acquire a large number of images containing transmission towers within a certain area, meeting the needs for rapid monitoring and accurate detection.

[0181] Example 2:

[0182] like Figure 9 As shown, this invention application proposes a deep learning-based synthetic aperture radar image power pole detection system, comprising:

[0183] The acquisition module is used to acquire synthetic aperture radar images of power poles to be identified;

[0184] The power pole detection result acquisition module is used to input the synthetic aperture radar image to be identified into the trained power pole detection model for inference, and obtain the power pole detection result in the synthetic aperture radar image to be identified.

[0185] The power pole detection model first uses historical synthetic aperture radar images to train a detection network based on generalized convolutional attention to obtain an initial detection model, and then uses knowledge distillation to train the initial detection model again.

[0186] Furthermore, the system also includes: a power pole detection model construction module, used for:

[0187] Historical synthetic aperture radar (SAR) images of power poles are acquired and used as input for training data.

[0188] The historical power pole labels in the historical synthetic aperture radar images are used as the output of the training data.

[0189] Using the input and output of the training data, a detection network based on generalized convolutional attention is trained to obtain an initial detection model;

[0190] The initial detection model is used as the teacher network; the number of feature channels in the initial detection model is reduced, and the reduced initial detection model is used as the student network.

[0191] Using knowledge distillation, the teacher network guides the student network in training, and the trained student network is then identified as the power pole detection model.

[0192] Furthermore, the acquisition module includes:

[0193] The preprocessing submodule is used to preprocess the synthetic aperture radar image to be identified;

[0194] The preprocessing includes: metadata tagging, radiometric calibration, multi-view processing, and geocoding.

[0195] Furthermore, the initial detection model includes:

[0196] The prediction output acquisition unit is used to predict the input of the training data using a detection network based on generalized convolutional attention, and obtain the prediction output.

[0197] The loss function value calculation unit is used to calculate the loss function value based on the output of the training data and the prediction output;

[0198] The initial detection model is determined by the first unit, which is used to determine the detection network based on generalized convolutional attention corresponding to the model parameters when the iteration termination condition is reached, if the current iteration termination condition is reached. The iteration termination condition includes the loss function value reaching a preset loss threshold and / or the current iteration number reaching a preset iteration number.

[0199] The initial detection model determines the second unit, which is used to adjust the model parameters of the detection network based on generalized convolutional attention according to the loss function value if the iteration termination condition is not met, and then use the detection network based on generalized convolutional attention with adjusted model parameters to re-predict the input of the training data until the iteration termination condition is met, thus obtaining the initial detection model.

[0200] Furthermore, the detection network based on generalized convolutional attention has multiple levels, and each level has a generalized convolutional attention module;

[0201] The initial detection model also includes:

[0202] The enlarged feature tensor acquisition unit is used to take the input of the training data as the input of the detection network based on generalized convolutional attention; take the first level among the multiple levels as the current level, take the feature tensor output by the generalized convolutional attention module in the current level as the initial feature tensor, pass the initial feature tensor to the generalized convolutional attention module of the next level, and then enlarge the size of the feature tensor output by the generalized convolutional attention module of the next level to form the enlarged feature tensor;

[0203] The progressively expanding feature tensor forming unit is used to take the next level as the new current level, and the expanded feature tensor as the feature tensor of the generalized convolutional attention module of the new current level, and pass it to the generalized convolutional attention module of the next level of the new current level, until the feature tensor output by the generalized convolutional attention module of the last level among the multiple levels is obtained, thus forming a feature tensor that is progressively expanded according to the multiple levels.

[0204] The initial detection model acquisition unit is used to use the gradually expanding feature tensor as the prediction output, and to train the detection network based on generalized convolutional attention based on the output of the training data and the prediction output to obtain the initial detection model.

[0205] Furthermore, the gradually expanding feature tensor forming unit includes:

[0206] An initial feature tensor processing subunit is used to pass the initial feature tensor to the next level of the generalized convolutional attention module, and use the next level of the generalized convolutional attention module to process the initial feature tensor to obtain the processed initial feature tensor.

[0207] An initial feature tensor segmentation subunit is used to segment the initial feature tensor and the processed initial feature tensor into multiple partial tensors according to the same segmentation rules;

[0208] Sub-feature tensor forming sub-units are used to concatenate each part tensor of the initial feature tensor with the corresponding part tensor in the processed initial feature tensor to form multiple sub-feature tensors;

[0209] The expanded feature tensor acquisition subunit is used to concatenate multiple sub-feature tensors; the concatenation result is used as the expanded feature tensor output by the next level of generalized convolutional attention module.

[0210] Furthermore, the enlarged feature tensor satisfies the following formula:

[0211] Out=X·MLP(flatten(X))·σ·(ConvB(X))

[0212] In the above formula, MLP represents the next level of generalized convolutional attention module, ConvB() represents the single-channel weight matrix generated by performing convolution operation on the convolutional blocks in the generalized convolutional attention module; σ represents the nonlinear function used for activation; flatten() represents the concatenation of multiple sub-feature tensors; X represents multiple sub-feature tensors; and Out represents the expanded feature tensor.

[0213] Furthermore, the power pole detection model includes:

[0214] An intermediate feature acquisition unit is used to input the training data into the teacher network and acquire the intermediate features output by the teacher network; and to input the training data into the student network and acquire the intermediate features output by the student network.

[0215] The distillation loss function value determination unit is used to calculate the distance metric between the intermediate features of the student network and the intermediate features of the teacher network; and to determine the distance metric between the intermediate features of the student network and the intermediate features of the teacher network as the distillation loss function value.

[0216] The trained student network is designated as the first unit, which is used to determine the student network corresponding to the model parameters when the distillation loss function value reaches the preset distillation loss threshold as the trained student network if the distillation loss function value reaches the preset distillation loss threshold.

[0217] The trained student network determines a second unit, which is used to adjust the model parameters of the student network according to the distillation loss function value if the distillation loss function value does not reach the preset distillation loss threshold. The intermediate features re-output by the student network after adjusting the model parameters are used to iterate repeatedly until the distillation loss function value reaches the preset distillation loss threshold, thus obtaining the trained student network.

[0218] The feature dimension extraction unit is used to extract the feature dimensions in the fully convolutional adaptive layers of the teacher network and the trained student network, respectively.

[0219] The power pole detection model confirmation unit is used to align the feature dimensions of the trained student network and the teacher network based on the feature dimensions in the fully convolutional adaptive layers of the teacher network and the trained student network; and to use the student network with aligned feature dimensions as the power pole detection model.

[0220] Furthermore, the expression for the distillation loss function is as follows:

[0221]

[0222] In the above formula, loss dis N represents the distillation loss value of the distillation loss function. p f is the size of the feature. cov W represents the total number of feature dimensions used to calculate the distillation loss; H represents the width of the feature map output in the fully convolutional adaptive layer; c represents the height of the feature map output in the fully convolutional adaptive layer; c represents the number of the c-th feature channel; C represents the total number of feature channels; S represents the intermediate features of the student network; t represents the intermediate features of the teacher network; and i and j represent the two-dimensional spatial coordinates of the feature map.

[0223] In response to the problem that the identification of power poles in high-resolution synthetic aperture radar (SAR) images is easily affected by clutter interference and local incident angle differences, posing a significant challenge to the power pole detection task, this invention introduces a detection network based on generalized convolutional attention. This network utilizes SAR images to acquire SAR images of power poles to be identified, and proposes a deep learning-based SAR image power pole detection system. This system exhibits significant advantages in the following two aspects: (1) To enhance the feature extraction capability and improve the detection accuracy of the power pole detection model, the feature extraction of the detection network based on generalized convolutional attention is improved by gradually expanding the feature tensor, thereby enhancing the accuracy and stability of the power pole detection model in identifying SAR images. (2) A knowledge distillation algorithm is used to reduce the resource consumption of the complex power pole detection model, improving the operating efficiency of the power pole detection model while maintaining high-precision detection.

[0224] Example 3:

[0225] like Figure 10 As shown, the present invention also provides an electronic device, which may be a computer device, a microcontroller device, a smart mobile device, etc. The electronic device in this embodiment may include a processor, a memory, a transceiver component, etc. The memory, processor, and transceiver component are connected via a bus; the memory can be used to store executable programs, and an exemplary executable program may include instructions; the processor is used to execute the instructions stored in the memory. The memory can also be used to store data, which can be accessed and / or modified when instructions are executed.

[0226] The processor may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, and it is suitable for implementing one or more instructions. Specifically, it is suitable for loading and executing one or more instructions in the storage medium to realize the corresponding method flow or corresponding function, so as to realize the steps of the deep learning synthetic aperture radar image power pole tower detection method in the above embodiments.

[0227] Example 4:

[0228] Based on the same inventive concept, this invention also provides a readable storage medium, specifically an electronic device readable storage medium (Memory). This readable storage medium is a memory device within an electronic device used to store programs and data. It is understood that the storage medium here can include both built-in storage media within the electronic device and extended storage media supported by the electronic device. The storage medium provides storage space, which stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more executable programs (including program code). It should be noted that the storage medium here can be high-speed RAM or non-volatile memory, such as at least one disk storage device. Loading and executing one or more instructions stored in the storage medium by the processor can implement the steps of the deep learning-based synthetic aperture radar image power pole detection method described in the above embodiments.

[0229] Those skilled in the art will understand that embodiments of this invention can be provided as methods, systems, or computer program products. Therefore, this invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this invention can take the form of a computer program product embodied 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.

[0230] This invention application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0231] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1The function specified in one or more boxes.

[0232] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0233] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit its protection scope. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that after reading the present invention, they can still make various changes, modifications or equivalent substitutions to the specific implementation of the application, but these changes, modifications or equivalent substitutions are all within the protection scope of the claims pending approval.

Claims

1. A deep learning synthetic aperture radar image power pole tower detection method, characterized in that, The method comprises the following steps: acquire a to-be-identified synthetic aperture radar image of a power tower; input the to-be-identified synthetic aperture radar image into a trained power tower detection model for inference to obtain a power tower detection result in the to-be-identified synthetic aperture radar image; wherein the power tower detection model is first trained using historical synthetic aperture radar images to obtain an initial detection model, and then the initial detection model is trained again using knowledge distillation to obtain a trained power tower detection model; the training process of the power tower detection model comprises: acquire historical synthetic aperture radar images of a power tower and input the historical synthetic aperture radar images as training data; output historical power tower labels in the historical synthetic aperture radar images as training data; train a detection network based on generalized convolution attention using the input and output of the training data to obtain an initial detection model; use the initial detection model as a teacher network, reduce the number of feature channels of the initial detection model, and use the reduced initial detection model as a student network; train the student network using the teacher network in a knowledge distillation manner, and determine the trained student network as a power tower detection model; the detection network based on generalized convolution attention has multiple levels, and each level has a generalized convolution attention module; the training of the detection network based on generalized convolution attention using the input and output of the training data to obtain an initial detection model comprises: input the training data as the input of the detection network based on generalized convolution attention; take the first level of the multiple levels as the current level, take the feature tensor output by the generalized convolution attention module in the current level as the initial feature tensor, and pass the initial feature tensor to the generalized convolution attention module in the next level; then expand the feature tensor output by the generalized convolution attention module in the next level to form an expanded feature tensor; take the next level as the new current level, take the expanded feature tensor as the feature tensor of the generalized convolution attention module in the new current level, and pass the feature tensor to the generalized convolution attention module in the next level of the new current level, until the feature tensor output by the generalized convolution attention module in the last level of the multiple levels is obtained to form a gradually expanded feature tensor according to the multiple levels; take the gradually expanded feature tensor as the prediction output, train the detection network based on generalized convolution attention based on the output of the training data and the prediction output to obtain an initial detection model; the process of taking the next level as the new current level, taking the expanded feature tensor as the feature tensor of the generalized convolution attention module in the new current level, and passing the feature tensor to the generalized convolution attention module in the next level of the new current level until the feature tensor output by the generalized convolution attention module in the last level of the multiple levels is obtained to form a gradually expanded feature tensor according to the multiple levels comprises: The initial feature tensor is transmitted to a next-level generalized convolution attention module, the initial feature tensor is processed by using the next-level generalized convolution attention module, and a processed initial feature tensor is obtained; The initial feature tensor and the processed initial feature tensor are divided into a plurality of partial tensors according to the same segmentation rule; Each partial tensor of the initial feature tensor is spliced with a corresponding partial tensor in the processed initial feature tensor to form a plurality of sub-feature tensors; The plurality of sub-feature tensors are spliced, and a spliced result is taken as an expanded feature tensor output by the next-level generalized convolution attention module.

2. The method of claim 1, wherein, After the power tower to-be-identified synthetic aperture radar image is acquired, the following steps are included: The to-be-identified synthetic aperture radar image is preprocessed. The preprocessing includes metadata labeling, radiation scaling, multi-view processing, and geographic coding.

3. The method of claim 1, wherein, The training of the generalized convolution attention-based detection network is performed by using the input and output of the training data to obtain an initial detection model, including: The input of the training data is predicted by using the generalized convolution attention-based detection network to obtain a predicted output; A loss function value is calculated based on the output of the training data and the predicted output; If the current iteration termination condition is reached, the generalized convolution attention-based detection network corresponding to the model parameters at the time when the iteration termination condition is reached is determined as the initial detection model; the iteration termination condition includes that the loss function value reaches a preset loss threshold and / or the current iteration number reaches a preset iteration number; If the current iteration termination condition is not reached, the model parameters of the generalized convolution attention-based detection network are adjusted according to the loss function value, and the input of the training data is predicted again by using the generalized convolution attention-based detection network with adjusted model parameters until the iteration termination condition is reached to obtain the initial detection model.

4. The method of claim 1, wherein, The expanded feature tensor satisfies the following formula: In the above formula, represents a generalized convolution attention module of the next level, represents a convolution operation on the convolution block in the generalized convolution attention module, and a single-channel weight matrix is generated; represents a nonlinear function for activation; () represents splicing of a plurality of sub-feature tensors; X represents a plurality of sub-feature tensors; represents an expanded feature tensor.

5. The method of claim 1, wherein, The student network is trained by using the teacher network in a knowledge distillation manner, and the trained student network is determined as a power tower detection model, including: The training data is input into the teacher network to obtain intermediate features output by the teacher network, and the training data is input into the student network to obtain intermediate features output by the student network; A distance measure between the intermediate features of the student network and the intermediate features of the teacher network is calculated; The distance measure between the intermediate features of the student network and the intermediate features of the teacher network is determined as a distillation loss function value; If the distillation loss function value reaches a preset distillation loss threshold, the student network corresponding to the model parameters at the time when the preset distillation loss threshold is reached is determined as the trained student network. If the distillation loss function value does not reach the preset distillation loss threshold, the model parameters of the student network are adjusted according to the distillation loss function value, and the intermediate features output by the student network after the model parameters are adjusted are used for repeated iteration until the distillation loss function value reaches the preset distillation loss threshold, so as to obtain the trained student network; Respectively extract the feature dimensions in the full convolution adaptive layer of the teacher network and the trained student network; Align the feature dimensions of the trained student network and the teacher network based on the feature dimensions in the full convolution adaptive layer of the teacher network and the trained student network; and the student network after the alignment of the feature dimensions is used as the power tower detection model.

6. The method of claim 5, wherein, The expression of the distillation loss function is as follows: In the above formula, is the total number of feature dimensions for calculating the distillation loss, is the width of the output feature map in the fully convolutional adaptive layer; is the height of the output feature map in the fully convolutional adaptive layer, and c is the number of feature channels, is the total number of feature channels; is the intermediate feature of the student network; t is an intermediate feature of the teacher network, and denotes a two-dimensional spatial coordinate position of the feature map.

7. A deep learning synthetic aperture radar image power pole detection system, characterized by, It comprises: An acquisition module is configured to acquire a to-be-identified synthetic aperture radar image of a power tower; A power tower detection result acquisition module is configured to input the to-be-identified synthetic aperture radar image into the trained power tower detection model for inference to obtain a power tower detection result in the to-be-identified synthetic aperture radar image. The power tower detection model first trains a detection network based on generalized convolution attention using historical synthetic aperture radar images to obtain an initial detection model, and then trains the initial detection model again using knowledge distillation to obtain a trained detection model. The system further comprises a power tower detection model construction module configured to: Acquire historical synthetic aperture radar images of a power tower and input the historical synthetic aperture radar images as training data; Output historical power tower labels in the historical synthetic aperture radar images as training data; Train a detection network based on generalized convolution attention using the input and output of the training data to obtain an initial detection model; Use the initial detection model as a teacher network, reduce the number of feature channels of the initial detection model, and use the reduced initial detection model as a student network; Use the teacher network to guide the training of the student network in a knowledge distillation manner, and determine the trained student network as a power tower detection model; The detection network based on generalized convolution attention has multiple levels, and each level has a generalized convolution attention module; The initial detection model comprises: An expanded feature tensor acquisition unit is configured to input the training data as input of the detection network based on generalized convolution attention; use a first level in the multiple levels as a current level, output a feature tensor of the generalized convolution attention module in the current level as an initial feature tensor, transmit the initial feature tensor to a generalized convolution attention module in a next level, and then expand the size of a feature tensor output by the generalized convolution attention module in the next level to form an expanded feature tensor. The gradually expanded feature tensor forming unit is configured to pass, as a new current level and an expanded feature tensor of a generalized convolution attention module of the new current level, the next level of the new current level to a generalized convolution attention module of the next level of the new current level, until a feature tensor output by a generalized convolution attention module of a last level of the multiple levels is obtained, thereby forming the gradually expanded feature tensor according to the multiple levels; The initial detection model obtaining unit is configured to take the gradually expanded feature tensor as a prediction output, train the generalized convolution attention-based detection network based on an output of the training data and the prediction output, and obtain an initial detection model; The gradually expanded feature tensor forming unit comprises: An initial feature tensor processing subunit configured to pass the initial feature tensor to a generalized convolution attention module of a next level, and process the initial feature tensor by using the generalized convolution attention module of the next level to obtain a processed initial feature tensor; An initial feature tensor segmentation subunit configured to segment the initial feature tensor and the processed initial feature tensor into multiple partial tensors according to a same segmentation rule; A sub-feature tensor forming subunit configured to respectively splice each partial tensor of the initial feature tensor with a corresponding partial tensor of the processed initial feature tensor to form multiple sub-feature tensors; An expanded feature tensor obtaining subunit configured to splice the multiple sub-feature tensors, and take a splicing result as an expanded feature tensor output by the generalized convolution attention module of the next level.

Images