A high-resolution satellite remote sensing image transmission tower extraction method

Abstract

The application provides a high-resolution satellite remote sensing image transmission tower extraction method, comprising the following steps: acquiring satellite remote sensing images labeled with to-be-extracted target position information; constructing a small target extraction model, the small target extraction model being composed of a feature extraction network, a feature fusion network and a target detection head; wherein the feature extraction network captures feature maps from the satellite remote sensing images and weights the feature maps through an attention mechanism, and the feature fusion network performs feature fusion on the weighted feature maps; the small target extraction model is trained; the trained small target extraction model is used to extract to-be-extracted target information in real-time satellite remote sensing images, and accurate monitoring and management of a power system are completed according to the to-be-extracted target information. The application has the beneficial effect that the feature extraction network is used to extract and enhance to-be-extracted target features, and the feature fusion network is used to gradually fuse features layer by layer, so that the network can accurately extract to-be-extracted targets in high-resolution satellite images.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of image recognition, and particularly relates to a high-resolution satellite remote sensing image power transmission tower extraction method. BACKGROUND

[0002] Power transmission line inspection is an important means to ensure the stable operation of the power system, high-resolution satellite remote sensing technology makes non-contact, large-scale environmental and infrastructure monitoring possible, and provides a new solution for power transmission line inspection, and power tower detection is the basis for power transmission line inspection;

[0003] Traditional optical satellite remote sensing image target detection methods usually rely on machine learning and image processing technology, involving key steps such as region selection, feature extraction and classifier design, however, the generalization ability of the detection model based on the traditional image processing method is poor in different scenes, and the deep learning method shows a certain effectiveness in detecting power towers in satellite remote sensing images, but it is still difficult to be directly applied to the detection task of power transmission towers in complex backgrounds in high-resolution satellite images, the reason is that:

[0004] In the target detection method of deep learning, the feature extraction network is used to extract high-dimensional features, however, in the detection of small targets in satellite remote sensing images, the proportion of the extracted features in the output feature map is small, which makes it difficult for the network to recognize;

[0005] In the detection of power transmission towers in satellite remote sensing images, there is a challenge of multi-scale detection, in the benchmark model, high-level features need to be propagated and interacted through multiple intermediate scales before being fused with low-level features, in this propagation and interaction process, the semantic information of high-level features may be lost or degraded. SUMMARY

[0006] Therefore, the application aims to provide a high-resolution satellite remote sensing image power transmission tower extraction method, so as to solve at least one of the above technical problems.

[0007] To achieve the above purpose, the technical scheme of the application is as follows:

[0008] A high-resolution satellite remote sensing image power transmission tower extraction method, comprising the following steps:

[0009] Obtaining satellite remote sensing images labeled with target position information to be extracted;

[0010] Constructing a small target extraction model, the small target extraction model is composed of a feature extraction network, a feature fusion network and a target detection head; wherein the feature extraction network captures feature maps from satellite remote sensing images and weights the feature maps through an attention mechanism, and the feature fusion network performs feature fusion on the weighted feature maps;

[0011] training the small target extraction model;

[0012] extracting the to-be-extracted target information in the real-time satellite remote sensing image using the trained small target extraction model, and completing the accurate monitoring and management of the power system according to the to-be-extracted target information.

[0013] Further, the process of obtaining the satellite remote sensing image labeled with the to-be-extracted target position information is specifically as follows:

[0014] Obtain multiple high-resolution satellite remote sensing images that capture the to-be-extracted target, and group the multiple satellite remote sensing images into a data set;

[0015] Label the positions of the to-be-extracted target in all images of the data set, and divide the labeled data set into a verification set and a training set according to a 1:9 ratio.

[0016] Further, the feature extraction network is composed of multiple convolutional layers and LSKA-AFF modules, and the satellite remote sensing image processed by the multiple convolutional layers is used as an original feature map, and the following operations are performed:

[0017] The original feature map sequentially passes through a batch normalization layer, a 1x1 convolution, a Gaussian error linear unit, an LSKA module, and a 1x1 convolution to obtain a first feature map;

[0018] The original feature map and the first feature map are input into the AFF module to obtain a second feature map;

[0019] The second feature map sequentially passes through a batch normalization layer and a convolution feedforward layer to obtain a third feature map;

[0020] The second feature map and the third feature map are input into the AFF module to obtain an output feature map.

[0021] Further, the LSKA module receives the feature map output by the Gaussian error linear unit and performs the following operations:

[0022] Convolve the feature map with two layers of split depth convolution to obtain output

[0023] Convolve the with two layers of split depth separable convolution, and pass through a 1x1 convolution to obtain output A C ;

[0024] The element product of the output feature map F and A C .

[0025] Further, the AFF module receives two feature maps and performs the following operations:

[0026] Element-wise sum of the two feature maps to obtain a weight array;

[0027] Element-wise multiplication of the weight array and one feature map, and element-wise multiplication of the reverse array and the other feature map, the reverse array = 1 - weight array;

[0028] Add the results of the two element-wise multiplications to obtain a weighted feature map.

[0029] Further, the feature fusion network is composed of a progressive feature pyramid network and a path aggregation network, and the working process of the feature fusion network is as follows:

[0030] Obtain the features of multiple convolution layers of the feature extraction network;

[0031] Input the low-level features into the front progressive feature pyramid network;

[0032] Input the output of the front progressive feature pyramid network and the high-level features into the rear progressive feature pyramid network to obtain multi-scale features;

[0033] Input the multi-scale features into the path aggregation network to obtain the detection result of the target to be extracted.

[0034] Further, the P-PANet network includes an ASFF module, which receives the weighted feature map and performs the following operations:

[0035] Extract feature vectors of the same position in feature maps with different weights;

[0036] Adjust the weight corresponding to each feature vector until the weights of all feature vectors meet the constraint condition, the constraint condition being that the sum of all weights is 1;

[0037] Weighted fusion of feature vectors corresponding to each weight.

[0038] Further, in the process of training the small target extraction model:

[0039] Set the batch size of the GPU to 16; select cosine annealing as the decay strategy; select the stochastic gradient descent algorithm as the optimizer; set the initial learning rate to 0.01; set the dynamic parameter to 0.937; set the number of training rounds to 350 iterations; monitor the performance indicators of the small target extraction model, and when the performance indicators do not significantly improve within 100 consecutive iterations, end and complete the training.

[0040] Further, before extracting the target information to be extracted in the real-time satellite remote sensing image, verify the performance indicators of the small target extraction model, if the verification result is up to standard, put the small target extraction model into use, otherwise adjust the small target extraction model until the performance indicators of the small target extraction model meet the standard.

[0041] Further, the performance indicators include:

[0042] Precision, the proportion of correct prediction results in samples predicted as positive classes by the small target extraction model;

[0043] Recall, the proportion of correct prediction by the small target extraction model in samples that are actually positive classes;

[0044] F1 score, the harmonic mean of precision and recall;

[0045] mAP0.5, used to measure the accuracy of the model in the detection task.

[0046] Compared with the prior art, the high-resolution satellite remote sensing image transmission tower extraction method has the following beneficial effects:

[0047] The feature extraction network expands the receptive field by dynamically enhancing the target region features, effectively captures global and local features, and solves the problem of uneven target size and generally small size in complex background; through multi-scale feature fusion, the detection ability of different scale targets is enhanced, thereby improving the performance of the model in complex background; the feature fusion network further improves the target detection ability of the model in complex background through feature gradual fusion and adaptive spatial weighting. BRIEF DESCRIPTION OF DRAWINGS

[0048] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application, and are incorporated in and constitute a part of this application. The embodiments of the present application, and their

[0049] Figure 1 The flowchart of the high-resolution satellite remote sensing image to be extracted target extraction method according to the embodiment of the present application is shown in the figure;

[0050] Figure 2 The LSKAFF-YOLO structure diagram according to the embodiment of the present application is shown in the figure;

[0051] Figure 3 The LSKA-AFF structure diagram according to the embodiment of the present application is shown in the figure;

[0052] Figure 4 The LSKA structure diagram according to the embodiment of the present application is shown in the figure;

[0053] Figure 5 The AFF structure diagram according to the embodiment of the present application is shown in the figure;

[0054] Figure 6 The P-PANet structure diagram according to the embodiment of the present application is shown in the figure;

[0055] Figure 7 The schematic diagram of the ASFF structure is described in the embodiments of the present application. DETAILED DESCRIPTION

[0056] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0057] In the description of the present application, it should be understood that the terms "center", "longitudinal", "transverse", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the purpose of facilitating the description of the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first", "second" and the like are only for the purpose of description and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" and the like can explicitly or implicitly include one or more of the features. In the description of the present application, unless otherwise specified, the meaning of "a plurality of" is two or more.

[0058] In the description of the present application, it should be noted that unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connection" should be understood broadly, for example, it can be fixed connection, or detachable connection, or integrally connected; it can be mechanical connection, or electrical connection; it can be directly connected, or indirectly connected through an intermediate medium, or the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood through specific circumstances.

[0059] The present application will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0060] As shown in the figure, a high-resolution satellite remote sensing image transmission tower extraction method comprises the following steps: Figure 1

[0061] S1, obtaining satellite remote sensing images labeled with position information of the target to be extracted.

[0062] The target to be extracted is a transmission tower, and the process of obtaining satellite remote sensing images labeled with position information of the target to be extracted in step S1 is as follows:

[0063] Obtain a plurality of high-resolution satellite remote sensing images of the target to be extracted, and form a data set by combining the plurality of satellite remote sensing images;

[0064] ​The positions of the targets to be extracted in all images of the labeled data set are marked, and the labeled data set is divided into a verification set and a training set according to a 1:9 ratio.

[0065] In some embodiments, the specific implementation process of step S1 is as follows:

[0066] At least 3000 high-resolution satellite images in which the target to be extracted is photographed are used to form a data set, Labelme tools are used to mark the positions of the power towers in the pictures in the data set, and all data is divided into a training set and a verification set according to a 9:1 ratio. Thus, a labeled data set is created, forming a training set and a test set for training the model.

[0067] S2, constructing a small target extraction model, the small target extraction model being composed of a feature extraction network, a feature fusion network, and a target detection head;

[0068] The feature extraction network captures feature maps from satellite remote sensing images and weights the feature maps through an attention mechanism, and the feature fusion network performs feature fusion on the weighted feature maps.

[0069] As shown in FIG. 1, Figure 2 In some embodiments, the small target extraction model is specifically an LSKAFF-YOLO network, which is composed of a feature extraction network, a feature fusion network, and a target detection head, and the specific parameters of the network are as shown in the following table:

[0070]

[0071]

[0072] The feature extraction network in step S2 is composed of a plurality of convolutional layers and an LSKA-AFF module, and the satellite remote sensing images processed by the plurality of convolutional layers are used as original feature maps, and the following operations are performed:

[0073] The original feature maps sequentially pass through a batch normalization layer, a 1x1 convolution, a Gaussian error linear unit, an LSKA module, and a 1x1 convolution to obtain a first feature map;

[0074] The original feature maps and the first feature maps are input into an AFF module to obtain a second feature map;

[0075] The second feature map sequentially passes through a batch normalization layer and a convolution feedforward layer to obtain a third feature map;

[0076] The second feature map and the third feature map are input into an AFF module to obtain an output feature map.

[0077] The LSKA module receives the feature map output by the Gaussian error linear unit and performs the following operations:

[0078] The feature map is convoluted with two layers of split deep convolution to obtain output

[0079] The is convoluted with two layers of split deep separable convolution, and the output A is obtained through 1x1 convolution C ;

[0080] The element product of the output feature map F and A C .

[0081] The AFF module receives two feature maps and performs the following operations:

[0082] The element sum of the two feature maps is obtained to obtain a weight array;

[0083] The element product of the weight array and one feature map, and the element product of the reverse array and the other feature map are obtained, where the reverse array = 1-weight array;

[0084] The results of the two element products are added to obtain a weighted feature map.

[0085] In some embodiments, as shown in Figure 3 , an LSKA-AFF structure is constructed in the feature extraction network, which has a total of 9 layers for enhancing the features of the target to be extracted. The module mainly consists of a BN (batch normalization) layer, an LSKA module, an AFF module, a GELU (Gaussian Error Linear Unit) and a CFFN (Convolution Feedforward) layer.

[0086] The structure of the LSKA module is as shown in Figure 4 , and the output of the LSKA module is as shown below. The combination of large convolution kernel and attention mechanism significantly expands the receptive field of the model, which can capture long-distance spatial dependencies, enhance the extraction ability of global and local features of the target region to be extracted, and reduce the misidentification of background noise and complex features.

[0087] Given an input feature map F∈R C×H×W , where C is the number of input channels, H and W represent the height and width of the feature map respectively, the output of LSKA can be obtained through formulas (1) to (4), and the input and output calculation formulas are as follows:

[0088]

[0089] A C =W 1×1 *Z C , (3)

[0090]

[0091] Among them, the symbols * and These represent convolution and element-wise multiplication, respectively.

[0092] In formula (1), d represents the hole ratio. The output of DW-Conv can be obtained by convolving the feature map F with two convolution kernels W of size (2d-1). This convolution is used to capture local spatial information and compensate for network effects in the subsequent two layers of decomposed DW-D-Conv (Equation (2)).

[0093] It is important to note that each channel C of the feature map F is convolved with the corresponding channel of the convolution kernel W. The size of the convolution kernel in DW-D-Conv is then... in These represent the floor operation, which is responsible for capturing... The global spatial information, where k also represents the maximum receptive field of the convolution kernel W, increases the model's ability to distinguish different features and effectively handle complex backgrounds and targets.

[0094] Attention map A is obtained by finely convolving the output of the DW-D-Conv layer with a 1×1 convolution kernel. C LSKA output It is attention map A C F and the input feature map C The product of elements.

[0095] The structure of the AFF module is as follows: Figure 5 As shown, the AFF module is used as a feature fusion module at the connection point, focusing on improving the ability to aggregate contextual information and the quality of feature fusion;

[0096] The AFF module can optimize the distribution of feature weights by adjusting the attention weights along the channel dimension, thereby extracting target features more accurately and suppressing irrelevant background information. Its input-output formula is as follows:

[0097]

[0098] in, The feature fusion method for summing elements, Represents element-wise product;

[0099] It is important to note the fusion weights. It consists of real numbers between 0 and 1. Similarly, this allows the network to perform soft selection or weighted averaging between X and Y, which are two distinct feature maps.

[0100] The feature fusion network in step S2 is composed of an incremental feature pyramid network and a path aggregation network, and the working process of the feature fusion network is as follows:

[0101] obtaining features of multiple convolutional layers of the feature extraction network;

[0102] inputting the low-level features into the pre-position incremental feature pyramid network;

[0103] inputting the output of the pre-position incremental feature pyramid network and the high-level features into the post-position incremental feature pyramid network to obtain multi-scale features;

[0104] inputting the multi-scale features into the path aggregation network to obtain a detection result of the target to be extracted.

[0105] The feature fusion network comprises an ASFF module, which receives the weighted feature maps and performs the following operations:

[0106] extracting feature vectors at the same position in feature maps with different weights;

[0107] readjusting the weights corresponding to each feature vector until the weights of all feature vectors satisfy a constraint condition, the constraint condition being that the sum of all weights is 1;

[0108] weighting and fusing the feature vectors corresponding to each weight.

[0109] In some embodiments, the feature fusion network is a P-PANet network, and the structure of the P-PANet is as shown in Figure 6 The P-PANet network is used for layer-by-layer incremental fusion of features, and the network has 17 layers in total. The P-PANet mainly comprises an incremental feature pyramid network (AFPN) and a path aggregation network (PAN), which promotes direct feature fusion between non-adjacent layers, thereby preventing loss or degradation of feature information during transmission and interaction;

[0110] features of layers 4, 6 and 10 are extracted from the feature extraction network to obtain a set of features with different scales, denoted as {P1, P2, P3}. When performing feature fusion, low-level features P1 and P2 are first input into the feature pyramid network, and then P3 is added. After the feature fusion step, a set of multi-scale features {P4, P5, P6} is generated. The semantic gap between non-adjacent layer features is larger than that between adjacent layer features, especially between bottom and top features;

[0111] Therefore, direct fusion of non-adjacent layer features has poor effect, and it is unreasonable to directly use P1, P2 and P3 for feature fusion. The architecture of the AFPN is incremental, which will make the semantic information of different levels of features closer during the incremental fusion process, thereby alleviating the above problems.

[0112] The ASFF module is arranged in the P-PANet network, and the structure of the ASFF module is as shown in the figure Figure 7 In the process of multi-level feature fusion, the ASFF assigns different spatial weights to features of different levels, enhances the importance of key levels, and reduces the influence of contradictory information from different targets. The input and output are as follows:

[0113] Let represent the feature vector from level n to level l at position (i, j), and the feature vector obtained by adaptive spatial fusion of multi-level features is is a linear combination of feature vectors and , defined as follows:

[0114]

[0115] wherein, and represent the spatial weight of the last level feature at level l, and satisfy the constraint condition

[0116] S3, train the small target extraction model.

[0117] In the process of training the small target extraction model in step S3:

[0118] Set the batch size of the GPU to 16; select the cosine annealing strategy; select the stochastic gradient descent algorithm as the optimizer; set the initial learning rate to 0.01; set the dynamic parameter to 0.937; and set the number of training rounds to 350 iterations.

[0119] Use the early stopping strategy to regulate the training process of the small target extraction model, specifically:

[0120] During the training process, monitor the performance indicators of the small target extraction model. When the performance indicators do not significantly improve within 100 consecutive iterations, end and complete the training.

[0121] S4, use the trained small target extraction model to extract the to-be-extracted target information in the real-time satellite remote sensing image, and complete the accurate monitoring and management of the power system according to the to-be-extracted target information.

[0122] Before extracting the to-be-extracted target information in the real-time satellite remote sensing image, verify the performance indicators of the small target extraction model. If the verification result is up to standard, the small target extraction model is put into use, otherwise the small target extraction model is adjusted until the performance indicators of the small target extraction model are up to standard.

[0123] The performance indicators include:

[0124] Precision, the proportion of correct prediction results in the samples predicted as positive classes by the small target extraction model;

[0125] Recall, the proportion of correct prediction by the small target extraction model in the actual positive class samples;

[0126] F1 score, the harmonic mean of precision and recall;

[0127] mAP0.5, used to measure the accuracy of the model in the detection task.

[0128] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the above-mentioned high-resolution satellite remote sensing image power transmission tower extraction method.

[0129] Compared with the prior art, the high-resolution satellite remote sensing image to-be-extracted target extraction method has the beneficial effects that:

[0130] The LSKA-AFF module is constructed, the feature extraction performance of the network is good, and the feature enhancement module has more accurate recognition ability for the features of the to-be-extracted target.

[0131] The P-PANet feature fusion network structure is constructed, which effectively avoids the loss or degradation of information in the feature transmission and interaction process, and realizes the multi-scale feature fusion of the to-be-extracted target in a complex background.

[0132] In order to realize the extraction of the to-be-extracted target in the high-resolution satellite remote sensing image, the above-mentioned LSKAFF-YOLO network is proposed, and the main conclusions are as follows:

[0133] In order to expand the receptive field of the model and effectively utilize the context information of the satellite remote sensing image, the LSKA-AFF is constructed in the feature extraction network, which utilizes the attention feature fusion module (AFF) to enhance the residual connection, and the LSKA-AFF utilizes the large convolution kernel after decomposition to expand the receptive field of the model, and capture the long-distance dependence relationship between global and local features, so as to realize accurate positioning of the to-be-extracted target in a complex background.

[0134] In order to solve the problem that the semantic difference is not aligned in the fusion process of the feature pyramid (FPN) due to simple upsampling and horizontal connection, the feature fusion network of the benchmark model is replaced by P-PANet, P-PANet gradually fuses more fine features layer by layer, effectively reduces the semantic difference between different scale features, and reduces the information loss and feature degradation problem.

[0135] The performance of LSKAFF-YOLO is evaluated on the public SRSPTD dataset and the self-built GFTD dataset. The experimental results show that the mAP0.5 reaches 88.8% and 94.6% respectively, and LSKAFF-YOLO significantly improves the detection accuracy of the target to be extracted in high-resolution satellite remote sensing images.

[0136] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, but not limited to them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and they should be covered in the scope of the claims and description of the present application.

[0137] The above is only the preferred embodiment of the present application, and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A high-resolution satellite remote sensing image transmission tower extraction method, characterized in that, The method comprises the following steps: Obtaining satellite remote sensing images marked with to-be-extracted target position information; Constructing a small target extraction model, which is composed of a feature extraction network, a feature fusion network and a target detection head; wherein the feature extraction network captures feature maps from the satellite remote sensing images and weights the feature maps through an attention mechanism, and the feature fusion network performs feature fusion on the weighted feature maps; Training the small target extraction model; Using the trained small target extraction model to extract to-be-extracted target information in real-time satellite remote sensing images, and completing accurate monitoring and management of the power system according to the to-be-extracted target information; The feature extraction network is composed of multiple convolution layers and LSKA-AFF modules, and the satellite remote sensing images processed by the multiple convolution layers are used as original feature maps, and the following operations are performed: The original feature maps pass through a batch normalization layer, a 1x1 convolution, a Gaussian error linear unit, a LSKA module and a 1x1 convolution in sequence to obtain first feature maps; The original feature maps and the first feature maps are input into an AFF module to obtain second feature maps; The second feature maps pass through a batch normalization layer and a convolution feedforward layer in sequence to obtain third feature maps; The second feature maps and the third feature maps are input into the AFF module to obtain output feature maps. 2.The method of claim 1, wherein, The process of obtaining satellite remote sensing images marked with to-be-extracted target position information is as follows: Obtaining multiple high-resolution satellite remote sensing images that capture to-be-extracted targets, and grouping the multiple satellite remote sensing images into a data set; Labeling the positions of the to-be-extracted targets in all images of the data set, and dividing the labeled data set into a verification set and a training set at a ratio of 1:

9. 3.The method of claim 1, wherein, The LSKA module receives the feature maps output by the Gaussian error linear unit and performs the following operations: convolving the feature map with two layers of split depth convolutions to obtain an output Will The two depthwise separable convolutions are then convolved, and the output A is obtained after a 1×1 convolution. C ; The output feature map F is multiplied by the elements of A C .

4. The method of claim 1, wherein the method further comprises: The AFF module receives two feature maps and performs the following operations: Element-wise summation of the two feature maps to obtain a weight array; Element-wise multiplication of the weight array and one feature map, and element-wise multiplication of the inverse array and the other feature map, wherein the inverse array = 1-weight array; Adding the results of the two element-wise multiplications to obtain a weighted feature map.

5. The method of claim 1, wherein the method further comprises: The feature fusion network is composed of a progressive feature pyramid network and a path aggregation network, and the working process of the feature fusion network is as follows: Obtaining features of multiple convolution layers of the feature extraction network; Inputting low-level features into a front progressive feature pyramid network; Inputting the output of the front progressive feature pyramid network and high-level features into a rear progressive feature pyramid network to obtain multi-scale features; Inputting the multi-scale features into a path aggregation network to obtain detection results of the to-be-extracted targets.

6. The method of claim 1, wherein the method further comprises: The feature fusion network includes an ASFF module, which receives the weighted feature maps and performs the following operations: Extracting feature vectors of the same position in feature maps with different weights; Re-adjusting the weights corresponding to each feature vector until the weights of all feature vectors satisfy the constraint condition, wherein the constraint condition is that the sum of all weights is 1; Weighted fusion of feature vectors corresponding to each weight.

7. The method of claim 1, wherein the method further comprises: determining a height of the transmission tower; and determining a distance between the transmission tower and the satellite. During the training of the small target extraction model, The attenuation strategy is selected as cosine annealing, the optimizer is selected as a stochastic gradient descent algorithm, and a performance index of the small target extraction model is monitored; when the performance index does not have a significant improvement within 100 continuous iterations, the training is ended and completed.

8. The high-resolution satellite remote sensing image transmission tower extraction method according to claim 1, characterized in that: Before extracting the to-be-extracted target information in the real-time satellite remote sensing image, the performance index of the small target extraction model is verified; if the verification result is up to standard, the small target extraction model is put into use; otherwise, the small target extraction model is adjusted until the performance index of the small target extraction model is up to standard.

9. The method of claim 8, wherein the method further comprises: The performance index includes: The accuracy is a proportion of correct prediction results in samples predicted as positive classes by the small target extraction model; the recall rate is a proportion of samples correctly predicted by the small target extraction model in actual positive classes; the F1 score is a harmonic mean of the accuracy and the recall rate; and the mAP0.5 is used to measure the accuracy of the model in the detection task.

Images