Power transmission tower remote sensing target detection method, device, equipment, storage medium and product

Abstract

The application discloses a power transmission tower remote sensing target detection method, device, equipment, storage medium and product, relates to the electric power computer vision technical field, and the power transmission tower remote sensing target detection method comprises the steps that a satellite remote sensing image is input into a power transmission tower detection model to obtain a target detection result, the power transmission tower detection model comprises a detection head, a pyramid structure, a bidirectional weighted feature fusion module and a coordinate attention module, the bidirectional weighted feature fusion module is used for fusing high-level information and low-level information in the pyramid structure into multi-scale features, and the coordinate attention module is used for weighting the multi-scale features based on a multi-direction attention graph to obtain weighted multi-scale features.The application improves the sensitivity to smaller power transmission towers in the satellite remote sensing image by fusing high-level information and low-level information, enhances important direction features by reweighting the multi-scale features, improves the spatial selectivity near the target, and thus improves the accuracy of power transmission tower detection.

Description

Technical Field

[0001] This application relates to the field of power computer vision technology, and in particular to methods, devices, equipment, storage media and products for remote sensing target detection of transmission towers. Background Technology

[0002] As a crucial part of power infrastructure, the effective monitoring of transmission towers is essential for ensuring power grid security. Intelligent inspection of power lines based on satellite remote sensing enables large-scale, operational inspections of power transmission channels, significantly improving efficiency and targeting, and has gradually become an important research direction.

[0003] Current methods utilize computer vision and deep learning to detect power transmission towers in satellite remote sensing images. However, satellite remote sensing images often have complex backgrounds, with many power transmission towers sharing similar features and textures with their surroundings. Furthermore, the towers occupy relatively little pixel area compared to the background, leading to lower detection accuracy. Additionally, satellite remote sensing images typically have low resolution, which decreases further as the number of layers in CNN models increases. Many detailed features in the image become more blurred after multiple convolutional operations, further affecting the accuracy of power transmission tower detection. Summary of the Invention

[0004] The main purpose of this application is to provide a method, apparatus, equipment, storage medium, and product for remote sensing target detection of transmission towers, aiming to solve the technical problem of low detection accuracy of transmission towers.

[0005] To achieve the above objectives, this application proposes a method for remote sensing target detection of transmission towers, the method comprising:

[0006] Acquire satellite remote sensing images containing power transmission towers;

[0007] The satellite remote sensing image is input into a preset transmission tower detection model to obtain the target detection result. The transmission tower detection model is trained based on remote sensing image samples with transmission tower annotations and different backgrounds. The transmission tower detection model includes a detection head, a pyramid structure, a bidirectional weighted feature fusion module, and a coordinate attention module. The bidirectional weighted feature fusion module is used to fuse high-level and low-level information in the pyramid structure into multi-scale features. The high-level and low-level information are feature information extracted from the satellite remote sensing image. The coordinate attention module is used to weight the multi-scale features based on the attention map to obtain weighted multi-scale features, which are then used by the detection head to obtain the target detection result. The attention map is an attention map that determines multiple directions based on the multi-scale features.

[0008] In one embodiment, before the step of inputting the satellite remote sensing image into a preset transmission tower detection model to obtain the target detection result, the method further includes:

[0009] Acquire the remote sensing image samples with transmission tower labels and different backgrounds;

[0010] The remote sensing image samples are input into a preset training model. Based on the efficient feature extraction module in the training model, features are extracted from the remote sensing image samples through different feature extraction depths to obtain the initial features corresponding to each node in the pyramid structure.

[0011] Based on the bidirectional weighted feature fusion module, the initial features of each layer in the pyramid structure are fused with the initial features of the adjacent layers to obtain the multi-scale features;

[0012] The weighted multi-scale features are obtained by weighting the multi-scale features based on the coordinate attention module.

[0013] Based on the detection head and the weighted multi-scale features, the power transmission towers in the remote sensing image samples are detected to obtain the detection results. Based on the detection results and the power transmission tower annotations, the model to be trained is adjusted to obtain the power transmission tower detection model.

[0014] In one embodiment, the step of weighting the multi-scale features based on the coordinate attention module to obtain the weighted multi-scale features includes:

[0015] Global average pooling is performed on the multi-scale features in both the height and width directions to obtain two directional descriptors.

[0016] The two direction descriptors are concatenated, and the concatenated descriptors are convolved by a preset shared convolution module to obtain a shared embedding.

[0017] The shared embedding is split into two shared sub-embedded layers, and based on the shared convolution module and a preset activation function, the shared sub-embedded layers are activated and convolved sequentially to obtain a horizontal attention map and a vertical attention map.

[0018] Based on the horizontal attention map and the vertical attention map, the multi-scale features are weighted to obtain the weighted multi-scale features.

[0019] In one embodiment, the bidirectional weighted feature fusion module includes multiple bidirectional weighted feature fusion units, and the adjacent layer is the layer in the pyramid structure adjacent to the layer where the current node is located. The step of fusing the initial features of each layer in the pyramid structure with the initial features of the adjacent layers based on the bidirectional weighted feature fusion module to obtain the multi-scale features includes:

[0020] For each node in each layer of the pyramid structure, the target feature fusion unit corresponding to the current node is determined from the bidirectional weighted feature fusion unit;

[0021] Determine whether the current node is the preset target node;

[0022] If the current node is the target node, then the adjacent nodes corresponding to the target feature fusion unit are determined from the adjacent layers, and the initial node of the layer where the current node is located is determined;

[0023] Based on preset learnable weights and preset coefficients, the first feature corresponding to the initial node, the second feature corresponding to the current node, and the third feature corresponding to the adjacent node are fused to obtain the multi-scale feature.

[0024] In one embodiment, the learning weights include a first weight, a second weight, and a third weight. The step of fusing the first feature corresponding to the initial node, the second feature corresponding to the current node, and the third feature corresponding to the neighboring nodes based on preset learnable weights and preset coefficients to obtain the multi-scale features includes:

[0025] The initial weighted features are obtained by adding the product of the first weight and the first feature, the product of the second weight and the second feature, and the product of the third weight and the third feature.

[0026] The first weight, the second weight, the third weight, and the preset coefficient are added together to obtain the sum of the characteristic coefficients;

[0027] The target weighted feature is obtained by calculating the quotient of the weighted feature and the sum of the feature coefficients.

[0028] The target weighted features are convolved to obtain the multi-scale features.

[0029] In one embodiment, the transmission tower detection model includes a lightweight convolution module, which includes depthwise separable convolution units. Before the step of fusing the initial features of each layer in the pyramid structure with the initial features of adjacent layers based on the bidirectional weighted feature fusion module, the model further includes:

[0030] The initial features are convolved to obtain a first feature with a corresponding number of channels;

[0031] Based on the depthwise separable convolutional unit, the first feature is convolved to obtain a second feature with the same number of channels as the first feature;

[0032] Based on the first feature and the second feature, the spliced ​​features are obtained by concatenating them along the channel dimension;

[0033] The order of the channels in the splicing feature is shuffled to obtain the processed initial feature.

[0034] Furthermore, to achieve the above objectives, this application also proposes a remote sensing target detection device for transmission towers, the remote sensing target detection device for transmission towers comprising:

[0035] The image acquisition module is used to acquire satellite remote sensing images containing power transmission towers;

[0036] The target detection module is used to input the satellite remote sensing image into a preset transmission tower detection model to obtain the target detection result. The transmission tower detection model is trained based on remote sensing image samples with transmission tower annotations and different backgrounds. The transmission tower detection model includes a detection head, a pyramid structure, a bidirectional weighted feature fusion module, and a coordinate attention module. The bidirectional weighted feature fusion module is used to fuse high-level and low-level information in the pyramid structure into multi-scale features. The high-level and low-level information are feature information extracted from the satellite remote sensing image. The coordinate attention module is used to weight the multi-scale features based on the attention map to obtain weighted multi-scale features, so that the detection head can obtain the target detection result based on the multi-scale features. The attention map is an attention map that determines multiple directions based on the multi-scale features.

[0037] In addition, to achieve the above objectives, this application also proposes a remote sensing target detection device for transmission towers, the device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the remote sensing target detection method for transmission towers as described above.

[0038] In addition, to achieve the above objectives, this application also proposes a storage medium, which is a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the steps of the remote sensing target detection method for transmission towers as described above.

[0039] In addition, to achieve the above objectives, this application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of the remote sensing target detection method for transmission towers as described above.

[0040] One or more technical solutions proposed in this application have at least the following technical effects:

[0041] A satellite remote sensing image containing a power transmission tower is acquired, and the image is input into a preset power transmission tower detection model to obtain a target detection result. The power transmission tower detection model is trained based on remote sensing image samples with tower annotations and different backgrounds. The model includes a detection head, a pyramid structure, a bidirectional weighted feature fusion module, and a coordinate attention module. The bidirectional weighted feature fusion module fuses high-level and low-level information from the pyramid structure into multi-scale features. These high-level and low-level features are extracted from the satellite remote sensing image. The coordinate attention module weights the multi-scale features based on an attention map to obtain weighted multi-scale features, which the detection head uses to obtain the target detection result. The attention map is an attention map that determines multiple directions based on the multi-scale features.

[0042] To address the challenges of complex backgrounds in satellite remote sensing images, where power transmission towers occupy relatively little pixel area compared to the background, and the decreasing resolution of satellite remote sensing images with increasing CNN model layers in deep learning, leading to blurred details after multiple convolutional operations and consequently low detection accuracy for power transmission towers, this application improves the accuracy of power transmission tower detection by employing a power transmission tower detection model that includes a bidirectional weighted feature fusion module and a coordinate attention module. Specifically, this application extracts high-level and low-level features from satellite remote sensing images using a pyramid structure. The bidirectional weighted feature fusion module then fuses features from adjacent levels within the feature pyramid, strengthening the interaction between high-level and low-level features, preserving inter-level information, and enhancing sensitivity to smaller power transmission towers in satellite remote sensing images, thereby improving the accuracy of power transmission tower detection. Furthermore, this application reweights the multi-scale features based on the attention maps determined by the multi-scale features in multiple directions. By weighting, the channels in important directions can be emphasized, while the responses of channels in unimportant directions can be weakened, thereby improving the spatial selectivity near the target. Spatial selection can reduce false alarms caused by repetitive textures and improve the accuracy of transmission tower detection. Attached Figure Description

[0043] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart illustrating an embodiment of the remote sensing target detection method for transmission towers provided in this application.

[0046] Figure 2 This is an illustration of the transmission tower detection effect provided in Embodiment 1 of the remote sensing target detection method for transmission towers in this application;

[0047] Figure 3 This is an overall detection flowchart provided for Embodiment 1 of the remote sensing target detection method for power transmission towers in this application;

[0048] Figure 4 This is a schematic diagram of the overall structure of the model provided in Embodiment 1 of the remote sensing target detection method for transmission towers in this application;

[0049] Figure 5 This is a schematic diagram of the lightweight convolutional structure provided in Embodiment 1 of the remote sensing target detection method for transmission towers in this application;

[0050] Figure 6 This is a flowchart illustrating Embodiment 2 of the remote sensing target detection method for transmission towers provided in this application;

[0051] Figure 7 This is a schematic diagram of the coordinate attention module structure provided in Embodiment 2 of the remote sensing target detection method for transmission towers in this application;

[0052] Figure 8 This is a flowchart illustrating Embodiment 3 of the remote sensing target detection method for transmission towers provided in this application;

[0053] Figure 9 This is a comparative schematic diagram of the bidirectional weighted feature fusion module structure provided in Embodiment 3 of the remote sensing target detection method for transmission towers in this application;

[0054] Figure 10 This is a schematic diagram of the module structure of the remote sensing target detection device for transmission towers according to an embodiment of this application;

[0055] Figure 11 This is a schematic diagram of the equipment structure of the hardware operating environment involved in the remote sensing target detection method for transmission towers in the embodiments of this application;

[0056] Figure 12 This is a schematic diagram illustrating the data acquisition consent process involved in the remote sensing target detection method for power transmission towers in this application embodiment.

[0057] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0058] It should be understood that the specific embodiments described herein are merely illustrative of the technical solutions of this application and are not intended to limit this application.

[0059] To better understand the technical solution of this application, a detailed description will be provided below in conjunction with the accompanying drawings and specific implementation methods.

[0060] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a tablet computer, personal computer, or mobile phone, or an electronic device capable of performing the above functions, such as a remote sensing target detection device for transmission towers. The following description uses a remote sensing target detection device for transmission towers as an example to illustrate this embodiment and the subsequent embodiments.

[0061] With the continuous development and application of artificial intelligence technology, power line inspection is gradually shifting from digitalization to intelligentization. As a crucial part of power infrastructure, the effective monitoring of transmission towers is essential for ensuring power grid security. Compared to drone inspections, intelligent power line inspection based on satellite remote sensing can achieve large-scale, operational power channel inspections, significantly improving efficiency and targeting, and has gradually become an important research direction.

[0062] Currently, computer vision technology is commonly used to identify power transmission towers in satellite inspection images. However, in current satellite remote sensing images, transmission towers may occupy a smaller pixel area compared to the background, and the target scale varies greatly, making it difficult to improve detection accuracy. Furthermore, the resolution of satellite remote sensing images is inherently insufficient for target detection; as the number of layers in CNN models increases, the image resolution continuously decreases, and many detailed features become more blurred after multiple convolutional operations, resulting in insufficient detection accuracy for small transmission towers. The background of satellite remote sensing images is also complex; many transmission towers share similar features and textures with surrounding ground objects, further hindering the improvement of detection accuracy.

[0063] Based on this, embodiments of this application provide a method for remote sensing target detection of transmission towers, referring to... Figure 1 , Figure 1 This is a flowchart illustrating the first embodiment of the remote sensing target detection method for power transmission towers in this application.

[0064] In this embodiment, the remote sensing target detection method for power transmission towers includes steps S10 to S20:

[0065] Step S10: Obtain satellite remote sensing images containing power transmission towers;

[0066] It should be noted that satellite remote sensing images are data that are observed and recorded from space by sensors carried by satellites. They can contain information on various surface features such as terrain, vegetation, and buildings. In this embodiment, the images contain information on power transmission towers and their surrounding environment.

[0067] Step S20: Input the satellite remote sensing image into a preset transmission tower detection model to obtain the target detection result. The transmission tower detection model is trained based on remote sensing image samples with transmission tower annotations and different backgrounds. The transmission tower detection model includes a detection head, a pyramid structure, a bidirectional weighted feature fusion module, and a coordinate attention module. The bidirectional weighted feature fusion module is used to fuse high-level and low-level information in the pyramid structure into multi-scale features. The high-level and low-level information are feature information extracted from the satellite remote sensing image. The coordinate attention module is used to weight the multi-scale features based on the attention map to obtain weighted multi-scale features, so that the detection head can obtain the target detection result based on the multi-scale features. The attention map is an attention map that determines multiple directions based on the multi-scale features.

[0068] It should be noted that the power transmission tower detection model is a deep learning model used to automatically identify and locate power transmission towers from remote sensing images. Its structure includes a detection head, a pyramid structure, a bidirectional weighted feature fusion module, and a coordinate attention module. Target detection results refer to the model's output information about the location of the power transmission tower in the image, represented in this embodiment as bounding boxes, accompanied by category labels and confidence scores. Remote sensing image samples are satellite remote sensing image data with annotations used to train the model. The presence of power transmission tower annotations indicates that the locations of power transmission towers in the image have been marked manually or semi-automatically, and different backgrounds indicate that these samples cover diverse geographical environments.

[0069] The detection head is the sub-network in the object detection model responsible for ultimately predicting the object's category and location. The pyramid structure is a structure in deep neural networks used to extract multi-scale features. By downsampling the input image at different levels, it generates feature maps with different resolutions and semantic levels; higher-level features have stronger semantic information but lower spatial resolution, while lower-level features retain more detail but have weaker semantics.

[0070] The bidirectional weighted feature fusion module is used to weight and integrate high-level features (strong semantics, low resolution) and low-level features (rich details, weak semantics) in the pyramid structure to generate multi-scale features that balance semantics and detail. High-level information refers to features extracted from higher levels of the pyramid structure, which have strong semantic abstraction capabilities but less spatial detail. Low-level information refers to features extracted from lower levels of the pyramid structure, which retain more spatial details of the original image but have weaker semantic expressive capabilities.

[0071] Multi-scale features are feature representations obtained by fusing information from different levels (high and low), containing both semantic information and preserving spatial details, making them suitable for target detection in complex scenes. The coordinate attention module is an attention mechanism module that generates an attention map by modeling attention weights in the horizontal and vertical directions of the feature map, which is used to spatially weight the multi-scale features. The attention map is a two-dimensional weight map generated by the coordinate attention module, reflecting the importance of different spatial locations for the target detection task, and is used to weight and enhance the features. Weighted multi-scale features are feature representations obtained after the coordinate attention module uses the attention map to weight the multi-scale features, strengthening the response to areas related to the transmission tower.

[0072] It is understandable that, since the transmission tower detection model adopts a pyramid structure to extract features at different levels, and fuses high-level and low-level information into multi-scale features through a bidirectional weighted feature fusion module, the model can capture the semantic features of the transmission tower while preserving its fine spatial structure.

[0073] Furthermore, the coordinate attention module in this embodiment generates attention maps in multiple directions based on multi-scale features, and performs weighted enhancement on key areas, thereby highlighting the salience of transmission towers in complex backgrounds. This effectively improves the detection accuracy and positioning precision of transmission towers in diverse backgrounds, reducing missed detections and false detections. Moreover, since the samples used for model training themselves contain diverse backgrounds, the combination of the aforementioned multi-scale fusion and attention mechanism further enhances the model's generalization ability, enabling it to maintain stable performance even in newly encountered image regions. The transmission tower detection effect of this embodiment can be referenced... Figure 2 , Figure 2 This includes the original images, real labels, comparison model detection results (YOLOv11 model) for urban, suburban, and rural areas respectively, and the detection results of the transmission tower detection model of this application.

[0074] In one feasible implementation, the specific implementation method prior to the step of inputting the satellite remote sensing image into a preset transmission tower detection model to obtain the target detection result can also be:

[0075] Remote sensing image samples with different backgrounds and labeled transmission towers are acquired. These samples are then input into a pre-defined training model. Based on the efficient feature extraction module within the training model, features are extracted from the remote sensing image samples at different feature extraction depths to obtain initial features corresponding to each node in the pyramid structure. Using a bidirectional weighted feature fusion module, the initial features of each layer in the pyramid structure are fused with the initial features of adjacent layers to obtain multi-scale features. The multi-scale features are then weighted using a coordinate attention module to obtain weighted multi-scale features. Based on the detection head and the weighted multi-scale features, transmission towers in the remote sensing image samples are detected to obtain detection results. The training model is then adjusted based on the detection results and the transmission tower labels to obtain the transmission tower detection model.

[0076] It should be noted that the model to be trained refers to the initial neural network model that has not yet completed parameter optimization during the training phase. Its structure includes an efficient feature extraction module, a pyramid structure, a bidirectional weighted feature fusion module, a coordinate attention module, and a detection head, which is used to gradually optimize the model into the final transmission tower detection model by learning from remote sensing image samples. The efficient feature extraction module is a sub-network in the model used to quickly and effectively extract basic features from the input remote sensing image samples, supporting feature extraction at different depths to adapt to the multi-level requirements of the pyramid structure.

[0077] Feature extraction depth refers to the number of network layers or abstraction levels used by the efficient feature extraction module when processing images. Different depths correspond to different levels of semantic and detailed information, used to generate the initial features required for each node in the pyramid structure. The initial features are the raw feature maps output by the efficient feature extraction module at a specific feature extraction depth, serving as the basic input for each node in the pyramid structure, and have not yet undergone cross-layer fusion or attention weighting processing.

[0078] The overall process of target detection in this embodiment can be referred to Figure 3 Among them, in Figure 3 In the model training process, this embodiment divides the dataset into a training set and a validation set according to a preset ratio, and uses cross-validation to train the dataset. The model's detection head follows the decoupled classification and regression design of YOLOv11 (an object detection model). The optimal model obtained from the training is used to detect and identify transmission towers and evaluate the model performance. The evaluation metrics for model performance evaluation include precision, recall, and average accuracy.

[0079] The overall structure of the model in this embodiment can be referred to Figure 4 .according to Figure 4As can be seen, the overall structure of the model in this embodiment includes a backbone network, a neck network, and a head network. The backbone network includes a Spatial Pyramid Pooling-Fast (SPPF) module and a C2PSA (a common attention enhancement module in YOLOv11, used to enhance the representation of key regions and improve the ability to extract small targets). The efficient feature extraction module in this embodiment can be the c3k2 module (a variant of the C3 series feature extraction block in the YOLO11 / YOLOv11 architecture (k2 structure), used for efficient feature extraction and fusion).

[0080] Understandably, this embodiment uses remote sensing image samples with transmission tower annotations and different backgrounds during the model training phase. An efficient feature extraction module generates initial features for each node of the pyramid structure at different feature extraction depths, enabling the model to learn the common and dissimilar features of transmission towers from diverse environments. A bidirectional weighted feature fusion module fuses the initial features of adjacent layers, effectively integrating semantic information and spatial details to form more discriminative multi-scale features. Furthermore, a coordinate attention module applies direction-aware weighting to the fused multi-scale features, further enhancing the feature response of the area where the transmission towers are located.

[0081] Therefore, this embodiment enables the training model to fully adapt to the changes in the form of transmission towers under different geographical backgrounds through the training process, significantly improving its generalization ability and detection accuracy. By comparing the detection results with the real annotations and continuously optimizing the model parameters, it ensures that the final transmission tower detection model has high accuracy and strong robustness in practical applications.

[0082] In one feasible implementation, the transmission tower detection model includes a lightweight convolution module, which includes depthwise separable convolution units. The specific implementation prior to fusing the initial features of each layer in the pyramid structure with the initial features of adjacent layers based on the bidirectional weighted feature fusion module can also be:

[0083] The initial feature is convolved to obtain a first feature with a corresponding number of channels. Based on the depthwise separable convolutional unit, the first feature is convolved to obtain a second feature with the same number of channels as the first feature. The first feature and the second feature are concatenated along the channel dimension to obtain a concatenated feature. The order of the channels in the concatenated feature is shuffled to obtain the processed initial feature.

[0084] It's important to note that lightweight convolution modules are convolutional structures that reduce computational cost and parameter count, improving inference efficiency while maintaining model performance. Depthwise separable convolutional units are the core component of lightweight convolution modules, decomposing standard convolution into channel-wise spatial convolution and pointwise channel-wise hybrid convolution, significantly reducing computational complexity.

[0085] The overall process of the lightweight convolution module in this embodiment can be referred to... Figure 5 The formula for calculating standard convolution is:

[0086]

[0087] Where W represents the width of the spatial dimension and H represents the height of the spatial dimension, K1 and K2 represent the height and width of the convolution kernel, and C1 and C2 are the number of input and output channels, respectively. This indicates the computational cost of standard convolution.

[0088] The formulas for calculating depthwise separable convolution and channel shuffling are as follows:

[0089]

[0090] in, It represents the computational cost of depthwise separable convolution and channel shuffling.

[0091] The first feature is an intermediate feature representation obtained by performing a regular convolution operation on the initial feature. Its number of channels is set according to task requirements and is used for subsequent lightweight processing. The second feature is the output feature after processing the first feature through a depthwise separable convolutional unit. Its number of channels is the same as the first feature, but its computational cost is lower and it retains key spatial information. The concatenated feature is a combined feature formed by connecting the first and second features along the channel dimension, fusing the complementary information of standard convolution and lightweight convolution. The processed initial feature is then subjected to channel shuffling on the concatenated feature to obtain the final feature, which enhances the interaction between features and improves the expressive power of the subsequent fusion module.

[0092] Understandably, this embodiment introduces a lightweight convolution module containing depthwise separable convolutional units before fusion. The initial features are processed sequentially using standard convolution and depthwise separable convolution, and the results are concatenated along the channel dimension. This allows the processed features to simultaneously integrate the highly expressive standard convolutional response and the computationally inefficient lightweight convolutional response. By shuffling the channel order of the concatenated features, information interaction between features extracted from different convolutional paths is promoted, alleviating the redundancy problem between channels. This improves the generalization ability of the features while controlling model complexity, enabling the subsequent bidirectional weighted feature fusion module to perform cross-layer fusion based on higher-quality input features, thereby improving the accuracy and efficiency of transmission tower detection.

[0093] In summary, this embodiment acquires satellite remote sensing images containing power transmission towers, inputs these images into a preset power transmission tower detection model, and obtains target detection results. The power transmission tower detection model is trained based on remote sensing image samples with tower annotations and different backgrounds. The model includes a detection head, a pyramid structure, a bidirectional weighted feature fusion module, and a coordinate attention module. The bidirectional weighted feature fusion module fuses high-level and low-level information from the pyramid structure into multi-scale features. These high-level and low-level information are feature information extracted from the satellite remote sensing images. The coordinate attention module weights the multi-scale features based on an attention map to obtain weighted multi-scale features, which the detection head uses to obtain the target detection results. The attention map is an attention map that determines multiple directions based on the multi-scale features.

[0094] To address the challenges of complex backgrounds in satellite remote sensing images, where power transmission towers occupy relatively little pixel area compared to the background, and the decreasing resolution of satellite remote sensing images with increasing CNN model layers in deep learning, leading to blurred details and low detection accuracy, this embodiment improves the accuracy of power transmission tower detection by employing a power transmission tower detection model that includes a bidirectional weighted feature fusion module and a coordinate attention module. Specifically, this embodiment extracts high-level and low-level features from satellite remote sensing images using a pyramid structure. The bidirectional weighted feature fusion module then fuses features from adjacent levels within the feature pyramid, enhancing the interaction between high-level and low-level features, preserving inter-level information, and improving sensitivity to smaller power transmission towers in satellite remote sensing images, thereby increasing the accuracy of power transmission tower detection. Furthermore, this embodiment reweights the multi-scale features based on the attention maps determined by the multi-scale features in multiple directions. By weighting, the channels in important directions can be emphasized, while the responses of channels in unimportant directions can be weakened, thereby improving the spatial selectivity near the target. Spatial selection can reduce false alarms caused by repetitive textures and improve the accuracy of transmission tower detection.

[0095] Based on the first embodiment of this application, in the second embodiment of this application, the content that is the same as or similar to that in Embodiment 1 above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 6 The step of weighting the multi-scale features based on the coordinate attention module to obtain the weighted multi-scale features further includes steps S100~S400:

[0096] Step S100: Perform global average pooling on the multi-scale features in the height and width directions respectively to obtain two direction descriptors;

[0097] It should be noted that the direction descriptor is a one-dimensional vector obtained by performing global average pooling on multi-scale features in the height or width directions respectively. It is used to represent the global context information in that direction. The height direction descriptor reflects the feature response distribution in the vertical dimension, and the width direction descriptor reflects the feature response distribution in the horizontal dimension.

[0098] Understandably, this embodiment performs global average pooling on multi-scale features in both the height and width directions, thereby capturing the overall distribution characteristics of transmission towers in the vertical and horizontal spatial dimensions, forming two complementary direction descriptors. This separate processing preserves direction-specific information, avoiding the defect of traditional channel attention mechanisms where spatial information is completely compressed, resulting in the loss of direction sensitivity. This allows the coordinate attention module to generate attention maps based on more refined spatial structural cues, identifying transmission towers with distinct directional structures, improving the model's ability to perceive target geometric characteristics, and thus enhancing the response intensity of transmission tower-related regions in the weighted multi-scale features while suppressing irrelevant background interference.

[0099] Step S200: The two direction descriptors are concatenated, and the concatenated descriptors are convolved by a preset shared convolution module to obtain a shared embedding;

[0100] It should be noted that the shared convolutional module refers to a lightweight convolutional sub-network within the coordinate attention module used to process the concatenated orientation descriptors. Its parameters are shared in both directions to generate a unified semantic embedding representation. The shared embedding is an intermediate feature vector output by the shared convolutional module after performing a convolution operation on the concatenated orientation descriptors. It integrates contextual information from both the height and width directions and serves as the basis for subsequent branch attention generation.

[0101] Understandably, this embodiment concatenates the height and width descriptors and processes them jointly through a parameter-sharing convolutional module. This allows the spatial context information of the two directions to interact and merge in the same semantic space, avoiding information fragmentation caused by independent modeling. By introducing a shared convolutional module, the number of model parameters is reduced, computational efficiency is improved, and the generated shared embedding retains both direction specificity and consistent semantic expression capabilities.

[0102] Step S300: The shared embedding is split into two shared sub-embedded layers, and based on the shared convolution module and a preset activation function, the shared sub-embedded layers are activated and convolved sequentially to obtain a horizontal attention map and a vertical attention map.

[0103] It should be noted that shared embedding refers to splitting the shared embedding according to the channel dimension or other preset methods to obtain two sub-vectors, which are used to generate attention maps in the horizontal and vertical directions, respectively. The horizontal attention map is a two-dimensional weight map that reflects the importance of multi-scale features at each position in the width (horizontal) direction, and the vertical attention map is a two-dimensional weight map that reflects the importance of multi-scale features at each position in the height (vertical) direction.

[0104] It is understood that this embodiment splits the shared embedding into two shared sub-embedded layers, and generates independent horizontal and vertical attention maps through activation functions and shared convolution modules, respectively. This allows the model to model the spatial distribution characteristics of transmission towers in the horizontal and vertical directions while maintaining parameter efficiency. The introduction of activation functions enhances the nonlinear expressive power of attention weights, making the high-response region more focused on the actual location of the transmission tower.

[0105] Step S400: Based on the horizontal attention map and the vertical attention map, the multi-scale features are weighted to obtain the weighted multi-scale features.

[0106] It should be noted that weighted multi-scale features refer to new feature representations obtained by applying horizontal and vertical attention maps to multi-scale features for spatial weighting adjustments. This feature representation can better highlight the structural information of the target object in a specific direction, thereby improving the model's ability to identify and locate the target. The coordinate attention module structure in this embodiment can be referred to... Figure 7 .

[0107] Understandably, this embodiment utilizes the generated horizontal and vertical attention maps to perform directional decoupling weighting operations on multi-scale features, which can respectively enhance the key response regions of the transmission tower in the width and height directions. The weighting process in this embodiment can be implemented through element-wise multiplication, which is computationally efficient and preserves the channel structure of the original features. The resulting weighted multi-scale features, while maintaining multi-scale semantic information, significantly enhance the representation ability of transmission towers with obvious directional structures, improve the accuracy of subsequent detection head localization and recognition, and exhibit stronger robustness in complex remote sensing backgrounds.

[0108] In one embodiment, after the step of weighting the multi-scale features based on the horizontal attention map and the vertical attention map to obtain the weighted multi-scale features, before performing target detection using the weighted multi-scale features and the detection head, this embodiment further includes the following steps:

[0109] Based on the typical aspect ratio prior of transmission towers in remote sensing images, the horizontal attention map and the vertical attention map are proportionally constrained and corrected to obtain corrected horizontal attention map and corrected vertical attention map. Based on the corrected horizontal attention map and the corrected vertical attention map, the multi-scale features are re-weighted to obtain the final weighted multi-scale features.

[0110] It should be noted that the aspect ratio prior refers to the fact that transmission towers in satellite remote sensing images typically appear as tall, slender structures with a height significantly greater than their width. Therefore, their bounding boxes or principal axis directions have a statistically stable aspect ratio range, which can be used as a geometric constraint to optimize the attention distribution. Proportion constraint correction is a post-processing operation on the attention map based on the aspect ratio prior. By suppressing directional responses that do not conform to typical transmission tower morphology (such as excessively wide or low regions), it makes the attention map more consistent with the geometric characteristics of the real target. The corrected horizontal attention map is obtained after proportion constraint correction, compressed or smoothed in the width direction to match the typical narrow-width characteristics of transmission towers. The corrected vertical attention map is obtained after proportion constraint correction, enhanced or extended in the height direction to match the typical tall characteristics of transmission towers.

[0111] Understandably, this embodiment analyzes the spatial distribution characteristics of the horizontal and vertical attention maps, calculating their effective support width and height. If the horizontal attention response is too wide or the vertical response is too short, exceeding the preset reasonable aspect ratio range (e.g., the width accounts for more than 1 / 3 of the total dimension, or the height is less than twice the width), the attention weight of that region is attenuated or smoothed. Specifically, a horizontal Gaussian window constraint is applied to the horizontal attention map to suppress response peaks with excessive width, and a vertical stretching compensation is applied to the vertical attention map to strengthen continuous high-response regions, thus obtaining the corrected horizontal and vertical attention maps. These two corrected attention maps are used to re-weight the original multi-scale features, generating the final weighted multi-scale features for use by the detection head. This embeds the prior geometric structure unique to the transmission tower into the attention mechanism, making the feature weighting more closely match the actual target shape.

[0112] Since transmission towers have a highly significant and stable aspect ratio in remote sensing images, and the general attention mechanism may produce false responses that do not conform to this geometric law due to background interference, this embodiment introduces aspect ratio priors to perform proportional constraint correction on the horizontal and vertical attention maps. This not only significantly suppresses false detections caused by wide-body interference objects, but also enhances the perception of vertical continuity of partially occluded or low-contrast transmission towers. Thus, without increasing the model parameters, the detection accuracy is improved through priors.

[0113] In one embodiment, after the step of performing proportional constraint correction on the horizontal attention map and the vertical attention map based on the typical aspect ratio prior of the transmission tower in the remote sensing image to obtain the corrected horizontal attention map and the corrected vertical attention map, the method further includes:

[0114] Based on the continuity characteristics of the response intensity in the vertical attention map, potential transmission tower trunk areas are identified; if the horizontal attention map has symmetrically distributed local peaks on both sides of the trunk area, the horizontal attention response of that area is retained; otherwise, the horizontal attention weight of that area is attenuated to obtain the corrected horizontal attention map.

[0115] It should be noted that the continuity characteristic of response intensity refers to the fact that in the vertical attention map, the actual transmission tower trunk typically appears as a continuous, unbroken high-response band along the height direction, rather than isolated or discontinuous peaks. The potential transmission tower trunk region is the longitudinal region determined by the high-response columns in the vertical attention map that meet the minimum length and continuity thresholds, representing the location that may contain the transmission tower body. Symmetrically distributed local peaks are two local maxima appearing at approximately equidistant positions on the left and right sides of the trunk region in the horizontal attention map, corresponding to the typical symmetrical layout of the transmission tower crossarm (the transverse structure used to erect conductors) in the remote sensing image. Horizontal attention weight attenuation reduces the weight values ​​of horizontal attention map regions that do not conform to the symmetry prior by a preset ratio or function to suppress atypical structural responses.

[0116] Understandably, this embodiment analyzes the vertical attention map, extracts high-response columns with strong continuity along the height direction, and marks them as potential transmission tower backbone regions. On the horizontal profile corresponding to this backbone region, it detects whether there are two local peaks in the horizontal attention map that are approximately symmetrical about the backbone center. If they exist, the region is determined to conform to the typical symmetrical structure of the transmission tower crossarm, and its original horizontal attention response is retained. If they do not exist (only one side has a response, no peak, or the peak is asymmetrical), the region is considered more likely to be an interference object, and its horizontal attention weight is attenuated. Finally, the corrected horizontal attention map after verification of this structural consistency is output. Thus, based on the proportional constraint, the crossarm symmetry prior unique to the transmission tower is further introduced to further correct the attention map.

[0117] Because transmission towers in remote sensing images not only have high aspect ratio characteristics but also generally exhibit a structural pattern of continuous main trunk and symmetrical crossarms, single-direction attention is easily interfered with when the tower is partially obscured or at low resolution. This embodiment improves the accuracy of transmission tower detection by jointly analyzing the continuity of the vertical attention map and the symmetry of the horizontal attention map. The main trunk is located through the vertical map, and the crossarms are verified through the horizontal map. This effectively eliminates targets that are easily misjudged, such as isolated utility poles and tilted tower cranes, which have similar aspect ratios but asymmetrical structures.

[0118] In summary, this embodiment performs global average pooling on the multi-scale features in both the height and width directions to obtain two directional descriptors. These two directional descriptors are then concatenated, and a preset shared convolution module is used to convolve the concatenated descriptors to obtain a shared embedding. This shared embedding is then split into two shared sub-embedded features, and based on the shared convolution module and a preset activation function, these sub-embedded features are sequentially activated and convolved to obtain a horizontal attention map and a vertical attention map. Finally, based on the horizontal and vertical attention maps, the multi-scale features are weighted to obtain the weighted multi-scale features.

[0119] This embodiment introduces a lightweight processing flow based on ordinary convolution and depthwise separable convolution units before feature fusion, and enhances feature diversity by channel splicing and channel order shuffling. In the attention mechanism, a directional decoupling approach is adopted to model the global context in the height and width directions respectively. Efficient fusion and weight generation are achieved through shared convolution and parameter sharing strategies, which enables the model to significantly improve the feature representation ability and positioning accuracy of highly directional structural targets such as transmission towers while maintaining low computational overhead.

[0120] Based on the first and second embodiments of this application, in the third embodiment of this application, the content that is the same as or similar to that in embodiments one and two above can be referred to the above description, and will not be repeated hereafter. Based on this, please refer to... Figure 8 The bidirectional weighted feature fusion module includes multiple bidirectional weighted feature fusion units. The adjacent layer is the layer in the pyramid structure that is adjacent to the layer where the current node is located. The step of fusing the initial features of each layer in the pyramid structure with the initial features of the adjacent layers based on the bidirectional weighted feature fusion module to obtain the multi-scale features further includes steps A10 to A40:

[0121] Step A10: For each node in each layer of the pyramid structure, determine the target feature fusion unit corresponding to the current node from the bidirectional weighted feature fusion unit;

[0122] It should be noted that the bidirectional weighted feature fusion unit constitutes the basic processing unit of the bidirectional weighted feature fusion module. It is used to perform weighted fusion of the initial features of the current layer node and its neighboring layer nodes, supporting bidirectional flow of information between higher and lower layers. The current node refers to the feature node being processed in a certain layer of the pyramid structure, which corresponds to an initial feature with a specific spatial resolution and semantic level.

[0123] The target feature fusion unit is a specific fusion unit selected from multiple bidirectional weighted feature fusion units for the current node, used to perform feature fusion operations between the node and its adjacent layers.

[0124] It is understandable that, due to differences in semantic abstraction and spatial resolution among nodes at different levels in a pyramid structure, using a uniform fusion method might lead to a decrease in information transmission efficiency. This embodiment dynamically assigns a corresponding target feature fusion unit to each current node, allowing the fusion operation to adapt to the characteristics of the node's level, thereby improving the alignment accuracy and fusion efficiency of high-level semantic information and low-level detailed information in cross-layer interactions.

[0125] Step A20: Determine whether the current node is the preset target node;

[0126] It should be noted that the target node is a specific node that is pre-specified in the pyramid structure and requires special processing or serves as the basis for the fusion output.

[0127] Understandably, since not all nodes in the pyramid structure contribute equally to the final detection result, by introducing a mechanism to determine whether the current node is the target node, the fusion process can be selectively activated, focusing on the scale level most representative of the transmission tower detection task. This effectively reduces unnecessary feature fusion operations, lowers computational overhead, and improves model inference efficiency while ensuring multi-scale feature representation capabilities.

[0128] Step A30: If the current node is the target node, then determine the adjacent nodes corresponding to the target feature fusion unit from the adjacent layers, and determine the initial node of the layer where the current node is located;

[0129] It should be noted that the initial node refers to the original feature node of the current node in the pyramid structure that has not been merged, and its corresponding feature is the initial feature of that layer.

[0130] It is understandable that, since this embodiment only locates its corresponding neighboring nodes and the initial nodes of this layer after confirming that the current node is the target node, it ensures that feature fusion is carried out only between key levels and between nodes with clear structural correspondence, avoiding semantic confusion or loss of details caused by misalignment of levels or participation of irrelevant nodes in fusion, and improving the spatial alignment accuracy and semantic consistency of multi-scale features in cross-layer transmission.

[0131] Step A40: Based on preset learnable weights and preset coefficients, the first feature corresponding to the initial node, the second feature corresponding to the current node, and the third feature corresponding to the adjacent node are fused to obtain the multi-scale feature.

[0132] It should be noted that the learnable weights are parameters that are automatically optimized through backpropagation during model training, used to dynamically adjust the contribution ratio of different features in the fusion process. The preset coefficients are used in multi-scale fusion calculations to prevent division by zero during multi-scale fusion. The first feature is the feature representation obtained by pre-processing the initial features corresponding to the initial node. The second feature is the feature representation of the current node itself after feature processing, reflecting the current state of the node in the pyramid structure. The third feature is the feature representation output by adjacent nodes after corresponding processing, used to inject cross-layer contextual information into the current node.

[0133] In one feasible implementation, the learning weights include a first weight, a second weight, and a third weight. The specific implementation of fusing the first feature corresponding to the initial node, the second feature corresponding to the current node, and the third feature corresponding to the neighboring nodes based on preset learnable weights and preset coefficients to obtain the multi-scale features can also be:

[0134] The product of the first weight and the first feature, the product of the second weight and the second feature, and the product of the third weight and the third feature are added together to obtain the initial weighted feature. The first weight, the second weight, the third weight, and the preset coefficient are added together to obtain the feature coefficient sum. The quotient of the weighted feature and the feature coefficient sum is calculated to obtain the target weighted feature. The target weighted feature is convolved to obtain the multi-scale feature.

[0135] It should be noted that the first, second, and third features are learnable parameters used for the features corresponding to the initial weighted nodes. The initial weighted features are intermediate fusion results obtained by multiplying the first weight by the first feature, the second weight by the second feature, and the third weight by the third feature, respectively, and then summing the results. The sum of feature coefficients is the sum of the first weight, the second weight, the third weight, and a preset coefficient, used to normalize the initial weighted features. The target weighted features are the normalized fusion features obtained by dividing the initial weighted features by the sum of feature coefficients, used to eliminate the influence of weight scale. In this embodiment, convolution refers to applying a small 1×1 convolution operation to the target weighted features for channel adjustment or nonlinear enhancement, outputting the final multi-scale features.

[0136] For a comparison of the processing logic of the Bidirectional Weighted Feature Fusion (BiFPN) module in this embodiment with other methods, please refer to [link / reference]. Figure 9 , Figure 9 This is a comparison of the bidirectional weighted feature fusion module and path aggregation network (PAN) of this application, as well as the feature pyramid network (FPN).

[0137] In the picture to Indicates different number of floors, Indicates the first Layer input, The first in the top-down propagation path The middle node of the layer, The first in the top-down propagation path The middle node of the layer, For the first The output of the layer, For the first The output of the layer, , and It's the weight.

[0138] The formula for calculating nodes is as follows:

[0139]

[0140] in, It is a feature of the i-th node. It is the j-th weight. The preset coefficients are the preset coefficients.

[0141] It is understood that this embodiment explicitly calculates the weighted contributions of the first, second, and third weights to the first, second, and third features, respectively, and introduces preset coefficients to participate in the normalization denominator. By dividing the initial weighted features by the sum of the feature coefficients, the feature scale drift problem caused by excessively large or small weight amplitudes is effectively suppressed. Finally, the feature representation is further optimized through convolution operations, which enhances the numerical stability and semantic consistency of the generated multi-scale features, thereby improving the generalization ability and detection accuracy of the transmission tower detection model in complex remote sensing scenarios.

[0142] In summary, in this embodiment, for each node in each layer of the pyramid structure, the target feature fusion unit corresponding to the current node is determined from the bidirectional weighted feature fusion unit. It is then determined whether the current node is a preset target node. If the current node is the target node, the adjacent nodes corresponding to the target feature fusion unit are determined from the adjacent layers, and the initial node of the layer where the current node is located is determined. Based on preset learnable weights and preset coefficients, the first feature corresponding to the initial node, the second feature corresponding to the current node, and the third feature corresponding to the adjacent node are fused to obtain the multi-scale feature.

[0143] This embodiment dynamically adjusts the contribution ratio of the initial information, current state, and context of adjacent layers by learning weights, and combines preset coefficients to achieve normalization constraints, effectively balancing adaptability and training stability. While ensuring the semantic richness and spatial accuracy of multi-scale features, it significantly improves the model's ability to represent slender, highly directional targets such as power transmission towers, and also takes into account computational efficiency, enabling the entire detection system to have both high accuracy and strong robustness in complex remote sensing contexts.

[0144] It should be noted that the above examples are only for understanding this application and do not constitute a limitation on the remote sensing target detection method for transmission towers in this application. Any simple modifications based on this technical concept are within the protection scope of this application.

[0145] This application also provides a remote sensing target detection device for transmission towers; please refer to... Figure 10 The remote sensing target detection device for power transmission towers includes:

[0146] Furthermore, to achieve the above objectives, this application also proposes a remote sensing target detection device for transmission towers, the remote sensing target detection device for transmission towers comprising:

[0147] Image acquisition module 10 is used to acquire satellite remote sensing images containing power transmission towers;

[0148] The target detection module 20 is used to input the satellite remote sensing image into a preset transmission tower detection model to obtain the target detection result. The transmission tower detection model is trained based on remote sensing image samples with transmission tower annotations and different backgrounds. The transmission tower detection model includes a detection head, a pyramid structure, a bidirectional weighted feature fusion module, and a coordinate attention module. The bidirectional weighted feature fusion module is used to fuse high-level and low-level information in the pyramid structure into multi-scale features. The high-level and low-level information are feature information extracted from the satellite remote sensing image. The coordinate attention module is used to weight the multi-scale features based on the attention map to obtain weighted multi-scale features, so that the detection head can obtain the target detection result based on the multi-scale features. The attention map is an attention map that determines multiple directions based on the multi-scale features.

[0149] In one embodiment, the remote sensing target detection device for transmission towers further includes:

[0150] The sample acquisition module is used to acquire remote sensing image samples with transmission tower labels and different backgrounds;

[0151] The feature extraction module is used to input the remote sensing image samples into a preset training model. Based on the efficient feature extraction module in the training model, the remote sensing image samples are extracted with different feature extraction depths to obtain the initial features corresponding to each node in the pyramid structure.

[0152] The feature fusion module is used to fuse the initial features of each layer in the pyramid structure with the initial features of the adjacent layers based on the bidirectional weighted feature fusion module to obtain the multi-scale features;

[0153] The feature weighting module is used to weight the multi-scale features based on the coordinate attention module to obtain the weighted multi-scale features;

[0154] The transmission tower detection module is used to detect transmission towers in the remote sensing image samples based on the detection head and the weighted multi-scale features, obtain detection results, and adjust the training model based on the detection results and the transmission tower annotations to obtain the transmission tower detection model.

[0155] In one embodiment, the feature weighting module further includes:

[0156] The pooling submodule is used to perform global average pooling on the multi-scale features in the height and width directions, respectively, to obtain two directional descriptors.

[0157] The embedding submodule is used to concatenate the two direction descriptors and perform convolution on the concatenated descriptors through a preset shared convolution module to obtain a shared embedding.

[0158] The attention map generation submodule is used to split the shared embedding into two shared sub-embedded modules, and based on the shared convolution module and a preset activation function, to sequentially activate and convolve the shared sub-embedded modules to obtain a horizontal attention map and a vertical attention map.

[0159] The feature weighting submodule is used to weight the multi-scale features based on the horizontal attention map and the vertical attention map to obtain the weighted multi-scale features.

[0160] In one embodiment, the feature fusion module further includes:

[0161] The fusion unit determination submodule is used to determine the target feature fusion unit corresponding to the current node from the bidirectional weighted feature fusion unit for each node in each layer of the pyramid structure.

[0162] The node determination submodule is used to determine whether the current node is a preset target node;

[0163] The node determination submodule is used to determine the adjacent nodes corresponding to the target feature fusion unit from the adjacent layers if the current node is the target node, and to determine the initial node of the layer where the current node is located;

[0164] The feature fusion submodule is used to fuse the first feature corresponding to the initial node, the second feature corresponding to the current node, and the third feature corresponding to the adjacent node based on preset learnable weights and preset coefficients to obtain the multi-scale feature.

[0165] In one embodiment, the feature fusion submodule further includes:

[0166] The feature weighting unit is used to add the product of the first weight and the first feature, the product of the second weight and the second feature, and the product of the third weight and the third feature to obtain the initial weighted feature;

[0167] The coefficient summation unit is used to add the first weight, the second weight, the third weight, and the preset coefficient to obtain the sum of characteristic coefficients;

[0168] The weighted feature calculation unit is used to calculate the quotient of the weighted feature and the sum of the feature coefficients to obtain the target weighted feature;

[0169] A multi-scale feature determination unit is used to convolve the target weighted features to obtain the multi-scale features.

[0170] In one embodiment, the remote sensing target detection device for transmission towers further includes:

[0171] The first feature determination module is used to perform convolution on the initial features to obtain a first feature with a corresponding number of channels;

[0172] The second feature determination module is used to convolve the first feature based on the depth-separable convolutional unit to obtain a second feature with the same number of channels as the first feature.

[0173] The feature splicing module is used to splice the first feature and the second feature along the channel dimension to obtain spliced ​​features;

[0174] The channel shuffling module is used to shuffle the order of the channels in the splicing feature to obtain the processed initial feature.

[0175] The transmission tower remote sensing target detection device provided in this application, employing the transmission tower remote sensing target detection method in the above embodiments, can solve the technical problem of low detection accuracy of transmission towers. Compared with the prior art, the beneficial effects of the transmission tower remote sensing target detection device provided in this application are the same as those of the transmission tower remote sensing target detection method provided in the above embodiments, and other technical features in the transmission tower remote sensing target detection device are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0176] This application provides a remote sensing target detection device for power transmission towers, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the remote sensing target detection method for power transmission towers in Embodiment 1 described above.

[0177] The following is for reference. Figure 11The diagram illustrates a structural schematic suitable for implementing the remote sensing target detection device for transmission towers in the embodiments of this application. The remote sensing target detection device for transmission towers in the embodiments of this application may include, but is not limited to, mobile terminals such as mobile phones, tablets, laptops, digital broadcast receivers, PDAs (Personal Digital Assistants), PMPs (Portable Media Players), vehicle-mounted terminals (e.g., vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. Figure 11 The remote sensing target detection device for transmission towers shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0178] like Figure 11 As shown, the remote sensing target detection device for power transmission towers may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the remote sensing target detection device for power transmission towers. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows the transmission tower remote sensing target detection equipment to exchange data wirelessly or via wired communication with other devices. Although the figure shows transmission tower remote sensing target detection equipment with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0179] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0180] The remote sensing target detection device for power transmission towers provided in this application, employing the remote sensing target detection method for power transmission towers in the above embodiments, can solve the technical problem of low detection accuracy of power transmission towers. Compared with the prior art, the beneficial effects of the remote sensing target detection device for power transmission towers provided in this application are the same as those of the remote sensing target detection method for power transmission towers provided in the above embodiments, and other technical features of this remote sensing target detection device are the same as those disclosed in the method of the previous embodiment, and will not be repeated here.

[0181] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0182] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0183] This application provides a computer-readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the transmission tower remote sensing target detection method in the above embodiments.

[0184] The computer-readable storage medium provided in this application may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0185] The aforementioned computer-readable storage medium may be included in the remote sensing target detection equipment for transmission towers; or it may exist independently and not be assembled into the remote sensing target detection equipment for transmission towers.

[0186] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the remote sensing target detection device for power transmission towers, cause the remote sensing target detection device for power transmission towers to perform the aforementioned remote sensing target detection method for power transmission towers.

[0187] Computer program code for performing the operations of this application can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0188] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0189] The modules described in the embodiments of this application can be implemented in software or hardware. The names of the modules do not necessarily limit the functionality of the unit itself.

[0190] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described remote sensing target detection method for power transmission towers, thereby solving the technical problem of low detection accuracy of power transmission towers. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the remote sensing target detection method for power transmission towers provided in the above embodiments, and will not be repeated here.

[0191] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method for detecting remote sensing targets of transmission towers.

[0192] The computer program product provided in this application can solve the technical problem of low detection accuracy of transmission towers. Compared with the prior art, the beneficial effects of the computer program product provided in this application are the same as those of the remote sensing target detection method for transmission towers provided in the above embodiments, and will not be repeated here.

[0193] All user-related data involved in this application was obtained with the user's permission or consent, as per [reference]. Figure 12 In other words, when this application is applied to a specific product or technology, user permission is required to acquire and process the relevant data, and the processing of the relevant data must comply with the relevant laws, regulations and regulatory standards of the relevant countries and regions.

[0194] The above description is only a part of the embodiments of this application and does not limit the scope of protection of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A method for remote sensing target detection of transmission towers, characterized in that, The method includes: Acquire satellite remote sensing images containing power transmission towers; Satellite remote sensing images are input into a preset power transmission tower detection model to obtain target detection results. The power transmission tower detection model is trained based on remote sensing image samples with power transmission tower annotations and different backgrounds. The power transmission tower detection model includes a detection head, a pyramid structure, a bidirectional weighted feature fusion module, and a coordinate attention module. The bidirectional weighted feature fusion module is used to fuse high-level and low-level information in the pyramid structure into multi-scale features. The high-level and low-level information are feature information extracted from the satellite remote sensing images. The coordinate attention module is used to weight the multi-scale features based on the attention map to obtain weighted multi-scale features, which are then used by the detection head to obtain the target detection results. The attention map is an attention map that determines multiple directions based on the multi-scale features. The step of weighting the multi-scale features based on the attention map to obtain weighted multi-scale features includes: Global average pooling is performed on the multi-scale features in both the height and width directions to obtain two directional descriptors. The two direction descriptors are concatenated, and the concatenated descriptors are convolved by a preset shared convolution module to obtain a shared embedding. The shared embedding is split into two shared sub-embedded layers, and based on the shared convolution module and a preset activation function, the shared sub-embedded layers are activated and convolved sequentially to obtain a horizontal attention map and a vertical attention map. Based on the horizontal attention map and the vertical attention map, the multi-scale features are weighted to obtain the weighted multi-scale features.

2. The method as described in claim 1, characterized in that, Before the step of inputting the satellite remote sensing image into a preset transmission tower detection model to obtain the target detection result, the method further includes: Acquire the remote sensing image samples with transmission tower labels and different backgrounds; The remote sensing image samples are input into a preset training model. Based on the efficient feature extraction module in the training model, features are extracted from the remote sensing image samples through different feature extraction depths to obtain the initial features corresponding to each node in the pyramid structure. Based on the bidirectional weighted feature fusion module, the initial features of each layer in the pyramid structure are fused with the initial features of the adjacent layers to obtain the multi-scale features; The weighted multi-scale features are obtained by weighting the multi-scale features based on the coordinate attention module. Based on the detection head and the weighted multi-scale features, the power transmission towers in the remote sensing image samples are detected to obtain the detection results. Based on the detection results and the power transmission tower annotations, the model to be trained is adjusted to obtain the power transmission tower detection model.

3. The method as described in claim 2, characterized in that, The bidirectional weighted feature fusion module includes multiple bidirectional weighted feature fusion units. The adjacent layer is the layer in the pyramid structure that is adjacent to the layer where the current node is located. The step of fusing the initial features of each layer in the pyramid structure with the initial features of the adjacent layers based on the bidirectional weighted feature fusion module to obtain the multi-scale features includes: For each node in each layer of the pyramid structure, the target feature fusion unit corresponding to the current node is determined from the bidirectional weighted feature fusion unit; Determine whether the current node is the preset target node; If the current node is the target node, then the adjacent nodes corresponding to the target feature fusion unit are determined from the adjacent layers, and the initial node of the layer where the current node is located is determined; Based on preset learnable weights and preset coefficients, the first feature corresponding to the initial node, the second feature corresponding to the current node, and the third feature corresponding to the adjacent node are fused to obtain the multi-scale feature.

4. The method as described in claim 3, characterized in that, The learning weights include a first weight, a second weight, and a third weight. The step of fusing the first feature corresponding to the initial node, the second feature corresponding to the current node, and the third feature corresponding to the neighboring nodes based on preset learnable weights and preset coefficients to obtain the multi-scale features includes: The initial weighted features are obtained by adding the product of the first weight and the first feature, the product of the second weight and the second feature, and the product of the third weight and the third feature. The first weight, the second weight, the third weight, and the preset coefficient are added together to obtain the sum of the characteristic coefficients; The target weighted feature is obtained by calculating the quotient of the weighted feature and the sum of the feature coefficients. The target weighted features are convolved to obtain the multi-scale features.

5. The method as described in claim 2, characterized in that, The transmission tower detection model includes a lightweight convolution module, which includes depthwise separable convolution units. Before the step of fusing the initial features of each layer in the pyramid structure with the initial features of adjacent layers based on the bidirectional weighted feature fusion module, the model further includes: The initial features are convolved to obtain a first feature with a corresponding number of channels; Based on the depthwise separable convolutional unit, the first feature is convolved to obtain a second feature with the same number of channels as the first feature; Based on the first feature and the second feature, the spliced ​​features are obtained by concatenating them along the channel dimension; The order of the channels in the splicing feature is shuffled to obtain the processed initial feature.

6. A remote sensing target detection device for transmission towers, characterized in that, The device includes: The image acquisition module is used to acquire satellite remote sensing images containing power transmission towers; The target detection module is used to input the satellite remote sensing image into a preset transmission tower detection model to obtain the target detection result. The transmission tower detection model is trained based on remote sensing image samples with transmission tower annotations and different backgrounds. The transmission tower detection model includes a detection head, a pyramid structure, a bidirectional weighted feature fusion module, and a coordinate attention module. The bidirectional weighted feature fusion module is used to fuse high-level and low-level information in the pyramid structure into multi-scale features. The high-level and low-level information are feature information extracted from the satellite remote sensing image. The coordinate attention module is used to weight the multi-scale features based on the attention map to obtain weighted multi-scale features, so that the detection head can obtain the target detection result based on the multi-scale features. The attention map is an attention map that determines multiple directions based on the multi-scale features. The target detection module is further configured to perform global average pooling on the multi-scale features in the height and width directions of the multi-scale features respectively to obtain two direction descriptors. The two direction descriptors are concatenated, and the concatenated descriptors are convolved by a preset shared convolution module to obtain a shared embedding. The shared embedding is split into two shared sub-embedded layers, and based on the shared convolution module and a preset activation function, the shared sub-embedded layers are activated and convolved sequentially to obtain a horizontal attention map and a vertical attention map. Based on the horizontal attention map and the vertical attention map, the multi-scale features are weighted to obtain the weighted multi-scale features.

7. A remote sensing target detection device for power transmission towers, characterized in that, The device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the remote sensing target detection method for transmission towers as described in any one of claims 1 to 5.

8. A storage medium, characterized in that, The storage medium is a computer-readable storage medium, and a computer program is stored on the storage medium. When the computer program is executed by a processor, it implements the steps of the remote sensing target detection method for transmission towers as described in any one of claims 1 to 5.

9. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the remote sensing target detection method for transmission towers as described in any one of claims 1 to 5.

Images