A method, device and equipment for detecting a crossbeam of a power transmission tower

Abstract

The application discloses a power transmission tower beam detection method, device and equipment, the method comprises the following steps: preprocessing the obtained power transmission tower inspection image to obtain a target beam image; based on a three-branch detection head and a score fusion mechanism, a model is constructed and trained according to a direction strip convolution module, a node alignment sampling module and a continuity constraint module, an optimized detection model is generated, the continuity constraint module includes a hollow convolution and a residual gate unit; the target beam image is input into the optimized detection model for beam detection, and a beam detection result is obtained, the beam detection result includes a beam prediction probability. Therefore, the application can solve the technical problems that the prior art lacks the perception ability of the beam structure essence, is easy to cause misjudgment, has insufficient multi-scale adaptability, and has more similar interference, so that the detection lacks precision.

Description

Technical Field

[0001] This application relates to the field of transmission tower inspection technology, and in particular to a method, device and equipment for detecting the crossbeams of transmission towers. Background Technology

[0002] With the continuous expansion of power transmission lines, autonomous drone inspection has become a core method for the operation and maintenance of transmission towers. When drones operate near towers, they need to accurately identify tower components to complete tasks such as path planning, obstacle avoidance, precise landing, and inspection of bolt and pin defects. Crossbeams, as stable and structurally significant components of transmission towers, are key visual references for drone spatial positioning and operational navigation. High-precision automatic crossbeam detection is an important technological foundation for improving the automation level of drone inspections, reducing manual interpretation costs, and ensuring safety during near-tower operations.

[0003] Current beam detection technologies are mainly divided into two categories. One category is traditional digital image processing methods, which extract beam structures based on manually designed features such as Canny edge detection and Hough line transform. The principle is simple, but it relies on manual parameter tuning and has poor robustness in complex tower scenarios. The other category is deep learning object detection methods, represented by Faster R-CNN, SSD, and the YOLO series. Among them, YOLOv12 has become the mainstream application solution due to its balanced accuracy and speed. Existing improvements are mostly focused on general feature extraction and multi-scale fusion optimization, and have not yet fully combined the unique structural geometric characteristics of transmission tower beams.

[0004] Existing detection solutions still have core shortcomings that make them unsuitable for power transmission tower scenarios. First, they only learn the appearance and texture features of the beams and lack the ability to perceive their structural essence, resulting in a high misjudgment rate when there is partial occlusion or the texture is similar to that of the diagonal members. Second, the discrete feature pyramid cannot cover the dramatic scale changes in the near and far views captured by drones, resulting in the loss of small-scale details and insufficient large-scale global information. Third, similar components such as the diagonal members and auxiliary materials of the tower body cause strong interference, and there is a lack of targeted differentiation mechanisms, leading to frequent false detections in densely packed pole areas, which cannot meet the real-time inference requirements of drones. Summary of the Invention

[0005] This application provides a method, apparatus, and equipment for detecting crossbeams of transmission towers, which addresses the technical problems of existing technologies lacking the ability to perceive the essential structure of crossbeams, which easily leads to misjudgments, coupled with insufficient multi-scale adaptability and numerous similar interferences, resulting in a lack of accuracy in detection.

[0006] In view of this, the first aspect of this application provides a method for detecting the crossbeams of transmission towers, comprising: The acquired transmission tower inspection images are preprocessed to obtain the target beam image; The model is constructed and trained based on the three-branch detection head and score fusion mechanism according to the directional strip convolution module, the node alignment sampling module and the continuity constraint module to generate an optimized detection model. The continuity constraint module includes dilated convolution and residual gating unit. The target beam image is input into the optimized detection model for beam detection, and the beam detection result is obtained, which includes the beam prediction probability.

[0007] Preferably, the preprocessing of the acquired transmission tower inspection images to obtain the target beam image includes: Images of power transmission towers are acquired using pre-set image acquisition equipment, which includes a high-definition camera, an inspection robot, a vehicle-mounted inspection device, and a multispectral imaging device. The transmission tower inspection image is cropped and scaled to obtain the initial beam image; The initial beam image is filtered and denoised using preset filtering techniques to obtain a denoised beam image. The preset filtering techniques include mean filtering, Gaussian filtering, and bilateral filtering. The denoised beam image is enhanced by a preset enhancement technique, which includes histogram equalization, linear grayscale transformation and Gamma correction. The enhanced beam image is intelligently optimized using a preset optimization model to generate the target beam image.

[0008] Preferably, the step of building and training a model based on a three-branch detection head and score fusion mechanism using a directional strip convolution module, a node alignment sampling module, and a continuity constraint module to generate an optimized detection model further includes: Based on the cosine annealing restart strategy, an adaptive learning rate is configured for each of the three branch detection heads; Configure the comprehensive loss function based on the preset bounding box loss function and the preset orientation matching loss function; Model training is achieved by combining the SGD optimizer, the adaptive learning rate, and the comprehensive loss function.

[0009] Preferably, the step of building and training a model based on a three-branch detection head and score fusion mechanism using a directional strip convolution module, a node alignment sampling module, and a continuity constraint module to generate an optimized detection model further includes: The optimized detection model is compressed using INT8 quantization technology. The optimized detection model is pruned using a pre-defined structured channel pruning scheme.

[0010] Preferably, the step of inputting the target beam image into the optimized detection model for beam detection to obtain a beam detection result, wherein the beam detection result includes a beam prediction probability, characterized in that it includes: The target beam image is input into the optimized detection model for multi-scale feature extraction to obtain an initial feature map; The directional strip convolution module in the model is used to perform lateral length continuous feature enhancement and vertical thickness and edge local feature enhancement on the initial feature map based on separable convolution to obtain a directional enhanced feature map; Based on the enhanced directional feature map, the orientation of the structural centerline of the candidate beam and its corresponding two endpoints are estimated. The node alignment sampling module performs adaptive sampling and feature aggregation analysis based on the direction of the structural centerline and the two end points along the extension direction of the beam to determine the sampling feature map. The dilated convolution in the continuity constraint module is used to expand the receptive field of the sampled feature map. At the same time, the residual gating unit is combined to enhance and correct the continuous edge response in the beam length direction to generate a shared feature map. The shared feature map is input into the three-branch detection head for detection and analysis to obtain the three-branch detection result, which includes visual detection information, structural consistency score, and non-beam suppression penalty term. Based on the scoring fusion mechanism, a weighted score is calculated according to the detection results of the three branches, and the beam detection result is generated.

[0011] The second aspect of this application provides a transmission tower beam detection device, comprising: The image acquisition unit is used to preprocess the acquired transmission tower inspection images to obtain the target beam image; The model generation unit is used to construct and train a model based on the three-branch detection head and score fusion mechanism according to the directional strip convolution module, the node alignment sampling module and the continuity constraint module, and generate an optimized detection model. The continuity constraint module includes dilated convolution and residual gating units. A beam detection unit is used to input the target beam image into the optimized detection model for beam detection and obtain beam detection results, including beam prediction probabilities.

[0012] Preferably, the image acquisition unit is specifically used for: Images of power transmission towers are acquired using pre-set image acquisition equipment, which includes a high-definition camera, an inspection robot, a vehicle-mounted inspection device, and a multispectral imaging device. The transmission tower inspection image is cropped and scaled to obtain the initial beam image; The initial beam image is filtered and denoised using preset filtering techniques to obtain a denoised beam image. The preset filtering techniques include mean filtering, Gaussian filtering, and bilateral filtering. The denoised beam image is enhanced by a preset enhancement technique, which includes histogram equalization, linear grayscale transformation and Gamma correction. The enhanced beam image is intelligently optimized using a preset optimization model to generate the target beam image.

[0013] Preferably, it further includes an optimization configuration unit, used for: Based on the cosine annealing restart strategy, an adaptive learning rate is configured for each of the three branch detection heads; Configure the comprehensive loss function based on the preset bounding box loss function and the preset orientation matching loss function; Model training is achieved by combining the SGD optimizer, the adaptive learning rate, and the comprehensive loss function.

[0014] Preferably, the beam detection unit is specifically used for: The target beam image is input into the optimized detection model for multi-scale feature extraction to obtain an initial feature map; The directional strip convolution module in the model is used to perform lateral length continuous feature enhancement and vertical thickness and edge local feature enhancement on the initial feature map based on separable convolution to obtain a directional enhanced feature map; Based on the enhanced directional feature map, the orientation of the structural centerline of the candidate beam and its corresponding two endpoints are estimated. The node alignment sampling module performs adaptive sampling and feature aggregation analysis based on the direction of the structural centerline and the two end points along the extension direction of the beam to determine the sampling feature map. The dilated convolution in the continuity constraint module is used to expand the receptive field of the sampled feature map. At the same time, the residual gating unit is combined to enhance and correct the continuous edge response in the beam length direction to generate a shared feature map. The shared feature map is input into the three-branch detection head for detection and analysis to obtain the three-branch detection result, which includes visual detection information, structural consistency score, and non-beam suppression penalty term. Based on the scoring fusion mechanism, a weighted score is calculated according to the detection results of the three branches, and the beam detection result is generated.

[0015] A third aspect of this application provides a transmission tower beam detection device, the device including a processor and a memory; The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the transmission tower beam detection method described in the first aspect according to the instructions in the program code.

[0016] As can be seen from the above technical solutions, the embodiments of this application have the following advantages: This application provides a method for detecting crossbeams on power transmission towers, comprising: preprocessing acquired images of power transmission tower inspections to obtain target crossbeam images; constructing and training a model based on a three-branch detection head and scoring fusion mechanism using a directional strip convolution module, a node alignment sampling module, and a continuity constraint module to generate an optimized detection model, wherein the continuity constraint module includes dilated convolution and residual gating units; inputting the target crossbeam image into the optimized detection model for crossbeam detection to obtain crossbeam detection results, wherein the crossbeam detection results include crossbeam prediction probabilities.

[0017] The transmission tower beam detection method provided in this application designs an optimized detection model for targeted detection of transmission tower beams. The directional strip convolution module can extract the structural responses in the length and thickness directions of the beam, enhancing the continuous feature representation of the beam's slender structure. Combined with a continuity constraint module containing dilated convolution and residual gating units, it can enhance long-distance continuous edge responses and suppress fracture pseudo-features, enabling the model to accurately capture the slender and continuous structural essence of the beam, rather than relying on appearance, thus improving recognition stability in occluded scenarios. Furthermore, the node alignment sampling module performs adaptive sampling along the beam structure, avoiding the problem of feature dimensionality reduction disrupting structural continuity caused by existing fixed-grid downsampling. This mechanism can compress feature dimensions while preserving complete beam structural information, balancing the capture of small-scale details and large-scale global structures. In addition, the use of a three-branch detection head and a scoring fusion mechanism allows for comprehensive confidence analysis from multiple perspectives, reducing the impact of similar interference and ensuring the accuracy of the detection results, thereby adapting to the detection requirements of UAVs. Therefore, this application can solve the technical problems of existing technologies that lack the ability to perceive the essence of the beam structure, which easily leads to misjudgment, coupled with insufficient multi-scale adaptability and a large number of similar interferences, resulting in a lack of accuracy in detection. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a method for detecting the crossbeams of a transmission tower provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of a transmission tower beam detection device provided in an embodiment of this application; Figure 3 A schematic diagram illustrating the beam detection and analysis process using the optimized detection model provided in this embodiment of the application. Figure 4 Example of beam inspection results provided in this application embodiment Figure 1 ; Figure 5 Example of beam inspection results provided in this application embodiment Figure 2 ; Figure 6 Example of beam inspection results provided in this application embodiment Figure 3 ; Figure 7 Example of beam inspection results provided in this application embodiment Figure 4 ; Figure 8 Example of beam inspection results provided in this application embodiment Figure 5 . Detailed Implementation

[0019] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0020] For easier understanding, please refer to Figure 1 An embodiment of a method for detecting the crossbeams of a transmission tower provided in this application includes: Step 101: Preprocess the acquired transmission tower inspection images to obtain the target beam image; Step 102: Based on the directional strip convolution module, node alignment sampling module and continuity constraint module, build and train a model based on the three-branch detection head and score fusion mechanism to generate an optimized detection model. The continuity constraint module includes dilated convolution and residual gating units. Step 103: Input the target beam image into the optimized detection model to detect the beam and obtain the beam detection result, which includes the beam prediction probability.

[0021] It should be noted that the images of power transmission tower inspections contain information about multiple beam structures and other impurities. Therefore, basic cropping and scaling can be performed on the acquired images to generate initial beam images for subsequent beam detection. In addition to high-definition cameras, inspection robots, vehicle-mounted inspection equipment, and multispectral imaging equipment, other pre-set image acquisition devices can be selected. For example, handheld inspection devices can be used for manual acquisition, drone aerial photography can be used for mobile devices, and spectral imaging equipment can include infrared thermal imaging equipment, ultraviolet imaging equipment, and polarization imaging equipment, etc. Specific options are not limited here.

[0022] To reduce interference from redundant background information in the image from the source, this embodiment performs a series of preprocessing steps on the acquired initial beam image, including denoising, enhancement, correction, and intelligent optimization based on filtering. This not only minimizes noise interference but also enhances the visibility features of the beam target, providing reliable image source data for subsequent target detection.

[0023] In addition, image normalization, pixel value standardization, and basic image quality correction can be performed to adjust the image to the fixed resolution and numerical distribution required for model input, resulting in a target measurement image that can be directly input into the model for detection. Specifically, size normalization can use methods such as proportional scaling with padding, direct resizing, and multi-scale random cropping; pixel standardization can employ strategies such as ImageNet mean-variance normalization and sample-by-sample zero-mean normalization; image quality correction can use basic algorithms such as histogram equalization, Gamma brightness correction, and Gaussian denoising, flexibly selected according to the lighting and noise levels of the inspected image.

[0024] It is understandable that the initial beam image is generally an outdoor image, so it contains various types of noise such as salt and pepper noise, Gaussian noise, cloud noise, and wind and sand interference. In this case, to improve the image purity, two or more noise filtering techniques can be combined to eliminate impurities at different levels and improve image quality. Therefore, the preset filtering techniques in this embodiment include, but are not limited to, mean filtering, Gaussian filtering, and bilateral filtering, or a combination of the three. Other filtering techniques, such as median filtering, can also be added, which are not limited here.

[0025] Since the initial beam image in this embodiment is used for transmission tower inspection tasks, it is necessary to detect specific beam targets. Therefore, enhancement methods are needed to address brightness issues such as excessive darkness, underexposure, and low contrast caused by backlighting, shadows, cloudy days, and low illumination at night during inspections. Thus, this embodiment selects different preset enhancement techniques. Histogram equalization can automatically stretch the grayscale range of the image and optimize the sense of layering in beam images taken on cloudy days or against the light. Linear grayscale transformation can adjust the overall image brightness and contrast, improving the clarity of beam images in dim scenes. Gamma correction can correct image brightness and darkness deviations, optimizing the display effect of details in strong light shadows and low light environments.

[0026] This embodiment utilizes a pre-defined optimization model for intelligent optimization, employing artificial intelligence algorithms to achieve high-precision, intelligent image optimization. This ensures that target detection can be adapted to repair low-quality images in various complex and challenging scenarios. The architecture of the pre-defined optimization model can be designed and selected according to actual conditions, and image intelligent optimization training can be completed for use in real-world scenarios. The model can specifically achieve image super-resolution reconstruction, AI image deblurring, background blurring and segmentation, and defect feature enhancement. It can not only comprehensively improve image clarity but also magnify and repair the detailed features of the beam structure, highlighting the presence of the detected target.

[0027] It is understood that this embodiment can sequentially perform the above-mentioned different preprocessing optimizations on the initial beam image. Furthermore, in addition to the preprocessing operations proposed in this embodiment, other related techniques can be added to improve image optimization. For example, image sharpening enhancement processing can employ techniques such as Sobel sharpening, Laplacian sharpening, Canny edge detection enhancement, and high-frequency component enhancement to highlight image details, textures, and the edge features of the beam outline, while preserving high-frequency detail features and weakening low-frequency background interference, thereby comprehensively improving the clarity of the beam's details. Since the image needs to be input into the model for target detection, necessary grayscale and binarization processing can also be performed on the obtained target beam image. This is not specifically limited here and can be designed according to the actual situation.

[0028] It should be noted that the optimized detection model in this embodiment uses YOLOv12 as the base network to generate candidate boxes and performs a two-stage architecture of three-branch secondary verification for structural semantic analysis. This model performs thorough detection and analysis on the beam image through four core modules. The directional strip convolution module can enhance the continuous response in the beam length direction and the edge features in the thickness direction by combining strip convolutions. The node alignment sampling module can perform adaptive sampling along the beam length direction, replacing fixed grid downsampling, which can avoid destroying the continuity of the beam structure. The continuity constraint module can enhance the long-distance continuous edge response and suppress the pseudo response caused by fracture and intersection structures through dilated convolution and residual gating units. The three-branch detection head and scoring fusion mechanism can output specific parameters of vision, structure and non-beam, and filter out beam targets by calculating the fused score.

[0029] The optimized detection model in this embodiment can retain the continuous structural features of the beam and ensure good detection stability even in scenarios with partial occlusion and changes in scale. Moreover, it is modularly designed based on the YOLOv12 framework, with controllable computational complexity, compatibility with existing inspection equipment, and real-time deployment on the UAV side, thus meeting inspection requirements.

[0030] The model building process includes core network construction and parameter fitting. In fact, it integrates three major feature enhancement modules—directional strip convolution, node alignment sampling, and continuity constraint—on the basic detection network, and configures a three-branch detection head and a scoring fusion mechanism. The optimized detection model is generated by training with labeled data. The continuity constraint module consists of dilated convolution and residual gating units, which are used to enhance the long-distance continuous structural features of the beam.

[0031] In practice, directional strip convolution can be implemented using variations such as depthwise separable strip convolution and standard strip convolution; node alignment sampling can employ strategies such as equidistant sampling along the centerline and dense sampling at endpoints; in continuity constraints, dilated convolution can utilize single-diffraction or multi-diffraction parallel structures, and residual gating can be implemented using Sigmoid gating, gated linear units (GLUs), etc.; optimizers can be AdamW, SGD, Lion, etc., and learning rate scheduling can employ strategies such as cosine annealing and adaptive decay; the loss function can be a weighted combination of EIoU / SIoU bounding box loss with classification loss and structural constraint loss. This is only an example and is not a limitation; it can be designed according to actual needs.

[0032] The loss function in this embodiment is a comprehensive loss function, which is generated by combining a preset bounding box loss function and a preset direction matching loss function. The bounding box loss function adds independent penalties for width and height dimensions to the CIoU (Content of Integrity) calculation. This configuration is more suitable for bounding box regression of slender beam targets, improving aspect ratio prediction accuracy. The direction matching loss function introduces a target angle matching term, which better matches the predominantly lateral directional attribute of transmission tower beams, thus optimizing the directional fit of the bounding box. Furthermore, the comprehensive loss function in this embodiment can also incorporate other loss functions, such as classification loss functions. Only two types of loss functions are used as examples here, without specifying any particular type.

[0033] This embodiment uses the SGD optimizer for model optimization training because the SGD optimizer is suitable for full pre-training on large-scale transmission tower datasets. Its convergence process is stable and its generalization ability is strong, which can solidify the beam appearance feature extraction capability of the basic network. Of course, other optimizers can also be used for model training, such as the AdamW optimizer, which is adapted to multi-branch joint loss training scenarios and can suppress the risk of overfitting under small sample sizes. This is only an example and is not a limitation.

[0034] In addition to the parameters configured for model training mentioned above, regularization mechanisms, such as Dropout regularization, can also be configured. Through regularization, DropPath random depth dropping can be added to directional strip convolution and three-branch head; the classification branch adopts a label smoothing strategy; and the entire network is equipped with L2 weight decay to jointly suppress overfitting.

[0035] Geometric constraint NMS can also be performed. During non-maximum suppression, aspect ratio and orientation angle checks can be added to filter candidate boxes that do not conform to the beam's geometric properties. Edge continuity checks can also be added to the output boxes to remove falsely detected broken targets. If the network is large, a sampling transfer fine-tuning strategy can be used for layered freezing fine-tuning. For example, the backbone network can be frozen first, and only newly added modules and branches can be trained. During the later stages of training, the neck and higher-level backbones can be gradually unfrozen to achieve model convergence in stages. Specific optimization parameter configurations are not limited in this embodiment and can be selected according to actual conditions.

[0036] If the optimized detection model of this embodiment needs to be deployed to an inspection terminal, the model size cannot be too large; otherwise, it cannot be deployed on mobile terminals and will be difficult to adapt to the inspection and detection needs of flexible application environments. Therefore, when deploying the optimized detection model of this embodiment to a specific terminal, optional FP32, FP16, or INT8 quantization modes can be used for model simplification. The score fusion can be performed using fixed weights or adaptive dynamic weight fusion calculation. The output beam detection result can be a single-frame detection result or a batch detection result, both of which can be achieved by the model of this embodiment.

[0037] In the process of detecting crossbeams of transmission towers, the target crossbeam image is input into the optimized detection model, which is then subjected to directional strip convolution feature enhancement, node alignment sampling dimensionality reduction, and continuity constraint refinement. The three-branch detection head outputs visual confidence, structural consistency score, and non-crossbeam penalty term in parallel. The crossbeam prediction probability of each candidate target is calculated through a score fusion mechanism. Finally, the crossbeam detection result containing the detection box and the corresponding prediction probability can be output.

[0038] Specifically, the target beam image input optimization detection model in this embodiment undergoes initial multi-scale feature extraction in the backbone coding layer, directional enhancement processing by the directional strip convolution module, sampling aggregation analysis by the node alignment sampling module, length direction enhancement and correction by the continuity constraint module, and three-branch detection score calculation. Specifically, the directional strip convolution module in this embodiment can enhance the feature representation of the beam's lateral length extension and longitudinal thickness structure. In practice, it uses lateral strip separable convolution and longitudinal strip separable convolution to process candidate features. Lateral strip separable convolution enhances the continuous response in the beam's extension direction, while longitudinal strip separable convolution supplements local information in the lateral thickness direction and edge direction.

[0039] Based on the obtained directional enhancement feature map, the orientation of the candidate beam's structural centerline and the corresponding end point regions can be estimated. In this embodiment, an adaptive sampling module can be used to perform sampling along the possible extension direction of the beam based on the orientation of the structural centerline and the end points. Specifically, several sampling points are selected to form a sampling sequence consistent with the beam's structural orientation. The sampling points can be distributed along the candidate box centerline, the connection regions at both ends, and regions with strong edge responses. Subsequently, the features at these sampling points can be aggregated to express the overall structural state of the candidate beam, generating a sampling feature map that expresses the overall structure. Since conventional fixed-mesh downsampling or pooling operations can easily destroy the continuity of the beam structure, this embodiment designs a node-aligned sampling module to perform sampling processing of structural characteristics, preserving the continuity features of the beam structure, which is more in line with the actual situation.

[0040] The transmission tower beam detection method provided in this application employs an optimized detection model for targeted detection of transmission tower beams. The directional strip convolution module extracts structural responses along the beam's length and thickness, enhancing the continuous feature representation of the beam's slender structure. Combined with a continuity constraint module containing dilated convolution and residual gating units, it enhances long-distance continuous edge responses and suppresses false fracture features, enabling the model to accurately capture the slender, continuous structural essence of the beam, rather than relying on appearance, thus improving recognition stability in occluded scenarios. Furthermore, the node alignment sampling module performs adaptive sampling along the beam structure, avoiding the structural continuity loss caused by feature dimensionality reduction in existing fixed-grid downsampling. This mechanism compresses feature dimensions while preserving complete beam structural information, balancing the capture of small-scale details and large-scale global structures. Additionally, the use of a three-branch detection head and a scoring fusion mechanism allows for comprehensive confidence analysis, reducing the impact of similar interference and ensuring the accuracy of detection results, thus adapting to the detection requirements of UAVs. Therefore, the embodiments of this application can solve the technical problems of the lack of perception of the essence of the beam structure in the prior art, which easily leads to misjudgment, coupled with insufficient multi-scale adaptability and a large number of similar interferences, resulting in a lack of accuracy in detection.

[0041] For ease of understanding, another embodiment of the transmission tower beam detection method provided in this application includes: Step 201: Use preset image acquisition equipment to acquire images of the transmission tower inspection. The preset image acquisition equipment includes a high-definition camera, an inspection robot, a vehicle-mounted inspection equipment, and a multispectral imaging device. Step 202: Crop and scale the transmission tower inspection image to obtain the initial beam image; Step 203: The initial beam image is filtered and denoised using preset filtering techniques to obtain a denoised beam image. The preset filtering techniques include mean filtering, Gaussian filtering, and bilateral filtering. Step 204: Enhance the denoised beam image using preset enhancement techniques to obtain an enhanced beam image. The preset enhancement techniques include histogram equalization, linear grayscale transformation, and Gamma correction. Step 205: Use a preset optimization model to intelligently optimize the enhanced beam image and generate the target beam image.

[0042] It should be noted that the images of power transmission tower inspections contain information about multiple beam structures and other impurities. Therefore, basic cropping and scaling can be performed on the acquired images to generate initial beam images for subsequent beam detection. In addition to high-definition cameras, inspection robots, vehicle-mounted inspection equipment, and multispectral imaging equipment, other pre-set image acquisition devices can be selected. For example, handheld inspection devices can be used for manual acquisition, drone aerial photography can be used for mobile devices, and spectral imaging equipment can include infrared thermal imaging equipment, ultraviolet imaging equipment, and polarization imaging equipment, etc. Specific options are not limited here.

[0043] To reduce interference from redundant background information in the image from the source, this embodiment performs a series of preprocessing steps on the acquired initial beam image, including denoising, enhancement, correction, and intelligent optimization based on filtering. This not only minimizes noise interference but also enhances the visibility features of the beam target, providing reliable image source data for subsequent target detection.

[0044] In addition, image normalization, pixel value standardization, and basic image quality correction can be performed to adjust the image to the fixed resolution and numerical distribution required for model input, resulting in a target measurement image that can be directly input into the model for detection. Specifically, size normalization can use methods such as proportional scaling with padding, direct resizing, and multi-scale random cropping; pixel standardization can employ strategies such as ImageNet mean-variance normalization and sample-by-sample zero-mean normalization; image quality correction can use basic algorithms such as histogram equalization, Gamma brightness correction, and Gaussian denoising, flexibly selected according to the lighting and noise levels of the inspected image.

[0045] It is understandable that the initial beam image is generally an outdoor image, so it contains various types of noise such as salt and pepper noise, Gaussian noise, cloud noise, and wind and sand interference. In this case, to improve the image purity, two or more noise filtering techniques can be combined to eliminate impurities at different levels and improve image quality. Therefore, the preset filtering techniques in this embodiment include, but are not limited to, mean filtering, Gaussian filtering, and bilateral filtering, or a combination of the three. Other filtering techniques, such as median filtering, can also be added, which are not limited here.

[0046] Since the initial beam image in this embodiment is used for transmission tower inspection tasks, it is necessary to detect specific beam targets. Therefore, enhancement methods are needed to address brightness issues such as excessive darkness, underexposure, and low contrast caused by backlighting, shadows, cloudy days, and low illumination at night during inspections. Thus, this embodiment selects different preset enhancement techniques. Histogram equalization can automatically stretch the grayscale range of the image and optimize the sense of layering in beam images taken on cloudy days or against the light. Linear grayscale transformation can adjust the overall image brightness and contrast, improving the clarity of beam images in dim scenes. Gamma correction can correct image brightness and darkness deviations, optimizing the display effect of details in strong light shadows and low light environments.

[0047] This embodiment utilizes a pre-defined optimization model for intelligent optimization, employing artificial intelligence algorithms to achieve high-precision, intelligent image optimization. This ensures that target detection can be adapted to repair low-quality images in various complex and challenging scenarios. The architecture of the pre-defined optimization model can be designed and selected according to actual conditions, and image intelligent optimization training can be completed for use in real-world scenarios. The model can specifically achieve image super-resolution reconstruction, AI image deblurring, background blurring and segmentation, and defect feature enhancement. It can not only comprehensively improve image clarity but also magnify and repair the detailed features of the beam structure, highlighting the presence of the detected target.

[0048] It is understood that this embodiment can sequentially perform the above-mentioned different preprocessing optimizations on the initial beam image. Furthermore, in addition to the preprocessing operations proposed in this embodiment, other related techniques can be added to improve image optimization. For example, image sharpening enhancement processing can employ techniques such as Sobel sharpening, Laplacian sharpening, Canny edge detection enhancement, and high-frequency component enhancement to highlight image details, textures, and the edge features of the beam outline, while preserving high-frequency detail features and weakening low-frequency background interference, thereby comprehensively improving the clarity of the beam's details. Since the image needs to be input into the model for target detection, necessary grayscale and binarization processing can also be performed on the obtained target beam image. This is not specifically limited here and can be designed according to the actual situation.

[0049] Step 206: Based on the directional strip convolution module, node alignment sampling module and continuity constraint module, construct and train a model based on the three-branch detection head and score fusion mechanism to generate an optimized detection model. The continuity constraint module includes dilated convolution and residual gating units.

[0050] It should be noted that the optimized detection model in this embodiment uses YOLOv12 as the base network to generate candidate boxes and performs a two-stage architecture of three-branch secondary verification for structural semantic analysis. This model performs thorough detection and analysis on the beam image through four core modules. The directional strip convolution module can enhance the continuous response in the beam length direction and the edge features in the thickness direction by combining strip convolutions. The node alignment sampling module can perform adaptive sampling along the beam length direction, replacing fixed grid downsampling, which can avoid destroying the continuity of the beam structure. The continuity constraint module can enhance the long-distance continuous edge response and suppress the pseudo response caused by fracture and intersection structures through dilated convolution and residual gating units. The three-branch detection head and scoring fusion mechanism can output specific parameters of vision, structure and non-beam, and filter out beam targets by calculating the fused score.

[0051] The optimized detection model in this embodiment can retain the continuous structural features of the beam and ensure good detection stability even in scenarios with partial occlusion and changes in scale. Moreover, it is modularly designed based on the YOLOv12 framework, with controllable computational complexity, compatibility with existing inspection equipment, and real-time deployment on the UAV side, thus meeting inspection requirements.

[0052] Furthermore, the training process for optimizing the detection model can be further designed and optimized, specifically including: Based on the cosine annealing restart strategy, adaptive learning rates are configured for the three-branch detection heads; Configure the comprehensive loss function based on the preset bounding box loss function and the preset orientation matching loss function; Model training is achieved by combining the SGD optimizer, adaptive learning rate, and comprehensive loss function.

[0053] It should be noted that the initial detection model built based on YOLOv12 needs to be optimized and trained using a beam image training set for object detection. The training process can be tailored to specific parameters suitable for the beam detection task, such as the learning rate, loss function, and optimizer. In this embodiment, the learning rate is an adaptively changing parameter. Specifically, for the three-branch detection head, a cosine annealing restart strategy is used to configure the learning rate for each branch. This ensures that the output structure of the three-branch detection head better reflects the actual situation, improving prediction accuracy. Adaptive learning rate configuration strategies can also be designed and selected according to actual conditions; this is only an example and not a limitation.

[0054] The loss function in this embodiment is a comprehensive loss function, which is generated by combining a preset bounding box loss function and a preset direction matching loss function. The bounding box loss function adds independent penalties for width and height dimensions to the CIoU (Content of Integrity) calculation. This configuration is more suitable for bounding box regression of slender beam targets, improving aspect ratio prediction accuracy. The direction matching loss function introduces a target angle matching term, which better matches the predominantly lateral directional attribute of transmission tower beams, thus optimizing the directional fit of the bounding box. Furthermore, the comprehensive loss function in this embodiment can also incorporate other loss functions, such as classification loss functions. Only two types of loss functions are used as examples here, without specifying any particular type.

[0055] This embodiment uses the SGD optimizer for model optimization training because the SGD optimizer is suitable for full pre-training on large-scale transmission tower datasets. Its convergence process is stable and its generalization ability is strong, which can solidify the beam appearance feature extraction capability of the basic network. Of course, other optimizers can also be used for model training, such as the AdamW optimizer, which is adapted to multi-branch joint loss training scenarios and can suppress the risk of overfitting under small sample sizes. This is only an example and is not a limitation.

[0056] In addition to the parameters configured for model training mentioned above, regularization mechanisms, such as Dropout regularization, can also be configured. Through regularization, DropPath random depth dropping can be added to directional strip convolution and three-branch head; the classification branch adopts a label smoothing strategy; and the entire network is equipped with L2 weight decay to jointly suppress overfitting.

[0057] Geometric constraint NMS can also be performed. During non-maximum suppression, aspect ratio and orientation angle checks can be added to filter candidate boxes that do not conform to the beam's geometric properties. Edge continuity checks can also be added to the output boxes to remove falsely detected broken targets. If the network is large, a sampling transfer fine-tuning strategy can be used for layered freezing fine-tuning. For example, the backbone network can be frozen first, and only newly added modules and branches can be trained. During the later stages of training, the neck and higher-level backbones can be gradually unfrozen to achieve model convergence in stages. Specific optimization parameter configurations are not limited in this embodiment and can be selected according to actual conditions.

[0058] Furthermore, to achieve efficient and reliable terminal deployment, the optimized detection model can be simplified, specifically including: The optimized detection model is compressed using INT8 quantization technology. The optimized detection model is pruned using a pre-defined structured channel pruning scheme.

[0059] It should be noted that if the optimized detection model in this embodiment needs to be deployed to an inspection terminal, the model size cannot be too large; otherwise, it will be impossible to deploy on mobile terminals and adapt to the inspection needs of flexible application environments. Therefore, this embodiment can also use quantization and pruning techniques to compress and prune the model, minimizing its size as much as possible without changing its core structure and function. It is understood that, in addition to quantization compression and model pruning, knowledge distillation and operator fusion acceleration can also be used to achieve model lightweighting; these are merely examples and not limitations.

[0060] Step 207: Input the target beam image into the optimized detection model for multi-scale feature extraction to obtain the initial feature map; Step 208: Use the directional strip convolution module in the model to perform lateral length continuous feature enhancement and vertical thickness and edge local feature enhancement on the initial feature map based on separable convolution to obtain the directional enhanced feature map; Step 209: Estimate the orientation of the structural centerline and the corresponding two endpoints of the candidate beam based on the orientation enhancement feature map; Step 210: Using the node alignment sampling module, adaptive sampling and feature aggregation analysis are performed based on the direction of the structural centerline and the two end points along the extension direction of the beam to determine the sampling feature map; Step 211: Expand the receptive field of the sampled feature map using dilated convolution in the continuity constraint module, and at the same time combine the residual gating unit to enhance and correct the continuous edge response in the beam length direction to generate a shared feature map. Step 212: Input the shared feature map into the three-branch detection head for detection and analysis to obtain the three-branch detection results. The three-branch detection results include visual detection information, structural consistency score, and non-beam suppression penalty term. Step 213: Based on the scoring fusion mechanism, calculate the weighted score according to the three-branch detection results and generate the beam detection results.

[0061] Please see Figure 3 In this embodiment, the target beam image input optimization detection model undergoes initial multi-scale feature extraction in the backbone coding layer, directional enhancement processing by the directional strip convolution module, sampling aggregation analysis by the node alignment sampling module, length direction enhancement and correction by the continuity constraint module, and three-branch detection score calculation. Specifically, the directional strip convolution module in this embodiment can enhance the feature representation of the beam's lateral length extension and longitudinal thickness structure. This is achieved by using lateral and longitudinal strip separable convolutions to process candidate features. The lateral strip separable convolution enhances the continuous response in the beam's extension direction, while the longitudinal strip separable convolution supplements local information in the lateral thickness direction and edge direction.

[0062] Based on the obtained directional enhancement feature map, the orientation of the candidate beam's structural centerline and the corresponding end point regions can be estimated. In this embodiment, an adaptive sampling module can be used to perform sampling along the possible extension direction of the beam based on the orientation of the structural centerline and the end points. Specifically, several sampling points are selected to form a sampling sequence consistent with the beam's structural orientation. The sampling points can be distributed along the candidate box centerline, the connection regions at both ends, and regions with strong edge responses. Subsequently, the features at these sampling points can be aggregated to express the overall structural state of the candidate beam, generating a sampling feature map that expresses the overall structure. Since conventional fixed-mesh downsampling or pooling operations can easily destroy the continuity of the beam structure, this embodiment designs a node-aligned sampling module to perform sampling processing of structural characteristics, preserving the continuity features of the beam structure, which is more in line with the actual situation.

[0063] Since beams are typically long-distance continuous structures, while diagonal members or background interference often exhibit localized, intersecting, or discontinuous responses, the continuity constraint module in this embodiment includes dilated convolution and residual gating units to determine whether the linear responses within the candidate beam region are continuous. Specifically, dilated convolution can be used to expand the receptive field of the sampled feature map, while residual gating units are combined to enhance and correct the continuous edge responses in the transverse length direction of the beam. When the transverse responses within the candidate region are continuously distributed along the length direction, this mechanism can improve the beam structure consistency score.

[0064] The generated shared feature map is input into a three-branch detection head for detection and analysis. The three-branch detection head includes a detection head, a structural verification head, and a non-beam suppression head. The detection head outputs the beam bounding box, category, and visual confidence score; the structural verification head outputs the orientation angle, aspect ratio, endpoint connection strength, and structural consistency score; and the non-beam suppression head outputs the penalty term for diagonal members or similar components, i.e., the non-beam suppression penalty term.

[0065] The scoring fusion mechanism involves weighting the outputs of the three-branch detection heads to obtain a comprehensive score. This comprehensive score is then compared to a scoring threshold to determine the beam detection result. If the score is lower than the threshold, the beam is classified as either a non-beam or a low-confidence candidate. Specifically, the weighted calculation includes visual confidence, structural consistency score, and non-beam suppression penalty. The weighted calculation process can be expressed as follows:

[0066] in, S This represents the overall score. Indicates visual confidence level. Indicates the structural consistency score. This indicates a non-beam suppression penalty term. , These are weighting coefficients. They are only considered in the overall score. S Only when the score exceeds the threshold is it determined to be a beam, and a beam detection result is obtained; otherwise, it is determined to be a non-beam or a low-confidence candidate. The comprehensive detection mechanism of this embodiment can avoid misjudgment caused by over-reliance on visual confidence judgment, so that the detection result simultaneously satisfies the appearance recognition condition and the structural semantic condition.

[0067] This embodiment compares the performance of the conventional Yolov12 algorithm with the two core metrics of mAP50 and mAP50-95. Please refer to Table 1. The optimized detection model Yolov12_new provided in this embodiment demonstrates a stable improvement in both metrics. Furthermore, the algorithm in this embodiment exhibits strong stability, is unaffected by background interference, and can maintain high detection accuracy in various complex environments. For details, please refer to [link / reference]. Figures 4-8 .

[0068] Table 1. Comparison of mAP50 and mAP50-95 metrics for different algorithms

[0069] The transmission tower beam detection method provided in this application employs an optimized detection model for targeted detection of transmission tower beams. The directional strip convolution module extracts structural responses along the beam's length and thickness, enhancing the continuous feature representation of the beam's slender structure. Combined with a continuity constraint module containing dilated convolution and residual gating units, it enhances long-distance continuous edge responses and suppresses false fracture features, enabling the model to accurately capture the slender, continuous structural essence of the beam, rather than relying on appearance, thus improving recognition stability in occluded scenarios. Furthermore, the node alignment sampling module performs adaptive sampling along the beam structure, avoiding the structural continuity loss caused by feature dimensionality reduction in existing fixed-grid downsampling. This mechanism compresses feature dimensions while preserving complete beam structural information, balancing the capture of small-scale details and large-scale global structures. Additionally, the use of a three-branch detection head and a scoring fusion mechanism allows for comprehensive confidence analysis, reducing the impact of similar interference and ensuring the accuracy of detection results, thus adapting to the detection requirements of UAVs. Therefore, the embodiments of this application can solve the technical problems of the lack of perception of the essence of the beam structure in the prior art, which easily leads to misjudgment, coupled with insufficient multi-scale adaptability and a large number of similar interferences, resulting in a lack of accuracy in detection.

[0070] For easier understanding, please refer to Figure 2 This application provides an embodiment of a transmission tower beam detection device, comprising: The image acquisition unit 201 is used to preprocess the acquired transmission tower inspection images to obtain the target beam image; The model generation unit 202 is used to build and train a model based on the three-branch detection head and score fusion mechanism according to the directional strip convolution module, the node alignment sampling module and the continuity constraint module, and generate an optimized detection model. The continuity constraint module includes dilated convolution and residual gating units. The beam detection unit 203 is used to input the target beam image into the optimized detection model to perform beam detection and obtain beam detection results, which include beam prediction probabilities.

[0071] Furthermore, the image acquisition unit 201 is specifically used for: Images of power transmission towers are acquired using pre-set image acquisition equipment, which includes high-definition cameras, inspection robots, vehicle-mounted inspection equipment, and multispectral imaging equipment. The images of the transmission tower inspection are cropped and scaled to obtain the initial beam image; The initial beam image is filtered and denoised using preset filtering techniques to obtain a denoised beam image. The preset filtering techniques include mean filtering, Gaussian filtering, and bilateral filtering. The denoised beam image is enhanced by using preset enhancement techniques, including histogram equalization, linear grayscale transformation, and Gamma correction. A preset optimization model is used to intelligently optimize the enhanced beam image to generate the target beam image.

[0072] Furthermore, it also includes an optimization configuration unit 204, used for: Based on the cosine annealing restart strategy, adaptive learning rates are configured for the three-branch detection heads; Configure the comprehensive loss function based on the preset bounding box loss function and the preset orientation matching loss function; Model training is achieved by combining the SGD optimizer, adaptive learning rate, and comprehensive loss function.

[0073] Furthermore, it also includes a model simplification unit for: The optimized detection model is compressed using INT8 quantization technology. The optimized detection model is pruned using a pre-defined structured channel pruning scheme.

[0074] Furthermore, the beam detection unit 203 is specifically used for: The target beam image is input into the optimized detection model for multi-scale feature extraction to obtain an initial feature map. The directional strip convolution module in the model is used to perform lateral length continuous feature enhancement and vertical thickness and edge local feature enhancement on the initial feature map based on separable convolution, resulting in a directional enhanced feature map; The orientation of the structural centerline and the corresponding two endpoints of the candidate beam are estimated based on the direction-enhanced feature map. The node alignment sampling module performs adaptive sampling and feature aggregation analysis based on the direction of the structural centerline and the two end points along the extension direction of the beam to determine the sampling feature map. The receptive field of the sampled feature map is expanded by using dilated convolution in the continuity constraint module, and the continuous edge response in the beam length direction is enhanced and corrected by combining residual gating unit to generate shared feature map; The shared feature map is input into the three-branch detection head for detection and analysis to obtain the three-branch detection results, which include visual detection information, structural consistency score, and non-beam suppression penalty term. Based on the scoring fusion mechanism, a weighted score is calculated according to the three-branch detection results, and the beam detection results are generated.

[0075] It should be noted that the images of power transmission tower inspections contain information about multiple beam structures and other impurities. Therefore, basic cropping and scaling can be performed on the acquired images to generate initial beam images for subsequent beam detection. In addition to high-definition cameras, inspection robots, vehicle-mounted inspection equipment, and multispectral imaging equipment, other pre-set image acquisition devices can be selected. For example, handheld inspection devices can be used for manual acquisition, drone aerial photography can be used for mobile devices, and spectral imaging equipment can include infrared thermal imaging equipment, ultraviolet imaging equipment, and polarization imaging equipment, etc. Specific options are not limited here.

[0076] To reduce interference from redundant background information in the image from the source, this embodiment performs a series of preprocessing steps on the acquired initial beam image, including denoising, enhancement, correction, and intelligent optimization based on filtering. This not only minimizes noise interference but also enhances the visibility features of the beam target, providing reliable image source data for subsequent target detection.

[0077] In addition, image normalization, pixel value standardization, and basic image quality correction can be performed to adjust the image to the fixed resolution and numerical distribution required for model input, resulting in a target measurement image that can be directly input into the model for detection. Specifically, size normalization can use methods such as proportional scaling with padding, direct resizing, and multi-scale random cropping; pixel standardization can employ strategies such as ImageNet mean-variance normalization and sample-by-sample zero-mean normalization; image quality correction can use basic algorithms such as histogram equalization, Gamma brightness correction, and Gaussian denoising, flexibly selected according to the lighting and noise levels of the inspected image.

[0078] It is understandable that the initial beam image is generally an outdoor image, so it contains various types of noise such as salt and pepper noise, Gaussian noise, cloud noise, and wind and sand interference. In this case, to improve the image purity, two or more noise filtering techniques can be combined to eliminate impurities at different levels and improve image quality. Therefore, the preset filtering techniques in this embodiment include, but are not limited to, mean filtering, Gaussian filtering, and bilateral filtering, or a combination of the three. Other filtering techniques, such as median filtering, can also be added, which are not limited here.

[0079] Since the initial beam image in this embodiment is used for transmission tower inspection tasks, it is necessary to detect specific beam targets. Therefore, enhancement methods are needed to address brightness issues such as excessive darkness, underexposure, and low contrast caused by backlighting, shadows, cloudy days, and low illumination at night during inspections. Thus, this embodiment selects different preset enhancement techniques. Histogram equalization can automatically stretch the grayscale range of the image and optimize the sense of layering in beam images taken on cloudy days or against the light. Linear grayscale transformation can adjust the overall image brightness and contrast, improving the clarity of beam images in dim scenes. Gamma correction can correct image brightness and darkness deviations, optimizing the display effect of details in strong light shadows and low light environments.

[0080] This embodiment utilizes a pre-defined optimization model for intelligent optimization, employing artificial intelligence algorithms to achieve high-precision, intelligent image optimization. This ensures that target detection can be adapted to repair low-quality images in various complex and challenging scenarios. The architecture of the pre-defined optimization model can be designed and selected according to actual conditions, and image intelligent optimization training can be completed for use in real-world scenarios. The model can specifically achieve image super-resolution reconstruction, AI image deblurring, background blurring and segmentation, and defect feature enhancement. It can not only comprehensively improve image clarity but also magnify and repair the detailed features of the beam structure, highlighting the presence of the detected target.

[0081] It is understood that this embodiment can sequentially perform the above-mentioned different preprocessing optimizations on the initial beam image. Furthermore, in addition to the preprocessing operations proposed in this embodiment, other related techniques can be added to improve image optimization. For example, image sharpening enhancement processing can employ techniques such as Sobel sharpening, Laplacian sharpening, Canny edge detection enhancement, and high-frequency component enhancement to highlight image details, textures, and the edge features of the beam outline, while preserving high-frequency detail features and weakening low-frequency background interference, thereby comprehensively improving the clarity of the beam's details. Since the image needs to be input into the model for target detection, necessary grayscale and binarization processing can also be performed on the obtained target beam image. This is not specifically limited here and can be designed according to the actual situation.

[0082] It should be noted that the optimized detection model in this embodiment uses YOLOv12 as the base network to generate candidate boxes and performs a two-stage architecture of three-branch secondary verification for structural semantic analysis. This model performs thorough detection and analysis on the beam image through four core modules. The directional strip convolution module can enhance the continuous response in the beam length direction and the edge features in the thickness direction by combining strip convolutions. The node alignment sampling module can perform adaptive sampling along the beam length direction, replacing fixed grid downsampling, which can avoid destroying the continuity of the beam structure. The continuity constraint module can enhance the long-distance continuous edge response and suppress the pseudo response caused by fracture and intersection structures through dilated convolution and residual gating units. The three-branch detection head and scoring fusion mechanism can output specific parameters of vision, structure and non-beam, and filter out beam targets by calculating the fused score.

[0083] The optimized detection model in this embodiment can retain the continuous structural features of the beam and ensure good detection stability even in scenarios with partial occlusion and changes in scale. Moreover, it is modularly designed based on the YOLOv12 framework, with controllable computational complexity, compatibility with existing inspection equipment, and real-time deployment on the UAV side, thus meeting inspection requirements.

[0084] It should be noted that the initial detection model built based on YOLOv12 needs to be optimized and trained using a beam image training set for object detection. The training process can be tailored to specific parameters suitable for the beam detection task, such as the learning rate, loss function, and optimizer. In this embodiment, the learning rate is an adaptively changing parameter. Specifically, for the three-branch detection head, a cosine annealing restart strategy is used to configure the learning rate for each branch. This ensures that the output structure of the three-branch detection head better reflects the actual situation, improving prediction accuracy. Adaptive learning rate configuration strategies can also be designed and selected according to actual conditions; this is only an example and not a limitation.

[0085] The loss function in this embodiment is a comprehensive loss function, which is generated by combining a preset bounding box loss function and a preset direction matching loss function. The bounding box loss function adds independent penalties for width and height dimensions to the CIoU (Content of Integrity) calculation. This configuration is more suitable for bounding box regression of slender beam targets, improving aspect ratio prediction accuracy. The direction matching loss function introduces a target angle matching term, which better matches the predominantly lateral directional attribute of transmission tower beams, thus optimizing the directional fit of the bounding box. Furthermore, the comprehensive loss function in this embodiment can also incorporate other loss functions, such as classification loss functions. Only two types of loss functions are used as examples here, without specifying any particular type.

[0086] This embodiment uses the SGD optimizer for model optimization training because the SGD optimizer is suitable for full pre-training on large-scale transmission tower datasets. Its convergence process is stable and its generalization ability is strong, which can solidify the beam appearance feature extraction capability of the basic network. Of course, other optimizers can also be used for model training, such as the AdamW optimizer, which is adapted to multi-branch joint loss training scenarios and can suppress the risk of overfitting under small sample sizes. This is only an example and is not a limitation.

[0087] In addition to the parameters configured for model training mentioned above, regularization mechanisms, such as Dropout regularization, can also be configured. Through regularization, DropPath random depth dropping can be added to directional strip convolution and three-branch head; the classification branch adopts a label smoothing strategy; and the entire network is equipped with L2 weight decay to jointly suppress overfitting.

[0088] Geometric constraint NMS can also be performed. During non-maximum suppression, aspect ratio and orientation angle checks can be added to filter candidate boxes that do not conform to the beam's geometric properties. Edge continuity checks can also be added to the output boxes to remove falsely detected broken targets. If the network is large, a sampling transfer fine-tuning strategy can be used for layered freezing fine-tuning. For example, the backbone network can be frozen first, and only newly added modules and branches can be trained. During the later stages of training, the neck and higher-level backbones can be gradually unfrozen to achieve model convergence in stages. Specific optimization parameter configurations are not limited in this embodiment and can be selected according to actual conditions.

[0089] It should be noted that if the optimized detection model in this embodiment needs to be deployed to an inspection terminal, the model size cannot be too large; otherwise, it will be impossible to deploy on mobile terminals and adapt to the inspection needs of flexible application environments. Therefore, this embodiment can also use quantization and pruning techniques to compress and prune the model, minimizing its size as much as possible without changing its core structure and function. It is understood that, in addition to quantization compression and model pruning, knowledge distillation and operator fusion acceleration can also be used to achieve model lightweighting; these are merely examples and not limitations.

[0090] Please see Figure 3In this embodiment, the target beam image input optimization detection model undergoes initial multi-scale feature extraction in the backbone coding layer, directional enhancement processing by the directional strip convolution module, sampling aggregation analysis by the node alignment sampling module, length direction enhancement and correction by the continuity constraint module, and three-branch detection score calculation. Specifically, the directional strip convolution module in this embodiment can enhance the feature representation of the beam's lateral length extension and longitudinal thickness structure. This is achieved by using lateral and longitudinal strip separable convolutions to process candidate features. The lateral strip separable convolution enhances the continuous response in the beam's extension direction, while the longitudinal strip separable convolution supplements local information in the lateral thickness direction and edge direction.

[0091] Based on the obtained directional enhancement feature map, the orientation of the candidate beam's structural centerline and the corresponding end point regions can be estimated. In this embodiment, an adaptive sampling module can be used to perform sampling along the possible extension direction of the beam based on the orientation of the structural centerline and the end points. Specifically, several sampling points are selected to form a sampling sequence consistent with the beam's structural orientation. The sampling points can be distributed along the candidate box centerline, the connection regions at both ends, and regions with strong edge responses. Subsequently, the features at these sampling points can be aggregated to express the overall structural state of the candidate beam, generating a sampling feature map that expresses the overall structure. Since conventional fixed-mesh downsampling or pooling operations can easily destroy the continuity of the beam structure, this embodiment designs a node-aligned sampling module to perform sampling processing of structural characteristics, preserving the continuity features of the beam structure, which is more in line with the actual situation.

[0092] Since beams are typically long-distance continuous structures, while diagonal members or background interference often exhibit localized, intersecting, or discontinuous responses, the continuity constraint module in this embodiment includes dilated convolution and residual gating units to determine whether the linear responses within the candidate beam region are continuous. Specifically, dilated convolution can be used to expand the receptive field of the sampled feature map, while residual gating units are combined to enhance and correct the continuous edge responses in the transverse length direction of the beam. When the transverse responses within the candidate region are continuously distributed along the length direction, this mechanism can improve the beam structure consistency score.

[0093] The generated shared feature map is input into a three-branch detection head for detection and analysis. The three-branch detection head includes a detection head, a structural verification head, and a non-beam suppression head. The detection head outputs the beam bounding box, category, and visual confidence score; the structural verification head outputs the orientation angle, aspect ratio, endpoint connection strength, and structural consistency score; and the non-beam suppression head outputs the penalty term for diagonal members or similar components, i.e., the non-beam suppression penalty term.

[0094] The scoring fusion mechanism involves weighting the outputs of the three-branch detection heads to obtain a comprehensive score. This comprehensive score is then compared to a scoring threshold to determine the beam detection result. If the score is lower than the threshold, the beam is classified as either a non-beam or a low-confidence candidate. Specifically, the weighted calculation includes visual confidence, structural consistency score, and non-beam suppression penalty. The weighted calculation process can be expressed as follows:

[0095] in, S This represents the overall score. Indicates visual confidence level. Indicates the structural consistency score. This indicates a non-beam suppression penalty term. , These are weighting coefficients. They are only considered in the overall score. S Only when the score exceeds the threshold is it determined to be a beam, and a beam detection result is obtained; otherwise, it is determined to be a non-beam or a low-confidence candidate. The comprehensive detection mechanism of this embodiment can avoid misjudgment caused by over-reliance on visual confidence judgment, so that the detection result simultaneously satisfies the appearance recognition condition and the structural semantic condition.

[0096] This embodiment compares the performance of the conventional Yolov12 algorithm with the two core metrics of mAP50 and mAP50-95 (see Table 1). The optimized detection model provided in this embodiment demonstrates a stable improvement in both metrics. Furthermore, the algorithm in this embodiment exhibits strong stability, is unaffected by background interference, and can maintain high detection accuracy in various complex environments. For details, please refer to... Figures 4-8 .

[0097] This application also provides a transmission tower beam detection device, the device including a processor and a memory; The memory is used to store program code and transfer the program code to the processor; The processor is used to execute the transmission tower beam detection method in the above method embodiment according to the instructions in the program code.

[0098] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0099] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0100] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0101] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for executing all or part of the steps of the methods described in the various embodiments of this application through a computer device (which may be a personal computer, server, or network device, etc.). The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.

[0102] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for detecting the crossbeams of a transmission tower, characterized in that, include: The acquired transmission tower inspection images are preprocessed to obtain the target beam image; The model is constructed and trained based on the three-branch detection head and score fusion mechanism according to the directional strip convolution module, the node alignment sampling module and the continuity constraint module to generate an optimized detection model. The continuity constraint module includes dilated convolution and residual gating unit. The target beam image is input into the optimized detection model for beam detection, and the beam detection result is obtained, which includes the beam prediction probability.

2. The method for detecting the crossbeam of a transmission tower according to claim 1, characterized in that, The preprocessing of the acquired transmission tower inspection images to obtain the target beam image includes: Images of power transmission towers are acquired using pre-set image acquisition equipment, which includes a high-definition camera, an inspection robot, a vehicle-mounted inspection device, and a multispectral imaging device. The transmission tower inspection image is cropped and scaled to obtain the initial beam image; The initial beam image is filtered and denoised using preset filtering techniques to obtain a denoised beam image. The preset filtering techniques include mean filtering, Gaussian filtering, and bilateral filtering. The denoised beam image is enhanced by a preset enhancement technique, which includes histogram equalization, linear grayscale transformation and Gamma correction. The enhanced beam image is intelligently optimized using a preset optimization model to generate the target beam image.

3. The method for detecting the crossbeam of a transmission tower according to claim 1, characterized in that, The step of building and training a model based on a three-branch detection head and score fusion mechanism using a directional strip convolution module, a node alignment sampling module, and a continuity constraint module to generate an optimized detection model also includes: Based on the cosine annealing restart strategy, an adaptive learning rate is configured for each of the three branch detection heads; Configure the comprehensive loss function based on the preset bounding box loss function and the preset orientation matching loss function; Model training is achieved by combining the SGD optimizer, the adaptive learning rate, and the comprehensive loss function.

4. The method for detecting the crossbeam of a transmission tower according to claim 1, characterized in that, The step of building and training a model based on a three-branch detection head and score fusion mechanism using a directional strip convolution module, a node alignment sampling module, and a continuity constraint module to generate an optimized detection model also includes: The optimized detection model is compressed using INT8 quantization technology. The optimized detection model is pruned using a pre-defined structured channel pruning scheme.

5. The method for detecting the crossbeam of a transmission tower according to claim 1, characterized in that, The step involves inputting the target beam image into the optimized detection model for beam detection, obtaining beam detection results, wherein the beam detection results include beam prediction probabilities, characterized in that they include: The target beam image is input into the optimized detection model for multi-scale feature extraction to obtain an initial feature map; The directional strip convolution module in the model is used to perform lateral length continuous feature enhancement and vertical thickness and edge local feature enhancement on the initial feature map based on separable convolution to obtain a directional enhanced feature map; Based on the enhanced directional feature map, the orientation of the structural centerline of the candidate beam and its corresponding two endpoints are estimated. The node alignment sampling module performs adaptive sampling and feature aggregation analysis based on the direction of the structural centerline and the two end points along the extension direction of the beam to determine the sampling feature map. The dilated convolution in the continuity constraint module is used to expand the receptive field of the sampled feature map. At the same time, the residual gating unit is combined to enhance and correct the continuous edge response in the beam length direction to generate a shared feature map. The shared feature map is input into the three-branch detection head for detection and analysis to obtain the three-branch detection result, which includes visual detection information, structural consistency score, and non-beam suppression penalty term. Based on the scoring fusion mechanism, a weighted score is calculated according to the detection results of the three branches, and the beam detection result is generated.

6. A transmission tower crossbeam detection device, characterized in that, include: The image acquisition unit is used to preprocess the acquired transmission tower inspection images to obtain the target beam image; The model generation unit is used to construct and train a model based on the three-branch detection head and score fusion mechanism according to the directional strip convolution module, the node alignment sampling module and the continuity constraint module, and generate an optimized detection model. The continuity constraint module includes dilated convolution and residual gating units. A beam detection unit is used to input the target beam image into the optimized detection model for beam detection and obtain beam detection results, including beam prediction probabilities.

7. The transmission tower beam detection device according to claim 6, characterized in that, The image acquisition unit is specifically used for: Images of power transmission towers are acquired using pre-set image acquisition equipment, which includes a high-definition camera, an inspection robot, a vehicle-mounted inspection device, and a multispectral imaging device. The transmission tower inspection image is cropped and scaled to obtain the initial beam image; The initial beam image is filtered and denoised using preset filtering techniques to obtain a denoised beam image. The preset filtering techniques include mean filtering, Gaussian filtering, and bilateral filtering. The denoised beam image is enhanced by a preset enhancement technique, which includes histogram equalization, linear grayscale transformation and Gamma correction. The enhanced beam image is intelligently optimized using a preset optimization model to generate the target beam image.

8. The transmission tower beam detection device according to claim 6, characterized in that, It also includes an optimized configuration unit for: Based on the cosine annealing restart strategy, an adaptive learning rate is configured for each of the three branch detection heads; Configure the comprehensive loss function based on the preset bounding box loss function and the preset orientation matching loss function; Model training is achieved by combining the SGD optimizer, the adaptive learning rate, and the comprehensive loss function.

9. The transmission tower beam detection device according to claim 6, characterized in that, The beam detection unit is specifically used for: The target beam image is input into the optimized detection model for multi-scale feature extraction to obtain an initial feature map; The directional strip convolution module in the model is used to perform lateral length continuous feature enhancement and vertical thickness and edge local feature enhancement on the initial feature map based on separable convolution to obtain a directional enhanced feature map; Based on the enhanced directional feature map, the orientation of the structural centerline of the candidate beam and its corresponding two endpoints are estimated. The node alignment sampling module performs adaptive sampling and feature aggregation analysis based on the direction of the structural centerline and the two end points along the extension direction of the beam to determine the sampling feature map. The dilated convolution in the continuity constraint module is used to expand the receptive field of the sampled feature map. At the same time, the residual gating unit is combined to enhance and correct the continuous edge response in the beam length direction to generate a shared feature map. The shared feature map is input into the three-branch detection head for detection and analysis to obtain the three-branch detection result, which includes visual detection information, structural consistency score, and non-beam suppression penalty term. Based on the scoring fusion mechanism, a weighted score is calculated according to the detection results of the three branches, and the beam detection result is generated.

10. A transmission tower beam testing device, characterized in that, The device includes a processor and a memory; The memory is used to store program code and transmit the program code to the processor; The processor is used to execute the transmission tower beam detection method according to any one of claims 1-5 according to the instructions in the program code.

Images