Electric tower metal anomaly detection method based on unmanned aerial vehicle and related device

Abstract

The invention provides an electric tower metal anomaly detection method based on an unmanned aerial vehicle and a related device. A visible light image is input into a first feature extraction network to extract a first feature, and an infrared image is input into a second feature extraction network to extract a second feature; performing cross splicing operation and depth separable convolution operation on the first feature and the second feature to obtain a third feature; performing up-sampling on the plurality of third features of different scales along an up-sampling path and performing down-sampling on the plurality of third features of different scales along a down-sampling path to obtain a plurality of fourth features of different scales; performing convolution processing on each fourth feature through a plurality of convolution paths to obtain a plurality of fifth features, and fusing the plurality of fifth features to obtain a sixth feature; performing detection according to the sixth feature to obtain an electric tower metal anomaly detection result; and precision, efficiency, light weight and generalization capability can be considered.

Description

Technical Field

[0001] This application relates to the field of image processing, and more particularly to a method and related apparatus for detecting metal anomalies in power towers based on unmanned aerial vehicles (UAVs). Background Technology

[0002] Traditional high-voltage power tower inspection relies on manual climbing or single-spectrum imaging, which suffers from low efficiency, high risk, and incomplete coverage. A single spectrum (such as visible light) cannot comprehensively characterize multiple types of defects, such as metal corrosion. Traditional manual inspection relies on visual inspection or basic image processing techniques, which suffers from low efficiency, high subjectivity, and high false negative rates, especially in high-speed continuous production scenarios where real-time requirements are difficult to meet. Furthermore, traditional image inspection techniques suffer from the problem of difficulty in modeling the global spatial correlation of fine cracks due to local receptive fields. Summary of the Invention

[0003] The following is an overview of the topics described in detail in this article.

[0004] The purpose of this application is to at least partially solve one of the technical problems existing in the related technologies. The embodiments of this application provide a method and related device for detecting metal anomalies in power towers based on UAVs, which can take into account accuracy, efficiency, lightweight and generalization ability.

[0005] An embodiment of the first aspect of this application provides a method for detecting metal anomalies in power towers based on unmanned aerial vehicles (UAVs), comprising: Acquire images of power towers, including visible light images and infrared images; The visible light image is input into a first feature extraction network to extract a first feature, and the infrared image is input into a second feature extraction network to extract a second feature, wherein the first feature extraction network and the second feature extraction network share weights; The first feature and the second feature are combined with a cross-concatenation operation and a depthwise separable convolution operation to obtain the third feature; Multiple third features at different scales are upsampled along an upsampling path and downsampled along a downsampling path to obtain multiple fourth features at different scales. Each fourth feature is processed by multiple convolutional paths to obtain multiple fifth features, and the multiple fifth features are fused to obtain a sixth feature; Based on the sixth feature, the detection results of the power tower metal anomaly are obtained.

[0006] According to certain embodiments of the first aspect of this application, the image of the power tower is obtained by taking pictures with a camera mounted on a drone, which flies along a planned path; The planned path was obtained using the following method: Obtain the preset power tower monitoring points; A first path is generated based on the power tower monitoring points; A second path is generated by swapping the positions of any number of power tower monitoring points in the first path. When the total length of the second path is less than the total length of the first path, update the first path to the second path; When the total length of the second path is greater than the total length of the first path, a probability value is obtained based on the current temperature coefficient. The first path is then updated to the second path based on the probability value. Alternatively, the first path can be kept unchanged based on the difference between the probability value and the second path. Update the current temperature coefficient based on the preset cooling coefficient; If the current temperature coefficient is less than the preset temperature coefficient threshold or reaches the preset iteration number threshold, the current first path will be used as the planned path.

[0007] According to certain embodiments of the first aspect of this application, the power tower image is labeled with a time stamp and a location label for the power tower monitoring point.

[0008] According to certain embodiments of the first aspect of this application, the method further includes: The fourth feature is decoupled into a multi-scale subgraph; An attention mechanism is applied to the multi-scale subgraph.

[0009] According to certain embodiments of the first aspect of this application, upsampling of multiple third features at different scales along an upsampling path includes: The third feature is convolved and upsampled and then integrated with the second feature to obtain the first integrated feature. The result of convolution and upsampling of the first integrated feature is convolved with the first third feature and the first integrated feature to obtain the third upsampled output feature. The result of convolving the third feature is used as the first upsampled output feature, and the result of convolving the first integrated feature is used as the second upsampled output feature.

[0010] According to certain embodiments of the first aspect of this application, multiple third features at different scales are downsampled along a downsampling path, including: The result of convolving the third upsampled output feature is used as the first downsampled output feature; The result of convolving the third upsampled output feature is downsampled to obtain the first downsampled result. The first downsampled result is then integrated with the second upsampled output feature and convolved to obtain the second downsampled output feature. The second downsampled output feature is integrated with the first upsampled output feature and the integration result of the first downsampled output feature and the second upsampled output feature to obtain the second integrated result. The second integrated result is convolved and then integrated with the second downsampled result to obtain the third integrated result. The convolution result of the third integrated result is used as the third downsampled output feature.

[0011] According to certain embodiments of the first aspect of this application, the step of performing convolution processing on each fourth feature through multiple convolutional paths to obtain multiple fifth features, and fusing the multiple fifth features to obtain a sixth feature, includes: The fourth feature is processed by convolution through the first convolution path to obtain the first fifth feature. The first convolution path includes a 1x1 convolutional layer, a 3x3 convolutional layer and a 1x1 convolutional layer connected in sequence. The fourth feature is processed by convolution through the second convolution path to obtain the second fifth feature. The second convolution path includes a 1x1 convolutional layer, a 5x5 convolutional layer and a 1x1 convolutional layer connected in sequence. The fourth feature is processed by convolution through the third convolution path to obtain the third fifth feature. The third convolution path includes a 1x1 convolutional layer, a 7x7 convolutional layer and a 1x1 convolutional layer connected in sequence. The fourth feature is processed by convolution through the fourth convolution path to obtain the fourth fifth feature. The fourth convolution path includes a 1x1 convolutional layer. The first, second, third, and fourth fifth features are fused and then passed through a 1x1 convolutional layer and a 3x3 convolutional layer to obtain the sixth feature.

[0012] According to certain embodiments of the first aspect of this application, the step of detecting based on the sixth feature to obtain a metal anomaly detection result for the power tower includes: The sixth feature is subjected to dilated convolution to obtain dilated convolution features; The sixth feature is subjected to depthwise separable convolution to obtain depthwise separable convolution features; The fusion results of the hollow convolution feature and the depth-separable convolution feature are subjected to pointwise convolution and standard convolution to obtain the metal anomaly detection results of the power tower.

[0013] According to a second aspect of this application, an electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the UAV-based method for detecting metal anomalies in power towers as described in the first aspect of this application.

[0014] According to a third aspect of this application, a computer storage medium stores computer-executable instructions for performing the UAV-based method for detecting metal anomalies in power towers as described in an embodiment of the first aspect of this application.

[0015] The above scheme has at least the following beneficial effects: First features are obtained by inputting a visible light image into a first feature extraction network, and second features are obtained by inputting an infrared image into a second feature extraction network; third features are obtained by cross-splicing the first and second features and performing depthwise separable convolution; multiple third features at different scales are upsampled along an upsampling path and downsampled along a downsampling path to obtain multiple fourth features at different scales; multiple fifth features are obtained by convolution processing each fourth feature through multiple convolution paths, and multiple fifth features are fused to obtain a sixth feature; detection is performed based on the sixth feature to obtain the power tower metal anomaly detection result; it can balance accuracy, efficiency, lightweight design, and generalization ability. Attached Figure Description

[0016] The accompanying drawings are used to provide a further understanding of the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application and do not constitute a limitation on the technical solutions of this application.

[0017] Figure 1 This is a flowchart illustrating the steps of a method for detecting metal anomalies in power towers based on unmanned aerial vehicles (UAVs). Figure 2 This is a schematic diagram of the power tower monitoring points; Figure 3 This is a schematic diagram of the power tower monitoring points; Figure 4 This is the front view of the power tower monitoring point after multi-view mapping; Figure 5 This is a schematic diagram of the planned route; Figure 6 This is a structural diagram of the power tower metal anomaly detection model; Figure 7 This is a structural diagram of the feature fusion module; Figure 8 This is a structural diagram of the pyramid feature aggregation module; Figure 9 This is a structural diagram of the feature perception module. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, or the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0020] The embodiments of this application will be further described below with reference to the accompanying drawings.

[0021] The embodiments of this application provide a method and related apparatus for detecting metal anomalies in power towers based on unmanned aerial vehicles (UAVs).

[0022] Reference Figure 1 Methods for detecting metal anomalies in power towers include, but are not limited to, the following steps: Step S100: Obtain an image of the power tower; Step S200: Input the visible light image into the first feature extraction network to extract the first feature, and input the infrared image into the second feature extraction network to extract the second feature; Step S300: The first feature and the second feature are cross-stitched and depthwise separable convolutional to obtain the third feature; Step S400: Upsample multiple third features at different scales along the upsampling path and downsample them along the downsampling path to obtain multiple fourth features at different scales. Step S500: Perform convolution processing on each fourth feature through multiple convolution paths to obtain multiple fifth features, and fuse the multiple fifth features to obtain the sixth feature; Step S600: Detection is performed based on the sixth feature to obtain the detection result of metal anomaly of the power tower.

[0023] For step S100, the image of the power tower is obtained by taking pictures with a camera mounted on a drone, which flies along a planned path.

[0024] The planned path is obtained as follows: Preset power tower monitoring points are obtained; a first path is generated based on the power tower monitoring points; the positions of any number of power tower monitoring points in the first path are swapped to generate a second path; when the total length of the second path is less than the total length of the first path, the first path is updated to the second path; when the total length of the second path is greater than the total length of the first path, a probability value is obtained based on the current temperature coefficient, and the first path is updated to the second path based on the probability value, while the first path remains unchanged based on the difference between the probability values ​​and the first path; the current temperature coefficient is updated based on a preset cooling coefficient; if the current temperature coefficient is less than a preset temperature coefficient threshold or reaches a preset iteration number threshold, the current first path is used as the planned path.

[0025] Specifically, refer to Figure 2 The red dots represent the interactively customized tower monitoring points. Through interactive calibration, professional knowledge can be integrated into the calibration process, ensuring that inspection personnel fully consider their expertise during calibration, thereby obtaining more accurate and reasonable results.

[0026] Reference Figure 3 In the mapping diagram, the yellow dot represents the power tower monitoring point T; the green dot represents point M0, which is the point with the minimum distance from the mapping circular curve to the power tower monitoring point; the blue dots represent M1 and M2, which are the multi-view mapping solutions of the power tower monitoring point. To ensure that the vector formed by the two mapping points and the monitoring point is uniformly distributed in the monitoring point space, the included angle between the vectors is 120 degrees.

[0027] When monitoring power tower monitoring points, single-view mapping cannot display the three-dimensional depth of objects, making it difficult to understand the spatial relationships of complex structures. Furthermore, a single viewpoint cannot display all the details of an object, especially occluded parts. (Refer to...) Figure 4 Based on the front view after multi-view mapping, all mapping points are located in the safety inspection surface to ensure a safe distance between the drone and the power transmission tower.

[0028] Reference Figure 5 A multi-view mapping strategy was used to map power tower monitoring points to UAV inspection points. Taking multi-view mapping as an example, the path planning problem for UAV inspection of power towers was solved by minimizing the total flight distance of the UAV traversing all mapped points. If there are n power tower monitoring points on the power tower surface, the objective function is expressed as: ;in, ; The distance is the Euclidean distance. A first path is generated based on the monitoring points of the power towers. A second path is generated by swapping the positions of any two monitoring points in the first path. The total length of the paths is calculated based on the objective function. If the total length of the second path is less than the total length of the first path, the first path is updated to the second path. If the total length of the second path is greater than the total length of the first path, a probability value is obtained based on the length difference between the two paths and the current temperature coefficient. The probability value is expressed as: The first path is updated to the second path based on the probability value, while the first path remains unchanged based on the difference between the probability values ​​and the first path. The current temperature coefficient is updated according to the preset cooling coefficient, which is the product of the previous temperature coefficient and the cooling coefficient. If the current temperature coefficient is less than the preset temperature coefficient threshold or reaches the preset iteration number threshold, the current first path is used as the planned path.

[0029] The images of the power tower include both visible light and infrared images. A drone equipped with a dual-light pod (visible light + infrared thermal imager) flies along a pre-set route. Visible light acquisition uses a high-resolution camera (e.g., ≥24 megapixels), adjusting shooting parameters for different monitoring points within a range of 5 to 10 meters from the target: multi-angle close-ups (subject body ≥60%) are used for details such as connection points to ensure small defects are identifiable. Infrared acquisition uses a device with a temperature difference sensitivity ≤0.5℃ to capture temperature anomalies and employs an automatic temperature gradient mode to avoid overexposure. The acquired data is transmitted back to the ground station via a 5G image transmission system. If blurry images or abnormal temperature data are detected, the system can automatically trigger a re-flight mechanism. All data is tagged with timestamps and the location of the power tower monitoring point for easy comparison with historical records.

[0030] Reference Figure 6 The image of the power tower is input into the power tower metal anomaly detection model.

[0031] For step S200, the visible light image is input into the first feature extraction network to extract the first feature, and the infrared image is input into the second feature extraction network to extract the second feature. The first feature extraction network and the second feature extraction network share weights.

[0032] Specifically, a visible light image and an infrared light image are input. The two H×W×3 dimensional images are input into a ResNet50 feature extraction network with shared weights. The feature extraction network F outputs the first feature in the H×W×C dimension and the second feature in the H×W×C dimension, respectively. C is determined by the number of convolutional kernels of the feature extraction network.

[0033] The first feature includes the first feature of the first layer, the first feature of the second layer, and the first feature of the third layer; the second feature includes the second feature of the first layer, the second feature of the second layer, and the second feature of the third layer. The first feature of the first layer and the second feature of the first layer are concatenated, the first feature of the second layer and the second feature of the second layer are concatenated, and the first feature of the third layer and the second feature of the third layer are concatenated. The three concatenated results are input into the feature fusion module.

[0034] For step S300, the first feature and the second feature are cross-stitched and depthwise separable convolutional to obtain the third feature through the feature fusion module.

[0035] Reference Figure 7Specifically, the feature fusion module starts with multi-branch input features and achieves fine-grained interaction and recombination of features in the channel dimension through channel-dimensional cross-mixing operations. It leverages the spatial detail mining capabilities of single-channel features through channel-level convolution, while simultaneously promoting complementary fusion of cross-channel features through the cross-mixing mechanism, enhancing the diversity and discriminative power of feature representation. The subsequent 1×1 point-wise convolution, based on the dimensional transformation characteristics of point-by-point convolution, efficiently completes linear mapping of the channel dimension and optimizes computational load while maintaining the integrity of feature spatial information, achieving reasonable control of feature dimensions. The entire module, through the collaborative design of hierarchical convolutional operations (channel-level interaction and point-by-point dimensional transformation), ensures deep information fusion of features in both channel and spatial dimensions, while balancing model complexity and feature representation capability through lightweight convolutional operations (1×1 PW), providing precisely enhanced and dimensionally adapted output features for subsequent tasks such as surface defect detection.

[0036] For step S400, the pyramid feature aggregation module upsamples multiple third features of different scales along the upsampling path and downsamples them along the downsampling path to obtain multiple fourth features of different scales.

[0037] Reference Figure 8 The method involves upsampling multiple third features at different scales along an upsampling path, including: convolving and upsampling the third third feature and integrating it with the second third feature to obtain a first integrated feature; convolving the result of convolving and upsampling the first integrated feature with the first third feature and the first integrated feature to obtain a third upsampled output feature; using the result of convolving the third third feature as the first upsampled output feature, and using the result of convolving the first integrated feature as the second upsampled output feature.

[0038] Multiple third features at different scales are downsampled along a downsampling path, including: convolving the third upsampled output features as the first downsampled output features; downsampling the convolved third upsampled output features to obtain the first downsampled result; integrating the first downsampled result with the second upsampled output features and convolving the result to obtain the second downsampled output features; integrating the second downsampled result of the second downsampled output features with the first upsampled output features and the integration result of the first downsampled result with the second upsampled output features to obtain the second integrated result; convolving the second integrated result and integrating it with the second downsampled result to obtain the third integrated result; and using the convolution result of the third integrated result as the third downsampled output features.

[0039] Specifically, the pyramid feature aggregation module constructs a multi-level feature interaction neural network architecture. Its core idea lies in effectively capturing and integrating cross-scale visual information through a hierarchical feature fusion mechanism. This architecture uses the C3-C5 level feature maps output by the feature fusion module as input. The data flow first fuses features from different levels through channel concatenation, and then extracts the deep semantic information contained therein through 3×3 convolutional layers. Subsequently, its pyramid decomposition strategy decouples the feature map into multi-scale sub-maps, while the embedded attention mechanism dynamically allocates weights by modeling the spatial correlation between pixels within the feature map. This collaborative design can simultaneously model cross-scale contextual relationships and enhance the detailed representation ability of key regions. Cross-level feature integration and concatenation are achieved through channel dimension stacking, complementing the bidirectional feature propagation path. The top-down upsampling path gradually restores spatial resolution using transposed convolution, while the bottom-up downsampling path compresses feature dimensions through stride convolution. This bidirectional information flow constructs a closed-loop multi-scale feature interaction system, ensuring the full integration of high-level semantics and low-level details. In particular, skip connections directly inject shallow features rich in low-level texture information from the original backbone network into high-level semantic features, significantly mitigating the inherent spatial detail loss problem during upsampling. The final output P3-P5 feature pyramid achieves standardized processing of heterogeneous features through a unified dimension, maintaining the high-level semantic abstraction capability of the P5 layer while enabling the P3 layer to accurately represent the spatial structural information of the target (e.g., precise localization of object contours in instance segmentation tasks). This architecture design, which balances global semantic understanding with local detail awareness, demonstrates significant advantages in tasks requiring fine-grained localization and semantic understanding in collaboration, such as surface defect detection.

[0040] Reference Figure 9 For step S500, the feature perception module performs convolution processing on each fourth feature through multiple convolutional paths to obtain multiple fifth features, and the multiple fifth features are fused to obtain the sixth feature, including the following steps: The fourth feature is processed by convolution through the first convolution path to obtain the first fifth feature. The first convolution path includes a 1x1 convolutional layer, a 3x3 convolutional layer and a 1x1 convolutional layer connected in sequence. The fourth feature is processed by convolution through the second convolution path to obtain the second fifth feature. The second convolution path includes a 1x1 convolutional layer, a 5x5 convolutional layer and a 1x1 convolutional layer connected in sequence. The fourth feature is processed by convolution through the third convolution path to obtain the third fifth feature. The third convolution path includes a 1x1 convolutional layer, a 7x7 convolutional layer, and a 1x1 convolutional layer connected in sequence. The fourth feature is processed by convolution through the fourth convolution path to obtain the fourth fifth feature. The fourth convolution path includes a 1x1 convolutional layer. The first, second, third, and fourth fifth features are fused and then passed through a 1x1 convolutional layer and a 3x3 convolutional layer to obtain the sixth feature.

[0041] Specifically, the feature perception module employs a multi-branch, multi-scale convolutional feature extraction and fusion module architecture. It achieves efficient integration of multi-scale semantic and detailed information of the input features through parallel extraction of features from multiple receptive fields. The input feature map is processed through four parallel branches: the first three branches follow a sequence logic of "1×1 convolution - large-size convolutional kernel (3×3 / 5×5 / 7×7) - 1×1 convolution". The core function of the initial 1×1 convolution is channel dimension reduction (reducing the number of feature channels through linear transformation) to reduce the computational cost of subsequent large convolutional kernel operations, balancing feature expressiveness and computational cost. The convolutional kernels of different sizes in the middle (3×3 for small receptive fields, 5×5 and 7×7 for larger receptive fields) are used for multi-scale feature capture. Small convolutional kernels focus on local texture details, while large convolutional kernels capture the global contextual structure. The final 1×1 convolution is responsible for channel dimension recovery, ensuring the consistency of the output feature dimensions of each branch, preparing for subsequent fusion. The fourth branch retains only 1×1 convolutions to supplement lightweight feature extraction at a single scale, complementing the first three branches to form a "multi-scale + lightweight" feature representation. The feature maps output from each branch are concatenated along the channel dimension. This operation directly superimposes the multi-scale features from multiple branches along the channel dimension, maximizing the preservation of the feature specificity of each branch and forming a multi-scale feature collection. The subsequent 1×1 convolution + 3×3 convolution on the right is a secondary fusion and refinement of the concatenated features: the 1×1 convolution is used for secondary adjustment of the channel dimension, while the 3×3 convolution further integrates the correlation of multi-scale features through local spatial convolution. The fused sixth feature provides a feature representation with both multi-scale information and discriminative power for downstream detection tasks. Overall, this architecture avoids the limitations of a single convolutional kernel on receptive field and computational cost, and strengthens the network's ability to represent targets in complex scenes through the complementary integration of multi-scale features.

[0042] For step S600, the high-efficiency dual-source collaborative detection head performs detection based on the sixth feature to obtain the power tower metal anomaly detection result, including the following steps: performing dilated convolution on the sixth feature to obtain dilated convolution feature; performing depthwise separable convolution on the sixth feature to obtain depthwise separable convolution feature; and performing pointwise convolution and standard convolution on the fusion result of dilated convolution feature and depthwise separable convolution feature to obtain the power tower metal anomaly detection result.

[0043] Specifically, the dual-source collaborative architecture employed by this object detection head embodies the collaborative optimization principle of feature reuse and task decoupling in deep neural networks at the fundamental level. Its symmetrical parallel structure begins with two sets of parameter-shared 3×3 standard convolutional layers and dilated convolutional layers, adhering to the local receptive characteristics and parameter efficiency principles of convolutional neural networks: standard convolutions capture spatial semantic features through local receptive fields, while dilated convolutions systematically expand the receptive field while maintaining the feature map resolution. Together, they achieve cross-scale contextual information fusion. This architecture effectively avoids spatial information loss caused by traditional downsampling operations, and is particularly suitable for the core requirement of feature fidelity in small object detection. Four consecutive convolutional operations constitute a nonlinear feature enhancement network, achieving high-order feature interaction through deeply stacked convolutional layers. Within the framework of cascaded nonlinear transformation theory of convolutional neural networks, this process gradually constructs discriminative hierarchical representations, while maintaining gradient stability through inter-layer normalization and residual connections, avoiding the problem of blind expansion of the receptive field caused by increased network depth.

[0044] After feature processing, the network splits into a dual-branch structure for bounding box regression and object classification, reflecting the parameter decoupling concept in multi-task learning. In the regression branch, 1×1 convolutions expand the number of channels to 4×reg_max dimension, transforming continuous coordinate prediction into discrete distribution integrals, thereby enhancing localization accuracy by improving the representational power of regression parameters. By focusing on the classification task, the number of channels output by its 1×1 convolution corresponds to the total number of target categories, directly establishing a category response mapping in the feature space. The two branches are respectively connected to Bbox Loss and Cls Loss modules, using task-specific loss functions for end-to-end optimization: the regression branch refines the geometric matching degree between the predicted box and the ground truth box through DIoU loss or Gaussian distribution loss, while the classification branch uses Focal Loss or cross-entropy loss to balance the weights of positive and negative samples. The entire architecture achieves general feature reuse through parameter sharing of the basal convolutional layers, significantly reducing model complexity; while the task-specific convolutional layers give the two branches independent feature transformation spaces, ensuring that the geometrically sensitive features required for localization and the semantic discriminative features required for classification do not interfere with each other, enabling the real-time object detection system to achieve both accuracy and efficiency.

[0045] This method for detecting metal anomalies in power towers combines a multimodal hierarchical detection network, employing both visible and infrared light modes. An embedded channel attention mechanism enhances the sensitivity of metal corrosion areas, while a feature selection mechanism maintains detection accuracy. It supports real-time acquisition of optical images of the thin film surface, which are pre-processed and input into the model for defect localization and classification. The detection results are then fed back to the quality inspection terminal. While ensuring detection accuracy, this method effectively saves significant manpower and resources, achieving better work results.

[0046] Embodiments of this application provide an electronic device. The electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the UAV-based method for detecting metal anomalies in power towers as described above.

[0047] This electronic device can be any smart terminal, including computers.

[0048] In general, for the hardware structure of electronic devices, the processor can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, to execute relevant programs and implement the technical solutions provided in the embodiments of this application.

[0049] The memory can be implemented in the form of read-only memory (ROM), static storage device, dynamic storage device, or random access memory (RAM). The memory can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory and is called and executed by the processor.

[0050] Input / output interfaces are used to implement information input and output.

[0051] The communication interface is used to enable communication and interaction between this device and other devices. Communication can be achieved through wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0052] The bus transmits information between various components of a device, such as the processor, memory, input / output interfaces, and communication interfaces. The processor, memory, input / output interfaces, and communication interfaces communicate with each other within the device via the bus.

[0053] Embodiments of this application provide a computer storage medium. The computer storage medium stores computer-executable instructions for performing the UAV-based power tower metal anomaly detection method described above.

[0054] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium. In the foregoing description of this specification, references to terms such as "one embodiment," "another embodiment," or "some embodiments," etc., indicate that a specific feature, structure, material, or characteristic described in connection with an embodiment or example is included in at least one embodiment or example of this application. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0055] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0056] The units described above 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.

[0057] 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.

[0058] 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 multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0059] 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 the units described above 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 between each other may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms. Although embodiments of this application have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of this application, the scope of which is defined by the claims and their equivalents.

[0060] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for detecting metal anomaly of an electric tower based on a UAV, characterized in that, The method comprises the following steps: acquiring an electric tower image, the electric tower image comprising a visible light image and an infrared image; inputting the visible light image into a first feature extraction network to extract a first feature, and inputting the infrared image into a second feature extraction network to extract a second feature, the first feature extraction network and the second feature extraction network sharing weights; performing a cross-stitch operation and a depth separable convolution operation on the first feature and the second feature to obtain a third feature; up-sampling and down-sampling a plurality of third features of different scales along an up-sampling path and a down-sampling path to obtain a plurality of fourth features of different scales; performing convolution processing on each fourth feature along a plurality of convolution paths to obtain a plurality of fifth features, and fusing the plurality of fifth features to obtain a sixth feature; detecting according to the sixth feature to obtain an electric tower metal anomaly detection result. 2.The unmanned aerial vehicle based electric tower metal anomaly detection method according to claim 1, characterized in that, The electric tower image is obtained by a shooting device carried on a UAV, and the UAV flies according to a planned path; The planned path is obtained by the following method: acquiring a preset electric tower monitoring point; generating a first path according to the electric tower monitoring point; swapping the positions of any plurality of electric tower monitoring points in the first path to generate a second path; when the total length of the second path is less than the total length of the first path, updating the first path to the second path; when the total length of the second path is greater than the total length of the first path, obtaining a probability value according to a current temperature coefficient, performing updating the first path to the second path according to the probability value, and performing keeping the first path unchanged according to a difference between one and the probability value; updating the current temperature coefficient according to a preset temperature reduction coefficient; if the current temperature coefficient is less than a preset temperature coefficient threshold or reaches a preset iteration number threshold, taking the current first path as the planned path. 3.The unmanned aerial vehicle based electric tower metal anomaly detection method of claim 2, wherein, The electric tower image is labeled with a time label and a position label of an electric tower monitoring point. 4.The unmanned aerial vehicle based electric tower metal anomaly detection method of claim 1, wherein, Further comprising: decoupling the fourth feature into multi-scale subgraphs; performing an attention mechanism on the multi-scale subgraphs. 5.The unmanned aerial vehicle based electric tower metal anomaly detection method of claim 1, wherein, The up-sampling of the plurality of third features of different scales along the up-sampling path comprises: integrating the third third feature after convolution and up-sampling with the second third feature to obtain a first integrated feature; performing convolution on the result of convolution and up-sampling of the first integrated feature with the first third feature and the first integrated feature to obtain a third up-sampling output feature; taking the result of convolution of the third third feature as a first up-sampling output feature, and taking the result of convolution of the first integrated feature as a second up-sampling output feature. 6.The unmanned aerial vehicle based electric tower metal anomaly detection method of claim 5, wherein, The down-sampling of the plurality of third features of different scales along the down-sampling path comprises: taking the result of convolution of the third up-sampling output feature as a first down-sampling output feature; performing down-sampling on the result of convolution of the third up-sampling output feature to obtain a first down-sampling result, and taking the result of integration and convolution of the first down-sampling result with the second up-sampling output feature as a second down-sampling output feature; The second down-sampling result of the second down-sampling output feature is integrated with the integration result of the first up-sampling output feature and the first down-sampling result and the second integration result is obtained, the third integration result is obtained by integrating the convolution result of the second integration result and the second down-sampling result, and the convolution result of the third integration result is taken as the third down-sampling output feature. 7.The unmanned aerial vehicle based electric tower metal anomaly detection method of claim 1, wherein, The fourth feature is convoluted through a plurality of convolution paths to obtain a plurality of fifth features, and the sixth feature is obtained by fusing the plurality of fifth features, comprising: The fourth feature is convoluted through a first convolution path to obtain a first fifth feature, and the first convolution path comprises a 1x1 convolution layer, a 3x3 convolution layer and a 1x1 convolution layer connected in sequence; The fourth feature is convoluted through a second convolution path to obtain a second fifth feature, and the second convolution path comprises a 1x1 convolution layer, a 5x5 convolution layer and a 1x1 convolution layer connected in sequence; The fourth feature is convoluted through a third convolution path to obtain a third fifth feature, and the third convolution path comprises a 1x1 convolution layer, a 7x7 convolution layer and a 1x1 convolution layer connected in sequence; The fourth feature is convoluted through a fourth convolution path to obtain a fourth fifth feature, and the fourth convolution path comprises a 1x1 convolution layer; The first fifth feature, the second fifth feature, the third fifth feature and the fourth fifth feature are fused through a 1x1 convolution layer and a 3x3 convolution layer to obtain the sixth feature. 8.The unmanned aerial vehicle based electric tower metal anomaly detection method of claim 1, wherein, The sixth feature is convoluted to obtain a hollow convolution feature; The sixth feature is depth separable convoluted to obtain a depth separable convolution feature; The hollow convolution feature and the depth separable convolution feature are fused and point-by-point convoluted and standard convoluted to obtain the electric tower metal anomaly detection result. Comprising:

9. An electronic device, comprising: A memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the unmanned aerial vehicle-based electric tower metal anomaly detection method according to any one of claims 1 to 8. Computer executable instructions are stored, and the computer executable instructions are used to execute the unmanned aerial vehicle-based electric tower metal anomaly detection method according to any one of claims 1 to 8.

10. A computer storage medium, characterized in that, ​

Images