Anti-motion blur unmanned aerial vehicle power line strand breakage real-time identification method and system

Abstract

The invention discloses an anti-motion blur unmanned aerial vehicle power line broken strand real-time identification method and system, and belongs to the technical field of image processing, and the method comprises the steps: obtaining a power line image shot by an unmanned aerial vehicle as an input image; performing block processing on the input image to generate a plurality of to-be-processed image blocks; performing selective filtering on the to-be-processed image block to obtain an enhanced image block; extracting the enhanced image block to obtain a shallow semantic feature map and a deep semantic feature map; fusing the shallow semantic feature map and the deep semantic feature map to generate multi-scale fusion features; and based on the multi-scale fusion features, identifying the broken strand position of the power line and outputting broken strand position information. According to the method, a selective filtering technology is adopted, and block processing and a multi-scale feature fusion mechanism are combined, so that motion blur interference can be effectively resisted, and high-precision recognition of the broken strand of the power line is realized.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for real-time identification of broken strands in power lines of unmanned aerial vehicles (UAVs) that resists motion blur. Background Technology

[0002] Unmanned aerial vehicle (UAV) power line inspection is an important means of ensuring the safe and stable operation of the power grid. A key technology involved is the automatic identification of various defects in power lines and their associated components using image processing techniques from the large amounts of aerial images or video streams captured by UAVs. Broken strands in power lines are a typical and highly dangerous defect, and their real-time and accurate identification is crucial for preventing line faults and ensuring power supply safety. In the field of image processing technology, defect detection of such slender targets has always been a research hotspot and a challenge.

[0003] Currently, computer vision-based methods for identifying broken strands in power lines primarily rely on deep learning models, especially convolutional neural networks. These methods typically input images captured by drones directly into object detection or image segmentation networks, allowing the network to learn the visual features of the broken strand defect and thus locate and identify the broken strand region in the image. Some methods employ general image enhancement algorithms, such as global histogram equalization or sharpening filtering, to preprocess the input image in order to improve the model's recognition performance.

[0004] However, existing technical solutions have significant shortcomings in practical applications. During drone inspections, motion is inevitable, resulting in directional motion blur in the captured images. This blur severely damages the fine edge information at broken strands in power lines, causing a sharp decline in the recognition performance of deep learning models that rely on visual features. Conventional image enhancement methods are usually global and cannot specifically address directional blur; they may even amplify background noise, interfering with recognition. Furthermore, most existing recognition models remain at the pixel pattern matching level, lacking an understanding of the structural continuity of power lines. This makes them prone to misclassifying objects in the background that resemble broken strands as defects, resulting in poor robustness. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method and system for real-time identification of broken power line strands in unmanned aerial vehicles (UAVs) that resists motion ambiguity. By employing selective filtering technology combined with block processing and multi-scale feature fusion mechanisms, it can effectively resist motion ambiguity interference and achieve high-precision identification of broken power line strands.

[0006] The above objectives can be achieved through the following approach: A method and system for real-time identification of broken strands in power lines using a drone to resist motion blur includes: acquiring power line images captured by a drone as input images; segmenting the input images into blocks to generate multiple image blocks to be processed; selectively filtering the image blocks to be processed to obtain enhanced image blocks; extracting shallow semantic feature maps and deep semantic feature maps from the enhanced image blocks; fusing the shallow and deep semantic feature maps to generate multi-scale fusion features; and identifying the location of broken strands in the power line based on the multi-scale fusion features and outputting the broken strand location information.

[0007] Optionally, the step of selectively filtering the image block to be processed to obtain an enhanced image block includes: analyzing the gradient distribution of the image block to be processed to generate a blurred direction-aware result; obtaining direction-aware filtering parameters based on the blurred direction-aware result; and performing adaptive filtering on the image block to be processed based on the direction-aware filtering parameters to output an enhanced image block.

[0008] Optionally, the step of analyzing the gradient distribution of the image patch to be processed and generating a fuzzy direction perception result includes: obtaining a gradient direction field structure based on the gradient distribution of the image patch to be processed; and performing fuzzy direction intensity mapping on the gradient direction field structure to generate a fuzzy direction perception result.

[0009] Optionally, the step of extracting the enhanced image patch to obtain a shallow semantic feature map and a deep semantic feature map includes: extracting features from the enhanced image patch and outputting an initial feature map; inputting the initial feature map into a preset shallow branch network and a deep branch network respectively, and generating shallow semantic feature maps and deep semantic feature maps in parallel.

[0010] Optionally, the step of extracting features from the enhanced image patch and outputting an initial feature map includes: extracting local directional feature responses from the enhanced image patch to obtain a multi-channel directional feature tensor; performing directional clustering compression on the multi-channel directional feature tensor to output an initial feature map.

[0011] Optionally, identifying the location of a broken strand in a power line based on the multi-scale fusion features and outputting the broken strand location information includes: performing geometric feature structure topology modeling on the multi-scale fusion features to obtain a power line geometric topology tensor; performing eigenvalue decomposition and vector field divergence analysis on the power line geometric topology tensor to identify the location of a broken strand in the power line and outputting the pixel coordinate matrix of the broken strand location.

[0012] Optionally, the step of segmenting the input image to generate multiple image blocks to be processed includes: performing local structural sharpness assessment on the input image to obtain a power line region mask and discontinuous segment boundaries; and generating multiple image blocks to be processed based on the power line region mask and discontinuous segment boundaries.

[0013] Optionally, generating multiple image blocks to be processed based on the power line region mask and the discontinuous block boundaries includes: performing structural constraint topological subdivision on the power line region mask to obtain a fault-aware topological sub-map; calculating the geometric fracture probability field on the discontinuous block boundaries to generate a boundary fracture weight grid; and generating multiple image blocks to be processed based on the fault-aware topological sub-map and the boundary fracture weight grid.

[0014] Optionally, the step of performing geometric feature structure topology modeling on the multi-scale fused features to obtain the electric line geometric topology tensor includes: obtaining a set of electric line structure feature points based on the local geometric features of the multi-scale fused features; using the set of electric line structure feature points as nodes, establishing connecting edges based on spatial proximity and feature similarity to obtain a electric line topology graph; and performing vector synthesis on the electric line topology graph to obtain the electric line geometric topology tensor.

[0015] Based on the same inventive concept, this invention also provides a real-time identification system for broken strands of power lines by unmanned aerial vehicles (UAVs) to resist motion blur. The system includes: an image acquisition module for acquiring images of power lines captured by a UAV as input images; a block processing module for segmenting the input image into blocks to generate multiple image blocks to be processed; a selective filtering module for selectively filtering the image blocks to be processed to obtain enhanced image blocks; a feature extraction module for extracting features from the enhanced image blocks to obtain shallow semantic feature maps and deep semantic feature maps; a multi-scale feature fusion module for fusing the shallow and deep semantic feature maps to generate multi-scale fused features; and a broken strand location identification module for identifying the location of broken strands in the power line based on the multi-scale fused features and outputting the broken strand location information.

[0016] Compared with the prior art, the present invention has the following advantages: This invention effectively overcomes the damage to image quality caused by motion blur in UAV aerial photography by introducing a preprocessing mechanism that combines intelligent block segmentation with adaptive filtering. This method can sense and selectively enhance image areas where details are unclear due to motion blur, restoring clear outlines of power lines while avoiding over-sharpening and noise amplification of non-blurred areas. This ensures the quality of the input data from the source, laying a solid foundation for subsequent accurate recognition.

[0017] This invention employs a parallel dual-branch feature extraction network and a multi-scale feature fusion strategy to achieve comprehensive and deep perception of power line features. The shallow branch preserves the fine geometric edge details at the broken strand defects, while the deep branch extracts the overall trend and contextual semantic information of the power line. The effective fusion of these two approaches allows the system to simultaneously ensure the accuracy of local details and the robustness of global semantics, thus accurately distinguishing between real broken strands and artifact interference even in complex contexts.

[0018] This invention innovatively elevates the identification task from pixel-level pattern matching to the level of structural topology analysis. By constructing a geometric topological tensor of the electric field line and performing fracture feature decomposition on it, this method can identify breakpoints from the perspective of assessing structural integrity. This method does not rely on the fixed appearance of broken strands, but rather focuses on the structural discontinuities they cause. Therefore, it has strong adaptability and generalization ability to various forms of broken strands, partial occlusion, and changes in illumination, significantly improving the reliability of identification.

[0019] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating a method for real-time identification of broken strands in unmanned aerial vehicle power lines that resists motion fuzzing, according to an embodiment of the present invention.

[0022] Figure 2 This is a multi-scale fusion feature map according to an embodiment of the present invention.

[0023] Figure 3 This is a gradient direction field structure diagram of the image block to be processed according to an embodiment of the present invention.

[0024] Figure 4 This is a feature extraction result diagram of an embodiment of the present invention.

[0025] Figure 5 This is a diagram for identifying broken strands in power lines according to an embodiment of the present invention.

[0026] Figure 6This is a schematic diagram of the structure of a real-time identification system for broken strands of power lines in a drone that resists motion fuzzing, according to an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Reference Figure 1 One embodiment of the present invention proposes a real-time identification method for broken power line strands in unmanned aerial vehicles (UAVs) that resists motion fuzzing. It employs selective filtering technology, combined with block processing and multi-scale feature fusion mechanism, which can effectively resist motion fuzzing interference and achieve high-precision identification of broken power line strands.

[0029] The method described in this embodiment specifically includes: The image of the power lines captured by the drone is used as the input image; The input image is divided into blocks to generate multiple image blocks to be processed; Selective filtering is performed on the image block to be processed to obtain an enhanced image block; The enhanced image patch is extracted to obtain a shallow semantic feature map and a deep semantic feature map; By fusing the shallow semantic feature map and the deep semantic feature map, a multi-scale fused feature is generated; Based on the multi-scale fusion features, the location of broken strands in the power line is identified and the broken strand location information is output.

[0030] Specifically, firstly, an intelligent segmentation strategy is used to decompose the original UAV aerial image into multiple image blocks containing complete electric field line structures, avoiding the loss of key information during segmentation. Next, addressing the image blurring problem often caused by UAV motion, the method employs a selective filtering technique. This technique can perceive the blur direction of each image block and apply directional sharpening, effectively restoring the clarity of the electric field lines lost due to motion blur without damaging other details. After obtaining the enhanced image blocks, the system uses a parallel dual-branch network structure to simultaneously extract shallow features representing fine geometric details and deep semantic feature maps containing global structural information. These two complementary features are effectively fused to generate accurate and robust multi-scale fused features, such as... Figure 2As shown. The specific fusion process is as follows: First, the shallow semantic feature map (size H×W×C1) and the deep semantic feature map (size H×W×C2) are concatenated in the channel dimension to form a feature map of size H×W×(C1+C2), where H and W are the height and width of the feature map, and C1 and C2 are the number of channels of the shallow and deep features, respectively. The concatenated feature map is processed by a 1×1 convolutional layer (with 1 kernel) and a sigmoid activation function to generate an attention weight map of size H×W×1. The weight values ​​range from [0,1], and the higher the value, the more critical the feature at that position is to the segment breakage identification. The shallow semantic feature map and the attention weight map are multiplied element-wise to obtain a weighted shallow feature (H×W×C1). The deep semantic feature map is multiplied element-wise with (1-attention weight map) to obtain a weighted deep feature (H×W×C2). The features are added together while ensuring that the number of channels is the same to obtain a feature map with the number of channels C1 (or C2) (H×W×C1, where C1=C2). This method transforms the identification problem from traditional pixel pattern matching into an analysis of the integrity of the power line topology. It constructs a geometric model of the power line based on multi-scale fusion features, and identifies discontinuities in the structure by analyzing the intrinsic connectivity of the model, ultimately accurately outputting the location of the broken strand.

[0031] Optionally, the step of selectively filtering the image patch to be processed to obtain an enhanced image patch includes: Analyze the gradient distribution of the image block to be processed to generate a blurred direction perception result; Based on the fuzzy direction perception results, the direction perception filtering parameters are obtained; Based on the direction-aware filtering parameters, the image block to be processed is subjected to adaptive filtering processing, and an enhanced image block is output.

[0032] Specifically, to achieve targeted blur removal, the system first needs to accurately perceive the direction and degree of blur. This process involves analyzing the gradient distribution of the image patch to be processed and generating a blur direction perception result. The gradient direction field structure of the image patch to be processed is as follows: Figure 3 As shown. Specifically, the system uses gradient operators, such as the Sobel or Scharr operators, to calculate the gradient magnitude and direction of each pixel in the image block to be processed, thereby constructing a complete gradient direction field structure. In images with motion blur, the gradient directions of most edges will be concentrated in the direction perpendicular to the blurred motion trajectory. Based on this physical characteristic, the system performs blur direction intensity mapping on the gradient direction field structure and analyzes its distribution by statistically plotting the histogram of gradient directions. The angle corresponding to the peak in the histogram is the most dominant gradient direction, denoted as . Since the principal gradient direction is perpendicular to the blur direction, the core parameter of motion blur, namely the blur direction angle, can be calculated. Fuzzy direction angle It can be obtained through the following relation: , Here, The principal gradient direction angle is obtained through gradient direction histogram statistics. This fuzzy direction angle... Together with the fuzziness intensity, they constitute the fuzzy direction perception result.

[0033] Secondly, based on the acquired fuzzy direction perception results, the system generates corresponding direction perception filtering parameters.

[0034] The process includes the following steps: fuzzy parameter quantization; fuzzy direction angle. Added from the principal gradient direction Obtain, standardize to ; fuzziness intensity Quantification is achieved by using the peak height of the gradient histogram and the half-width ratio: , The calibration coefficient, ranging from 1.2 to 1.8, was used for UAV inspection images (resolution 2048×1080 to 4096×2160, motion blur equivalent to 3-20 pixel displacement). Calibration was performed using 1000 sample images labeled with the actual blurred pixel length: For each image, the following calculation was performed. (Gradient histogram peak height / half width at half maximum), take all samples The mean was used as the final value (the experimental statistical mean was 1.5, which can be used directly). The peak height of the gradient direction histogram is represented by W, where W represents the full width at half maximum (FWHM) of the histogram peak; Oriented PSF construction: along Direction generation Matrix, pixel values ​​on line segments (Normalized), all others are 0; Wiener filter parameters: obtained through PSF Fourier transform Noise power spectrum Signal power spectrum Calculate the transfer function: , Used for image patch enhancement, where Represents frequency coordinates. The Fourier transform of the motion blur point spread function (PSF) reflects the characteristics of motion blur in the frequency domain (determined by the blur direction angle and blur intensity). express The conjugate complex number is used to correct phase information in the frequency domain, ensuring phase consistency during the filtering process. The power spectrum of the signal (effective information of electric field lines) in the image block reflects the energy distribution of the signal in the frequency domain. Using the electric field line region mask, the pixel region within the mask (considered as the region containing the effective signal) is extracted, and a Fourier transform is performed on the region. After calculating its power spectrum, Gaussian smoothing (window 3×3) is performed to obtain the signal power spectrum. The power spectrum of noise in the image patch reflects the energy distribution of noise in the frequency domain. The background area outside the mask (a region with no electric field lines and a flat gray level) is selected, and Fourier transform is performed on the region. The power spectrum is calculated and the mean is taken as the noise power spectrum (the background area is automatically identified through connected component analysis, and the area is not less than 50×50 pixels). express The squared modulus, or power spectrum of the PSF, is used to quantify the degree of attenuation of the frequency domain signal by the blur. These parameters are designed to construct a filter kernel specifically for canceling blur in a particular direction. This filter kernel, such as the inverse model of a linear point spread function (PSF) or the parameters of a Wiener filter, has a shape directly determined by the blur direction angle. The direction angle is determined by the blur intensity. The spatial orientation of the filter kernel is determined, while the blur intensity determines its scale or length. Finally, using the generated orientation-aware filtering parameters, the system performs adaptive filtering on the original image patch to be processed. This process is essentially an image deconvolution, operating the image patch with an orientation-aware filter kernel customized based on its own blur characteristics. Because this filter is tailored to the specific blur direction and intensity of the current image patch, it can maximally recover the signal along the blur direction, sharpening the blurred power line edges while avoiding damage to image details in other directions. After the operation, the output is an enhanced image patch with clearer structure, free from motion blur.

[0035] Optionally, analyzing the gradient distribution of the image patch to be processed and generating a blurred direction-aware result includes: Based on the gradient distribution of the image patch to be processed, the gradient direction field structure is obtained; The gradient direction field structure is subjected to fuzzy direction intensity mapping to generate fuzzy direction perception results.

[0036] Specifically, firstly, based on the gradient distribution of the image block to be processed, the gradient direction field structure is obtained. The goal of this operation is to transform the pixel grayscale information of the image into structured data that can represent edge and texture directions. In specific implementation, the system uses gradient operators, such as the Sobel operator, to perform convolution operations on the input image block I to be processed. For each pixel in the image... Calculate the gradient components in the horizontal direction respectively. gradient components in the vertical direction Based on these two components, the gradient magnitude at that point can be further calculated. With gradient direction Gradient direction The calculation formula is as follows: , In this formula, For pixels gradient direction angle at that location, The function ensures that the angle calculation results cover the entire range from -π to π, thus uniquely determining the direction of the gradient vector. and They are pixels The vertical and horizontal gradient values ​​obtained through convolution. The gradient directions of all pixels. Together, they form a direction map with the same size as the original image patch, which is the gradient orientation field structure. This structure intuitively reflects the orientation distribution of all edges and details in the image. Then, the gradient orientation field structure is mapped to blur direction intensity to generate a blur direction perception result. The core idea of ​​this operation is to use the statistical characteristics of gradient directions in motion-blurred images to infer the blur attributes. In images with single-direction motion blur, the gradient directions of most sharp edges will be concentrated in the direction perpendicular to the blurred motion trajectory. Therefore, by statistically analyzing the gradient orientation field structure, the dominant gradient direction can be determined, and thus the blur direction can be inferred. Specifically, the system constructs a gradient orientation histogram H. To enhance the robustness of the results, only the gradient magnitude is statistically analyzed. Pixels with values ​​greater than a preset threshold are excluded to eliminate noise interference in flat areas. The construction of this histogram H can be represented as the accumulation of gradient magnitudes within a specific directional range. Subsequently, the peak with the highest count value is found in the histogram H; the angle corresponding to this peak is the principal gradient direction in the image patch, denoted as . Since the principal gradient direction is perpendicular to the fuzzy direction, the fuzzy direction angle is... It can be obtained through the following relationship: , Here This is a key parameter in the fuzzy direction perception result, precisely indicating the direction of motion blur. Simultaneously, the sharpness and height of the peaks in the histogram can serve as a measure of blur intensity, together forming a three-dimensional fuzzy direction perception result, providing a precise basis for subsequently generating direction perception filter parameters.

[0037] Optionally, the step of extracting the enhanced image patch to obtain a shallow semantic feature map and a deep semantic feature map includes: Extract the features of the enhanced image patch and output the initial feature map; The initial feature maps are input into preset shallow branch networks and deep branch networks respectively, and shallow semantic feature maps and deep semantic feature maps are generated in parallel.

[0038] Specifically, the first step is to extract features from the input enhanced image patch, outputting a shared initial feature map. The feature extraction result is as follows: Figure 4As shown. In practice, the enhanced image patch is input into a shared backbone network consisting of several convolutional layers. These initial convolutional layers are responsible for performing basic feature extraction tasks, transforming the raw pixel data into a multi-channel feature representation containing basic visual elements such as edges, corners, and textures. This feature representation output by the shared backbone network is the initial feature map. It serves as the common input to the two subsequent parallel branches, ensuring that the extraction of both features is based on the same fundamental information source, which is beneficial for subsequent fusion. Next, this initial feature map is simultaneously fed into two network branches with different structures and functions. The first is a shallow branch network. This network is characterized by a small number of layers and uses little or no downsampling operations, such as pooling layers or convolutions with a stride greater than 1. Its purpose is to preserve the spatial resolution of the feature map to the maximum extent. By performing a small number of convolutional operations in the shallow branch network, local details in the initial feature map can be further refined, such as the subtle edge contours of electric field lines, surface texture variations, and fracture morphology at broken strands. The output of this branch is the shallow semantic feature map, characterized by its large size and precise spatial location information, providing a basis for accurate localization of broken segments. The shallow branch network structure includes three convolutional layers. The first convolutional layer has a 3×3 kernel size, a stride of 1, and the same padding method, using ReLU as the activation function. Its purpose is to initially extract local detail features such as edges and textures of the image while preserving the image's spatial resolution. Since shallow features need to retain more details, a smaller convolutional kernel can capture this information more accurately. The second convolutional layer also uses a 3×3 kernel, a stride of 1, the same padding method, and ReLU as the activation function, further refining and enhancing the initially extracted features. The third convolutional layer still has a 3×3 kernel size, a stride of 1, the same padding method, and ReLU as the activation function. Through these three convolutional operations, the shallow features of the image are gradually extracted and integrated. No pooling layer: Considering that the goal of the shallow branch network is to preserve spatial resolution, no pooling layer is set to avoid the loss of spatial information due to downsampling. After processing through these three convolutional layers, the output feature map is the shallow semantic feature map. This feature map retains more of the original spatial details of the image, providing a foundation for subsequent fusion with the deep semantic feature map. The second is the deep branch network, which works in parallel with the shallow branch network. Unlike the shallow branch, the deep branch network has a deeper network structure, containing more convolutional layers and downsampling operations. Through progressive convolution and downsampling, the network can gradually expand its receptive field, abstracting higher-level, more global semantic information from local features. For example, the network can learn the overall shape of an electric field line as a linear structure in an image, and the semantic concept of a broken strand as a specific structural anomaly. The output of this branch is the deep semantic feature map.Its characteristic is a lower spatial resolution, but each feature vector contains rich contextual information, which helps the model understand the image content and accurately distinguish real fracture defects from background noise or other interference. The deep branching network structure includes convolutional layers: the deep branching network contains 5 convolutional layers and 2 pooling layers. The first two convolutional layers have similar structures. The first convolutional layer has a 3×3 kernel size, a stride of 1, same padding, and uses the ReLU activation function to perform preliminary feature extraction on the input image. The second convolutional layer also has a 3×3 kernel size, a stride of 1, same padding, and uses the ReLU activation function to further enhance the features. Then, a max-pooling layer is applied, with a 2×2 kernel size and a stride of 2. Max pooling is used to downsample the feature map, reducing its spatial resolution while preserving the main features of the image and reducing computational cost. Next, two more convolutional layers, each with a 3×3 kernel size, a stride of 1, same padding, and the ReLU activation function, perform deeper feature extraction on the downsampled feature map. Then, a second max-pooling layer is applied, with a 2×2 kernel size and a stride of 2, for further downsampling. Through this deeper network structure and multiple downsampling operations, the deep branching network can gradually extract more abstract and semantic features from the image. Fully Connected Layer: After 5 convolutional layers and 2 pooling layers, a fully connected layer is applied. The number of neurons in the fully connected layer can be set according to the actual situation, for example, 512, and the activation function is ReLU. The fully connected layer integrates the features extracted by the previous convolutional and pooling layers, further abstracting the deep semantic feature map of the image, providing a more representative feature representation for the final output.

[0039] Optionally, extracting features from the enhanced image patch and outputting an initial feature map includes: Extract the local directional feature response of the enhanced image patch to obtain a multi-channel directional feature tensor; The multi-channel directional feature tensor is subjected to directional clustering compression to output an initial feature map.

[0040] Specifically, firstly, the system extracts the local directional feature responses of the enhanced image patch to obtain a multi-channel directional feature tensor. The core of this operation is to utilize a set of direction-selective filters to detect structural information in various directions within the image. Specifically, the system presets a set of... A filter bank consisting of basis filters, where each filter has the strongest response to edges or lines in a specific direction. These directions uniformly cover a range of 0 to 180 degrees. The input enhanced image patch is then compared with this... Each basis filter is convolved, and each convolution generates a response map. The value of each pixel in this response map reflects the matching strength between the original image at that location and the direction represented by the filter. Stacking response maps along the channel dimension creates a patch with the same height and width as the original enhanced image patch, but with a different number of channels. A multi-channel directional feature tensor is used. Each channel in this tensor encodes all feature response information of the image in a specific direction. Secondly, given that directly using this multi-channel directional feature tensor would lead to enormous computational costs and information redundancy in subsequent networks, the system needs to perform directional clustering compression to output a compact and robust initial feature map. The purpose of this operation is to reduce the dimensionality of the features while preserving key directional information. Specifically, the system will... The finely divided directional channels are as follows A broader category of directions ( much smaller For example, the orientations are divided into four main categories: horizontal, vertical, 45-degree angle, and 135-degree angle. Then, for each pixel location, the system selects the maximum value from multiple channel responses belonging to the same orientation category as the final response value for that category at that pixel. This process can be represented by the following formula: , In this formula, The initial feature map representing the output at pixel location Passing the exam The value on the new channel. The value range is from 1 to ,represent There are directional categories. It is the first The original elements contained in each directional category A set of indices for each directional channel. It is the original multi-channel directional feature tensor at position and the The response values ​​on each original directional channel. By analyzing all... When this maximum pooling operation is performed on each directional category, the system will then pool one... The channel tensor is compressed into a The channel tensor, this The channel tensor is the initial feature map of the final output.

[0041] Optionally, identifying the location of a broken strand in a power line based on the multi-scale fusion features and outputting the broken strand location information includes: Geometric feature structure topology modeling is performed on the multi-scale fused features to obtain the electric line geometric topology tensor; The eigenvalue decomposition and vector field divergence analysis are performed on the tensor of the electric power line geometric topology to identify the location of the broken strand and output the pixel coordinate matrix of the broken strand location.

[0042] Specifically, firstly, geometric feature structure topology modeling is performed on the multi-scale fused features to obtain the electric field line geometric topology tensor. The multi-scale fused feature is an information-rich feature map, where pixel values ​​reflect the probability that a location is an electric field line. To analyze its inherent geometric structure, this method computes a structure tensor at each pixel point or its neighborhood in the feature map. This tensor is constructed by analyzing the local gradients of the multi-scale fused feature map and can be represented as: , In this matrix, and These represent the partial derivatives of the multi-scale fused feature map in the horizontal and vertical directions, respectively, i.e., the gradient components, which can be calculated through methods such as differencing. (Symbols) This indicates summation within a local window centered on the current pixel. This is a 2x2 matrix. The electric field line geometry topology tensor describes the direction and consistency of local features at each pixel. For linear structures, one eigenvalue of this tensor will be much larger than another. By performing this operation on the entire feature map, a tensor field is generated. Subsequently, eigenvalue decomposition and vector field divergence analysis are performed on the electric field line geometry topology tensor to ultimately identify the location of broken strands in the electric field line and output the pixel coordinate matrix of the broken strand location. Electric field line broken strand identification is as follows: Figure 5 As shown. For each structure tensor in the tensor field Eigenvalue decomposition is performed to obtain two eigenvalues ​​and their corresponding eigenvectors. In the electric field line region, the eigenvector corresponding to the larger eigenvalue indicates the local orientation of the electric field line. Connecting the principal eigenvectors of all points in the electric field line region forms a vector field covering the orientation of the electric field line. On a continuous and intact electric field line, this vector field is smooth, the vectors are approximately parallel to each other, and its divergence is close to zero. However, at the location of a break in the electric field line, this continuity is broken, and the vector field forms a sink or endpoint at this location. This sharp contraction of the local vectors manifests as significant negative divergence. Therefore, by calculating the divergence of this vector field and locating those pixels with strongly negative divergence values, the location of the break in the electric field line can be accurately identified. Finally, the two-dimensional coordinates of all pixels identified as break locations are summarized into a list or matrix, namely the pixel coordinate matrix of the break locations, as the final identification result output.

[0043] Optionally, the step of dividing the input image into blocks to generate multiple image blocks to be processed includes: The input image is subjected to local structural sharpness evaluation to obtain the electric field line region mask and the boundary of discontinuous blocks; Based on the electric field line region mask and the non-continuous block boundaries, multiple image blocks to be processed are generated.

[0044] Specifically, the system first needs to perform a local structural sharpness assessment on the input raw power line image. The purpose of this assessment is to identify the regions containing power lines in the image and find potential structural discontinuities, which are often locations prone to strand breakage. Sharpness assessment can employ methods such as gradient magnitude, Laplacian variance, or frequency domain analysis. For linear structures like power lines, their edges are sharp, thus appearing as regions with high local gradient magnitudes in the image. By calculating the gradient magnitude map of the entire image and setting an appropriate threshold, the system can generate a binary power line region mask. In this mask, pixels with a value of 1 represent power line regions, while those with a value of 0 represent the background. Simultaneously, during sharpness calculation, the system pays special attention to boundaries where sharpness changes drastically, such as a sudden transition from a sharp line to a blurred background. These boundaries are marked as discontinuous block boundaries, suggesting possible breaks in the power line structure. Secondly, based on the power line region mask and discontinuous block boundaries obtained in the previous step, the system generates multiple image patches to be processed. This generation process is not a simple uniform grid division, but a content-adaptive intelligent segmentation strategy. The system performs sliding window-style segmentation along the centerline or skeleton of the electric field line region mask. The window size is preset to ensure that each image block contains a complete electric field line segment. This segmentation method along the target structure ensures that each image block to be processed is rich in effective information related to electric fields. More importantly, when the sliding window encounters a previously marked discontinuous segment boundary, the system forces the generation of an independent image block containing that boundary at that boundary. This strategy ensures that the most diagnostically valuable potential breakpoints can be completely captured within an image block, facilitating subsequent focused analysis. Through this combination method, a series of image blocks of similar size and focused content are ultimately generated for processing.

[0045] Optionally, generating multiple image blocks to be processed based on the electric field region mask and the boundaries of discontinuous blocks includes: The power line region mask is structurally constrained and topologically partitioned to obtain a fault-aware topological sub-map. Perform geometric fracture probability field calculation on the discontinuous block boundary to generate a boundary fracture weight grid; Based on the fault-aware topological sub-map, and combined with the boundary fracture weighted grid, multiple image blocks to be processed are generated.

[0046] Specifically, the generated electric field line region mask is first subjected to structurally constrained topology partitioning. The implementation of structurally constrained topology partitioning is as follows: Superpixel node construction: The SLIC superpixel segmentation algorithm is used to segment the electric field line region mask, dividing the mask into uniformly sized superpixel blocks (superpixel side length is set to 10-15 pixels to ensure coverage of the local electric field line structure). The center coordinates of each superpixel block are (…). , As nodes in the graph, node features are the average gradient direction and mean gray value of pixels within the superpixel block. Neighborhood edge definition: For any two nodes, if their corresponding superpixel blocks are spatially related as 4-neighborhoods or 8-neighborhoods (i.e., edges overlap), then an undirected edge is established between the two nodes. Edge weight calculation: The edge weight consists of two parts, as shown in the formula: , in The normalized Euclidean distance between the centers of the two superpixels (range [0,1]); The difference between the features of the two nodes (range [0,1]) is obtained by weighting the normalized cosine value of the gradient direction angle with the absolute difference of the gray mean value; As a balance coefficient (set to 0.3 to prioritize spatial continuity), based on experimental verification in 1000 sample images, Values ​​between 0.2 and 0.4 ensure the continuity of the topology partitioning, while a value of 0.3 balances spatial proximity (avoiding over-cutting) and feature similarity (ensuring the aggregation of similar structures). Minimum cut partitioning: Using the "electric line region mask edge" as the constraint boundary (cut edge weights are set to infinity to avoid cross-mask partitioning), the GraphCut algorithm is used to partition the above graph. Each connected subgraph corresponds to a fault-aware topology subgraph, ensuring the continuity of the electric line structure within the subgraph. This partitioning is not a simple pixel region segmentation, but rather treats the mask as a graphical structure and analyzes it to identify independent, continuous electric line segments. This can be achieved by first finding the centerline of the mask using a skeleton extraction algorithm, and then decomposing the entire electric line network into a set of topologically connected edges and nodes based on the endpoints and intersections of the skeleton. Each independent edge or small connected subgraph constitutes a topology subgraph. This set of subgraphs is the fault-aware topology subgraph set, which decomposes the originally connected mask into multiple meaningful electric line structure units that should be analyzed as a whole. Simultaneously, the system calculates the geometric break probability field for the discontinuous block boundaries obtained in the previous step. The core idea is to assign a "break cost" to each potential cutting location. For any point on the discontinuous block boundary, the system calculates the distance from that point to the nearest electric field line region mask pixel. The smaller this distance, the more likely a cut is to sever the electric field line structure, thus resulting in a higher "geometric break probability." This probability can be quantified as a weight value; for example, the weight is inversely proportional to the distance. By performing this calculation on all boundary points, the system generates a boundary break weight raster. This raster visually represents where image segmentation is safe and where segmentation would damage critical structures. We can use the following function to calculate the break probability field for any boundary point. fracture weight Modeling: , in, From the boundary point Mask to power line area The Euclidean distance to the nearest point in the middle. It is a positive scaling factor used to adjust the sensitivity of distance to the weights. The value of is essentially a feature-driven parameter, and its adaptive mechanism is achieved through dynamic calculation of the physical dimensions of the electric field lines. This formula ensures that when the boundary points... Its breaking weight when far from power lines It approaches a relatively large value (indicating low breakage cost), while when When near power lines, Approaching zero (indicating extremely high fracture cost). Finally, the system generates multiple final image patches to be processed based on the fault-aware topological subgraph set and combined with the boundary fracture weight grid. This generation process is an optimization problem, the goal of which is to cut along the path with the lower weight (i.e., lower fracture probability) in the boundary fracture weight grid, under the constraint that each topological subgraph must be completely contained within a certain image patch. This can be achieved using the minimum cut algorithm in graph theory. The image is preprocessed; for targets such as power line images, edge detection algorithms can be used to extract edge information from the image first. These edge pixels are used as the basic nodes for constructing the graph. At the same time, to reduce the number of nodes and improve computational efficiency, a superpixel segmentation algorithm can be combined to segment the image into several superpixels, and the center pixel or representative feature pixel of the superpixel can be selected as an additional node. For node connections, if the pixels or superpixels corresponding to two nodes are spatially adjacent (4-neighborhood or 8-neighborhood), then an edge is established between them. In this way, the image pixels and boundary information are transformed into a graph with a node and edge structure, which can reflect the local structure and spatial relationships of the image and is suitable for the minimum cut algorithm. The system treats image pixels and boundaries as nodes and edges of a graph, and uses boundary break weights as edge costs. By finding the minimum-cost cutting scheme, the system ultimately generates a series of appropriately sized and complete image patches, each containing at least one complete power line topology subgraph.

[0047] Optionally, the step of performing geometric feature structure topology modeling on the multi-scale fused features to obtain the electric power line geometric topology tensor includes: Based on the local geometric features of the multi-scale fusion features, a set of power line structure feature points is obtained; Using the set of power line structural feature points as nodes, connection edges are established based on spatial proximity and feature similarity to obtain a power line topology graph; Vector synthesis is performed on the power line topology diagram to obtain the power line geometric topology tensor.

[0048] Specifically, the first step is to obtain a set of power line structure feature points based on the local geometric features of the multi-scale fusion feature map. The multi-scale fusion feature map is an activation map that emphasizes the power line regions. To extract discrete representative structural points, non-maximum suppression algorithms or keypoint detection algorithms such as Harris corner detection can be used. These algorithms analyze the gradient changes or autocorrelation matrices of each pixel and its neighborhood in the multi-scale fusion feature map to select points with the strongest local responses that best represent the backbone or key turning points of the linear structure. These selected points collectively constitute the power line structure feature point set, each point accompanied by its pixel coordinates in the image and its feature vector in the multi-scale fusion feature map. Next, the power line structure feature point set is used as nodes, and connecting edges are established based on spatial proximity and feature similarity to obtain the power line topology graph. This step aims to reconstruct the connection relationships between power line segments. For any two nodes in the power line structure feature point set, the following criteria are used to determine whether to establish a connecting edge between them. First, spatial proximity is considered. The Euclidean distance between two nodes is calculated, and a potential connection is only established if this distance is less than a preset spatial threshold. Second, feature similarity is considered. The cosine similarity between the feature vectors associated with two nodes is calculated. This similarity reflects whether the features of the local regions where the two points are located are consistent. A connection is confirmed only if the feature similarity is higher than a preset similarity threshold. By performing this judgment on all possible point pairs in the point set, an undirected graph, namely the power line topology graph, is constructed. The nodes in this graph are the structural feature points of the power lines, and the edges represent the strong correlation between these points in the physical space and feature space, effectively simulating the continuous structure of the power lines. Finally, vector synthesis is performed on the power line topology graph to obtain the final power line geometric topology tensor. This step transforms the discrete graph structure back into a continuous tensor field representation for subsequent differential analysis. For each pixel location in the image, the value of its geometric topology tensor is calculated by aggregating information from its neighboring power line topology graphs. Specifically, for a pixel, all graph nodes and connecting edges within a certain range around it are examined. The direction and length of each edge can be represented as a vector. By weighted summing or tensor product synthesis of these neighboring edge vectors, a 2x2 structure tensor describing the local structure's orientation and intensity can be constructed at that pixel location. Therefore, the weighting formula can be expressed as: Let the target pixel be... A certain neighboring edge The coordinates of the midpoint are ,side Feature similarity is (If the range is [0,1], then:) , Among them, spatial distance weight : , The scale parameter is set based on the pixel width of the electric field lines; the feature similarity weights are... for: , The weighted summation formula is as follows: Let the edges be... The direction unit vector is Length (strength) The tensor is then generated by a weighted sum of the outer products of the edge vectors: ], The final generated 2x2 geometric topology tensor corresponds to the pixels in the image. The local structural information at that location is used to describe the topological features of the power lines, such as their direction and continuity. Represents pixels The set of neighboring edges (such as edges within a 3x3 neighborhood). This represents a specific edge in the set, and these edges come from the "power line topology graph" constructed earlier; Representing an edge For pixels The contribution weights comprehensively reflect "spatial proximity" and "feature similarity"; Representing an edge The length (intensity) reflects the structural salience of the electric field line segment represented by that edge (e.g., the longer the length or the stronger the characteristic response). (The larger the value) Representing an edge Directional unit vector This indicates the direction of the power line segment (such as horizontal, vertical, or inclined direction). The outer product matrix representing the direction vector is the core of transforming one-dimensional direction information into a two-dimensional structural description.

[0049] This synthesis process can be viewed as an interpolation or smoothing operation from a discrete image to a continuous field, ultimately generating a smoothly varying tensor field across the entire image, namely the electric field line geometry tensor. This tensor exhibits strong anisotropy in the electric field line region, with its principal direction aligned with the direction of the electric field lines, while it exhibits isotropy in the background region.

[0050] Based on the same inventive concept, such as Figure 6 As shown, the present invention also provides a real-time identification system for broken strands in unmanned aerial vehicles (UAVs) power lines that resists motion fuzzing, the system comprising: The image acquisition module is used to acquire images of power lines taken by the drone as input images; The block processing module is used to divide the input image into blocks to generate multiple image blocks to be processed; A selective filtering module is used to selectively filter the image block to be processed to obtain an enhanced image block; The feature extraction module is used to extract the enhanced image blocks to obtain shallow semantic feature maps and deep semantic feature maps; A multi-scale feature fusion module is used to fuse the shallow semantic feature map and the deep semantic feature map to generate multi-scale fused features; The broken strand location identification module is used to identify the broken strand location of the power line based on the multi-scale fusion features and output the broken strand location information.

[0051] To verify the feasibility of this invention in practice, it was applied to a regional power grid company's autonomous drone inspection mission of 500kV high-voltage transmission lines. Most of the lines maintained by this power grid company are located in mountainous areas and are frequently affected by strong winds, resulting in varying degrees of motion blur in the power line images captured by the drones. Traditional manual identification or conventional algorithmic identification methods suffer from low recognition rates and serious false alarms and missed alarms. This embodiment aims to verify the invention's real-time and accurate identification capability of broken strands in motion-blurred images of power lines under complex environments.

[0052] In this embodiment, the system is deployed on an industrial drone's onboard computing platform equipped with a high-definition gimbal camera. During an inspection flight targeting power lines in a mountainous area, the drone flew at a speed of 15 meters per second, capturing a series of power line video streams with a resolution of 4096x2160 pixels. The system extracts image frames from these streams as input images for processing.

[0053] First, for an input image that is partially blurred due to gusts of wind, the block processing module of this invention begins operation. The system first performs a local structural sharpness assessment on the input image, generating a power line region mask to accurately identify the main regions containing power lines in the image, while simultaneously identifying smooth backgrounds such as the sky and distant mountains. Based on this mask, the system further performs structurally constrained topological decomposition, decomposing the complex power line network into multiple independent fault-aware topological sub-maps. Simultaneously, the system calculates the geometric fracture probability field for low-resolution regions in the image, generating a boundary fracture weight grid to assess the risk of cutting the image at different locations. Finally, the system combines the topological sub-maps and the weight grid to intelligently divide the original image into 12 image blocks containing complete power line segments, discarding 7 pure background image blocks. This avoids the problem of cutting a potential break point into different image blocks and reduces the amount of data for subsequent processing. Subsequently, an image block containing significant motion blur is sent to the selective filtering module. This module first analyzes the gradient distribution of the image patch. By constructing a gradient direction field structure and mapping the blur direction intensity, it calculates that the main blur direction is approximately 5 degrees horizontally, and the blur intensity is equivalent to the linear motion of 12 pixels. Based on this blur direction perception result, the system generates corresponding direction-aware filtering parameters and performs adaptive deconvolution filtering on the image patch. In the enhanced image patch output after processing, the previously blurred details of the electric field lines become clear and sharp, laying the foundation for subsequent feature extraction. Next, the enhanced image patch is fed into the feature extraction module. The system first extracts the local steering feature response of the enhanced image patch, generating a multi-channel direction feature tensor containing 16 channels, each channel corresponding to a linear structure response in a specific direction. Subsequently, through direction clustering compression, this tensor is compressed into a 4-channel initial feature map, representing the structural intensity in the horizontal, vertical, and two diagonal directions, respectively. This initial feature map is simultaneously fed into parallel shallow and deep branch networks. The shallow network outputs a shallow semantic feature map containing the precise edges and strand endpoints of the power line; the deep network outputs a deep semantic feature map that understands the overall structure and contextual relationships of the power line.

[0054] The multi-scale feature fusion module fuses the two feature maps mentioned above to generate a multi-scale fused feature that combines precise positioning information with high-level semantic information. Finally, the broken strand location identification module processes this fused feature. The system first extracts a set of power line structural feature points based on the fused feature and constructs a power line topology map based on spatial proximity and feature similarity. Then, it transforms this map into a power line geometric topology tensor through vector synthesis. In analyzing this tensor field, the system discovers that the eigenvalues ​​of the structural tensor in a certain region... and The ratio dropped sharply from 8.5 for a normal line segment to 1.2, indicating a significant reduction in anisotropy. Simultaneously, divergence analysis of the eigenvector field in this region revealed a sharp local maximum. Combining these two indicators, the system determined the point to be the location of the broken strand and output its precise pixel coordinate matrix. Through the implementation of this invention, the system successfully identified a minor defect caused by lightning strikes, involving only two broken steel-cored aluminum strands, during this inspection. This defect was almost impossible for the human eye or traditional algorithms to detect in the original blurred image. To quantitatively verify the beneficial effects of this invention, we compared it with a traditional deep learning object detection method (baseline method) that lacks adaptive deblurring and topology analysis capabilities.

[0055] Table 1 Comparison of Anti-motion blurring effects

[0056] Table 2 Comparison of performance in identifying broken strands

[0057] Table 3 Comparison of System Processing Efficiency

[0058] As can be seen from the data in Tables 1-3 above, this invention has significant advantages in resisting motion blur and identifying broken strands. Table 1 clearly shows that the selective filtering module of this invention far surpasses the baseline method in image restoration quality. For severely blurred images (such as IMG_002), this invention can improve the image sharpness index PSNR to 30.5dB, a qualitative leap compared to the baseline method's 23.1dB, successfully restoring the strand details obscured by blur. The data in Table 2 further confirms the recognition performance of this invention. In scenarios with severe blur caused by strong winds, the accuracy of the baseline method drops sharply to 68.7%, generating a large number of false alarms. However, this invention, with its strong anti-blurring capability and robust topology analysis, still maintains a high accuracy of 97.8%, with an extremely low number of false alarms. This proves the high reliability of this invention under harsh inspection conditions. Table 3 shows the advantages of this invention in processing efficiency. By employing intelligent block-based processing, this invention avoids redundant calculations on a large number of irrelevant background areas, resulting in a total processing time of only 210 milliseconds for a single 4K image. Compared to the baseline method's 350 milliseconds for processing the entire image, this represents an efficiency improvement of approximately 40%, fully meeting the application requirements for real-time UAV recognition. In summary, this embodiment fully demonstrates that this invention can effectively solve the motion blur problem in UAV inspections, achieving high-precision, high-efficiency, and highly robust real-time recognition of power line strand breakage defects.

[0059] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0060] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and practicing the disclosure of this invention. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. A method for real-time identification of broken strands in unmanned aerial vehicle (UAV) power lines, characterized in that, The method includes: The image of the power lines captured by the drone is used as the input image; The input image is divided into blocks to generate multiple image blocks to be processed; Selective filtering is applied to the image block to be processed to obtain an enhanced image block; this includes: analyzing the gradient distribution of the image block to be processed to generate a blurred direction-aware result; obtaining direction-aware filtering parameters based on the blurred direction-aware result; and performing adaptive filtering on the image block to be processed based on the direction-aware filtering parameters to output an enhanced image block. The enhanced image patch is extracted to obtain a shallow semantic feature map and a deep semantic feature map; By fusing the shallow semantic feature map and the deep semantic feature map, a multi-scale fused feature is generated; Based on the multi-scale fusion features, the location of broken strands in the power line is identified and the broken strand location information is output. The step of analyzing the gradient distribution of the image patch to be processed and generating a fuzzy direction perception result includes: obtaining a gradient direction field structure based on the gradient distribution of the image patch to be processed; and performing fuzzy direction intensity mapping on the gradient direction field structure to generate a fuzzy direction perception result.

2. The method for real-time identification of broken strands in UAV power lines with anti-motion fuzziness as described in claim 1, characterized in that, The extraction of the enhanced image patch to obtain the shallow semantic feature map and the deep semantic feature map includes: Extract the features of the enhanced image patch and output the initial feature map; The initial feature maps are input into preset shallow branch networks and deep branch networks respectively, and shallow semantic feature maps and deep semantic feature maps are generated in parallel.

3. The method for real-time identification of broken strands in UAV power lines with anti-motion fuzziness as described in claim 2, characterized in that, The step of extracting features from the enhanced image patch and outputting an initial feature map includes: Extract the local directional feature response of the enhanced image patch to obtain a multi-channel directional feature tensor; The multi-channel directional feature tensor is subjected to directional clustering compression to output an initial feature map.

4. The method for real-time identification of broken strands in UAV power lines with anti-motion fuzziness as described in claim 1, characterized in that, The step of identifying the location of broken strands in power lines and outputting the broken strand location information based on the multi-scale fusion features includes: Geometric feature structure topology modeling is performed on the multi-scale fused features to obtain the electric line geometric topology tensor; The eigenvalue decomposition and vector field divergence analysis are performed on the tensor of the electric power line geometric topology to identify the location of the broken strand and output the pixel coordinate matrix of the broken strand location.

5. The method for real-time identification of broken strands in UAV power lines with anti-motion fuzziness as described in claim 1, characterized in that, The step of dividing the input image into blocks to generate multiple image blocks to be processed includes: The input image is subjected to local structural sharpness evaluation to obtain the electric field line region mask and the boundary of discontinuous blocks; Based on the electric field line region mask and the non-continuous block boundaries, multiple image blocks to be processed are generated.

6. The method for real-time identification of broken strands in UAV power lines with anti-motion fuzziness as described in claim 5, characterized in that, The process of generating multiple image blocks to be processed based on the electric field region mask and the boundaries of discontinuous blocks includes: The power line region mask is structurally constrained and topologically partitioned to obtain a fault-aware topological sub-map. Perform geometric fracture probability field calculation on the discontinuous block boundary to generate a boundary fracture weight grid; Based on the fault-aware topological sub-map, and combined with the boundary fracture weighted grid, multiple image blocks to be processed are generated.

7. The method for real-time identification of broken strands in UAV power lines with anti-motion fuzziness as described in claim 4, characterized in that, The geometric feature structure topology modeling of the multi-scale fused features to obtain the electric line geometric topology tensor includes: Based on the local geometric features of the multi-scale fusion features, a set of power line structure feature points is obtained; Using the set of power line structural feature points as nodes, connection edges are established based on spatial proximity and feature similarity to obtain a power line topology graph; Vector synthesis is performed on the power line topology diagram to obtain the power line geometric topology tensor.

8. A real-time identification system for broken strands in unmanned aerial vehicles (UAVs) power lines, characterized in that, The system includes: The image acquisition module is used to acquire images of power lines taken by the drone as input images; The block processing module is used to divide the input image into blocks to generate multiple image blocks to be processed; A selective filtering module is used to selectively filter the image block to be processed to obtain an enhanced image block; The feature extraction module is used to extract the enhanced image blocks to obtain shallow semantic feature maps and deep semantic feature maps; A multi-scale feature fusion module is used to fuse the shallow semantic feature map and the deep semantic feature map to generate multi-scale fused features; The broken strand location identification module is used to identify the broken strand location of the power line based on the multi-scale fusion features and output the broken strand location information.

Images