A method and system for distinguishing between cracks and light spots in insulators

Abstract

The application belongs to the technical field of power distribution line inspection, and provides an insulator crack-light spot distinguishing method and system. In aerial unmanned aerial vehicle inspection, through planning a closed track and synchronous shooting of three exposure levels, cooperating with a double-branch detection network, chain geometry topology constraint and one-dimensional multi-flow graph convolution, the candidate crack detection is realized by using high light / dark line double-branch lightweight convolution to access the double-branch detection network. The candidate frame is aligned with the disc edge and the three-exposure local is cropped by geometric matching, and the multi-dimensional features are extracted. After the saturation mark is used to attenuate the high light channel, the features are assembled into a chain graph in physical order, and the cross-disc coherent dark lines are captured. Finally, the crack probability is output by the full connection layer, the mirror reflection is suppressed, the real cracks are realized, the false positives and false negatives are effectively reduced, and accurate basis is provided for power outage maintenance decision.

Description

Technical Field

[0001] This invention belongs to the field of power distribution line inspection technology, specifically relating to a method and system for distinguishing insulator cracks from light spots. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Currently, a common technical approach in power distribution line inspection involves using multi-rotor drones to execute pre-defined polygonal or spiral trajectories around insulator strings to acquire continuous visible light images. These images are then analyzed using a convolutional neural network-based defect detection model. Mainstream methods typically transfer general detection frameworks such as YOLO and Faster-RCNN to insulator targets, using transfer learning to shorten the training cycle and batch data augmentation to improve adaptability to different lighting conditions and shooting angles. This type of single-frame algorithm, leveraging high-resolution sensors, can output the location and confidence level of potential cracks in a short time and is widely integrated into existing mobile terminals or cloud-based analysis platforms of power distribution companies.

[0004] To reduce false detection rates under strong light conditions, some researchers have implemented hardware-based light suppression measures on the airborne end. For example, adding a variable polarizing filter to the gimbal reduces specular reflection by rotating the polarization angle; or using local high dynamic range imaging, combining pixels from different exposures of the same scene into a single HDR image, thereby compressing the brightness gradient before processing by the detection network. Other solutions introduce a real-time metering module at the flight control level, dynamically adjusting the shutter speed or gain based on the histogram distribution to reduce the occurrence of saturation areas.

[0005] At the algorithm level, some researchers have proposed various optimization models or introduced attention mechanisms and other modal data to improve the reliability of defect identification.

[0006] In summary, existing technologies mainly focus on two main lines: improving input quality and optimizing feature extraction of single frames or short sequences. They mitigate the interference of specular reflection through hardware dimming, HDR synthesis, attention branching, and detection-tracking coupling, and provide maintenance personnel with automated crack location references while ensuring the real-time performance of the algorithm.

[0007] However, most existing technologies rely on additional hardware or offline compositing processes, making it difficult to complete synchronously within the time window of a typical aerial photography mission. Hardware-based light suppression solutions are significantly affected by changes in the sun's incident angle and gimbal attitude fluctuations, and insufficient real-time adjustment can still leave saturation patches. HDR compositing requires pixel-level registration between multiple frames, and micro-displacements caused by high-altitude vibrations often introduce new artifacts. While multi-branch networks and attention mechanisms improve feature separation capabilities, when using only single-frame or short-sequence information, the grayscale and morphology of specular reflections and crack details still highly overlap, blurring the network's discrimination boundaries. The false positive and false negative rates remain significant in strong sunlight scenarios. Temporal models combining optical flow or Kalman filtering require long, continuous video streams to operate stably. Power line inspections often involve fixed-point, slow, panoramic shooting, with limited inter-frame perspective differences, making it difficult to provide sufficient motion cues. Summary of the Invention

[0008] To address the aforementioned issues, this invention proposes a method and system for distinguishing between insulator cracks and reflective spots. While maintaining the existing UAV flight path and operation procedures, this invention can distinguish between micro-cracks and speckled reflective spots on the same aerial image in real time. This allows for the output of only the necessary crack information to the maintenance end during the inspection loiter time, thereby reducing false alarms, missed alarms, and the burden of subsequent manual verification.

[0009] According to some embodiments, the present invention adopts the following technical solution: A method for distinguishing insulator cracks from light spots includes the following steps: Three frames of images of the insulator under high, medium and low exposure levels were acquired. The sampling trajectory of the images was a concentric closed curve trajectory planned with the axis of the insulator string as the center. Three frames of images are input into a pre-trained dual-branch target detection network. The highlight component of the high-exposure image and the dark texture component of the same frame are input into the first and second branches of the target detection network, respectively. The specular reflection samples are given weak negative labels and the loss weights are adaptively adjusted according to the batch ratio. The target detection network outputs a set of crack candidate boxes. The grayscale images in the same frame are processed to obtain the set of insulator disks along the center of the circle. The main chain is formed by sorting the main components according to the direction of the principal components. The candidate boxes in the crack candidate box set are matched one-to-one with the center of the main chain. Redundant candidate boxes are removed and a list of candidate boxes arranged in the order of disk number is generated. Each candidate position in the candidate box list is a tile of the same size as the high, medium and low exposure frame cropping. Extract each exposure texture vector and calculate the saturation ratio. Mark the high exposure saturation stream, perform adjacent disk compensation for the missing stream, and construct a chain node object carrying the three stream textures. The chain node objects are arranged into a one-dimensional chain graph according to the physical disk number, high exposure saturation marks are injected into the nodes, and the crack probability of each node is calculated. Calculate the local mean of the crack probability for each node. If the probability of a single node is lower than the preset proportion of its local mean and the high exposure saturation is marked as true, then the node is marked as specular reflection. If the probabilities of three adjacent nodes are all high and the low exposure flow is complete, then the probability of one of the nodes is increased. Finally, output a list of crack or spot results sorted by batch number.

[0010] As an alternative implementation, acquiring three frames of images of the insulator under high, medium, and low exposure levels, and the process of mapping the image sampling trajectory as a concentric closed curve trajectory planned with the insulator string axis as the center, includes: planning the UAV's flight trajectory as a circle with the insulator string axis as the center and a radius of... The concentric closed curves, in which Indicates the horizontal distance from the fuselage to the insulator axis; sets the constant flight speed and fixed focal length and pitch angle; Maintain the aircraft moving at a constant speed along a closed curve and keep the pitch angle constant, control the track overlap rate to be fixed above the set threshold, and ensure that each insulator always falls on the center area of ​​the sensor for imaging. For each preset track sampling point on the closed curve, three exposure levels (high, medium, and low) are used in succession with continuous shutter speeds to complete three frame captures within a single exposure cycle, generating a complementary image triplet. During three consecutive exposures, the contour contrast structure on the inspection line is used as a visual reference. Real-time template matching is performed on the elements of the triplet of the three frames to calculate the reference offset vector, and the reference offset vector is mapped to the gimbal fine-tuning command.

[0011] As an alternative implementation, the dual-branch target detection network, based on the YOLOv13 target detection network, reconstructs its input layer from a single channel to a dual-branch system. The first branch receives the highlight component from the high-exposure image, denoted as... The second branch receives the dark texture component from the same frame image, denoted as... ; right and Each of the entry points is connected in series with a lightweight convolutional kernel group, so that the network decouples the speckle and crack texture at the shallowest layer. Keep the parameters of subsequent layers of the YOLOv13 object detection network unchanged.

[0012] As an alternative implementation, during the training process, the target detection network introduces weak negative labels to the specular reflective areas in the training set, marking these areas as background categories that need to be identified but are not critical to the power outage decision. The weight of this category in the loss function is set as the batch adaptive coefficient. And after each forward propagation, it is dynamically updated based on the proportion of specular reflective samples in the current batch. It automatically weakens the gradient contribution when the number of reflective samples surges; The established online reweighting strategy causes the network to focus more attention on the dark pattern branches. The tiny crack edges in the mirror do not deviate from the learning direction due to the high contrast of the bright areas.

[0013] As an alternative implementation, the process of the target detection network outputting a set of crack candidate boxes includes: extracting the area and confidence of the candidate boxes from the original detection outputs of the first branch and the second branch respectively; and comprehensively considering the area-confidence comparison of the two branches, calculating an adaptive threshold online for the dark ripple branch. Dark texture boxes with a confidence level greater than an adaptive threshold are imported into the priority list, while only highlight boxes with a confidence level greater than a fixed threshold are imported into the backup list. Perform structured reasoning on the priority list; simultaneously store the backup list; An object detection network is used to perform pixel-by-pixel reasoning on the same frame of images, output a set of candidate boxes, and simultaneously extract the visual feature vectors of the corresponding level.

[0014] Furthermore, considering the area-confidence comparison of the two branches, the process of calculating the adaptive threshold online for the dark ripple branch includes:

[0015] in, An adaptive confidence threshold for dark ripple branches; This is the system's baseline threshold. This is the threshold adjustment coefficient; , For the first The first dark pattern frame and the first Confidence level of each highlight frame; , This represents the pixel area of ​​the corresponding candidate box; To prevent tiny constants with a denominator of zero.

[0016] As an alternative implementation, the process of processing the grayscale image of the same frame to obtain the set of center points along the insulator disk includes: performing opening and closing operations on the grayscale image of the same frame using circular structuring elements; performing Hough circle search on the processed image within a set radius range to obtain a set of candidate arcs; calculating the center point and average curvature of each element in each candidate arc set; and sequentially removing pseudo arcs with an average curvature higher than a threshold and semicircles with radius defects caused by conductor shading to obtain a set of center points covering the entire insulator disk edge and free from shadows and reflections.

[0017] As an alternative implementation, the process of forming the main chain based on the principal component direction sorting includes: performing principal component direction analysis on the set of circle centers, extracting the first principal axis vector as the pitch baseline of the insulator string; calculating the projection scalar of each circle center on the first principal axis vector, and sorting the projection scalars in ascending order of their values ​​to generate a physical order index. At the same time, the shooting side is determined based on the orientation of the first main axis and the airborne attitude information to ensure that the indexing direction from bottom to top or from top to bottom is consistent with the on-site installation direction; Connect adjacent circle centers sequentially. If the line segment length falls within the rated disc distance tolerance range, it is marked as a valid edge and the two are merged into the same chain; otherwise, it is judged as a broken chain and the current branch is terminated. For multiple chains appearing in the same frame, select the one with the most edges as the main chain, and at the same time remove duplicate or isolated circle centers that fall into the background to ensure that only one complete geometric skeleton that runs through the entire disc edge is retained, laying a unique topology for the construction of exposure multi-flow map.

[0018] As an alternative implementation, the process of matching candidate boxes in the crack candidate box set with the main chain center one-to-one, removing redundant candidate boxes, and generating a candidate box list arranged in disk order includes: associating the candidate box set with the main chain center sequence: Based on the center coordinates, size parameters, and crack confidence of each candidate box, as well as the coordinates and size parameters of each circle center, a composite cost matrix is ​​constructed using two types of features, and the minimum cost match is found using Hungarian assignment. Determine the matching result for each circle center. If the corresponding value is higher than the threshold, it is determined that there are no reliable crack candidates in this plate and it is marked as missing. If the value is lower than the threshold, the matching box is written into the circle center label, and the remaining boxes whose distance from the center of this box is less than the set value are uniformly marked as redundant to prevent duplicate counting in the same plate. The filtered matching boxes are rearranged according to the main chain order to generate a one-dimensional sequential list, i.e., the candidate box list, and the missing and redundant position indices are recorded simultaneously. The disk number and tower identification number are added sequentially to the sequential list, and the candidate box coordinates, size parameters, crack confidence and deep visual features are encapsulated together into a structured sequence object.

[0019] As an alternative implementation, the process of cropping equal-sized patches from high, medium, and low exposure frames for each candidate position in the candidate box list includes: mapping the center coordinates of each candidate box in the candidate box list to three images at the same time: a low exposure frame, a medium exposure frame, and a high exposure frame; and cropping local patches at the mapped coordinates using a pixel window of a fixed size.

[0020] As an alternative implementation, the process of extracting each exposure texture vector and calculating the saturation ratio, and marking the high exposure saturation stream includes: extracting a fixed-length texture vector for each stream patch using a lightweight convolutional encoder with shared weights; simultaneously reading the average gray level and saturation pixel ratio of the patch during the encoding process, and writing them together with the fixed-length texture vector into the node label, and uniformly mapping strong reflection peaks and weak crack ridges to the same feature dimension. The three texture vectors of the same node are concatenated in low-medium-high exposure order, and the disk number and chain index are appended to package them into a multi-stream node object; The three stream tiles are sequentially input into the multi-threshold brightness consistency operator. The multi-threshold brightness consistency operator is used to calculate the saturation ratio of the exposure. Then, the specular saturation is judged based on the cross-stream difference. If the saturation ratio is the set value, it is determined that the current node has specular saturation in the high-exposure stream, and the multi-threshold brightness consistency operator then marks this stream as missing.

[0021] As an alternative implementation, the process of constructing a chain node object carrying three-stream texture for adjacent disk compensation of missing streams includes: treating the exposure stream marked as missing as the target to be compensated, establishing an up and down search window for each node of the chain index, and retaining only candidate nodes within the three-step range that have complete information on the missing stream. For each candidate, the texture vector of the non-missing flow is calculated and a distance decay coefficient is introduced to form a weighted similarity; then the co-flow vector of the highest-scoring candidate is scaled and copied to the current node, and the adjacent disk similar texture compensation operator is used to perform segmented mapping to achieve real-time compensation; By using segmented mapping, the node with the most similar texture and the closest physical distance is selected as the compensation source within the chain structure to suppress the difference in lighting angle caused by excessive span, while cosine similarity is used to ensure the consistency of material texture. If no available source is found within the three-step window, the missing state will remain.

[0022] Furthermore, in the piecewise mapping using the adjacent disk similar texture compensation operator, the mapping relationship is as follows:

[0023] in, For the exposure stream after compensation Texture vector; For the original exposure stream Texture vector; , For nodes , The spliced ​​texture vectors on both streams without missing elements; This is the distance attenuation coefficient; This is an adaptive scaling factor for the copied texture; For the physical order index of the nodes; and For vector dot product and Norm.

[0024] As an alternative implementation, the process of assembling the chain node objects into a one-dimensional chain graph according to the physical disk order includes: assembling the multi-flow node objects into a one-dimensional chain graph in physical order; and retaining only the two directed edges pointing to the adjacent upper and lower disks for each multi-flow node object in the node set. Each edge in the edge set is injected with two attributes: the actual disk spacing and the average brightness difference of the three streams. The node body is appended with the three stream texture vector, disk number and high exposure saturation mark, and uniformly expanded into a fixed-length feature table. The chain graph follows the insulator installation sequence.

[0025] As an alternative implementation, the process of injecting high-exposure saturation markers into nodes and calculating the crack probability of each node includes: sending the high-exposure texture vector in the node features into an exposure gating unit; the exposure gating unit consists of two fully connected network layers, taking the saturation marker and the three-stream grayscale mean as inputs, and outputting a gating coefficient of zero to one; the gating coefficient is used to attenuate the high-exposure texture vector and concatenate it with the medium and low-exposure vectors to form the updated node features; The first layer of chain graph convolution is applied to the gated node features: the features of the current node and its preceding and following neighboring nodes are weighted and averaged according to the weights generated by the linear mapping of edge attributes, and a new node representation is output, which fuses the texture of the current disk with the continuous information of the adjacent disks; In the second layer of graph convolution, cross-slice residuals are introduced: in addition to directly aggregating first-order adjacency information, features of distances to two disk nodes are incorporated and superimposed on the current node through the residual path to obtain the expanded domain representation. Layer normalization is applied to the outputs of the two layers to balance the contributions of different distances. A coherence suppressor is applied to the node representation after convolution stacking: the bidirectional difference between the predicted scores of the current node and the nodes before and after it is calculated. If the bidirectional difference is higher than the dynamic threshold at the same time, the node is identified as an isolated peak lacking coherence and multiplied by the suppression factor. The dynamic threshold is estimated in real time based on the mean and variance of the scores on the chain. The node vectors after coherence suppression are fed into the classification head, and a fully connected layer with a sigmoid function is used to output the crack probability. The disk number and exposure index are then packaged together.

[0026] As an alternative implementation, the process of calculating the local mean of the crack probability of each node includes: recording the crack probabilities arranged in disk order as a vector, where each element in the vector represents the crack probability of the insulator disk corresponding to the chain index. A sliding window of length three is used. For each central node, the window mean is calculated by taking the probabilities of its three adjacent nodes above and below, while retaining the node's own probability. The window mean is used as a local continuous reference.

[0027] As an alternative implementation, if the probability of a single node is lower than a preset proportion of its local mean and the high exposure saturation is marked as true, then the process of marking this node as specular reflection includes: For each node, a deviation judgment is performed. If the probability of a node crack is less than 50% of the average value of its corresponding window, and the saturation flag of the high exposure channel of this node is equal to one, then this node is inferred to be an isolated highlight caused by specular reflection. Set its crack probability to zero and write the spot identifier in the node label.

[0028] As an alternative implementation, the process of enhancing the node probability includes applying an enhancement rule to each consecutive three-node segment in the chain vector: when the crack probability of the three nodes is higher than the mean of their respective windows, and at least two nodes show complete dark patterns in the low exposure stream, i.e., the low exposure saturation ratio is less than a preset threshold, it is inferred that there is a real cross-disk crack. At this time, the crack probability of the middle node is multiplied by a weighting coefficient to highlight the defect center, while keeping the original values ​​of the nodes on both sides. The weakened and enhanced crack probabilities are repackaged into a list of classification results, and then the shooting timestamp and exposure level information are added. The results are sorted from high to low according to the revised crack probabilities.

[0029] An insulator crack-spot differentiation system includes: The insulator image acquisition module is configured to acquire three frames of images of the insulator under high, medium and low exposure levels. The sampling trajectory of the image is a concentric closed curve trajectory planned with the axis of the insulator string as the center. The target detection module is configured to input three frames of images into a pre-trained dual-branch target detection network. The highlight component of the high-exposure image and the dark texture component of the same frame are respectively input into the first and second branches of the target detection network. The specular reflection samples are given weak negative labels and the loss weights are adaptively adjusted according to the batch ratio. The target detection network outputs a set of crack candidate boxes. The candidate box preprocessing module is configured to process the grayscale image of the same frame, obtain the set of insulator disks along the center, sort them according to the principal component direction to form the main chain, match the candidate boxes in the crack candidate box set with the center of the main chain one-to-one, remove redundant candidate boxes and generate a list of candidate boxes arranged in the order of disk number. The adjacent disk compensation module is configured to extract each exposure texture vector and calculate the saturation ratio for each candidate position in the candidate box list, cropping a patch of the same size from high, medium, and low exposure frames, marking the high exposure saturation stream, performing adjacent disk compensation for the missing stream, and constructing a chain node object carrying the texture of the three streams. The node crack probability calculation module is configured to form a one-dimensional chain graph of the chain node objects according to the physical disk number, inject high exposure saturation markers into the nodes, and calculate the crack probability of each node. The differentiation module is configured to calculate the local mean of the crack probability of each node. If the probability of a single node is lower than the preset proportion of its local mean and the high exposure saturation is marked as true, then the node is marked as specular reflection. If the probabilities of three adjacent nodes are all high and the low exposure flow is complete, then the probability of one of the nodes is increased. Finally, a list of crack or spot results sorted by batch number is output.

[0030] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention innovatively provides a method for rapid differentiation of insulator cracks and light spots on drones. At the drone end, the insulator string axis is set as the center of a circle to plan a concentric closed curve trajectory, with constant flight speed, fixed focal length, and pitch angle. At the trajectory sampling point, three frames are captured by continuously triggering high, medium, and low exposures on the insulator. The offset vector is calculated through structural template matching and the gimbal is fine-tuned to ensure that the field of view of the three frames remains consistent within the sub-pixel range, improving the accuracy of image acquisition and contributing to improved differentiation accuracy in later stages.

[0031] This invention innovatively provides a method for rapid differentiation of insulator cracks and light spots in UAVs. During target detection, the method improves the target detection network by forming a dual-branch input layer. This decouples the specular reflection spots from the crack texture at the shallowest layer while keeping other layers unchanged, ensuring direct compatibility of the overall computation graph with existing embedded inference frameworks. This provides a separable feature tensor flow for chain graph discrimination. Furthermore, by introducing weak negative labels in the specular reflection area during training, the network focuses more on the tiny crack edges in the dark crack branch, preventing it from deviating from its learning direction due to the high contrast of the specular highlight area. This preserves stable and crack-related gradient signals for subsequent candidate generation. The dual-branch outputs are used to retain candidates based on a dark crack priority and a reflection delay approach. This allows for crack localization during the UAV's flight time, followed by background cleanup of potential false positives from reflections, ensuring that the judgment is triggered only by genuine defects.

[0032] This invention innovatively provides a method for rapid differentiation of cracks and light spots in UAV insulators. This method combines chain-like geometric topological constraints and multi-flow graph neural networks. A main chain is formed based on the principal component direction, and candidate boxes are matched one-to-one with the center of the main chain. Redundant candidates are removed, and a candidate list arranged in disk order is generated. For each candidate position in the candidate list, a cropped image is used to extract each exposure texture vector and calculate the saturation ratio. High-exposure saturation flows are marked using a multi-threshold brightness consistency operator. Missing flows are compensated for by neighboring disks according to the principles of nearest neighbor priority, texture similarity, and distance attenuation. A chain-like node object carrying three-flow textures is constructed, effectively realizing the correspondence between candidate information and actual disk order, eliminating false detections due to empty background clutter, ensuring that the chain-like inference always has complete crack information, suppressing illumination angle differences caused by excessively long spans, and using cosine similarity to ensure the consistency of material texture.

[0033] This invention innovatively provides a method for rapid differentiation of insulator cracks and speckles using drones. It can distinguish between micro-cracks and speckled reflections on insulators in real time within the same aerial image. This allows for the output of only the necessary crack information to the maintenance end during the inspection loiter time, thereby reducing false alarms, missed alarms, and the burden of subsequent manual verification. It provides an accurate basis for power outage maintenance decisions and has a wide range of applications.

[0034] This invention innovatively proposes a candidate box precision processing method that combines dual-branch decoupled detection and chain-like geometric topology constraints. It addresses the dual problems in existing insulator defect detection: single-branch networks cannot decouple highlight reflections from dark crack features, and candidate boxes lack geometric constraints, leading to numerous redundant false detections and an inability to correlate with the physical disk order of insulators. This invention optimizes candidate box quality through a two-stage collaborative process: the first stage reconstructs the input layer of the target detection network into a dual-branch system. The first branch receives the highlight component of the high-exposure image, and the second branch receives the dark crack component from the same frame. Lightweight convolutional kernels are cascaded at the entrances of the two branches to achieve the shallowest feature decomposition. In the first stage, during training, specular reflection samples are labeled with weak negative labels, and their loss weights are set as batch adaptive coefficients (dynamically updated according to the proportion of reflective samples, weakening the gradient when there is a surge), ensuring that the network focuses on the smallest cracks in the dark ripple branch. In the second stage, the grayscale image of the same frame is processed by circular structuring element opening / closing operations and Hough circle retrieval within a set radius interval to obtain the set of disk edge centers. Principal component analysis is used to extract the pitch baseline of the insulator string to generate a physical order index. The chain with the most edges is selected as the main chain. Then, a composite cost matrix and Hungarian allocation algorithm are used to achieve one-to-one matching between candidate boxes and the center of the main chain, eliminating redundant boxes and rearranging them according to disk order. This method avoids reflection interference to output stable crack candidates, eliminates background stray false detections, and establishes the correlation between candidate boxes and physical disk order, laying the foundation for subsequent processing.

[0035] This invention innovatively proposes a nearest-neighbor priority-texture-similar missing exposure stream compensation method. Addressing the problem in existing multi-exposure feature processing where missing streams caused by high-exposure saturation are directly discarded, resulting in incomplete features for subsequent inference, this invention constructs a compensation process around the integrity of multi-exposure features: First, for each candidate position in the candidate box list, equal-sized image patches are cropped from high, medium, and low-exposure frames. A shared-weight lightweight convolutional encoder is used to extract fixed-length texture vectors and simultaneously calculate the saturation ratio. Next, a multi-threshold brightness consistency operator is used to determine specular saturation based on the high-exposure saturation ratio and cross-exposure difference, marking the missing high-exposure streams. Finally, a three-step search window is established for the missing streams, retaining only candidate nodes with complete information. The weighted similarity of the texture vectors of non-missing streams is calculated (introducing a distance attenuation coefficient), and the highest-scoring candidate's co-stream vector is scaled and copied to the current node to achieve compensation. This method prioritizes nodes with similar textures and the closest physical distance as compensation sources, suppressing the influence of illumination angle differences while ensuring the integrity of the three-stream texture features, providing continuous and reliable feature support for subsequent graph convolutional inference.

[0036] This invention innovatively proposes a crack-spot differentiation inference method combining multi-stream chain graph convolution with local mean re-evaluation. Addressing the problem that existing inference processes only output single-node crack probabilities and lack cross-disk coherent information modeling, leading to difficulty in eliminating isolated reflection false positives, this invention achieves accurate differentiation through a three-level process of graph construction, feature optimization, and inference re-evaluation. First, chain nodes carrying three-stream textures are assembled into a one-dimensional chain graph according to physical disk order. The actual disk spacing and the average brightness difference of the three streams are injected into each edge. High-exposure saturation markers are added to nodes and expanded into a fixed-length feature table. Second, the high-exposure texture vector is fed into an exposure gating unit, using the saturation markers and the average grayscale value of the three streams to output a gating coefficient that attenuates the high-light interference vector. The data is then concatenated with medium / low exposure vectors to update features. This is followed by two layers of chained graph convolution (the first layer integrates neighboring disk information, and the second layer introduces cross-disk residuals to expand the receptive field) and a coherence suppressor (calculating the difference in prediction scores between nodes and their preceding and following neighbors to weaken isolated peaks) to output the initial crack probability. Finally, a sliding window of length three is used to calculate the local mean. If the probability of a single node is lower than the local mean setpoint and high exposure saturation occurs, it is marked as specular reflection. If the probabilities of three adjacent nodes are all high and the low exposure flow is complete, the probability of the middle node is increased. Through cross-disk coherence information modeling and local mean re-evaluation, accurate distinction between specular reflection and real cracks is achieved, significantly reducing false positives and false negatives. The final output is a reliable list of results sorted by disk number.

[0037] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0038] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0039] Figure 1 This is a schematic diagram comparing the existing technology with the method provided in this embodiment; Figure 2 One example is a method flowchart provided in this embodiment; Figure 3 This is a network architecture diagram implemented using a method provided in one embodiment; Figure 4 This is a schematic diagram of step S3 of the method provided in one embodiment; Figure 5 This is a schematic diagram of step S4 of the method provided in one embodiment; Figure 6 This is a schematic diagram of step S5 of a method provided in one embodiment; Figure 7This is a schematic diagram of step S6 of the method provided in one embodiment. Detailed Implementation

[0040] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0041] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0042] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0043] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0044] Example 1 As described in the background section, under the combined conditions of high contrast, frequent specular reflection, and strict inspection time limits, existing technologies still lack a reliable means to instantly detect cracks and light spots.

[0045] To address the aforementioned issues, this embodiment provides a method for rapidly distinguishing between microcracks and sunlight reflections on insulators using unmanned aerial vehicles (UAVs). During high-altitude UAV inspections, by planning closed flight paths and simultaneously capturing images with three exposure levels, combined with a dual-branch detection network, chain-like geometric topology constraints, and one-dimensional multi-flow graph convolution, the method can distinguish between microcracks in insulators and sunlight reflections in real time, reducing false alarms and missed alarms, and providing accurate information for power outage maintenance decisions.

[0046] Specifically, a method for rapid differentiation of cracks and light spots in drone insulators, such as... Figure 1 , Figure 2 As shown, it includes the following steps: S1. On the UAV, set the insulator string axis as the center of the circle and plan a concentric closed curve trajectory, and set a constant speed, fixed focal length and pitch angle; at the trajectory sampling point, continuously trigger high, medium and low exposure to the insulator to complete three frames of shooting, calculate the offset vector through structural template matching and drive the gimbal to fine adjust, so that the three frames of field of view are consistent within the sub-pixel range; S2. Input the three frames of images into the improved target detection network. The highlight component of the high-exposure image and the dark texture component of the same frame are input into the first branch and the second branch respectively. The specular reflection samples are given weak negative labels and the loss weight is adaptively adjusted according to the batch ratio, so that the network outputs a set of crack candidate boxes. S3. Perform morphological processing and Hough circle retrieval on the grayscale image of the same frame to obtain the set of insulator disks along the center of the circle, sort them according to the direction of the principal component to form the main chain, and match the candidate box with the center of the main chain one-to-one, remove redundant candidates and generate a candidate list arranged in the order of disk number. S4. At each candidate position, crop the same size patch from high, medium and low exposure frames, extract each exposure texture vector and calculate the saturation ratio, mark the high exposure saturation stream according to the multi-threshold brightness consistency operator, and compensate the missing stream according to the principle of nearest neighbor priority, texture similarity and distance attenuation, and construct a chain node object carrying the texture of the three streams. S5. The node objects are arranged into a one-dimensional chain graph according to the physical disk order. High exposure saturation marks are injected into the nodes and the saturation channels are attenuated by the exposure gating unit. Then, the nodes are processed by two layers of chain graph convolution with cross-film residuals and coherence suppressor to obtain the crack probability of each node. S6. Apply a sliding window of length three to calculate the local mean of the crack probability. If the probability of a single node is lower than the preset proportion of its local mean and the high exposure saturation is marked as true, then mark this node as specular reflection. If the probabilities of three adjacent nodes are all high and the low exposure flow is complete, then increase the probability of one of the nodes. Finally, output a list of crack or spot results sorted by batch number.

[0047] In the implementation process of this embodiment, as follows: Figure 3 As shown, a network architecture is adopted. The overall architecture first uses a lightweight convolution with two branches for highlight / dark texture to connect to the improved YOLOv13 backbone to realize candidate crack detection; geometric matching aligns the candidate boxes with the disk edge and clips the three-exposure local, and the texture encoder extracts 128-dimensional features; exposure gating uses saturation markers to attenuate the highlight channel, and then assembles the features into a chain graph in physical order, which is fed into two layers of GraphConv containing cross-disk residuals to capture cross-disk coherent dark textures; finally, the fully connected layer outputs the crack probability, realizing the suppression of specular reflection and the detection of real cracks.

[0048] Of course, other network model structures or modules can also be used in other embodiments. These will not be elaborated upon here.

[0049] The following is a detailed description of each sub-step.

[0050] Step S1 includes the following sub-steps: S11. Plan the flight trajectory of the drone as a circle with the insulator string axis as the center and a radius of... The concentric closed curves, in which This indicates the horizontal distance from the fuselage to the insulator axis; simultaneously, a constant airspeed is written to the flight controller. and fixed focal length and pitch angle The above three parameters are used to limit... (1) Spatial displacement of adjacent frames, (2) Imaging magnification, (3) Perpendicularity of lens optical axis to insulator axis.

[0051] When the body moves at a constant speed along a closed curve and maintains When the angle of side-incident sunlight remains constant, it is compressed to a minimum range, and specular reflection can only appear within a limited field of view, thus reducing reflection interference at the source of the image; simultaneously, by fixing the track overlap rate at... The above ensures that each insulator always falls in the center area of ​​the sensor for imaging.

[0052] S12. For each preset trajectory sampling point on the closed curve ( ) adopt high ( ),middle( ),Low( Three exposure sequences of continuous shutter speeds, among which , , These represent the sets of exposure parameters for the same image sensor at three gain levels; triggered by hardware during a single exposure cycle. Three frames are captured within the time frame to generate a brightness-complementary image triplet. In this triplet, Micro-cracks in dark areas were completely preserved. Suppressing high-saturation regions, Provides texture references, mapping crack details and specular reflections to the most suitable exposure settings.

[0053] S13. During three consecutive exposures, high-contrast structures such as tower crossarms, joint fittings, and insulator pins are used as visual references for the three frames. , , Perform real-time template matching and calculate the reference offset vector. .Will Mapped to gimbal fine-tuning commands ,in For yaw compensation angle, For pitch compensation angle; Step S2 includes the following sub-steps: S21. Reconstruct the input layer of the improved target detection network from a single channel to a dual-branch system: The first branch receives the highlight component from the high-exposure image, denoted as... ; The second branch receives the dark stripe component from the same frame image, denoted as... .

[0054] right and Each of the ingress cascaded lightweight convolutional kernel groups and This decouples the specular reflections from the crack texture at the shallowest layer of the network.

[0055] In addition to the above-mentioned entry point changes, the parameters of subsequent layers of the original YOLOv13 backbone network will remain unchanged to ensure that the overall computation graph is directly compatible with the existing embedded inference framework, providing a separable feature tensor flow for chain graph discrimination.

[0056] S22. Introduce weak negative labels to the specular reflective areas in the training set, marking these areas as background categories that need to be identified but are not critical to the power outage decision.

[0057] The weight of this category in the loss function is set as the batch adaptive coefficient. And after each forward propagation, it is dynamically updated based on the proportion of specular reflective samples in the current batch. It automatically weakens the gradient contribution when the number of reflective samples surges.

[0058] The established online reweighting strategy causes the network to focus more attention on the dark pattern branches. The learning direction is not deviated due to the high contrast of the mirror bright area, which is the edge of the smallest crack in the mirror, so as to retain a stable gradient signal related to the crack for the subsequent candidate generation process.

[0059] S23. The dual-branch outputs are selected for candidate retention based on the principle of prioritizing dark patterns and delaying reflections. The specific technical approach is as follows: First, regarding the dark pattern branches With the high-light branch The original detection output is used to extract candidate boxes. , area and confidence level Online calculation of adaptive thresholds for dark ripple branches The following formula is used to comprehensively consider the area-confidence comparison of the two branches:

[0060] in, An adaptive confidence threshold for dark ripple branches; This is the system's baseline threshold. This is the threshold adjustment coefficient; , For the first The first dark pattern frame and the first Confidence level of each highlight frame; , This represents the pixel area of ​​the corresponding candidate box; To prevent tiny constants with a denominator of zero.

[0061] Second, by Import the dark pattern border into the priority list and only for Import highlight frames into the backup list This enables proactive release of weak responses to cracks and immediate suppression of strong responses to pure reflection. A fixed threshold is used to maintain the repulsion strength against large-area speckles.

[0062] Third, according to the chained interface from the priority list to the downstream topology sequence, Directly input the data into the next processing unit for structured reasoning; at the same time... storage.

[0063] This deferred mechanism allows for crack location to be completed during the drone's loiter time, followed by background cleanup of potential false positives due to reflections, ensuring that the assessment is triggered only by a genuine defect.

[0064] S24. Using the improved network described above, perform pixel-by-pixel inference on the same frame image and output a candidate box set in real time. ,in The coordinates of the candidate box center are: , Let be the side length. The candidate confidence level is determined, and visual feature vectors of the corresponding level are extracted simultaneously. Considering the real-time requirements of high-altitude operations, a small number of false detections of specular reflections are allowed in the dataset, which will be quickly compressed in the subsequent topology sequence filtering stage.

[0065] like Figure 4 As shown, step S3 includes the following sub-steps: S31. Denote the set of candidate box center coordinates output in step S2 as... ,in Indicates the first The center pixel position of each candidate box; for the grayscale image of the same frame. Using a radius of The circular structural elements are subjected to opening and closing operations to enhance the grayscale contrast between the edge of the insulator disk and the background, so that the contour response can still be stably locked in high-contrast scenes.

[0066] The radius range for processing the image is: Hough circle search yields a candidate set of arcs. For each Calculate the center of the circle and mean curvature and remove them in turn. Above the threshold The pseudo-circular arc and the semicircle with radius loss due to conductor obstruction. The final output is a set of circles covering the entire edge of the insulator disc and unaffected by shadows or reflections. .

[0067] S32, Set the centers of the circles Input principal component direction analysis and extract the first principal axis vector. As the elevation baseline of the insulator string; for each center of the circle Calculate its in Projected scalar on and will Generate a physical order index by sorting the values ​​in ascending order. At the same time, based on The orientation and airborne attitude information are used to determine the shooting side, ensuring that the indexing direction from bottom to top or from top to bottom is consistent with the on-site installation direction, and avoiding subsequent crack positioning deviations caused by disc replacement.

[0068] S33. Perform single-chain tracking on the sorting results: sequentially link adjacent center points... and Connect the lines, if the length of the line segment is... Falling within the rated disc spacing tolerance range Then mark them as valid edges and merge them into the same chain; if If the boundary exceeds the specified range, the chain is considered broken, and the current branch is terminated. For multiple chains appearing in the same frame, the chain with the most edges is selected as the main chain. At the same time, duplicate or isolated circles falling into the background are removed to ensure that only a complete geometric skeleton running through the entire edge of the disk is retained, laying a unique topology for the construction of the exposure multi-flow map.

[0069] S34. The set of candidate boxes output in step S2. With the main chain center sequence obtained in step S33 Link: A. Each candidate box The center coordinates are Width and height are recorded as The confidence level of the crack is denoted as ; center of each circle The coordinates are as The corresponding radius of the arc is denoted as Construct a composite cost matrix using two types of features. The Hungarian allocation is used to find the minimum cost match. :

[0070] in, It is an injective set from the center index to the candidate box index; To prevent the stability constant from having a denominator of zero; Align weights with radius; The meanings of the remaining symbols remain consistent with those above.

[0071] B. For each center point Read the matching results If the corresponding replacement value is higher than the threshold If no reliable crack candidate exists in this area, it is considered missing; if the cost is below the threshold, the matching box is written to the center label, and the distance from the center of this box to the center is less than the threshold. The remaining boxes are uniformly marked as redundant to prevent duplicate counting on the same disk.

[0072] C. Arrange the filtered matching boxes in the main chain order. Rearrange to generate a one-dimensional sequential list. Within the same step, the missing and redundant bit indices are recorded synchronously to provide a position mask for graph neural network compensation and reflection removal.

[0073] At this point, the candidate information corresponds to the actual number of sessions, and false positives due to empty background straying are eliminated.

[0074] S35, For sequential lists Add disk numbers sequentially pole identification number The candidate box coordinates, size, crack confidence, and deep visual features are encapsulated together into a structured sequence object. .Will It is then directly fed into the next step of the chain-like graph neural network module to connect the candidate layer with the graph inference layer.

[0075] like Figure 5 As shown, step S4 includes the following sub-steps: S41. Sequentialize the list from step S3. The center coordinates of each candidate box are mapped to the low-exposure frame at the same time. , medium exposure frame High-exposure frames Three images; a fixed size is used at the mapped coordinates. The pixel window is used to crop a local patch, and the results are obtained respectively. .

[0076] S42. For each flow block ( Lightweight convolutional encoder with shared weights Extracting fixed-length texture vectors During the encoding process, the average gray level of this image block is read simultaneously. With saturation pixel ratio and with The strong reflection peaks and weak crack ridges are mapped to the same feature dimension, preserving interpretable brightness indicators for subsequent missing value determination and preventing light spots from obscuring the crack clues needed for backend inference.

[0077] S43. Arrange the three texture vectors of the same node according to... The exposure order is concatenated as And add disk number With linked indexes Packaged into multi-stream node objects .

[0078] S44. Input the three-flow maps sequentially into the multi-threshold brightness consistency operator proposed in this application to adjust the exposure. Calculate saturation ratio (Refers to the pixel ratio where the grayscale within the window fills 5% of the sensor's upper limit), and then the specular saturation is judged based on the cross-current difference. The core judgment index is:

[0079] in, Boolean flag for mirror saturation; Low, medium, and high exposure saturation ratios; This is the high exposure saturation threshold; This is the threshold for cross-exposure difference.

[0080] like If the current node is found to have specular saturation in the high-exposure stream, the operator will then mark the current stream as missing, prompting subsequent steps to perform texture compensation without changing the data of the other two streams, thus ensuring that the chained inference always has complete crack information.

[0081] S45. Treat the exposure streams marked as missing in step S44 as targets to be compensated, and implement the neighbor-disk similarity texture compensation operator proposed in this application according to the three-level logic of nearest neighbor priority, texture similarity, and distance attenuation.

[0082] Specifically, firstly, for chained indexes... nodes Create vertical search window Only retain the three-step range and in the missing flow Candidate nodes with complete information.

[0083] For each candidate The texture vectors of the non-missing flows are calculated and a distance decay coefficient is introduced to form a weighted similarity. Then, the highest-scoring candidate co-flow vectors are scaled and copied to the current node to achieve real-time compensation. The overall mapping relationship is as follows: ; in, For the exposure stream after compensation Texture vector; For the original exposure stream Texture vector; , For nodes , The spliced ​​texture vectors on both streams without missing elements; This is the distance attenuation coefficient; This is an adaptive scaling factor for the copied texture; For the physical order index of the nodes; and For vector dot product and Norm.

[0084] Through the above segmented mapping, the node with the most similar texture and the closest physical distance is preferentially selected as the compensation source within the chain structure. Suppress the difference in lighting angle caused by excessive span, and use cosine similarity to ensure the consistency of material texture; If no available source is found within the three-step window, the missing source will remain in the missing state and will be automatically ignored by the backend discrimination layer.

[0085] like Figure 6 As shown, step S5 includes the following sub-steps: S51. The multi-stream node object output in step S4 According to physical order Assembled into a one-dimensional chain diagram ; For the set of nodes Each of them Only retain the pointers to the adjacent upper and lower plates. , The two directed edges are used to avoid interference from side branches.

[0086] Set of opposite edges Each of the items Inject two attributes: one is the actual disk spacing. Secondly, the average brightness difference among the three streams. Add three-stream texture vectors to the node body. Disk Number and high exposure saturation mark And uniformly expand into a fixed-length feature table .

[0087] The chain diagram follows the insulator installation sequence, allowing subsequent diagram convolutions to be recursively applied along physically connected paths, thus fully preserving the extension clues of cracks on the continuous disk surface.

[0088] S52, Node Features High-exposure texture vectors in Feed into exposure gating unit This unit consists of two fully connected layers, labeled with saturation. Compared with the average gray level of the three streams As input, output a gating coefficient between zero and one. .use right Attenuation and matching with medium and low exposure vectors , splicing to form updated node features .

[0089] By utilizing the unique saturation characteristics of specular reflection in the high-exposure channel, the impact of flash spots on subsequent inference is actively reduced, while ensuring that crack and dark pattern information is fully preserved in the medium and low-exposure channels, thus matching the high-contrast scene on site.

[0090] S53, Features of gated nodes Implement the first layer of chain graph convolution: Combine the features of the current node and its preceding and following neighboring nodes according to edge attributes. Weights generated by linear mapping Perform a weighted average and output the new node representation. By integrating the texture of this disc with the continuous information of adjacent discs, the continuity of cracks appearing in sheet-like patterns along the cracks is enhanced; High-exposure channels have been gated and suppressed, and isolated reflections lacking cross-disk support will be automatically diluted during aggregation, providing a consistent basis for judgment.

[0091] S54. Introduce cross-disk residuals in the second-layer graph convolution: In addition to directly aggregating first-order adjacency information, incorporate the residuals from nodes between the two disks. Features and residual paths are used to connect with the current node. Superposition yields the extended field representation. Since cracks often span three to four disk lengths, the receptive field is lengthened within the network to maintain edge gradient continuity. Layer normalization is applied to the outputs of the two layers to balance contributions from different distances, preventing distant anomalous flashes from being mistakenly amplified and ensuring that the network continuously focuses on structural dark ripple cues.

[0092] S55. Node representation after convolution stacking Apply a coherence suppressor: Calculate the bidirectional difference between the prediction scores of the current node and the nodes before and after it. ,like and At the same time, it is higher than the dynamic threshold. If this node is considered an isolated peak lacking corresponding upper and lower peaks, it will be multiplied by a suppression factor. Threshold It is estimated in real time from the mean and variance of the on-chain score.

[0093] S56. The node vector after coherence suppression... Send to terminal sorting head A fully connected layer with a sigmoid function is used to output the crack probability. and the disk number With exposure index Package. For nodes weakened by the coherence suppressor but still above the empirical lower limit, add an additional reflective hint field for quick verification by the operations team.

[0094] like Figure 7 As shown, step S6 includes the following sub-steps: S61. Denote the crack probabilities output from step S5, arranged in disk order, as a vector. ,in The linked index is represented as The probability of cracking in the insulator disc; using a sliding window of length three, for each center node Take the probabilities of its upper and lower adjacent nodes, for a total of three terms. Calculate window mean And at the same time retain the node's own probability. .Will As a local continuous reference, it enables subsequent steps to revise single-point anomalies based on the constraint that the distance between physically adjacent disks is constant.

[0095] S62. Perform deviation determination for each node: if the node crack probability... Below its corresponding window mean Fifty percent, and the high-exposure channel saturation marker of this node. If the value is equal to one, then this node is inferred to be an isolated highlight caused by specular reflection; Set its crack probability to zero and write the spot identifier in the node label.

[0096] S63. For each consecutive three-node segment in a chained vector Application enhancement rule: When the probability of cracks at all three nodes is higher than the mean of their respective windows. Furthermore, at least two nodes exhibit complete dark patterns in the low-exposure stream (i.e., low-exposure saturation ratio). If the probability of a crack at an intermediate node is less than a preset threshold, then a true cross-disk crack is inferred to exist. In this case, the crack probability at the intermediate node is multiplied by a weighting coefficient. To highlight the defect center while maintaining the original values ​​of the nodes on both sides; S64. Repackage the weakened and enhanced crack probabilities into a classification result list. ,in Number the towers, For disk number, The revised crack probability, Identify cracks or light spots; then add the shooting timestamp and exposure level information. (Press...) After sorting from high to low, the list is pushed to the operation and maintenance visualization interface and highlighted in real time on the corresponding blocks.

[0097] To verify the effectiveness of the above method, the following verification embodiment is provided.

[0098] A provincial power company, in conjunction with a smart inspection technology center, conducted a pilot project on rapid insulator defect identification under high-contrast conditions on the K-II line 110kV distribution corridor from March 2023 to February 2024. The pilot section was 38.4km long, with 128 towers and 9860 equivalent insulator discs. In 2022, the false alarm rate due to specular reflection on the same line reached as high as 18.7%. To thoroughly evaluate the defect identification effect under specular interference, the project team introduced the technical method of this application as the core algorithm.

[0099] During the data acquisition phase, the project team used two hexacopter UAVs to complete panoramic photography of each tower side according to the concentric closed curve specified in this application. The airborne camera was equipped with a global shutter CMOS sensor and fixed three exposure sequences: low exposure 1 / 4000s, reference exposure 1 / 1600s, and high exposure 1 / 400s. Three complementary images were captured continuously within 120ms at each sampling point, and the distance between the airframe and the insulator axis was recorded by laser ranging. To ensure training diversity, the reference exposure was manually reduced by one stop in the V-band where shadow interference was severe. During the pilot phase, a total of 41,275 sets of three-stream images were collected. After automatic dehazing and rotor stripe removal, a training set of 6,412 images with marked crack areas and 15,038 images with specular reflection areas, and a test set of 7,118 images with mixed crack and reflective surfaces were generated.

[0100] During model training, the improved YOLOv13 was run on an NVIDIA Jetson AGX with an input resolution of 768×768 and 16 kernels in each of the first layer of the dual-branch convolutional network. The weak negative label mirror weight λ was initially set to 0.4 and then dynamically adjusted according to the batch ratio. After 30 rounds of convergence, the object detection network output crack candidates, which were then subjected to two layers of 64-dimensional convolution and coherence suppression by a one-dimensional chained multi-flow graph neural network, with the total frame-level latency controlled at 142ms. The final model parameters were locked on November 15, 2023, with the detection threshold θ set to 0.28 and the coherence suppression threshold τ calculated in real time from the chain mean.

[0101] During the on-site deployment phase, the system underwent a full-scale re-flight test on the K-II line in January 2024, completing image acquisition, on-the-fly calculation, and simultaneous result transmission for the entire line in just 3.9 hours. The algorithm output a total of 152 crack alerts and 487 spot alerts. The on-site acceptance process employed a two-level verification process involving infrared thermography and tower sampling. Verification showed a crack detection accuracy of 92.8%, a recall rate of 90.4%, and a false alarm rate of 4.6% for specular reflection. This represents a significant improvement compared to the conventional single-frame YOLO detection solution in 2022 (accuracy of 81.9%, false alarm rate of 18.7%). The average decision-making closed-loop time was reduced from 5.5 days in the traditional process to 18 hours.

[0102] To visually demonstrate the algorithm's performance, the project team selected three detection schemes for comparison and evaluation on the same line from 2022 to 2024: Table 1 Performance Comparison of Technical Methods in this Application and Comparative Solutions

[0103] As shown in the table, the technology presented in this application maintains high accuracy and significantly suppresses false alarms in high-contrast mirror environments, while the inference latency meets the requirements for online computing by UAVs. The maintenance team can confirm the power outage maintenance plan by the evening of the same day, and the inspection-decision-work order flow has been compressed to a closed loop within one day for the first time, reducing the amount of manual secondary verification by more than 70%. The project team plans to further integrate this model into the provincial company's standard aerial survey kit and promote its application to routine inspection tasks of 35kV and above lines.

[0104] Example 2 An insulator crack-spot differentiation system includes: The insulator image acquisition module is configured to acquire three frames of images of the insulator under high, medium and low exposure levels. The sampling trajectory of the image is a concentric closed curve trajectory planned with the axis of the insulator string as the center. The target detection module is configured to input three frames of images into a pre-trained dual-branch target detection network. The highlight component of the high-exposure image and the dark texture component of the same frame are respectively input into the first and second branches of the target detection network. The specular reflection samples are given weak negative labels and the loss weights are adaptively adjusted according to the batch ratio. The target detection network outputs a set of crack candidate boxes. The candidate box preprocessing module is configured to process the grayscale image of the same frame, obtain the set of insulator disks along the center, sort them according to the principal component direction to form the main chain, match the candidate boxes in the crack candidate box set with the center of the main chain one-to-one, remove redundant candidate boxes and generate a list of candidate boxes arranged in the order of disk number. The adjacent disk compensation module is configured to extract each exposure texture vector and calculate the saturation ratio for each candidate position in the candidate box list, cropping a patch of the same size from high, medium, and low exposure frames, marking the high exposure saturation stream, performing adjacent disk compensation for the missing stream, and constructing a chain node object carrying the texture of the three streams. The node crack probability calculation module is configured to form a one-dimensional chain graph of the chain node objects according to the physical disk number, inject high exposure saturation markers into the nodes, and calculate the crack probability of each node. The differentiation module is configured to calculate the local mean of the crack probability of each node. If the probability of a single node is lower than the preset proportion of its local mean and the high exposure saturation is marked as true, then the node is marked as specular reflection. If the probabilities of three adjacent nodes are all high and the low exposure flow is complete, then the probability of one of the nodes is increased. Finally, a list of crack or spot results sorted by batch number is output.

[0105] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0106] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0107] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0108] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0109] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for distinguishing insulator cracks from light spots, characterized in that, Includes the following steps: Three frames of images of the insulator under high, medium and low exposure levels were acquired. The sampling trajectory of the images was a concentric closed curve trajectory planned with the axis of the insulator string as the center. Three frames of images are input into a pre-trained dual-branch target detection network. The highlight component of the high-exposure image and the dark texture component of the same frame are input into the first and second branches of the target detection network, respectively. The specular reflection samples are given weak negative labels and the loss weights are adaptively adjusted according to the batch ratio. The target detection network outputs a set of crack candidate boxes. The grayscale images in the same frame are processed to obtain the set of insulator disks along the center of the circle. The main chain is formed by sorting the main components according to the direction of the principal components. The candidate boxes in the crack candidate box set are matched one-to-one with the center of the main chain. Redundant candidate boxes are removed and a list of candidate boxes arranged in the order of disk number is generated. Each candidate box in the candidate box list is placed in a high, medium, and low exposure frame clipping size tile. Each exposure texture vector is extracted and the saturation ratio is calculated. The high exposure saturation stream is marked, and the missing stream is compensated by adjacent disks. A chain node object carrying the three stream textures is constructed. The chain node objects are arranged into a one-dimensional chain graph according to the physical disk number, high exposure saturation marks are injected into the nodes, and the crack probability of each node is calculated. Calculate the local mean of the crack probability for each node. If the probability of a single node is lower than the preset proportion of its local mean and the high exposure saturation is marked as true, then the node is marked as specular reflection. If the probabilities of three adjacent nodes are all high and the low exposure flow is complete, then the probability of one of the nodes is increased. Finally, output a list of crack or spot results sorted by batch number.

2. The insulator crack-spot differentiation method as described in claim 1, characterized in that, The process of acquiring three frames of images of the insulator under high, medium, and low exposure conditions, and using the sampling trajectory of the images as a concentric closed curve trajectory planned with the insulator string axis as the center, includes: planning the UAV's flight trajectory as a circle with the insulator string axis as the center and a radius of... The concentric closed curves, in which Indicates the horizontal distance from the fuselage to the insulator axis; sets the constant flight speed and fixed focal length and pitch angle; Maintain the aircraft moving at a constant speed along a closed curve and keep the pitch angle constant, control the track overlap rate to be fixed above the set threshold, and ensure that each insulator always falls on the center area of ​​the sensor for imaging. For each preset track sampling point on the closed curve, three exposure levels (high, medium, and low) are used in succession with continuous shutter speeds to complete three frame captures within a single exposure cycle, generating a complementary image triplet. During three consecutive exposures, the contour contrast structure on the inspection line is used as a visual reference. Real-time template matching is performed on the elements of the triplet of the three frames to calculate the reference offset vector, and the reference offset vector is mapped to the gimbal fine-tuning command.

3. The insulator crack-spot differentiation method as described in claim 1, characterized in that, The dual-branch target detection network, based on the YOLOv13 target detection network, reconstructs its input layer from a single channel to a dual-branch system. The first branch receives the highlight component from the high-exposure image, denoted as... The second branch receives the dark texture component from the same frame image, denoted as... ; right and Each of the entry points is connected in series with a lightweight convolutional kernel group, so that the network decouples the speckle and crack texture at the shallowest layer. Keep the parameters of subsequent layers of the YOLOv13 object detection network unchanged.

4. The insulator crack-spot differentiation method as described in claim 1, characterized in that, During the training process, the target detection network introduces weak negative labels to the specular reflective areas in the training set, marking these areas as background categories that need to be identified but are not critical to the power outage decision. The weight of this category in the loss function is set as the batch adaptive coefficient. And after each forward propagation, it is dynamically updated based on the proportion of specular reflective samples in the current batch. It automatically weakens the gradient contribution when the number of reflective samples surges; The established online reweighting strategy causes the network to focus more attention on the dark pattern branches. The tiny crack edges in the mirror do not deviate from the learning direction due to the high contrast of the bright areas.

5. The insulator crack-spot differentiation method as described in claim 1, characterized in that, The process of the target detection network outputting a set of crack candidate boxes includes: extracting the area and confidence of the candidate boxes from the original detection outputs of the first and second branches respectively; and comprehensively considering the area-confidence comparison of the two branches, calculating an adaptive threshold online for the dark crack branches. Dark texture boxes with a confidence level greater than an adaptive threshold are imported into the priority list, while only highlight boxes with a confidence level greater than a fixed threshold are imported into the backup list. Perform structured reasoning on the priority list; simultaneously store the backup list; An object detection network is used to perform pixel-by-pixel reasoning on the same frame of images, output a set of candidate boxes, and simultaneously extract the visual feature vectors of the corresponding level.

6. The insulator crack-spot differentiation method as described in claim 5, characterized in that, Taking into account the area-confidence comparison of the two branches, the process of calculating the adaptive threshold online for the dark ripple branch includes: in, An adaptive confidence threshold for dark ripple branches; This is the system's baseline threshold. This is the threshold adjustment coefficient; , The first The first dark pattern frame and the first Confidence level of each highlight frame; , These are the candidate bounding boxes for the dark texture branch and the highlight branch, respectively. , Candidate boxes , The pixel area; To prevent tiny constants with a denominator of zero.

7. The insulator crack-spot differentiation method as described in claim 1, characterized in that, The process of processing the grayscale image of the same frame to obtain the set of circle centers along the insulator disk includes: performing opening and closing operations on the grayscale image of the same frame using circular structuring elements; performing Hough circle search on the processed image within a set radius range to obtain a set of candidate arcs; calculating the center and average curvature of each element in each candidate arc set; and sequentially removing pseudo arcs with an average curvature higher than the threshold and semicircles with radius defects caused by conductor shading to obtain a set of circle centers that covers the entire insulator disk edge and is not affected by shadows or reflections.

8. The insulator crack-spot differentiation method as described in claim 1, characterized in that, The process of forming the main chain based on the principal component direction sorting includes: performing principal component direction analysis on the set of circle centers, extracting the first principal axis vector as the pitch baseline of the insulator string; calculating the projection scalar of each circle center on the first principal axis vector, and sorting the projection scalars in ascending order of their values ​​to generate a physical order index. At the same time, the shooting side is determined based on the orientation of the first main axis and the airborne attitude information to ensure that the indexing direction from bottom to top or from top to bottom is consistent with the on-site installation direction; Connect adjacent circle centers sequentially. If the line segment length falls within the rated disc distance tolerance range, it is marked as a valid edge and the two are merged into the same chain; otherwise, it is judged as a broken chain and the current branch is terminated. For multiple chains appearing in the same frame, select the one with the most edges as the main chain, and at the same time remove duplicate or isolated circle centers that fall into the background to ensure that only one complete geometric skeleton that runs through the entire disc edge is retained, laying a unique topology for the construction of exposure multi-flow map.

9. The insulator crack-spot differentiation method as described in claim 1, characterized in that, The process of matching candidate boxes in the crack candidate box set with the main chain center one-to-one, removing redundant candidate boxes, and generating a candidate box list arranged in disk order includes: associating the candidate box set with the main chain center sequence. Based on the center coordinates, size parameters, and crack confidence of each candidate box, as well as the coordinates and size parameters of each circle center, a composite cost matrix is ​​constructed using two types of features, and the minimum cost match is found using Hungarian assignment. Determine the matching result for each circle center. If the corresponding value is higher than the threshold, it is determined that there are no reliable crack candidates in this plate and it is marked as missing. If the value is lower than the threshold, the matching box is written into the circle center label, and the remaining boxes whose distance from the center of this box is less than the set value are uniformly marked as redundant to prevent duplicate counting in the same plate. The filtered matching boxes are rearranged according to the main chain order to generate a one-dimensional sequential list, i.e., the candidate box list, and the missing and redundant position indices are recorded simultaneously. The disk number and tower identification number are added sequentially to the sequential list, and the candidate box coordinates, size parameters, crack confidence and deep visual features are encapsulated together into a structured sequence object.

10. The insulator crack-spot differentiation method as described in claim 1, characterized in that, in The process of cropping a block of each candidate frame in the candidate box list to a high, medium, and low exposure frame includes: mapping the center coordinates of each candidate frame in the candidate box list to three images at the same time: a low exposure frame, a medium exposure frame, and a high exposure frame; and cropping a local block at the mapped coordinates using a pixel window of a fixed size.

11. The insulator crack-spot differentiation method as described in claim 1, characterized in that, The process of extracting each exposure texture vector and calculating the saturation ratio, and labeling high exposure saturation streams includes: extracting fixed-length texture vectors for each stream patch using a lightweight convolutional encoder with shared weights; simultaneously reading the average gray level and saturation pixel ratio of the patch during the encoding process, and writing them into the node label along with the fixed-length texture vector; and mapping strong reflection peaks and weak crack ridges to the same feature dimension. The three texture vectors of the same node are concatenated in low-medium-high exposure order, and the disk number and chain index are appended to package them into a multi-stream node object; The three stream tiles are sequentially input into the multi-threshold brightness consistency operator. The multi-threshold brightness consistency operator is used to calculate the saturation ratio of the exposure. Then, the specular saturation is judged based on the cross-stream difference. If the saturation ratio is the set value, it is determined that the current node has specular saturation in the high-exposure stream, and the multi-threshold brightness consistency operator then marks this stream as missing.

12. The insulator crack-spot differentiation method as described in claim 1, characterized in that, The process of constructing a chain of node objects carrying three-stream textures to compensate for missing streams includes: treating the exposure streams marked as missing as the targets to be compensated, establishing upper and lower search windows for each node in the chain index, and retaining only candidate nodes within the three-step range that have complete information on the missing streams. For each candidate, the texture vector of the non-missing flow is calculated and a distance decay coefficient is introduced to form a weighted similarity; then the co-flow vector of the highest-scoring candidate is scaled and copied to the current node, and the adjacent disk similar texture compensation operator is used to perform segmented mapping to achieve real-time compensation; By using segmented mapping, the node with the most similar texture and the closest physical distance is selected as the compensation source within the chain structure to suppress the difference in lighting angle caused by excessive span, while cosine similarity is used to ensure the consistency of material texture. If no available source is found within the three-step window, the missing state will remain.

13. The insulator crack-spot differentiation method as described in claim 12, characterized in that, In piecewise mapping using the adjacent disk similarity texture compensation operator, the mapping relationship is as follows: in, For the exposure stream after compensation Texture vector; For the original exposure stream Texture vector; , For nodes , The spliced ​​texture vectors on both streams without missing elements; This is a vertical search window; This is the distance attenuation coefficient; This is an adaptive scaling factor for the copied texture; For the physical order index of the nodes; and For vector dot product and Norm.

14. The insulator crack-spot differentiation method as described in claim 1, characterized in that, The process of assembling the chain node objects into a one-dimensional chain graph according to the physical disk order includes: assembling the multi-flow node objects into a one-dimensional chain graph in physical order; and retaining only the two directed edges pointing to the adjacent upper and lower disks for each multi-flow node object in the node set. Each edge in the edge set is injected with two attributes: the actual disk spacing and the average brightness difference of the three streams. The node body is appended with the three stream texture vector, disk number and high exposure saturation mark, and uniformly expanded into a fixed-length feature table. The chain graph follows the insulator installation sequence.

15. The insulator crack-spot differentiation method as described in claim 1, characterized in that, The process of injecting high-exposure saturation markers into nodes and calculating the crack probability of each node includes: sending the high-exposure texture vector in the node features into the exposure gating unit; the exposure gating unit consists of two fully connected network layers, taking the saturation marker and the average gray level of the three streams as inputs, and outputting a gating coefficient of zero to one; the gating coefficient is used to attenuate the high-exposure texture vector and concatenate it with the medium and low-exposure vectors to form the updated node features; The first layer of chain graph convolution is applied to the gated node features: the features of the current node and its preceding and following neighboring nodes are weighted and averaged according to the weights generated by the linear mapping of edge attributes, and a new node representation is output, which fuses the texture of the current disk with the continuous information of the adjacent disks; In the second layer of graph convolution, cross-slice residuals are introduced: in addition to directly aggregating first-order adjacency information, features of distances to two disk nodes are incorporated and superimposed on the current node through the residual path to obtain the expanded domain representation. Layer normalization is applied to the outputs of the two layers to balance the contributions of different distances. A coherence suppressor is applied to the node representation after convolution stacking: the bidirectional difference between the predicted scores of the current node and the nodes before and after it is calculated. If the bidirectional difference is higher than the dynamic threshold at the same time, the node is identified as an isolated peak lacking coherence and multiplied by the suppression factor. The dynamic threshold is estimated in real time based on the mean and variance of the scores on the chain. The node vectors after coherence suppression are fed into the classification head, and a fully connected layer with a sigmoid function is used to output the crack probability. The disk number and exposure index are then packaged together.

16. The insulator crack-spot differentiation method as described in claim 1, characterized in that, The process of calculating the local mean of the crack probability at each node includes: recording the crack probabilities arranged in disk order as a vector, where each element in the vector represents the crack probability of the insulator disk corresponding to the chain index; A sliding window of length three is used. For each central node, the window mean is calculated by taking the probabilities of its three adjacent nodes above and below, while retaining the node's own probability. The window mean is used as a local continuous reference.

17. The insulator crack-spot differentiation method as described in claim 1, characterized in that, If the probability of a single node is lower than a preset proportion of its local mean and the high exposure saturation is marked as true, then the process of marking this node as specular reflection includes: For each node, a deviation judgment is performed. If the probability of a node crack is less than 50% of the average value of its corresponding window, and the saturation flag of the high exposure channel of this node is equal to one, then this node is inferred to be an isolated highlight caused by specular reflection. Set its crack probability to zero and write the spot identifier in the node label.

18. The insulator crack-spot differentiation method as described in claim 1, characterized in that, The process of enhancing the probability of nodes includes applying enhancement rules to each consecutive three-node segment in the chain vector: when the crack probability of the three nodes is higher than the mean of their respective windows, and at least two nodes present complete dark patterns in the low exposure stream, i.e., the low exposure saturation ratio is less than a preset threshold, it is inferred that there is a real cross-disk crack. At this time, the crack probability of the middle node is multiplied by a weighting coefficient to highlight the defect center, while keeping the original values ​​of the nodes on both sides. The weakened and enhanced crack probabilities are repackaged into a list of classification results, and then the shooting timestamp and exposure level information are added. The results are sorted from high to low according to the revised crack probabilities.

19. An insulator crack-spot differentiation system, characterized in that, include: The insulator image acquisition module is configured to acquire three frames of images of the insulator under high, medium and low exposure levels. The sampling trajectory of the image is a concentric closed curve trajectory planned with the axis of the insulator string as the center. The target detection module is configured to input three frames of images into a pre-trained dual-branch target detection network. The highlight component of the high-exposure image and the dark texture component of the same frame are respectively input into the first and second branches of the target detection network. The specular reflection samples are given weak negative labels and the loss weights are adaptively adjusted according to the batch ratio. The target detection network outputs a set of crack candidate boxes. The candidate box preprocessing module is configured to process the grayscale image of the same frame, obtain the set of insulator disks along the center, sort them according to the principal component direction to form the main chain, match the candidate boxes in the crack candidate box set with the center of the main chain one-to-one, remove redundant candidate boxes and generate a list of candidate boxes arranged in the order of disk number. The adjacent disk compensation module is configured to place each candidate position in the candidate box list into a high, medium, and low exposure frame crop size patch, extract each exposure texture vector and calculate the saturation ratio, mark the high exposure saturation stream, perform adjacent disk compensation for the missing stream, and construct a chain node object carrying the texture of the three streams. The node crack probability calculation module is configured to form a one-dimensional chain graph of the chain node objects according to the physical disk number, inject high exposure saturation markers into the nodes, and calculate the crack probability of each node. The differentiation module is configured to calculate the local mean of the crack probability of each node. If the probability of a single node is lower than the preset proportion of its local mean and the high exposure saturation is marked as true, then the node is marked as specular reflection. If the probabilities of three adjacent nodes are all high and the low exposure flow is complete, then the probability of one of the nodes is increased. Finally, a list of crack or spot results sorted by batch number is output.

Images