Airline wire breakage repair visual inspection method and system

Abstract

The present application belongs to the technical field of airborne wire harness maintenance and detection, and relates to a kind of airborne wire breakage repair visual detection method and system, method includes: collecting repair site image and carrying out distortion correction and brightness normalization to obtain standard image;The standard image is input into the target detection model to obtain the positioning result of the stripping area and the indentation area;Based on the positioning result, determine the segmentation input area and extract the area image, input the area image into the instance segmentation model to obtain the stripping mask and the indentation mask;Based on the mask, extract the wire axial direction and perform posture normalization, establish the conversion relationship between pixel quantity and physical quantity through scale calibration, determine the stripping length, the axial distance between the stripping end face and the indentation starting boundary, the number of indentations and the fracturing state;Compare the process characteristic quantity with the preset rule library to output the qualified, recheck or unqualified result.The technical scheme of the present application can improve the stability of the test criterion and the consistency of the quantitative review, and form a comparable and traceable test conclusion.

Description

Technical Field

[0001] This invention belongs to the field of airborne wiring harness repair and inspection technology, and specifically relates to a visual inspection method and system for repairing broken airborne wires. Background Technology

[0002] In existing technologies, the quality inspection of airborne power cable breakage repairs typically relies on manual visual inspection and simple measuring tool verification. This involves maintenance personnel observing and measuring the stripping condition, crimping marks, and related dimensional relationships on-site to confirm the consistency of the repair process and flight safety requirements. However, existing methods for inspecting the quality of airborne power cable breakage repairs have significant shortcomings in terms of the consistency of visual judgment and dimensional verification.

[0003] In practical applications, maintenance sites are often accompanied by fluctuations in lighting conditions, differences in shooting and observation angles, changes in line posture, and surface reflections. Although manual visual inspection or rough interpretation based on a single image can complete basic confirmation, the stability of the morphological discrimination criteria for stripping and indentation boundaries is weak, and the overall consistency of quantitative verification of key dimensional relationships such as stripping length and position spacing is low. This results in low comparability and traceability of inspection results from different personnel and different batches.

[0004] Therefore, it is evident that existing technologies often suffer from problems such as weak stability in the overall morphological identification criteria and low consistency in the quantitative verification of key dimensional relationships during the quality inspection of airborne power line breakage repairs. These are the shortcomings of existing technologies.

[0005] In view of this, it is very necessary to provide a visual inspection method and system for repairing broken airborne power lines in order to solve the above-mentioned defects in the prior art. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies in the quality inspection of airborne power line breakage repair, namely, the weak stability of the overall morphological discrimination criteria and the low consistency of quantitative verification of key dimensional relationships. This invention provides a visual inspection method and system for airborne power line breakage repair to solve the aforementioned technical problems.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] In a first aspect, this application provides a visual inspection method for repairing broken airborne power lines, including:

[0009] Images of the repaired area are acquired and preprocessed to obtain standardized images. The preprocessing includes distortion correction and brightness normalization.

[0010] The standardized image is input into the target detection model to obtain the localization results, which include the localization results of the stripped area and the indentation area.

[0011] Based on the localization results, the segmentation input region is determined. The corresponding region image is extracted from the standardized image and input into the instance segmentation model to obtain the stripping mask and the indentation mask.

[0012] Based on the stripping mask and indentation mask, the axial direction of the line is extracted and the orientation is normalized. The conversion relationship between the number of pixels and physical quantities is obtained through scale calibration. The stripping length, the axial distance between the stripping end face and the indentation start boundary, the number of indentations and the cracking state are determined and written into the process feature quantity.

[0013] The process feature quantities are compared with the preset rule library and the results are output as qualified, required to be reviewed, or unqualified.

[0014] By adopting the above technical solution, and through standardized processing of the images of the repaired area, combined with an integrated process of detection, segmentation and measurement, stable identification and consistent quantification of the key process states of airborne power line breakage repair can be achieved. This enables the formation of comparable and traceable inspection conclusions, meeting the requirements of high stability of morphological discrimination criteria and high consistency of quantitative verification of key dimensional relationships in the quality inspection of airborne power line breakage repair.

[0015] Among these methods, preprocessing based on distortion correction and brightness normalization ensures that the input image maintains relatively consistent imaging characteristics under illumination fluctuations, viewing angle differences, and line pose changes, thereby improving the reliability of subsequent recognition and measurement from the source. Target detection-based localization of the stripped and indented regions provides clear spatial constraints for determining the region of interest and reduces the impact of background interference on discrimination consistency. Extracting region images based on the localization results and performing instance segmentation to obtain stripped and indented masks provides higher precision in the boundary representation of stripped and indented areas and establishes a unified morphological criterion, thus improving the inspection performance across different batches. The comparability of results is ensured; further, the axial direction of the line is extracted based on the mask and the orientation is normalized so that the length and spacing measurements are established under a consistent axial reference. Then, the conversion relationship between pixel quantity and physical quantity is obtained by combining scale calibration, so that the process feature quantities such as stripping length, axial spacing between the stripping end face and the indentation starting boundary, number of scratches and cracking state have a unified quantitative expression. Finally, the conclusion of qualified, requiring review or unqualified is output by comparing through the rule base, so that the judgment basis and output conclusion present regular and traceable characteristics, and overall meet the requirements of conclusion stability and review consistency under the conditions of rapid field inspection.

[0016] Preferably, the steps for acquiring images of the repair area include:

[0017] A set of image capture poses is established for the repaired area. Images of the repaired area corresponding to each image capture pose in the set are collected and pose labels are generated.

[0018] The posture identifiers are associated with and stored with the corresponding images. When the set of posture identifiers corresponding to the repaired area does not meet the preset integrity conditions, the images corresponding to each imaging posture in the imaging posture set of the repaired area are re-acquired until the set of posture identifiers corresponding to the repaired area meets the preset integrity conditions.

[0019] By adopting the above technical solution, multi-pose acquisition of the repair area is performed using the image posture set and integrity closed-loop verification is performed using posture identifiers. This achieves coverage constraints on key observation angles, which can reduce boundary discrimination and measurement deviations caused by missing angles and improve the comparability and stability of test results from different batches.

[0020] Preferably, the preprocessing also includes: performing reflection highlight suppression after distortion correction, performing sharpness evaluation after brightness normalization, and when the sharpness evaluation does not meet the preset sharpness conditions, marking the corresponding standardized image as a verification image and associating the verification image identifier in the output result.

[0021] By adopting the above technical solution, an image quality gating mechanism is formed by using reflection highlight suppression and sharpness evaluation to achieve reliable identification and associated output of reflective and blurred scenes. This can improve the effectiveness and traceability of standardized images and reduce the probability of misjudgment caused by low-quality input.

[0022] Preferably, the steps of inputting the standardized image into the target detection model to obtain the localization result include:

[0023] Multi-scale inference is performed on the standardized image to obtain a candidate localization set, and overlap suppression is performed on the candidate localization set to obtain the localization results of the stripped area and the indentation area.

[0024] Confidence indicators are generated for the location results of the stripped area and the indentation area. When any confidence indicator does not meet the preset confidence conditions, the corresponding location result is marked as a verification location result and associated with the confidence indicator in the output result.

[0025] By adopting the above technical solution, combining multi-scale inference and overlap suppression to generate positioning results and introducing confidence labels and verification marks, explicit management of positioning uncertainty can be achieved, which can improve the consistency and robustness of positioning results and make the output in the case of low confidence more in line with the verification consistency requirements.

[0026] Preferably, the steps of determining the segmentation input region based on the localization result, extracting the region image corresponding to the segmentation input region from the standardized image, and inputting the region image into the instance segmentation model to obtain the stripping mask and indentation mask include:

[0027] Based on the location results of the stripped area and the indentation area, a segmentation input region is generated and boundary expansion is performed to obtain an expanded segmentation input region.

[0028] Background suppression is performed within the extended segmentation input region to obtain a region image. The region image is then input into the instance segmentation model to obtain a stripping mask and an indentation mask.

[0029] Calculate the coverage consistency index between the stripping mask and the corresponding positioning result for the stripping mask and the indentation mask respectively, and write the coverage consistency index into the process feature quantity.

[0030] By adopting the above technical solution, a more focused segmentation input is constructed through boundary expansion and background suppression, and the coverage consistency index is calculated and written into the process feature quantity. This achieves structured consistency constraints on the localization and segmentation results, which can improve the stability and interpretability of the mask quality and enhance the reliability of the process feature quantity calculation.

[0031] Preferably, the step of extracting the wire axis based on the stripping mask and indentation mask and performing attitude normalization to determine the axial distance between the stripping end face and the indentation starting boundary includes:

[0032] Extract the center line of the line and determine the axis of the line within the union of the stripping mask and the indentation mask;

[0033] Perform coordinate rotation normalization on the standardized image based on the line axis;

[0034] In the rotated and normalized coordinate system, the stripping end face is determined based on the end boundary of the stripping mask in the axial projection, the indentation start boundary is determined based on the end boundary of the indentation mask in the axial projection, and the axial spacing is obtained based on the interval between the stripping end face and the indentation start boundary in the axial projection.

[0035] By adopting the above technical solution, the center line is extracted based on the union range and the interval between the end face and the starting boundary is defined by axial projection, so as to realize a unified geometric measurement caliber for axial spacing. This can reduce the impact of attitude differences on distance measurement and improve the consistency and comparability of quantitative verification of key dimension relationships.

[0036] Preferably, the steps of determining the number of notches and writing the process feature quantities include:

[0037] Within the area defined by the stripping mask, a scratch detection sub-region is constructed, and fine line structure enhancement is performed within the scratch detection sub-region to obtain a set of fine line candidates;

[0038] Based on the axial direction of the line, a candidate set of indentations is obtained by performing directional consistency screening on the candidate set of thin lines.

[0039] The number of notches is obtained by performing connectivity analysis and length consistency screening on the candidate notch set, and the number of notches is written into the process feature quantity.

[0040] By adopting the above technical solution, and combining fine line structure enhancement and direction consistency screening with connectivity and length consistency analysis, stable extraction and quantity statistics of the scratch target can be achieved. This can reduce the interference of texture noise and background fine lines on the counting, and improve the reliability and consistency of the scratch quantity representation.

[0041] Preferably, the steps of comparing process feature quantities with a preset rule base and outputting qualified, required, or unqualified results include:

[0042] Perform consistency checks on the process feature quantities obtained from different images according to the posture identifier. When the consistency check does not meet the preset consistency conditions, output the result that needs to be reviewed and record the inconsistency type identifier.

[0043] When the consistency check meets the preset conditions, the process feature quantities are matched one by one according to the hierarchical comparison order of the preset rule base, and the qualified or unqualified results are output, and the hit rule identifier is recorded.

[0044] By adopting the above technical solution, the consistency verification of the posture dimension and the hierarchical rule matching output are used to record the inconsistency type and the hit rule identifier, so as to realize the rule-based judgment and source expression under the fusion of multi-view results, which can improve the accuracy of the triggers that need to be reviewed and the traceability of the conclusions.

[0045] Secondly, this application also provides a visual inspection system for airborne power line breakage repair, comprising:

[0046] The imaging unit is used to acquire images of the repaired area;

[0047] The preprocessing unit is used to perform preprocessing on the image to obtain a standardized image. The preprocessing includes distortion correction and brightness normalization.

[0048] The localization unit is used to input the standardized image into the target detection model to obtain the localization results, which include the localization results of the stripped area and the indentation area.

[0049] The segmentation unit is used to determine the segmentation input region based on the localization result, extract the region image corresponding to the segmentation input region from the standardized image, and input the region image into the instance segmentation model to obtain the stripping mask and the indentation mask.

[0050] The measurement unit is used to extract the axial direction of the line body based on the stripping mask and the indentation mask and perform orientation normalization. It obtains the conversion relationship between pixel quantity and physical quantity through scale calibration, determines the stripping length, the axial distance between the stripping end face and the indentation start boundary, the number of scratches and the cracking state, and writes them into the process feature quantity.

[0051] The judgment unit is used to compare the process feature quantity with the preset rule library and output the qualified result, the result that needs to be reviewed, or the unqualified result.

[0052] Preferably, the segmentation unit includes a region generation subunit, a boundary expansion subunit, a background suppression subunit, and a mask output subunit, and the determination unit includes a consistency calculation subunit and an index writing subunit, wherein...

[0053] The region generation subunit is used to generate a segmented input region based on the location results of the stripping region and the indentation region.

[0054] The boundary expansion subunit is used to perform boundary expansion on the segmented input region to obtain an expanded segmented input region.

[0055] The background suppression subunit is used to perform background suppression within the extended segmented input region to obtain a region image;

[0056] The mask output subunit is used to input the region image into the instance segmentation model to obtain the stripping mask and the indentation mask;

[0057] The consistency calculation subunit is used to calculate the coverage consistency index of the stripping mask and the stripping area positioning result, as well as the indentation mask and the indentation area positioning result, respectively.

[0058] The indicator writing sub-unit is used to write the coverage consistency indicator into the process feature quantity.

[0059] As can be seen from the above technical solutions, the present invention has the following advantages:

[0060] This application provides a visual inspection method and system for airborne power line breakage repair. By standardizing the image of the repair area and combining it with an integrated process of detection, segmentation and measurement judgment, it achieves stable identification and consistent quantification of key process states in airborne power line breakage repair. It can form comparable and traceable inspection conclusions, meeting the requirements of high stability of morphological discrimination criteria and high consistency of quantitative verification of key dimensional relationships in the quality inspection of airborne power line breakage repair.

[0061] Furthermore, the design principle of this invention is reliable, the structure is simple, and it has a very wide range of application prospects.

[0062] Therefore, it is evident that the present invention has outstanding substantive features and significant progress compared with the prior art, and the beneficial effects of its implementation are also obvious. Attached Figure Description

[0063] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a flowchart of a visual inspection method for repairing broken airborne power lines provided by the present invention;

[0065] Figure 2 This is a schematic diagram of a visual inspection system for repairing broken airborne power lines provided by the present invention.

[0066] The system comprises: 1. Image acquisition unit; 2. Preprocessing unit; 3. Positioning unit; 4. Segmentation unit; 5. Measurement unit; and 6. Judgment unit. Detailed Implementation

[0067] Various embodiments of this disclosure are described more fully below with reference to the accompanying drawings. This disclosure may have various embodiments, and adjustments and changes may be made therein. However, it should be understood that there is no intention to limit the various embodiments of this disclosure to the specific embodiments disclosed herein, but rather this disclosure should be understood to cover all adjustments, equivalents, and / or alternatives falling within the spirit and scope of the various embodiments of this disclosure.

[0068] In the following, the terms “comprising” or “may include”, which may be used in various embodiments of this disclosure, indicate the presence of the disclosed functions, operations, or elements, and do not limit the addition of one or more functions, operations, or elements. Furthermore, as used in various embodiments of this disclosure, the terms “comprising,” “having,” and their cognates are intended only to indicate a particular feature, number, step, operation, element, component, or combination of the foregoing, and should not be construed as primarily excluding the presence of one or more other features, numbers, steps, operations, elements, components, or combinations of the foregoing, or the possibility of adding one or more combinations of the foregoing.

[0069] It should be noted that, in various embodiments of this disclosure, the expression "or" or "at least one of A and / or B" includes any combination or all combinations of the words listed simultaneously. For example, the expression "A or B" or "at least one of A and / or B" may include A, may include B, or may include both A and B.

[0070] It should be noted in advance that, in order to facilitate a clear and accurate description of the technical solutions in the embodiments of this application, the following is a brief explanation of some terms and related technologies involved in the embodiments of this application:

[0071] 1. Object detection: A task in computer vision that typically takes an input image as the object and outputs the category and location range of the object in the image. The location range is usually represented by a bounding box, and a confidence score is given to indicate the reliability of the localization and classification results.

[0072] 2. Instance segmentation: This is a visual task that divides each target instance in an image into pixels. Its output is usually a mask or probability mask corresponding to the target, which can accurately represent the contour boundary of the target and support subsequent morphological analysis and size measurement.

[0073] 3. Scale calibration: This is the process of establishing a mapping relationship between the pixel scale of an image and the actual physical scale. It is usually done by obtaining the conversion coefficient between pixels and length units through a calibration board or known size characteristics. It can also be combined with camera parameters to correct the scale deviation caused by imaging distortion in order to support physical measurement.

[0074] To address the problems of weak stability in morphological judgment and low consistency in quantitative verification of key dimensional relationships caused by the long-term reliance on manual visual inspection and experience in the quality inspection of airborne power line breakage repair, and the fact that inspection conclusions are easily affected by lighting conditions, viewing angle, and changes in line posture, making it difficult to meet the actual needs of rapid, stable, and comparable inspection results under field conditions, this application discloses a visual inspection method and system for airborne power line breakage repair. By introducing a computer vision model to automatically analyze images of the repaired area, a unified identification and quantitative judgment of the stripping and indentation states is formed, thereby improving the stability and consistency of repair quality inspection results, increasing field inspection efficiency, reducing reliance on manual labor, and enhancing the comparability and traceability of inspection conclusions.

[0075] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0076] like Figure 1 As shown in the figure, this embodiment provides a visual inspection method for repairing broken airborne power lines, including:

[0077] Step S1: Acquire images of the repair area and preprocess the images to obtain standardized images. Preprocessing includes distortion correction and brightness normalization.

[0078] Step S2: Input the standardized image into the target detection model to obtain the localization results, which include the localization results of the stripped area and the indentation area;

[0079] Step S3: Determine the segmentation input region based on the localization result, extract the region image corresponding to the segmentation input region from the standardized image, and input the region image into the instance segmentation model to obtain the stripping mask and the indentation mask;

[0080] Step S4: Extract the axial direction of the line body based on the stripping mask and the indentation mask and perform orientation normalization. Obtain the conversion relationship between pixel quantity and physical quantity through scale calibration, determine the stripping length, the axial distance between the stripping end face and the indentation start boundary, the number of indentations and the cracking state, and write them into the process feature quantity.

[0081] Step S5: Compare the process feature quantities with the preset rule library and output the qualified result, the result that needs to be reviewed, or the unqualified result.

[0082] This embodiment standardizes the images of the repaired area to maintain consistent imaging quality under varying lighting conditions, viewing angles, and changes in line posture, providing stable visual input for subsequent judgment. By combining computer vision models to automatically locate and finely characterize the stripping and indentation-related areas, the morphological criteria for stripping and indentation boundaries are transformed from subjective experience into unified algorithmic criteria, reducing the impact of human differences on the judgment results. Quantitative characterization of key process states under unified reference constraints ensures consistent feature representation of inspection elements such as stripping length, the relationship between the stripping end and the indentation position, and surface defect status, improving the comparability and traceability of inspection results from different batches. The introduction of rule-based judgment logic to output qualified, requiring review, and unqualified conclusions provides a clear review path and engineering feasibility, ensuring the reliability of repair quality confirmation while also considering the efficiency of rapid field inspection. Overall, this improves the stability and consistency of the airborne power cable breakage repair quality inspection process, reduces reliance on manual labor, and enhances the continuity and standardization of field operations.

[0083] Hereinafter, steps S1 to S5 will be specifically described according to embodiments of this application.

[0084] In step S1, the core task is to generate standardized image inputs for the airborne power line break repair area, which can be used for subsequent localization and segmentation processing, and to ensure image coverage integrity and quality consistency under multi-pose imaging conditions. The input is the original image sequence obtained from the repair area under different imaging poses, and the output is a standardized image set that satisfies geometric and brightness consistency. When quality or pose coverage does not meet the constraints, traceable verification image identifiers and pose identifiers are generated, ensuring that subsequent processing obtains a stable, comparable, and traceable image foundation under the same data caliber.

[0085] Specifically, the image of the repaired area can be represented as a two-dimensional pixel matrix. ,in, Indicates the image capture pose index. Represents pixel coordinates, Indicates the channel index.

[0086] In some embodiments of this application, a set of imaging postures can be pre-established for the repair area to address the constraint of consistent image coverage. This is used to define the camera's observation direction, distance, and pitch range relative to the line axis, ensuring that the indented and stripped areas obtain fully visible texture and boundary information in at least one orientation. This is used when acquiring images of the repaired area in a set of image-taking orientations. Each image taking posture When dealing with corresponding images, pose labels can be generated for each image. This is used to express the correspondence between the image and a certain pose in the image pose set, and to establish an association storage with the corresponding image so that there is a direct index entry when comparing by pose, verifying cross-pose consistency, or tracing and verifying the source of the image.

[0087] Furthermore, attitude coverage integrity can be determined by preset integrity conditions. Constraints are imposed, among which This represents the set of attitude identifiers obtained for the repaired area. The preset integrity condition can be a combination of "attitude identifier deduplication count reaches a threshold and key attitudes are mandatory," for example, denoting the subset of key attitudes as... And order:

[0088]

[0089] in, This represents the deduplication operator. This represents the minimum pose coverage. If the pose identifier set corresponding to the repaired area does not meet the preset integrity condition, it is necessary to re-acquire images of the repaired area corresponding to each of the image capture poses in the image capture pose set. Alternatively, only the images corresponding to the missing poses can be re-acquired until the preset integrity condition is met. This ensures that the standardized image set has closed coverage in the pose dimension, avoiding subsequent misjudgments and incomparability caused by single-view occlusion, reflection, or incomplete boundaries.

[0090] Based on this, the image needs to be preprocessed, including distortion correction and brightness normalization, to obtain a standardized image.

[0091] Specifically, regarding distortion correction, considering that radial distortion of the lens will change the geometric morphology of the stripping and indentation boundaries when capturing images at close range, in this embodiment, distortion correction can be based on the camera intrinsic parameter matrix. and distortion coefficient vector Construct pixel remapping function , original pixel coordinates Mapping to Corrected Coordinates And obtain the distortion-corrected image. In engineering implementation, this remapping can be generated once by looking up a table and called frame by frame, so that the computational cost of distortion correction is consistent with the frame rate constraint of multi-pose imaging.

[0092] In some embodiments of this application, the metal terminals and indentation areas of the repaired region are prone to saturation and high brightness under directional lighting, leading to texture loss and affecting subsequent boundary extraction. Therefore, after distortion correction, reflection highlight suppression can be further performed to reduce the brightness component. Perform compression mapping to obtain the luminance component after highlight suppression. It can be written as:

[0093]

[0094] in, Indicates the high compressive strength coefficient. This represents the highlight threshold, used to suppress saturation in the highlight area and preserve gradient details around the indentation, so that the contrast between the fracturing texture and the indentation boundary can still be stably characterized after suppression.

[0095] In terms of brightness normalization, to reduce brightness shifts caused by different poses and exposures, brightness normalization can perform linear normalization on the brightness components after highlight suppression to obtain the brightness-normalized brightness components. The following methods can be used:

[0096]

[0097] in, Indicates the gain coefficient. The bias coefficients are obtained by matching the current image brightness statistics with the target brightness statistics, so that the brightness distribution under different imaging postures converges under the same calibration aperture, thereby forming a standardized image. For direct use in subsequent processing.

[0098] Furthermore, sharpness evaluation can be performed after brightness normalization to avoid interference from exposure differences in sharpness measurement. Sharpness evaluation can employ a gradient energy index based on the Laplacian operator. It can be written as:

[0099]

[0100] in, Represents the Laplace operator. Represents the variance operator. Used to characterize the intensity of high-frequency details and reflect the resolvability of indentation boundaries and fine line structures in the current pose. If the sharpness evaluation does not meet the preset sharpness conditions... If so, the standardized image is marked as the verification image, and a verification image identifier is generated. Establish a correlation with the output results so as to retain traceable evidence of image quality anomalies in subsequent outputs and support manual review or re-imaging to close the loop.

[0101] For example, the image capture posture set may include frontal view along the body axis and yaw view along the body axis. Radial pitch along the line Several postures, with preset integrity conditions, can make The key attitude subset includes the frontal and lateral yaw attitudes; highlight threshold. The upper quantile value of the luminance quantile can be used to adapt to different lighting conditions. Adjustable within a range without introducing significant contrast collapse; preset sharpness conditions. It can be obtained by combining camera resolution and working distance calibration to ensure that the texture gradient near the indentation start boundary still has a stable response in the normalized brightness domain.

[0102] Thus, step S1 establishes a closable image set under integrity constraints by performing multi-pose coverage acquisition on the repair area and forming a posture identifier and image association storage. At the same time, it forms a standardized image generation mechanism with geometric and brightness consistency in the distortion correction, reflection highlight suppression and brightness normalization links. When the sharpness evaluation does not meet the preset sharpness conditions, it forms a quality tracking entry point for the verification image identifier, thereby providing a directly usable and consistent image input basis for the subsequent localization result generation and segmentation input area construction.

[0103] In step S2, the core task is to rapidly locate the stripped and indented areas of the airborne power line breakage repair site based on the standardized image set formed in step S1, and simultaneously output the location results and quality tracking information that can be used for subsequent instance segmentation, cropping, and measurement standardization. The input is a single standardized image and its attitude identifier; the output is the location results of the stripped and indented areas, along with a confidence level associated with each location result. When the location quality is insufficient to support subsequent boundary extraction and size calculation, the location results need to be included in the verification process to avoid the transmissive amplification of incorrect cropping and measurement.

[0104] In this embodiment, to ensure stable response of the target detection model to both small-scale texture targets and local structural targets at the repair site, the target detection model can be constructed by cascading a feature extraction backbone network, a feature fusion network, and a detection head. The backbone network maps the standardized image into multi-layer semantic features, the feature fusion network performs top-down and bottom-up information interaction across different scales, and the detection head outputs candidate localization sets on feature maps at each scale. For example, the backbone network can employ a residual structure formed by alternating stacks of convolutional layers, normalization layers, and nonlinear activation layers, allowing shallow layers to retain edge details and deep layers to aggregate texture semantics. The feature fusion network can employ a feature pyramid structure, enabling the high-contrast local morphology of the indentation region and the elongated boundary features of the stripped area to obtain detectable representations at different resolutions. The detection head can employ a parallel structure of classification and regression branches. The classification branch outputs a confidence score indicating whether the target category is a stripped area or an indentation region, while the regression branch outputs the center coordinates and width / height parameters of the corresponding bounding box, ensuring that the localization result simultaneously possesses category attribution and geometric boundary.

[0105] Specifically, the input tensor of the target detection model maintains a one-to-one correspondence with the standardized image. The standardized image can be scaled proportionally and padded as necessary according to the model input size to maintain geometric proportion consistency. The candidate localization set output by the model can be split into a stripping candidate set and an indentation candidate set according to categories for subsequent overlap suppression and confidence label generation.

[0106] In some embodiments of this application, to perform multi-scale inference on a standardized image to obtain a candidate localization set, the multi-scale feature layer output by the feature fusion network can be denoted as... ,in Indicates the scale layer index. Indicates the scale level; at each scale level, the detector head outputs candidate box parameters. With confidence score ,in This represents the candidate index at this scale level. These represent the stripping line category and the indentation category, respectively. For example, the candidate location set can be written as:

[0107]

[0108] in, Representing scale layer The number of candidates on Used to unify the candidate localization results of the same category obtained from multi-scale reasoning.

[0109] After the candidate localization set is formed, to avoid multiple overlapping boxes for the same target at adjacent scales or locations, which could affect the stability of subsequent clipping, overlap suppression needs to be performed on the candidate localization set. Overlap suppression can be based on the intersection-union ratio (IU), denoted as IU function. When the overlap between two candidate boxes exceeds a threshold, the one with the higher confidence score is retained; thus, the set of results for locating the stripped area can be obtained. Set of indentation area location results For example, the intersection-union ratio can be written as:

[0110]

[0111] in, Represents the area operator of a region. and These represent intersecting regions and merging regions, respectively.

[0112] Furthermore, to explicitly bind the positioning quality to the output and form a traceable link, confidence labels need to be generated for the positioning results of the stripped area and the indentation area, respectively. The confidence label can be composed of the confidence score corresponding to the positioning result and the pose label, so that it can reflect both the confidence level of the target detection model in the current positioning and trace which imaging pose in the imaging pose set the positioning originated from. For example, the generation of the confidence label can adopt the method of "score quantization + pose stitching", where score quantization can... The mapping is performed as discrete levels or interval labels to enable filtering, sorting, and verification aggregation on the output side. When any confidence indicator does not meet the preset confidence conditions, the corresponding positioning result needs to be marked as a verification positioning result and associated with the confidence indicator in the output result. This allows subsequent processes to identify that the positioning belongs to low-confidence positioning and trigger verification or re-acquisition strategies when performing instance segmentation and pruning, thereby avoiding pruning offset and boundary error propagation caused by low-quality positioning.

[0113] For example, the target detection model can use a three-scale output to cover the long boundary of the striped area and the local texture of the indentation area. The model input resolution can be kept at a fixed ratio with the normalized image in step S1 to reduce scale drift. The overlap suppression threshold can be matched with the typical projection size of the target in the repair area to reduce duplicate boxes. The pre-set confidence conditions can be obtained by statistical analysis of the validation set to keep the confidence score distribution and the review trigger ratio stable under different imaging postures, thereby forming a controllable balance between accuracy and review cost.

[0114] Thus far, step S2 has constructed a localization result generation link based on standardized image input, consisting of multi-scale inference, candidate localization set organization, and overlap suppression. It binds the localization quality and pose source to the output result through confidence labeling. When the confidence does not meet the preset conditions, a tracking entry point for verifying the localization result is formed, thereby providing stable spatial constraints and quality boundary conditions for subsequent cropping, instance segmentation, and size measurement based on the localization result.

[0115] In step S3, the core task is to construct a segmentation input region that matches the local structure of the repair area based on the location results of the stripping area and the indentation area output in step S2, and to form a region image input with purer texture and weaker background interference within this region, thereby driving the instance segmentation model to stably generate the stripping mask and the indentation mask. At the same time, the consistency between the segmentation result and the location result is solidified into a traceable process feature quantity in the form of a coverage consistency index, so as to constrain the input boundary and confidence level of subsequent size measurement and defect discrimination.

[0116] In this embodiment, the bounding boxes of the stripped area and the indentation area positioning results can be used as the geometric constraint entry points to determine the segmentation input region. To maintain the relative positional relationship between the stripped area and the indentation area within the local field of view, the segmentation input region can be constructed from the bounding rectangle of the union of the two types of positioning boxes, and associated one-to-one with the pose identifier to avoid mask scale inconsistencies caused by clipping aperture drift across poses. This segmentation input region can be represented as a rectangular window. Its coordinates are from the top left corner. and the coordinates of the bottom right corner It is confirmed that the coordinates are all taken in the pixel coordinate system of the standardized image, and use the same scale reference as the positioning box coordinates output in step S2.

[0117] After the segmented input region is formed, to ensure that the end boundaries of the stripping lines, the starting boundaries of the indentations, and the crack texture retain their complete context after cropping, boundary expansion needs to be performed on the segmented input region. Boundary expansion is implemented at the pixel level or proportionally to the region size, with the expansion amount applied to the four boundaries (left, right, top, and bottom), and image boundary cropping is performed for protection. The expanded segmented input region after boundary expansion is denoted as... ,in, and These represent the horizontal and vertical expansion amounts, respectively, used to preserve the transition areas on both sides of the indentation and the fine texture of the stripped surface without introducing too much background, thus avoiding truncated missegmentation of instances in the boundary neighborhood.

[0118] Furthermore, after determining the extended segmentation input region, a window can be cropped from the standardized image to extract the region image corresponding to the segmentation input region. The cropped region image is denoted as... Simultaneously, to reduce the interference of localized high reflectivity, background texture, and tooling edges on the mask boundary at the repaired area, background suppression needs to be performed within the expanded segmentation input region to create a more focused input. Background suppression can be performed jointly in the luminance and color domains: first, an adaptive threshold is constructed using the luminance component of the region image to separate excessively dark backgrounds; then, local gradient magnitude suppression is combined to suppress large flat areas, making the retained area more concentrated near the wire body, stripped surface, and terminal indentation. The region image obtained after background suppression is denoted as... And maintain consistency with the input instance segmentation model before input instance segmentation. The window coordinate mapping relationship is used to subsequently backproject the mask back into the normalized image coordinate system.

[0119] It's important to note that the instance segmentation model inputting a region image needs to maintain stable pixel-level classification capabilities even when both thin, elongated boundaries from peeling lines and local textures from indentations coexist. To this end, the instance segmentation model can consist of an encoder, a decoder, and a mask prediction head. The encoder obtains multi-scale semantic features by stacking multiple levels of convolutional downsampling layers. The decoder restores spatial resolution by progressively upsampling and fusing features from the encoder at the same scale. The mask prediction head outputs a pixel-level class probability map on the fused high-resolution feature map. The model input is a region image after background suppression. The model outputs two probability maps of the same size as the region image, corresponding to the stripping line category and the indentation category, respectively. Simultaneously, to ensure consistency with subsequent geometric measurements, the probability maps are thresholded to obtain a mask, while maintaining the pixel coordinates of the mask. The window coordinates are consistent.

[0120] In some embodiments of this application, the background-suppressed region image and the pose label together constitute the inference input of the instance segmentation model, making the region images under different imaging poses more consistent in terms of background interference intensity, thereby reducing the random swaying of the mask boundary.

[0121] Furthermore, after the instance segmentation model completes inference, the output stripping mask is denoted as... The indentation mask is recorded as And press it in the window The coordinates are mapped back to the standardized image coordinate system to form a mask representation that can be aligned with the positioning results; a closed inference chain is completed here, and the mask becomes the pixel-level basis for subsequent boundary extraction, axis vector measurement and defect discrimination.

[0122] After mask generation, to suppress the cumulative impact of positioning errors or segmentation drift on subsequent cutting and measurement, it is necessary to calculate the coverage consistency index between the mask and the corresponding positioning result, and write it into the process feature quantity for subsequent use. Specifically, the coverage consistency index can be expressed as the intersection-union ratio (IUU) of the mask area and the corresponding positioning box area, calculated separately for stripping and indentation. Let the area corresponding to the stripping positioning box be... The area corresponding to the indentation positioning frame is Then the coverage consistency index can be written as:

[0123]

[0124] in, and These are used to characterize the consistency of the mask boundaries and positioning boundaries in spatial coverage for stripping and indentation, respectively. The coverage consistency index, as part of the process characteristic quantity, can be stored in association with the attitude identifier and confidence identifier, enabling subsequent calculations of stripping length, distance between the stripping end and the indentation location, number of scratches, and fracturing determination to filter inputs or trigger a review process under the same quality caliber.

[0125] For example, boundary expansion and The positioning frame width and height can be set in a fixed ratio, allowing the extended field of view to adapt to different wire diameters and working distances; background suppression can suppress the influence of tooling background and highly reflective areas while maintaining the fine texture of the stripped surface and the gradient of the indentation edge; the coverage consistency index can be used for cross-pose comparison, when the same repaired area is under different imaging poses. or When the difference is too large, the mask result corresponding to the pose can be included in the verification link to reduce the interference of abnormal samples on subsequent measurements.

[0126] Thus far, step S3 completes the construction and boundary expansion of the segmentation input region under the constraints of the localization result. Within the expanded segmentation input region, background suppression is used to form a region image with the same input aperture as the instance segmentation model, thereby generating a stripping mask and an indentation mask. At the same time, the coverage consistency index between the mask and the localization result is written into the process feature quantity to form quality boundary conditions, thereby providing a stable, traceable and cross-pose consistent pixel-level input foundation for subsequent boundary extraction and size measurement based on the mask.

[0127] In step S4, the core task is to establish a unified geometric measurement caliber based on the stripping mask and indentation mask formed in step S3. This allows the stripping length, the axial distance between the stripping end face and the indentation starting boundary, the number of indentations, and the fracturing state to be stably obtained in the same axial coordinate system of the line body. Furthermore, pixel quantities are converted into physical quantities through dimensional calibration to meet the constraints of dimensional accuracy in process judgment. The inputs are the stripping mask, the indentation mask, and their pixel representations in the standardized image coordinate system. The outputs are the dimensional results and defect results written with process feature quantities. Simultaneously, the outputs are the line body axial and rotational normalization parameters and dimensional calibration parameters strongly correlated with the measurement, enabling subsequent judgment and verification to trace the measurement caliber and error sources.

[0128] In this embodiment, the line axis can be extracted and its orientation normalized based on the stripping mask and indentation mask. Specifically, it is necessary to construct a centerline representation of the line in the image plane to eliminate the influence of image capture orientation differences on the axial distance.

[0129] In some embodiments of this application, to avoid centerline jitter caused by noise at the boundary of a single mask, the union region of the stripping mask and the indentation mask can be denoted as... and in Extract the centerline of the line body within the coverage area. Centerline extraction can be achieved by... Refine to obtain the skeleton pixel set And then The axial direction vector of the line body is obtained by performing least-squares line fitting, denoted as . It can be determined by the direction of the fitted line and normalized to ensure the scale consistency of subsequent projection calculations.

[0130] After obtaining the line axis, coordinate rotation normalization is performed on the standardized image based on the line axis, aligning the line axis with the rotated and normalized horizontal axis. This transforms the "axial spacing" from a two-dimensional Euclidean distance into a one-dimensional projection difference and reduces the systematic errors introduced by attitude changes. Specifically, let the rotation angle... And construct a two-dimensional rotation matrix Coordinate rotation can be written as:

[0131]

[0132] in, Represents the normalized image pixel coordinates. This represents the pixel coordinates after rotation normalization. This represents the coordinates of the rotation center point, which can be represented by a union mask. The geometric center is determined to reduce the risk of boundary trimming after rotation. After rotation normalization, the line axis is aligned with... Axial alignment makes "axial projection" numerically equivalent to the alignment of axes. One-dimensional statistics of coordinates provide a stable caliber for end boundary extraction.

[0133] Furthermore, in the rotation-normalized coordinate system, the stripping end face and the indentation starting boundary need to be transformed into measurable axial positions. To this end, the axial projected end boundaries can be calculated on the pixel sets of the stripping mask and the indentation mask respectively within the rotation-normalized coordinate system. Let the pixel set of the stripping mask in the rotation-normalized coordinate system be... The set of indentation mask pixels is Based on this, the position of the stripping end face can be determined by the maximum end of the axial projection of the stripping mask, and the position of the indentation starting boundary can be determined by the minimum end of the axial projection of the indentation mask, that is:

[0134]

[0135] in, This indicates the pixel position of the stripped end face in the axial direction. This indicates the pixel position of the indentation start boundary in the axial direction. This ensures that the diameter of the stripping end face and the indentation start boundary does not change with orientation.

[0136] Once the end position is determined, the axial spacing can be obtained based on the axial projection of the stripping end face and the indentation starting boundary:

[0137]

[0138] in, This represents the number of pixels for axial spacing. A positive value indicates that the indentation starting boundary is located axially behind the stripping end face, while a negative value indicates that the two overlap or the indentation intrudes into the stripping area. This information can be used as input for subsequent verification and anomaly detection. Consistent with the axial spacing calculation method, the stripping length can also be obtained in a rotated normalized coordinate system through the range of the axial projections of the stripping mask. Let:

[0139]

[0140] in, The number of pixels representing the stripping length ensures that the length and spacing are calculated based on the same axial coordinate system and have consistent error propagation characteristics.

[0141] It should be noted that, to obtain the physical correspondence of the above measurement results, it is necessary to obtain the conversion relationship between pixel quantity and physical quantity through scale calibration, thereby uniformly converting pixel quantity into physical quantities such as millimeters. Scale calibration can be based on a calibration plate or known dimensional features to form a scaling factor. It is stored in association with the rotation normalization parameter, and the conversion relationship can be written as:

[0142]

[0143]

[0144] in, A physical quantity representing the length of wire stripped. A physical quantity representing axial spacing, a proportionality factor. It remains stable under the same imaging and assembly conditions and is used to ensure measurement consistency across postures.

[0145] In some embodiments of this application, after completing the axial vector measurement, the number of scratches can be calculated, and the direction consistency screening can be achieved by limiting the stripping area and combining it with the axial direction of the line.

[0146] Specifically, a scratch detection sub-region can be constructed within the area defined by the stripping mask to improve the detection rate and reduce background interference. This scratch detection sub-region can be composed of a strip-shaped region between the inward-shrinking region obtained by morphological erosion of the stripping mask and the outward-expanding region obtained by morphological dilation of the stripping mask. This covers the texture bands where scratches may appear on the stripped surface and avoids the jagged noise at the mask boundaries, thus ensuring that subsequent fine-line enhancement only affects the surface area related to the scratches.

[0147] It should be noted that scratches appear as elongated, high-contrast structures in the image. Therefore, to enhance this type of structure and form a selectable candidate set, fine-line structure enhancement based on second-order derivative and directional filtering can be performed within the scratch detection sub-region to obtain a fine-line response map. The thin-line enhancement can be achieved by convolving the brightness components with a multi-directional filter bank and taking the maximum response, so that thin lines in different directions can be enhanced; subsequently, threshold segmentation is performed on the thin-line response map to obtain a candidate set of thin lines. This provides an input set for subsequent direction consistency screening.

[0148] After the candidate set of fine lines is formed, the key distinction between nicks and noisy fine lines lies in the consistency of their direction with the axis of the line body. Therefore, a candidate set of nicks can be obtained by performing direction consistency filtering on the candidate set of fine lines based on the axis of the line body; for example, the principal direction vector can be calculated for each connected component of the fine line. and the axial direction vector of the line body Calculate the included angle. Where, the included angle... It can be calculated from the inner product and filtered using a threshold constraint:

[0149]

[0150] in, The directional consistency threshold is used to eliminate random textures and reflective edges that have excessive angles with the line axis, thereby converging the candidate set into a score candidate set. .

[0151] Furthermore, after direction filtering, the candidate set of notches can be further refined. Connectivity analysis and length consistency screening are performed to suppress fragmented lines and occasional noise, thereby determining the number of notches. Specifically, the skeleton length of each candidate notch connected component is calculated and matched with a preset length interval. Then, connected components that meet the length interval and do not overlap or merge with adjacent candidates are counted as the number of notches. After the counting is completed, the number of scratches and the aforementioned axial spacing, stripping length, and other measurement results can be written into the process feature quantity in a unified recording format, so that subsequent process judgments can simultaneously access the dimensional quantity and the defect quantity and perform correlation analysis within the same data structure.

[0152] Furthermore, it can also be applied to indentation masks. Within a defined indentation area, the fracturing state is determined and recorded as a process characteristic, thus incorporating structural defects in the indentation area into the same judgment criterion. Specifically, in a rotated normalized coordinate system... Lowering the normalized luminance component ,exist Internal calculation of crack response and generation of fracturing strength index It can be written as:

[0153]

[0154]

[0155] in, Represents the gradient operator, Indicates the gradient magnitude. Indicates an indicator function, Indicates the crack response threshold. Used to highlight the fine, high-contrast structure within the indentation area. Used to characterize the cumulative intensity of the overthreshold crack response within the indentation region. Based on Determine the fracturing state with a preset threshold range And write the process feature values, and at the same time... with attitude markings Associated storage enables subsequent cross-posture consistency checks and rule base matching to directly reference the fracturing state.

[0156] For example, the scaling factor of scale calibration The pixel length in the image can be estimated from a known length reference and kept consistent with the image capture pose set; boundary expansion and rotation normalization parameters can be stored in association with pose identifiers to ensure cross-pose traceability; orientation consistency threshold. It can suppress axially irrelevant fine lines while ensuring the preservation of the scratches. The length consistency screening range can be matched with the common scratch scales in wire stripping processes, thereby maintaining the stability of scratch count under different wire diameters and different surface materials.

[0157] At this point, step S4 extracts the center line of the line body and determines the axis of the line body by using the union range of the stripping mask and the indentation mask. The axial vector measurement is converted into the projection difference in the rotation-normalized coordinate system by coordinate rotation normalization. The conversion relationship between pixel quantity and physical quantity is established by combining scale calibration, thereby obtaining the stripping length and the axial distance from the stripping end face to the indentation starting boundary. At the same time, within the stripping area, the number of scratches is obtained by strengthening the fine line structure, filtering the direction consistency and the connectivity and length consistency. The fracturing state is further determined and written into the process feature quantity, thus providing a unified, stable and traceable measurement input basis for subsequent threshold judgment and anomaly review based on the process feature quantity.

[0158] In step S5, the core task is to place the process feature quantities written in step S4 under a unified judgment caliber to complete the conclusion output. This allows the process feature quantities corresponding to multi-pose images to form a deterministic qualified or unqualified conclusion based on a preset rule base when the consistency check passes, and to form a conclusion requiring review and a traceable inconsistency type identifier and hit rule identifier when the consistency check fails. The input consists of a set of process feature quantities associated with the pose identifier and a preset rule base. The output is a qualified result, a result requiring review, or an unqualified result, along with identifier information closely linked to the conclusion to support subsequent review and traceability.

[0159] In this embodiment, process feature quantities can be compared with a preset rule base to output qualified, required-for-review, or unqualified results. In this process, the comparability issue between multi-pose results can be addressed first, followed by the deterministic issue of rule matching. Specifically, process feature quantities can be denoted as... ,in This is the image capture pose index corresponding to the pose identifier. As a dimension of process features, the process feature components can include stripping length, axial distance between the stripping end face and the indentation starting boundary, number of indentations, fracturing state, and quality indicators related to segmentation coverage consistency. Simultaneously, to suppress outliers caused by reflections, occlusions, or incomplete boundaries in individual poses under cross-pose conditions, consistency checks need to be performed on the process features obtained from different images using pose identifiers as indexes, and these consistency checks are used as gating conditions before entering the rule base for matching.

[0160] In some embodiments of this application, consistency checks can be performed on process feature quantities obtained from different images according to their orientation. The consistency check can focus on whether the critical dimensions and critical defect quantities of the same repair site under different orientations are within the allowable deviation range. For example, let the set of indices for the critical process feature quantum set be... And let the reference posture be The reference pose can be obtained by jointly selecting confidence indicators and coverage consistency indicators, making the process feature quantities corresponding to the reference pose more stable in terms of positioning and segmentation quality. For any key component... It can calculate the cross-attitude deviation. ,in, For a set of attitude identifiers, This is used to characterize the maximum deviation of this component under cross-attitude conditions. Furthermore, the consistency verification criterion, i.e., the preset consistency condition, can be defined as a joint constraint on the deviations of each key component, namely:

[0161]

[0162] in, Representing components The normalization scale is used to normalize process characteristic quantities of different dimensions to the same comparison caliber. This indicates a preset consistency threshold. This criterion allows phenomena such as "excessive drift in size and attitude," "excessive fluctuation in the number of scratches," and "inconsistent fracturing conditions" to be included in the consistency check in a unified manner.

[0163] When a consistency check fails to meet the preset consistency conditions, a traceable cause needs to be generated for review and location. At this point, the review result should be output, and the inconsistency type identifier should be recorded. The inconsistency type identifier can be generated based on the triggering component set, where the triggering set is... Inconsistency type identifier can be determined by The mapping allows the output results to clearly characterize the sources of inconsistencies, such as axial dimension drift, low mask coverage consistency, scratch count fluctuations, or fracturing state conflicts. This enables targeted re-image acquisition or manual verification strategies to be adopted according to the type during the review process.

[0164] Once the consistency check meets the preset consistency conditions, the system proceeds to the rule base matching stage. It is necessary to aggregate cross-attitude process features into a single judgment input to avoid multiple versions of process feature definitions for the same repair area during rule matching. Furthermore, the aggregation method can be based on a reference attitude. The system prioritizes robust fusion of other parameters, such as median fusion for dimensional parameters, mode fusion for discrete counts, and minimum constraint for quality indicators, to ensure that the converged process features are robust. It maintains representativeness while being constrained by unfavorable postures.

[0165] Furthermore, the pre-defined rule base needs to have a hierarchical comparison order to ensure the determinism of the output conclusion. For example, the rule base can be represented as a set of rules. And assign hierarchical priority to each rule. The hierarchical comparison order is as follows The rules are executed in descending order of priority, ensuring that "hard rejection rules" take precedence over "acceptance confirmation rules," and allowing "review-triggered rules" to still be triggered based on quality indicators or abnormal combinations of conditions after consistency verification has passed. Based on this, process feature quantities can be matched one by one according to the hierarchical comparison order of the preset rule base, outputting acceptable or unacceptable results. Furthermore, the rules can be... Defined as to Logical decision function ,when The time indicates that a rule has been hit; matching is performed one rule at a time according to the hierarchical comparison order, and the matching stops at the first rule that has been hit, in order to avoid conflicting conclusions caused by multiple rules being hit at the same time.

[0166] After a rule is matched, the rule matching identifier needs to be recorded, and the matching information needs to be embedded in the output to support traceability. The rule matching identifier can be a rule number or a rule code string, and it is associated with the qualified or unqualified output results for storage, enabling direct identification of the specific rule conditions that triggered the judgment and the corresponding threshold. If no rule is matched after each rule is matched, this situation can be handled according to the fallback strategy defined in the rule base as a result requiring review or an unqualified result, and the fallback rule identifier should also be recorded to maintain output closure.

[0167] For example, a preset consistency condition threshold is used. The settings can be combined with camera resolution, scale calibration error and mask boundary jitter statistics to ensure that the dimensional fluctuation under different imaging postures does not exceed the allowable deviation of the process; the rule base hierarchy can place rules such as fracturing state rejection, axial spacing exceeding the limit, and stripping length exceeding the limit at the high priority level, and place rules such as abnormal number of scratches and low quality indicators at the medium priority level, so that the output results maintain the interpretability and consistency of the judgment logic while meeting key safety and quality constraints.

[0168] Thus, step S5 establishes a unified standard for cross-attitude results through consistency verification of process feature quantities of attitude identifier index, and generates a result to be reviewed and an inconsistency type identifier when the consistency does not meet the preset consistency conditions. When the consistency meets the preset conditions, it completes the matching and output of qualified or unqualified results according to the hierarchical comparison order of the preset rule base and records the hit rule identifier, thereby forming a rule judgment system with traceable conclusion output, closed judgment link and cross-attitude robustness, providing a deterministic basis for the reliable output of the final process conclusion.

[0169] In summary, this method, through multi-pose overlay imaging and image geometry-lighting standardization processing, combined with a collaborative localization and segmentation mechanism of target detection and instance segmentation, achieves quantifiable extraction of stripping length, axial spacing, number of scratches, and cracking state under a unified line axial coordinate system. Furthermore, it generates traceable quality conclusions through cross-pose consistency verification and hierarchical rule base judgment. This significantly reduces false positives and false negatives, as well as subjective interpretation differences, under complex lighting and reflective interference conditions in the field. It also reduces reliance on manual inspection and experience thresholds, improving detection efficiency, dimensional measurement stability, and judgment consistency, thereby enhancing the reliability and controllability of airborne power line breakage repair quality.

[0170] It should be noted that, although the embodiments in this application are based on... Figure 1 Steps S1 to S5 are described sequentially, but this does not mean that steps S1 to S5 must be performed in a strict order. The reason this embodiment follows this order is... Figure 1 The order in which steps S1 to S5 are described is provided to facilitate understanding of the technical solutions of the embodiments of this application by those skilled in the art. In other words, in the embodiments of this application, the order of steps S1 to S5 can be appropriately adjusted according to actual needs.

[0171] In some embodiments of this application, the "Visual Inspection Method for Airborne Wire Breakage Repair" is applied to the field quality assessment of aircraft wire breakage repair sites. The inspection targets cover the stripped section and the terminal indentation area. The output results are used to provide a qualified conclusion on the repair quality and support verification and traceability. This embodiment focuses on the collaborative processing of multi-angle imaging, target detection, and instance segmentation, and quantifies and judges process characteristics such as stripping length, axial distance between the stripped end face and the indentation starting boundary, number of stripping marks, and indentation cracking state.

[0172] A complete implementation process may include the following steps:

[0173] Step 1: The camera establishes an image-taking pose set for the area to be repaired. For each image capture pose Acquire raw images and generate pose labels To form the original image sequence To ensure complete image capture of the indentation operation area, an integrity check is performed on the attitude identifier set. If the attitude identifier set does not meet the preset integrity conditions, the image corresponding to the missing attitude is re-acquired until the integrity conditions are met, so that subsequent cross-attitude consistency checks have closed-loop input.

[0174] Step two, for each original image Perform preprocessing to obtain a standardized image Preprocessing first involves performing distortion correction on the image based on the camera calibration parameters to obtain a distortion-corrected image. After distortion correction, reflection highlight suppression was applied to the luminance component to preserve indentation boundaries and crack details. Subsequently, luminance normalization was performed to obtain a luminance-normalized image, and sharpness evaluation was completed. Sharpness evaluation is indicated by... ,when If the preset clarity condition is not met, a verification image identifier is generated. and with attitude identifier Associative storage enables the tracking of the source of image quality anomalies at that pose in the output results.

[0175] Step 3, standardize the image Inputting the target detection model yields a set of localization results. The target detection model consists of a feature extraction backbone network, a feature fusion network, and a detection head connected together. The detection head outputs a set of candidate localization results. Candidate locations for the stripping line category and the indentation category are generated under multi-scale inference. After performing overlap suppression on the candidate location set, the stripping line region localization result is obtained. Results of indentation area localization And generate confidence level labels respectively. and When any confidence level indicator does not meet the preset confidence level conditions, the corresponding positioning result is marked as a verification positioning result and stored in association with the confidence level indicator, so that low-confidence inputs can be gated and traced in the subsequent clipping and measurement stages.

[0176] Step four: Construct a segmented input region based on the localization results. and The segmented input region is obtained by taking the bounding rectangle of the union of the pixels in the pixel coordinate system. and to Perform boundary expansion to obtain an expanded segmented input region. It can be written as:

[0177]

[0178] in, and This is the boundary expansion amount. Cropping from the normalized image. Corresponding region image and in Intra-process background suppression yields region images .Will The input is an instance segmentation model, which consists of an encoder, a decoder, and a mask prediction head connected together. The output is a stripped mask. With indentation mask The masking results are then mapped back to the normalized image coordinate system to maintain geometric alignment with the positioning results.

[0179] Step 5: Calculate the coverage consistency index for both the stripping mask and the indentation mask, and write it into the process feature values. The coverage consistency index is denoted as follows: and The intersection-over-union ratio (IoU) of the mask region and the corresponding bounding box region is used:

[0180]

[0181] in, This is an area operator. The coverage consistency index, along with the attitude identifier and confidence index, are jointly written into the process feature quantity. This allows process characteristics to simultaneously carry a unified standard of "measurement results + quality boundaries".

[0182] Step six: Extract the axis of the line and perform orientation normalization to complete dimensional measurement and conversion. and Extract the centerline of the line body within the union range and determine the axial direction vector of the line body. Based on this, coordinate rotation normalization is performed on the standardized image. The rotation angle is denoted as . The rotation normalization relation can be written as:

[0183]

[0184] in, These are the coordinates of the rotation center point. After rotation normalization, the position of the stripping end face is determined based on the end boundary of the stripping mask projected along the axial direction. The starting boundary position of the indentation is determined based on the end boundary of the indentation mask projected axially. The axial spacing in pixels is obtained based on the interval between the two projections along the axial direction. Simultaneously, the stripping length in pixels is obtained from the axial projection range of the stripping mask within the rotated normalized coordinate system. The conversion relationship between pixel count and physical quantity is obtained through scale calibration, and the scaling factor is denoted as... Convert the number of pixels into physical quantities:

[0185]

[0186]

[0187] And and Write process feature quantity .

[0188] Step 7: Determine the number of scratches and write them into the process feature values. Construct a scratch detection sub-region within the area defined by the stripping mask, and perform fine line structure enhancement within the scratch detection sub-region to obtain a candidate set of fine lines. Based on the axial direction of the line, a directional consistency screening process is performed on the candidate set of thin lines to obtain the candidate set of indentations. Then, connectivity analysis and length consistency screening are performed on the candidate set of notches to obtain the number of notches. and will Write process feature quantity In this embodiment, the direction consistency screening is constrained by the threshold of the angle between the main direction of the scratch and the axis of the line, and the length consistency screening is constrained by the length range of the scratch skeleton, so that the number of scratches can still be stably output even when reflective edges, background fine lines and random noise are present.

[0189] Step 8: Determine the fracturing state and write it into the process feature quantity. Based on the indentation area defined by the indentation mask, fracturing discrimination is performed on the fine line texture and boundary discontinuities within the indentation area, and the fracturing state is recorded as follows: And write the process feature quantity In this embodiment, fracturing detection and attitude identification are performed. Coverage consistency indicators When used together, when the coverage consistency of the indentation area is low or the clarity evaluation does not meet the preset clarity conditions, the fracturing status and the corresponding markings together form a verification evidence link.

[0190] Step nine: Compare the process feature quantities with the preset rule base and output the judgment result. Process feature quantities obtained from different images according to posture identifiers. A consistency check is performed. If the consistency check fails to meet the preset consistency conditions, a result requiring review is output, and an inconsistency type identifier is recorded. If the consistency check meets the preset conditions, the process feature quantities are matched one by one according to the hierarchical comparison order of the preset rule base, and a qualified or unqualified result is output, with the matching rule identifier recorded. In this embodiment, the preset rule base includes the existence of stripping lines, the existence of indentations, the stripping length range constraint, the lower limit constraint of axial spacing, the fracturing rejection constraint, and the review trigger constraint related to quality indicators, so that the output result has both deterministic conclusions and traceable evidence.

[0191] Through the complete implementation process described above, this method organically combines multi-pose coverage imaging, positioning and segmentation collaborative processing, axial normalization measurement, and rule-based judgment closed-loop, enabling key dimensions such as stripping length and axial spacing to be stably output under a unified geometric caliber. It also solidifies quality information such as coverage consistency, clarity, and confidence into a traceable evidence link, which can significantly reduce the risk of misjudgment and omission under complex lighting and reflective interference conditions in the field, reduce reliance on human visual experience, improve detection efficiency, measurement consistency, and judgment interpretability, thereby improving the reliability and controllability of airborne power line breakage repair quality and reducing flight safety hazards.

[0192] It should be understood that the step numbers identified by "Step 1, Step 2" and other similar forms in the above embodiments are only used to distinguish different steps and do not limit the steps to be executed in the order of these numbers. The specific execution order of each step can be adjusted according to its functional requirements and the inherent logic in the actual application scenario. The above step numbers should not be interpreted as a limitation on the implementation process of the embodiments of this application.

[0193] like Figure 2 As shown, the following is an embodiment of a visual inspection system for airborne wire breakage repair provided by this disclosure. This airborne wire breakage repair visual inspection system and the airborne wire breakage repair visual inspection methods of the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the airborne wire breakage repair visual inspection system, please refer to the embodiments of the above airborne wire breakage repair visual inspection methods.

[0194] Based on the same concept, another embodiment of this application provides a visual inspection system for airborne power line breakage repair, comprising:

[0195] Imaging unit 1 is used to acquire images of the repaired area;

[0196] Preprocessing unit 2 is used to perform preprocessing on the image to obtain a standardized image. The preprocessing includes distortion correction and brightness normalization.

[0197] The positioning unit 3 is used to input the standardized image into the target detection model to obtain the positioning results, which include the positioning results of the stripped area and the positioning results of the indentation area.

[0198] Segmentation unit 4 is used to determine the segmentation input region based on the localization result, extract the region image corresponding to the segmentation input region from the standardized image, and input the region image into the instance segmentation model to obtain the stripping mask and the indentation mask;

[0199] Measurement unit 5 is used to extract the axial direction of the line body based on the stripping mask and the indentation mask and perform attitude normalization. It obtains the conversion relationship between pixel quantity and physical quantity through scale calibration, determines the stripping length, the axial distance between the stripping end face and the indentation start boundary, the number of scratches and the cracking state, and writes them into the process feature quantity.

[0200] Judgment unit 6 is used to compare process feature quantities with preset rule base and output qualified results, results requiring review, or unqualified results.

[0201] In some embodiments of this application, the segmentation unit 4 includes a region generation subunit, a boundary expansion subunit, a background suppression subunit, and a mask output subunit; the determination unit 6 includes a consistency calculation subunit and an index writing subunit, wherein...

[0202] The region generation subunit is used to generate a segmented input region based on the location results of the stripping region and the indentation region.

[0203] The boundary expansion subunit is used to perform boundary expansion on the segmented input region to obtain an expanded segmented input region.

[0204] The background suppression subunit is used to perform background suppression within the extended segmented input region to obtain a region image;

[0205] The mask output subunit is used to input the region image into the instance segmentation model to obtain the stripping mask and the indentation mask;

[0206] The consistency calculation subunit is used to calculate the coverage consistency index of the stripping mask and the stripping area positioning result, as well as the indentation mask and the indentation area positioning result, respectively.

[0207] The indicator writing sub-unit is used to write the coverage consistency indicator into the process feature quantity.

[0208] In summary, this system, through the collaborative efforts of image acquisition unit 1, preprocessing unit 2, positioning unit 3, segmentation unit 4, measurement unit 5, and judgment unit 6, achieves stable acquisition of images of the repaired area, standardized processing under unified imaging conditions, and automatic identification and quantitative characterization of key process states such as stripping and indentation. This enables a holistic shift in the quality inspection of airborne power line breakage repair from experience-based judgment to rule-based quantitative judgment. It can maintain the consistency and comparability of morphological criteria and dimensional measurement results under conditions of lighting changes, viewing angle differences, and line posture fluctuations, providing a stable and traceable quantitative basis for the rapid confirmation and verification of repair quality under field conditions.

[0209] The above-disclosed embodiments are merely preferred embodiments of the present invention, but the present invention is not limited thereto. Any non-creative variations that can be conceived by those skilled in the art, as well as any improvements and modifications made without departing from the principles of the present invention, should fall within the protection scope of the present invention.

Claims

1. A visual inspection method for repairing broken airborne power lines, characterized in that, include: Images of the repaired area are acquired and preprocessed to obtain standardized images. The preprocessing includes distortion correction and brightness normalization. The standardized image is input into the target detection model to obtain the localization results, which include the localization results of the stripped area and the indentation area. Based on the localization results, the segmentation input region is determined. The corresponding region image is extracted from the standardized image and input into the instance segmentation model to obtain the stripping mask and the indentation mask. Based on the stripping mask and indentation mask, the axial direction of the line is extracted and the orientation is normalized. The conversion relationship between the number of pixels and physical quantities is obtained through scale calibration. The stripping length, the axial distance between the stripping end face and the indentation start boundary, the number of indentations and the cracking state are determined and written into the process feature quantity. The process feature quantities are compared with the preset rule library and the results are output as qualified, required to be reviewed, or unqualified.

2. The visual inspection method for repairing broken airborne power lines as described in claim 1, characterized in that, The steps for acquiring images of the repaired area include: A set of image capture poses is established for the repaired area. Images of the repaired area corresponding to each image capture pose in the set are collected and pose labels are generated. The posture identifiers are associated with and stored with the corresponding images. When the set of posture identifiers corresponding to the repaired area does not meet the preset integrity conditions, the images corresponding to each imaging posture in the imaging posture set of the repaired area are re-acquired until the set of posture identifiers corresponding to the repaired area meets the preset integrity conditions.

3. The visual inspection method for airborne power line breakage repair as described in claim 1, characterized in that, Preprocessing also includes: performing reflection highlight suppression after distortion correction, performing sharpness evaluation after brightness normalization, and when the sharpness evaluation does not meet the preset sharpness conditions, marking the corresponding standardized image as a verification image and associating the verification image identifier in the output results.

4. The visual inspection method for repairing broken airborne power lines as described in claim 1, characterized in that, The steps for inputting a standardized image into a target detection model to obtain the localization result include: Multi-scale inference is performed on the standardized image to obtain a candidate localization set, and overlap suppression is performed on the candidate localization set to obtain the localization results of the stripped area and the indentation area. Confidence indicators are generated for the location results of the stripped area and the indentation area. When any confidence indicator does not meet the preset confidence conditions, the corresponding location result is marked as a verification location result and associated with the confidence indicator in the output result.

5. The visual inspection method for repairing broken airborne power lines as described in claim 1, characterized in that, Based on the localization results, the segmentation input region is determined. The corresponding region image is extracted from the standardized image, and this region image is input into the instance segmentation model to obtain the stripping mask and indentation mask. The steps include: Based on the location results of the stripped area and the indentation area, a segmentation input region is generated and boundary expansion is performed to obtain an expanded segmentation input region. Background suppression is performed within the extended segmentation input region to obtain a region image. The region image is then input into the instance segmentation model to obtain a stripping mask and an indentation mask. Calculate the coverage consistency index between the stripping mask and the corresponding positioning result for the stripping mask and the indentation mask respectively, and write the coverage consistency index into the process feature quantity.

6. The visual inspection method for airborne power line breakage repair as described in claim 1, characterized in that, The steps for extracting the axial direction of the wire body based on the stripping mask and indentation mask, performing attitude normalization, and determining the axial distance between the stripping end face and the indentation starting boundary include: Extract the center line of the line and determine the axis of the line within the union of the stripping mask and the indentation mask; Perform coordinate rotation normalization on the standardized image based on the line axis; In the rotated and normalized coordinate system, the stripping end face is determined based on the end boundary of the stripping mask in the axial projection, the indentation start boundary is determined based on the end boundary of the indentation mask in the axial projection, and the axial spacing is obtained based on the interval between the stripping end face and the indentation start boundary in the axial projection.

7. The visual inspection method for repairing broken airborne power lines as described in claim 6, characterized in that, The steps for determining the number of notches and writing them into the process feature quantities include: Within the area defined by the stripping mask, a scratch detection sub-region is constructed, and fine line structure enhancement is performed within the scratch detection sub-region to obtain a set of fine line candidates; Based on the axial direction of the line, a candidate set of indentations is obtained by performing directional consistency screening on the candidate set of thin lines. The number of notches is obtained by performing connectivity analysis and length consistency screening on the candidate notch set, and the number of notches is written into the process feature quantity.

8. The visual inspection method for repairing broken airborne power lines as described in claim 2, characterized in that, The steps of comparing process feature quantities with a preset rule base and outputting qualified, required, or unqualified results include: Perform consistency checks on the process feature quantities obtained from different images according to the posture identifier. When the consistency check does not meet the preset consistency conditions, output the result that needs to be reviewed and record the inconsistency type identifier. When the consistency check meets the preset conditions, the process feature quantities are matched one by one according to the hierarchical comparison order of the preset rule base, and the qualified or unqualified results are output, and the hit rule identifier is recorded.

9. A visual inspection system for repairing broken airborne power lines, characterized in that, include: The imaging unit is used to acquire images of the repaired area; The preprocessing unit is used to perform preprocessing on the image to obtain a standardized image. The preprocessing includes distortion correction and brightness normalization. The localization unit is used to input the standardized image into the target detection model to obtain the localization results, which include the localization results of the stripped area and the indentation area. The segmentation unit is used to determine the segmentation input region based on the localization result, extract the region image corresponding to the segmentation input region from the standardized image, and input the region image into the instance segmentation model to obtain the stripping mask and the indentation mask. The measurement unit is used to extract the axial direction of the line body based on the stripping mask and the indentation mask and perform orientation normalization. It obtains the conversion relationship between pixel quantity and physical quantity through scale calibration, determines the stripping length, the axial distance between the stripping end face and the indentation start boundary, the number of scratches and the cracking state, and writes them into the process feature quantity. The judgment unit is used to compare the process feature quantity with the preset rule library and output the qualified result, the result that needs to be reviewed, or the unqualified result.

10. The airborne power line breakage repair visual inspection system as described in claim 9, characterized in that, The segmentation unit includes a region generation subunit, a boundary expansion subunit, a background suppression subunit, and a mask output subunit. The decision unit includes a consistency calculation subunit and an index writing subunit. The region generation subunit is used to generate a segmented input region based on the location results of the stripping region and the indentation region. The boundary expansion subunit is used to perform boundary expansion on the segmented input region to obtain an expanded segmented input region. The background suppression subunit is used to perform background suppression within the extended segmented input region to obtain a region image; The mask output subunit is used to input the region image into the instance segmentation model to obtain the stripping mask and the indentation mask; The consistency calculation subunit is used to calculate the coverage consistency index of the stripping mask and the stripping area positioning result, as well as the indentation mask and the indentation area positioning result, respectively. The indicator writing sub-unit is used to write the coverage consistency indicator into the process feature quantity.

Images