Method for eliminating the shed of v-shaped suspension composite insulator based on multiple nonlinear transformation

Abstract

The present application belongs to the cross technical field of power equipment state monitoring and computer vision, and discloses a V-shaped suspension composite insulator umbrella skirt elimination method based on multiple nonlinear transformation. A large number of V-shaped suspension composite insulator infrared inspection images covering multi-line tower scenes are collected, a pre-trained lightweight single-stage target detection model is used to detect the V-shaped suspension composite insulator string inspection images in the data set, the alignment frame parameters of the core rod are output, and the core rod ROI region is obtained based on the alignment frame cutting. The core rod umbrella skirt interference is removed and the axial posture is calculated, and then the core rod axial posture is adjusted. Finally, segmentation and result mapping are realized. The present application preferentially eliminates the umbrella skirt shielding interference through mathematical modeling, and then accurately controls the ROI extraction, angle calculation, rotation correction and result mapping process, balancing the segmentation accuracy, real-time performance and positioning accuracy.

Description

Technical Field

[0001] This invention relates to the field of power equipment condition monitoring and computer vision interdisciplinary technology, and in particular to a method for eliminating the sheds of V-shaped suspension composite insulators based on multiple nonlinear transformations. Background Technology

[0002] V-shaped suspension composite insulator strings are widely used in 110kV and above high-voltage transmission lines due to their anti-icing and anti-wind deflection characteristics. The structural integrity of their internal core rods directly determines the safety and stability of the transmission system. The sheds of this type of insulator have an inherent insulation structure, which will form a surrounding shielding of the core rod. UAV inspection technology has become the mainstream method for condition monitoring of suspension composite insulators due to its wide coverage and high efficiency. However, inspection images often show the core rod tilted at a large angle (>45°) due to the long shooting distance and angle deviation. In addition, background interference (towers, conductors, vegetation) is serious, and the core rod features in the inspection images are easily obscured.

[0003] Existing segmentation techniques cannot be adapted to V-shaped string scenarios, with the following specific shortcomings:

[0004] Traditional segmentation methods (such as thresholding and edge detection) are based on pixel grayscale differences to achieve segmentation. These methods are easily affected by background noise, and the segmentation integrity of large-angle mandrels is less than 30%.

[0005] Deep learning segmentation methods learn high- and low-resolution mappings and semantic features through deep networks, but the general model is not optimized for the linear long structure and large tilt angle characteristics of suspension composite insulator core rods, and has the following shortcomings:

[0006] Large-angle target segmentation failure: The model relies on horizontal / vertical feature extraction. After the core is tilted, the features are misaligned, and the segmentation integrity is ≤32%.

[0007] Global correction computation is large: Existing global rotation correction methods (such as affine transformation global processing) account for more than 90% of the computation, and the processing time for a single image is ≥125ms, which cannot meet the real-time requirements of UAV inspection (frame rate ≥15fps).

[0008] Inconsistent segmentation and positioning: The segmentation results do not reflect the original image coordinates, and the positioning error is ≥11 pixels, making it difficult to accurately match the actual position of the core rod. Therefore, a dedicated segmentation method for the core rod of V-shaped suspension composite insulator strings is urgently needed. Summary of the Invention

[0009] To overcome the technical defects of existing technologies, this invention provides a method for eliminating the shading of V-shaped suspension composite insulator skirts based on multiple nonlinear transformations. The method prioritizes eliminating the shading interference of the skirts through mathematical modeling, and then precisely controls the ROI extraction, angle calculation, rotation correction and result mapping process to balance segmentation accuracy, real-time performance and positioning accuracy.

[0010] The technical solution adopted in this invention is: a method for eliminating the sheds of V-shaped suspension composite insulators based on multiple nonlinear transformations, comprising the following steps:

[0011] Step 1: Collect a large number of infrared inspection images of V-shaped suspension composite insulators covering multiple line tower scenarios, and perform image preprocessing and mandrel annotation to construct a dataset;

[0012] Step 2: Use a pre-trained lightweight single-stage target detection model to detect the V-shaped suspension composite insulator string inspection images in the dataset, output the alignment box parameters of the core rod, and crop the core rod ROI region based on the alignment box.

[0013] Step 3: Locate the main body of the umbrella skirt within the ROI region through the periodic regularity of the umbrella skirt, then perform a nonlinear transformation, combine angle histogram analysis to identify the dominant angle of interference of the umbrella skirt, and remove the interference edges according to the pixel gradient direction to obtain the purified edge image. Finally, perform a second nonlinear transformation on the purified edge image and extract the straight line, and calculate the current axis posture of the core rod by weighted average of the line length.

[0014] Step 4: Construct an affine transformation matrix with the center of the alignment frame as the rotation center and the axial orientation as the rotation angle. Rotation correction is performed on the ROI region to obtain an upright mandrel image;

[0015] Step 5: Input the rotated and corrected upright mandrel image into the semantic segmentation model to obtain a pixel-level classification mask; by calculating the inverse matrix of the original pose correction transformation, map the segmentation mask from the correction coordinate system back to the original image coordinate system to realize the localization of the detection result in the original inspection image.

[0016] Furthermore, in step 3, the specific process of identifying the dominant angle of interference from the umbrella skirt by combining angle histogram analysis is as follows:

[0017] The polar angles of all candidate lines extracted by the first nonlinear transformation are counted, and an angle histogram covering 0°-180° is constructed.

[0018] Calculate the average of all peaks in the histogram. Filter the angle histogram to find peak height ≥ The angle of the umbrella skirt is the dominant angle set for interference.

[0019] Furthermore, the specific method for removing interfering edges based on the pixel gradient direction in step 3 is as follows:

[0020] Calculate the gradient direction of each edge pixel in the original edge image;

[0021] If the edge image The gradient direction of the middle pixel and the dominant angle of any umbrella-shaped interference If the difference is within ±5°, the corresponding pixel will be removed from the edge image.

[0022] Furthermore, in the operation of second nonlinear transformation and line extraction, lines with large inclination angles with an absolute slope value > 1 are selected.

[0023] Furthermore, the method for calculating the current axial orientation of the mandrel in step 3 using a line length weighted average is to calculate the final tilt angle of the mandrel using a line length weighted average. The formula is used for calculation, and is expressed as:

[0024] ;

[0025] In the formula, This represents the total number of straight lines extracted to represent the edges of the mandrel; For the first The pixel length of a straight line; For the first The angle of inclination corresponding to each straight line.

[0026] Furthermore, the affine transformation matrix in step 4... for:

[0027] ;

[0028] In the formula, Indicates the center coordinates.

[0029] Furthermore, the inverse matrix in step 5 for:

[0030] .

[0031] Furthermore, the collection of a large number of infrared inspection images of V-shaped suspension composite insulators covering multiple line tower scenarios includes sunny days, cloudy days, light rain, and scenarios covering multiple line towers.

[0032] Furthermore, the image preprocessing includes image cropping, preserving the core area of ​​the insulator, and standardizing the size.

[0033] Furthermore, the method for marking the core rod is to use manual marking combined with an edge masking algorithm to generate a core rod alignment box and a binary edge mask matrix for core rod marking.

[0034] The beneficial effects of this invention are as follows: Addressing the unique problems of "mutual shading between upper and lower skirts and strong interference from periodic sawtooth edges" inherent in V-shaped suspension composite insulator strings, as well as the shortcomings of existing methods for segmenting the core rod at the bottom, such as breakage, high computational load, and inconsistency between segmentation and positioning, this invention constructs a technical system for dynamic ROI extraction, core rod skirt interference removal, axial attitude calculation and adjustment, and precise core rod segmentation. Specifically, the core rod skirt interference removal and axial attitude calculation detect the skirt interference angle and use an angle histogram to remove periodic edges; the core rod axial attitude adjustment and segmentation precisely extract the core rod's straight line, and through related technical processing, achieve high-precision segmentation of the V-shaped string core rod. Attached Figure Description

[0035] One or more embodiments are illustrated by way of example with reference numerals in the accompanying drawings. These illustrations do not constitute a limitation on the embodiments. Elements with the same reference numerals in the drawings are denoted as similar elements. Unless otherwise stated, the figures in the drawings are not to be limited by scale.

[0036] Figure 1 This is a schematic diagram of the process of the present invention;

[0037] Figure 2 This is a schematic diagram illustrating the elimination of the core rod umbrella skirt according to the present invention;

[0038] Figure 3 This is a schematic diagram of the mandrel segmentation according to the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of this invention clearer, the various embodiments of this invention will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the various embodiments of this invention to facilitate a better understanding of this application. However, the technical solutions claimed in the claims of this application can be implemented even without these technical details and with various variations and modifications based on the following embodiments.

[0040] like Figures 1 to 3 As shown, this embodiment provides a method for eliminating the sheds of a V-shaped suspension composite insulator based on multiple nonlinear transformations, including the following steps:

[0041] Step 1: Collect a large number of infrared inspection images of V-shaped suspension composite insulators covering multiple line tower scenarios, and perform image preprocessing and mandrel annotation to construct a dataset;

[0042] The collection included a large number of infrared inspection images of V-shaped suspension composite insulators covering multiple line tower scenarios, including sunny days, cloudy days, light rain, and scenarios covering multiple line towers.

[0043] Image preprocessing operations are as follows:

[0044] Trimming: Preserve the core area of ​​the insulator and maintain a uniform size; this helps reduce interference from irrelevant background and ensures consistent input data size, facilitating model processing;

[0045] Normalization: Maps pixel values ​​to the [0,1] interval; helps improve the training stability and convergence speed of the model; This represents the normalized pixel value. Represents the original pixel value;

[0046] Data augmentation: The dataset was expanded using methods such as "angle perturbation (-15°~15°), illumination perturbation (±20%), and occlusion simulation (randomly adding umbrella skirt fragments)" to increase the model's generalization ability and avoid overfitting;

[0047] Core annotation: Manual annotation combined with edge masking algorithm is used to generate core alignment boxes and binary edge mask matrices for training object detection and segmentation models.

[0048] Step 2: Use a pre-trained lightweight single-stage target detection model to detect the V-shaped suspension composite insulator string inspection images in the dataset, output the alignment box parameters of the core rod, and crop the core rod ROI region based on the alignment box.

[0049] Specifically, this involves processing the inspection images of the V-shaped suspension composite insulator strings to be processed. Input a pre-trained lightweight single-stage object detection model (such as the YOLO series models).

[0050] A lightweight single-stage object detection model is adopted, which extracts features through 4 levels of convolution:

[0051] Level 1 convolution: , This indicates a convolution operation with a 3×3 kernel and 64 output channels. This represents the feature map output after the first-level convolution operation.

[0052] Levels 2-4 convolution: , , Indicates the first The feature map output after a multi-level convolution operation. Indicates the first The feature map output after a multi-level convolution operation. This indicates a batch normalization operation. express Activation function;

[0053] Final output feature map Mandrel testing accuracy AP ≥ 85.5%;

[0054] ROI trimming: Model output mandrel alignment box parameters , Indicates the center coordinates, , Representing width and height respectively, the coordinate set is used to clip the ROI region based on the alignment box, satisfying:

[0055] ;

[0056] In the formula, This represents the combination of all pixel coordinates of the core ROI region after cropping, i.e., the image region to be retained; Represents the x-coordinate of a pixel within the ROI region; This represents the ordinate of a pixel within the ROI region.

[0057] It uses the "alignment box" output by the target detection model to accurately extract the region containing only the target core rod from the original inspection image containing a complex background.

[0058] Step 3: Locate the main body of the umbrella skirt within the ROI region based on the periodic pattern of the umbrella skirt. Then, through a nonlinear transformation, identify the dominant angle of interference of the umbrella skirt by combining angle histogram analysis. Remove the interference edges according to the pixel gradient direction to obtain the purified edge image. Finally, perform a second nonlinear transformation on the purified edge image and extract straight lines. Calculate the current axis orientation of the core rod by weighted average of line lengths.

[0059] The core of this invention lies in the application of multiple nonlinear transformations, a progressive optimization process from coarse to fine. The first transformation (coarse screening and diagnosis) aims not to directly identify the core rod, but to intelligently diagnose interference sources. By analyzing the angle histograms of all candidate lines, the dominant angles of the periodic edges of the umbrella skirt are systematically identified. This is equivalent to first understanding the distribution pattern of the target, laying the foundation for subsequent precise removal.

[0060] The second transformation (refinement and extraction): After removing known interference angles, the "cleaned" edge image is transformed again. At this point, noise and interference in the image are greatly reduced, allowing the algorithm to more effectively extract long straight-line features representing the true axis of the core rod. This mechanism decomposes the complex problem of "finding the target in noise" into three controllable steps: "first identifying noise, then eliminating noise, and finally accurately locating the target," significantly improving the robustness and reliability of the algorithm.

[0061] Perform umbrella skirt unit positioning: Calculate the average vertical spacing between adjacent umbrella skirts within the ROI. Fitting the radius of the umbrella skirt Define the size of the periodic unit. , , Indicates the width of a periodic unit. Indicates the height of a periodic unit;

[0062] Slide window to traverse ROI and filter by core pixel percentage The window is used as the valid periodic unit, and the set of valid periodic units is output.

[0063] Target subject identification:

[0064] For each valid class of periodic units The local threshold is calculated using the Otsu algorithm. Binarization yields the binarized image. Then, all the binarized images are merged into a single binarized image. The overall binarized image obtained by 3×3 rectangular kernel erosion Obtain corrosion image Differential positioning insulator edge image .

[0065] Among them, the Otsu algorithm (also known as the "Otsu algorithm" or "max inter-class difference method") is a widely used algorithm in image processing for automatically determining the optimal threshold for image binarization.

[0066] Step 4: Construct an affine transformation matrix with the center of the alignment frame as the rotation center and the axial orientation as the rotation angle. Rotation correction is performed on the ROI region to obtain an upright mandrel image.

[0067] Step 5: Input the rotated and corrected upright mandrel image into the semantic segmentation model to obtain a pixel-level classification mask; by calculating the inverse matrix of the original pose correction transformation, map the segmentation mask from the correction coordinate system back to the original image coordinate system to achieve accurate positioning of the detection result in the original inspection image.

[0068] The key difference is that this invention performs localized processing, whereas traditional methods involve rotating and correcting the entire inspection image. A high-resolution infrared image has a massive number of pixels (e.g., 1920×1080), making affine transformation computationally extremely complex. This invention rotates only the cropped core region of interest (ROI). The ROI typically occupies only a small portion of the original image (e.g., 200×800 pixels), drastically reducing the computational load. This is the key reason for reducing single-image processing time from 125ms to 68ms (a reduction of approximately 45%), enabling this method to meet the real-time processing requirements of ≥15fps for UAV inspections.

[0069] Using the center of the detection box as the rotation center means that the rotation operation is performed around the "centroid of the current mandrel region." This ensures that the rotated mandrel remains centered in the image, avoiding the problem of the target being rotated out of the image boundary that might occur when rotating around the image origin. Since the rotation is based on the target's own position and orientation, the rotated ROI can completely contain the corrected mandrel without secondary cropping, avoiding pixel loss or target incompleteness caused by multiple croppings, and maintaining target integrity.

[0070] By using rotation correction, tilted mandrels at arbitrary angles are uniformly "aligned" into a near-vertical upright posture, significantly reducing the learning and recognition difficulty for subsequent semantic segmentation models. The linear structural features of the upright mandrel better match the features learned by the model. This allows the model to more accurately capture the continuous, slender edges of the mandrel. This directly leads to significant improvements in segmentation integrity (from 65% to 91%) and segmentation IoU (from 78.5% to 90.5%), effectively solving the core challenge of "fragmented segmentation of targets at large tilt angles."

[0071] The method for identifying the angular distribution of the umbrella skirt is to analyze the edge image. Performing a nonlinear transformation is represented as:

[0072] ;

[0073] In the formula, This represents a composite objective function used to evaluate the set of candidate skirt lines. The quality is determined by combining data fitting terms and periodic prior terms; For the periodic parameters of the umbrella skirt (such as the average spacing between adjacent umbrella skirts); This represents the standard deviation of the pixel error (controlling the tolerance range of the error between the pixel and the line fitting). The weighting coefficients represent the relative importance of the "data fitting term" and the "periodic prior term" in the balancing formula. This is an indicator function used to determine whether a point is near a line; Representing edge image The first in 1 pixel; Represents the set of candidate umbrella skirt lines The Middle A straight line; For pixels The weights; Represents pixels and The degree of matching; express The Middle A straight line; express The Middle A straight line; express , The periodic parameters of the umbrella skirt The probability of; The set of all edge points;

[0074] Represented as:

[0075] ;

[0076] In the formula, This represents the tolerance range from the point to the line. , express The horizontal and vertical coordinates; express Polar coordinate distance parameters (from origin to...) (vertical distance); express Polar coordinate angle parameters (from origin to...) The angle between the vertical line segment and the x-axis.

[0077] Gradient magnitude weight Represented as:

[0078] ;

[0079] In the formula, Representing an image In pixels gradient at the point, image The image of the core-skirt region obtained after preprocessing (such as skirt unit localization and local threshold binarization). It is the gradient operator in image processing, which represents the gradient of image I at pixel p.

[0080] Represented as:

[0081] ;

[0082] In the formula, for The gradient direction.

[0083] Represented as:

[0084] ;

[0085] In the formula, , They represent , The polar coordinate distance parameter.

[0086] The second nonlinear transformation operation in the cleaned edge image and line extraction is the same as the first nonlinear transformation algorithm. Integration allows the algorithm to utilize both low-level image features (such as edge intensity) and high-level semantics (such as the periodicity of the skirt), thereby accurately distinguishing core rod and skirt interference in complex backgrounds.

[0087] Weights are assigned to edge points based on gradient magnitude to ensure that high-contrast edges (such as mandrel contours) dominate the transformation and reduce the impact of noise.

[0088] By using the cosine square term to measure the degree of matching between the gradient direction and the straight line normal, the contribution of directionally consistent edges is enhanced, and false detections of irrelevant edges are avoided.

[0089] Among them, the optimal set of lines for:

[0090] ;

[0091] In the formula, Given a coarse set of candidate umbrella-shaped lines containing all possible edge lines, directly using this set for analysis is akin to searching for a weak signal amidst massive noise, highly prone to failure. Maximizing... The optimal set of lines is obtained. Select the group that best meets the standard from all the straight lines to achieve precise interference removal.

[0092] Extract a set of 50 candidate lines , to capture the angular distribution of the periodic edges of the umbrella skirt.

[0093] In step 3, the specific process of identifying the dominant angle of interference from the umbrella skirt by combining angle histogram analysis is as follows:

[0094] Statistical candidate line set Polar angles of all straight lines Construct an angle histogram covering 0° to 180°;

[0095] Calculate the average of all peaks in the angle histogram. Filter the peak height in the angle histogram ≥ The angle of the umbrella skirt is the dominant angle set for interference.

[0096] The specific method for removing interfering edges based on pixel gradient direction in step 3 is as follows:

[0097] Calculate edge image The gradient direction of each pixel;

[0098] If the edge image The gradient direction of the middle pixel and any dominant interference angle If the difference is ≤ ±5°, then the corresponding pixel is set to 0 to obtain a cleaned edge image. .

[0099] The method for calculating the current axis attitude of the mandrel in step 3 using line length weighted average is to calculate the final tilt angle of the mandrel using line length weighted average. The calculation is performed using the following formula:

[0100] ;

[0101] In the formula, This represents the total number of straight lines extracted to represent the edges of the mandrel; For the first The pixel length of a straight line; For the first The tilt angle corresponding to each straight line has a calculation error of ≤ ±1.0°. Straight lines with large absolute slope values ​​and large tilt angles are specifically selected, and the final angle is calculated using a line length-weighted average. This ensures that the mandrel's attitude can be accurately perceived even at extreme angles that are close to vertical or horizontal, fundamentally solving the problem of fragmented segmentation of large-tilt targets caused by biased training data in general models.

[0102] In step 4, an affine transformation matrix is ​​constructed with the center of the alignment frame as the rotation center and the axial orientation as the rotation angle. The specific operation for performing rotation correction on the ROI region to obtain an upright mandrel image is to align it with the center of the bounding box. With the center of rotation as the rotation angle, construct a rotation angle of... affine transformation matrix :

[0103] .

[0104] It's not a simple rotation matrix, but a complete 2D affine transformation matrix. It encapsulates both rotation and translation transformations. , Element: precisely describes the rotation operation around the origin. Translation terms: These terms are crucial; they ensure that rotation is performed around a specified center point (i.e., the center of the alignment box), rather than the image's default origin (0,0). This prevents the target object from running out of the viewport after rotation.

[0105] It achieves the complex operation of "rotating around an arbitrary point" through a single, compact mathematical expression, ensuring that the mandrel is fully and centrally displayed in the corrected ROI after rotation.

[0106] This provides the prerequisite for pixel-level accurate inverse transformation: since T is a structurally standard transformation matrix with known parameters, it is absolutely invertible as long as its determinant is not zero (this condition must be satisfied for pure rotation). This strict mathematical invertibility is the fundamental prerequisite for perfectly mapping the segmentation results back to the original image. If approximate or irreversible transformations (such as certain nonlinear distortion corrections) are used, the re-mapping accuracy cannot be guaranteed. This lays the mathematical foundation for achieving the excellent metric of "positioning error < 2 pixels".

[0107] Calculate the dimensions of the rotated ROI: Calculate the width of the rotated ROI using geometric relationships. With Gao Ensure the core rod is intact and uncut:

[0108] ;

[0109] .

[0110] The minimum rectangular size that can completely contain the rotated mandrel is dynamically calculated, ensuring that the entire mandrel, after pose adjustment, is within the new ROI image, with no part accidentally cut off due to rotation. This is a prerequisite for achieving high segmentation integrity (91%). If the mandrel is cropped at both ends, no matter how powerful the subsequent segmentation model is, it cannot recover the lost parts. The complete mandrel image provides the segmentation model with continuous contextual information, helping the model better understand the global structure and local features of the mandrel, thereby making more accurate judgments.

[0111] The new ROI size ensures integrity while avoiding the introduction of excessive irrelevant background regions. This controls the amount of data input to the segmentation model and contributes positively to improving processing speed.

[0112] Bilinear interpolation optimization is performed: bilinear interpolation is used to remap the rotated pixels, preserving the edge sharpness of the core rod. The pixel value calculation formula is as follows:

[0113] ;

[0114] In the formula, This indicates the coordinates in the corrected (rotated) upright mandrel image. Pixel value at; These represent the weighting coefficients of bilinear interpolation. , Both represent the neighboring pixel indices during interpolation; This represents the coordinates within the original core rod ROI region. The neighboring pixel values ​​at the location; the edge sharpness retention rate ≥95%, finally yielding an upright mandrel image. .

[0115] Compared to the simplest nearest neighbor interpolation (which results in noticeable jagged edges), bilinear interpolation produces a smooth transition by taking a weighted average of four pixels in a 2x2 region around the target point.

[0116] Segmentation models, especially deep learning-based models, are highly sensitive to the quality of the input image. Clear, smooth edges are more conducive to the model extracting accurate boundary features. Blurred or severely jagged edges can confuse the model and reduce segmentation accuracy.

[0117] Maintaining the subtle gaps between the mandrel and the skirt: The boundary between the mandrel and the skirt is a crucial segmentation boundary. Bilinear interpolation better preserves the continuity of these subtle gaps, preventing them from becoming fragmented and increasing the difficulty of segmentation.

[0118] Image of an upright mandrel Input an improved U-Net model (with the decoder incorporating a CBAM attention mechanism to focus on the fine edges of the core):

[0119] Encoder: 4-level convolution + max pooling (output channels 64→128→256→512) to extract multi-scale features;

[0120] Decoder: 4-level transposed convolution (3×3 kernel, stride 2) + feature concatenation to restore resolution;

[0121] Output: Core binary segmentation mask , segmentation IoU≥92%.

[0122] Symmetrical structure designed specifically for pixel-level segmentation:

[0123] Encoder (downsampling): Through four levels of convolution and pooling, the receptive field is progressively expanded to extract multi-scale information from low-level edge textures to high-level semantic features. This enables the model to understand "what is the core" and "the difference between the core, the skirt, and the background."

[0124] Decoder (upsampling): Through four levels of transposed convolution, spatial resolution is gradually restored, combining high-level semantic features with precise localization information. This is the key to achieving pixel-level output.

[0125] Skip connections: By concatenating the high-resolution feature map of the encoder with the corresponding layer of the decoder, this effectively compensates for the details lost during downsampling (such as the precise edges of the mandrel), solving the problem of decreased localization accuracy in deep networks. The mandrel is a typical long, thin, continuous target. U-Net's encoder can capture its global coherence, while the decoder and skip connections ensure the continuity and integrity of its segmentation results, effectively avoiding the "segmentation breakage" problem common in traditional methods.

[0126] The core innovation of this improved model lies in its paradigm shift from "average effort" to "precise focus," automatically assessing the importance of each feature channel. The model learns which feature maps are crucial for identifying the "core" (e.g., those responsible for extracting vertical edges and specific textures) and assigns them higher weights. When facing complex infrared backgrounds (such as towers and wires), the model actively suppresses redundant feature channels related to the background while enhancing channels related to the core's essential features, making the model more robust to interference. On each feature map, it automatically calculates which spatial locations require focused attention. It generates an attention heatmap, where bright areas represent regions the model needs to focus on—key to achieving high-precision edge segmentation. The spatial attention mechanism significantly strengthens the response of subtle but crucial areas such as the junction between the core and the skirt, and the core's own contour, guiding the model to allocate computational resources to clarifying these difficult-to-distinguish boundaries. Simultaneously, it weakens the weight of uniform or irrelevant background areas in the image, further refining the segmentation environment.

[0127] Finally, the inverse affine transformation is reiterated: Constructing the affine transformation matrix in step three. inverse matrix , segmentation mask Repost to the original inspection image:

[0128] .

[0129] yes It performs a precise inverse operation. It can map each pixel in the corrected coordinate system back to the coordinate system of the original image one-to-one without deviation. The segmentation model works on the corrected "upright" image, and its output mask is also located in this corrected coordinate system. It acts like a precise navigator, guiding each pixel in the "upright world" safely back to its home in the "original world." This is the core technology for solving the industry pain point of "segmentation and localization inconsistency." It ensures that what the segmentation model sees is what is ultimately located on the original image, achieving true pixel-level alignment.

[0130] and Together, they form a perfect mathematical closed loop. This loop ensures that only coordinate system transformations occur during information processing, without any information loss or distortion. This improves the reliability and repeatability of the entire method. As long as the input and parameters are the same, the result is deterministic, which is crucial for industrial inspection applications.

[0131] Throughout the entire process, the ROI image only underwent one resampling (i.e., from...). (Performed rotation correction). When mapping the mask back, the method used is... Instead of resampling the image data a second time, this invention performs coordinate transformation. Rotating and re-applying the segmented mask image itself introduces a second interpolation error, potentially leading to blurred edges or pixel misalignment. This invention, however, uses... By simply recalculating the coordinate relationships, the binary data of the mask itself remains unchanged, preserving the edge clarity of the segmentation result to the maximum extent, so that the edges of the reconstructed core are "continuous and clear, without jagged edges or breaks".

[0132] Reply rules: , The final segmentation result has a positioning error of <2 pixels. In the original inspection image, coordinates The final segmentation mask value at that location is used to mark whether that location in the original inspection image belongs to the mandrel region.

[0133] Reply rules are based on the inverse matrix This process firmly pins the "virtual" segmentation result back to its original physical location in the original inspection image. If a crack or breakage point is found on the segmented mandrel, this location can be precisely circled on the original image, and the inspection personnel can be informed: "There is a defect at a certain coordinate near the third skirt from the bottom on tower A, insulator string B." Measurement is possible: Only with accurate coordinate correspondence can the physical dimensions be estimated (e.g., estimating the actual crack length using pixel distance and camera parameters).

[0134] The entire transformation process (from the original ROI to the upright ROI) is carried out by a single, parameterized affine transformation matrix. Description. This transformation is reversible.

[0135] Since the center of rotation and the angle are known and precisely calculated, their inverse matrix can be easily obtained. This method is used to map the segmentation results back to the original image coordinate system without loss and with high precision. This localized transformation, based on the same reference frame (alignment box), ensures that the coordinate relationship before and after processing is strictly corresponding. This is the fundamental guarantee for achieving the excellent indicator of "positioning error < 2 pixels". If a global rotation is performed and then reversed for positioning, the error will accumulate and be amplified.

[0136] and Together, a two-way channel was established from the original image to the processing space and then perfectly returned. Guaranteed accuracy transfer: The localization accuracy of the object detection step, the angular accuracy of the pose calculation, and the pixel accuracy of the segmentation model are losslessly transferred and integrated into the final output. Elegant systems engineering: Using simple yet powerful mathematical tools, the discrete steps in image processing (detection, correction, segmentation, and reassembly) are linked into a coherent, rigorous, and automated pipeline. Therefore, and It's not just the two formulas for achieving rotation and inverse rotation, but also the soul of the entire invention that unifies "high-precision segmentation" and "high-precision positioning".

[0137] Experimental verification was conducted by selecting 500 inspection images of V-shaped suspension composite insulators and comparing them with existing methods (bicubic interpolation + YOLOv5, global rotation + U-Net). The verification results are shown in Table 1.

[0138] Table 1 Comparison between the present invention and existing methods

[0139]

[0140] The above experiments confirm that:

[0141] 1. The segmentation integrity of the method of the present invention is improved by 59-184% compared with the existing methods, the single image processing time is reduced by 45-46%, the positioning error is reduced by 83-89%, the segmentation IoU is improved by 15-25%, and the edge error is reduced by 66-78%, which is comprehensively superior to the comparison methods;

[0142] 2. The reconstructed core rod edges by the method of the present invention are continuous and clear, without jagged edges or breaks, and can retain the subtle gap features between the core rod and the skirt. The background interference pixel ratio is ≤4.2%, while the comparison method has obvious edge blurring (bicubic interpolation) or excessive smoothing (global rotation) problems.

[0143] Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0144] Those skilled in the art will understand that the above embodiments are specific examples of implementing the present invention, and in practical applications, various changes in form and detail may be made without departing from the spirit and scope of the present invention.

Claims

1. A method for eliminating the shed skirts of a V-shaped suspension composite insulator based on multiple nonlinear transformations, characterized in that, Includes the following steps: Step 1: Collect a large number of infrared inspection images of V-shaped suspension composite insulators covering multiple line tower scenarios, and perform image preprocessing and mandrel annotation to construct a dataset; Step 2: Use a pre-trained lightweight single-stage target detection model to detect the V-shaped suspension composite insulator string inspection images in the dataset, output the alignment box parameters of the core rod, and crop the core rod ROI region based on the alignment box. Step 3: Locate the main body of the umbrella skirt within the ROI region through the periodic regularity of the umbrella skirt, then perform a nonlinear transformation, combine angle histogram analysis to identify the dominant angle of interference of the umbrella skirt, and remove the interference edges according to the pixel gradient direction to obtain the purified edge image. Finally, perform a second nonlinear transformation on the purified edge image and extract the straight line, and calculate the current axis posture of the core rod by weighted average of the line length. Step 4: Construct an affine transformation matrix with the center of the alignment frame as the rotation center and the axial orientation as the rotation angle. Rotation correction is performed on the ROI region to obtain an upright mandrel image; Step 5: Input the rotated and corrected upright mandrel image into the semantic segmentation model to obtain a pixel-level classification mask; by calculating the inverse matrix of the original pose correction transformation, map the segmentation mask from the correction coordinate system back to the original image coordinate system to realize the localization of the detection result in the original inspection image; In step 3, the specific process of identifying the dominant angle of interference from the umbrella skirt by combining angle histogram analysis is as follows: The polar angles of all candidate lines extracted by the first nonlinear transformation are counted, and an angle histogram covering 0°-180° is constructed. Calculate the average of all peaks in the histogram. Filter the angle histogram to find peak height ≥ The angle, as the set of dominant angles for interference from the umbrella skirt; The specific method for removing interfering edges based on pixel gradient direction in step 3 is as follows: Calculate the gradient direction of each edge pixel in the original edge image; If the edge image The gradient direction of the middle pixel and the dominant angle of any umbrella-shaped interference If the difference is within ±5°, the corresponding pixel will be removed from the edge image.

2. The method for eliminating V-shaped suspension composite insulator skirts based on multiple nonlinear transformations according to claim 1, characterized in that: In the operation of quadratic nonlinear transformation and line extraction, lines with large inclination angles with an absolute slope value greater than 1 are selected.

3. The method for eliminating V-shaped suspension composite insulator skirts based on multiple nonlinear transformations according to claim 2, characterized in that: The method for calculating the current axial orientation of the mandrel in step 3 by weighted averaging of line lengths is to calculate the final tilt angle of the mandrel using weighted averaging of line lengths. The formula is used for calculation, and is expressed as: ； In the formula, This represents the total number of straight lines extracted to represent the edges of the mandrel; For the first The pixel length of a straight line; For the first The angle of inclination corresponding to each straight line.

4. The method for eliminating V-shaped suspension composite insulator skirts based on multiple nonlinear transformations according to claim 3, characterized in that: The affine transformation matrix in step 4 for: ； In the formula, Indicates the center coordinates.

5. The method for eliminating V-shaped suspension composite insulator skirts based on multiple nonlinear transformations according to claim 4, characterized in that: The inverse matrix in step 5 for: 。 6. The method for eliminating V-shaped suspension composite insulator skirts based on multiple nonlinear transformations according to claim 1, characterized in that: The aforementioned collection of a large number of infrared inspection images of V-shaped suspension composite insulators covering multiple line tower scenarios includes sunny days, cloudy days, light rain, and scenarios covering multiple line towers.

7. The method for eliminating the sheds of a V-shaped suspension composite insulator based on multiple nonlinear transformations according to claim 1, characterized in that: The image preprocessing includes image cropping, preserving the core area of ​​the insulator, and standardizing the size.

8. The method for eliminating V-shaped suspension composite insulator skirts based on multiple nonlinear transformations according to claim 1, characterized in that: The method for marking the core rod is to use manual marking combined with an edge masking algorithm to generate a core rod alignment box and a binary edge mask matrix for core rod marking.

Images