Online detection method for surface defects of corrugated outer conductor in radio frequency coaxial cable

By acquiring multi-view cylindrical images using a ring-shaped array of area cameras, removing overlapping edge pixels and performing pixel-level coordinate projection transformation, a flattened panoramic image of the corrugated surface is generated. This solves the dynamic distortion problem in the detection of surface defects on the outer conductor of radio frequency coaxial cables, and achieves accurate defect separation and real-time interception.

CN122887917BActive Publication Date: 2026-11-10HANGZHOU LINAN GUANGDA CABLE
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611381584.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-08
Publication Date
2026-11-10
Estimated Expiration
2046-09-08

AI Technical Summary

Technical Problem

During the high-speed continuous extrusion molding process of the corrugated outer conductor of the radio frequency coaxial cable, the static filter template cannot match the trough shadow of dynamic distortion, resulting in the leakage of normal structural periodic shadows as a high-frequency feature, leading to defect detection failure and amplification of parameter accumulation error.

Method used

A ring-shaped array of area cameras is used to acquire multi-view cylindrical images. Overlapping edge pixels are removed, and pixel-level coordinate projection transformation is performed to generate a flattened wavy panoramic image. The mean axial brightness projection and time axis distortion offset are calculated to construct a two-dimensional spatial phase drift field. The reference image is stitched together and decomposed into a three-dimensional reference tensor block. Defect features are extracted through a quadtree structure to generate a real-time interception signal.

Benefits of technology

It effectively eliminates spatial redundancy and grayscale jumps caused by multi-view stitching, smooths non-rigid local distortions, and achieves accurate separation of wavy backgrounds and abnormal defects without relying on prior defect templates, thereby improving the accuracy of image data processing and its engineering application value.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122887917B_ABST
    Figure CN122887917B_ABST
Patent Text Reader

Abstract

This invention relates to the field of image processing technology and discloses an online detection method for surface defects on the corrugated outer conductor of a radio frequency coaxial cable. The method includes: acquiring multi-view cylindrical images and removing overlapping edges, then transforming the images by projection to obtain a flattened panoramic image of the corrugation; calculating the time axis distortion offset to form a two-dimensional spatial phase drift field; using this drift field to correct the flattened image and stitching it into a three-dimensional reference tensor block and decomposing it into a sparse anomalous defect tensor; mapping it into a quadtree data structure and merging connected components; and generating a real-time interception signal based on the physical size of the defect. This invention eliminates nonlinear geometric distortion and shadow noise leakage, and achieves the separation of the corrugated background and the defect, as well as the precise closed-loop conversion from pixel-level defect response to physical space engineering interception actions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of image processing technology and discloses an online detection method for surface defects on the corrugated outer conductor of a radio frequency coaxial cable. Background Technology

[0002] With the development of intelligent manufacturing and machine vision technology, cable production systems typically use image processing-based surface inspection equipment to achieve automated identification of surface defects on the corrugated outer conductor of radio frequency coaxial cables. In the process of detecting surface defects on the corrugated outer conductor of radio frequency coaxial cables, the surface inspection equipment needs to continuously acquire two-dimensional images of the surface of the corrugated outer conductor of the radio frequency coaxial cable, and then extract periodic corrugated image features through frequency domain filtering algorithms. This allows for the separation of background image information of normal corrugations from high-frequency abnormal pixel features, ultimately completing pixel-level segmentation of the surface defect area of ​​the corrugated outer conductor of the radio frequency coaxial cable.

[0003] However, in high-speed cable routing scenarios involving continuous extrusion forming, the static filter template cannot match the dynamically distorted trough shadows due to the technical limitations of the static global spatial alignment assumption. This causes normal structural periodic shadows to leak as high-frequency features into the abnormal space, resulting in feature separation failure and amplified parameter accumulation errors. Therefore, how to eliminate the optical shadow noise leakage introduced by the periodic phase drift of the corrugated outer conductor during online inspection of surface defects in RF coaxial cables using surface inspection equipment is a problem that urgently needs to be solved in this field. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide an online detection method for surface defects on the corrugated outer conductor of a radio frequency coaxial cable. To achieve the above objectives, the present invention provides the following technical solution: An online detection method for surface defects on the corrugated outer conductor of a radio frequency coaxial cable includes: Obtain the original multi-view cylindrical image, extract and remove overlapping edge pixels from the original multi-view cylindrical image to obtain the corrected multi-view cylindrical image, and perform pixel-level coordinate projection transformation on the corrected multi-view cylindrical image to obtain the flattened wavy panoramic image. The mean axial brightness projection is calculated based on the flattened wavy panoramic image and combined into an axial brightness projection sequence. The time axis distortion offset is calculated using the axial brightness projection sequence and combined into a time axis distortion offset sequence. The column offset of the flattened wavy panoramic image is calculated based on the time axis distortion offset sequence. The column offsets are then arranged in a matrix to form a two-dimensional spatial phase drift field. A reference image is obtained by offset correction of the flattened wavy panoramic image using a two-dimensional spatial phase drift field. The reference image is then stitched together to obtain a three-dimensional reference tensor block, and the three-dimensional reference tensor block is decomposed to obtain a sparse anomaly defect tensor. The sparse anomaly defect tensor is mapped to a quadtree data structure. The bottom-level nodes are extracted and removed from the quadtree data structure to obtain the defect feature quadtree. Connected component merging is performed on the defect feature quadtree to obtain the target defect connected component set. The physical size of the defect is calculated based on the target defect connected component set, and a comparison and judgment are made based on the physical size of the defect to generate a real-time interception signal.

[0005] Furthermore, the method for obtaining the original multi-view cylindrical image includes: By detecting the circular array of area cameras at the workstation, exposure and acquisition are completed simultaneously based on the synchronous trigger signal output by the rotary encoder, and the original multi-view cylindrical image of the corrugated outer conductor of the RF coaxial cable is obtained. The data structure of the original multi-view cylindrical image is multiple two-dimensional grayscale matrices, with row indices corresponding to pixel positions in the physical radial circumferential direction and column indices corresponding to pixel positions in the physical axial direction.

[0006] Furthermore, the method for obtaining the corrected multi-view cylindrical image includes: Extract the pixel array region of the overlapping arc segment from each two-dimensional grayscale matrix as the overlapping edge image region; Set a candidate threshold to divide the overlapping edge image region into low grayscale class and high grayscale class, and calculate the pixel ratio and grayscale mean of the two classes. Based on the pixel ratio and grayscale mean, calculate the inter-class variance corresponding to the candidate threshold, and extract the candidate threshold that maximizes the inter-class variance as the optimal segmentation threshold. Obtain the pixel grayscale values ​​in the overlapping edge image region, and remove pixels whose pixel grayscale values ​​are less than the optimal segmentation threshold, while retaining pixels whose grayscale values ​​are equal to or greater than the optimal segmentation threshold, to obtain the corrected multi-view cylindrical image.

[0007] Furthermore, the method for acquiring the flattened wavy panoramic image includes: Obtain the cylindrical projection coordinate transformation matrix, expand the pixel position of the pixel in the two-dimensional grayscale matrix in the corrected multi-view cylindrical image to homogeneous coordinates, multiply it by the cylindrical projection coordinate transformation matrix and normalize it to obtain the circumferential coordinates and axial coordinates of the pixel. The corrected multi-view cylindrical image after coordinate projection transformation is stitched together end to end along the circumferential coordinate direction. For positions where the circumferential and axial coordinates are non-integer values, four adjacent integer coordinate pixels in a preset regular pixel grid are extracted. The distance ratio of each pixel is calculated as a weight, and the pixel gray values ​​of the four integer coordinate pixels are weighted by bilinear interpolation and summed to obtain the resampled gray values. After traversing and assigning values, a flattened wavy panoramic image is obtained.

[0008] Furthermore, the method for obtaining the axial brightness projection sequence includes: Traverse each column of the flattened wavy panoramic image, sum the integer pixel grayscale values ​​of all rows contained in the current column, and divide by the total number of rows to obtain the average axial brightness projection value corresponding to that column. Arrange the axial brightness projection mean values ​​of all columns in ascending order to generate an axial brightness projection sequence.

[0009] Furthermore, the method for obtaining the time axis distortion offset sequence includes: Obtain the standard ripple sequence, define each array element in the standard ripple sequence as a standard sampling point, and define each array element in the axial brightness projection sequence as a projection sampling point; Construct a distance matrix where the values ​​of each element are equal to the absolute values ​​of the algebraic differences between the mean axial brightness projection of the projected sampling points and the values ​​of the standard sampling points. Construct a cumulative distance matrix. The elements in the cumulative distance matrix are recursively calculated by adding the minimum value of the distance matrix element to the left, above, and upper left of the current position. Based on the cumulative distance matrix, obtain a regular path. Traverse all projected sampling points and search for standard sampling points that match the current projected sampling point on the regular path. When there are multiple matches, take the standard sampling point with the largest sequence index. Subtract the sequence index of the current projected sampling point from the largest sequence index and use the difference as the time axis distortion offset. After traversal, combine them into a time axis distortion offset sequence.

[0010] Furthermore, the method for constructing the two-dimensional spatial phase drift field includes: The time axis distortion offset of each projected sampling point in the time axis distortion offset sequence is assigned to all pixels in the corresponding column of the flattened wavy panoramic image, as the column offset of that column. For column positions in the column index that are not directly covered by regularized paths, find the column positions and their column offsets that are directly covered by regularized paths immediately to the left and right of the column, and complete the column offsets of the column using linear interpolation. Arrange the column offsets of all pixels in a matrix according to their row and column index positions to form a two-dimensional spatial phase drift field.

[0011] Furthermore, the method for acquiring the reference image includes: Define the column index of each pixel in the flattened wavy panoramic image as the axial coordinate component, and define the combination of the row index and column index as the original pixel coordinate. Add the column offset at the same row and column index position in the two-dimensional spatial phase drift field to the axial coordinate component of the original pixel coordinates to obtain the target alignment coordinates; For each integer coordinate point in the coordinate grid formed by the target alignment coordinates, look up its corresponding mapped floating-point coordinates in the flattened wavy panoramic image, extract the four integer pixels that are immediately adjacent to the mapped floating-point coordinates, perform bilinear interpolation weighted summation with distance ratio as the weight, obtain the pixel grayscale value after resampling of the current integer coordinate point, and obtain the reference image after traversing and assigning values.

[0012] Furthermore, the method for constructing the three-dimensional reference tensor block includes: Set the frame number, extract multiple consecutive reference images equal to the frame number according to the order in which the synchronization trigger signal is generated, and define them as consecutive multi-frame reference images. Define a time frame dimension, which uses the order in which the synchronization trigger signals are generated as the sorting index; Multiple consecutive reference images are orthogonally stacked and stitched along the time frame dimension to generate a three-dimensional reference tensor block. The first order of the three-dimensional reference tensor block corresponds to the physical radial circumferential direction, the second order corresponds to the physical axial direction, and the third order corresponds to the time frame dimension.

[0013] Furthermore, the method for calculating the sparse anomaly defect tensor includes: Introduce and initialize the low-rank ripple background tensor, the sparse anomaly defect tensor, the Lagrange multiplier tensor, and the quadratic penalty factor; The calculation is performed based on the three-dimensional reference tensor block, sparse anomaly defect tensor, Lagrange multiplier tensor and quadratic penalty factor. The calculation results are expanded along the three dimensions, the singular value soft thresholding shrinkage operation is performed, and after folding and restoring, the arithmetic mean is taken to update the low-rank ripple background tensor. The calculation is performed based on the three-dimensional reference tensor block, the low-rank ripple background tensor, the Lagrange multiplier tensor and the quadratic penalty factor. The calculation results are then subjected to element-wise soft thresholding shrinkage to update the sparse anomaly defect tensor. Calculate the residual tensor between the 3D reference tensor block and the low-rank ripple background tensor and the sparse anomaly defect tensor, and update the Lagrange multiplier tensor by combining the quadratic penalty factor and the residual tensor. The process is repeated until the residual tensor norm is less than the preset convergence threshold, and then the updated sparse anomaly defect tensor is output.

[0014] Furthermore, the method for generating the real-time interception signal includes: Extract the union of the non-zero tensor elements of the sparse anomaly defect tensor to obtain the two-dimensional defect response plane. Use the two-dimensional defect response plane as the root node to perform recursive space partitioning and generate a quadtree data structure. Delete the lowest-level node in the quadtree data structure whose number of non-zero pixels is less than the preset noise threshold and rebuild the parent-child index to obtain the defect feature quadtree; Merge adjacent bottom-level nodes in the defect feature quadtree into the same connected component, and summarize to obtain the target defect connected component set. By using the inverse of the cylindrical projection coordinate transformation matrix, the non-zero pixels of each connected domain in the target defect connected domain set are reverse-mapped to obtain the axial physical coordinates and the circumferential physical coordinates. The difference between the maximum and minimum values ​​of the axial physical coordinates and the circumferential physical coordinates of non-zero pixels in each connected domain is calculated to form the physical size of the defect. If the physical size of the defect is greater than the preset tolerance, a real-time interception signal is generated.

[0015] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: This invention solves the problem that the closed circular surface of the outer conductor of a radio frequency coaxial cable cannot be completely imaged with low distortion from a single viewpoint by using original multi-view cylindrical images and flattened panoramic images of corrugations, thus eliminating spatial redundancy and grayscale jumps caused by multi-view stitching. Secondly, after elastic alignment in a one-dimensional reduced space, a two-dimensional spatial phase drift field is smoothly generated, eliminating non-rigid local distortions with extremely low computational overhead and avoiding the offset step caused by direct assignment. Furthermore, the drift field is used to correct the flattened image and stitch it into a three-dimensional reference tensor block, solving the technical problems of trough shadow leakage in dynamic distortion and the insufficient information in a single frame of two-dimensional image to separate the structural corrugated background from the real defect features. This achieves accurate separation of the corrugated background from abnormal defects without relying on any prior defect template. Finally, the abnormal defect tensor is mapped to a quadtree structure and connected component restoration is performed, solving the problem of massive invalid traversal in large defect-free areas, and achieving a precise closed-loop conversion from abstract pixel-level defect response to actual production line physical space engineering interception actions. The overall solution is based on coordinate projection transformation in image space, pixel-level resampling, and spatiotemporal feature dimension decomposition, which greatly improves the accuracy of image data processing and visual analysis calculations and enhances its engineering application value. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are 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.

[0017] Figure 1 This is a flowchart of the online detection method for surface defects on the corrugated outer conductor of a radio frequency coaxial cable provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of a planar arrangement of a field-array camera provided in an embodiment of the present invention; Figure 3This is a schematic diagram of overlapping edge pixel removal provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the distance matrix and normalized path provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of a stacked three-dimensional reference tensor block provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the defect response plane quadtree partitioning provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of connected component merging provided in an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1 Please see Figure 1 As shown, this embodiment provides an online detection method for surface defects on the corrugated outer conductor of a radio frequency coaxial cable, including: Step S10: Obtain the original multi-view cylindrical image, extract and remove overlapping edge pixels from the original multi-view cylindrical image to obtain the corrected multi-view cylindrical image, and perform pixel-level coordinate projection transformation on the corrected multi-view cylindrical image to obtain the flattened wavy panoramic image.

[0020] Further, step S10 includes: Step S11: Simultaneously acquire original multi-view cylindrical images of the outer corrugated conductor of the radio frequency coaxial cable using a ring-shaped array of area cameras.

[0021] In the continuous extrusion production process of radio frequency (RF) coaxial cables, the outer layer is typically formed by continuously extruding and rolling metal strips into an axially periodically undulating annular corrugated outer conductor. This annular corrugated structure exhibits alternating peaks and troughs along the axial direction. When in the detection field of view, the peaks form structurally bright reflections under external light source illumination, while the troughs form structural optical shadows. However, in the high-speed routing scenario of continuous cable extrusion, tension fluctuations during cable winding and mechanical traction vibrations easily cause high-frequency axial micro-stretching and radial offset of the physical cable in the detection field of view, resulting in nonlinear spatial phase drift of the corrugated structure mapped in the two-dimensional image. To accurately capture the global surface features of the RF coaxial cable's corrugated outer conductor and eliminate optical shadow noise leakage introduced by the periodic phase drift of the corrugations, a ring-distributed multi-view synchronous imaging framework is established. The aim is to transform passively distorted local single-view images into original multi-view cylindrical images of the main body with global geometric references.

[0022] Specifically, a ring-shaped array of area array cameras is deployed at the inspection station between the downstream of the traction section and the upstream of the take-up section in the continuous cable extrusion production line to simultaneously acquire original multi-view cylindrical images of the corrugated outer conductor of the RF coaxial cable. The ring-shaped area array camera refers to an imaging device composed of multiple area array cameras evenly arranged at equal central angle intervals within the same inspection plane perpendicular to the RF coaxial cable axis, centered on the RF coaxial cable axis. This device is used to achieve seamless image capture of the circumferential surface of the corrugated outer conductor of the RF coaxial cable. Specifically, in underground mines and cable workshops with high humidity and dust, a single camera cannot cover the entire circumferential surface of the outer corrugated conductor of the RF coaxial cable in a single exposure. Furthermore, the circumferential surface suffers from severe perspective compression at oblique angles. For example, six area array cameras are arranged in a ring at equal central angles of 60 degrees. The optical axis of each area array camera points towards the axis of the RF coaxial cable. The angle of the circumferential arc segment covered by a single area array camera is greater than 60 degrees, ensuring overlapping arc segments between adjacent area array cameras and guaranteeing complete coverage of the entire circumferential surface. The ring-distributed area array camera array simultaneously completes exposure and image acquisition based on a synchronous trigger signal provided by a rotary encoder mounted on a traction wheel. This signal forces a uniform shutter opening time across multiple area array cameras within a microsecond timescale, severing the circumferential pixel spatial sampling misalignment link caused by the underlying hardware response delay of different area array cameras. The rotary encoder outputs a pulse as a synchronization trigger signal whenever the RF coaxial cable advances axially by a preset axial step. The preset axial step is determined based on the axial field of view width of the area array camera array. For example, the axial step is set to half the axial field of view width of the area array camera array to ensure that adjacent exposures are seamlessly connected axially. The ring-shaped area array camera array simultaneously completes exposure and acquisition when it receives the same pulse, and the output image is the original multi-view cylindrical image of the RF coaxial cable's corrugated outer conductor at the same axial position. The data structure of the original multi-view cylindrical image is a two-dimensional grayscale matrix corresponding to the number of area array cameras. The row index of each two-dimensional grayscale matrix corresponds to the pixel position in the physical radial circumferential direction, and the column index corresponds to the pixel position in the physical axial direction. The matrix elements take integer pixel grayscale values ​​ranging from 0 to 255. See also... Figure 2 This is a schematic diagram of an area array camera arrangement provided by an embodiment of the present invention. A circle is drawn at the center to represent the corrugated outer conductor of an RF coaxial cable, and the center of the circle is marked with the axis of the RF coaxial cable. For example, six trapezoidal icons representing area array cameras are evenly placed around the perimeter of the circle at 60-degree central angle intervals. Each area array camera extends two dashed lines towards the center of the circle to form a fan-shaped field of view. The six fields of view form an arc segment covering the entire circumference of the circle, and the arcs of adjacent fields of view overlapping are called overlapping arc segments.

[0023] Step S12: Extract the overlapping edge pixels of the original multi-view cylindrical image, and use the Otsu's method to remove the overlapping edge pixels to obtain the corrected multi-view cylindrical image.

[0024] After acquiring the original multi-view cylindrical images, overlapping arcs exist between the circular arcs covered by adjacent area array cameras. This overlapping region is repeatedly sampled in the two-dimensional grayscale matrices of multiple cameras. If spatial stitching is performed directly, these redundant pixels, which suffer from severe viewing angle distortion, will form abrupt grayscale changes at the stitching point, which may be mistakenly identified as abnormally high-frequency features. To eliminate the spatial redundancy and grayscale jumps caused by multi-view stitching, overlapping edge pixels in the original multi-view cylindrical images are extracted, and the Otsu's method is used to accurately remove these overlapping edge pixels.

[0025] Specifically, the pixel array regions corresponding to the overlapping arc segments in each two-dimensional grayscale matrix are extracted and defined as overlapping edge image regions; the overlapping edge pixels are the pixels located within the overlapping edge image regions. For the overlapping edge image regions of each two-dimensional grayscale matrix, the frequency of occurrence of each integer pixel grayscale value is counted, and the frequency is divided by the total number of pixels in the overlapping edge image regions to obtain the probability of occurrence of each integer pixel grayscale value; a grayscale histogram characterizing the brightness distribution features of the overlapping edge image regions is constructed with the integer pixel grayscale values ​​on the x-axis and the probability of occurrence on the y-axis. A candidate threshold is set. Since the matrix elements of the original multi-view cylindrical image range from 0 to 255 integer pixel grayscale values, the optimization space for the candidate threshold is set to the set of integers from 0 to 255. Following an increasing order from 0 to 255, each integer is sequentially selected as a candidate threshold for the current iteration. For each candidate threshold, the overlapping edge pixels in the overlapping edge image region are divided into a low-grayscale class (pixel grayscale values ​​less than the current candidate threshold) and a high-grayscale class (pixel grayscale values ​​greater than or equal to the current candidate threshold). The pixel proportion and mean grayscale value of the low-grayscale class, and the pixel proportion and mean grayscale value of the high-grayscale class are calculated. The pixel proportion of the low-grayscale class is the ratio of the number of low-grayscale pixels in the overlapping edge image region to the total number of pixels. The mean grayscale value of the low-grayscale class is the ratio of the sum of the grayscale values ​​of all low-grayscale pixels in the overlapping edge image region to the number of low-grayscale pixels. The pixel proportion of the high-grayscale class is the ratio of the number of high-grayscale pixels in the overlapping edge image region to the total number of pixels. The mean grayscale value of the high-grayscale class is the ratio of the sum of the grayscale values ​​of all high-grayscale pixels in the overlapping edge image region to the number of high-grayscale pixels. Based on the pixel proportion and grayscale mean, the inter-class variance corresponding to the candidate threshold being traversed is calculated. This inter-class variance rigorously quantifies the separability between high-grayscale frontal imaging pixel groups and low-grayscale oblique imaging pixel groups. A larger inter-class variance indicates a more significant difference in optical energy between the two types of pixels due to different viewing angles, and a clearer boundary. The calculation method is as follows: the inter-class variance equals the pixel proportion of the low-grayscale class multiplied by the pixel proportion of the high-grayscale class, and then multiplied by the square of the difference between the grayscale mean of the low-grayscale class and the grayscale mean of the high-grayscale class. This calculation method is based on the principles of variance analysis in statistics and the optical attenuation law of Lambert's cosine law. At the physical optics level, when the corrugated outer conductor surface of the same RF coaxial cable is imaged at a near-normal angle in adjacent area array cameras, the light incident direction is close to the surface normal, resulting in a large amount of reflected light entering the camera, exhibiting a high grayscale class. However, when imaged at a larger oblique angle, the light incident direction deviates from the surface normal, and the light flux is significantly attenuated according to Lambert's cosine law, exhibiting a low grayscale class. After traversing all candidate thresholds in the range of 0 to 255, the candidate threshold corresponding to the maximum inter-class variance is extracted and determined as the optimal segmentation threshold.In the overlapping edge image region, pixels with grayscale values ​​less than the optimal segmentation threshold are identified as redundant pixels with severe perspective distortion and optical attenuation and are removed. Pixels with grayscale values ​​equal to or greater than the optimal segmentation threshold are retained. The resulting image set, after setting the redundant pixels in the overlapping edge image region of each two-dimensional grayscale matrix to empty, is defined as the corrected multi-view cylindrical image. See also. Figure 3 This is a schematic diagram of overlapping edge pixel removal provided in an embodiment of the present invention. The main body of the image is a pixel grid matrix representing the overlapping edge image area. The grid is divided into two parts by a wavy dividing line: one part of the grid is filled with dark gray, representing low grayscale classes with perspective distortion and optical attenuation, which are redundant pixels to be removed; the other part of the grid is filled with light gray, representing high grayscale classes, which are retained.

[0026] Step S13: Perform pixel-level coordinate projection transformation on the corrected multi-view cylindrical image to obtain a flattened wavy panoramic image.

[0027] After removing overlapping edge pixels and obtaining the corrected multi-view cylindrical image, the images output by each stage array camera still exist in their own independent image plane coordinate system. Due to oblique imaging, the circumferential surface suffers from perspective compression, and the same physical wave period exhibits unequal axial distribution in images from different viewpoints, making it impossible to directly extract a regular periodic brightness pattern along the axis. Therefore, the image plane coordinate system of each stage array camera is mapped to the same two-dimensional plane coordinate system based on the unfolded cylindrical surface. This corrects the curved cylindrical surface, which was originally perspective compressed due to oblique imaging, into a proportionally unfolded planar image, ensuring that the peaks and troughs of the same period exhibit consistent physical spacing along the axis. This lays the geometric foundation for subsequent extraction of a regular periodic brightness pattern along the axis.

[0028] The cylindrical surface projection coordinate transformation matrix is ​​set. Specifically, during the system calibration phase, a standard cylindrical calibration target with known-spacing circumferential angle scale lines and known-spacing axial length scale lines printed on its surface is placed at the testing station, ensuring that the axis of the standard cylindrical calibration target coincides with the axis of the RF coaxial cable. The circumferential angle scale lines refer to longitudinal scale lines uniformly distributed along the axial direction of the standard cylindrical calibration target, used to mark known angular positions in the circumferential direction; the axial length scale lines refer to transverse scale lines uniformly distributed along the circumferential direction of the standard cylindrical calibration target, used to mark known physical length positions in the axial direction; the homogeneous coordinates of the area array camera image plane refer to the expansion of the two-dimensional image plane coordinates of the area array camera into a ternary column vector expression form, used to support the unified description of linear transformations such as translation, rotation, and projection using matrix multiplication. For each area array camera in a ring-shaped array, using the single cylindrical calibration image corresponding to that camera in the multi-view cylindrical calibration images it acquires as input, a set of point pairs is established between the known physical coordinates of the intersection points of the circumferential angle scale lines and the axial length scale lines on the standard cylindrical calibration target surface and their corresponding pixel coordinates in the single cylindrical calibration image of that camera. The single cylindrical calibration image refers to the two-dimensional grayscale matrix independently acquired by the current area array camera in the multi-view cylindrical calibration image. Its data structure is consistent with the two-dimensional grayscale matrix output by a single area array camera in the original multi-view cylindrical image, with row indices corresponding to pixels in the physical radial circumferential direction. The position and column index correspond to the pixel position along the physical axis. The matrix elements are integer pixel grayscale values ​​ranging from 0 to 255. The only difference is that the imaging object is a standard cylindrical calibration target rather than the corrugated outer conductor of an online coaxial RF cable. Using the unfolded cylindrical surface as a reference, the circumferential direction of the unfolded cylindrical surface is defined as the row coordinate axis, and the axial direction is defined as the column coordinate axis, establishing a two-dimensional planar coordinate system. Based on the set of point pairs, the least squares method is used to solve the 3x3 projection matrix that maps the homogeneous coordinates of the image plane of the stage array camera to the two-dimensional planar coordinates of the unfolded cylindrical surface. The solved projection matrix is ​​determined as the cylindrical surface projection coordinate transformation matrix corresponding to the stage array camera. Each stage array camera corresponds to a cylindrical surface projection coordinate transformation matrix. The cylindrical surface projection coordinate transformation matrices of all stage array cameras are stored uniformly in the calibration parameter file of the detection system after calibration.

[0029] For each single-view corrected cylindrical image in the corrected multi-view cylindrical image, a pixel-level coordinate projection transformation is performed. The process is as follows: for each non-overlapping edge pixel retained in the current single-view corrected cylindrical image, its image plane coordinates are expanded to homogeneous coordinates of the area array camera image plane, then multiplied by the cylindrical projection coordinate transformation matrix of the area array camera corresponding to the single-view corrected cylindrical image, and the homogeneous coordinates of the transformation result are normalized to two-dimensional plane coordinates to obtain the circumferential and axial coordinates of the non-overlapping edge pixel on the unfolded cylindrical surface. According to the adjacency relationship of the circumferential arc segments covered by each area array camera, all transformed images are stitched together end to end along the circumferential coordinate direction. Since the circumferential and axial coordinates obtained by each non-overlapping edge pixel after the cylindrical projection coordinate transformation are usually non-integer values, that is, their landing points are located between adjacent integer coordinate pixels in the two-dimensional plane coordinate system of the unfolded cylindrical surface, rather than pixel grid points that are exactly aligned to integer coordinates, such landing point positions are defined as non-integer grid positions. To resample the grayscale values ​​at non-integer grid locations to a regular pixel grid based on integer coordinates, where the regular pixel grid is an array of pixels arranged at uniform intervals using integer row and column coordinates as indices in a two-dimensional planar coordinate system of an unfolded cylinder, for each non-integer grid location, the four nearest integer coordinate pixels in the regular pixel grid are taken. The distance ratios between the non-integer grid location and the four integer coordinate pixels in the row and column directions are calculated respectively. The grayscale values ​​of the four integer coordinate pixels are weighted and summed. The resulting weighted sum is used as the resampled grayscale value of the current integer coordinate pixel in the target regular pixel grid, thus completing the grayscale assignment at the integer coordinate point and performing bilinear interpolation resampling. Specifically, assuming that the row and column coordinates of the non-integer grid location are both non-integer values, the floor value of the row coordinate and the floor value of the column coordinate of the non-integer grid location are taken, and the four nearest integer coordinate pixels are defined as the top-left, top-right, bottom-left, and bottom-right points, respectively. The top-left point is located at the integer coordinate determined by the rounded-down values ​​of the row and column coordinates. The top-right point is located at the integer coordinate determined by the rounded-down values ​​of the row and column coordinates plus one. The bottom-left point is located at the integer coordinate determined by the rounded-down values ​​of the row and column coordinates plus one. The bottom-right point is located at the integer coordinate determined by the rounded-down values ​​of the row and column coordinates plus one. The row-direction decimal offset is defined as the decimal part of the row coordinate of the non-integer grid position minus its rounded-down value. The column-direction decimal offset is defined as the decimal part of the column coordinate of the non-integer grid position minus its rounded-down value. The values ​​of both the row-direction and column-direction decimal offsets are greater than or equal to zero and less than one.Based on the decimal offsets in the row and column directions, the bilinear interpolation weights for each of the four integer coordinate pixels are calculated: the weight of the top-left pixel is one minus the decimal offset in the row direction multiplied by one minus the decimal offset in the column direction; the weight of the top-right pixel is one minus the decimal offset in the row direction multiplied by the decimal offset in the column direction; the weight of the bottom-left pixel is the decimal offset in the row direction multiplied by one minus the decimal offset in the column direction; and the weight of the bottom-right pixel is the decimal offset in the row direction multiplied by the decimal offset in the column direction. The grayscale value of each of the four integer coordinate pixels is multiplied by its corresponding weight and then summed to obtain the resampled grayscale value for that non-integer grid position. The weights indicate that the closer the non-integer grid position is to the integer coordinate pixel in the row and column directions, the greater the contribution of the integer coordinate pixel to the resampled grayscale value, thus maintaining the local smoothness of the image during continuous grayscale transitions. After traversing all non-integer grid positions, a flattened wavy panoramic image is obtained. The data structure of the flattened wavy panoramic image is a two-dimensional grayscale matrix. The row index corresponds to the unfolded position in the physical radial circumferential direction, and the column index corresponds to the unfolded position in the physical axial direction. The matrix elements take integer pixel grayscale values ​​from 0 to 255.

[0030] Step S10 addresses the issue of the inability to completely and with low distortion imagery of the closed circumferential surface of the corrugated outer conductor of an RF coaxial cable from a single viewpoint by using the original multi-view cylindrical image, the corrected multi-view cylindrical image, and the flattened corrugated panoramic image. Specifically, the original multi-view cylindrical image ensures temporal consistency of images from different views along the circumference at the same axial position, severing the spatial sampling misalignment link in the circumferential direction caused by hardware response delays. The corrected multi-view cylindrical image eliminates spatial redundancy and grayscale jumps caused by multi-view stitching, preventing redundant low-quality pixels from forming grayscale abrupt bands at subsequent stitching points and being mistaken for abnormal high-frequency features. The flattened corrugated panoramic image corrects the curved cylindrical surface, which was originally compressed by perspective due to oblique imaging, into a proportionally scaled planar unfolded image, ensuring that the peaks and troughs of the same period exhibit consistent physical spacing along the axial direction, laying the geometric foundation for subsequent extraction of regular periodic brightness patterns along the axial direction.

[0031] Step S20: Calculate the mean axial brightness projection based on the flattened wavy panoramic image and combine them into an axial brightness projection sequence. Calculate the time axis distortion offset using the axial brightness projection sequence and combine them into a time axis distortion offset sequence. Calculate the column offset of the flattened wavy panoramic image based on the time axis distortion offset sequence. Arrange the column offsets into a matrix to form a two-dimensional spatial phase drift field.

[0032] Further, step S20 includes: Step S21: Calculate the mean axial brightness projection based on the flattened wavy panoramic image and combine them into an axial brightness projection sequence.

[0033] After acquiring the flattened panoramic image of the ripples, since the image has losslessly unfolded the circumferential closed surface into a plane and the peaks and troughs of the same period exhibit a consistent physical spacing along the axial direction, directly calculating the dense optical flow field in the two-dimensional image space or using a spatial transformation network for feature alignment is computationally expensive. This not only fails to meet the millisecond-level real-time requirements of high-speed routing scenarios but also easily gets trapped in local optima caused by the periodic shadows of the ripples, mismatching the trough shadows of normal ripples as trough shadows of adjacent periods. To achieve robust periodic phase matching with extremely low computational cost, the two-dimensional flattened panoramic image of the ripples is averaged along the circumferential direction to reduce its dimensionality into a one-dimensional sequence. This cancels out random illumination unevenness and local noise along the circumferential direction while preserving the essential characteristic of the periodic alternation of ripple peak highlights and trough shadows along the axial direction.

[0034] Specifically, the column index of the flattened wavy panoramic image corresponds to the unfolded position along the physical axis. Pixels in the same column have the same physical axial position coordinates but correspond to different unfolded positions along the physical radial circumferential direction. Each column of the flattened wavy panoramic image is traversed. For the current column, the integer pixel grayscale values ​​of all rows contained therein are summed. The sum is then divided by the total number of rows in the flattened wavy panoramic image to calculate the mean axial brightness projection value corresponding to that column. The mean axial brightness projection values ​​of all columns of the flattened wavy panoramic image are arranged sequentially in ascending order of column index to generate an axial brightness projection sequence. The mean axial brightness projection value refers to the arithmetic mean of the integer pixel grayscale values ​​of all pixels at a specific physical axial position coordinate along the physical radial circumferential unfolded position of the flattened wavy panoramic image. Its physical meaning is to filter out local optical abrupt changes and random image noise in the radial direction through mean calculation, thereby uniquely quantifying the global optical reflection intensity at that physical axial position in a one-dimensional reduced-dimensional space. The data structure of the axial brightness projection sequence is a one-dimensional real number array with a length equal to the total number of columns in the flattened panoramic image of the corrugated surface. The array elements range from 0 to 255. Since the peaks of the outer conductor of the RF coaxial cable form bright reflections under external light and the troughs form optical shadows, the axial brightness projection sequence exhibits a stable periodic oscillating waveform along the physical axis. The peaks correspond to the high values ​​of the array elements, and the troughs correspond to the low values. This axial brightness projection sequence is a highly condensed brightness characteristic representation of the axial period of the outer conductor of the RF coaxial cable.

[0035] Step S22: Obtain the standard ripple sequence, calculate the time axis distortion offset of the axial brightness projection sequence and the standard ripple sequence, and combine them into a time axis distortion offset sequence.

[0036] After obtaining the axial brightness projection sequence, it is aligned with an ideal, distortion-free reference. Due to the axial micro-stretching and radial offset of the corrugations caused by tension fluctuations in the take-up and release lines and mechanical traction vibrations, the axial brightness projection sequence will experience nonlinear local stretching and spatial phase drift along the physical axis. If a rigid linear matching or cross-correlation algorithm is used directly, it cannot adapt to this non-rigid local distortion, which can easily lead to phase misalignment between peaks and troughs. In order to accurately establish a periodic phase match between peaks and between troughs even when the sequence undergoes nonlinear local deformation, a dynamic time warping algorithm is introduced to perform low-computational-efficiency and highly robust elastic alignment in the one-dimensional sequence space.

[0037] Specifically, a standard corrugation sequence is obtained. This standard corrugation sequence refers to the axial brightness projection sequence obtained during the system calibration phase. A standard RF coaxial cable with a defect-free, mechanically undisturbed outer conductor and uniform corrugation period is placed at the testing station. After undergoing the same image acquisition, flattening, and axial brightness projection mean calculation processing, this sequence is saved as a reference benchmark with a strictly uniform period. The data structure of the standard corrugation sequence is a one-dimensional real number array, with array elements ranging from 0 to 255. Each array element in the standard corrugation sequence is defined as a standard sampling point, and each array element in the axial brightness projection sequence is defined as a projection sampling point.

[0038] Construct a distance matrix between the axial brightness projection sequence and the standard ripple sequence. Let N be the total number of projection sampling points in the axial brightness projection sequence and M be the total number of standard sampling points in the standard ripple sequence. Construct an N x M distance matrix. Let i be the row index of the distance matrix, ranging from 1 to N. Row index i also serves as the sequence index of the axial brightness projection sequence. Let j be the column index of the distance matrix, ranging from 1 to M. Column index j also serves as the sequence index of the standard ripple sequence. The element in the i-th row and j-th column of the distance matrix is ​​equal to the absolute value of the algebraic difference between the mean axial brightness projection value of the i-th projection sampling point in the axial brightness projection sequence and the value of the j-th standard sampling point in the standard ripple sequence. The distance matrix is ​​an N x M non-negative real number matrix. The elements in the distance matrix reflect the degree of phase mismatch between the axial brightness projection sequence and the standard ripple sequence at corresponding sampling point positions based on local grayscale differences.

[0039] Find the regularized path with the minimum cumulative distance based on the distance matrix. The regularized path is a monotonically progressive path that starts from the element in the 1st row and 1st column of the distance matrix and terminates at the element in the Nth row and Mth column, where each step can only move to the right, down, or to the right-lower adjacent element. An N x M cumulative distance matrix is ​​constructed using dynamic programming, where the row and column indices are i and j, respectively. The element in the i-th row and j-th column of the cumulative distance matrix is ​​recursively calculated using the following formula: ,in, This represents the function that takes the minimum value. Indicates from Take the minimum value from the middle; A and B represent the cumulative distance matrix and the distance matrix, respectively. The element in the i-th row and j-th column of the cumulative distance matrix. Let be the element in the i-th row and j-th column of the distance matrix, and let the recursive boundary be . After the recursion is completed, starting from the element in the Nth row and Mth column of the cumulative distance matrix, backtracking step by step along the predecessor element that minimizes the cumulative distance to the element in the 1st row and 1st column of the cumulative distance matrix. The sequence of elements traversed during backtracking is the normalized path. Physically, the normalized path corresponds to the optimal periodic phase matching method between the axial brightness projection sequence and the standard ripple sequence, with peaks matching peaks and troughs matching troughs.

[0040] The time axis distortion offset of the axial brightness projection sequence is extracted based on the regularized path. For the i-th projection sampling point of the axial brightness projection sequence, a standard sampling point in the standard ripple sequence that matches the i-th projection sampling point is searched on the regularized path. When multiple matching standard sampling points exist, the standard sampling point with the largest sequence index j is selected, and the difference between the largest sequence index j and the sequence index i of the projection sampling point is taken as the time axis distortion offset of the i-th projection sampling point. After traversing all projection sampling points of the axial brightness projection sequence, the time axis distortion offset sequence is obtained. The data structure of the time axis distortion offset sequence is a one-dimensional integer array with a length equal to the length of the axial brightness projection sequence. Each array element in the time axis distortion offset sequence is the time axis distortion offset of the corresponding projection sampling point. When the time axis distortion offset is positive, it indicates that the periodic phase of the corresponding projected sampling point relative to the standard ripple sequence is axially compressed and needs to be shifted backward for alignment. The standard phase direction is defined as the axial physical unfolding direction based on the sequence index of the standard sampling point in the standard ripple sequence. When the time axis distortion offset is negative, it indicates that the periodic phase of the corresponding projected sampling point relative to the standard ripple sequence is axially stretched and needs to be shifted forward for alignment. See also Figure 4 This diagram illustrates the distance matrix and regularized path provided in this embodiment of the invention. A two-dimensional grid matrix represents the distance matrix. Both the row and column directions of the matrix are arrays of squares. Within the matrix, a path is highlighted with dark blocks, starting from the upper left corner and extending continuously to the lower right corner. Each step on this path involves moving only to the adjacent squares in the right, down, or lower right direction. The entire continuous highlighted path represents a regularized path.

[0041] Step S23: Calculate the column offset of the flattened wavy panoramic image based on the time axis distortion offset sequence, and arrange the column offsets into a matrix to form a two-dimensional spatial phase drift field.

[0042] After obtaining the time axis distortion offset sequence, this sequence only provides the offset of each projected sampling point column by column in the physical axis dimension. However, the flattened wavy panoramic image is a two-dimensional grayscale matrix, requiring precise axial offset for each pixel before subsequent pixel-by-pixel resampling correction can be performed. Therefore, the discrete time axis distortion offset sequence is smoothly expanded into a continuous two-dimensional deformation field covering the entire flattened wavy panoramic image, thus providing a continuously valued axial offset for each pixel. Finally, spatial alignment of the wavy period is achieved through coordinate offset correction.

[0043] Specifically, since the axial brightness projection sequence is obtained by averaging the values ​​of each column of the flattened wave-like panoramic image, the total number of projection sampling points in the axial brightness projection sequence is numerically the same as the total number of columns in the flattened wave-like panoramic image. Therefore, the column index of the flattened wave-like panoramic image is also represented by i, with a value range of 1 to N. The total number of rows in the flattened wave-like panoramic image is defined as H, and the row index ranges from 1 to H. The row index is represented by c. The time axis distortion offset of the i-th projection sampling point in the time axis distortion offset sequence is assigned to all pixels in the i-th column of the flattened wave-like panoramic image with an upper row index of 1 to H as the column offset of the i-th column. The column offset of each row in the same column is kept consistent along the physical radial dimension, i.e., the row direction. Along the physical axial dimension, i.e., the column direction, for column positions in column indices 1 to N that are not directly covered by the regularized path, the column offset of the column position is calculated and completed by linear interpolation based on the distance ratio between the column position and the column positions directly covered by the adjacent regularized paths on both sides. Specifically, for column positions in the column index that are not directly covered by a normalized path, the column position directly covered by a normalized path immediately to the left of that column position is found in the time axis warp offset sequence, and its column index is defined as q. The column position directly covered by a normalized path immediately to the right of that column position is found, and its column index is defined as r. The column index of the currently uncovered column position is defined as s, and the value of s satisfies the following conditions: Furthermore, q and r are both column indices from 1 to N that are directly covered by the regularized path. The time axis distortion offset corresponding to column q is defined as the left offset, and the time axis distortion offset corresponding to column r is defined as the right offset. Then, the linear interpolation column offset of column s is equal to the left offset multiplied by the difference between the column indices of column r and column s divided by the difference between the column indices of column r and column q, plus the right offset multiplied by the difference between the column indices of column s and column q divided by the difference between the column indices of column r and column q, thereby achieving a linear transition of offset between adjacent regularized path covered columns.

[0044] After calculating the column offsets of all columns in the flattened wave pattern panoramic image, the column offset of the i-th column is uniformly assigned to all pixels with indexes 1 to H in the upper row of the i-th column along the physical radial circumferential dimension, i.e., the row direction. The assigned value to each pixel is defined as the axial offset corresponding to the pixel, thus ensuring that the axial offsets of all pixels within the same column remain consistent. After traversing all column positions from index 1 to N, each pixel in the flattened wave pattern panoramic image obtains a continuously valued axial offset. The axial offsets of all pixels are arranged in a matrix according to the row and column index positions of the pixels in the flattened wave pattern panoramic image, forming a two-dimensional spatial phase drift field. The data structure of the two-dimensional spatial phase drift field is a two-dimensional real number matrix of the same size as the flattened wave pattern panoramic image. The matrix elements of the two-dimensional spatial phase drift field are the axial offsets of the corresponding pixels, used to represent the amount of displacement required for the corresponding pixel to move along the physical axial direction to align with the periodic phase of the standard wave pattern sequence. Linear interpolation smoothly extends the periodic phase mismatch at discrete sampling points into a continuous deformation field covering the entire image, avoiding the offset step caused by direct assignment.

[0045] Step S20 solves the technical problems of huge computational cost, easy getting trapped in local optimal solutions of wavy periodic shadows, and difficulty in accurately matching the nonlinear spatial phase drift of wavy caused by mechanical traction vibration when directly calculating dense optical flow fields in two-dimensional image space or using spatial transformation networks for feature alignment, by flattening the wavy panoramic image and two-dimensional spatial phase drift field. In particular, the flattened wavy panoramic image is smoothly expanded into a two-dimensional spatial phase drift field after elastic alignment in one-dimensional reduced space, which eliminates non-rigid local distortion with extremely low computational cost and avoids the step shift of offset caused by direct assignment.

[0046] Step S30: Use the two-dimensional spatial phase drift field to offset and correct the flattened wavy panoramic image to obtain a reference image, stitch the reference image to obtain a three-dimensional reference tensor block, and decompose the three-dimensional reference tensor block to obtain a sparse anomaly defect tensor.

[0047] Further, step S30 includes: Step S31: The flattened wavy panoramic image is offset-corrected and resampled using a two-dimensional spatial phase drift field to obtain a reference image.

[0048] After obtaining the two-dimensional spatial phase drift field, although the axial offset required to align each pixel in the flattened corrugated panoramic image to the periodic phase of the standard corrugation sequence has been determined, the flattened corrugated panoramic image itself remains unchanged, retaining the dynamic distortion generated in the high-speed cable routing scenario. If tensor decomposition or background stripping is directly performed on the flattened corrugated panoramic image containing nonlinear spatial phase drift, the trough shadows of the dynamic distortion will not be aligned with the static background, thus leaking as high-frequency features into the abnormal defect space. To eliminate this spatial misalignment, the coordinate offset correction and resampling of the flattened corrugated panoramic image are performed using the two-dimensional spatial phase drift field, forcing the dynamic distortion corrugated structure to be deformed in the reverse direction to compensate for the standard periodic phase, thereby generating a reference image with strict spatial periodic alignment properties.

[0049] Specifically, the column index of each pixel in the flattened wavy panoramic image is defined as the axial coordinate component, and the combination of the row and column indices of each pixel in the flattened wavy panoramic image is defined as the original pixel coordinate. For each original pixel coordinate in the flattened wavy panoramic image, the axial coordinate component of the original pixel coordinate is added to the axial offset at the same row and column index position in the two-dimensional spatial phase drift field to obtain the corrected target pixel coordinate, which is defined as the target alignment coordinate. This establishes a mapping relationship between the target alignment coordinate and the original pixel coordinate. Since the axial offset in the two-dimensional spatial phase drift field is a continuous real number, and the axial coordinate component of the target alignment coordinate usually contains decimals, it cannot be directly used as a discrete index of the digital image for image resampling processing. For each integer coordinate point in the coordinate grid formed by the target alignment coordinates, its corresponding coordinate point with a decimal coordinate value in the flattened wavy panoramic image is looked up according to the mapping relationship, and defined as the mapped floating-point coordinate. The four nearest integer pixels surrounding this mapped floating-point coordinate are extracted. The distance ratios between this mapped floating-point coordinate and the four nearest integer pixels in the physical radial circumferential direction and the physical axial direction are calculated respectively. Using the distance ratios as interpolation weights, the integer pixel grayscale values ​​of the four nearest integer pixels are weighted and summed using bilinear interpolation to obtain the resampled pixel grayscale value of the current integer coordinate point. After traversing all integer coordinate points and assigning pixel grayscale values, the reference image is obtained. The data structure of the reference image is a two-dimensional grayscale matrix of the same size as the flattened wavy panoramic image, and the matrix elements of the reference image take integer pixel grayscale values ​​ranging from 0 to 255. After coordinate offset correction and image resampling, the structural bright reflections and structural optical shadows in the reference image are realigned to the periodic phase specified by the standard ripple sequence, so that the nonlinear geometric distortion error caused by mechanical stretching and vibration is eliminated in the ripple structure of the reference image, achieving pixel-by-pixel spatial tight alignment.

[0050] Step S32: Perform continuous multi-frame stitching on the reference image to generate a three-dimensional reference tensor block.

[0051] After acquiring the reference image, the nonlinear geometric distortion errors caused by mechanical stretching and vibration have been eliminated from the corrugated structure, achieving pixel-by-pixel spatial alignment. However, relying solely on the two-dimensional pixel information of a single frame of the reference image, it is difficult to separate the corrugated background, such as structural high-brightness reflections and structural optical shadows, from the real defect features such as micro-cracks and indentations on the outer conductor surface of the RF coaxial cable. Since the RF coaxial cable is continuously extruded and routed at high speed, if multiple reference images acquired under consecutive synchronous trigger signals and processed with the same offset correction are superimposed along the time axis, the corrugated background will exhibit extremely high temporal repeatability and stability at the same original pixel coordinate position, while defects such as micro-cracks and indentations will appear as irregular transient changes on the time axis due to the randomness of the defect spatial distribution. By stitching together multiple consecutive frames of reference images in chronological order, a three-dimensional tensor containing the time frame dimension can be constructed.

[0052] Specifically, the number of frames is set, determined based on the production line speed and the physical coverage length of a single frame reference image along the physical axis. The principle for setting the number of frames is to ensure that the total physical length accumulated by multiple consecutive reference images along the physical axis reaches at least twenty times the period length of a single corrugation in the RF coaxial cable's ring corrugated structure, thus guaranteeing sufficient corrugated background samples that satisfy statistical stability in the time dimension. For example, it is set to 60 frames. Multiple reference images are continuously output at the times corresponding to the generation of multiple consecutive synchronous trigger signals. Multiple reference images, equal to the number of frames, are extracted according to the order in which the synchronous trigger signals corresponding to the reference images are generated, and defined as consecutive multi-frame reference images. The time frame dimension is defined, which refers to the data dimension used to characterize the evolution of pixel grayscale values ​​at different sampling times, with the order in which the synchronous trigger signals corresponding to the reference images are generated as an index. The extracted consecutive multi-frame reference images are sequentially orthogonally stacked and stitched along the time frame dimension to generate a three-dimensional reference tensor block. The data structure of the three-dimensional reference tensor block is a third-order real tensor. The first order of the three-dimensional reference tensor block corresponds to the physical radial circumferential direction of the reference image, the second order corresponds to the physical axial direction of the reference image, and the third order corresponds to the time frame dimension. The tensor element values ​​of the three-dimensional reference tensor block are consistent with the matrix elements of the reference image, being integer pixel grayscale values ​​from 0 to 255. Due to the pre-processing of coordinate offset correction and image resampling, the three-dimensional reference tensor block corresponds to a wave structure with absolutely consistent periodic phase at the same spatial row and column coordinate positions in the time frame dimension. Therefore, normal structural highlight reflections and structural optical shadows are highly repetitive in the time frame dimension, possessing a very strong low-rank property in the mathematical tensor space. See also Figure 5This is a schematic diagram of a stacked three-dimensional reference tensor block provided in an embodiment of the present invention. Multiple identical rectangular planes represent reference images. These rectangular planes are neatly arranged along the same direction and stacked layer by layer to form a cuboid block, representing a three-dimensional reference tensor block. Coordinate axes are drawn on the three-dimensional edges of the cuboid, corresponding to the physical radial circumferential direction, the physical axial direction, and the time frame dimension, respectively.

[0053] Step S33: Establish an optimization objective based on the three-dimensional reference tensor block and use the alternating direction multiplier method for iterative solution to obtain the sparse anomaly defect tensor.

[0054] After acquiring the 3D reference tensor block, due to the prior coordinate offset correction and image resampling processing, the same spatial row and column coordinate positions in the 3D reference tensor block in the time frame dimension correspond to a wave structure with absolutely consistent periodic phase. Normal structural bright reflections and structural optical shadows are highly repetitive in the time frame dimension, possessing a strong low-rank property in the mathematical tensor space. However, real defects such as micro-cracks and indentations, due to their random and localized spatial distribution on the surface of the corrugated outer conductor of the RF coaxial cable, exhibit irregular transient changes in the time frame dimension of the 3D reference tensor block, possessing a sparse property in the mathematical tensor space. Based on the difference between the low-rank and sparse properties mentioned above, the two are separated through low-rank sparse decomposition. The 3D reference tensor block is decomposed into the sum of a low-rank wave background tensor and a sparse anomalous defect tensor. The low-rank wave background tensor absorbs the highly repetitive wave background components such as structural bright reflections and structural optical shadows in the time frame dimension, while the sparse anomalous defect tensor retains the gray-level deviation response corresponding to the real defects.

[0055] Specifically, we establish the optimization objective for low-rank sparse decomposition. Let the three-dimensional baseline tensor block be denoted as tensor T, the low-rank ripple background tensor to be solved be denoted as tensor L, and the sparse anomaly defect tensor to be solved be denoted as tensor D. Under the constraint that the sum of tensor L and tensor D equals tensor T, the optimization objective is to minimize the weighted sum of the tensor nuclear norm of tensor L and the first norm of tensor D. The optimization problem is as follows: Under the constraint that tensor L plus tensor D equals tensor T, minimize the tensor nuclear norm of tensor L plus the tradeoff coefficient multiplied by the first norm of tensor D. Wherein, the tensor kernel norm of tensor L is defined as the sum of all singular values ​​of the three expansion matrices obtained by expanding tensor L along the first, second and third order dimensions respectively; the first norm of tensor D is defined as the sum of the absolute values ​​of all tensor elements of tensor D; a tradeoff coefficient λ is set, which is used to balance the relative weights of low-rank terms and sparse terms in the optimization objective, and its value is based on the reciprocal of the square root of the larger of the first and second order dimensions of the three-dimensional reference tensor block.

[0056] The above optimization problem is solved iteratively using the alternating direction multiplier method. A Lagrange multiplier tensor Y of the same size as tensor T is introduced, and all tensor elements of tensors L, D, and Y are initialized to zero. An augmented Lagrange function is constructed. A quadratic penalty factor, denoted as μ, is introduced, with its initial value being the reciprocal of the maximum absolute value of all tensor elements in the three-dimensional reference tensor block. An augmented Lagrange function is constructed based on tensors L, D, Y, and μ. This augmented Lagrange function is equal to the tensor nuclear norm of tensor L plus the tradeoff coefficient multiplied by the first norm of tensor D, plus the inner product of tensor Y and tensor T minus tensor L and then tensor D (the residual tensor), plus the quadratic penalty factor μ divided by two multiplied by the square of the norm of the residual tensor obtained by subtracting tensor L and then tensor D from tensor T. Here, the inner product of tensor Y and the residual tensor is the sum of the products of corresponding tensor elements of two tensors of the same size, used to introduce the constraint tensor L plus tensor D equal to tensor T into the objective function; the quadratic penalty factor μ divided by two and multiplied by the square of the residual tensor norm constitutes a quadratic penalty term, used to gradually increase the penalty on the constraint condition as μ increases during the iteration process, driving the sum of tensor L and tensor D to converge to tensor T. The following three sub-steps are executed alternately until the convergence condition is met. The first sub-step is to update tensor L. Fix the values ​​of tensor D and tensor Y in the current iteration round, calculate the result of tensor T minus tensor D plus tensor Y divided by μ, expand the result along the first, second, and third order dimensions to obtain three expansion matrices, and perform singular value soft thresholding shrinkage operation on each expansion matrix. The process of the singular value soft thresholding contraction operation is as follows: Perform singular value decomposition on the current expanded matrix to obtain a left singular vector matrix, a singular value diagonal matrix, and a right singular vector matrix; subtract the soft threshold from each singular value in the singular value diagonal matrix, and retain the original sign of the non-negative part to obtain the contracted singular value diagonal matrix, where the soft threshold is a weighting coefficient divided by μ; multiply the left singular vector matrix, the contracted singular value diagonal matrix, and the right singular vector matrix to reconstruct the matrix; after performing singular value soft thresholding contraction on each of the three dimensions, fold the three reconstructed matrices along their corresponding dimensions to restore them to third-order tensors of the same size as tensor T; then take the arithmetic mean of these three restored third-order tensors element-wise, and use the resulting average tensor as the tensor L updated in this iteration; the second sub-step is to update tensor D. Fix the updated values ​​of tensor L and tensor Y in the current iteration, calculate the result of tensor T minus tensor L plus tensor Y divided by μ, and perform element-wise soft thresholding contraction on each element of this result.The element-wise soft thresholding shrinkage operation is as follows: Take the absolute value of the current tensor element, subtract the soft threshold from its non-negative part, and multiply by the sign of the current tensor element value to obtain the shrunken tensor element value. The soft threshold is a weighting coefficient divided by μ. After traversing all tensor elements to complete the element-wise soft thresholding shrinkage, the result is used as the tensor D updated in this iteration. The third sub-step is to update the Lagrange multiplier tensor Y. Fix the updated values ​​of tensor L and tensor D in the current iteration, add μ to tensor Y, multiply by tensor T, subtract tensor L, and then subtract tensor D. This result is used as the tensor Y updated in this iteration.

[0057] After each of the above three sub-steps, a convergence criterion is calculated. The convergence criterion is the norm of the residual tensor obtained by subtracting tensor L and then tensor D from tensor T. The norm of the residual tensor is compared with a convergence threshold, which is the norm of tensor T multiplied by a convergence coefficient, where the convergence coefficient is 0.0000001. When the norm of the residual tensor is less than the convergence threshold, the iteration is considered converged, and the iteration terminates; otherwise, the next iteration continues. After the iteration terminates, the tensor D obtained in the current iteration is output as the sparse anomaly defect tensor. The data structure of the sparse anomaly defect tensor is a third-order real tensor of the same size as the three-dimensional reference tensor block. Most tensor elements of the sparse anomaly defect tensor are zero, and non-zero tensor elements correspond to the grayscale deviation of the defect pixel relative to the normal ripple grayscale value corresponding to the low-rank ripple background tensor in the corresponding time frame.

[0058] Step S30 addresses the technical challenges of dynamically distorted trough shadows failing to align with the static background and thus leaking as high-frequency features into the anomalous defect space, and the insufficient information from a single-frame two-dimensional image to separate structural ripple backgrounds from real defect features, through a reference image, a three-dimensional reference tensor block, and a sparse anomalous defect tensor. Specifically, the reference image eliminates nonlinear geometric distortion errors caused by mechanical stretching and vibration, achieving pixel-by-pixel spatial alignment; the three-dimensional reference tensor block ensures high repetition of normal structural highlight reflections and structural optical shadows across time frames, constructing a strong low-rank property in the mathematical tensor space, while simultaneously making the transient changes of real defects exhibit sparse properties; the sparse anomalous defect tensor incorporates all highly repetitive ripple background components across time frames into the low-rank ripple background tensor, retaining only the grayscale deviation response corresponding to real defects, thus achieving separation of the ripple background from anomalous defects without relying on any prior defect template.

[0059] Step S40: Map the sparse anomaly defect tensor to a quadtree data structure, extract and remove the bottom-level nodes from the quadtree data structure to obtain the defect feature quadtree, perform connected component merging on the defect feature quadtree to obtain the target defect connected component set, calculate the defect physical size based on the target defect connected component set, and compare and judge based on the defect physical size to generate a real-time interception signal.

[0060] Further, step S40 includes: Step S41: Divide the sparse anomaly defect tensor into multiple quadrant nodes to generate a quadtree data structure.

[0061] After obtaining the sparse anomaly defect tensor, most of its tensor elements are zero. The non-zero tensor elements correspond to the grayscale deviation of the defect pixel relative to the normal ripple grayscale value of the low-rank ripple background tensor in the corresponding time frame. If we directly traverse all the tensor elements of the sparse anomaly defect tensor pixel by pixel to extract the spatial distribution of defects, under the premise that most areas of the outer conductor surface of the RF coaxial cable are defect-free, a large number of zero-value tensor elements corresponding to defect-free areas will generate a huge amount of invalid traversal, which is difficult to meet the stringent real-time requirements of high-speed routing scenarios. In order to avoid invalid traversal in large areas of defect-free regions and achieve efficient spatial hierarchical indexing of defect pixels, we first extract the union of the non-zero tensor elements of each spatial row and column coordinate position of the sparse anomaly defect tensor along the time frame dimension to obtain a two-dimensional defect response plane that reflects the spatial distribution of all defects. The row index of the two-dimensional defect response plane corresponds to the physical radial circumferential direction of the reference image, and the column index corresponds to the physical axial direction of the reference image. Then, we introduce a quadtree data structure to recursively partition the non-zero pixels in the two-dimensional defect response plane.

[0062] Specifically, taking the overall rectangular region of the two-dimensional defect response plane as the root node, the root node is divided into four quadrant nodes of equal area along the midline of the physical radial circumference and the physical axial direction: upper left, upper right, lower left, and lower right. Each quadrant node containing non-zero pixels is further divided into four sub-quadrant nodes in the same manner. Quadrant nodes without non-zero pixels are stopped from being divided and marked as empty nodes, until the side length of the quadrant nodes reaches a preset minimum node side length. The minimum node side length is determined based on the minimum pixel size of the defect to be detected; for example, the minimum node side length can be 2 pixels. The quadtree data structure is a tree structure where each non-leaf node has four child nodes. Each node records the coordinate range of the rectangular region it covers in the two-dimensional defect response plane and the set of non-zero pixels it contains. The recursive division organizes non-zero pixels into a tree structure indexed by spatial location hierarchy. Regions with densely distributed non-zero pixels are subdivided to the lowest level node, while blank regions are marked as empty nodes at higher levels and no longer subdivided, thus avoiding invalid traversal of large areas without defects. See also Figure 6 This is a schematic diagram of the quadtree partitioning of the defect response plane provided in an embodiment of the present invention. A large square represents a two-dimensional defect response plane. Cross lines inside the large square divide it into four sub-quadrants. In the quadrant containing scattered black dots, cross lines are drawn to further subdivide it, and the scattered black dots represent non-zero pixels. In the large quadrant without any dots, the interior remains blank, representing empty nodes.

[0063] Step S42: Extract the lowest level node from the quadtree data structure and remove it to obtain the defect feature quadtree.

[0064] After generating the quadtree data structure, the bottom-level nodes in the quadtree structure have organized the non-zero pixels in the two-dimensional defect response plane into a tree structure indexed by spatial location hierarchy. However, after iteratively solving the sparse anomalous defect tensor using the alternating direction multiplier method, individual isolated single-pixel-level random responses may still remain in its non-zero tensor elements. These random responses are not caused by real micro-cracks or indentation unevenness defects, but by extremely weak numerical perturbations in the tensor decomposition residuals. They do not constitute a continuous physical structure in space, corresponding to isolated bottom-level nodes containing only a very small number of non-zero pixels in the quadtree data structure. If not removed, these isolated bottom-level nodes will introduce false defect responses in subsequent connected component merging and physical size calculations, interfering with the final interception decision. Therefore, a preset noise threshold is used to screen the number of non-zero pixels contained in the bottom-level nodes, and isolated bottom-level nodes that do not meet the threshold condition are removed.

[0065] Specifically, the bottom-level node refers to the leaf node in the quadtree data structure that no longer has child nodes and whose side length is equal to the side length of the smallest node. The quadtree data structure is traversed in a depth-first manner, visiting each bottom-level node sequentially. For the currently visited bottom-level node, the number of non-zero pixels in the two-dimensional defect response plane within its covered rectangular area is counted, which is taken as the non-zero pixel count of the current bottom-level node. The noise threshold is determined based on the area array camera resolution and the minimum physical size of the actual defect to be detected. Half of the pixel area corresponding to the minimum physical size is taken as the preset noise threshold. Bottom-level nodes with a non-zero pixel count less than the noise threshold are deleted from the quadtree data structure. Simultaneously, all upper-level nodes whose four child nodes are empty after deletion are also deleted. The upper-level nodes refer to nodes that directly contain the current bottom-level node in the quadtree data structure. The quadtree data structure is defined as follows: A parent node and all its parent nodes tracing back along the root node are considered. The parent node refers to the node directly above the current node generated by a single equal partition in the quadtree data structure. The lowest-level node with a non-zero pixel count greater than or equal to the noise threshold and its parent nodes at all levels are retained. The parent-child index relationship is reconstructed for all retained nodes: For each retained non-lowest-level node, its four child nodes directly generated by equal partitioning in the original quadtree data structure are checked. Child nodes still in the retained state are registered as the current child nodes of that node, and deleted child nodes are removed from the node's child node list. Based on the spatial inclusion relationship between the coordinate range of the rectangular area covered by each retained node and the coordinate range of the rectangular area covered by its current child node, the parent-child index between each node is confirmed and updated level by level to obtain the defect feature quadtree. The data structure of the defect feature quadtree is the same as the quadtree data structure, but only the lowest-level node with a non-zero pixel count greater than or equal to the noise threshold and its parent nodes at all levels are retained, thus filtering out isolated random noise while preserving the continuous response of the real crack and indentation unevenness.

[0066] Step S43: Merge the connected components of the defect feature quadtree to obtain the target defect connected component set, calculate the physical size of the defect based on the target defect connected component set, and make a comparison and judgment based on the physical size of the defect to generate a real-time interception signal.

[0067] After obtaining the defect feature quadtree, the non-zero pixels in the sparse anomaly defect tensor are organized into a tree structure indexed by spatial location hierarchy. However, limited by the geometric boundaries of the recursive partitioning of the defect feature quadtree, the same complete micro-crack or indentation uneven physical defect is often split into adjacent bottom-level nodes. Directly measuring these bottom-level nodes scattered across different branches will cause a single complete defect to be repeatedly split into multiple independent objects, leading to false negatives and false negatives. At the same time, the pixel coordinates in the bottom-level nodes are still in the pixel dimension of the two-dimensional defect response plane, which cannot directly reflect the true physical size. In order to restore the scattered bottom-level nodes into complete defect objects with physical continuity and to convert pixel-level detection results into engineering physical quantities, adjacent bottom-level nodes are merged using connected component merging operations, and the pixel coordinates are mapped back to the three-dimensional physical coordinate system through an inverse matrix, thereby restoring the physical spatial shape of the true defect and outputting a real-time interception signal.

[0068] Specifically, all bottom-level nodes in the defect feature quadtree are obtained. Each bottom-level node is initialized as an independent connected component, which is a set of non-zero pixels that are spatially connected in the two-dimensional defect response plane. The eight-adjacency criterion is used to determine whether any two bottom-level nodes are adjacent. The eight-adjacency criterion is defined as follows: if any one of the eight directions (up, down, left, right, upper left, upper right, lower left, lower right) of the rectangular area covered by one bottom-level node shares a common edge or a common vertex with the rectangular area covered by another bottom-level node, then the two bottom-level nodes are considered adjacent. A common edge refers to a line segment region in the two-dimensional defect response plane where the geometric boundary line segments of the rectangular areas covered by the two bottom-level nodes completely or partially overlap; a common vertex refers to the corner point of the geometric boundary of the rectangular areas covered by the two bottom-level nodes having completely identical horizontal and vertical coordinate values ​​in the two-dimensional defect response plane. Traverse all bottom-level nodes in the defect feature quadtree. If any two bottom-level nodes are determined to be adjacent, merge their connected components into the same connected component. Summarize all connected components obtained after all merging operations to obtain the target defect connected component set. The data structure of the target defect connected component set is a set of connected components, where each connected component records the coordinate range of the merged rectangular region of all bottom-level nodes it contains and the set of all non-zero pixels it contains.

[0069] The set of connected components of the target defect is mapped to a three-dimensional physical coordinate system using the inverse of the cylindrical projection coordinate transformation matrix, and the physical dimensions of the defect are calculated. The inverse of the cylindrical projection coordinate transformation matrix is ​​used to reverse-map the two-dimensional coordinates on the two-dimensional plane coordinate system of the unfolded cylindrical surface back to the three-dimensional physical coordinates of the outer conductor of the RF coaxial cable corrugated. The mapping to the three-dimensional physical coordinate system is as follows: for each non-zero pixel in the set of connected components of the target defect, its coordinates in the two-dimensional defect response plane are expanded to homogeneous coordinates, then multiplied on the left by the inverse of the cylindrical projection coordinate transformation matrix and normalized to obtain the axial and circumferential physical coordinates of the non-zero pixel in the three-dimensional physical coordinate system based on the RF coaxial cable axis. Based on the axial and circumferential physical coordinates, for each connected component, the difference between the maximum and minimum values ​​of the axial physical coordinates of all its non-zero pixels is taken as the axial physical length of the defect; the arc span corresponding to the difference between the maximum and minimum values ​​of the circumferential physical coordinates of all its non-zero pixels is taken as the circumferential physical width of the defect; the axial physical length and circumferential physical width of the defect together constitute the physical dimensions of the defect. A real-time interception signal is generated when the physical size of a defect exceeds a preset tolerance. The preset tolerance refers to the maximum physical size limit of defects allowed by the product quality standard for the corrugated outer conductor of RF coaxial cables; for example, it is set as an upper limit of 0.5 mm for the axial physical length and 0.3 mm for the circumferential physical width. The physical size of the defect in each connected region is compared and calculated against the preset tolerance. When the axial physical length of the defect exceeds the upper limit of the preset tolerance, or the circumferential physical width of the defect exceeds the upper limit of the preset tolerance, a real-time interception signal is generated, carrying the three-dimensional physical coordinates of the non-zero pixels in the current connected region and the physical size of the defect. The real-time interception signal is sent to the programmable logic controller (PLC) of the cable continuous extrusion production line, so that the PLC can mark the axial physical coordinate position corresponding to the defect and trigger the rejection and slitting action of the downstream take-up segment, thereby converting the pixel-level detection result into actual engineering control actions for the production line. See also... Figure 7 This is a schematic diagram of connected component merging provided in an embodiment of the present invention. Multiple adjacent small squares are drawn on a mesh background to represent the bottom-level nodes. Several small squares are connected together by perfectly fitting edges, representing common edges, or intersecting angles, representing common vertices, forming an irregular contiguous region. This contiguous region is outlined by a bold dashed line, corresponding to the target defect connected component set.

[0070] Step S40 solves the technical problems of generating a huge amount of invalid traversal in a large area of ​​defect-free region, isolated random noise interference left by tensor decomposition, and the inability to perform accurate physical measurement and engineering control due to the splitting of the same physical defect by spatial partitioning boundary by using sparse anomaly defect tensor and real-time interception signal. Among them, the sparse anomaly defect tensor is mapped into real-time interception signal after spatial hierarchical indexing and connected component restoration, realizing the accurate closed-loop conversion from abstract pixel-level defect response to actual production line physical space engineering interception action.

[0071] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0072] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An online detection method for surface defects on the corrugated outer conductor of a radio frequency coaxial cable, characterized in that, The method includes: Obtain the original multi-view cylindrical image, extract and remove overlapping edge pixels from the original multi-view cylindrical image to obtain the corrected multi-view cylindrical image, and perform pixel-level coordinate projection transformation on the corrected multi-view cylindrical image to obtain the flattened wavy panoramic image. The mean axial brightness projection is calculated based on the flattened wavy panoramic image and combined into an axial brightness projection sequence. The time axis distortion offset is calculated using the axial brightness projection sequence and combined into a time axis distortion offset sequence. The column offset of the flattened wavy panoramic image is calculated based on the time axis distortion offset sequence. The column offsets are then arranged in a matrix to form a two-dimensional spatial phase drift field. A reference image is obtained by offset correction of the flattened wavy panoramic image using a two-dimensional spatial phase drift field. The reference image is then stitched together to obtain a three-dimensional reference tensor block, and the three-dimensional reference tensor block is decomposed to obtain a sparse anomaly defect tensor. The sparse anomaly defect tensor is mapped to a quadtree data structure. The bottom-level nodes are extracted and removed from the quadtree data structure to obtain the defect feature quadtree. Connected component merging is performed on the defect feature quadtree to obtain the target defect connected component set. The physical size of the defect is calculated based on the target defect connected component set, and a comparison and judgment are made based on the physical size of the defect to generate a real-time interception signal.

2. The online detection method for surface defects of the corrugated outer conductor of an RF coaxial cable according to claim 1, characterized in that, The method for obtaining the original multi-view cylindrical image includes: By detecting the circular array of area cameras at the workstation, exposure and acquisition are completed simultaneously based on the synchronous trigger signal output by the rotary encoder, and the original multi-view cylindrical image of the corrugated outer conductor of the RF coaxial cable is obtained. The data structure of the original multi-view cylindrical image is multiple two-dimensional grayscale matrices, with row indices corresponding to pixel positions in the physical radial circumferential direction and column indices corresponding to pixel positions in the physical axial direction.

3. The online detection method for surface defects of the corrugated outer conductor of an RF coaxial cable according to claim 1, characterized in that, The method for obtaining the corrected multi-view cylindrical image includes: Extract the pixel array region of the overlapping arc segment from each two-dimensional grayscale matrix as the overlapping edge image region; Set a candidate threshold to divide the overlapping edge image region into low grayscale class and high grayscale class, and calculate the pixel ratio and grayscale mean of the two classes. Based on the pixel ratio and grayscale mean, calculate the inter-class variance corresponding to the candidate threshold, and extract the candidate threshold that maximizes the inter-class variance as the optimal segmentation threshold. Obtain the pixel grayscale values ​​in the overlapping edge image region, and remove pixels whose pixel grayscale values ​​are less than the optimal segmentation threshold, while retaining pixels whose grayscale values ​​are equal to or greater than the optimal segmentation threshold, to obtain the corrected multi-view cylindrical image.

4. The online detection method for surface defects of the corrugated outer conductor of an RF coaxial cable according to claim 3, characterized in that, The method for acquiring the flattened wavy panoramic image includes: Obtain the cylindrical projection coordinate transformation matrix, expand the pixel position of the pixel in the two-dimensional grayscale matrix in the corrected multi-view cylindrical image to homogeneous coordinates, multiply it by the cylindrical projection coordinate transformation matrix and normalize it to obtain the circumferential coordinates and axial coordinates of the pixel. The corrected multi-view cylindrical image after coordinate projection transformation is stitched together end to end along the circumferential coordinate direction. For positions where the circumferential and axial coordinates are non-integer values, four adjacent integer coordinate pixels in a preset regular pixel grid are extracted. The distance ratio of each pixel is calculated as a weight, and the pixel gray values ​​of the four integer coordinate pixels are weighted by bilinear interpolation and summed to obtain the resampled gray values. After traversing and assigning values, a flattened wavy panoramic image is obtained.

5. The online detection method for surface defects of the corrugated outer conductor of an RF coaxial cable according to claim 4, characterized in that, The method for obtaining the axial brightness projection sequence includes: Traverse each column of the flattened wavy panoramic image, sum the integer pixel grayscale values ​​of all rows contained in the current column, and divide by the total number of rows to obtain the average axial brightness projection value corresponding to that column. Arrange the axial brightness projection mean values ​​of all columns in ascending order to generate an axial brightness projection sequence.

6. The online detection method for surface defects of the corrugated outer conductor of an RF coaxial cable according to claim 5, characterized in that, The method for obtaining the time axis warp offset sequence includes: Obtain the standard ripple sequence, define each array element in the standard ripple sequence as a standard sampling point, and define each array element in the axial brightness projection sequence as a projection sampling point; Construct a distance matrix where the values ​​of each element are equal to the absolute values ​​of the algebraic differences between the mean axial brightness projection of the projected sampling points and the values ​​of the standard sampling points. Construct a cumulative distance matrix. The elements in the cumulative distance matrix are recursively calculated by adding the minimum value of the distance matrix element to the left, above, and upper left of the current position. Based on the cumulative distance matrix, obtain a regular path. Traverse all projected sampling points and search for standard sampling points that match the current projected sampling point on the regular path. When there are multiple matches, take the standard sampling point with the largest sequence index. Subtract the sequence index of the current projected sampling point from the largest sequence index and use the difference as the time axis distortion offset. After traversal, combine them into a time axis distortion offset sequence.

7. The online detection method for surface defects of the corrugated outer conductor of an RF coaxial cable according to claim 6, characterized in that, The method for constructing the two-dimensional spatial phase drift field includes: The time axis distortion offset of each projected sampling point in the time axis distortion offset sequence is assigned to all pixels in the corresponding column of the flattened wavy panoramic image, as the column offset of that column. For column positions in the column index that are not directly covered by regularized paths, find the column positions and their column offsets that are directly covered by regularized paths immediately to the left and right of the column, and complete the column offsets of the column using linear interpolation. Arrange the column offsets of all pixels in a matrix according to their row and column index positions to form a two-dimensional spatial phase drift field.

8. The online detection method for surface defects of the corrugated outer conductor of a radio frequency coaxial cable according to claim 7, characterized in that, The method for obtaining the reference image includes: Define the column index of each pixel in the flattened wavy panoramic image as the axial coordinate component, and define the combination of the row index and column index as the original pixel coordinate. Add the column offset at the same row and column index position in the two-dimensional spatial phase drift field to the axial coordinate component of the original pixel coordinates to obtain the target alignment coordinates; For each integer coordinate point in the coordinate grid formed by the target alignment coordinates, look up its corresponding mapped floating-point coordinates in the flattened wavy panoramic image, extract the four integer pixels that are immediately adjacent to the mapped floating-point coordinates, perform bilinear interpolation weighted summation with distance ratio as the weight, obtain the pixel grayscale value after resampling of the current integer coordinate point, and obtain the reference image after traversing and assigning values.

9. The online detection method for surface defects of the corrugated outer conductor of a radio frequency coaxial cable according to claim 2, characterized in that, The method for constructing the three-dimensional reference tensor block includes: Set the frame number, extract multiple consecutive reference images equal to the frame number according to the order in which the synchronization trigger signal is generated, and define them as consecutive multi-frame reference images. Define a time frame dimension, which uses the order in which the synchronization trigger signals are generated as the sorting index; Multiple consecutive reference images are orthogonally stacked and stitched along the time frame dimension to generate a three-dimensional reference tensor block. The first order of the three-dimensional reference tensor block corresponds to the physical radial circumferential direction, the second order corresponds to the physical axial direction, and the third order corresponds to the time frame dimension.

10. The online detection method for surface defects of the corrugated outer conductor of an RF coaxial cable according to claim 1, characterized in that, The method for calculating the sparse anomaly defect tensor includes: Introduce and initialize the low-rank ripple background tensor, the sparse anomaly defect tensor, the Lagrange multiplier tensor, and the quadratic penalty factor; The calculation is performed based on the three-dimensional reference tensor block, sparse anomaly defect tensor, Lagrange multiplier tensor and quadratic penalty factor. The calculation results are expanded along the three dimensions, the singular value soft thresholding shrinkage operation is performed, and after folding and restoring, the arithmetic mean is taken to update the low-rank ripple background tensor. The calculation is performed based on the three-dimensional reference tensor block, the low-rank ripple background tensor, the Lagrange multiplier tensor and the quadratic penalty factor. The calculation results are then subjected to element-wise soft thresholding shrinkage to update the sparse anomaly defect tensor. Calculate the residual tensor between the 3D reference tensor block and the low-rank ripple background tensor and the sparse anomaly defect tensor, and update the Lagrange multiplier tensor by combining the quadratic penalty factor and the residual tensor. The process is repeated until the residual tensor norm is less than the preset convergence threshold, and then the updated sparse anomaly defect tensor is output.

Citation Information

Patent Citations

  • Pipeline defect detection method and system based on multi-scale attention

    CN122200136A

  • Water quality monitoring and early warning system and method based on image multi-feature fusion

    CN122591552A