Method for accurately extracting micro-defects of seamless steel pipe based on point cloud features

Abstract

The application discloses a seamless steel pipe micro-defect accurate extraction method based on point cloud features, which comprises the following steps: collecting a pipe surface three-dimensional point cloud, converting to a point set in a pipe coordinate system; resampling to form a height map and establishing an index mapping; sequentially performing three-level background suppression on the height map: sliding window robust fitting residual stripping macro-morphology, multi-scale decomposition low-frequency stripping mesoscale fluctuation, frequency domain high-pass filtering and adaptive notch stripping periodic texture, to obtain a bottom map; constructing a height domain enhancement channel based on the bottom map, constructing a geometric domain enhancement channel based on the point cloud, and mapping and fusing to obtain a defect saliency map; adaptive threshold segmentation to obtain a candidate region; sub-pixel boundary positioning in a one-dimensional profile along the normal direction of the candidate region boundary, and sub-pixel depth positioning in the region; calculating geometric features according to the sub-pixel boundary and depth, and judging the defect type. The application solves the problem that the existing seamless steel pipe defect detection technology cannot stably detect micro-defects.

Description

Technical Field

[0001] This invention relates to the field of industrial nondestructive testing technology, specifically to a method for accurately extracting minute defects in seamless steel pipes based on point cloud features. Background Technology

[0002] During the processes of piercing, rolling, sizing, straightening, and cold drawing, seamless steel pipes inevitably develop various defects on their outer surface. Among these defects, tiny, shallow defects with a depth of 0.1mm to 0.3mm and a width of 0.2mm (most typically shallow scratches and blue lines) have long been a major area of ​​missed detection in the industry. This is not due to a lack of attention from on-site operators, but rather because existing testing methods cannot detect these tiny, shallow defects.

[0003] Firstly, there is manual visual inspection. Depressions or protrusions of 0.1mm in size barely create a noticeable difference in light and shadow on the pipe surface. Inspectors need to monitor moving steel pipes at a rapid pace on the production line, making it common for such defects to be missed. Furthermore, the judgment standards differ between different inspection teams, and this tendency to miss defects is exacerbated by inspector fatigue.

[0004] Secondly, there's two-dimensional machine vision. Grayscale imaging relies on the contrast between light and shadow at the defect, but shallow defects cause very little height variation, resulting in image contrast close to noise levels. Even more problematic is that oxide stains, watermarks, and oil spots on the pipe surface appear as large areas of grayscale anomalies in the image, leading to a trade-off between false positives and false negatives, making it difficult to achieve a perfect balance regardless of the detection threshold setting.

[0005] Third, electromagnetic detection (eddy current, magnetic leakage). This type of method is effective for internal defects or deep volumetric defects, but it is not sensitive enough for surface defects with small openings and very shallow depths; even if it can detect them, it cannot provide the accurate outline and size of the defect, and cannot support fine-grained classification.

[0006] Fourth, existing attempts at 3D point cloud detection. Laser triangulation can acquire the 3D topography of the pipe surface, unaffected by surface color, and is theoretically the most promising approach. However, three problems arise during implementation: (1) The steel pipe is not an ideal cylinder. The macroscopic fluctuations caused by the bending of the whole pipe, ellipticity, and local warping are large, and the amplitude is one to two orders of magnitude larger than that of small defects, and the defect signal is directly submerged; (2) Regular medium-scale undulations such as rolling roll marks and spiral textures partially overlap with minor defects in the frequency band, and a single filtering method cannot address both. (3) Point clouds are discrete sampling, and the defect boundary can only be determined down to the sampling point level. Taking a point spacing of 0.2 mm as an example, the boundary positioning error is close to the width of the tiny defect itself, making size measurement meaningless and the judgment threshold impossible to set accurately.

[0007] This has created an awkward situation in the industry: point cloud data can be collected, but minute defects cannot be extracted. To reliably extract these minute defects, three things must be addressed simultaneously: removing the complex background, amplifying the weak defect signal to a level that allows for stable segmentation, and measuring the boundary and depth below the sampling interval. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, a method for accurately extracting minute defects in seamless steel pipes based on point cloud features is provided to solve the problem that existing defect detection technologies for seamless steel pipes cannot reliably detect minute defects.

[0009] To achieve the above objectives, a method for accurately extracting minute defects in seamless steel pipes based on point cloud features is provided, comprising the following steps: A three-dimensional point cloud of the circumferential surface of a seamless steel pipe is acquired, and the three-dimensional point cloud is transformed into the pipe coordinate system to obtain a point set represented by circumferential angle, axial position and radial measurement. The point set is resampled along the circumferential and axial directions with the tube axis as a reference to form a gridded height map. The values ​​of the grid points in the height map are the radial measurements, and the grid points form an index mapping with the three-dimensional point cloud. The first, second and third levels of background suppression are sequentially performed on the height map to obtain the base map. The first level of background suppression uses a sliding window to robustly fit the local reference surface and take the residual to remove the macroscopic morphology. The second level of background suppression uses multi-scale decomposition to discard low-frequency background components to remove the mesoscale undulations. The third level of background suppression uses frequency domain high-pass filtering and adaptive notch filtering to remove periodic textures. A height domain enhancement channel is constructed based on the base map, and a geometry domain enhancement channel is constructed based on the 3D point cloud. The response of the geometry domain enhancement channel is mapped to the base map and weighted and fused with the response of the height domain enhancement channel to obtain a defect saliency map. Adaptive threshold segmentation is performed on the defect saliency map to obtain candidate defect regions; Subpixel boundary localization is performed on a one-dimensional cross-section along the normal direction of the candidate defect region boundary, and subpixel depth localization is performed within the candidate region to output the subpixel level boundary and depth of the minute defect; The geometric features of the minute defect are obtained by calculating the subpixel boundary and depth, and the type of the minute defect is determined based on the geometric features.

[0010] Furthermore, the height domain enhancement channel includes a speckle enhancement sub-channel and a linear enhancement sub-channel; Based on the base map, highly correlated features are extracted. The speckle enhancement sub-channel uses a multi-scale Gaussian difference operator to filter the highly correlated features and outputs the speckle response. The linear enhancement sub-channel uses the Hessian matrix operator to perform multi-directional filtering on the highly correlated features, outputting a linear response.

[0011] Furthermore, the geometric domain enhancement channel includes a normal deflection angle sub-channel and a curvature anomaly sub-channel; The normal deflection angle sub-channel outputs the deflection angle response based on the spatial angle between the normal vector of each point in the three-dimensional point cloud and the normal vector of the ideal tube surface. The curvature anomaly subchannel outputs a curvature anomaly response based on the degree of deviation between the principal curvature of each point in the three-dimensional point cloud and the ideal pipe surface curvature.

[0012] Furthermore, during the first-level background suppression, a sliding window is opened along the axial direction. Within the sliding window, the local reference surface is obtained by RANSAC random sampling consistency fitting. Defect points are excluded as outliers. The reference parameters of two adjacent sliding windows are smoothly transitioned by interpolation. The residual map of the height map relative to the local reference surface is obtained point by point.

[0013] Furthermore, when performing the second-level background suppression, the residual map is decomposed using two-dimensional wavelet decomposition, the low-frequency approximate subband is discarded, and the mid-to-high frequency map is obtained by reconstructing the mid-to-high frequency map using the mid-to-high frequency detail subband.

[0014] Furthermore, during the third-level background suppression, a two-dimensional Fourier transform is performed on the mid-to-high frequency image, and after high-pass filtering, the position of the spectral peak is automatically detected. Narrowband notch filtering is then applied to the fundamental frequency and harmonics to obtain the background image.

[0015] Furthermore, the step of locating sub-pixel boundaries on a one-dimensional profile along the normal direction of the candidate defect region boundary includes: The one-dimensional cross-section is taken on the base map along the normal direction of each point along the boundary of the candidate defect region; After interpolation and encryption, the one-dimensional profile is subjected to curve fitting on the edge segment. The intersection of the fitted curve and the local background horizontal is taken as the sub-pixel edge point. Traverse all points of the candidate defect region boundary and connect all the sub-pixel edge points to form a closed sub-pixel boundary.

[0016] Furthermore, the subpixel depth localization step within the candidate region includes performing local surface fitting on the base image within the candidate defect region, and taking the extreme value of the fitted surface as the subpixel depth.

[0017] This invention provides an electronic device, comprising: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method for accurate extraction of minute defects in seamless steel pipes based on point cloud features as described in any of the above designs.

[0018] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method for accurate extraction of minute defects in seamless steel pipes based on point cloud features as described in any of the above designs.

[0019] The beneficial effects of this invention are that the method for accurate extraction of minute defects in seamless steel pipes based on point cloud features collects the point cloud of the steel pipe in the entire circumference. The preprocessing, three-level background suppression, enhancement fusion, segmentation and sub-pixel extraction are all processed in a segmented and parallel manner to achieve "inspection while passing through". It can directly replace the high-intensity manual visual inspection position and greatly improve the detection efficiency and consistency.

[0020] The meshed height map in the method for accurate extraction of minute defects in seamless steel pipes based on point cloud features of this invention facilitates the use of mature two-dimensional image algorithms, while geometric quantities such as normal vectors and curvature are always calculated on the original three-dimensional point cloud. The two maintain a strict correspondence through index mapping, which not only ensures the convenience of signal processing, but also avoids the loss of geometric information caused by meshing. This is one of the substantial innovations of this technical solution that distinguishes it from conventional pure image or pure point cloud methods.

[0021] The three-level background suppression method of the seamless steel pipe micro-defect accurate extraction method based on point cloud features of the present invention targets the steel pipe bending, ellipticity, local undulation (millimeter-level macroscopic deformation), roll marks and wavy lines (medium-scale undulations) and spiral rolling texture (periodic interference) step by step. Each method is used for a scale segment of each pipe and the serial order is optimized and determined, so that the algorithm is not sensitive to field conditions such as steel pipe specification changes, batch differences, and production line speed fluctuations, the false alarm rate is controllable, and the maintenance cost is low.

[0022] This invention presents a method for accurately extracting minute defects in seamless steel pipes based on point cloud features. It constructs a complementary enhancement channel combining a height domain and a geometric domain, providing a dual chain of evidence. Based on a base map, height domain enhancement and geometric domain enhancement are fused in parallel. The geometric channel directly utilizes the normal vector and curvature information of the original 3D point cloud, exhibiting significantly higher contrast than the height map for defects with extremely shallow depths but still deflected edges (such as shallow scratches and blue lines). This multi-dimensional fusion mechanism enables shallow defects to obtain a stable response on the saliency map, overcoming the fundamental problem of unreliable segmentation caused by weak single height signals.

[0023] This invention provides a method for accurately extracting minute defects in seamless steel pipes based on point cloud features, overcoming the limitations of point cloud sampling intervals and achieving sub-pixel-level boundary and depth measurements. A one-dimensional profile is taken along the normal direction at the boundary of the candidate defect region. After interpolation and curve fitting, the intersection of the fitted curve and the background horizontal plane is used as the sub-pixel edge point, achieving boundary positioning accuracy of 1 / 5 to 1 / 10 of the sampling interval. Simultaneously, local surface fitting is performed inside the defect, and the extreme value positioning depth resolution also reaches the sub-pixel level. This allows for the accurate online quantification of the geometric dimensions of the defect, such as length, width, area, and depth, providing a reliable quantitative basis for subsequent defect classification and judgment. Attached Figure Description

[0024] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating the method for accurately extracting minute defects in seamless steel pipes based on point cloud features, according to an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of the steel pipe measurement cross section in Embodiment 1 of the present invention.

[0026] Figure 3 This is the original radial cross-sectional view of the steel pipe in Embodiment 1 of the present invention.

[0027] Figure 4 This is a radial cross-section of the steel pipe after first-level background suppression in Embodiment 1 of the present invention.

[0028] Figure 5 This is a radial cross-section of the steel pipe after the second-level background suppression in Embodiment 1 of the present invention.

[0029] Figure 6 This is a radial cross-section of the steel pipe after third-level background suppression in Embodiment 1 of the present invention.

[0030] Figure 7 This is a cross-sectional view of the height residual of the steel pipe after being enhanced by the height domain enhancement channel in Embodiment 1 of the present invention.

[0031] Figure 8 This is a cross-sectional view of the normal deflection angle of the steel pipe after being reinforced by the geometric domain enhancement channel in Embodiment 1 of the present invention.

[0032] Figure 9 This is a schematic diagram of the principle of candidate defect region boundary extraction in Embodiment 1 of the present invention.

[0033] Figure 10 This is a schematic diagram illustrating the depth measurement principle of the candidate defect region in Embodiment 1 of the present invention.

[0034] Figure 11This is a schematic diagram illustrating the classification of minute defect morphologies in Embodiment 1 of the present invention.

[0035] Figure 12 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention.

[0036] Figure label: 1. Seamless steel pipe; 11. Laser profile sensor; Processor 210, communication interface 220, memory 230, communication bus 240. Detailed Implementation

[0037] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0038] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0039] Reference Figure 1 This invention provides a method for accurately extracting minute defects in seamless steel pipes based on point cloud features, comprising the following steps: S1. Collect the three-dimensional point cloud of the circumferential surface of the seamless steel pipe (hereinafter referred to as the steel pipe), and convert the three-dimensional point cloud to the pipe coordinate system to obtain a point set represented by circumferential angle, axial position and radial measurement.

[0040] In this embodiment, refer to Figure 2 As shown, a measurement section is selected on the conveying path of the steel pipe, and N sets of laser profile sensors (N is a positive integer from 6 to 12, and N is 8 sets in this embodiment) are evenly distributed around the circumference, with each set covering a 360° / N arc segment. A 5% to 10% overlap is reserved between the fields of view of two adjacent sets of laser profile sensors for splicing and registration.

[0041] The steel pipe passes through the measurement section at a uniform speed along the axial direction. The speed encoder on the roller conveyor used to transport the steel pipe outputs pulses, which synchronously trigger N sets of laser profile sensors to collect three-dimensional point cloud data of the circumferential surface (or outer surface) of the seamless steel pipe. Each frame of the N sets of laser profile sensors obtains a segment of circumferential profile line. The frame sequence is spliced ​​according to the axial position to obtain the complete point cloud of the entire outer surface of the steel pipe.

[0042] The point cloud is represented in the tube coordinate system as {Pi=(θi, zi, ri)}, where θ is the circumferential angle, z is the axial coordinate, and r is the radial measurement.

[0043] This embodiment employs a ring array laser profile sensor that passes through in one pass with no blind spots in the circumference, better meeting the actual needs of steel pipe production lines. For calibration, standard bars are used to periodically calibrate the pose parameters of N sets of laser profile sensors to a unified pipe coordinate system.

[0044] S2. Using the tube axis as a reference, the point set is resampled along the circumferential and axial directions to form a gridded height map. The values ​​of the grid points in the height map are radial measurements, and the grid points form an index mapping with the 3D point cloud.

[0045] Specifically, using the tube axis as a reference, the three-dimensional point cloud is resampled to a regular grid according to (θ, z), and the radial measurement r is used as the grid value of the grid point of the regular grid to form a height map H(θ, z).

[0046] Resampling uses inverse distance weighted interpolation, with grid spacing comparable to the original point density (0.2 mm in both circumferential and axial directions), and the index mapping from grid points to original 3D points is preserved.

[0047] The essence of this height map H(θ,z) is to unfold the circumference of the steel pipe into a two-dimensional topographic map.

[0048] The horizontal axis (θ) of the height diagram represents the circumferential angle of the steel pipe (0°~360°), which is equivalent to cutting and flattening the outer wall of the pipe along the generatrix.

[0049] The vertical axis (z) of the height diagram represents the axial length of the steel pipe (along the direction in which the steel pipe moves).

[0050] The grid points are the intersections of a regular two-dimensional grid, with a spacing of 0.2mm × 0.2mm (comparable to the point cloud density).

[0051] The value r (radial measurement) stored at the grid point is the actual physical radius of the grid point from the central axis of the steel pipe.

[0052] If the steel pipe were a perfect, ideal cylinder, this height map would be a uniform gray plane with equal height everywhere.

[0053] Defects (scratches, holes, blue lines) on steel pipes will have a smaller radius at the defect location (surface concavity), which will appear as a dark / low grayscale area on the image.

[0054] The raised defects (scabbing, steel adhesion) on the steel pipe have an increased radius (surface protrusion), which are represented as bright / high grayscale areas on the image.

[0055] Height maps transform messy, irregularly arranged 3D point clouds (discrete points) into regular, ordered 2D matrices (images). This allows mature 2D image signal processing algorithms such as filtering, wavelet decomposition, Fourier transform, and difference of Gaussian (DoG) to run directly on the map without having to deal with complex 3D topological relationships.

[0056] S3. Perform first-level, second-level, and third-level background suppression on the height map in sequence to obtain the base map. The first-level background suppression uses a sliding window to robustly fit the local reference surface and take the residual to remove the macroscopic morphology. The second-level background suppression uses multi-scale decomposition to discard low-frequency background components to remove mesoscale undulations. The third-level background suppression uses frequency domain high-pass filtering and adaptive notch filtering to remove periodic textures.

[0057] When performing the first-level background suppression, a sliding window is opened along the axis. Within the sliding window, a local reference plane is obtained by RANSAC random sampling consistency fitting. Defect points are excluded as outliers. The reference parameters of two adjacent sliding windows are smoothly transitioned by interpolation. The residual map is obtained by taking the residual map of the height map relative to the local reference plane point by point.

[0058] In this embodiment, the first-level background suppression is a sliding window robust reference surface fitting. The first-level background suppression is performed by opening a sliding window along the axis (window length of 200mm, half-window overlap), and fitting a local cylindrical surface within the window using RANSAC random sampling consistency: repeatedly randomly sampling subsets to estimate the cylinder parameters, taking the consistency set with the most interior points as the standard, and setting the interior point threshold to 0.05mm.

[0059] Defects are naturally excluded as outliers and do not contaminate the baseline.

[0060] The reference parameters of adjacent windows are smoothly transitioned using spline interpolation to obtain a local reference surface field that continuously varies with the tube length.

[0061] Point-by-point residuals: R(θ,z)=H(θ,z) Hbase(θ,z).

[0062] The entire tube bending, ellipticity, and slow whitishness are stripped away by the first-level background suppression.

[0063] The first-level background suppression input is the height map H(θ,z). The height map H contains the original map with all the information, including macroscopic curvature (millimeter-level), mesoscale undulations (roll printing), periodic textures, and tiny defect signals.

[0064] A local reference plane Hbase, varying with pipe length, is fitted using a sliding window and RANSAC. Then, the original height map is subtracted from this reference plane (H-Hbase) to output the residual map R(θ,z).

[0065] The output residual map R(θ,z) has been stripped of macroscopic curvature, ellipticity, and warping, leaving mainly mesoscale undulations (wavy lines / roller marks) and defect signals. At this point, the residual amplitude has been reduced from the millimeter level to the sub-millimeter level.

[0066] When performing the second-level background suppression, the residual map is decomposed into two-dimensional wavelet decomposition, the low-frequency approximate subband is discarded, and the mid-to-high frequency map is obtained by reconstructing the mid-to-high frequency map using the mid-to-high frequency detail subband.

[0067] Two-dimensional wavelet decomposition is performed on the residual map R(θ, z). The low-frequency approximate subband corresponds to the mesoscale background and is directly discarded. The mid-to-high frequency detail subbands where the defect energy is concentrated are reconstructed to obtain R2(θ, z).

[0068] There is only one principle for choosing wavelet basis and layer number: ensure that the characteristic scale of the defect falls within the frequency band of the reserved subband.

[0069] The second-level background suppression takes as input the residual image R(θ,z) from the first-level background suppression. The second-level background suppression performs a two-dimensional wavelet decomposition on R (e.g., 4-layer db4), splitting the image into "low-frequency approximation (large-scale waves)" and "mid-to-high-frequency details (small-scale abrupt changes)". Then, it directly discards the low-frequency approximation subband and reconstructs the mid-to-high-frequency residual image R2(θ,z) using only the mid-to-high-frequency detail subband.

[0070] The output mid-to-high frequency residual map R2(θ,z) strips away the mesoscale undulations (the general trend of waves and roller printing textures) of several millimeters to tens of millimeters in wavelength. What remains is mainly the defect signal (energy concentrated in the mid-to-high frequency) + the residue of strong periodic textures (such as the fundamental frequency and harmonics of spiral roller printing).

[0071] When performing the third-level background suppression, a two-dimensional Fourier transform is performed on the mid-to-high frequency image, and after high-pass filtering, the position of the spectral peak is automatically detected. The fundamental frequency and harmonics are processed by narrowband notch filtering to obtain the background image.

[0072] Because the multi-scale decomposition of the second-level background suppression is not thorough for strongly periodic textures, the fundamental frequency and harmonics of the spiral roller printing appear as sharp spectral peaks in the spectrum, which are easily retained. The third-level background suppression performs a two-dimensional Fourier transform on R2, first using a high-pass filter to suppress the residual low frequencies (the cutoff wavelength is set according to the maximum characteristic scale of the defect), then automatically detecting the position of the spectral peaks, and performing narrowband notch filtering on the fundamental frequency and each harmonic. After inverse transformation, the background image B(θ,z) is obtained.

[0073] The input for the third-level background suppression is the mid-to-high frequency residual map R2(θ,z) (from the output of the second level).

[0074] The third-level background suppression performs a two-dimensional Fourier transform on R2 to the frequency domain. First, a high-pass filter is applied to suppress the residual extremely low frequencies. Then, sharp "energy spikes" in the spectrum (these spikes correspond to the fundamental frequency / harmonics of the periodic texture) are automatically detected. These spikes are then removed in a specific direction using a narrowband notch filter. Finally, the transformation is performed inversely back to the spatial domain to output the base map B(θ,z).

[0075] The periodic texture interference has been completely removed from the base image B(θ,z). At this point, only the pure, weak defect residual signal (0.1mm-level pit packets) and a very small amount of white noise that cannot be recognized by any filter remain in the B image.

[0076] The three-stage background suppression consists of three stages: a macroscopic stage for fitting, a mesoscale stage for decomposition, and a periodic texture stage for the frequency domain. Without any stage of background suppression, residual background will cover defects as small as 0.1 mm, which is the key difference between this cascaded structure and the conventional "fitting + filtering" approach.

[0077] S4. Construct a height domain enhancement channel based on the base map, construct a geometric domain enhancement channel based on the 3D point cloud, map the response of the geometric domain enhancement channel to the base map and weight-fuse it with the response of the height domain enhancement channel to obtain a defect saliency map.

[0078] The shallow defects in the base map B output in step S3 are still weak, making direct segmentation unreliable. Step S4 enhances the data using three complementary channels, followed by weighted fusion. Specifically, this includes constructing height-domain enhancement channels and geometric-domain enhancement channels.

[0079] Among them, the height domain enhancement channel includes a speckle enhancement sub-channel and a linear enhancement sub-channel.

[0080] Based on the base map, highly correlated features are extracted. The speckle enhancement sub-channel uses a multi-scale Gaussian difference operator to filter the highly correlated features and outputs the speckle response.

[0081] The multi-scale Gaussian difference (DoG) operator uses three levels of σ: 0.15 / 0.3 / 0.6 mm, to enhance the response of clustered targets such as holes, pits, scars, and steel adhesion.

[0082] The DoG (Dot Gauge) is essentially a "shape detector." It treats the base map B as a "topographic contour map," sliding a circular window across it and comparing the residual height value at the window's center point with the residual height values ​​of points around it. If the center point is a negative pit (residual height of -0.15mm) while the surrounding area is flat (close to 0), the DoG calculates a large positive response (marked as a dot). If the surrounding area is also pitted (residual noise), the responses cancel each other out. Therefore, it deals with the degree of spatial abrupt change in residual height.

[0083] The linear enhancement sub-channel uses the Hessian matrix operator to perform multi-directional filtering on highly correlated features, outputting a linear response.

[0084] Based on the linear structural response of the Hessian matrix eigenvalues, the maximum response is obtained point-by-point after filtering at multiple scales and directions. Scratches, blue lines, and cracks, regardless of their orientation, will always be captured by the filter in a certain direction. Random noise does not have directional consistency, and its response is naturally suppressed.

[0085] The Hessian matrix is ​​essentially a directional derivative. It calculates the rate of change of slope (second derivative) of each pixel in the base map B along different directions. On a residual height map, a scratch appears as a continuous, very shallow "groove" (a narrow strip with negative residual height). The Hessian calculates which direction the residual height decreases the fastest and exhibits a "ridge / valley" distribution. If the residual height at a certain point exhibits a "valley" shape along a certain angle (high on both sides, low in the middle), the linear channel will output a strong response at that point. Therefore, it deals with the "linear extension structure of residual height along a certain axis".

[0086] The geometric domain enhancement channel includes a normal deflection angle sub-channel and a curvature anomaly sub-channel.

[0087] The normal deflection angle sub-channel outputs the deflection angle response based on the spatial angle between the normal vector of each point in the 3D point cloud and the normal vector of the ideal pipe surface.

[0088] The curvature anomaly subchannel outputs a curvature anomaly response based on the degree of deviation between the principal curvature of each point in the 3D point cloud and the ideal pipe surface curvature.

[0089] The normal vector and principal curvature of each point are estimated on the original 3D point cloud, and a normal deflection angle map N(θ, z) and a curvature anomaly map K(θ, z) are constructed. Even defects with a depth of only 0.1 mm show a deflection of several to more than ten degrees relative to the pipe surface normal on the geometric map, resulting in a much higher contrast on the height map. Shallow scratches and blue lines are no longer missed under this scheme, primarily due to this channel.

[0090] During fusion, the three channels are normalized and then weighted and superimposed. The weights can be adjusted according to the defect types of concern to the production line. This embodiment provides a set of general weights (w1=0.30, w2=0.30, w3=0.25, w4=0.15), resulting in the defect saliency map S: S=w1·DoG(B)+w2·Line(B)+w3·N′+w4·K′.

[0091] The normal vector deflection angle N and curvature K calculated on the 3D point cloud are calculated in 3D space, but after the calculation, they are filled back into the 2D mesh map according to the mesh position corresponding to each point, forming the response map N'(θ,z) and K'(θ,z) of the geometric channel.

[0092] At this point, N' and K' have the exact same matrix size and coordinate mapping as the filtered base maps DoG(B) and Line(B).

[0093] Finally, the corresponding pixels of these four "images" are weighted and superimposed to complete the fusion and form the final defect saliency image S.

[0094] S5. Perform adaptive threshold segmentation on the defect saliency map to obtain candidate defect regions.

[0095] The defect saliency map S is divided into 32×32 grid blocks. The threshold of each block is determined by adding k times the local mean to the standard deviation (k is 3.5) or by OTSU adaptive determination within the block to resist residual uneven background. After binarization, connected components are marked, morphological opening and closing operations are performed to clean up fragments and fill holes, and isolated noise points are filtered out according to the lower limit of area and length to obtain a set of candidate defect regions.

[0096] S6. Perform sub-pixel boundary localization on a one-dimensional profile along the normal direction of the candidate defect region boundary, and perform sub-pixel depth localization within the candidate region to output the sub-pixel level boundary and depth of the minute defect.

[0097] The steps for sub-pixel boundary localization on a one-dimensional profile along the normal direction of the candidate defect region boundary include: Take a one-dimensional profile on the base map along the normal direction of each point at the boundary of the candidate defect region; After interpolation and encryption, the edge segments of the one-dimensional profile are fitted with curves, and the intersection of the fitted curve and the local background horizontal is taken as the sub-pixel edge points. Traverse all points on the boundary of the candidate defect region and connect all sub-pixel edge points to form a closed sub-pixel boundary.

[0098] The subpixel depth localization step within the candidate region includes fitting a local surface to the base image within the candidate defect region and taking the extreme value of the fitted surface as the subpixel depth.

[0099] The boundaries of the candidate defect regions segmented in step S5 are at the grid level. If the dimensions are directly calculated based on this, with a grid spacing of 0.2mm, the boundary positioning error would be close to the width of the tiny defect itself, rendering the measurement meaningless and the judgment threshold inaccurate. Therefore, this step employs sub-pixel extraction in two steps.

[0100] Subpixel boundary extraction: For each candidate defect region, a one-dimensional profile is taken on the base image B along the normal direction of each point on the boundary of the candidate defect region. The profile is then densified to 8 times the density of the original mesh using bicubic interpolation. A quadratic curve is fitted to the edge segment of the profile, and the intersection point of the fitted curve with the local background horizontal plane is obtained as the subpixel edge point in that normal direction. This process is repeated for all normal directions, and the resulting edge points are connected to form a closed subpixel boundary.

[0101] Subpixel depth measurement: A local quadratic surface is fitted to the base image B within the candidate defect region. The extreme points and extreme values ​​of the fitted surface are taken as the peak depth (for depressions) or peak height (for convexities) of the defect. The depth resolution is no longer limited by the grid spacing and can reach 1 / 5 to 1 / 10 of the sampling spacing.

[0102] At this point, the length, width, area, and depth of the defects are all output with sub-pixel precision, making subsequent classification and grading accurate.

[0103] S7. Calculate the geometric features of the micro-defects based on the sub-pixel boundaries and depth, and determine the type of micro-defects based on the geometric features.

[0104] Based on the sub-pixel boundaries and depths obtained in step S6, a set of geometric and morphological features are calculated for each defect region: area, major axis, minor axis, aspect ratio, orientation angle (relative to tube axis), peak depth (or height), volume, average edge slope (normal deflection angle statistics), residual sign composition (depression / protrusion / mixture), and internal texture density.

[0105] The classification of minor defects uses a rule tree (which can be entirely replaced by a lightweight classifier such as SVM). Typical forms of the nine defect types are shown below. Figure 11 The key points for judgment are as follows: Scratches: large aspect ratio (greater than 8), oriented close to the axis, shallow in depth, and with gentle edge slope.

[0106] Blue lines: large aspect ratio, oriented close to the axis or slightly spiral, narrower in width and shallower in depth—the main difference from scratches lies in the width and depth statistics, requiring a stable response of the geometric channel to reliably distinguish them.

[0107] Holes: Small in area, nearly circular (length-to-width ratio less than 2), concave, with steep edges. Pockmarks resemble holes but are smaller, shallower, and often appear in clusters; local density characteristics need to be introduced for differentiation.

[0108] Cracks: Extremely high aspect ratio, extremely steep edges, variable direction, often bifurcated; counting the bifurcations is a key characteristic. Rolled folds are linear, with the direction consistent with the helix angle of the rolling direction, and the cross-section shows a folded double peak. Dents are localized pits of medium area, with a ring of material extrusion forming a bulge at the edge; the residual sign shows a mixed pattern of "negative inside and positive outside".

[0109] External scars (external folds) are mainly positive residuals or a mixture of positive and negative, with a large area and irregular edges; steel-bonded structures are protruding, with a step-like change in surface height and steep edges.

[0110] The rule parameters can be calibrated once according to the production line specifications and used for a long time. The final output is the circumferential angle, axial position, type, length, width and depth dimensions and confidence level of each defect, and it is linked to the marking, alarm or sorting actuators.

[0111] The present invention provides a method for accurately extracting minute defects in seamless steel pipes based on point cloud features. This method uses a ring array of N laser contour sensors to collect point clouds from the pipe surface, which are then unfolded into a dual representation of a height map and a 3D point cloud in the pipe coordinate system. Subsequently, three levels of background suppression are performed, progressively stripping away macroscopic curvature, mesoscale undulations, and residual periodic textures. Further, multidimensional enhancement with complementary height and geometric domains is applied, followed by adaptive block segmentation to obtain candidate regions and sub-pixel boundaries and depth extraction. Finally, geometric features are calculated to classify nine types of defects and output their location, type, size, and confidence level online.

[0112] Mesh-based height maps are suitable for signal processing such as filtering, decomposition, and enhancement, while geometric quantities such as normal vectors and curvature must be calculated back to the original 3D point cloud. The two are linked through index mapping and neither can be omitted.

[0113] The three-level background suppression is cascaded, each controlling one scale segment. No single method can completely remove the background, and changing the order will cause mutual interference. This structure was determined after repeated experiments.

[0114] The point cloud feature-based method for precise extraction of minute defects in seamless steel pipes of the present invention uses a three-level background suppression series (macromorphic topography stripping, mesoscale undulation removal, and periodic texture notch filtering) to stably separate minute shallow defects (shallow scratches, blue lines, etc.) with a depth of 0.1mm to 0.3mm and a width of 0.2mm from the complex pipe surface background. This makes such defects, which are almost impossible to identify by traditional visual inspection and two-dimensional vision, detectable targets, and transforms the long-standing problem of missed detection in the industry into a controllable detection range.

[0115] The three-level background suppression method of the seamless steel pipe micro-defect accurate extraction method based on point cloud features of the present invention targets the steel pipe bending, ellipticity, local undulation (millimeter-level macroscopic deformation), roll marks and wavy lines (medium-scale undulations) and spiral rolling texture (periodic interference) step by step. Each method is used for a scale segment of each pipe and the serial order is optimized and determined, so that the algorithm is not sensitive to field conditions such as steel pipe specification changes, batch differences, and production line speed fluctuations, the false alarm rate is controllable, and the maintenance cost is low.

[0116] This invention presents a method for accurately extracting minute defects in seamless steel pipes based on point cloud features. This method constructs a complementary enhancement channel combining a height domain and a geometric domain, providing a dual chain of evidence. Based on the base map, height domain enhancement (DoG, Hessian) and geometric domain enhancement (normal deflection angle, curvature anomaly) are fused in parallel. The geometric channel directly utilizes the normal vector and curvature information of the original 3D point cloud, exhibiting significantly higher contrast than the height map for defects with extremely shallow depth but still deflected edges (such as shallow scratches and blue lines). This multi-dimensional fusion mechanism enables shallow defects to obtain a stable response on the saliency map, overcoming the fundamental problem of unreliable segmentation caused by weak single height signals.

[0117] This invention provides a method for accurately extracting minute defects in seamless steel pipes based on point cloud features, overcoming the limitations of point cloud sampling intervals and achieving sub-pixel-level boundary and depth measurements. A one-dimensional profile is taken along the normal direction at the boundary of the candidate defect region. After interpolation and curve fitting, the intersection of the fitted curve and the background horizontal plane is used as the sub-pixel edge point, achieving boundary positioning accuracy of 1 / 5 to 1 / 10 of the sampling interval. Simultaneously, local surface fitting is performed inside the defect, and the extreme value positioning depth resolution also reaches the sub-pixel level (0.02mm to 0.04mm). This allows for the accurate online quantification of the geometric dimensions of defects, such as length, width, area, and depth, for the first time, providing a reliable quantitative basis for subsequent defect classification and judgment.

[0118] The present invention provides a method for precise extraction of minute defects in seamless steel pipes based on point cloud features. This method employs a ring array laser contour sensor, which allows for full-circumferential point cloud acquisition by passing the steel pipe in a straight line along its axial direction in a single pass, eliminating blind spots. Point cloud preprocessing, three-level background suppression, enhancement fusion, segmentation, and sub-pixel extraction are all processed in a segmented, parallel flow. The time consumed per segment is lower than the lower limit of the production line conveyor cycle time, truly achieving "inspection while passing through." This method can directly replace high-intensity manual visual inspection, significantly improving inspection efficiency and consistency.

[0119] The method for precise extraction of minute defects in seamless steel pipes based on point cloud features of this invention ultimately outputs the circumferential angle, axial position, type (nine categories including holes, scratches, roll folds, external scars, blue lines, steel adhesion, cracks, pitting, and dents), length, width, and depth dimensions and confidence level of each defect. The results can be linked to marking, alarm, or sorting actuators to provide real-time data support for production line quality traceability and process optimization.

[0120] The gridded height map in the point cloud feature-based method for accurate extraction of minute defects in seamless steel pipes of this invention facilitates the use of mature two-dimensional image algorithms (filtering, wavelet, Fourier transform, etc.), while geometric quantities such as normal vectors and curvature are always calculated on the original three-dimensional point cloud. The two maintain a strict correspondence through index mapping, which not only ensures the convenience of signal processing, but also avoids the loss of geometric information caused by gridding. This is one of the substantial innovations of this technical solution that distinguishes it from conventional pure image or pure point cloud methods.

[0121] To further illustrate the specific implementation process of the method for accurate extraction of minute defects in seamless steel pipes based on point cloud features of the present invention, the following embodiments are provided for detailed explanation.

[0122] Example 1 This embodiment provides a method for accurately extracting minute defects in seamless steel pipes based on point cloud features. Specifically, it takes a hot-rolled seamless steel pipe production line as the implementation object. The seamless steel pipe has a specification of Φ114mm×8mm, a conveying speed of 1.5m / s, and passes through the measurement section in a straight line along the axial direction.

[0123] Eight sets of laser profile sensors are evenly distributed circumferentially across the measurement section, with 2048 profile points per set, a frame rate of 7.5kHz, corresponding to an axial sampling interval of 0.2mm and a circumferential point spacing of approximately 0.2mm. The overlap between adjacent fields of view is 8%. The position and pose of each sensor are calibrated using a standard bar, and the roller conveyor speed encoder outputs pulses to synchronously trigger data acquisition. Data is processed in 1m segments with 0.5m overlap sliding, using multi-threaded parallel processing, with the processing time for a single segment below the lower limit of the conveyor cycle time.

[0124] Step 1: Circular array point cloud acquisition.

[0125] Each frame acquired by eight laser contour sensors yields a circumferential contour line. The frame sequence is stitched together according to axial position to obtain a complete 3D point cloud of the entire outer surface of the steel pipe. The pose parameters of each sensor to a unified pipe coordinate system are calibrated using a standard rod. The point cloud is represented in the pipe coordinate system as follows: {Pi = (θi, zi, ri)}; Where θ is the circumferential angle, z is the axial position, and r is the radial measurement.

[0126] Step 2: Preprocessing and dual characterization construction.

[0127] First, the 3D point cloud data is preprocessed, including flying point removal and overlapping band fusion.

[0128] Flying point rejection is based on k-neighborhood statistics, where k is 16. If the distance from a point to its neighborhood centroid exceeds a set multiple, it is judged as a flying point and discarded to eliminate outliers caused by reflection and dust.

[0129] During overlap fusion, the repeated points in the overlapping area of ​​adjacent sensors are weighted and averaged according to their distances to the center of their respective sensors to eliminate seam steps.

[0130] Using the tube axis as a reference, the preprocessed 3D point cloud is resampled to a regular grid according to (θ, z), and inverse distance weighted interpolation is used with a grid spacing of 0.2mm × 0.2mm. The radial measurement r is used as the grid value to form a height map H(θ, z). An index mapping is established between the grid points and the original 3D point cloud for mutual value retrieval during subsequent geometric channel calculations.

[0131] Step 3: Three-level background suppression.

[0132] Combination Figure 3 and Figure 4 First-level background suppression: Sliding window robust reference surface fitting - input height map H(θ,z), output residual map R(θ,z).

[0133] A sliding window is opened along the axial direction, with a window length of 200 mm and an overlap of 100 mm. A local cylindrical surface within the window is fitted using RANSAC random sampling consistency, with an in-point threshold of 0.05 mm and a maximum iteration count of 200. Defect points are naturally excluded as outliers and do not contaminate the baseline. The baseline parameters of adjacent windows are smoothly transitioned using cubic spline interpolation to obtain a local baseline surface field Hbase(θ,z) that continuously varies with the pipe length. The residuals are then calculated point by point. R(θ,z)=H(θ,z)-Hbase(θ,z).

[0134] The bending, ellipticity, and slow whitishness of the entire tube are stripped away at this stage, and the residual amplitude is reduced to the sub-millimeter level.

[0135] Combination Figure 4 and Figure 5 Second-level background suppression: Multi-scale decomposition—input residual map R(θ,z), output mid-to-high frequency residual map R2(θ,z).

[0136] Two-dimensional wavelet decomposition (db4 wavelet, 4-level decomposition) is performed on R(θ,z). The low-frequency approximate subband, corresponding to the mesoscale background (undulations of several millimeters to tens of millimeters in wavelength, such as roller prints and wavy lines), is directly discarded; the mid-to-high frequency detail subbands with concentrated defect energy are reconstructed to obtain R2(θ,z). The mesoscale components such as roller prints and wavy lines are suppressed by more than an order of magnitude.

[0137] Combination Figure 5 and Figure 6 Third-level background suppression: frequency domain high-pass and adaptive notch filtering - input mid-to-high frequency residual map R2(θ,z), output base map B(θ,z).

[0138] A two-dimensional Fourier transform is performed on R2. First, a high-pass filter is used to suppress residual low frequencies, with the cutoff wavelength set to 30 mm based on the maximum characteristic scale of the defect. Then, the position of the spectral peaks is automatically detected, and a narrow-band notch filter is applied to the fundamental frequency and the first three harmonics. After inverse transform, the background suppression map B(θ,z) is obtained. On the output map B, defects with a depth of 0.1 mm are already higher than the residual background.

[0139] Step 4: Multidimensional enhancement of minute defects.

[0140] Based on the base map B and the original 3D point cloud, complementary height domain enhancement channels and geometric domain enhancement channels are used for enhancement, and then weighted fusion is performed.

[0141] See Figure 7 DoG (Dot Channel): Employs multi-scale Gaussian difference with σ at three levels: 0.15mm, 0.3mm, and 0.6mm, to enhance the response of clustered targets such as holes, pits, scars, and steel adhesion.

[0142] Linear channel (Hessian): Based on the linear structural response of the Hessian matrix eigenvalues, the maximum response is obtained point-by-point after filtering at 3 scales and 12 directions. Scratches, blue lines, and cracks, regardless of their direction, will always be captured by the filter in a certain direction.

[0143] See Figure 8 The geometric domain enhancement channel estimates the normal vector and principal curvature of each point on the original 3D point cloud (normal vector estimation is performed using a neighborhood radius of 3 points), constructing a normal deflection angle map N(θ,z) and a curvature anomaly map K(θ,z). Even a defect with a depth of only 0.1mm exhibits a deflection of several to over ten degrees relative to the pipe surface normal on its edge slope, resulting in a significantly higher contrast on the geometric map compared to the height map.

[0144] Fusion: After normalizing each of the three channels, a weighted summation is performed with weights of 0.30 / 0.30 / 0.25 / 0.15, resulting in the defect saliency map S. S=w1·DoG(B)+w2·Line(B)+w3·N′+w4·K′.

[0145] Step 5: Adaptive segmentation of candidate defect regions.

[0146] The defect saliency map S is divided into 32×32 grid blocks. For each block, a threshold is adaptively determined by adding k times the local mean and standard deviation (k is set to 3.5). After binarization, connected components are labeled, and morphological opening and closing operations are performed to clean up fragments and fill holes (opening operation kernel 3×3). Finally, the area is capped at 0.2 mm. 2 Isolated noise points are filtered out using a length lower limit of 0.5 mm to obtain a set of candidate defect regions.

[0147] Step 6: Subpixel level extraction.

[0148] See Figure 9 Boundary extraction: For each candidate defect region, a one-dimensional profile is taken on the base image B along the normal direction of each point on the boundary. The profile is then densified to 8 times the density of the original mesh using bicubic interpolation. A quadratic curve is fitted to the edge segment of the profile, and the intersection point of the fitted curve with the local background horizontal plane is obtained as the sub-pixel edge point in that normal direction. This process is repeated for all normal directions, and the resulting edge points are connected to form a closed sub-pixel boundary.

[0149] See Figure 10 Depth measurement: A local quadratic surface is fitted to the base image B within the candidate defect region. The extreme points and extreme values ​​of the fitted surface are taken as the peak depth (for depressions) or peak height (for protrusions) of the defect. The depth resolution reaches 0.02mm~0.04mm (i.e., 1 / 5~1 / 10 of the sampling interval), no longer limited by the grid spacing.

[0150] At this point, the length, width, area, and depth of the defect are all output with sub-pixel precision.

[0151] Step 7: Feature calculation and identification of nine types of defects.

[0152] For each defect region, calculate a set of geometric and morphological features: area, major axis, minor axis, aspect ratio, orientation angle (relative to the tube axis), peak depth (or height), volume, average edge slope (statistical normal deflection angle), residual sign composition (depression / protrusion / mixture), and internal texture density.

[0153] See Figure 12 The diagram shows a physical structure of an electronic device, such as... Figure 12 As shown, the electronic device may include a processor 210, a communication interface 220, a memory 230, and a communication bus 240. The processor 210, communication interface 220, and memory 230 communicate with each other via the communication bus 240. The processor 210 can call computer programs stored in the memory 230 to execute steps of a method for accurately extracting minute defects in seamless steel pipes based on point cloud features.

[0154] Furthermore, the logical instructions in the aforementioned memory 230 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0155] On the other hand, this application embodiment also provides a storage medium, which is a computer-readable storage medium storing a computer program. The computer program is used to cause a processor to execute the steps of the method for accurate extraction of minute defects in seamless steel pipes based on point cloud features provided in the above embodiments.

[0156] The computer-readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic storage (e.g., floppy disk, hard disk, magnetic tape, magneto-optical disk (MO)), optical storage (e.g., CD, DVD, BD, HVD), and semiconductor storage (e.g., ROM, EPROM, EEPROM, non-volatile memory (NAND FLASH), solid-state drive (SSD)).

[0157] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0158] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0159] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for accurately extracting minute defects in seamless steel pipes based on point cloud features, characterized in that, Includes the following steps: A three-dimensional point cloud of the circumferential surface of a seamless steel pipe is acquired, and the three-dimensional point cloud is transformed into the pipe coordinate system to obtain a point set represented by circumferential angle, axial position and radial measurement. The point set is resampled along the circumferential and axial directions with the tube axis as a reference to form a gridded height map. The values ​​of the grid points in the height map are the radial measurements, and the grid points form an index mapping with the three-dimensional point cloud. The first, second and third levels of background suppression are sequentially performed on the height map to obtain the base map. The first level of background suppression uses a sliding window to robustly fit the local reference surface and take the residual to remove the macroscopic morphology. The second level of background suppression uses multi-scale decomposition to discard low-frequency background components to remove the mesoscale undulations. The third level of background suppression uses frequency domain high-pass filtering and adaptive notch filtering to remove periodic textures. A height domain enhancement channel is constructed based on the base map, and a geometry domain enhancement channel is constructed based on the 3D point cloud. The response of the geometry domain enhancement channel is mapped to the base map and weighted and fused with the response of the height domain enhancement channel to obtain a defect saliency map. Adaptive threshold segmentation is performed on the defect saliency map to obtain candidate defect regions; Subpixel boundary localization is performed on a one-dimensional cross-section along the normal direction of the candidate defect region boundary, and subpixel depth localization is performed within the candidate region to output the subpixel level boundary and depth of the minute defect; The geometric features of the minute defect are obtained by calculating the subpixel boundary and depth, and the type of the minute defect is determined based on the geometric features.

2. The method for accurate extraction of minute defects in seamless steel pipes based on point cloud features according to claim 1, characterized in that, The height-domain enhancement channel includes a speckle enhancement sub-channel and a linear enhancement sub-channel; Based on the base map, highly correlated features are extracted. The speckle enhancement sub-channel uses a multi-scale Gaussian difference operator to filter the highly correlated features and outputs the speckle response. The linear enhancement sub-channel uses the Hessian matrix operator to perform multi-directional filtering on the highly correlated features, outputting a linear response.

3. The method for accurate extraction of minute defects in seamless steel pipes based on point cloud features according to claim 2, characterized in that, The geometric domain enhancement channel includes a normal deflection angle sub-channel and a curvature anomaly sub-channel; The normal deflection angle sub-channel outputs the deflection angle response based on the spatial angle between the normal vector of each point in the three-dimensional point cloud and the normal vector of the ideal tube surface. The curvature anomaly subchannel outputs a curvature anomaly response based on the degree of deviation between the principal curvature of each point in the three-dimensional point cloud and the ideal pipe surface curvature.

4. The method for accurate extraction of minute defects in seamless steel pipes based on point cloud features according to claim 1, characterized in that, When performing the first-level background suppression, a sliding window is opened along the axial direction. The local reference surface is obtained by fitting the RANSAC random sampling consensus within the sliding window. Defect points are excluded as outliers. The reference parameters of two adjacent sliding windows are smoothly transitioned by interpolation. The residual map is obtained by taking the residual map of the height map relative to the local reference surface point by point.

5. The method for accurate extraction of minute defects in seamless steel pipes based on point cloud features according to claim 4, characterized in that, When performing the second-level background suppression, the residual map is decomposed using two-dimensional wavelet decomposition, the low-frequency approximate subband is discarded, and the mid-to-high frequency map is obtained by reconstructing the mid-to-high frequency map using the mid-to-high frequency detail subband.

6. The method for accurate extraction of minute defects in seamless steel pipes based on point cloud features according to claim 5, characterized in that, When performing the third-level background suppression, a two-dimensional Fourier transform is performed on the mid-to-high frequency image, and after high-pass filtering, the position of the spectral peak is automatically detected. Narrowband notch filtering is performed on the fundamental frequency and harmonics to obtain the background image.

7. The method for accurate extraction of minute defects in seamless steel pipes based on point cloud features according to claim 1, characterized in that, The step of locating sub-pixel boundaries on a one-dimensional profile along the normal direction of the candidate defect region boundary includes: The one-dimensional cross-section is taken on the base map along the normal direction of each point along the boundary of the candidate defect region; After interpolation and encryption, the one-dimensional profile is subjected to curve fitting on the edge segment. The intersection of the fitted curve and the local background horizontal is taken as the sub-pixel edge point. Traverse all points of the candidate defect region boundary and connect all the sub-pixel edge points to form a closed sub-pixel boundary.

8. The method for accurate extraction of minute defects in seamless steel pipes based on point cloud features according to claim 7, characterized in that, The step of performing subpixel depth localization within the candidate region includes performing local surface fitting on the base image within the candidate defect region, and taking the extreme value of the fitted surface as the subpixel depth.

9. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method for accurate extraction of minute defects in seamless steel pipes based on point cloud features as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method for accurate extraction of minute defects in seamless steel pipes based on point cloud features as described in any one of claims 1 to 8.

Images