An ultra-thin strip plate type online detection and reconstruction method and system
By combining an imaging acquisition device and a semantic segmentation network on ultra-thin strips, the problem of laser line breakage caused by high reflectivity was solved, and high-precision automated extraction of strip shape data was achieved, improving yield and production efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TAIYUAN UNIVERSITY OF SCIENCE AND TECHNOLOGY
- Filing Date
- 2026-05-28
- Publication Date
- 2026-06-23
AI Technical Summary
In existing technologies, the high reflectivity of ultra-thin strips causes laser lines to break, making it impossible to accurately extract strip shape data, resulting in low yield and low accuracy due to reliance on manual repair.
The laser contour image of the strip surface is acquired laterally by an imaging acquisition device. Denoising and repair are performed using a pre-trained semantic segmentation network to generate a laser binarization mask. The laser center coordinates are calculated by selecting an appropriate algorithm to reconstruct the three-dimensional point cloud data. Full-width geometric topology analysis is then performed to extract the waveform and wavy components of the plate surface.
It improved the accuracy of plate shape detection, solved the problem of laser line breakage caused by high reflectivity, realized automated high-precision plate shape data extraction, and improved production efficiency.
Smart Images
Figure CN122265290A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of strip shape detection, and in particular to a method and system for online detection and reconstruction of ultra-thin strip shape. Background Technology
[0002] The definition of "ultra-thin strip" may vary across different fields. For example, the national standard GB / T 15574—2016, "Classification of Steel Products," broadly classifies plate and strip steel into thin plates (less than 3 mm) and thick plates (not less than 3 mm). In scientific research and production, plate and strip steel is classified into five categories based on thickness: ultra-thin strip, thin plate, medium plate, thick plate, and extra-thick plate. In this field, steel plates with a thickness of less than 0.2 mm are referred to as ultra-thin strip (this embodiment is particularly applicable to strip materials with a thickness of 0.1 mm and less than 0.1 mm; unless otherwise specified in the following embodiments, ultra-thin strip refers to strip materials with a thickness of less than or equal to 0.1 mm), steel plates with a thickness of 0.2-3 mm are referred to as thin plates, steel plates with a thickness of 3-20 mm are referred to as medium plates, steel plates with a thickness of 20-60 mm are referred to as thick plates, and steel plates with a thickness of more than 60 mm are referred to as extra-thick plates.
[0003] With the development of technology, the application of ultra-thin strips (thickness typically less than or equal to 0.2 mm, especially 0.1 mm) is becoming increasingly widespread, making online inspection of the strip shape particularly important for quality. For example, in the production of ultra-thin stainless steel strips (hand-tearable steel) with a thickness of 0.02 mm to 0.1 mm.
[0004] High reflectivity causes broken lines: The surface of stainless steel is as smooth as a mirror. Even with adjustments to the camera exposure, minute undulations or bends in certain areas can still cause instantaneous specular reflections to directly enter the lens, resulting in localized pixel saturation of the sensor. In overexposed areas, the grayscale information of the laser lines is lost, manifesting as "broken" laser lines in the image.
[0005] The ultimate goal of strip shape inspection is to obtain various waveform data. However, high reflectivity can cause laser lines to break and become unusable, resulting in inaccurate waveform data. To address this issue, current ultra-thin strip production relies heavily on manual experience to repair breaks. This manual judgment method has low accuracy and impacts production efficiency. Summary of the Invention
[0006] This application provides an online detection and reconstruction method and system for ultra-thin strips, which at least solves the problem of low yield caused by high reflectivity leading to laser line breakage and inability to extract in the prior art.
[0007] According to one aspect of this application, an online detection and reconstruction method for ultra-thin strip profiles is provided, comprising: acquiring a laser contour image of the strip surface laterally via an imaging acquisition device; inputting the acquired laser contour image into a pre-trained semantic segmentation network to obtain a laser binarization mask, wherein the semantic segmentation network is used to denoise the laser contour image and repair laser stripes broken by reflection in the denoised laser contour image to generate a continuous laser binarization mask; within the constraints of the laser binarization mask, selecting a corresponding algorithm based on the local grayscale distribution characteristics and signal-to-noise ratio of the laser stripes, and calculating and extracting the laser center coordinates based on the selected algorithm; reconstructing three-dimensional point cloud data of the strip surface based on the extracted laser center coordinates; performing full-width geometric topology analysis on the reconstructed three-dimensional point cloud data to obtain the strip surface waveform; decomposing the strip surface waveform into wave components of different orders, and calculating the differential elongation and wave height distribution vector distributed along the width direction of the strip.
[0008] Furthermore, the imaging acquisition device includes: a line laser projector; two industrial cameras arranged in a horizontal side-axis layout on the left and right peripheries of the strip running direction, with the angle between the camera optical axis and the horizontal plane set to 20° to 60°; and an orthogonal polarization filter assembly, which includes a polarizer disposed at the line laser projector and an analyzer disposed in front of the lens of the industrial camera, wherein the polarization directions of the polarizer and the analyzer are adjusted to a 90° orthogonal state.
[0009] Furthermore, the line laser projector is a blue-violet laser.
[0010] Furthermore, the denoising process for the laser contour image includes: perceiving the geometric shape of the light clusters in the laser contour image through multi-scale convolution kernels, identifying asymmetric false halo layers as background, and retaining only the energy skeleton area that conforms to the overall line trend; and using context awareness, marking isolated bright spots that lack linear features extending along the width direction of the strip as interference signals and removing them.
[0011] Furthermore, the repair of laser stripes broken by reflection in the denoised laser contour image includes: for the physical break of the laser stripes, based on the curvature trend and slope vector of the laser stripes on both sides of the break point, predicting the missing path at the image semantic level and generating a connecting bridge, which is used to repair the laser stripes.
[0012] Furthermore, based on the local grayscale distribution characteristics and signal-to-noise ratio of the laser stripes, at least one of the following algorithms is selected for use: number domain Gaussian fitting algorithm, three-point parabolic fitting algorithm, intensity-weighted centroid algorithm, and center positioning algorithm.
[0013] Furthermore, the corresponding algorithm is selected based on the local gray-scale distribution characteristics and signal-to-noise ratio of the laser stripe, including: when the gray-scale distribution of the light stripe cross section presents a complete bell-shaped curve and is not saturated, a three-point logarithmic domain Gaussian fitting algorithm is used; when the top of the light stripe cross section is flat or computational resources are limited, a three-point parabolic fitting algorithm is used; when specular reflection is detected, causing continuous pixel saturation, the algorithm is switched to intensity-weighted centroid algorithm or edge gradient-based center localization algorithm.
[0014] According to another aspect of the embodiments of this application, an online detection and reconstruction system for ultra-thin strip profiles is also provided, wherein the imaging acquisition device includes: a line laser projector; two industrial cameras arranged in a horizontal side-axis layout on the left and right peripheries of the strip running direction, the angle between the camera optical axis and the horizontal plane being set to 20° to 60°; an orthogonal polarization filter assembly, including a polarizer disposed at the line laser projector and an analyzer disposed in front of the lens of the industrial cameras, wherein the polarization directions of the polarizer and the analyzer are adjusted to a 90° orthogonal state; and the data processing unit is used to execute the above-described method.
[0015] According to another aspect of this application, an online detection and reconstruction system for ultra-thin strip profiles is also provided, comprising: an imaging acquisition module for acquiring laser contour images of the surface of an ultra-thin strip in operation; an image processing module for preprocessing the laser contour images and completing the laser stripe regions affected by reflection interference, noise interference, or stripe breakage to generate a binary mask corresponding to the laser stripes; a center extraction module for extracting the laser center coordinates based on local image features of the laser stripes within the area defined by the binary mask; a three-dimensional reconstruction module for reconstructing three-dimensional point cloud data of the strip surface based on binocular vision calibration parameters and the laser center coordinates; a strip profile calculation module for calculating the transverse strip profile curve based on the three-dimensional point cloud data and outputting strip profile parameters including wave height distribution, differential elongation, and / or wave shape components; and a control interface module for sending the strip profile parameters to the rolling mill strip profile control system.
[0016] According to another aspect of the embodiments of this application, a readable storage medium is also provided, on which computer instructions are stored, wherein the computer instructions, when executed by a processor, implement the above-described method steps.
[0017] According to another aspect of the embodiments of this application, a computer program product is also provided, which stores computer instructions, wherein the computer instructions, when executed by a processor, implement the above-described method steps.
[0018] In this embodiment, a laser contour image of the strip surface is acquired laterally along the side axis using an imaging acquisition device. The acquired laser contour image is input into a pre-trained semantic segmentation network to obtain a laser binarization mask. The semantic segmentation network is used to denoise the laser contour image and repair laser stripes broken by reflection in the denoised laser contour image, generating a continuous laser binarization mask. Within the constraints of the laser binarization mask, a corresponding algorithm is selected based on the local grayscale distribution characteristics and signal-to-noise ratio of the laser stripes. The laser center coordinates are calculated and extracted using the selected algorithm. Three-dimensional point cloud data of the strip surface is reconstructed based on the extracted laser center coordinates. Full-width geometric topology analysis is performed on the reconstructed three-dimensional point cloud data to obtain the plate surface waveform. The plate surface waveform is decomposed into wave components of different orders, and the differential elongation and wave height distribution vector distributed along the strip width direction are calculated. This application solves the problem of low yield caused by high reflectivity leading to broken laser lines that cannot be extracted in the prior art, thereby improving accuracy. Attached Figure Description
[0019] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are illustrative, and the parts and elements in the drawings are not necessarily drawn to scale. The accompanying drawings, which constitute a part of this application, are used to provide a further understanding of this application, and the illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application. In the drawings:
[0020] Figure 1 This is a system hardware structure diagram according to an embodiment of this application;
[0021] Figure 2 This is a top view of the camera placement position according to an embodiment of this application;
[0022] Figure 3 This is a main flowchart of the method according to an embodiment of this application;
[0023] Figure 4 This is a system control logic block diagram according to an embodiment of this application;
[0024] Figure 5 This is a flowchart of online detection and reconstruction of ultra-thin strip profile according to an embodiment of this application;
[0025] Figure 6 This is a schematic diagram of the point cloud reconstruction result according to an embodiment of this application;
[0026] Figure 7 This is a schematic diagram of the plate type parameter calculation results according to an embodiment of this application. Detailed Implementation
[0027] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. It should also be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowcharts, in some cases, the steps shown or described may be performed in a different order than that shown here.
[0028] The technical terms used in the following embodiments will be explained first.
[0029] A mask, used as a master for image transfer, functions similarly to a photographic negative. A binarization mask is a binary image of the same size as the original image, with pixel values typically between 0 and 1. In the mask, pixels with a value of 1 represent regions of interest, while pixels with a value of 0 represent regions to be ignored.
[0030] Subpixel resolution refers to a resolution smaller than a single pixel in image processing and display technology. It involves several aspects: Subpixel concept: The smaller units between two physical pixels are called subpixels; this concept is derived through computational methods. Subpixel precision: The subdivision between adjacent pixels, typically achieved through interpolation algorithms to improve image detail and quality. Applications in machine vision: In machine vision systems, subpixel precision can capture finer image features, thereby improving the accuracy of recognition and analysis.
[0031] Subpixel extraction is a technique in image processing designed to improve image resolution and edge detection accuracy. Its basic concept is to further subdivide pixels into smaller units, thereby achieving more precise image analysis.
[0032] Laser spot center point positioning is a technique that uses the light emitted by a laser to form a spot on a receiving plane, and then uses an algorithm to calculate its precise center position.
[0033] A 3D point cloud is a data structure used to represent objects or scenes in three-dimensional space, consisting of a large number of 3D coordinate points (X, Y, Z). These points are obtained through discrete sampling of the object's surface or scene, and can accurately describe the object's geometry and spatial position. In addition to coordinate information, point clouds may also contain additional attributes such as color, intensity, and time.
[0034] Geometric topological analysis is a branch of mathematics that studies the shape and structure of space and its invariant properties under continuous transformations, involving core concepts such as topological spaces and manifolds.
[0035] Elongation is a key indicator for measuring a material's ability to elastically deform under tension.
[0036] Bending force refers to the force applied to the rolls during the rolling process, either hydraulically or mechanically, to change the shape of the rolls and thus control the geometry and quality of the rolled product. The application of bending force can effectively improve the crown and straightness of the strip, ensuring that the product quality meets standards.
[0037] Rolled texture refers to raised or recessed patterns or lines created on the surface of metal materials through a rolling process. This texture not only enhances the visual appeal of the metal but also improves its physical properties, such as wear resistance and corrosion resistance. Rolled textures are widely used in various industries, including construction, automotive, and home appliances, to improve the appearance and durability of products.
[0038] To address the problems in the prior art, the following embodiments provide a method for online detection and reconstruction of ultra-thin strip profiles. Figure 5 This is a flowchart of an online detection and reconstruction method for ultra-thin strip profiles according to an embodiment of this application, such as... Figure 5 As shown, the steps in this method are explained below.
[0039] Step S502: The laser contour image of the strip surface is acquired laterally by the imaging acquisition device.
[0040] In this step, a preferred device is used for optical interference resistance. The imaging acquisition device includes: a line laser projector; two industrial cameras arranged in a horizontal side-axis layout on the left and right sides of the strip running direction, with the angle between the camera optical axis and the horizontal plane set to 20° to 60°; and an orthogonal polarization filter assembly, which includes a polarizer located at the line laser projector and an analyzer located in front of the industrial camera lens, wherein the polarization directions of the polarizer and the analyzer are adjusted to be orthogonal at 90°.
[0041] To further reduce interference, the line laser projector is a blue-violet laser.
[0042] Step S504: Input the acquired laser contour image into a pre-trained semantic segmentation network to obtain a laser binarization mask. The semantic segmentation network is used to denoise the laser contour image and repair the laser stripes in the denoised laser contour image that are broken due to reflection, thereby generating a continuous laser binarization mask.
[0043] Denoising can be achieved through the following steps: First, by using multi-scale convolutional kernels to perceive the geometric shape of light clusters in the laser contour image, asymmetric false halo layers are identified as background, retaining only the energy skeleton region that conforms to the overall line trend. Second, using context awareness, isolated bright spots lacking linear features extending along the strip width are marked as interference signals and removed. Other denoising methods can also be used, which will not be elaborated upon here.
[0044] The repair can be performed as follows: For the physical break in the laser stripe, based on the curvature trend and slope vector of the laser stripes on both sides of the break point, predict the missing path at the image semantic level and generate a connecting bridge. This connecting bridge is used to repair the laser stripe. Other repair methods can also be used; the appropriate method should be selected based on the actual situation.
[0045] Step S506: Within the constraint range of the laser binarization mask, select the corresponding algorithm based on the local grayscale distribution characteristics and signal-to-noise ratio of the laser stripes, and calculate and extract the laser center coordinates based on the selected algorithm.
[0046] There are several algorithms available, such as the number domain Gaussian fitting algorithm, the three-point parabola fitting algorithm, the intensity-weighted centroid algorithm, and the center localization algorithm.
[0047] In one optional example, when the grayscale distribution of the light stripe cross section presents a complete bell-shaped curve and is not saturated, a three-point logarithmic domain Gaussian fitting algorithm is used; when the top of the light stripe cross section is flat or computational resources are limited, a three-point parabolic fitting algorithm is used; when specular reflection is detected, causing continuous pixel saturation, the algorithm is switched to intensity-weighted centroid algorithm or edge gradient-based center localization algorithm.
[0048] Step S508: Reconstruct the three-dimensional point cloud data of the strip surface based on the extracted laser center coordinates.
[0049] Step S510: Perform full-width geometric topology analysis on the reconstructed three-dimensional point cloud data to obtain the plate surface waveform; decompose the plate surface waveform into wave components of different orders, and calculate the differential elongation and wave height distribution vector distributed along the width direction of the strip.
[0050] The above steps solve the problem of low yield caused by laser line breakage due to high reflectivity in existing technologies, thereby improving accuracy.
[0051] Corresponding to the above method, an online detection and reconstruction system for ultra-thin strip profiles is also provided, comprising: an imaging acquisition device and a data processing unit, wherein the imaging acquisition device includes: a line laser projector; two industrial cameras arranged in a horizontal side-axis layout on the left and right peripheries of the strip's running direction, with the angle between the camera optical axis and the horizontal plane set to 20° to 60°; and an orthogonal polarization filter assembly, comprising a polarizer disposed at the line laser projector and an analyzer disposed in front of the industrial camera lens, wherein the polarization directions of the polarizer and the analyzer are adjusted to a 90° orthogonal state; the data processing unit is used to execute the above method. Simultaneously, a readable storage medium is also provided, storing computer instructions thereon, wherein when executed by a processor, the computer instructions implement the above method steps. A computer program product is also provided, storing computer instructions thereon, wherein when executed by a processor, the computer instructions implement the above method steps.
[0052] In this embodiment, a software system is also provided to implement the functions of the data processing unit. The system includes:
[0053] The imaging acquisition module is used to acquire laser contour images of the surface of an ultra-thin strip during operation;
[0054] The image processing module is used to preprocess the laser contour image and complete the laser stripe region affected by reflection interference, noise interference or stripe breakage to generate a binarized mask corresponding to the laser stripe.
[0055] The center extraction module is used to extract the laser center coordinates based on the local image features of the laser stripes within the area defined by the binarized mask.
[0056] The 3D reconstruction module is used to reconstruct the 3D point cloud data of the strip surface based on the binocular vision calibration parameters and the laser center coordinates.
[0057] The strip profile calculation module is used to calculate the transverse strip profile curve based on the three-dimensional point cloud data, and output the strip profile parameters such as wave height distribution, differential elongation and / or wave shape components.
[0058] The control interface module is used to send the plate shape parameters to the rolling mill plate shape control system.
[0059] The following describes one method provided in this embodiment. It should be noted that some steps in this method are optional.
[0060] Step S1: Acquire a laser profile image of the strip surface using an imaging acquisition system. This step employs lateral side-axis acquisition.
[0061] Optionally, in this step, an anti-reflective imaging strategy can be employed to better acquire the laser contour image. The anti-reflective imaging acquisition device adopts a wavelength-polarization-geometric three-dimensional collaborative anti-reflective strategy, which includes the selection of the laser projector and the configuration of the industrial camera, etc. The anti-reflective strategy is illustrated below with an example.
[0062] The following scheme is adopted in this strategy.
[0063] For laser projectors, a line laser projector can be selected, and a blue-violet laser with a wavelength of 405nm is preferred.
[0064] Two industrial cameras are arranged in a horizontal side-axis layout on the left and right sides of the strip's running direction, with the angle between the camera's optical axis and the horizontal plane set to 20° to 60°.
[0065] The orthogonal polarization filter assembly includes a polarizer disposed at the line laser projector and an analyzer disposed in front of the industrial camera lens, wherein the polarization directions of the polarizer and the analyzer are adjusted to be orthogonal at 90°.
[0066] In this example, the anti-reflective imaging acquisition system includes the above-mentioned components: a line laser projector, two industrial cameras, and an orthogonal polarization filter assembly. By constructing the anti-reflective imaging acquisition system, the two industrial cameras are respectively positioned on the left and right peripheries of the ultra-thin strip's running direction. Using a side-axis perspective in conjunction with the line laser projector, the laser contour image of the strip surface is acquired.
[0067] The above example employs a three-dimensional collaborative anti-reflective strategy of "wavelength-polarization-geometry" in the anti-reflective imaging acquisition system. Specifically, the light source wavelength optimization uses a 405nm blue-violet laser instead of a traditional red laser, leveraging the stronger scattering efficiency of the short-wavelength beam on the metal's microscopic surface and its better penetration of the oil film to improve the signal-to-noise ratio of the diffuse reflection signal. Orthogonal polarization filtering involves setting a polarizer at the laser projection end and an analyzer at the camera imaging end, adjusting their polarization directions to 90° orthogonal. Utilizing the physical property that specular reflection maintains its polarization state while diffuse reflection undergoes depolarization, strong specular reflection light entering the lens is physically blocked. The side-axis spatial layout adopts a transverse side-axis arrangement, placing the optical axes of the two cameras on either side of the strip's running direction, with the angle between the camera optical axes and the horizontal plane set to 20° to 60°. By utilizing the anisotropic scattering characteristics of the metal surface, the system completely avoids the strong specular reflection cones distributed along the rolling texture direction, picking up only high-quality lateral scattered light signals.
[0068] Step S2: Input the acquired laser contour image into a pre-trained semantic segmentation network, which is called a laser stripe semantic segmentation network. The laser stripe semantic segmentation network is used to denoise the laser contour image and repair the broken laser lines in the denoised laser contour image due to reflection, generating a continuous laser binarization mask.
[0069] In this step, pixel-level semantic analysis is performed on the laser contour image by segmentation to identify and remove background noise. At the same time, the laser lines that are broken due to reflection are intelligently repaired to generate a continuous laser binarization mask. Therefore, this step is called the neural network signal enhancement step.
[0070] The laser stripe semantic segmentation network is based on an encoder-decoder architecture. Its intelligent repair and denoising steps include: halo stripping: using multi-scale convolutional kernels to perceive the geometric shape of light clusters in the laser contour image, identifying asymmetric false halo layers as background, and retaining only the energy skeleton area that conforms to the overall line trend; semantic denoising: using context awareness, isolated bright spots that lack linear features extending along the strip width direction are marked as interference signals and removed; intelligent fracture repair: for physical fractures of laser lines, based on the curvature trend and slope vector of the laser lines on both sides of the fracture point, the missing path is predicted at the image semantic level and a connecting bridge is generated.
[0071] In the above steps, halo stripping targets the bright spots caused by local specular reflection on the strip surface. The network is trained to learn the core features of the Gaussian distribution and topological continuity of laser lines. The network does not rely solely on pixel brightness, but perceives the geometry of the light spots through multi-scale convolutional kernels. It identifies the divergent and asymmetrical "false halo layer" at the edge of the light spot as background, retaining only the "energy skeleton region" inside the light spot that conforms to the overall line trend. This allows for the extraction of sub-pixel level effective signal paths from large overexposed spots.
[0072] Semantic denoising leverages the network's context-aware capabilities to perform morphological discrimination on bright spots in the image. Isolated bright noise points caused by surface oil stains or scratches, lacking linear features extending along the strip width, are marked as interference signals and removed by the network to prevent subsequent extraction algorithms from being influenced by noise.
[0073] Intelligent fracture repair addresses physical fractures of laser lines caused by abrupt changes in the mirror reflection angle. The network predicts the missing path at the semantic level of the image based on the curvature trend and slope vector of the laser lines on both sides of the fracture point, generating continuous mask connection bridges to ensure that subsequent 3D reconstruction data is not interrupted.
[0074] The final output laser binarization mask is designed as a compact narrowband mask. This mask acts as a spatial filter, forcing the subsequent sub-pixel extraction algorithm to perform calculations only within the "skeleton" of the laser line, thus completely shielding the interference of the light cluster edge halo on the center positioning accuracy from the algorithm level.
[0075] Step S3: Within the constraint range of the laser binarization mask, based on the local grayscale distribution characteristics and signal-to-noise ratio of the laser stripes, an adaptive sub-pixel extraction operator is selected to calculate the laser center coordinates. This step is called adaptive multi-strategy sub-pixel extraction.
[0076] In this step, there are many methods for adaptively selecting sub-pixel extraction operators. For example, adaptive selection of sub-pixel extraction operators includes: when the gray-scale distribution of the light stripe cross section presents a complete bell-shaped curve and is not saturated, a three-point logarithmic domain Gaussian fitting algorithm is used; when the top of the light stripe cross section is flat or computational resources are limited, a three-point parabolic fitting algorithm is used; when specular reflection is detected, causing continuous pixel saturation, the method is switched to intensity-weighted centroid method or edge gradient-based center localization method.
[0077] This step, within the constraints of the laser binarization mask, performs multiple rounds of iterative optimization extraction, from coarse positioning to fine refinement. In each round of calculation, the system evaluates the local grayscale distribution characteristics and signal-to-noise ratio of the laser stripes in real time, dynamically adapts the optimal sub-pixel extraction operator (including Gaussian fitting, parabolic fitting, centroid method, or edge method), and integrates intelligent smoothing and continuity constraint mechanisms. In this way, the laser center coordinates with micrometer-level accuracy can be calculated.
[0078] The following example illustrates multiple iterations.
[0079] First, pixel-level coarse extraction is performed within the mask area to quickly locate the energy ridge of the light stripe. Then, a dynamic search window is constructed based on this, and bilateral filtering or morphological filtering is performed on the original grayscale data within the window to suppress speckle noise on the metal surface while preserving the edge gradient, providing a clean data source for the next round of fine extraction.
[0080] Morphological evaluation is performed on the filtered light stripe cross-section, and the extraction algorithm is automatically switched based on the evaluation results: when the cross-sectional gray-level distribution presents a complete bell-shaped curve and is unsaturated, a three-point logarithmic domain Gaussian fitting is used, and the nonlinear Gaussian model is linearized through logarithmic transformation, while the center position and peak width parameters are estimated. To achieve the highest accuracy in extraction; when the top of the cross section is slightly flat or computational resources are limited, a three-point parabolic fitting is used to obtain the sub-pixel center with extremely low computational cost; when strong specular reflection causes continuous pixel saturation, the Gaussian model fails, and the system automatically switches to the intensity-weighted centroid method or the edge gradient-based center positioning method, using the gradient symmetry of the light spot edge to infer the geometric center and prevent data from flying away.
[0081] In the above example, the traditional "extract first, then filter" approach is abandoned, and online smoothing is adopted in the extraction stage. A continuity constraint function is introduced, and the trend of neighboring points is referenced when calculating the coordinates of the current point. Abrupt jumps that violate the continuity assumption are penalized or corrected, thereby directly outputting a smooth, continuous, and uninterrupted laser center curve at the algorithm level.
[0082] Step S4: Based on the principle of binocular stereo vision and combined with system calibration parameters, the extracted laser center coordinates are converted into three-dimensional point cloud data of the strip surface. This step is called three-dimensional point cloud reconstruction.
[0083] For example, "3D point cloud reconstruction" is performed based on the principle of binocular stereo vision. Epipolar correction is applied to the laser center points extracted by the first and second cameras using pre-calibrated binocular intrinsic and extrinsic parameters. Based on epipolar constraints, source image points are matched, and spatial coordinates are calculated using triangulation principles. Subsequently, the point clouds reconstructed by the left and right cameras were uniformly transformed to the world coordinate system, and the overlapping field of view was weighted and fused to eliminate the blind spots that may exist in a single observation, and synthesize a complete cross-sectional profile covering the full width of the strip.
[0084] Figure 6 This is a schematic diagram of the point cloud reconstruction result according to an embodiment of this application. Figure 6 The coordinate unit in the text is mm. Figure 6 The reconstructed point cloud data is shown.
[0085] Figure 7 This is a schematic diagram of the plate shape parameter calculation results according to an embodiment of this application. Figure 7 The curve involves the peak curvature, peak waveform, and valley waveform of the left and right edges.
[0086] Step S5: Perform full-width geometric topology analysis on the reconstructed 3D point cloud data, decompose the plate waveform into wave components of different orders, and calculate the differential elongation and wave height distribution vector distributed along the strip width direction. The waveform data includes: wave components, differential elongation, and wave height distribution vector. This step is called plate shape parameter calculation.
[0087] In this step, the reconstructed high-precision 3D point cloud data undergoes full-width geometric topology analysis, decomposing the complex plate waveform into wave components of different orders (such as center wave, edge wave, and quarter wave), and calculating the differential elongation and wave height distribution vector distributed along the strip width direction, thereby generating comprehensive wave data containing local and overall deformation characteristics, providing quantitative feedback for subsequent rolling mill pressure distribution, bending roll force adjustment, and tension elongation setting.
[0088] In this example, the waveform data obtained from "wave shape feature analysis" provides refined control parameters for the rolling process. First, the 3D point cloud data is mapped to several control channels divided along the strip width. Legendre polynomials are used to orthogonally decompose the strip profile curve, separating first- and / or second-order straightness components representing symmetrical deformation (corresponding to bending roll and pressure control) and higher-order components representing asymmetrical deformation (corresponding to tilting roll control). Then, based on "fiber strip theory," the differential elongation and residual stress distribution of each channel relative to the reference length are calculated, outputting a structured data package containing the wave height value, maximum relative elongation difference, and characteristic wave shape phase for each channel. This data package is directly connected to the mill's strip profile control system as a feedforward or feedback input variable for calculating the dynamic pressure correction value for each stand and the total elongation setting, thereby achieving closed-loop precise control of the ultra-thin strip profile.
[0089] The following describes one possible implementation of each step in this example.
[0090] The Legendre polynomial is used to orthogonally decompose the cross-sectional curve of the plate, as shown in the following formula:
[0091]
[0092] in, The coefficient of the first-order term represents the tilt of the plate. The corresponding quadratic coefficient represents the mid-wave or the edge wave;
[0093] Divide the strip into N control channels along its width (e.g., the number of segments corresponding to the cooling rolls of the rolling mill). Calculate the actual arc length of the fiber strip within each channel. Compared with the reference length Differential elongation :
[0094]
[0095] This can then be converted to I units: .
[0096] This embodiment also provides a system, Figure 1 This is a system hardware structure diagram according to an embodiment of this application. The system solves the high reflectivity problem of ultra-thin stainless steel strips through a specific optical layout, such as... Figure 1 As shown, the system includes a gantry spanning over the strip, on which a wired laser projector and two sets of industrial camera assemblies are mounted.
[0097] The line laser projector is mounted directly above the gantry, projecting a laser line vertically downwards. To overcome the high light transmittance of the metal surface and the interference of the oil film, a blue-violet laser with a wavelength of 405nm is preferred. Compared to the traditional 650nm red light, the 405nm wavelength has higher photon energy, improving the scattering efficiency on the metal micro-lattice surface by approximately 30%, and can effectively penetrate the surface rolling oil film, illuminating the strip surface to form a clear diffuse reflection light stripe.
[0098] Figure 2 This is a top view of the camera placement position according to an embodiment of this application, such as... Figure 2 As shown, to avoid the specular reflection light cone, the first and second cameras are symmetrically mounted on the left and right outer edges of the strip's running direction, respectively. The angle between the camera's optical axis and the horizontal plane... The angle is set between 20° and 60°. At this angle, strong specular light reflected along the normal direction is physically avoided, and the camera only receives side-scattered light, thus ensuring that the image sensor is not saturated by strong light.
[0099] For example, in orthogonal polarization filtering, a linear polarizer is installed at the output port of the laser projector, adjusting its polarization direction to the P direction; and analyzers are installed in front of the lenses of two industrial cameras, adjusting their polarization direction to the S direction. By utilizing the physical properties that specular reflected light maintains its polarization state but is blocked by the S-direction analyzer, and that diffuse reflected light undergoes depolarization and can pass through, residual specular noise is further filtered out.
[0100] This method is based on the hardware system of an anti-reflective imaging acquisition system. In addition to the anti-reflective imaging acquisition device for projecting structured light onto the surface of the strip and acquiring images, the system also includes a data processing unit connected to the anti-reflective imaging acquisition device for performing the processing and calculation steps in the above method.
[0101] Figure 3 This is a main flowchart of the method according to an embodiment of this application. Figure 4 This is a system control logic block diagram according to an embodiment of this application, which is described below in conjunction with... Figure 3 and Figure 4 The above steps will be explained.
[0102] Step S1: Control the anti-reflective imaging acquisition device to acquire the laser contour image of the strip surface, and simultaneously capture the laser stripe image of the strip surface using the left and right cameras. Due to the influence of the on-site environment, the original image may contain oil mist noise, speckle, and laser line "breakage" caused by extreme local bending of the strip.
[0103] Step S2: Input the acquired image into the pre-trained laser stripe semantic segmentation network, perform pixel-level semantic analysis on the image, identify and remove background noise, and intelligently repair the laser lines that are broken due to reflection to generate a continuous laser binarization mask.
[0104] For example, the original grayscale image is input into a pre-trained "laser stripe semantic segmentation network." An improved network based on encoder-decoder U-Net is employed. A dilated convolutional module is introduced deep into the encoder to expand the receptive field, enabling the network to perceive the contextual geometric trends on both sides of the broken laser line. The network performs pixel-level classification of the image. Isolated bright noise points that do not conform to the texture features of the laser line are identified as background; the edge diffusion regions of bright light clusters are identified, retaining only the central high-energy region as the "skeleton"; for broken regions, the network predicts missing path connections based on the slope and curvature of the light stripes on both sides of the break, generating a continuous binary mask. A clean, continuous laser stripe binary mask image is output as the search constraint region for subsequent algorithms.
[0105] In the intelligent fracture repair process, for the laser line fracture area caused by high reflectivity, the effective center point sequence on both sides of the fracture edge is extracted. Let the coordinates of the discrete points on the effective curve be... Calculate the local slope vector at the breakpoint. With curvature The local slope k is obtained through first-order difference or least-squares fitting.
[0106]
[0107] Local curvature Based on the first derivative of discrete points and second derivative Approximate calculation:
[0108]
[0109] The semantic segmentation network combined with the calculated local slope With curvature As a geometric prior constraint, the coordinates of the missing path are predicted using polynomial interpolation or spline interpolation, thereby generating a smooth mask connection bridge to ensure that the repaired laser line conforms to the continuity characteristics of the physical deformation of the metal strip.
[0110] Step S3: Within the constraint range of the laser binarization mask, based on the local grayscale distribution characteristics and signal-to-noise ratio of the laser stripes, an adaptive sub-pixel extraction operator is selected to calculate the laser center coordinates.
[0111] For example, within the area defined by the mask, the system dynamically selects the optimal extraction operator based on the real-time grayscale distribution characteristics of the light stripe cross-section: when the grayscale distribution of the light stripe cross-section is unimodal and unsaturated, a logarithmic domain three-point Gaussian fitting algorithm is used. This algorithm assumes that the light intensity follows a Gaussian distribution, linearizes the nonlinear model through logarithmic transformation, and calculates the sub-pixel-level center coordinates. When continuous saturated pixels are detected at the center of the light stripe, the Gaussian model fails. The system automatically switches to the grayscale centroid method or the edge gradient-based center positioning method, using the gradient symmetry of the light spot edge to infer the center position. Smoothness constraint: A continuity constraint mechanism is introduced. If the currently calculated center point abruptly changes from its neighboring points, it is corrected by referring to historical data from the previous frame or neighborhood trends, and outliers are eliminated.
[0112] In the introduced continuity constraint mechanism, to prevent sub-pixel center coordinates from flying away due to local noise, a spatial continuity penalty function is set. Let the current column be... The sub-pixel center calculated at is The center of the preceding column is Then, the coordinate jump difference is calculated based on the first-order smoothness assumption:
[0113]
[0114] The system sets dynamic thresholds. ( (Based on the strip thickness and the maximum allowable wave height gradient). When When the value is 1, the point is determined to be a sudden jump point that violates the continuity assumption, triggering a correction mechanism. The correction algorithm utilizes the set of valid data points in its neighborhood. Perform weighted smoothing:
[0115]
[0116] in, The Gaussian weights are based on spatial distance, thus directly outputting a smooth and continuous laser center curve at the algorithm's underlying layer.
[0117] Step S4: Based on the principle of binocular stereo vision and combined with the system calibration parameters, the extracted laser center coordinates are converted into three-dimensional point cloud data of the strip surface.
[0118] For example, using the camera intrinsic parameter matrix K and the stereo extrinsic parameter matrix obtained in advance through Zhang Zhengyou's calibration method. The following operations are performed: Epipolar correction: Distortion and stereo correction are performed on the image coordinates of the left and right cameras to ensure that corresponding image points are in the same row. Stereo matching: Under epipolar constraints, the laser center points extracted by the left and right cameras are matched. Triangulation: Based on the parallax principle, the three-dimensional coordinates of each point on the laser line in the world coordinate system are calculated. This forms a cross-sectional point cloud profile covering the width of the strip.
[0119] Step S5 involves performing full-width geometric topology analysis on the reconstructed 3D point cloud data. The plate waveform is decomposed into wave components of different orders, and the differential elongation and wave height distribution vectors along the strip width direction are calculated. To convert the 3D data into control parameters executable by the rolling mill, the system performs the following calculations:
[0120] The measured cross-sectional curve of the plate Orthogonal decomposition using Legendre polynomials:
[0121]
[0122] in:
[0123] The coefficient of the first-order term represents the tilt (wedge shape) of the plate.
[0124] The corresponding quadratic coefficient represents the mid-wave (or edge wave).
[0125] The coefficient of the fourth term represents a quarter wave (compound wave).
[0126] Unit calculation: Divide the strip into N control channels along its width (e.g., the number of segments corresponding to the segmented cooling rolls of a rolling mill). Calculate the actual arc length of the fiber strip within each channel. Compared with the reference length Differential elongation :
[0127]
[0128]
[0129] This can then be converted to I units: .
[0130] Data output: The system will calculate the wave coefficients of each order. The I-unit values of each channel are sent in real time to the secondary control system of the rolling mill via the PLC communication interface as feedback signals for the automatic shape control (AFC) system. This is used to dynamically adjust the bending roll force, tilting roll angle, and coolant flow rate to achieve closed-loop control of the plate shape.
[0131] In the above Figure 3 and Figure 4The steps shown break through the physical limits of imaging ultra-thin mirror-finish metal strips, significantly improving the quality of the original signal. This invention solves the detection problem caused by high reflectivity of 0.02mm to 0.1mm ultra-thin stainless steel strips (tearable steel) at the physical imaging level through a three-dimensional collaborative anti-reflection strategy of "wavelength-polarization-geometry". It uses 405nm blue-violet light instead of traditional red light, utilizing the high scattering rate of short wavelengths on the microscopic surface of the metal to effectively enhance diffuse reflection signals and penetrate surface oil film interference. Through a 20°-60° lateral axis geometric layout combined with orthogonal polarization filtering, it physically blocks most of the strong specular reflection light from entering the lens, eliminating image overexposure and light spot saturation from the source. This allows for the acquisition of clear and continuous laser stripe images even on highly reflective surfaces, greatly reducing the difficulty of subsequent algorithm processing.
[0132] In the above Figure 3 and Figure 4 The illustrated steps achieve intelligent repair and semantic-level noise reduction of broken laser lines, ensuring the continuity and robustness of detection. Addressing the potential physical breakage of laser lines in extremely thin strips under severe waviness or surface oil contamination, this invention introduces a pre-trained "laser stripe semantic segmentation network." Intelligent Repair: Unlike traditional algorithms that can only handle continuous lines, this system can predict and "repair" missing paths at the semantic level based on the curvature and slope vectors on both sides of the break point, ensuring the integrity of the 3D reconstructed data and avoiding false alarms or data loss due to localized reflections. Halo Removal: By identifying the "energy skeleton" of the laser stripes through a deep learning model, false halos at the edges of bright light clusters are effectively removed, preventing positional deviations caused by halo pull in the traditional grayscale centroid method.
[0133] In the above Figure 3 and Figure 4 The illustrated steps demonstrate a direct conversion from "visual images" to "process parameters." Through data-driven closed-loop control, this invention surpasses the traditional strip shape analyzer's function of merely "defect display." By employing full-range geometric topology analysis and Legendre orthogonal decomposition, it directly outputs key process parameters that guide mill adjustments. Quantitative guidance: The differential elongation, wave height distribution vector, and center / edge wave components calculated by the system can be directly correlated to adjustments in the mill's bending roll force, tilting roll, and downward pressure. Closed-loop feedback: The generated structured data package provides precise feedforward or feedback input variables for the mill's strip shape control system (AGC / AFC), resolving the long-standing reliance on manual experience to judge strip shape in ultra-thin strip production, significantly improving yield and production safety.
[0134] In the above embodiments, a method and system for online detection and reconstruction of ultra-thin strip profiles are disclosed, belonging to the field of online detection and strip profile control technology for metal strips. The method includes: acquiring laser contour images of the strip surface using a line laser projector and a dual industrial camera arranged laterally; preprocessing the laser contour images and completing laser stripe regions affected by high reflectivity, noise interference, or stripe breakage to generate a binary mask corresponding to the laser stripes; extracting the laser center coordinates based on local image features of the laser stripes within the area defined by the binary mask; reconstructing three-dimensional point cloud data of the strip surface based on binocular vision calibration parameters and the laser center coordinates; calculating the strip's transverse profile curve based on the three-dimensional point cloud data, and outputting profile parameters such as wave height distribution, differential elongation, and / or wave shape components. The system includes an imaging acquisition module, an image processing module, a center extraction module, a three-dimensional reconstruction module, a profile calculation module, and a control interface module. This invention can improve the continuity, stability and accuracy of online strip shape detection for highly reflective ultra-thin strips, and provide reliable data support for closed-loop control of strip shape during rolling and straightening processes.
[0135] The above embodiments solve the problem in the detection of ultra-thin strips in related technologies where the laser line breaks due to high reflectivity, making extraction impossible. This improves the physical limit of imaging ultra-thin mirror metal strips, enhances the quality and accuracy of the original signal, alleviates the pain point of relying on manual experience to judge the shape in the production of ultra-thin strips, and improves the yield and production safety.
[0136] The above are merely embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.
Claims
1. A method for online detection and reconstruction of ultra-thin strip profiles, characterized in that, include: Laser contour images of the strip surface are acquired laterally along the side axis using an imaging acquisition device; The acquired laser contour image is input into a pre-trained semantic segmentation network for preprocessing and stripe completion processing to obtain a laser binarization mask. The semantic segmentation network is used to denoise the laser contour image and repair the laser stripes that are broken due to reflection in the denoised laser contour image to generate a continuous laser binarization mask. Within the constraint range of the laser binarization mask, a corresponding algorithm is selected based on the local gray-scale distribution characteristics and signal-to-noise ratio of the laser stripes, and the laser center coordinates are calculated and extracted based on the selected algorithm. Based on the binocular vision calibration parameters and the laser center coordinates, the three-dimensional point cloud data of the strip surface is reconstructed; Full-width geometric topology analysis is performed on the reconstructed 3D point cloud data to obtain the plate surface waveform; the plate surface waveform is decomposed into wave components of different orders, and the differential elongation and wave height distribution vector distributed along the width direction of the strip are calculated.
2. The method according to claim 1, characterized in that, The imaging acquisition device includes: Line laser projector; Two industrial cameras are arranged in a horizontal side-axis layout on the left and right sides of the strip's running direction, with the angle between the camera's optical axis and the horizontal plane set to 20° to 60°. An orthogonal polarization filter assembly includes a polarizer disposed at the line laser projector and an analyzer disposed in front of the industrial camera lens, wherein the polarization directions of the polarizer and the analyzer are adjusted to be orthogonal at 90°.
3. The method according to claim 2, characterized in that, The line laser projector is a blue-violet laser.
4. The method according to claim 1, characterized in that, The denoising process for the laser contour image includes: By perceiving the geometric shape of light clusters in laser contour images through multi-scale convolution kernels, asymmetric false halo layers are identified as background, and only the energy skeleton area that conforms to the overall line trend is retained. By utilizing context awareness, isolated bright spots that lack linear features extending along the width of the strip are marked as interference signals and removed.
5. The method according to claim 4, characterized in that, Repairing broken laser stripes caused by reflection in the denoised laser contour image includes: For physical breaks in laser stripes, based on the curvature trend and slope vector of the laser stripes on both sides of the break point, the missing path is predicted at the image semantic level and a connecting bridge is generated, which is used to repair the laser stripes.
6. The method according to claim 1, characterized in that, Based on the local grayscale distribution characteristics of the laser stripes and the signal-to-noise ratio, at least one of the following algorithms shall be selected for use: Logarithmic domain Gaussian fitting algorithm, three-point parabolic fitting algorithm, intensity-weighted centroid algorithm, and center positioning algorithm.
7. The method according to claim 6, characterized in that, The algorithms selected based on the local grayscale distribution characteristics and signal-to-noise ratio of the laser stripes include: When the grayscale distribution of the light stripe cross section presents a complete bell-shaped curve and is not saturated, a logarithmic domain Gaussian fitting algorithm is used; when the top of the light stripe cross section is flat or computational resources are limited, a three-point parabolic fitting algorithm is used; when specular reflection is detected, causing continuous pixel saturation, the algorithm is switched to intensity-weighted centroid algorithm or edge gradient-based center localization algorithm.
8. An online detection and reconstruction system for ultra-thin strip profiles, characterized in that, include: The imaging acquisition module is used to acquire laser contour images of the surface of an ultra-thin strip during operation; The image processing module is used to preprocess the laser contour image and complete the laser stripe region affected by reflection interference, noise interference or stripe breakage to generate a binarized mask corresponding to the laser stripe. The center extraction module is used to extract the laser center coordinates based on the local image features of the laser stripes within the area defined by the binarized mask. The 3D reconstruction module is used to reconstruct the 3D point cloud data of the strip surface based on the binocular vision calibration parameters and the laser center coordinates. The strip profile calculation module is used to calculate the transverse strip profile curve based on the three-dimensional point cloud data, and output the strip profile parameters such as wave height distribution, differential elongation and / or wave shape components. The control interface module is used to send the plate shape parameters to the rolling mill plate shape control system.