Paper drum defect full-inspection method based on machine vision

Abstract

The invention relates to the technical field of industrial nondestructive testing, in particular to a paper drum defect full-inspection method based on machine vision, which comprises the following steps: step 1, constructing a curved surface self-adaptive optical environment: arranging annular low-angle strip-shaped light source arrays along the circumferential direction of a paper drum at equal intervals, and forming a dynamically adjusted set acute angle between the emergent direction of each light source and the surface normal of the paper drum, the acute angle is reduced along with the increase of the diameter of the paper barrel, and near-infrared coaxial diffused light sources are arranged at the two axial ends of the paper barrel; step 2, motion compensation three-dimensional reconstruction; 3, multi-modal defect hierarchical decision making: expanding the three-dimensional point cloud into a two-dimensional image along the circumferential direction, and constructing a dynamic reference grid with the grid cell size adaptively adjusted along with the pattern complexity on the two-dimensional image; cooperatively extracting geometric structure features, surface texture features and gluing seam features; and outputting defect classification through a three-level decision tree. By means of the method, efficient and accurate paper barrel full inspection can be achieved, and the automation level of a production line and the paper barrel quality control capacity are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of industrial nondestructive testing technology, and in particular to a method for full inspection of defects in paper drums based on machine vision. Background Art

[0002] In the field of quality inspection in paper drum production, existing machine vision inspection technologies have the following field-specific shortcomings: As a typical cylindrical container, the unfolding of a paper drum requires a coordinate transformation from a curved surface to a plane. Traditional two-dimensional image acquisition methods (such as orthographic projection using a planar array camera) neglect curvature changes during the unfolding process, leading to nonlinear distortion in the glued seam area. This distortion manifests as follows: The glued seam in the middle of the barrel is compressed, and minor cracks are smoothly covered; Stretching of the adhesive seam in the transition zone between the barrel opening and bottom caused a real defect to be misjudged as a distortion artifact.

[0003] Actual measurements show that when the diameter of the paper drum is greater than 300mm, the positioning error of the glued seam can reach more than ±2mm, resulting in a failure rate of more than 15% for cracking defects.

[0004] Furthermore, the surface of paper drums typically exhibits an alternating distribution of highly reflective printed areas and low-reflective solid-color areas (such as the original color of kraft paper). Single-light source solutions (such as top-mounted ring lights) cannot accommodate these differences in reflectivity. The ink reflection in the printing area produces specular reflection, leading to localized overexposure and loss of registration misalignment defects; Solid color areas have strong light absorption, and surface depressions cause texture contrast to fall below the detection threshold due to insufficient illumination.

[0005] Industry data shows that such misjudgments account for 30%-40% of the total number of defects removed.

[0006] Therefore, there is an urgent need for a machine vision-based method for full inspection of defects in paper drums to solve the above problems. Summary of the Invention

[0007] To achieve the above objectives, the present invention provides a machine vision-based method for full inspection of defects in paper drums, comprising: Step 1: Constructing an adaptive optics environment for curved surfaces: A ring-shaped low-angle strip light source array is arranged at equal intervals along the circumference of the paper drum. The emission direction of each light source forms a dynamically adjusted acute angle with the normal to the surface of the paper drum. This acute angle decreases as the diameter of the paper drum increases. Near-infrared coaxial diffuse light sources are deployed at both ends of the paper drum along its axial direction. Step 2: Motion-compensated 3D reconstruction: The conveyor line speed and paper drum rotation angular velocity are acquired in real time, and the acquisition frequency of the line scan camera is dynamically calculated based on the line speed to ensure that the overlap rate of adjacent scan zones meets the requirements of three-dimensional reconstruction. Perform pixel-level spatiotemporal calibration based on motion parameters: map the scanned image to a virtual cylindrical coordinate system and restore motion blur; Reconstruct the three-dimensional point cloud of the outer surface of the paper drum and establish a reference coordinate system with the geometric center of the drum as the origin and the axis of the drum as the Z-axis; Step 3: Multimodal defect hierarchy decision: The 3D point cloud is unfolded into a 2D image along the circumference, and a dynamic reference grid is constructed on the 2D image with the grid cell size adaptively adjusted according to the pattern complexity. Collaborative extraction of geometric structural features, surface texture features, and adhesive seam features; Defect classification is output using a three-level decision tree: Level I structural defects are determined by comparing the magnitude of curvature abrupt changes with the structural deformation limit. Level II determines printing defects based on the comparison of the Mahalanobis distance between the entropy anomaly area and the normal area and the texture difference threshold; Level III is used to determine adhesive defects based on the discontinuity of the differential peak value of the adhesive joint and the grayscale profile, which meet the conditions for poor adhesive bonding. The structural variation limit, texture difference threshold, and poor bonding conditions were all determined statistically using a defect-free sample set.

[0008] Preferably, the dynamic adjustment of the set acute angle includes: Establish a geometric mapping relationship between the diameter of the paper drum and the incident angle of the light source: when the diameter of the paper drum increases, reduce the angle between the light source emission direction and the normal to the drum surface to compensate for the illuminance attenuation in the middle of the drum. The radius of the paper drum is obtained in real time by a laser rangefinder, and the required acute angle value is queried according to the pre-calibrated radius-angle mapping curve. The pitch adjustment mechanism of the driving light source unit ensures that the error between the actual incident angle and the query value is less than the angle tolerance. The radius-angle mapping curve is generated through optical simulation and experimental calibration: under a standard light source array, the illuminance distribution on the surface of paper drums with different radii is measured, and the optimization target is to achieve a set ratio of illuminance between the middle and edge of the drum. The optimal incident angle is then iteratively solved.

[0009] Preferably, the dynamic calculation of the sampling frequency includes: Based on the angular velocity of the paper drum rotation and the minimum number of sampling points required in the circumferential direction, calculate the minimum number of image frames required for a single rotation. The number of axial scan rows is determined based on the conveyor line speed and the axial length of the paper drum; The final acquisition frequency is obtained by multiplying the product of the minimum number of frames and the number of scan lines by an overlap safety factor greater than 1. The minimum number of sampling points is determined by the accuracy requirements of 3D reconstruction: the sampling points are gradually increased during the reconstruction test, and the minimum number of points corresponding to the time when the deviation between the reconstructed model and the actual size is less than the set error is the required value.

[0010] Preferably, the motion blur restoration process includes: The motion blur direction and blur kernel length are calculated based on the rotational angular velocity and exposure time. A point spread function model matching the size of the blur kernel is constructed, which considers the nonlinearity of pixel displacement caused by the curved surface of the paper barrel; The image is restored using a regularized deconvolution algorithm, and the deblurring effect and noise suppression are balanced by adjusting the regularization coefficient; The determination of the regularization coefficient includes: acquiring degraded images on a standard motion blur test chart, using the combined score of edge sharpness and noise level of the restored image as an indicator, and iteratively optimizing the coefficient value.

[0011] Preferably, the generation of the dynamic reference mesh includes: Extract scale-invariant feature points of printed patterns from 2D unfolded images; An initial triangular mesh is generated using feature points as nodes; The weighted sum of the gray-level variance and the mean gradient magnitude within each grid cell is used as a complexity index. When the complexity index exceeds the adaptive threshold, recursively subdivide the grid until the detection accuracy requirement is met; The adaptive threshold is dynamically set based on the overall image complexity distribution: the image is divided into several blocks, and the percentage quantile of the highest complexity within each block is used as the threshold benchmark.

[0012] Preferably, the extraction of the curvature abrupt change amplitude includes: Take circumferential point cloud sections at equal intervals along the axial direction of the paper drum; Perform circle fitting on the point cloud of each section and calculate the sequence of fitting radii; Perform sliding window statistical analysis on the radius sequence and calculate the standard deviation of the radius within the window; When the number of windows in which the standard deviation continuously exceeds the fluctuation tolerance is greater than the set value, it is determined that there is a curvature change in the region; The fluctuation tolerance is determined by statistical analysis of the radius fluctuation of defect-free samples: the maximum standard deviation of all samples is calculated and multiplied by a safety factor to obtain the tolerance value.

[0013] Preferably, the determination of poor bonding conditions includes: Peak detection is performed on the differential curve of the adhesive joint to eliminate spurious peaks caused by noise. The effective peak spacing distribution is statistically analyzed, and an early warning is triggered when the spacing is less than the minimum adhesive discontinuity distance. Extract grayscale profiles at the warning locations and calculate the discontinuity index of the profile signal; When the discontinuity index exceeds the dynamic threshold, a bonding defect is confirmed. The minimum adhesive discontinuity distance was determined through destructive experiments: defect samples with different discontinuity lengths were prepared, and the minimum length that could be reliably detected by machine vision was used as the benchmark value.

[0014] Preferably, cross-modal verification is performed before output defect classification: Inversely map the coordinates of the defect region in the 2D image to the 3D point cloud space; Calculate the depth gradient of the defect region in three-dimensional space; When the depth change gradient is less than the optical interference tolerance, the original defect determination is cancelled. The determination of optical interference tolerance includes: applying simulated illumination interference to the surface of a defect-free sample, measuring the maximum value of the spurious depth change caused by it, and using twice that value as the tolerance benchmark.

[0015] Preferably, the determination of the structural variation limit includes: Collect a dataset of curvature abrupt change amplitudes from defect-free paper drum samples; Calculate the mean and standard deviation of the dataset; The tolerance value is the mean plus a certain number of standard deviations, and this multiple is adjusted according to the defect missed detection rate requirements. The defect missed detection rate is required to be converted into statistical confidence level through production line process standards, and then mapped to the standard deviation multiple.

[0016] Preferably, determining the texture difference threshold includes: Extract the texture feature vectors of defect-free samples and typical defective samples respectively; Calculate the Mahalanobis distance distribution between the two types of feature vectors; The initial threshold is set at the preset high quantile of the distance distribution of defect-free samples; The threshold is dynamically fine-tuned based on the actual false alarm rate. The preset high quantile is determined through ROC curve analysis: the defect detection rate and false alarm rate corresponding to different quantiles are tested on the validation set, and the minimum quantile that meets the false alarm rate requirement is selected.

[0017] The beneficial effects of this invention are: 1. This invention overcomes the problem of neglecting curvature changes, which leads to distortion in the glued seam area and thus affects the accuracy of defect detection, by dynamically adjusting the incident angle of the light source, using point cloud processing based on 3D reconstruction, and pixel-level spatiotemporal calibration. By arranging a low-angle strip light source array along the circumference of the paper drum and dynamically adjusting the light source exit angle, the uniformity of illumination across the entire drum is ensured, compression and tensile deformation of the glued seam is avoided, and defect areas of the paper drum are accurately located.

[0018] 2. This invention successfully eliminates the influence of reflectivity differences on detection results by configuring a near-infrared coaxial diffuse light source and employing a combination of dynamic adjustment of the light source incident angle and a light source array. Furthermore, by using 3D point cloud reconstruction and a multimodal decision-making mechanism based on texture differences, adaptive illumination compensation can be performed according to the reflectivity characteristics of different regions, thereby improving the detection accuracy of defects and reducing false positives and false negatives.

[0019] 3. This invention significantly improves detection accuracy. In particular, for the identification of minute cracks and poor bonding defects, it employs adaptive mesh generation and complexity indices to refine the detection area, ensuring accurate detection even of minute defects in the middle or transition zone of the barrel. Through a cross-modal verification method, it further reduces misjudgments caused by optical interference, ensuring high reliability of defect classification.

[0020] 4. This invention solves the detection problems caused by paper drum rotation and surface reflection differences by using techniques such as dynamic calculation of acquisition frequency, pixel-level spatiotemporal calibration, and motion-compensated image restoration. Furthermore, the defect classification method based on Mahalanobis distance and texture differences ensures more accurate defect identification, avoids misjudgments caused by changes in lighting and surface material differences, and further reduces the false negative and false positive rates. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the steps of the method of the present invention; Figure 2 This is a flowchart of the motion blur restoration process described in the method of the present invention; Figure 3 This is a flowchart illustrating the steps for determining poor bonding conditions according to the method of the present invention. Detailed Implementation

[0023] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. It should also be noted that, to make the embodiments more comprehensive, the following embodiments are the best and preferred embodiments, and those skilled in the art can use other alternative methods to implement some well-known technologies; moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0024] Please see Figures 1-3This invention provides a machine vision-based method for full defect inspection of paper drums. In step 1, to effectively overcome the coordinate transformation problem from curved surface to plane caused by the cylindrical shape of the paper drum, an annular low-angle strip light source array is first arranged at equal intervals around the circumference of the paper drum. The emission direction of each light source forms a dynamically adjusted acute angle with the normal to the surface of the paper drum. This acute angle decreases as the diameter of the paper drum increases, ensuring the uniformity and accuracy of the light source illumination. Near-infrared coaxial diffuse light sources are deployed at both ends of the paper drum along its axial direction to enhance the illumination effect on the transition area between the drum opening and the bottom. This design effectively avoids image distortion and reflection caused by uneven illumination, providing clear and stable image information for subsequent defect detection.

[0025] In step 2, the paper drum rotates at a certain speed on the production line and moves along the conveyor line. Traditional methods are prone to blurring when dealing with this motion, affecting detection accuracy. This invention acquires the conveyor line speed and the paper drum's rotational angular velocity in real time, and dynamically calculates the acquisition frequency of the line scan camera based on this, ensuring that the overlap rate of adjacent scan strips meets the requirements for 3D reconstruction. Through precise spatiotemporal calibration, the scanned image is mapped to a virtual cylindrical coordinate system, and motion blur is restored. This method effectively compensates for image deviations caused by the paper drum's motion during production, accurately reconstructs the 3D point cloud of the paper drum's outer surface, and ensures high accuracy in subsequent defect detection.

[0026] In step 3, the reconstructed 3D point cloud is unfolded into a 2D image along the circumference. An adaptive dynamic reference mesh is constructed, automatically adjusting the mesh cell size based on the pattern complexity. This method enables higher-resolution detection of complex surfaces or detailed areas. In the defect classification process, this invention combines geometric structural features, surface texture features, and glue seam features, employing a three-level decision tree for defect determination. Level I determines the presence of structural defects by comparing the magnitude of curvature abrupt changes with the structural deformation limit. Level II determines printing defects based on the Mahalanobis distance and texture difference between the entropy anomaly region and the normal region. Level III is determined based on the differential peak value of the adhesive joint and the discontinuity of the grayscale profile to identify adhesive defects.

[0027] These decision conditions were obtained through statistical analysis of a defect-free sample set, ensuring accurate identification and classification of each defect type.

[0028] Dynamic adjustment and adaptive design of the optical environment solve the problem of uneven illumination when processing curved surfaces of paper drums using traditional methods, providing high-quality image input. Secondly, motion-compensated 3D reconstruction technology overcomes the impact of motion blur in the paper drums, improving detection accuracy. Finally, combining multimodal defect hierarchy decision-making not only improves the accuracy of defect detection but also reduces false positives and false negatives, particularly excelling in identifying minute cracks, adhesive defects, and printing problems. This method enables efficient and accurate full inspection of paper drums, effectively enhancing the automation level of the production line and the quality control capabilities of the paper drums.

[0029] In one possible implementation, during the paper drum production process, due to the cylindrical shape of the paper drum, the lighting conditions in the center and edges of the drum differ. Particularly as the diameter of the paper drum increases, the illuminance in the center gradually decreases, affecting the quality of subsequent images and inspection results. To address this issue, this invention designs a method for dynamically adjusting the incident angle. Specifically, as the diameter of the paper drum increases, the angle between the emission direction of the light source and the normal to the surface of the paper drum gradually decreases. This adjustment compensates for the attenuation of illuminance in the center of the drum, ensuring uniform illumination across the entire surface of the paper drum, especially in the central region, guaranteeing uniform lighting and clear images.

[0030] To precisely control the incident angle of the light source, this invention uses a laser rangefinder to acquire the radius information of the paper drum in real time. By measuring the real-time radius value, the required acute angle value can be queried based on a pre-calibrated radius-angle mapping curve. This real-time feedback mechanism adjusts the incident angle of the light source according to changes in the paper drum, ensuring that the lighting conditions are always optimal during each detection process.

[0031] The light source unit is equipped with a pitch adjustment mechanism that automatically adjusts the incident angle of the light source based on the queried acute angle value, ensuring that the error between the light source and the queried value is less than a preset angle tolerance. Through this adjustment mechanism, the light source can automatically adapt to the lighting requirements of paper drums of different sizes, eliminating the inaccuracies caused by manual adjustment and ensuring precise control of the light source angle.

[0032] To achieve precise acute angle adjustment, the radius-angle mapping curve is generated through a combination of optical simulation and experimental calibration. First, under a standard light source array, the illuminance distribution on the surface of paper drums with different radii is measured, with the optimization goal being to achieve a set standard for the ratio of illuminance at the center to the edge of the drum. Through multiple iterative calculations, the optimal incident angle is determined, and the corresponding radius-angle mapping curve is generated. This curve ensures that the incident angle of the light source always meets the optimal illumination requirements for different paper drum diameters, achieving the most ideal detection effect.

[0033] By automating the adjustment of the light source's incident angle, the uniformity of illumination during the paper drum inspection process is significantly improved, particularly addressing the issue of illuminance attenuation in the center of the paper drum, thus ensuring image quality and inspection accuracy. Furthermore, the use of a laser rangefinder and dynamic adjustment mechanism not only enhances the automation and accuracy of the inspection but also reduces errors from manual adjustments, avoiding the influence of human factors on the inspection results. Ultimately, this method enables efficient and high-precision paper drum defect detection, improving the quality control capabilities of the production line and ensuring the consistency and stability of product quality.

[0034] In one possible implementation, the number of image frames to be acquired is first determined based on the rotational angular velocity of the paper drum and the minimum number of sampling points required in the circumferential direction. The rotational angular velocity reflects the degree of rotation of the paper drum per unit time, while the minimum number of sampling points is determined by the accuracy requirements of 3D reconstruction. Specifically, during model reconstruction, the number of sampling points needs to be gradually increased until the dimensional deviation between the reconstructed 3D model and the actual object is less than a predetermined error; the corresponding number of sampling points is the minimum number of sampling points. Therefore, based on the rotational angular velocity of the paper drum, the system can calculate the minimum number of image frames required for each rotation to meet the requirements of accurate 3D reconstruction.

[0035] Next, based on the conveyor speed and the axial length of the paper drum, the number of scan lines along the axial direction is determined. The number of axial scan lines can be calculated using the relationship between the axial length of the paper drum and the conveyor speed. To ensure that every area of ​​the paper drum is fully scanned, the number of axial scan lines must match the length of the paper drum to avoid omissions or duplicate scans.

[0036] Based on the aforementioned steps, the product of the minimum number of frames and the number of axial scan lines is used as a benchmark, multiplied by an overlap safety factor greater than 1 to obtain the final acquisition frequency. This overlap safety factor ensures sufficient overlap between adjacent image frames, thereby avoiding data loss or discontinuity due to insufficient acquisition frequency, thus guaranteeing the accuracy and reliability of 3D reconstruction.

[0037] The minimum number of sampling points is determined by the required accuracy of 3D reconstruction. In practical applications, the number of sampling points is gradually increased through repeated testing until the dimensional deviation between the reconstructed model and the actual object is less than a set error range. This value is considered the minimum number of points required to meet the accuracy requirements. This technical feature ensures that the sampling frequency meets the needs of reconstruction accuracy, avoiding inaccuracies in the 3D model due to insufficient sampling.

[0038] By dynamically calculating the acquisition frequency, the image acquisition speed can be automatically adjusted according to the movement state of the paper drum and the required reconstruction accuracy, ensuring the quality of image acquisition and the accuracy of the reconstructed model. Accurate calculation of the minimum number of sampling points, the number of axial scan rows, and the number of image frames avoids over-acquisition and under-acquisition. Ultimately, this adaptive acquisition frequency calculation method significantly improves the efficiency and accuracy of paper drum defect detection, reduces the need for manual intervention and adjustments, and ensures the accuracy and reliability of defect detection, enhancing the automation level of the production line and the quality control capabilities of the paper drums.

[0039] In one possible implementation, motion blur occurs during the rotation of the paper drum due to the interaction between the camera's exposure time and the drum's rotational angular velocity. To accurately restore the image, the direction of the motion blur and the length of the blur kernel must first be calculated based on the drum's rotational angular velocity and the camera's exposure time. The rotational angular velocity determines the direction of the blur, while the product of the exposure time and the rotational angular velocity affects the size of the blur kernel. By calculating these parameters, the specific characteristics of the blur can be obtained, serving as the basis for the restoration process.

[0040] To effectively restore the image, a point spread function (PSF) model matching the blur kernel size is needed. Since the surface of the paper drum is curved, its curvature causes nonlinear changes in pixel displacement within the image. Therefore, the PSF needs to consider this nonlinear effect to accurately describe the cause of the blur. By considering the geometric characteristics of the paper drum surface, a more realistic PSF can be established, thereby improving the accuracy of the restoration.

[0041] To recover a sharp image from a blurred image, a regularized deconvolution algorithm is employed. This algorithm infers the original image from the blurred image through a deconvolution process. In this process, the introduction of regularization coefficients balances the deblurring effect with noise suppression. The regularization term helps suppress noise during image restoration, avoiding new noise generated by excessive deblurring, resulting in a restored image with high sharpness without excessively amplifying noise.

[0042] The choice of regularization coefficient is crucial to the image restoration effect. This invention acquires degraded images on a standard motion blur test chart and uses the edge sharpness and noise level of the restored image for comprehensive scoring, which serves as the basis for optimizing the regularization coefficient. By iteratively optimizing the value of the regularization coefficient, the optimal balance between deblurring and noise suppression is ensured, thereby obtaining a clear image with low noise.

[0043] By accurately calculating the motion blur direction and blur kernel length, and constructing a point spread function model suitable for the paper drum surface, the blurring problem caused by rotation and curved surfaces is effectively overcome, significantly improving the accuracy of image restoration. The regularized deconvolution algorithm, by adjusting the regularization coefficient, can remove motion blur and suppress noise, ensuring the quality of the restored image. This method provides a clear image foundation for subsequent paper drum defect detection, effectively improving the accuracy and efficiency of detection.

[0044] In one possible implementation, firstly, scale-invariant feature points (algorithms such as SIFT and ORB can be used for feature point extraction) are extracted from the two-dimensional unfolded image of the printed pattern on the surface of the paper drum. These feature points reflect key structural information in the pattern and are unaffected by factors such as scale and rotation. By extracting these feature points, a reliable basis can be provided for subsequent mesh generation.

[0045] Based on the extracted feature points, an initial triangular mesh is generated using algorithms such as Delaunay triangulation. During mesh generation, feature points are used as nodes in the mesh, and the edges of the triangles connect these feature points. This step provides a structured mesh partition for image complexity analysis, making subsequent complexity calculations more accurate.

[0046] After triangulation, the gray-level variance and the mean gradient magnitude within each grid cell are calculated. The gray-level variance reflects the fluctuation of brightness changes within the region, while the mean gradient magnitude reflects the intensity of image edges. The weighted sum of these two values ​​serves as a complexity index for that grid cell, effectively distinguishing between complex and simple regions in a pattern.

[0047] When the complexity index of a certain grid cell exceeds a set adaptive threshold, the system will recursively subdivide the grid. By subdividing the grid, higher resolution can be provided in complex areas to better capture subtle defects in the pattern. The subdivision process continues until the complexity index of the grid drops below the threshold or meets the detection accuracy requirements.

[0048] To ensure the adaptive threshold can be flexibly adjusted, avoiding over- or under-subdivision, the threshold is dynamically adjusted based on the complexity distribution of the entire image. The image is first divided into several small blocks, and the highest complexity value is calculated within each block. The threshold for the entire image is then set based on the percentage quantiles of these highest complexity values. This dynamic adjustment ensures that complex regions receive sufficient resolution, while simple regions avoid unnecessary subdivision, thereby optimizing computational efficiency and accuracy.

[0049] By dynamically generating adaptive meshes, effective subdivision of complex regions in images can be achieved, avoiding overprocessing of simple regions and optimizing image processing resource and time consumption. Furthermore, the use of dynamically set adaptive thresholds allows the entire detection process to adapt to varying image complexity, improving the method's versatility and robustness. This approach ensures thorough capture of image details, significantly improving defect detection accuracy and efficiency, making it suitable for real-time monitoring in large-scale production environments.

[0050] In one possible implementation, firstly, several circumferential cross-sections are cut from the paper drum at equal intervals along its axial direction. Each cross-section represents point cloud data of the paper drum's surface at a certain location. The point cloud is acquired from a machine vision system and contains the three-dimensional coordinate information of each point on the paper drum's surface. This point cloud data provides raw data support for subsequent geometric fitting.

[0051] Circular fitting is performed on the point cloud data of each cross-section. Circular fitting uses algorithms such as least squares to fit the cross-sectional point cloud data into an optimal circular model, and then calculates the fitting radius for each cross-section. The change in the fitting radius reflects the change in the curvature of the paper drum surface, forming a radius sequence that serves as the basis for subsequent analysis.

[0052] Next, the radius sequence is input into a sliding window for statistical analysis. The sliding window can be set to a fixed length, and as the window slides, the standard deviation of all radius values ​​within the window is calculated. The standard deviation reflects the magnitude of radius changes within the window; a larger standard deviation indicates a significant change in curvature in that region. This step effectively captures the magnitude of abrupt changes in curvature.

[0053] When the number of windows within a sliding window whose standard deviation continuously exceeds the preset fluctuation tolerance is greater than a set value, it is determined that there is a curvature abrupt change in that area. In other words, if the radius change of a certain part of the paper drum exceeds the normal fluctuation range, it is considered that there is a curvature abrupt change in that part, which may be a defect caused by printing or other process problems.

[0054] The fluctuation tolerance is determined based on statistical analysis of the radius fluctuations of defect-free samples. First, the radius fluctuations of defect-free paper drum samples are statistically analyzed, and the maximum standard deviation of all samples is calculated. Then, this is multiplied by a safety factor to obtain the fluctuation tolerance value. This tolerance value serves as a standard to determine whether the curvature change in a certain area falls within the normal range during subsequent inspections, ensuring that areas with curvature changes exceeding this value can be accurately identified as defective areas.

[0055] Precise geometric analysis effectively identifies abrupt changes in curvature on the surface of paper drums, thereby uncovering potential defect areas. This method boasts high efficiency and accuracy, enabling rapid processing of paper drum data from large-scale production, real-time detection of curvature anomalies, and improved defect identification rates. Sliding window analysis meticulously captures curvature changes across each cross-section, avoiding potential missed detections in traditional methods. Furthermore, the dynamic setting of fluctuation tolerance ensures the method's flexibility and adaptability, accommodating the production characteristics of different batches of paper drums, further enhancing its robustness and applicability.

[0056] In one possible implementation, firstly, an image of the glue seam on the surface of the paper drum is acquired using a machine vision system, and then the differential curve of the glue seam is extracted using an image processing algorithm. The differential curve reflects the brightness variation of the glue seam at different locations. A peak detection algorithm is used to find the local maxima and minima in the curve; these peaks represent abrupt changes in the gluing process. However, the image acquisition process may be affected by noise, so it is necessary to remove spurious peaks caused by noise. Common noise removal methods include low-pass filtering and morphological processing, which can effectively filter out noise and retain true peak information.

[0057] After noise removal, the spacing of the remaining valid peaks is statistically analyzed. The peaks in the glue joint represent normal connection points during the bonding process, while the spacing between peaks reflects discontinuities in the bonding. If the peak spacing is too small, it may indicate a problem in the bonding process, such as overly dense bonding areas or discontinuous glue joints. A minimum bonding discontinuity distance threshold is set; when the spacing between valid peaks is less than this threshold, an early warning mechanism is triggered, indicating a potential bonding defect.

[0058] When the system triggers an alert, grayscale profile data is extracted for the alert location. This grayscale profile data reflects the brightness changes at that location in the image. Next, the discontinuity index of the grayscale profile signal is calculated; common methods include calculating the signal's standard deviation and gradient changes. If significant discontinuity (i.e., drastic grayscale changes) is observed in the profile signal, it further confirms the possibility of adhesive defects in that area. Large signal discontinuities indicate potential breakage or discontinuity in the adhesive joint.

[0059] To address the varying bonding characteristics of different paper drums, a dynamic threshold method is employed to identify bonding defects. This dynamic threshold is dynamically adjusted by comprehensively considering historical sample data and the current state of the paper drum. Through real-time monitoring of discontinuity indicators, the system confirms a bonding defect at that location when the value exceeds the set dynamic threshold. The dynamic threshold setting improves the robustness of the method and adapts to quality fluctuations in different paper drum samples.

[0060] The minimum glue discontinuity distance was determined through destructive testing. In the experiments, defect samples of varying lengths were created to simulate poor glue bonding in actual production. These samples were then inspected using a machine vision system to obtain the minimum reliably detectable discontinuity length, which served as the baseline value for the minimum glue discontinuity distance. Determining this baseline value ensured that the system's sensitivity to glue defects matched the requirements of actual production.

[0061] This invention enables efficient detection of gluing defects on the surface of paper drums, especially in complex gluing areas. Through peak detection and noise removal, accurate glue seam information can be extracted, avoiding noise interference and improving detection accuracy. By setting the minimum gluing discontinuity distance and adjusting the dynamic threshold, the system can flexibly adapt to different paper drum production characteristics and quality standards, ensuring high accuracy and low false alarm rate in defect detection. Furthermore, the discontinuity index of the grayscale profile signal allows for in-depth analysis of the specific location and nature of gluing defects, further improving the accuracy of defect localization. This method not only enables real-time monitoring of product quality on the production line but also significantly improves the production efficiency and quality control level of paper drums, possessing significant industrial application value.

[0062] In one possible implementation, the machine vision system first acquires preliminary information about the defects on the surface of the paper drum using a two-dimensional image and extracts the coordinates of the defect area. These coordinates represent the planar position of the defect in the two-dimensional image. Then, using a pre-established mapping relationship between the two-dimensional image and the three-dimensional point cloud space, the two-dimensional coordinates of the defect area are inversely mapped to the three-dimensional point cloud space. This step obtains the true coordinates of the defect area in three-dimensional space, ensuring that subsequent analysis can fully consider depth information.

[0063] After obtaining the three-dimensional coordinates, the depth gradient of the defect region is calculated. The depth gradient reflects the depth difference between the defect region and the surrounding normal region in three-dimensional space. This depth gradient is a key indicator for determining whether a defect is genuine. Generally, genuine defects will exhibit a significant depth change, while spurious defects caused by optical interference or noise usually have a smaller depth change. By calculating the depth gradient, genuine defects can be effectively distinguished from spurious defects caused by lighting or other factors.

[0064] If the calculated depth change gradient of the defect area is less than the set optical interference tolerance value, the defect determination for that area is revoked. This means that although defect markers may exist in a two-dimensional image, depth verification in three-dimensional space confirms that the depth change of the area does not conform to the characteristics of a defect, thus ruling out false defects caused by optical interference or other factors. This verification process effectively improves the accuracy of defect classification and avoids interference from optical errors.

[0065] The setting of the optical interference tolerance is crucial to the accuracy of this method. This tolerance is determined experimentally by applying simulated lighting interference (such as varying illumination angle, intensity, or reflection) to the surface of a defect-free sample and measuring the maximum value of the resulting spurious depth changes. Typically, illumination variations may cause minor fluctuations in surface depth, but these fluctuations should not be misinterpreted as genuine defects. The maximum value of the spurious depth changes measured experimentally is used as the tolerance benchmark, and the tolerance value is set to twice this maximum value to ensure that the tolerance is sufficiently lenient to exclude minor fluctuations caused by factors such as illumination.

[0066] By combining 2D images and 3D point cloud data for cross-modal verification, the accuracy of defect detection is significantly improved. Traditional 2D image-based defect detection is easily affected by external factors such as illumination and reflection, which may lead to the identification of false defects. By mapping the defect coordinates in the 2D image to the 3D point cloud space and calculating the depth gradient, the authenticity of defects can be verified more accurately. In addition, the setting of optical interference tolerance, through experimentally determined dynamic tolerance values, can effectively offset the effects of illumination interference, thereby reducing false positives and false negatives and ensuring the accuracy and reliability of the final defect classification.

[0067] This cross-modal verification method is particularly suitable for production environments that require high-precision testing. It can automatically screen out real defects without relying on manual intervention, thereby improving the automation level of product quality control.

[0068] In one possible implementation, to accurately determine the structural deformation of the paper drum, it is first necessary to collect sample data of defect-free paper drums. These defect-free samples represent the structural characteristics of normal paper drums. The surfaces of these defect-free paper drums are scanned using a machine vision system to extract the curvature abrupt change amplitude data. The curvature abrupt change amplitude reflects the drastic degree of change in the surface morphology of the paper drum, usually caused by structural deformation or defects. By inspecting a large number of defect-free paper drum samples, a curvature abrupt change amplitude dataset is constructed as a benchmark for subsequent tolerance setting.

[0069] After obtaining a dataset of curvature variation amplitudes from defect-free paper drum samples, statistical methods were used to analyze the data. First, the mean (representing the average curvature variation amplitude of normal paper drums) and standard deviation (measuring the range of data fluctuation) of the dataset were calculated. The mean and standard deviation provide a reference range for the structural deformation of defect-free paper drums, serving as the basis for setting tolerances.

[0070] When determining the tolerance limit for structural deformation, the tolerance value is set as "mean + several times the standard deviation". This setting method can automatically adjust according to the normal range of curvature abrupt changes, ensuring that paper drums with structural deformation exceeding this tolerance are identified as defective during normal production. The set multiple needs to be adjusted according to the specific defect miss rate requirements. This multiple reflects the tolerance for the miss rate; the larger the multiple, the greater the tolerance for error and the lower the miss rate; the smaller the multiple, the smaller the tolerance for error and the higher the miss rate.

[0071] The required defect false negative rate is determined based on the production line's process standards. In actual production, to ensure product quality, an acceptable false negative rate is usually set. For example, if the required false negative rate is no more than 1%, this rate can be mapped to a corresponding statistical confidence level (usually 99%) using statistical methods. Through statistical analysis, this confidence level is converted into a multiple of the standard deviation. Specifically, the relationship between confidence level and standard deviation can be determined using the properties of a normal distribution; for example, a 99% confidence level corresponds to approximately 2.33 times the standard deviation. In this way, the tolerance can be dynamically adjusted according to production process requirements, enabling the detection system to efficiently screen for defects while reducing the risk of false alarms and false negatives.

[0072] This method can precisely set the tolerance value for paper drum structural deformation, ensuring the efficiency and accuracy of the detection system. By collecting defect-free sample data and calculating the mean and standard deviation, a set of standards that dynamically adapt to production needs can be established. The tolerance value can be flexibly adjusted according to the defect false negative rate requirements, ensuring that the false negative rate and false alarm rate are within a reasonable range. The greatest advantage of this method is its ability to flexibly adjust the tolerance according to actual production process standards, meeting both the requirements for high-precision defect detection and ensuring the efficient operation of the production line, avoiding misjudgments and production line shutdowns caused by overly strict tolerances. Furthermore, by combining with process standards, the detection system has strong adaptability, capable of adapting to the detection needs of different production lines and under different production conditions. This not only improves the quality control level of paper drums but also reduces human intervention and increases the degree of production automation.

[0073] In one possible implementation, firstly, to determine the structural deformation of the paper drum, sample data of defect-free paper drums must be collected. These samples represent the structural characteristics of the paper drums under normal production conditions. The surface of the defect-free paper drums is scanned using a machine vision system to capture and extract data on curvature abrupt changes. The curvature abrupt change amplitude reflects drastic changes in the surface morphology of the paper drum, which are typically associated with structural deformation or defects. By inspecting a large number of defect-free samples, a curvature abrupt change amplitude dataset can be generated, providing a standard for subsequently setting tolerances.

[0074] Once sufficient defect-free paper drum sample data is collected, the mean and standard deviation of the dataset are calculated through statistical analysis. The mean represents the average curvature variation amplitude of normal paper drums, while the standard deviation reflects the fluctuation range of the curvature variation amplitude. The mean and standard deviation provide a reference basis for setting tolerance values, ensuring that subsequent defect detection can reasonably distinguish the characteristics of normal paper drums.

[0075] When determining the tolerance limit for structural deformation, the tolerance value is set as "mean + several times the standard deviation". This method ensures that during normal production, paper drums with structural deformation exceeding the tolerance limit are identified as defective. The tolerance multiple is adjusted based on the specific defect miss rate requirements; the allowable error range determines the miss rate. A larger multiple indicates a larger tolerance for error and a lower miss rate; a smaller multiple indicates a smaller tolerance for error and a higher miss rate.

[0076] The required false negative rate is determined based on the production line's process standards and converted into confidence levels using statistical methods. In actual production, an allowable false negative rate is usually set, for example, no more than 1%. Through statistical analysis, this false negative rate is converted into a corresponding confidence level (e.g., 99% confidence level). Then, based on the properties of the normal distribution, the confidence level is correlated with a multiple of the standard deviation. A 99% confidence level typically corresponds to 2.33 times the standard deviation. By adjusting the multiple of the standard deviation, the requirements of different production lines can be dynamically adapted to balance the false negative rate and the false positive rate.

[0077] By employing accurate statistical analysis methods, tolerance values ​​for paper drum structural deformation can be precisely set, ensuring the efficiency and accuracy of the detection system. Secondly, the dynamic adjustment mechanism of the tolerance values ​​allows the system to flexibly adapt to different production processes and quality control requirements. Adjusting the standard deviation multiple based on the defect miss rate requirements ensures that the miss rate and false alarm rate remain within a reasonable range, avoiding misjudgments or production line shutdowns caused by overly strict tolerances. By integrating with production process standards, this method not only improves the accuracy of paper drum quality control but also reduces human intervention and promotes production automation. Furthermore, this method effectively enhances the adaptability of the detection system, meeting the quality inspection needs of different production lines and conditions, contributing to improved overall production efficiency and product qualification rate, reduced production costs, and ultimately achieving more efficient and reliable paper drum defect detection.

[0078] The following examples will illustrate this in detail: This invention is applied to defect detection in paper drum production lines. The goal is to detect quality defects on the surface of paper drums, such as cracks, dents, or localized damage, using a machine vision system. This embodiment aims to use an algorithm based on the magnitude of curvature abrupt changes to analyze images of the paper drum surface and thus identify defects.

[0079] The system configuration includes: Camera: 4096x2160 resolution, 30fps, used to capture real-time images of the paper drum surface.

[0080] Light source: Uniform white light with an intensity of 300 lux is used to ensure stable image quality.

[0081] Computer: Equipped with an 8-core CPU, 32GB of memory and an NVIDIA 2080 GPU for image processing and algorithm calculations.

[0082] Images of the paper drum surface are captured by a camera and transmitted to a computer for processing. Three images are taken from different angles for each paper drum to ensure that the entire surface is covered.

[0083] The acquired images were processed as follows: Grayscale conversion: Converting a color image to a grayscale image using the standard formula: ; Gaussian filtering is used with a convolution kernel size of 5x5 and a standard deviation of 1 to remove noise interference.

[0084] Edge detection algorithms (such as Canny edge detection) are used to extract edge information from the surface of the paper drum, and then the curvature is calculated. The formula for calculating the curvature is: ; Where κ represents curvature, y is the vertical coordinate of a point on the surface, and x is the horizontal coordinate.

[0085] Collect curvature abrupt change data for 1000 defect-free paper drums. Using the process described above, extract the abrupt change amplitude for each paper drum and calculate its mean and standard deviation.

[0086] The mean value of the curvature abrupt change amplitude data for defect-free paper drums is: μ=0.15; The standard deviation is: σ=0.05; The tolerance is set to "mean + 2.33 standard deviations," which corresponds to a 99% confidence level, meaning 95% of the data will fall within this range. The tolerance value is calculated as follows: The threshold for defect detection is set at 0.267. When the curvature change exceeds this value, it is judged as a defect.

[0087] Select a newly acquired image of a paper bucket and perform the above curvature calculation and abrupt change amplitude extraction.

[0088] Calculate the curvature abrupt change of the paper drum and compare it with the tolerance value. For example, if the calculated abrupt change is 0.35, which is significantly higher than the set tolerance of 0.267, the paper drum is determined to be defective.

[0089] Based on the magnitude of the mutation, the defect type can be further analyzed: If the mutation amplitude is greater than 0.35, it indicates that there may be a large crack.

[0090] If the mutation range is between 0.267 and 0.35, it may indicate a slight surface dent or a minor crack.

[0091] To verify the effectiveness of the method of this invention, a comparative experiment was conducted with traditional detection methods. Traditional methods use a thresholding method based on surface images (simple grayscale comparison) with a tolerance value of 0.2.

[0092] In 100 randomly selected paper drum samples, the traditional method detected 80 defective samples, while the method of the present invention detected 90 defective samples, significantly reducing the false negative rate.

[0093] The traditional method has a false negative rate of 20% and a false positive rate of 5%.

[0094] The method of this invention reduces the false negative rate to 10% and keeps the false positive rate below 5%.

[0095] The method of this invention significantly reduces the false negative rate and greatly improves the accuracy.

[0096] The calculation and tolerance setting of curvature change amplitude can effectively distinguish different types of defects, improving the accuracy and stability of detection.

[0097] Based on the production line's quality requirements, further optimize the tolerance factor k (a multiple of the standard deviation) to balance the false negative rate and the false positive rate. For example, if the production line has extremely strict requirements for the false negative rate, the tolerance can be appropriately set to "mean + 2 times the standard deviation", that is: ; This adjustment will reduce the false negative rate, but may cause a slight increase in the false positive rate.

[0098] This embodiment demonstrates the specific application process of a paper drum defect detection method based on curvature abrupt change amplitude, and experimental data proves that this method has significant advantages in improving detection accuracy and reducing the false negative rate. Through specific parameter settings and algorithm calculations, this invention can accurately detect various defects on the surface of paper drums, and reasonably adjust the detection sensitivity according to the tolerance setting, thereby achieving efficient and accurate defect detection.

[0099] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0100] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for 100% defect inspection of paper drums based on machine vision, characterized in that, include: Step 1: Constructing an adaptive optics environment for curved surfaces: A ring-shaped low-angle strip light source array is arranged at equal intervals along the circumference of the paper drum. The emission direction of each light source forms a dynamically adjusted acute angle with the normal to the surface of the paper drum. This acute angle decreases as the diameter of the paper drum increases. Near-infrared coaxial diffuse light sources are deployed at both ends of the paper drum along its axial direction. Step 2: Motion-compensated 3D reconstruction: The conveyor line speed and paper drum rotation angular velocity are acquired in real time, and the acquisition frequency of the line scan camera is dynamically calculated based on the line speed to ensure that the overlap rate of adjacent scan strips meets the requirements of three-dimensional reconstruction. Perform pixel-level spatiotemporal calibration based on motion parameters: map the scanned image to a virtual cylindrical coordinate system and restore motion blur; Reconstruct the three-dimensional point cloud of the outer surface of the paper drum and establish a reference coordinate system with the geometric center of the drum as the origin and the axis of the drum as the Z-axis; Step 3: Multimodal defect hierarchy decision: The 3D point cloud is unfolded into a 2D image along the circumference, and a dynamic reference grid is constructed on the 2D image with the grid cell size adaptively adjusted according to the pattern complexity. Collaborative extraction of geometric structural features, surface texture features, and adhesive seam features; Defect classification is output using a three-level decision tree: Level I structural defects are determined by comparing the magnitude of curvature abrupt changes with the structural deformation limit. Level II determines printing defects based on the comparison of the Mahalanobis distance between the entropy anomaly area and the normal area and the texture difference threshold; Level III is used to determine adhesive defects based on the discontinuity of the differential peak value of the adhesive joint and the grayscale profile, which meet the conditions for poor adhesive bonding. The structural variation limit, texture difference threshold, and poor bonding conditions were all determined statistically using a defect-free sample set.

2. The method for full inspection of paper drum defects based on machine vision according to claim 1, characterized in that, The dynamic adjustment of the set acute angle includes: Establish a geometric mapping relationship between the diameter of the paper drum and the incident angle of the light source: when the diameter of the paper drum increases, reduce the angle between the light source emission direction and the normal to the drum surface to compensate for the illuminance attenuation in the middle of the drum. The radius of the paper drum is obtained in real time by a laser rangefinder, and the required acute angle value is queried according to the pre-calibrated radius-angle mapping curve. The pitch adjustment mechanism of the driving light source unit ensures that the error between the actual incident angle and the query value is less than the angle tolerance. The radius-angle mapping curve is generated through optical simulation and experimental calibration: under a standard light source array, the illuminance distribution on the surface of paper drums with different radii is measured, and the optimization target is to achieve a set ratio of illuminance between the middle and edge of the drum. The optimal incident angle is then iteratively solved.

3. The method for full inspection of paper drum defects based on machine vision according to claim 1, characterized in that, The dynamic calculation of the acquisition frequency includes: Based on the angular velocity of the paper drum rotation and the minimum number of sampling points required in the circumferential direction, calculate the minimum number of image frames required for a single rotation. The number of axial scan rows is determined based on the conveyor line speed and the axial length of the paper drum; The final acquisition frequency is obtained by multiplying the product of the minimum number of frames and the number of scan lines by an overlap safety factor greater than 1. The minimum number of sampling points is determined by the accuracy requirements of 3D reconstruction: the sampling points are gradually increased during the reconstruction test, and the minimum number of points corresponding to the time when the deviation between the reconstructed model and the actual size is less than the set error is the required value.

4. The method for full inspection of paper drum defects based on machine vision according to claim 1, characterized in that, The motion blur restoration process includes: The motion blur direction and blur kernel length are calculated based on the rotational angular velocity and exposure time. A point spread function model matching the size of the blur kernel is constructed, which considers the nonlinearity of pixel displacement caused by the curved surface of the paper barrel; The image is restored using a regularized deconvolution algorithm, and the deblurring effect and noise suppression are balanced by adjusting the regularization coefficient; The determination of the regularization coefficient includes: acquiring degraded images on a standard motion blur test chart, using the combined score of edge sharpness and noise level of the restored image as an indicator, and iteratively optimizing the coefficient value.

5. The method for full inspection of paper drum defects based on machine vision according to claim 1, characterized in that, The generation of the dynamic reference grid includes: Extract scale-invariant feature points of printed patterns from 2D unfolded images; An initial triangular mesh is generated using feature points as nodes; The weighted sum of the gray-level variance and the mean gradient magnitude within each grid cell is used as a complexity index. When the complexity index exceeds the adaptive threshold, recursively subdivide the grid until the detection accuracy requirement is met; The adaptive threshold is dynamically set based on the overall image complexity distribution: the image is divided into several blocks, and the percentage quantile of the highest complexity within each block is used as the threshold benchmark.

6. The method for full inspection of paper drum defects based on machine vision according to claim 1, characterized in that, The extraction of the curvature abrupt change amplitude includes: Take circumferential point cloud sections at equal intervals along the axial direction of the paper drum; Perform circle fitting on the point cloud of each section and calculate the sequence of fitting radii; Perform sliding window statistical analysis on the radius sequence and calculate the standard deviation of the radius within the window; When the number of windows in which the standard deviation continuously exceeds the fluctuation tolerance is greater than the set value, it is determined that there is a curvature change in the region; The fluctuation tolerance is determined by statistical analysis of the radius fluctuation of defect-free samples: the maximum standard deviation of all samples is calculated and multiplied by a safety factor to obtain the tolerance value.

7. The method for full inspection of paper drum defects based on machine vision according to claim 1, characterized in that, The determination of poor bonding conditions includes: Peak detection is performed on the differential curve of the adhesive joint to eliminate spurious peaks caused by noise. The effective peak spacing distribution is statistically analyzed, and an early warning is triggered when the spacing is less than the minimum adhesive discontinuity distance. Extract grayscale profiles at the warning locations and calculate the discontinuity index of the profile signal; When the discontinuity index exceeds the dynamic threshold, a bonding defect is confirmed. The minimum adhesive discontinuity distance was determined through destructive experiments: defect samples with different discontinuity lengths were prepared, and the minimum length that could be reliably detected by machine vision was used as the benchmark value.

8. The method for full inspection of paper drum defects based on machine vision according to claim 1, characterized in that, Perform cross-modal validation before outputting defect classification: Inversely map the coordinates of the defect region in the 2D image to the 3D point cloud space; Calculate the depth gradient of the defect region in three-dimensional space; When the depth change gradient is less than the optical interference tolerance, the original defect determination is cancelled. The determination of optical interference tolerance includes: applying simulated illumination interference to the surface of a defect-free sample, measuring the maximum value of the spurious depth change caused by it, and using twice that value as the tolerance benchmark.

9. A method for full inspection of paper drum defects based on machine vision according to claim 1, characterized in that, The determination of the structural variation limit includes: Collect a dataset of curvature abrupt change amplitudes from defect-free paper drum samples; Calculate the mean and standard deviation of the dataset; The tolerance value is the mean plus a certain number of standard deviations, and this multiple is adjusted according to the defect missed detection rate requirements. The defect missed detection rate is required to be converted into statistical confidence level through production line process standards, and then mapped to the standard deviation multiple.

10. A machine vision-based method for full inspection of defects in paper drums according to claim 1, characterized in that, The determination of the texture difference threshold includes: Extract the texture feature vectors of defect-free samples and typical defective samples respectively; Calculate the Mahalanobis distance distribution between the two types of feature vectors; The initial threshold is set at the preset high quantile of the distance distribution of defect-free samples; The threshold is dynamically fine-tuned based on the actual false alarm rate. The preset high quantile is determined through ROC curve analysis: the defect detection rate and false alarm rate corresponding to different quantiles are tested on the validation set, and the minimum quantile that meets the false alarm rate requirement is selected.

Images