Methods and systems for detecting surface defects in nonwoven fabrics

Abstract

The application discloses a non-woven fabric surface defect detection method and system, and relates to the field of image processing. The method comprises the following steps: synchronously collecting a non-woven fabric surface image and a transmission image through a double camera; constructing a space mapping coordinate system through a checkerboard calibration; realizing accurate association of two image pixels; based on the coordinate system, extracting the shrink ring width of calendered rhombus nodes and the diagonal length of the nodes from the surface image, and extracting the aperture of a transparent window and the number of frame bundle stacks from the transmission image; eliminating the influence of size difference by constructing a coupling vector; quantifying the imbalance degree of defects based on the coupling vector offset pointer; comparing the defect proportion with a preset threshold value to determine the risk level and generate a production control instruction; the scheme realizes cross-modal coupling analysis from the non-woven fabric surface to the internal structure, improves the defect detection precision, reduces the defect omission rate, and supports the automatic closed-loop management and control of production quality.

Description

Technical Field

[0001] This invention relates to the field of image processing, and more specifically to a method and system for detecting defects on the surface of nonwoven fabrics. Background Technology

[0002] Nonwoven fabrics, as flexible materials with properties such as breathability, flexibility, and acid and alkali resistance, are widely used in medical and health, packaging and protection fields. Nonwoven fabrics formed through calendering, in particular, possess both high tensile strength and breathability due to the regular diamond-shaped node structure formed on their surface. During production, the integrity of the calendered diamond nodes directly determines product performance: the width of the node shrinkage ring reflects the uniformity of fiber distribution, the aperture of the perforated window affects breathability, and the number of overlapping layers of the frame bundle is related to structural strength. The synergistic matching of these three factors is the core of ensuring nonwoven fabric quality. However, nonwoven fabric production often employs high-speed continuous production lines, making traditional inspection methods ill-suited for the real-time quality monitoring demands of high-speed production. Conventional machine vision inspection only analyzes surface images individually, making it difficult to establish the correlation between surface structure and internal pores, resulting in quality judgment relying on experience and limited accuracy.

[0003] The surface and internal pores of calendered rhomboid nodes are a synergistic manifestation of calendering process parameters, and an imbalance between the two directly reflects process abnormalities. However, existing detection technologies have not established a spatial mapping relationship between surface images and transmission images, making it impossible to accurately extract structural data and internal pore data of nonwoven fabric surfaces. Furthermore, existing detection technologies often extract nonwoven fabric image data from a single dimension, lacking correlation judgment. This results in a false negative rate of over 15% for surface node structures exceeding standard ranges and pore imbalance defects caused by uneven calendering forces during high-speed production, severely impacting the stringent quality requirements for nonwoven fabrics in high-end fields such as medical and health care. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a method and system for detecting surface defects in nonwoven fabrics, solving the problem that these technologies cannot accurately extract structural data and internal pore data from the surface of nonwoven fabrics.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] Acquire images of the nonwoven fabric surface and transmission images of the nonwoven fabric;

[0007] The shrinkage ring width and diagonal length of the calendered rhomboid node are extracted from the surface image of the nonwoven fabric. The calendered rhomboid node is a rhomboid texture unit formed by the calendering process of the nonwoven fabric, which includes a central open window and a surrounding fiber shrinkage ring. The diagonal direction of the calendered rhomboid node is consistent with the direction of calendering force transmission.

[0008] The aperture of the transparent window and the number of frame bundles are extracted from the nonwoven fabric transmission image. The transparent window is a pore structure formed at the center of the calendered rhomboid node due to the fiber being squeezed to the edge during the calendering process, including the boundary of the pore itself and the surrounding fiber frame.

[0009] A coupling vector is constructed based on the ratio of the shrinkage ring width to the node diagonal length and the ratio of the aperture of the transparent window to the number of stacked layers of the frame bundle;

[0010] The offset index is obtained by summing the differences based on the coupling vector, and the defect is determined by the offset index and the preset offset threshold.

[0011] Preferably, the surface image of the nonwoven fabric is acquired by a first linear array camera above the direction of nonwoven fabric movement, and the transmission image of the nonwoven fabric is acquired by a second linear array camera below the direction of nonwoven fabric movement.

[0012] A spatial mapping relationship between the nonwoven fabric surface image and the nonwoven fabric transmission image is established by using a pre-set checkerboard calibration plate, and a spatial mapping coordinate system is constructed based on the spatial mapping relationship.

[0013] Preferably, the surface calibration image and transmission calibration image of the checkerboard calibration plate are acquired simultaneously by the first line array camera and the second line array camera;

[0014] The pixel coordinates of checkerboard corner points in the surface calibration image and the transmission calibration image are extracted by the Harris corner detection algorithm, and the surface corner point set and the transmission corner point set are obtained respectively.

[0015] The homography matrix is ​​solved by the least squares method, and the spatial mapping relationship between the transmission corner point set and the surface corner point set is constructed by multiplying the homography matrix with the transmission corner point set. The spatial mapping relationship between the transmission corner point set and the surface corner point set is then converted into a spatial mapping table.

[0016] The pixels of the nonwoven fabric surface image are mapped to the checkerboard corner points of the surface calibration image, and the top left pixel of the nonwoven fabric surface image is set as the origin of the initial spatial mapping coordinate system. The nonwoven fabric running direction is set as the X-axis and the width direction is set as the Y-axis.

[0017] The pixels of the nonwoven fabric transmission image are mapped to the corner points of the transmission calibration image checkerboard, and then transformed to the initial spatial mapping coordinate system through a spatial mapping table to obtain a spatial mapping coordinate system that contains a one-to-one correspondence between the pixel coordinates of the nonwoven fabric surface image and the pixel coordinates of the nonwoven fabric transmission image.

[0018] Preferably, grayscale projection is performed on the nonwoven fabric surface image along the 45° and 135° diagonal directions of the nonwoven fabric running direction and the width direction, respectively, to extract brightness peaks that are distributed at 45° and 135°. For each brightness peak, the center line of the brightness peak is determined by Gaussian fitting of the center coordinate distribution of the peak, which serves as the diagonal of the calendering diamond node.

[0019] Based on the design spacing D of the rolled diamond nodes, each brightness peak is divided into a local window with a length of 1.5D. Within each local window, the local grayscale peak and the grayscale valley in its ±0.3D neighborhood are detected, and the peak-valley grayscale difference is calculated. When the peak-valley grayscale difference is ≥ the preset peak-valley difference threshold, the peak is determined as the center point of the corresponding rolled diamond node. The coordinates of the center point in each local window on the spatial mapping coordinate system are recorded, and the mean grayscale value G of the 3×3 neighborhood of the center point is calculated to obtain the set of center points containing the coordinates and the mean grayscale value.

[0020] Set the center point as the center of the circle, and expand outwards sequentially according to the preset radial step size to form K sampling circles. The radial step size is the difference in radius between adjacent sampling circles. Count the number of pixels N in each sampling circle and record the gray value corresponding to each pixel. When traversing to the Mth sampling circle, mark the number of pixels with gray value ≤ G on the circle as N1. When N1 ≥ N multiplied by the proportion threshold, the Mth sampling circle is determined as the inner boundary of the fiber shrinkage ring.

[0021] The shrinkage ring width W of the rolled rhomboid node is obtained by calculating the product of the radial step size and M, and the diagonal length L of the node is obtained by calculating the Euclidean distance between the opposite vertices on the diagonal.

[0022] Preferably, the coordinates of the center point of the calendered rhombus node in the nonwoven fabric surface image are mapped to the nonwoven fabric transmission image through a spatial mapping coordinate system, so as to obtain the coordinate set of the center point on the nonwoven fabric transmission image. The coordinate set is connected by a Deloitte triangulation, and the smallest rhombus quadrilateral formed by every four coordinates is taken as the search window.

[0023] The search window is divided into a central region and a ring-shaped region. The average gray value of the central region and the ring-shaped region are calculated respectively. When the value obtained by subtracting the average gray value of the ring-shaped region from the average gray value of the central region is greater than the preset gray value difference threshold, it is determined that the search window contains a transparent window.

[0024] Expand pixel by pixel outward from the edge of the central region, calculate the gray-level gradient of adjacent pixels in each direction, and stop pixel expansion in that direction when the gray-level gradient of N2 consecutive pairs of adjacent pixels in the same direction is less than or equal to the preset gradient threshold. Mark the coordinates of the first pixel in the N2 consecutive pairs of adjacent pixels as the boundary coordinates and the last coordinate as the border coordinates.

[0025] When pixel expansion stops in all directions, the closed contour formed by all marked boundary coordinates is taken as the transparency window boundary. The transparency window boundary is fitted by the minimum bounding rhombus algorithm, and the length of the short diagonal of the fitted minimum bounding rhombus is calculated to obtain the transparency window aperture V.

[0026] The region formed between the closed contour formed by all the bounding box coordinates and the boundary of the transparent window is constructed as an annular sampling band. The annular sampling band is divided into 4 regions. The fiber bundle texture period in different regions of the annular sampling band is extracted by Gabor filter. The number of local maxima of the Gabor filter response to the fiber bundle texture period in each region is counted. The average of the number of local maxima in the 4 regions is taken as the number of bounding box bundle stacks C.

[0027] Preferably, K1 is obtained by calculating the ratio of the width of the shrinkage ring corresponding to each center point in the center point set on the nonwoven fabric surface image to the diagonal length of the node;

[0028] The ratio of the aperture of the transparent window corresponding to each center point in the center point set on the nonwoven fabric transmission image to the number of layers of the frame bundle is calculated to obtain K2;

[0029] The coordinates of the center points corresponding to K1 and K2, and the values ​​of K1 and K2, form the coupling vector F.

[0030] Preferably, the offset index is obtained by calculating the sum of the absolute values ​​of the difference between K1 and the reference value in the coupling vector and the absolute values ​​of the difference between K2 and the calibration constant K0, where the reference value is the design ratio of the shrinkage ring width to the diagonal length of the node.

[0031] A preset offset threshold is set. If the offset exponent of the i-th coupling vector is greater than the offset threshold, the center point corresponding to the i-th coupling vector is marked as an imbalance defect; otherwise, it is marked as qualified.

[0032] Count the number of imbalance defects and implement risk control based on the number of imbalance defects.

[0033] Preferably, the number of center points marked as imbalance defects in the set of center points of the nonwoven fabric surface image is counted, and the number of center points marked as imbalance defects is divided by the total number of center points to obtain the defect ratio;

[0034] A preset defect threshold is set. When the defect ratio exceeds the defect threshold, it is judged as a level 1 risk, a control instruction is generated, and the coordinate set of the judgment time and the defect center point is recorded.

[0035] Preferably, a nonwoven fabric surface defect detection system is proposed to implement the defect detection method as described in any of the above, including:

[0036] The image acquisition module includes a first subunit, a second subunit, and a matching light source. The first subunit is a first linear array camera set above the nonwoven fabric running direction, and the second subunit is a second linear array camera set below the nonwoven fabric running direction. The matching light source is a backlight installed above the nonwoven fabric running direction. The first subunit and the second subunit acquire images of the nonwoven fabric surface and the nonwoven fabric transmission image simultaneously through a dual-camera synchronous triggering mechanism.

[0037] The data extraction module is used to construct a spatial mapping coordinate system for the acquired nonwoven fabric surface image and nonwoven fabric transmission image, and extract the shrinkage ring width and node diagonal length of the calendered rhomboid node from the nonwoven fabric surface image, and extract the aperture of the transparent window and the number of frame bundle stacks from the nonwoven fabric transmission image.

[0038] The defect determination module and the risk determination module are used to construct a coupling vector based on the extracted shrinkage ring width, node diagonal length, transparent window aperture and the number of frame bundle stacks, and calculate the offset index based on the coupling vector, and determine the defect based on the offset index.

[0039] The risk control module is used to statistically analyze the results of defect assessment, obtain the defect ratio, determine the risk level of the nonwoven fabric production line based on the defect ratio, and generate control instructions.

[0040] Compared with existing technologies, it has the following advantages:

[0041] The proposed nonwoven fabric surface defect detection method and system addresses the issue of spatial misalignment between the nonwoven fabric surface and internal features caused by differences in acquisition perspective. At the data acquisition level, it utilizes dual-camera synchronous triggering and checkerboard calibration to construct a spatial mapping coordinate system, controlling the pixel coordinate error between the surface image and the transmitted image within ±0.1 pixels. For example, the center point of the calendered rhomboid node in the surface image can be accurately mapped to the corresponding position in the transmitted image, providing a coordinate reference for cross-modal feature association and improving the subsequent feature matching accuracy to over 98%. At the feature extraction level, grayscale projection and circumferential sampling quantify the shrinkage ring width W and the node diagonal length L, achieving a quantitative description of the surface fiber distribution. Based on spatial mapping and positioning of the transparent window, grayscale gradient and Gabor filtering are used to extract the aperture V and the number of frame bundle stacks C, transforming the internal pore and fiber stacking features into calculable parameters. Compared to traditional visual inspection, the measurement errors of W, L, V, and C are ≤1 pixel, solving the industry pain point of abundant qualitative descriptions and limited quantitative analysis. At the coupling analysis level, coupling vectors K1 and K2 are constructed to eliminate the influence of node size differences, achieving cross-modal correlation between surface shrinkage and internal porosity and fiber layer number. At the risk control level, the degree of imbalance is quantified by the offset index, and by calculating the defect ratio and comparing it with a threshold, hierarchical management from single-point defects to global risks is achieved.

[0042] In summary, this solution achieves a closed-loop defect detection system for nonwoven fabrics by using a nonwoven fabric defect detection method. Based on the nonwoven fabric defect detection system, it upgrades nonwoven fabric defect detection from traditional qualitative judgment to quantitative, cross-modal, and fully automated control, significantly improving detection accuracy and production stability. It is especially suitable for quality monitoring in high-requirement fields such as medical and health care. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0044] Figure 2 This is a schematic diagram of the system framework of the present invention; Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] Please see Figure 1 This application provides a method for detecting surface defects in nonwoven fabrics;

[0047] The method specifically includes the following steps:

[0048] Step 1: Deploy a first line array camera above the nonwoven fabric running direction, in conjunction with a coaxial backlight, to acquire images of the nonwoven fabric surface. Deploy a second line array camera below the nonwoven fabric running direction to acquire images of the nonwoven fabric transmitted through the coaxial backlight. Specifically, the first and second line array cameras are exposed synchronously using the same trigger signal to ensure that the nonwoven fabric surface image and the nonwoven fabric transmitted image acquired at the same time correspond to the same area of ​​the nonwoven fabric (spatial position matching), thus avoiding image misalignment caused by the high-speed operation of the production line.

[0049] Step 2: Using a pre-set checkerboard calibration board (the spacing between the checkerboard corner points is known, with an accuracy of ±0.01mm), place it in the inspection area of ​​the nonwoven fabric production line. Use a first line scan camera to acquire a surface calibration image (the surface reflection image of the calibration board), and use a second line scan camera to acquire a transmission calibration image (the transmission image of the calibration board). Specifically, the checkerboard calibration board used should be made of a semi-transparent material. After acquiring the calibration images, remove it from the inspection area.

[0050] The pixel coordinates of checkerboard corner points in the surface calibration image and the transmission calibration image are extracted by the Harris corner detection algorithm, and the surface corner point set and the transmission corner point set are obtained respectively.

[0051] The homography matrix H (3×3 matrix) is solved by the least squares method. The spatial mapping relationship between the transmission corner point set and the surface corner point set is constructed by multiplying the homography matrix with the transmission corner point set. The spatial mapping relationship between the transmission corner point set and the surface corner point set is converted into a spatial mapping table (recording the coordinate mapping relationship between each pixel of the surface image and the corresponding pixel of the transmission image), realizing a one-to-one correspondence between the surface image pixels and the transmission image pixels.

[0052] The pixels of the nonwoven fabric surface image are mapped to the checkerboard corner points of the surface calibration image. The top left pixel of the nonwoven fabric surface image is set as the origin of the coordinate system. The nonwoven fabric running direction (along the length of the production line) is set as the X-axis, and the width direction (along the width of the production line) is set as the Y-axis. In this way, the initial spatial mapping coordinate system is constructed.

[0053] The pixels of the nonwoven fabric transmission image are mapped to the corner points of the transmission calibration image checkerboard, and then transformed to the initial spatial mapping coordinate system through the spatial mapping table, finally forming a spatial mapping coordinate system that contains a one-to-one correspondence between the pixel coordinates of the nonwoven fabric surface image and the pixel coordinates of the nonwoven fabric transmission image.

[0054] Specifically, by unifying the coordinate system, the pixel coordinates of the nonwoven fabric surface image and the nonwoven fabric transmission image are standardized, so that the subsequently extracted surface features (such as the width of the shrinkage ring and the diagonal length of the node) and transmission features (such as the aperture of the transparent window and the number of stacked layers of the frame bundle) can be associated in the same coordinate system, providing a unified spatial benchmark for the correlation analysis between the nonwoven fabric surface and the interior.

[0055] Step 3: Along the 45° and 135° diagonal directions of the nonwoven fabric running direction and width direction, perform grayscale projection on the nonwoven fabric surface image (accumulate the pixel grayscale values ​​along this direction) to obtain two intersecting grayscale projection curves. Extract the continuous areas with significantly higher grayscale values ​​than the background from the curves, which are the brightness peaks intersecting at 45° and 135° (corresponding to the diagonals of the calendering diamond nodes, as the fibers are dense and reflective at the diagonals). For each brightness peak, determine the brightness by Gaussian fitting the center coordinate distribution of the peak (the grayscale distribution of the peak conforms to a Gaussian distribution, and Gaussian fitting reduces the interference of noise on the peak center location). The center line of the peak array serves as the diagonal of the calendered rhombus node. Specifically, the calendered rhombus node is a rhombus texture unit formed by calendering (roller extrusion) of non-woven fabric. It includes a central open window (pore) and surrounding fiber shrinkage rings (border fiber bundles). Its diagonal direction is consistent with the direction of calendering force transmission (45°, 135°). By utilizing the directionality of the diagonal of the calendered rhombus node (45°, 135°), the peak array is accurately extracted through grayscale projection, solving the problem of difficult identification of rhombus structures in complex backgrounds. Gaussian fitting reduces the influence of noise, making the positioning error of the diagonal center line ≤0.5 pixels, providing an accurate benchmark for subsequent node segmentation.

[0056] Based on the design spacing D of the calendered rhombus nodes (i.e., the distance between the centers of adjacent nodes along the diagonal direction, determined by production process parameters), each brightness peak is divided into local windows of length 1.5D (ensuring each window contains one complete node). Within each local window, the local grayscale peak (candidate center point) and its ±0.3D neighborhood grayscale valley (low grayscale area around the peak) are detected, and the peak-valley grayscale difference is calculated. When the peak-valley grayscale difference is ≥ the preset peak-valley difference threshold (the peak-valley difference threshold is obtained by statistically analyzing the peak-valley differences of 100 qualified nodes, taking the mean + 3 times the standard deviation), the peak is determined as the center point of the corresponding calendered rhombus node. The coordinates of the center point in each local window on the spatial mapping coordinate system are recorded, and the 3×3 neighborhood grayscale mean G of the center point is calculated to obtain the set of center points containing the coordinates and grayscale mean. Specifically, the local window segmentation ensures that a single node is analyzed independently, avoiding interference from adjacent nodes. The peak-valley difference threshold effectively filters noise (such as false peaks caused by fiber lint), improving the accuracy of center point detection.

[0057] Set the center point as the center of the circle, and expand outwards sequentially according to the preset radial step size to form K sampling circles (K is the maximum number of expansions to ensure coverage of the node edge). The radial step size is the radius difference between adjacent sampling circles (which can be 1 pixel). Count the number of pixels N in each sampling circle and record the gray value corresponding to each pixel (starting from a center, form K circles around this center, with the number of pixels N on the circle gradually increasing). When traversing to the Mth sampling circle, mark the number of pixels on this circle with a gray value ≤ G (the gray value mean of the 3×3 neighborhood of the center point) as N1. When N1 ≥ N multiplied by the proportion threshold (the proportion threshold is obtained by collecting 100 samples of gray values ​​of the shrinkage ring on the surface of the nonwoven fabric, calculating the statistical relationship between the mean of its central neighborhood and the gray value of the inner boundary of the ring, and fitting the threshold), then the Mth sampling circle is determined as the inner boundary of the fiber shrinkage ring.

[0058] The shrinkage ring width W of the calendered rhombus node is obtained by calculating the product of the radial step size and M (W reflects the uniformity of edge fiber distribution; if W is too large, it indicates that the edge fibers are loose, and if it is too small, it may be over-compressed). The diagonal length L of the node is obtained by calculating the Euclidean distance between the opposite vertices on the diagonal. Specifically, along the diagonal (brightness peak line center line) of the calendered rhombus node, extending from the center point to both ends, the pixel position where the gray value first drops from ≥0.8G to ≤0.5G is detected as the opposite vertex on the diagonal. L includes two diagonals, L1 and L2, which are two diagonals along 45° and 135° in the calendered rhombus node, respectively. In this example, the diagonal L1 along 45° is selected as the diagonal length L of the node.

[0059] Specifically, this step uses the directionality of the rhombus nodes to accurately locate the structural skeleton and achieves objective judgment of features through peak-valley differences, avoiding reliance on human experience. The quantitative parameters provide a surface benchmark for joint analysis and are the key input for subsequent defect judgment. This step transforms the microstructural features of the rolling rhombus nodes into calculable digital parameters, solving the problem of more qualitative descriptions and less quantitative analysis in traditional detection.

[0060] Step 4: Map the coordinates of the center point of the calendered rhombus node in the nonwoven fabric surface image to the nonwoven fabric transmission image using a spatial mapping coordinate system. This yields the set of coordinates of the center point on the nonwoven fabric transmission image. Connect the coordinate sets using a Deloitte triangulation. Take the smallest rhombus quadrilateral formed by every four coordinates as the search window. Specifically, the search window is used to define the rhombus region within the detection range of the transparent window. Its side length is determined by the design spacing D between adjacent calendered rhombus nodes. For example, if the side length is 1.2D, it ensures coverage of the transparent window and its surrounding frame. The transparent window is a pore structure formed at the center of the calendered rhombus node due to the fibers being squeezed to the edge during the calendering process. It includes the boundary of the pore itself (defined by the gray-scale gradient change) and the surrounding fiber frame (related to the number of bundle layers).

[0061] The search window is divided into a central region occupying 40% of the search window area, which is a diamond-shaped core area centered on the mapping center point, and an annular region occupying the remaining 60% of the area (60% of the area can cover the annular sampling band). The grayscale mean values ​​of the central region and the annular region are calculated separately. When the grayscale mean value of the central region minus the grayscale mean value of the annular region is greater than a preset grayscale difference threshold (obtained by statistically analyzing 100 qualified transparent window samples and taking the mean value + 3 times the standard deviation), it is determined that the search window contains a transparent window. After the transparent window is transmitted through the light source, the central area is bright and the surrounding area is dark.

[0062] Expand pixel by pixel outward from the edge of the central region, calculate the gray-level gradient of adjacent pixels in each direction, and stop pixel expansion in that direction when the gray-level gradient of N2 consecutive pairs of adjacent pixels in the same direction is less than or equal to the preset gradient threshold. Mark the coordinates of the first pixel in the N2 consecutive pairs of adjacent pixels as the boundary coordinates and the last coordinate as the border coordinates.

[0063] Specifically, each direction represents the direction in which all pixels on the edge of the central region expand outward. Due to the existence of defects, the final expanded contour may not be the same as the shape of the central region. In this example, N2=5. The value of N2 cannot be greater than the border range of the transparent window (the border range is obtained from historical measurement data). Calculating continuous gray-level gradients can ensure stable gradient changes and avoid misjudgment caused by single-pixel noise. Multi-directional expansion can ensure the integrity of the boundary contour. By marking the coordinates of the first pixel in the adjacent pixels of N2 as the boundary coordinates and the last coordinate as the border coordinates, and then connecting all the boundary coordinates and all the border coordinates, a ring sampling band is formed. The ring sampling band is the border of the transparent window.

[0064] When pixel expansion stops in all directions, all marked boundary coordinates are fitted to a closed contour. The resulting closed contour is used as the boundary of the transparent window. The boundary of the transparent window is fitted by the minimum bounding rhombus algorithm. The length of the short diagonal of the fitted minimum bounding rhombus is calculated to obtain the aperture V of the transparent window.

[0065] Specifically, by fitting the boundary of the transparent window using the minimum bounding rhombus algorithm, the minimum bounding rhombus can match the actual shape of the transparent window, with an aperture calculation error of ≤2%. The lengths of the two diagonals of the minimum bounding rhombus are calculated. In this example, the shorter diagonal is selected as the aperture index, which can intuitively reflect whether the aperture of the nonwoven fabric transparent window is too large or too small.

[0066] The area formed between the closed contour formed by all border coordinates and the boundary of the transparent window is constructed as an annular sampling band. The annular sampling band is divided into 4 regions (a Cartesian coordinate system is established with the center region of the search window as the origin, and the 4 regions are located in the four quadrants of the coordinate system respectively). The fiber bundle texture period in different regions of the annular sampling band is extracted by a Gabor filter (wavelength of 8 pixels, direction consistent with the border direction of the region). The fiber bundle texture period is the periodic gray-scale change of fiber stacking. The number of local maxima of Gabor filter response to fiber bundle texture period in each region is counted (each maximum corresponds to one layer of fiber bundle). The average of the number of local maxima in the 4 regions is taken as the number of border bundle stacks C.

[0067] Specifically, this step solves the problem of difficult localization of transparent windows by using a search window based on spatial mapping. By combining grayscale gradient with texture analysis, it can simultaneously quantify pore size and fiber stacking density. The internal structure of the nonwoven fabric and the surface image complement each other, providing a collaborative basis for defect judgment. This step combines geometric constraints with image texture analysis to achieve automated and high-precision extraction of the microscopic features of transparent windows.

[0068] Step 5: Calculate the ratio of the shrinkage ring width to the diagonal length of the node corresponding to each center point in the set of center points on the nonwoven fabric surface image, to obtain K1. Specifically, K1 can reflect the proportion of the shrinkage ring width relative to the overall size of the node, and its physical meaning is the relative degree of surface fiber shrinkage. For example, if K1 is too large, it indicates that the shrinkage ring is too wide and the edge fiber distribution is loose.

[0069] The ratio of the aperture size of the transparent window to the number of stacked layers of the frame bundle for each center point in the center point set on the nonwoven fabric transmission image is calculated to obtain K2. Specifically, K2 can reflect the balance between the aperture size of the transparent window and the surrounding fiber stacking density. For example, if K2 is too large, it indicates that the aperture is too large and the fiber stacking is insufficient, which may pose a risk of fiber breakage.

[0070] The coupling vector F is formed by combining the coordinates of the center points corresponding to K1 and K2 and the values ​​of K1 and K2. Specifically, the coupling vector F obtained by combining the coordinates of K1, K2 and the corresponding center points in the spatial mapping coordinate system in a fixed format contains the external and internal structural parameters of the nonwoven fabric under the same physical node, providing a complete basis for subsequent determination of defect location and defect type by locating the defect through coordinates.

[0071] Step 6: The offset index is obtained by summing the absolute values ​​of the differences between K1 and the reference value (the reference value is the design ratio of the shrinkage ring width to the diagonal length of the node, which is directly determined by the calendering process design parameters) in the coupling vector, and the absolute values ​​of the differences between K2 and the calibration constant K0 (based on the parameters C and V of one hundred qualified calendering diamond nodes, the ratio of each qualified parameter C to V is calculated, and its arithmetic mean is obtained to obtain the calibration constant K0).

[0072] Specifically, the benchmark value is obtained directly from the process design parameters to ensure that the testing standards are consistent with the production goals. K0 is obtained through statistics of actual normal samples, automatically adapting to equipment characteristics and environmental differences to avoid deviations between theoretical values ​​and actual production. The calculated offset index represents the degree of comprehensive deviation of surface features and internal features in the coupling vector relative to the benchmark value. The larger the value, the more severe the imbalance.

[0073] A preset offset threshold is set (the offset threshold is the critical value that distinguishes between qualified and defective products. It is determined by analyzing the offset index distribution of historical defect samples, for example, taking the upper limit of the 95% confidence interval). If the offset index of the i-th coupling vector is greater than the offset threshold, the center point corresponding to the i-th coupling vector is marked as an imbalance defect (an imbalance defect is a node whose surface and internal features deviate from the threshold, corresponding to process problems such as uneven fiber distribution and calendering pressure fluctuations). Otherwise, it is marked as qualified.

[0074] Step 7: Count the number of center points marked as imbalance defects in the set of center points of the nonwoven fabric surface image, divide the number of center points marked as imbalance defects by the total number of center points to obtain the defect ratio;

[0075] A preset defect threshold is set. When the defect ratio is greater than the defect threshold, it is judged as a level 1 risk, a control instruction is generated, and the coordinate set of the judgment time and the defect center point is recorded.

[0076] When the defect ratio is less than or equal to the defect threshold, the condition is considered acceptable.

[0077] Specifically, in an acquired image of a nonwoven fabric surface, the number of defects on the image is calculated through the aforementioned series of steps. For example, if an image of a nonwoven fabric surface contains 1000 center points (corresponding to 1000 calendered diamond nodes), and its defect count is 100, then 100 of the calendered diamond nodes in the image have imbalance defects. Therefore, its defect ratio is 10%. The defect threshold is the critical value that distinguishes between the qualified state and the first-level risk state. However, different grades of nonwoven fabric require different defect thresholds for production. Therefore, the defect threshold can be determined through historical production data and quality standards. When the defect ratio is greater than the defect threshold, the nonwoven fabric defect surface detection system generates control commands, such as reducing the production line speed and adjusting the calendering roller pressure, and records the coordinates and judgment time of these imbalance defects. This facilitates manual location and analysis of defective nonwoven fabrics on the production line. When the defect ratio is less than or equal to the defect threshold, the current production parameters are maintained.

[0078] Furthermore, refer to Figure 2 As shown, a nonwoven fabric surface defect detection system is proposed to implement any of the defect detection methods described above, including:

[0079] The image acquisition module includes a first subunit, a second subunit, and a matching light source. The first subunit is a first linear array camera set above the nonwoven fabric running direction, the second subunit is a second linear array camera set below the nonwoven fabric running direction, and the matching light source is a backlight installed above the nonwoven fabric running direction. The first subunit and the second subunit acquire images of the nonwoven fabric surface and the nonwoven fabric transmission image simultaneously through a dual-camera synchronous triggering mechanism.

[0080] The data extraction module is used to construct a spatial mapping coordinate system for the collected nonwoven fabric surface image and nonwoven fabric transmission image, and extract the shrinkage ring width and node diagonal length of the calendered rhomboid node according to the nonwoven fabric surface image, and extract the aperture of the transparent window and the number of frame bundle stacks according to the nonwoven fabric transmission image.

[0081] The defect determination module, the risk determination module is used to construct a coupling vector based on the extracted shrinkage ring width, node diagonal length, transparent window aperture and frame bundle stacking number, and calculate the offset index based on the coupling vector, and determine the defect based on the offset index;

[0082] The risk control module is used to statistically analyze the results of defect determination, obtain the defect ratio, determine the risk level of the nonwoven fabric production line based on the defect ratio, and generate control instructions.

[0083] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for detecting surface defects of a nonwoven fabric, characterized by, The method comprises the following steps: acquiring a non-woven fabric surface image and a non-woven fabric transmission image; extracting the shrink ring width and the node diagonal length of the calendered rhombus node from the non-woven fabric surface image, the calendered rhombus node being a rhombus texture unit formed by the calendering process of the non-woven fabric, containing a central through window and a surrounding fiber shrink ring, and the diagonal direction of the calendered rhombus node being consistent with the calendering force conduction direction along the running direction of the non-woven fabric; performing gray projection on the non-woven fabric surface image at 45° and 135° along the running direction of the non-woven fabric, respectively, extracting the brightness peak columns distributed at 45° and 135°, for each brightness peak column, determining the center line of the brightness peak column as the diagonal line of the calendered rhombus node by Gaussian fitting the center coordinate distribution of the peak, and the diagonal direction being a fixed angle direction at 45° and 135° with the running direction of the non-woven fabric; extracting the through window aperture and the frame bundle stack number of the through window from the non-woven fabric transmission image, the through window being a pore structure formed at the center of the calendered rhombus node due to the fiber being extruded to the edge in the calendering process, containing the boundary of the pore itself and the fiber frame around the pore; constructing a coupling vector based on the ratio of the shrink ring width and the node diagonal length and the ratio of the through window aperture and the frame bundle stack number; calculating the sum of the differences of the coupling vector to obtain an offset index, and performing defect judgment according to a preset offset threshold and the offset index, the sum of the differences being the sum of the absolute value of the difference between the ratio of the shrink ring width and the node diagonal length and a reference value and the absolute value of the difference between the ratio of the through window aperture and the frame bundle stack number and a calibration constant, wherein the calibration constant is the arithmetic mean of the ratio of the through window aperture and the frame bundle stack number of the qualified calendered rhombus node.

2. The nonwoven fabric surface defect detection method according to claim 1, characterized by, The method comprises the following steps: acquiring a non-woven fabric surface image and a non-woven fabric transmission image; acquiring the non-woven fabric surface image by a first linear array camera above the running direction of the non-woven fabric, and acquiring the non-woven fabric transmission image by a second linear array camera below the running direction of the non-woven fabric; 3. The nonwoven fabric surface defect detection method according to claim 2, characterized by, establishing the spatial mapping relationship between the non-woven fabric surface image and the non-woven fabric transmission image by a preset checkerboard calibration board, and constructing a spatial mapping coordinate system based on the spatial mapping relationship. The method comprises the following steps: acquiring the non-woven fabric surface image by a first linear array camera above the running direction of the non-woven fabric, and acquiring the non-woven fabric transmission image by a second linear array camera below the running direction of the non-woven fabric; establishing the spatial mapping relationship between the non-woven fabric surface image and the non-woven fabric transmission image by a preset checkerboard calibration board, and constructing a spatial mapping coordinate system based on the spatial mapping relationship. The method comprises the following steps: synchronously acquiring the surface calibration image and the transmission calibration image of the checkerboard calibration board by the first linear array camera and the second linear array camera; extracting the pixel coordinates of the checkerboard corner points in the surface calibration image and the transmission calibration image by the Harris corner point detection algorithm, respectively obtaining a surface corner point set and a transmission corner point set; solving the homography matrix by the least square method, and constructing the spatial mapping relationship between the transmission corner point set and the surface corner point set by multiplying the homography matrix and the transmission corner point set, and converting the spatial mapping relationship between the transmission corner point set and the surface corner point set into a spatial mapping table; mapping the pixels of the non-woven fabric surface image to the checkerboard corner points of the surface calibration image, and setting the top-left corner pixel of the non-woven fabric surface image as the original point of the initial spatial mapping coordinate system, setting the running direction of the non-woven fabric as the X-axis, and setting the width direction as the Y-axis; The pixels of the non-woven fabric transmission image are mapped to transmission calibration image chessboard corner points, and are converted to an initial space mapping coordinate system through a space mapping table to obtain a space mapping coordinate system containing one-to-one correspondence between non-woven fabric surface image pixel coordinates and non-woven fabric transmission image pixel coordinates.

4. The nonwoven fabric surface defect detection method according to claim 1, characterized by, According to the non-woven fabric surface image, the shrink ring width and the node diagonal length of the calendered rhombic node are extracted, including: According to the design interval D of the calendered rhombic node, each brightness peak column is divided into a local window with a length of 1.5D; in each local window, the local gray scale peak and the gray scale valley in the neighborhood of the peak within a range of ±0.3D are detected, and the peak-valley gray scale difference is calculated; when the peak-valley gray scale difference is greater than or equal to a preset peak-valley difference threshold, the peak is determined as the center point of the corresponding calendered rhombic node, the coordinates of the center point in the space mapping coordinate system are recorded, and the 3*3 neighborhood gray scale average G of the center point is calculated to obtain a center point set containing coordinates and gray scale averages; The center point is set as the center of a circle, and is expanded outward in turn according to a preset radial step length to form K sampling circumferences, and the radial step length is the radius difference between adjacent sampling circumferences; the number of pixels N of each sampling circumference is counted, and the gray scale value corresponding to each pixel is recorded; when the Mth sampling circumference is reached, the number of pixels with a gray scale value less than or equal to G on the circumference is marked as N1; when N1 is greater than or equal to the threshold proportion of N, the Mth sampling circumference is determined as the inner boundary of the fiber shrink ring; The shrink ring width W of the calendered rhombic node is obtained by calculating the product of the radial step length and M, and the node diagonal length L is obtained by calculating the Euclidean distance between the relative vertices on the diagonal.

5. The nonwoven fabric surface defect detection method according to claim 4, characterized by, According to the non-woven fabric transmission image, the aperture of the transparent window and the number of frame beam stacks of the transparent window are extracted, including: The center point coordinates of the calendered rhombic node in the non-woven fabric surface image are mapped to the non-woven fabric transmission image through the space mapping coordinate system to obtain a coordinate set of the center point on the non-woven fabric transmission image, and the coordinate set is connected by a Delaunay triangulation network to form a search window in the form of a minimum rhombic quadrilateral with every four coordinates; The search window is divided into a central region and an annular region, and the gray scale averages of the central region and the annular region are calculated respectively; when the gray scale average of the central region minus the gray scale average of the annular region is greater than a preset gray scale difference threshold, it is determined that the search window contains a transparent window; The pixel expansion is stopped in the direction, and the coordinates of the first pixel in the continuous N2 pairs of adjacent pixels are marked as boundary coordinates, and the last coordinate is marked as a frame coordinate; When the pixel expansion in all directions is stopped, the closed contour formed by all the marked boundary coordinates is taken as the transparent window boundary, the transparent window boundary is fitted by a minimum circumscribed rhombus algorithm, the short diagonal length of the minimum circumscribed rhombus obtained by fitting is calculated, and the aperture V of the transparent window is obtained. The formed area between the closed contour formed by all the frame coordinates and the boundary of the through window is constructed as an annular sampling strip, the annular sampling strip is divided into four areas, the fiber bundle texture period in different areas of the annular sampling strip is extracted through a Gabor filter, the number of local maxima of the Gabor filter response fiber bundle texture period in each area is counted, and the average of the number of local maxima in the four areas is taken as the number of stack layers C of the frame bundle.

6. The nonwoven fabric surface defect detection method according to claim 5, characterized by, The coupling vector is constructed based on the ratio of the shrink ring width and the node diagonal length to the ratio of the through window aperture and the number of stack layers of the frame bundle, including: The ratio of the shrink ring width and the node diagonal length corresponding to each center point in the center point set on the non-woven fabric surface image is calculated to obtain K1; The ratio of the through window aperture and the number of stack layers of the frame bundle corresponding to each center point in the center point set on the non-woven fabric transmission image is calculated to obtain K2; The center point coordinates corresponding to K1 and K2 and the values of K1 and K2 form a coupling vector F.

7. The nonwoven fabric surface defect detection method according to claim 6, characterized by, The sum of the differences based on the coupling vector is calculated to obtain the offset index, and the defect is judged according to the preset offset threshold and the offset index, including: The offset index is obtained by calculating the sum of the absolute values of the differences between K1 and the reference value and the differences between K2 and the calibration constant K0, and the reference value is the designed ratio of the shrink ring width and the node diagonal length; The preset offset threshold, if the offset index of the i-th coupling vector is greater than the offset threshold, the center point corresponding to the i-th coupling vector is marked as an unbalanced defect, otherwise it is marked as qualified; The number of unbalanced defects is counted, and the risk control is carried out according to the number of unbalanced defects.

8. The nonwoven fabric surface defect detection method according to claim 7, characterized by, The number of unbalanced defects is counted, and the risk control is carried out according to the number of unbalanced defects, including: The number of center points marked as unbalanced defects in the center point set of the non-woven fabric surface image is counted, the number of center points marked as unbalanced defects is divided by the total number of center points to obtain the defect ratio; When the defect ratio is greater than the defect threshold, it is determined as a first-level risk, a control instruction is generated, and the time of judgment and the coordinate set of the defect center point are recorded.

9. A nonwoven fabric surface defect detection system for implementing the defect detection method according to any one of claims 1 to 8, characterized by, Including: The image acquisition module includes a first subunit, a second subunit and a matching light source, the first subunit is a first linear array camera arranged above the non-woven fabric running direction, the second subunit is a second linear array camera arranged below the non-woven fabric running direction, and the matching light source is a backlight source installed above the non-woven fabric running direction, the first subunit and the second subunit simultaneously collect the non-woven fabric surface image and the non-woven fabric transmission image through a double camera synchronous triggering mechanism; The data extraction module is used to construct a space mapping coordinate system for the collected non-woven fabric surface image and non-woven fabric transmission image, and to extract the shrink ring width and node diagonal length of the calendered rhombic node according to the non-woven fabric surface image, and to extract the through window aperture and the number of stack layers of the frame bundle according to the non-woven fabric transmission image; The defect judgment module is used to construct a coupling vector according to the extracted shrink ring width, node diagonal length, through window aperture and number of stack layers of the frame bundle, and to calculate the offset index according to the coupling vector, and to judge the defect according to the offset index; A risk control module is configured to count results of the defect determination, obtain a defect ratio, determine a risk level of the non-woven fabric production line according to the defect ratio, and generate a control instruction.

Images