Corrugated board thickness online detection method based on image recognition

Abstract

The present application relates to the field of image processing and online measurement technology, in particular to a corrugated board thickness online detection method based on image recognition. The method obtains a corrugated board end face image, analyzes the direction field of the end face extension direction, thickness direction and paperboard running direction to form a measurement coordinate system; generates an outer surface layer boundary candidate graph and a corrugated period skeleton graph in the coordinate system, configures a local thickness normal measurement window according to the phase interval of the corrugated unit, screens out abnormal boundary candidate points according to the boundary continuity, phase consistency and neighborhood thickness change relationship, and then fits the effective boundaries of the upper and lower surface layers and calculates the thickness along the local normal direction. The method constrains the outer boundary recognition and the corrugated period structure, reduces the misjudgment caused by the burr, glue layer shadow and core paper texture, and obtains the online thickness detection result distributed along the width direction of the paperboard.

Description

Technical Field

[0001] This invention relates to the field of image processing and online measurement technology, specifically to an online method for detecting the thickness of corrugated cardboard based on image recognition. Background Technology

[0002] Online thickness detection for corrugated cardboard is typically implemented during the cardboard forming, slitting, or conveying stages. A camera captures images of the cardboard's end face, and an image processing program identifies the outer contours of the upper and lower surfaces, converting the pixel distance between these contours into a thickness value. A common approach involves first converting the end face image to grayscale, filtering, and enhancing contrast, then using thresholding, edge detection, contour tracking, or line fitting to extract the outer boundary. To reduce the impact of local noise, some solutions select several measurement columns along the cardboard's width, searching for abrupt grayscale changes in each column. The upper grayscale change point is used as the upper surface boundary, and the lower grayscale change point as the lower surface boundary. The cardboard thickness is then obtained by averaging the results from multiple measurement columns or removing discrete points. Other solutions incorporate template matching or texture direction judgment to help confirm the distribution of corrugated lines, ensuring the detection area avoids obvious rough edges or shadowed areas. The basic idea of ​​the above scheme is still to use grayscale edges as the main measurement basis, regard the local brightness and darkness changes in the end face image as the basis for determining the outer boundary of the cardboard, and use the periodic texture of the internal corrugated core paper, the position of the flute peaks and valleys and the interlayer topological relationship as auxiliary identification information, without participating in the constraint and reverse verification of the outer boundary measurement area.

[0003] The main technical problem lies in the fact that the outer layer boundary and the internal corrugated core texture in the end-face image of corrugated cardboard are not always clearly separated. After the cardboard is cut, the end face may have fiber burrs, local indentations, adhesive layer shadows, and micro-tears at the cut, forming multiple gray-scale abrupt change bands near the outer boundary. The flute peaks, valleys, and folding ridges of the corrugated core will also produce periodic textures in the thickness direction, and the intensity of their gray-scale changes is sometimes close to that of the outer layer edge. Conventional edge detection in such images easily identifies internal flute textures, adhesive layer dark lines, or burr extensions as the outer layer boundary, causing the selection positions of the upper and lower boundaries to be offset. If a fixed flute template is used to limit the search range, when the cardboard has local compression, end-face tilt, or slight changes in flute pitch, the template phase deviates from the position of the actual flute unit, still causing the measurement window to fall into the wrong texture area. The essence of this problem is not insufficient sensitivity of the edge operator, but rather the lack of bidirectional constraint between the outer boundary recognition and the periodic structure of the corrugation. Simply relying on local gray-scale abrupt changes or a fixed template is difficult to stably determine the true thickness boundary in end-face images with interwoven burrs, shadows, and flute textures. Summary of the Invention

[0004] The purpose of this invention is to provide an online method for detecting the thickness of corrugated cardboard based on image recognition, which can solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] The online thickness detection method for corrugated cardboard based on image recognition includes: acquiring an image of the end face of the corrugated cardboard, performing orientation field analysis on the end face image in the end face extension direction, thickness direction and cardboard travel direction, and forming a measurement coordinate system;

[0007] Generate an outer surface boundary candidate map and a corrugated period skeleton map in the measurement coordinate system;

[0008] Based on the corrugated periodic skeleton diagram, the phase intervals of the corrugated units are divided, and local thickness normal measurement windows are configured for the phase intervals;

[0009] Within the local thickness normal measurement window, abnormal boundary candidate points are screened out based on boundary continuity, phase consistency, and neighborhood thickness variation relationship;

[0010] The remaining boundary candidate points are fitted with the effective boundaries of the upper and lower surfaces respectively, and the distance between the effective boundaries of the upper and lower surfaces is calculated along the local normal to obtain the online detection result of the corrugated cardboard thickness.

[0011] Preferably, forming the measurement coordinate system includes: extracting the main direction of the paper texture and coarsely locating the outer contour of the end face from the end face image, and determining the end face tilt based on the angle between the main direction of the paper texture and the direction of the outer contour of the end face;

[0012] The extension direction of the outer contour of the end face is used as the lateral coordinate reference, and the direction that is orthogonal to the lateral coordinate reference and passes through the thickness area of ​​the cardboard is used as the thickness coordinate reference.

[0013] The pixel positions in the end face image are mapped to the coordinate plane defined by the horizontal coordinate reference and the thickness coordinate reference according to the end face tilt amount, and the pixel correspondence before and after mapping is preserved.

[0014] Preferably, generating the outer surface boundary candidate map includes: calculating a gray-level gradient sequence, a texture abrupt change sequence, and an edge response sequence along the thickness direction in the measurement coordinate system, and merging spatially adjacent and oriented response points in the three types of sequences into a cluster of boundary candidate points;

[0015] Sub-pixel position estimation is performed on the candidate boundary point clusters, and a candidate boundary chain is established based on the extension continuity of the candidate boundary point clusters in the lateral coordinate.

[0016] Candidate boundary chains located in areas with dense texture in corrugated core paper, areas with adhesive layer shadows, or areas with rough edges that do not meet the outer surface layer orientation constraints are marked as boundary chains to be verified.

[0017] Preferably, generating the corrugated periodic skeleton diagram includes: extracting the core paper bending ridge line, flute peak candidate point, flute valley candidate point and interlayer transition zone in the measurement coordinate system, and constructing a flute candidate sequence according to the transverse arrangement order;

[0018] The adjacent spacing consistency check and upper and lower layer connection relationship check are performed on the candidate sequence of the corrugation pattern, and isolated candidate objects that do not match the paper layer topology are deleted;

[0019] The center line of the flute unit is determined based on the retained flute peak candidate points, flute valley candidate points, and core paper bending ridge lines, and the phase interval of the flute unit is divided with the center line of the flute unit as a reference.

[0020] Preferably, determining the end face tilt includes: dividing the end face image into multiple horizontally continuous local image bands, and calculating the paper texture direction distribution and outer contour direction distribution within each local image band;

[0021] When the directional difference between adjacent local image bands meets the condition of continuous change, segmented tilt compensation is used instead of a single global tilt compensation.

[0022] When there are rough edges or gaps in a local image band, the missing area is interpolated by using the directional distribution of adjacent local image bands, and the directional interpolation result is incorporated into the pixel correspondence.

[0023] Preferably, the labeling of the boundary chain to be verified includes: extracting the grayscale transition width, texture direction stability, and edge response peak position on both sides of the candidate boundary chain to form a boundary chain attribute sequence;

[0024] The boundary chain attribute sequence is compared with the core paper bending ridge attribute sequence near the same lateral position;

[0025] When the boundary chain attribute sequence is close to the core paper bending ridge in the texture direction and deviates from the outer surface layer thickness region in the normal position, the corresponding candidate boundary chain is set as the internal texture interference chain.

[0026] When the boundary chain attribute sequence satisfies the condition of continuous extension of the outer layer, the corresponding candidate boundary chain is set as a valid candidate boundary chain.

[0027] Preferably, the division of the phase interval of the flute unit includes: establishing a local periodic reference based on the lateral distance sequence between adjacent flute peak candidate points and flute valley candidate points, and excluding candidate point columns that deviate from the local periodic reference and simultaneously lack core paper bending ridge line connections;

[0028] The retained candidate point sequence is phase-labeled according to the peak segment, transition segment, and valley segment;

[0029] By matching the phase markers with the positional relationship of the interlayer transition zone, the phase boundary, center position, and measurable normal range corresponding to each lintel element are obtained.

[0030] Preferably, the process of filtering out abnormal boundary candidate points includes: projecting the effective boundary candidate chain into the corresponding linch cell phase interval, and comparing the changes in the normal position of the boundary candidate points on both sides of the phase boundary;

[0031] When there is a sudden change between a boundary candidate point and an adjacent in-phase interval, and the location of the sudden change coincides with the burr expansion area, the adhesive layer shadow area, or the internal texture interference chain, the boundary candidate point corresponding to the location of the sudden change is deleted.

[0032] When the boundary candidate points maintain the same offset in multiple consecutive phase intervals, the corresponding boundary candidate points are retained and participate in subsequent boundary fitting.

[0033] Preferably, configuring the local thickness normal measurement window includes: determining the window centerline based on the phase boundary, center position, and measurable normal range of each kerf element, and restricting the upper surface candidate region and the lower surface candidate region to the outer surface search zone on both sides of the window centerline;

[0034] When the phase label of an adjacent lintel unit is missing, the missing phase is interpolated based on the local periodic reference of the preceding and following lintel units.

[0035] The interpolated phase boundary and the outer surface search band together define the selection area of ​​the boundary candidate points.

[0036] Preferably, fitting the effective boundary of the upper surface layer and the effective boundary of the lower surface layer includes: performing segmented clustering of the retained boundary candidate points according to the lateral coordinate order, and taking the boundary candidate points that simultaneously satisfy phase consistency and normal position continuity within the same segment as the fitting point set;

[0037] A piecewise curve is established for the fitted point set, and connection constraints are set between adjacent piecewise curves;

[0038] Extract corresponding point pairs of two piecewise curves along the normal direction of the local thickness normal measurement window, and combine the normal distances between the corresponding point pairs into a thickness sequence distributed along the width direction.

[0039] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0040] 1. By adopting a detection method that combines corrugated periodic skeleton and outer layer boundary candidate constraints, thickness measurement no longer relies solely on local grayscale abrupt changes at the end face. The end face image, after orientation field analysis, forms a measurement coordinate system, incorporating the end face extension direction, thickness direction, and cardboard travel direction into the same coordinate reference, thus constraining projection deviations caused by end face tilt. The outer layer boundary candidate map provides possible upper and lower outer boundary positions, while the corrugated periodic skeleton map provides flute peaks, flute valleys, core paper bending ridges, and flute unit phase intervals. The local thickness normal measurement window is limited by the flute unit phase interval; boundary candidate points only participate in fitting if they simultaneously satisfy boundary continuity, phase consistency, and neighborhood thickness variation relationships. Therefore, even if rough edge extension, adhesive layer shadows, and internal core paper textures produce strong grayscale responses, they will be excluded due to inconsistencies with the corrugated phase or outer layer orientation. The effective boundaries of the upper and lower surface layers are piecewise fitted using remaining boundary candidate points, and the distance is calculated along the local normal, ensuring that the thickness result corresponds to the actual interlayer outer boundary position of the cardboard, reducing the risk of internal periodic textures being mistakenly taken as outer boundaries.

[0041] 2. Furthermore, a continuous processing chain is formed between the measurement coordinate system, the candidate map of the outer layer boundary, the corrugated periodic skeleton map, and the calculation of the local normal distance, so that interference from different sources in the end face image is processed in layers. The end face tilt is determined by the main direction of the paper layer texture and the direction of the end face outer contour. The segmented tilt compensation of the local image band can adapt to the slight bending of the end face or the unevenness of the cut edge. The boundary candidate chain is generated at the spatially consistent position of gray-level gradient, texture abrupt change and edge response, and the internal texture interference chain is reduced into the fitting process by comparing the candidate boundary chain attribute sequence with the core paper bending ridge attribute sequence. The phase interval of the flute unit is divided according to the flute peak, flute valley, interlayer transition zone and local periodic reference, so that the measurement window is adjusted with the actual flute position. Candidate points with continuous unidirectional offset are retained, and abrupt change points that coincide with the burr edge expansion area, glue layer shadow area or internal texture interference chain are deleted, so that local anomalies do not directly destroy the fitting of the entire boundary, nor simply erase the real continuous thickness change. The resulting thickness sequence is distributed along the width of the paperboard, which helps to reflect the location of local thickness changes and their continuous range. Attached Figure Description

[0042] Figure 1 This is an overall flowchart of an image recognition-based online thickness detection method for corrugated cardboard.

[0043] Figure 2 A flowchart for establishing the measurement coordinate system and end face tilt compensation is provided.

[0044] Figure 3 A flowchart for jointly generating the candidate map of the outer boundary layer and the corrugated periodic skeleton map;

[0045] Figure 4The flowchart shows the process of local thickness normal measurement window, screening of abnormal candidate points, and generation of thickness sequence. Detailed Implementation

[0046] In one embodiment, refer to Appendix Figure 1 An online thickness detection method for corrugated cardboard based on image recognition includes acquiring an end face image of the corrugated cardboard; performing orientation field analysis on the end face image in the end face extension direction, thickness direction, and cardboard travel direction to form a measurement coordinate system; generating an outer layer boundary candidate map and a corrugated periodic skeleton map in the measurement coordinate system; dividing the corrugated unit phase intervals according to the corrugated periodic skeleton map; configuring a local thickness normal measurement window for the phase intervals; filtering out abnormal boundary candidate points within the local thickness normal measurement window based on boundary continuity, phase consistency, and neighborhood thickness variation relationship; fitting the upper surface effective boundary and lower surface effective boundary to the remaining boundary candidate points respectively; and calculating the relationship between the upper surface effective boundary and the lower surface effective boundary along the local normal. The distance between the effective boundaries of the lower surface layer is used to obtain the online detection result of the corrugated cardboard thickness. In this embodiment, the end face image of the corrugated cardboard is used as the direct processing object. Without changing the cardboard production equipment and cardboard structure, a complete detection chain is formed by image coordinate normalization, boundary candidate extraction, corrugated periodic structure extraction and normal distance conversion. The determination of the outer surface layer boundary is constrained by the corrugated periodic skeleton. The corrugated periodic skeleton is then checked by the continuous direction of the outer surface layer boundary. This avoids the internal core paper texture being mistakenly taken as the thickness boundary due to relying solely on grayscale edges. In this embodiment, the end face direction, layered boundary, flute period and thickness normal relationship are incorporated into the same measurement logic, which can form a verifiable thickness sequence under the image conditions of rough edges, glue layer shadows and interwoven flutes.

[0047] In this embodiment, after acquiring the end face image of the corrugated cardboard, the processing program converts the image into a grayscale matrix indexed by pixel position. Each pixel in the grayscale matrix retains its original position, grayscale value, local gradient, texture direction, and candidate region label. The end face extension direction is determined by the direction of the outer contour major axis and the main direction of the paper layer texture. The thickness direction is determined by the normal of the end face extension direction. The cardboard travel direction is used to exclude the trailing texture formed along the conveying direction. When resolving the direction field, the position with the strongest change in brightness is not directly taken as the boundary. Instead, it is first determined whether the texture direction at that position conforms to the lateral extension characteristics of the outer layer boundary, and then it is determined whether there are periodic core paper bending ridges nearby. In this way, the outer layer boundary response and the internal corrugated texture response in the end face image are separated at the data level. In this embodiment, the boundary determination is changed from single-point grayscale response to region determination under direction constraints, which can reduce the interference of paper fiber burrs or cut shadows on subsequent thickness calculations.

[0048] In this embodiment, the orientation field analysis uses the local gradient covariance method to obtain the principal orientation of the paper texture. Let the grayscale of the end-face image be... The lateral gradient is The longitudinal gradient is The local neighborhood weight is The principal direction angle of the texture is Then we have the following formula:

[0049] ;

[0050] in, This indicates summation within the same local neighborhood. This indicates the amount of grayscale variation along the horizontal direction. This indicates the amount of grayscale variation along the vertical direction. This represents the weight of a pixel's contribution to the local orientation estimation. This represents the orientation angle of a local texture relative to the horizontal coordinate. If within a certain neighborhood... , ,but The angle is approximately 34.7 degrees, indicating that the texture of this neighborhood mainly extends along this angle, and is used to distinguish the direction of the outer layer boundary and the direction of the core paper bending texture.

[0051] In this embodiment, when forming the measurement coordinate system, the pixel band on the end face of the cardboard in the end face image is used as the initial processing area. The processing program first generates a direction histogram based on the set of main texture direction angles, and then compares the coarse positioning result of the end face outer contour with the peak value of the direction histogram. If there is an angle between the direction of the outer contour and the main direction of the paper texture, the angle is recorded as a component of the end face tilt. Then, the extension direction of the end face outer contour is used as the lateral coordinate reference, and the direction orthogonal to the lateral coordinate reference and passing through the thickness area of ​​the cardboard is used as the thickness coordinate reference. The original image pixels are mapped to the plane composed of the lateral coordinate and the thickness coordinate. After mapping, the correspondence between the original pixel position and the new coordinate position is still retained, which is used for subsequent backtracking of the rough edge expansion area, the glue layer shadow area and the internal texture interference chain. In this embodiment, the direction of thickness measurement is not directly affected by the end face tilt and the image framing deflection. The subsequent boundary fitting can be uniformly performed along the local normal.

[0052] In this embodiment, the candidate map of the outer layer boundary is generated by gray-level gradient, texture abrupt change, and edge response. The processing program reads the gray-level sequence at each lateral position along the thickness direction and calculates the gray-level change peak, texture direction abrupt change point, and edge response peak. When the three types of responses are adjacent in spatial position and have the same direction, these responses are merged into a cluster of candidate boundary points. After sub-pixel position estimation, the cluster of candidate boundary points forms a candidate boundary chain. The candidate boundary chain is continuously connected according to the lateral coordinate and records the chain length, normal position, direction change, response peak width, and neighborhood texture attributes. Candidate boundary chains that do not meet the outer layer direction constraint and are located in the dense texture area of ​​the corrugated core paper, the adhesive layer shadow area, or the rough edge expansion area are marked as boundary chains to be verified. In this embodiment, the boundary chains to be verified are not directly deleted, but their attributes are retained for use in corrugated periodic skeleton verification, so as to avoid the real outer boundary being eliminated prematurely when it is affected by local contamination or indentation.

[0053] In this embodiment, the comprehensive response of the candidate points at the outer boundary can be calculated using normalized attribute fusion. Let the candidate pixel be p, and the grayscale gradient normalization value be p. Texture mutation normalization value The edge continuous normalization value is The internal texture interference normalization value is The candidate response at the boundary is The weights of each item are Then we have the following formula:

[0054] ;

[0055] in, This represents the intensity of the change in brightness at point p. This indicates the degree of abrupt change in texture direction at point p. This indicates the degree to which point p forms a continuous boundary with its neighboring candidate points. This indicates the degree of similarity between point p and the internal core paper texture. , , , ,and , , , ,but This value is used to participate in the candidate chain sorting rather than to determine the thickness boundary independently.

[0056] In this embodiment, the corrugated periodic skeleton diagram is generated by the core paper bending ridge line, flute peak candidate points, flute valley candidate points, and interlayer transition zone. The processing program searches for areas with periodic curvature changes in the measurement coordinate system along the lateral direction. Line segments that curve and connect the textures near the upper and lower surfaces are selected as core paper bending ridge line candidates. Points where the texture bending direction changes and are located at the high or low position of the periodic unit are selected as flute peak candidate points and flute valley candidate points. Regions where the grayscale changes slowly but the directional attributes change abruptly between the face paper and the core paper are selected as interlayer transition zone candidates. Subsequently, a flute candidate sequence is constructed according to the lateral arrangement order. The consistency of adjacent spacing and the connection relationship between upper and lower layers are verified. Isolated candidate objects that do not match the paper layer topology are deleted. The retained flute peak candidate points, flute valley candidate points, and core paper bending ridge line are combined into the flute unit center line. In this embodiment, the corrugated structure is no longer a background texture, but a structural reference that constrains the search interval of the outer boundary.

[0057] In this embodiment, when dividing the phase interval of the flute unit according to the corrugated periodic skeleton diagram, the processing program uses the continuously appearing flute peak candidate points and flute valley candidate points as references to divide the transverse segment between adjacent flute peaks and adjacent flute valleys into flute peak segments, transition segments, and flute valley segments. The start and end positions, center positions, measurable normal range, and local periodic references of each segment are written into the phase record. The local periodic references are not given by a fixed template, but are estimated from the transverse distance sequence of continuous reliable flute units in the end face image. When a flute unit is locally compressed, causing the flute peak position to shift, the processing program uses the connection relationship of the core paper bending ridge line and the position of the upper and lower layer transition zone as the basis for phase correction, so that the phase interval is updated with the actual end face texture changes. In this embodiment, the measurement window follows the actual flute structure of the corrugated cardboard, reducing the situation where the fixed flute type window falls into the wrong texture area.

[0058] In this embodiment, a constraint relationship is established between the phase interval and the local thickness normal measurement window. The processing program uses the center line of the corrugated unit as the window center reference, limits the lateral search range with the phase boundary, and limits the thickness direction search range with the measurable normal range. The upper and lower surface candidate regions are constrained within the outer surface search zone. If a region has burrs extending outward, even if the burr pixels are located outside the thickness direction and produce a strong edge response, they will not enter the fitting point set because they are not in the outer surface search zone corresponding to the phase interval. If a region has internal core paper dark lines, the dark line pixels are classified into the internal texture interference chain because they are similar to the properties of the core paper bending ridge. In this embodiment, the boundary candidate points must simultaneously satisfy the constraints of the outer surface position, corrugated phase, and normal window. A single strong grayscale response cannot directly form a thickness boundary.

[0059] In this embodiment, the centerline of the local thickness normal measurement window can be calculated based on the phase boundary and the center position of the lintel element. Let the left phase boundary of the k-th lintel element be... The right boundary of the phase is The x-coordinate of the center of the lumbar element is The normal unit vector is The x-coordinate of any candidate point is The lateral deviation of the candidate point from the center of the window is Then we have the following formula:

[0060] ,

[0061] ;

[0062] in, and This indicates the boundary position of the same angular element in the transverse direction of the measurement coordinate system. This indicates the x-coordinate of the center of the local measurement window. This indicates the thickness normal direction corresponding to the rib element. This indicates the degree to which the candidate point deviates from the center of the window. , ,but ,when hour This deviation, together with the phase to which the candidate point belongs, is used to determine whether it is within the measurable window.

[0063] In this embodiment, when screening out abnormal boundary candidate points within the local thickness normal measurement window, the processing program simultaneously reads the phase of the candidate point, the position of the candidate point in the normal direction, the continuity of the candidate boundary chain to which the candidate point belongs, and the thickness change relationship within adjacent in-phase intervals. If a candidate point has an isolated normal abrupt change compared to candidate points in adjacent in-phase intervals, and the abrupt change location coincides with the burr expansion area, adhesive layer shadow area, or internal texture interference chain, then the candidate point is deleted from the fitting point set. If a candidate point maintains the same direction offset within multiple consecutive phase intervals, then the offset is retained as a candidate state of true boundary change and participates in subsequent fitting. The screening process is determined by phase consistency and neighborhood thickness change relationship, rather than relying on a fixed rejection threshold. In this embodiment, while local noise is eliminated, the continuous true thickness changes are not covered by ordinary smoothing processing.

[0064] In this embodiment, the determination of anomaly rejection can be made by calculating the phase consistency deviation. Let the phase code of candidate point q in the kth lint cell be... The normal position of the candidate point is The in-phase neighborhood reference normal position is The difference in normal values ​​between adjacent lateral positions is The phase consistency deviation is Neighborhood difference weight is Then we have the following formula:

[0065] ;

[0066] in, Indicates whether the candidate point belongs to the peak segment, transition segment, or valley segment. This indicates the position of the candidate point in the thickness direction. Indicates the reference normal position of adjacent in-phase candidate points. This represents the difference in normal vectors between a candidate point and its lateral neighboring candidate point. , , , ,but This value, together with the region attributes of the candidate point, determines whether the candidate point is included in the fitted point set.

[0067] In this embodiment, when fitting the effective boundaries of the upper and lower surfaces to the remaining boundary candidate points, the processing program performs segmented clustering of the retained candidate points according to the horizontal coordinate order. The candidate points within the same segment need to simultaneously satisfy phase consistency and normal position continuity. The upper and lower surface candidate points constitute different fitting point sets, and a segmented curve is established for each fitting point set. Connection constraints are set between the curves to ensure that adjacent segments maintain continuous direction at the horizontal connection. When a local area lacks available candidate points, the segmented curve does not directly cross the missing area to fill it with a straight line. Instead, it uses the boundary trend of the corrugated units with the same phase before and after the missing area and the local normal direction to generate a restricted connection, thus obtaining the effective boundaries of the upper and lower surfaces that can be used for thickness calculation. In this embodiment, the boundary fitting is both constrained by the corrugated periodic structure and retains the continuous geometric features of the outer surface of the cardboard.

[0068] In this embodiment, when calculating the thickness along the local normal, the processing program does not use the pixel difference perpendicular to the image edge as the thickness value. Instead, within the local thickness normal measurement window of each ridge unit, it extracts corresponding point pairs from the upper and lower surface piecewise curves, and sets the corresponding point on the upper surface as... The corresponding point on the lower surface is The local normal unit vector is The thickness at the k-th measurement position is The thickness sequence along the width direction is Then we have the following formula:

[0069] ,

[0070] ;

[0071] in, This represents the displacement vector between corresponding points on the upper and lower surfaces. Indicates the thickness normal direction, and K represents the number of locations involved in the measurement. , ,but This indicates that the pixel distance along the thickness normal at the measurement location is 48.

[0072] In this embodiment, when obtaining the online detection result of corrugated cardboard thickness, the processing program associates and stores the thickness sequence with the phase interval, candidate point confidence status, and anomaly screening record of each measurement position. The online detection result may include the thickness sequence distributed along the width direction of the cardboard, the representative thickness value converted from the thickness sequence, the effective boundary of the upper surface layer participating in the fitting, the effective boundary of the lower surface layer participating in the fitting, and the position of the deleted abnormal candidate point. The representative thickness value is composed of the measured values ​​in the thickness sequence that have passed the phase consistency check and do not belong to the boundary chain to be verified. Local missing regions are displayed with the limited connection results of adjacent phase consistent regions but are not used as the main representative thickness source. In this embodiment, the detection result maintains a correspondence with the specific structural position in the image, which is convenient for verifying whether the thickness anomaly comes from the real boundary change or local image interference.

[0073] In a preferred embodiment, refer to the appendix. Figure 2 The formation of the measurement coordinate system includes extracting the main direction of the paper texture and coarsely locating the outer contour of the end face image. The end face tilt is determined based on the angle between the main direction of the paper texture and the direction of the outer contour of the end face. The extension direction of the outer contour of the end face is used as the lateral coordinate reference, and the direction orthogonal to the lateral coordinate reference and passing through the thickness area of ​​the cardboard is used as the thickness coordinate reference. The pixel positions in the end face image are mapped to the coordinate plane defined by the lateral coordinate reference and the thickness coordinate reference according to the end face tilt, and the pixel correspondence before and after mapping is retained. Specifically, the coarse location of the outer contour of the end face is only used to determine the overall direction and is not directly used for the final thickness boundary. The main direction of the paper texture is used to correct the coordinate offset caused by the inclination of the cut, the slight sway of the cardboard and the change of the image viewing angle. The correspondence before and after mapping is used to write back the subsequently identified rough edges, shadows and internal texture interference to the original image area. In this embodiment, the end face image is transformed into a unified coordinate object suitable for thickness normal measurement, so that the subsequent phase window and boundary fitting have a consistent coordinate basis.

[0074] Furthermore, determining the end face tilt amount includes dividing the end face image into multiple horizontally continuous local image bands, calculating the paper texture direction distribution and outer contour direction distribution in each local image band, and when the direction difference between adjacent local image bands meets the condition of continuous change, a segmented tilt compensation amount is used instead of a single global tilt compensation amount. When there are burrs or gaps in a local image band, the direction distribution of adjacent local image bands is used to perform directional interpolation on the missing area, and the directional interpolation result is incorporated into the pixel correspondence. Specifically, the local image bands are continuously cropped according to the horizontal coordinate, and each local image band independently generates the texture direction distribution and outer contour direction distribution. If the direction difference between consecutive local image bands changes gradually, it indicates that there is a local bend or edge tilt in the end face, which should be mapped by the segmented tilt compensation amount respectively. If a burr in a local image band occludes the outer contour, it is filled by the extension result of the reliable direction in the adjacent local image band. In this embodiment, the global coordinate transformation is avoided to forcibly straighten the locally bent end face, causing boundary offset.

[0075] In a preferred embodiment, refer to the appendix. Figure 3 The generation of the outer layer boundary candidate map includes calculating the gray-level gradient sequence, texture abrupt change sequence, and edge response sequence along the thickness direction in the measurement coordinate system. Response points that are spatially adjacent and have the same direction in the three types of sequences are merged into a boundary candidate point cluster. Sub-pixel position estimation is performed on the boundary candidate point cluster, and candidate boundary chains are established based on the extension continuity of the boundary candidate point cluster in the lateral coordinate. Candidate boundary chains located in the dense texture area of ​​the corrugated core paper, the adhesive layer shadow area, or the rough edge expansion area, and which do not meet the outer layer orientation constraint, are marked as boundary chains to be verified. Specifically, the gray-level gradient sequence reflects the position of light-dark abrupt changes, the texture abrupt change sequence reflects the change in paper layer texture direction, and the edge response sequence reflects the output intensity of the boundary operator. The three are only merged into a candidate point cluster when they are spatially adjacent and the response direction conforms to the extension direction of the outer layer. Boundary chains to be verified are retained instead of being directly deleted, so that structural judgment can be performed later using the corrugated periodic skeleton. In this embodiment, the outer boundary candidate extraction is expanded from a single edge response to a multi-attribute joint judgment, reducing the possibility that the internal texture is directly included in the outer boundary fitting.

[0076] Preferably, the marking of the boundary chain to be verified includes extracting the grayscale transition width, texture direction stability, and edge response peak position on both sides of the candidate boundary chain to form a boundary chain attribute sequence. The boundary chain attribute sequence is compared with the attribute sequence of the core paper bending ridge near the same lateral position. When the boundary chain attribute sequence is close to the core paper bending ridge in the texture direction and deviates from the outer surface layer thickness area in the normal position, the corresponding candidate boundary chain is set as an internal texture interference chain. When the boundary chain attribute sequence meets the condition of continuous extension of the outer surface layer, the corresponding candidate boundary chain is set as a valid boundary candidate chain. Specifically, the grayscale transition width is used to distinguish the abrupt change of the outer contour of the cardboard and the slow transition of the adhesive layer shadow. The texture direction stability is used to distinguish the lateral extension of the outer surface layer and the bending core paper texture. The edge response peak position is used to determine whether the candidate chain is located within the outer surface layer thickness range. In this embodiment, the boundary chain to be verified can be subdivided into internal texture interference chains and valid boundary candidate chains to reduce the situation of accidentally deleting the real outer boundary or accidentally retaining the core paper texture.

[0077] In one embodiment, generating the corrugated periodic skeleton diagram includes extracting the core paper bending ridge line, flute peak candidate points, flute valley candidate points, and interlayer transition zone in the measurement coordinate system; constructing a flute pattern candidate sequence according to the lateral arrangement order; performing adjacent spacing consistency verification and upper and lower layer connection relationship verification on the flute pattern candidate sequence; deleting isolated candidate objects that do not match the paper layer topology; determining the center line of the flute unit based on the retained flute peak candidate points, flute valley candidate points, and core paper bending ridge line; and dividing the phase interval of the flute unit based on the center line of the flute unit. Specifically, the core paper bending ridge line is extracted through texture curvature and grayscale ridge line continuity; the flute peak candidate points and flute valley candidate points are determined through ridge line direction conversion points; and the interlayer transition zone is identified through texture abrupt changes at the junction of the face paper and core paper. Objects in the flute pattern candidate sequence must satisfy the lateral periodic relationship and upper and lower layer connection relationship. Isolated bright lines, glue spot edges, and cut dark lines are deleted even if they have local edge responses because they do not conform to the paper layer topology. In this embodiment, the corrugated periodic skeleton comes from the internal structural relationship of the cardboard rather than a fixed graphic template.

[0078] Furthermore, the division of the phase interval of the flute unit includes establishing a local periodic reference based on the lateral distance sequence between adjacent flute peak candidate points and flute valley candidate points, excluding candidate point sequences that deviate from the local periodic reference and lack core paper bending ridge connections, marking the remaining candidate point sequences according to flute peak segments, transition segments, and flute valley segments, and matching the phase markings with the positional relationship of the interlayer transition zone to obtain the phase boundary, center position, and measurable normal range corresponding to each flute unit. Specifically, the local periodic reference is jointly determined by the median lateral spacing and spacing variation trend of continuous reliable flute units. Point sequences that deviate from the local periodic reference and lack core paper bending ridge connections are considered not to belong to the true flute sequence. The flute peak segments, transition segments, and flute valley segments in the remaining candidate point sequences correspond to different outer surface layer search relationships. The measurable normal range is jointly defined by the interlayer transition zone and the outer surface layer candidate region. In this embodiment, the phase interval can adapt to local flute pitch changes and slight compression deformation.

[0079] In a preferred embodiment, refer to the appendix. Figure 4 The configuration of the local thickness normal measurement window includes determining the window centerline based on the phase boundary, center position, and measurable normal range of each kerf unit, and restricting the upper and lower surface candidate regions to the outer surface search bands on both sides of the window centerline. When the phase annotation of an adjacent kerf unit is missing, the missing phase is interpolated based on the local periodic reference of the preceding and following kerf units. The interpolated phase boundary and the outer surface search band together define the selection area of ​​the boundary candidate points. Specifically, the window centerline is not fixed at the geometric midline of the image, but moves with the center position and phase boundary of the kerf unit. The upper and lower surface candidate regions are located within the outer surface search band. If the phase annotation is missing due to local kerf peaks being occluded or crushed, the missing phase boundary is inferred using the local periodic reference of the preceding and following reliable kerf units. The interpolated phase only limits the selection area of ​​candidate points and does not directly generate the thickness boundary. In this embodiment, the measurement window can maintain continuous constraints when local texture is missing.

[0080] In this embodiment, filtering out abnormal boundary candidate points includes projecting the effective boundary candidate chain onto the corresponding kerf unit phase interval, comparing the normal position changes of the boundary candidate points on both sides of the phase boundary, and deleting the boundary candidate point corresponding to the abrupt change when there is an abrupt change between the boundary candidate point and the boundary candidate point in the adjacent phase interval, and the abrupt change position coincides with the burr expansion area, the adhesive layer shadow area, or the internal texture interference chain. When the boundary candidate point maintains the same direction offset in multiple consecutive phase intervals, the corresponding boundary candidate point is retained and participates in subsequent boundary fitting. Specifically, after the candidate point is projected, it is bound to a specific phase interval. The processing program compares the normal position between in-phase intervals instead of directly comparing the position across the entire image. Isolated abrupt changes are deleted if they coincide with the interference area, and continuous in-phase offsets are retained as actual boundary changes if they cross multiple phase intervals. In this embodiment, local abnormal points and real boundary changes are distinguished by the phase continuity relationship.

[0081] In this embodiment, fitting the effective boundaries of the upper and lower surface layers includes segmenting and clustering the retained boundary candidate points according to the lateral coordinate order, and using the boundary candidate points within the same segment that simultaneously satisfy phase consistency and normal position continuity as the fitting point set. A segmented curve is established for the fitting point set, and connection constraints are set between adjacent segmented curves. Corresponding point pairs of the two segmented curves are extracted along the normal direction of the local thickness normal measurement window, and the normal distances between the corresponding point pairs are combined into a thickness sequence distributed along the width direction. Specifically, the segmented clustering uses the phase interval as the basic boundary to avoid cross-phase mixing of candidate points in different lintel units. The segmented curve connection constraints limit the direction difference and normal position difference of adjacent curves at the connection point. Corresponding point pairs are determined by the same lateral measurement position and the same normal window. In this embodiment, the thickness sequence comes from the normal correspondence between the upper and lower outer surface layers, rather than from the vertical distance in the image coordinates.

[0082] In a preferred embodiment, the representative value of the thickness sequence can be obtained by weighted summation of reliable measurement points, where the thickness at the k-th measurement location is denoted as . The confidence weight of the candidate point at this position is The number of locations participating in the aggregation is K, representing a thickness of Then we have the following formula:

[0083] ;

[0084] in, It is determined by phase consistency, boundary chain continuity, and anomaly screening status. This represents the normal thickness distance at a single location, where K represents the total number of measurement locations involved in the thickness calculation. , , , , , ,but This result, along with the thickness sequence, serves as the online detection output.

[0085] In this embodiment, to clarify the source, constraints, and output relationships of each data object in the processing chain, the processing program manages measurement coordinates, boundary candidates, periodic skeletons, phase windows, and thickness sequences according to the data relationships listed in Table 1 below. Each type of object in Table 1 is generated by the previous processing stage and provides constraints for the next processing stage, avoiding boundary mis-selection caused by the lack of mutual verification between independent algorithm results. The specific relationships are shown in Table 1 below:

[0086] Table 1. Rules for Processing Paperboard Inspection Data

[0087] Measurement coordinate system End face extension direction, thickness direction, and paperboard travel direction Map the original pixels to the horizontal and thickness coordinate planes. Provides boundary extraction and normal measurement benchmarks Outer layer boundary candidate map Gray-scale gradient, texture abrupt change, edge response Merge spatially adjacent and oriented response points to form candidate boundary chains Provide candidate points for the upper and lower surfaces. Boundary Chain to be Verified Candidate chain position, direction, and texture attributes Compare with the properties of the core paper bending ridge line and distinguish between interfering chains and candidate chains. Participating in anomaly screening and fitting point selection Corrugated periodic skeleton diagram Corrugated peaks, corrugated valleys, core paper folding ridges, interlayer transition zones Verify horizontal cycle and upper / lower layer connection relationships. Dividing the phase interval of the lumbar unit Local thickness normal measurement window Phase boundary, center position, measurable normal range Limit the selection area of ​​candidate points and the direction of corresponding points Generate normal thickness distance Thickness sequence Point pairs corresponding to the upper and lower effective boundaries Calculate and associate the confidence weights based on the local normal. Output online detection results

[0088] In this embodiment, the data relationship shown in Table 1 embodies a process where sequential processing and reverse verification coexist. The measurement coordinate system provides a unified position reference for all pixels. The outer layer boundary candidate map provides a chain of points that can participate in thickness boundary fitting. The boundary chain to be verified retains intermediate states of suspected outer boundaries and internal interference. The corrugated periodic skeleton map provides the cardboard layered periodic constraint. The local thickness normal measurement window selects boundary candidate points within a specific phase interval. The thickness sequence binds the normal distance of the upper and lower effective boundaries with the confidence weight. If the thickness sequence shows an isolated offset in a certain area, the processing program can trace back the phase window, boundary chain attributes, and interference chain markers of that area to determine whether to re-screen candidate points. In this embodiment, each processing step retains traceable data, which facilitates locating the source of error in continuous detection.

[0089] In one embodiment, subpixel position estimation can be performed by curve fitting of the local response peaks of the boundary candidate point cluster. The processing program selects the response peak and its adjacent pixel response values ​​in the thickness direction of the candidate point cluster, and uses a parabola to approximate the position of the response peak to obtain the boundary candidate coordinates located between pixels. Subpixel position estimation is only used to improve the position expression accuracy of the candidate points in the measurement coordinate system, without changing the judgment logic that the boundary candidate points must satisfy gray-level gradient, texture abrupt change, edge response, phase consistency and normal window constraints. When there are multiple similar peaks near the local response peak, the processing program prioritizes retaining the peak coordinates that are continuous with the outer surface layer and far away from the internal texture interference chain. In this embodiment, while refining the boundary position, the refining process does not deviate from the corrugated periodic structure constraint.

[0090] In a preferred embodiment, when there is an adhesive layer shadow in the end face image, the processing program identifies the shadow area as a region with a large gray-level transition width, a slow change in texture direction, and an unstable edge response peak. The candidate boundary chains in the adhesive layer shadow area are included in the boundary chains to be verified and are compared with the outer surface candidate chains in the adjacent in-phase interval. If the candidate chain in the adhesive layer shadow area cannot be connected with the continuous candidate chain in the upper or lower surface outer surface search band, it will not be included in the fitting point set. If the shadow area covers a small section of the true outer surface boundary, the in-phase candidate chains before and after the area are used to generate a restricted connection by connecting the piecewise curve. In this embodiment, the adhesive layer shadow will not directly form a thickness boundary due to strong gray-level changes, nor will it cause the entire boundary to be interrupted due to local occlusion.

[0091] In a preferred embodiment, when there is burr expansion in the end face image, the processing program identifies the burr area as a region with messy grayscale texture on the outer side of the outer contour, low directional stability, and lack of continuity with the main direction of the paper layer texture. Even if the edge response in the burr expansion area is on the outer side of the cardboard, it still needs to be verified by the phase window and the outer surface layer search band. If the burr candidate point deviates from the outer surface layer thickness area in the normal direction and there is an isolated abrupt change with the candidate point in the adjacent phase interval, the corresponding candidate point is deleted. If the burr area overlaps with the true outer surface layer boundary, the processing program retains the inner boundary point in the continuous candidate chain of the true outer surface layer and excludes the outward burr edge. In this embodiment, the outer edge of the burr is not included in the thickness distance, and the true outer surface layer boundary of the cardboard can still be extracted through continuity and phase relationship.

[0092] In a preferred embodiment, when there is a region with strong internal core paper texture in the end face image, the processing program compares the core paper bending ridge attribute sequence and the candidate boundary chain attribute sequence. If a candidate chain bends along the texture direction with the flute peaks and valleys, and its normal position is located between the interlayer transition zones, then the candidate chain is set as an internal texture interference chain. The internal texture interference chain does not participate in the fitting of the upper or lower effective boundary, but it can participate in the corrugated period skeleton verification to correct the phase interval of the flute unit. If the internal texture interference chain is close to the outer layer candidate chain in a certain lateral position, the processing program retains the outer layer candidate chain according to the outer layer search zone and the boundary chain extension direction. In this embodiment, the internal core paper texture is transformed from an error source into a skeleton construction reference.

[0093] In a preferred embodiment, when a local corrugated unit is compressed, the processing program does not use a fixed corrugation pitch or fixed corrugation height template for forced matching. Instead, it adjusts the phase interval based on the local periodic reference of the adjacent reliable corrugated units before and after the local corrugated unit, the connection of the core paper bending ridge line, and the position of the interlayer transition zone. If the compression deformation causes the corrugation peak segment to shorten but the core paper bending ridge line can still connect the upper and lower layers, the corresponding corrugated unit is still included in the corrugated periodic skeleton diagram. If the candidate point of the outer layer boundary in the compressed area is continuously offset and spans multiple phase intervals, the offset is retained and participates in the piecewise fitting. If the offset only appears in a single phase interval and coincides with the burr or shadow, the corresponding candidate point is deleted. In this embodiment, the real thickness change under local compression is separated from image interference.

[0094] In a preferred embodiment, the processing program can use the verified local periodic reference from the previous image as the initial reference for the same batch of cardboard end faces in continuously detected images. However, in the current image, it is still necessary to re-verify through the candidate points of flute peaks, candidate points of flute valleys, the bending ridge line of the core paper, and the interlayer transition zone. Only the verified local periodic reference in the current image participates in the phase interval division. If the periodic reference in the previous image is inconsistent with the topological connection relationship in the current image, the phase interval is updated based on the topological connection relationship in the current image. The historical reference is only used as an interpolation constraint when the phase is missing, and does not directly replace the recognition result of the current image. In this embodiment, the phase recognition can be kept stable in continuous online detection, while avoiding historical data from covering the current real changes.

[0095] In this embodiment, the processing program retains a calculation source marker for each position in the thickness sequence. The source markers include direct measurement, measurement under phase interpolation constraints, measurement under piecewise curve connection, and measurement to be verified. Direct measurement indicates that the corresponding points of the upper and lower surface layers both come from a high-confidence candidate chain. Measurement under phase interpolation constraints indicates that at least one phase boundary is interpolated by the preceding and following local periodic references. Measurement under piecewise curve connection indicates that the boundary corresponding point comes from the restricted connection of adjacent curves. Measurement to be verified indicates that the candidate point has been retained but has a continuous offset from the adjacent in-phase interval. Different source markers correspond to different confidence weights when calculating the thickness representative value. All source markers are still retained when the thickness sequence is output. In this embodiment, the thickness result is associated with the measurement process, and local abnormal areas will not be confused with ordinary measurement areas.

[0096] In a preferred embodiment, after generating the effective boundaries of the upper and lower surfaces, the processing program verifies the lateral phase correspondence of the two boundary curves. During the verification, the phase interval of each corresponding point pair is matched with the phase label in the corrugated periodic skeleton diagram. If the corresponding point of the upper surface is located in the peak segment of a certain corrugated unit while the corresponding point of the lower surface is assigned to the valley segment of an adjacent corrugated unit, it indicates that the corresponding point pair has a phase error. The processing program then searches for candidate points of the lower surface along the local normal at that position or sets the corresponding point pair as to be verified for measurement. If both the upper and lower corresponding points fall within the measurable normal range of the same corrugated unit, the corresponding point pair is written into the thickness sequence. In this embodiment, the two ends of the thickness distance come from the same local structural unit to avoid mismatch causing thickness anomalies.

[0097] In one embodiment, the processing program can divide the end face image into a normal measurement area, a suspected interference area, and a missing verification area. In the normal measurement area, the boundary candidate chain and the corrugated periodic skeleton are continuous. In the suspected interference area, there are burr expansion areas, adhesive layer shadow areas, or internal texture interference chains. In the missing verification area, the phase annotation or the outer surface candidate chain is locally missing. The normal measurement area directly generates a set of fitting points. In the suspected interference area, abrupt candidate points are first deleted by anomaly screening rules before fitting. In the missing verification area, a restricted boundary is generated by front and rear phase interpolation and piecewise curve connection constraints. The three types of regions record different confidence states in the thickness sequence. In this embodiment, different processing rules are used for different image quality regions, and the same end face image does not lose its overall measurement capability due to local defects.

[0098] In this embodiment, the data flow in the implementation process begins with the end face image entering the measurement coordinate system. It is then generated in parallel through the outer layer boundary candidate image and the corrugated periodic skeleton image. The phase interval of the corrugated unit binds the two types of image results to the same local structural unit. The local thickness normal measurement window limits the candidate point selection area according to the phase boundary and the measurable normal range. Anomaly screening is completed based on the regional attributes of the candidate point, phase consistency, and neighborhood thickness change relationship. Segmented fitting generates upper and lower effective boundaries based on the retained candidate points. The normal distance calculation converts the corresponding point pairs into a thickness sequence. In the entire process, any local boundary result can be traced back to the original pixel, phase interval, and candidate chain attributes. In this embodiment, a closed verification chain is formed between the outer boundary recognition and the internal periodic structure recognition, which can solve the technical problem that the outer boundary and the internal corrugation are difficult to stably separate in the end face image.

Claims

1. A method for online detection of the thickness of corrugated board based on image recognition, characterized in that, include: Acquire an image of the corrugated cardboard end face, and perform orientation field analysis on the end face image in the end face extension direction, thickness direction, and cardboard travel direction to form a measurement coordinate system; Generate an outer surface boundary candidate map and a corrugated period skeleton map in the measurement coordinate system; Based on the corrugated periodic skeleton diagram, the phase intervals of the corrugated units are divided, and local thickness normal measurement windows are configured for the phase intervals; Within the local thickness normal measurement window, abnormal boundary candidate points are screened out based on boundary continuity, phase consistency, and neighborhood thickness variation relationship; The remaining boundary candidate points are fitted with the effective boundaries of the upper and lower surfaces respectively, and the distance between the effective boundaries of the upper and lower surfaces is calculated along the local normal to obtain the online detection result of the corrugated cardboard thickness. Generating the outer surface layer boundary candidate map includes: calculating grayscale gradient sequences, texture abrupt change sequences, and edge response sequences along the thickness direction in the measurement coordinate system; merging spatially adjacent and oriented response points in the three sequences into a boundary candidate point cluster; estimating the sub-pixel position of the boundary candidate point cluster; and establishing candidate boundary chains based on the extension continuity of the boundary candidate point cluster in the lateral coordinate; marking candidate boundary chains located in the corrugated core paper texture dense area, adhesive layer shadow area, or rough edge expansion area that do not meet the outer surface layer orientation constraint as boundary chains to be verified. Generating the corrugated periodic skeleton diagram includes: extracting the core paper bending ridge line, candidate points of flute peaks, candidate points of flute valleys, and interlayer transition zones in the measurement coordinate system, and constructing a candidate flute pattern sequence according to the transverse arrangement order; The adjacent spacing consistency check and upper and lower layer connection relationship check are performed on the candidate flute pattern sequence, and isolated candidate objects that do not match the paper layer topology are deleted; the center line of the flute unit is determined based on the retained flute peak candidate points, flute valley candidate points and core paper bending ridge line, and the phase interval of the flute unit is divided with the center line of the flute unit as the reference. The labeling of the boundary chain to be verified includes: extracting the grayscale transition width, texture direction stability, and edge response peak position on both sides of the candidate boundary chain to form a boundary chain attribute sequence; comparing the boundary chain attribute sequence with the core paper bending ridge attribute sequence near the same lateral position; when the boundary chain attribute sequence is close to the core paper bending ridge in the texture direction and deviates from the outer surface layer thickness region in the normal position, the corresponding candidate boundary chain is set as an internal texture interference chain; when the boundary chain attribute sequence satisfies the outer surface layer continuous extension condition, the corresponding candidate boundary chain is set as a valid boundary candidate chain.

2. The online thickness detection method for corrugated cardboard based on image recognition according to claim 1, characterized in that, Forming the measurement coordinate system includes: extracting the main direction of the paper texture and coarsely locating the outer contour of the end face from the end face image, and determining the inclination of the end face based on the angle between the main direction of the paper texture and the direction of the outer contour of the end face; The extension direction of the outer contour of the end face is used as the lateral coordinate reference, and the direction that is orthogonal to the lateral coordinate reference and passes through the thickness area of ​​the cardboard is used as the thickness coordinate reference. The pixel positions in the end face image are mapped to the coordinate plane defined by the horizontal coordinate reference and the thickness coordinate reference according to the end face tilt amount, and the pixel correspondence before and after mapping is preserved.

3. The online thickness detection method for corrugated cardboard based on image recognition according to claim 2, characterized in that, The determination of the end face tilt includes: dividing the end face image into multiple horizontally continuous local image bands, and calculating the paper texture direction distribution and outer contour direction distribution in each local image band; When the directional difference between adjacent local image bands meets the condition of continuous change, segmented tilt compensation is used instead of a single global tilt compensation. When there are rough edges or gaps in a local image band, the missing area is interpolated by using the directional distribution of adjacent local image bands, and the directional interpolation result is incorporated into the pixel correspondence.

4. The online thickness detection method for corrugated cardboard based on image recognition according to claim 3, characterized in that, The division of the phase interval of the flute unit includes: establishing a local periodic reference based on the lateral distance sequence between adjacent flute peak candidate points and flute valley candidate points, and excluding candidate point columns that deviate from the local periodic reference and simultaneously lack core paper bending ridge line connections; The retained candidate point sequence is phase-labeled according to the peak segment, transition segment, and valley segment; By matching the phase markers with the positional relationship of the interlayer transition zone, the phase boundary, center position, and measurable normal range corresponding to each lintel element are obtained.

5. The online thickness detection method for corrugated cardboard based on image recognition according to claim 4, characterized in that, The process of filtering out abnormal boundary candidate points includes: projecting the effective boundary candidate chain into the corresponding linch cell phase interval, and comparing the changes in the normal position of the boundary candidate points on both sides of the phase boundary; When there is a sudden change between a boundary candidate point and an adjacent in-phase interval, and the location of the sudden change coincides with the burr expansion area, the adhesive layer shadow area, or the internal texture interference chain, the boundary candidate point corresponding to the location of the sudden change is deleted. When the boundary candidate points maintain the same offset in multiple consecutive phase intervals, the corresponding boundary candidate points are retained and participate in subsequent boundary fitting.

6. The online thickness detection method for corrugated cardboard based on image recognition according to claim 5, characterized in that, Configuring the local thickness normal measurement window includes: determining the window centerline based on the phase boundary, center position, and measurable normal range of each ridge unit, and restricting the upper surface candidate region and the lower surface candidate region to the outer surface search zone on both sides of the window centerline; When the phase label of an adjacent lintel unit is missing, the missing phase is interpolated based on the local periodic reference of the preceding and following lintel units. The interpolated phase boundary and the outer surface search band together define the selection area of ​​the boundary candidate points.

7. The online thickness detection method for corrugated cardboard based on image recognition according to claim 6, characterized in that, Fitting the effective boundary of the upper surface layer and the effective boundary of the lower surface layer includes: performing segmented clustering of the retained boundary candidate points according to the order of the lateral coordinates, and taking the boundary candidate points that simultaneously satisfy phase consistency and normal position continuity within the same segment as the fitting point set; A piecewise curve is established for the fitted point set, and connection constraints are set between adjacent piecewise curves; Extract corresponding point pairs of two piecewise curves along the normal direction of the local thickness normal measurement window, and combine the normal distances between the corresponding point pairs into a thickness sequence distributed along the width direction.

Images