Printing paper defect detection system and method based on machine vision

Abstract

The invention particularly relates to a printing paper defect detection system and method based on machine vision, and relates to the technical field of machine vision detection. A multi-source data sensing module; and a defect candidate extraction and fusion module. According to the invention, through deep fusion of 2D texture and 3D morphology data, the problems of missed judgment and misjudgment of a traditional single detection mode are effectively solved; color difference and texture defects can be accurately captured through the 2D image, concave-convex defects can be accurately recognized through the 3D height map, and plane and three-dimensional all-dimensional defect feature description is formed after space-time alignment and pixel-level coordinate association; precise distinguishing of various defects such as stains, scratches, pits and keloids is achieved, and ineffective noise interference can be eliminated through quantitative features.

Description

Technical Field

[0001] This invention relates to the field of machine vision inspection technology, and in particular to a machine vision-based system and method for detecting defects in printed paper. Background Technology

[0002] As a basic material for industries such as packaging and printing, the surface defects of printing paper (such as stains, scratches, pits, and pulp lumps) directly affect the quality of downstream products.

[0003] Existing defect detection technologies are mainly divided into two categories: 2D detection and 3D detection.

[0004] 2D inspection identifies defects by acquiring grayscale and color information from images, but it cannot distinguish between uneven defects without texture differences (such as transparent glue dots), resulting in a high false negative rate.

[0005] 3D inspection identifies uneven defects by measuring surface morphology, but it is difficult to detect texture defects without morphological changes (such as light-colored stains), and it is easily affected by the inherent undulations of the paper surface, leading to misjudgments.

[0006] Meanwhile, printing paper production lines are characterized by high speed (up to 600m / min), wide width (usually 1.8m), and diverse defect types. A single detection method is difficult to balance detection accuracy, real-time performance, and industrial adaptability.

[0007] For example, traditional 2D inspection is prone to motion blur under high-speed motion, and 3D inspection has a large data processing volume, resulting in insufficient real-time performance. Both also have the problem of misjudgment, which cannot meet the needs of industrial production for high-quality inspection.

[0008] Therefore, a machine vision-based defect detection system and method for printed paper is proposed to address the aforementioned problems. Summary of the Invention

[0009] The purpose of this invention is to provide a machine vision-based defect detection system and method for printing paper in order to solve the above-mentioned problems.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] A machine vision-based printing paper defect detection system includes:

[0012] The system hardware architecture module is configured to build a hardware platform adapted to the needs of the printing paper production line, providing hardware support for high-quality acquisition of 2D texture and 3D morphology data;

[0013] The multi-source data perception module is configured to perform spatiotemporal aligned acquisition of 2D color images and 3D point cloud data to ensure that the 2D texture information and 3D shape information of the same paper area are correlated.

[0014] The defect candidate extraction and fusion module is configured to locate suspected defect areas from the preprocessed 2D image and 3D height map; after forming a comprehensive feature vector by associating 2D and 3D features, it performs defect classification and eliminates false positives.

[0015] Preferably, the system hardware architecture module specifically includes:

[0016] The resolution is calculated by working backward from the detection accuracy to determine the camera required for 2D high-definition color imaging; the light source is selected according to the type of defect.

[0017] A rotary encoder with 1024 pulses / revolution is selected and installed on a 100mm diameter conveyor roller, ensuring fixed-distance data acquisition;

[0018] A suitable line laser profilometer was selected for 3D topographic contour measurement.

[0019] The distance between the 2D camera and the production line is ≤50mm, and the alignment deviation of the center axis is ≤0.1mm; the same encoder trigger signal is shared.

[0020] Preferably, the multi-source data sensing module specifically includes:

[0021] After acquiring the corresponding 2D image data and 3D point cloud data respectively, we will ensure that the 2D and 3D acquisition delays are less than the preset standard in order to achieve time synchronization of data acquisition.

[0022] Using a checkerboard-patterned, raised-dot 3D calibration plate, four sets of 2D / 3D data from different angles were collected to calculate the 4×4 rigid body transformation matrix. This transforms 3D point clouds from the sensor coordinate system to the 2D pixel coordinate system to achieve spatial synchronization of data acquisition.

[0023] Preferably, the method further includes:

[0024] The data preprocessing and coordinate unification module is configured to eliminate noise and interference information in 2D images and 3D point clouds, and convert irregular 3D point clouds into height maps that are pixel-aligned with 2D images, thus clearing obstacles for subsequent defect candidate region extraction.

[0025] Preferably, the defect candidate extraction and fusion module specifically includes:

[0026] 2D Defect Screening: Constructing a Background Model: Collect 500 images of good products, and analyze each pixel. Calculate the mean of gray values with standard deviation The mean background image and standard deviation background image are obtained, and the model is updated every hour with 100 new good product images;

[0027] Defect detection: For each pixel of the current frame image Defective pixels are identified.

[0028] Preferably, the method further includes 3D morphological anomaly screening:

[0029] For each pixel of the heightmap Take a 3×3 neighborhood window and calculate the variance of the Z value. If satisfied > If so, it is marked as an abnormal pixel;

[0030] Using a Gaussian high-pass filter to preserve regions of dramatic height changes, if the absolute value of the Z-value after filtering is... If so, it is marked as an abnormal pixel;

[0031] Calculate the height and volume features for each suspected abnormal region.

[0032] Preferably, the method further includes defect determination:

[0033] Based on pixel-aligned height maps, the pixel coordinates of suspected 2D defect areas are... Directly map to a 3D height map and extract the 3D features of the corresponding region;

[0034] Conversely, 3D suspected regions can also be associated with 2D features to form a combined 2D and 3D feature vector for each candidate region: ;

[0035] in, The area of ​​the 2D defect region; For roundness; This represents the grayscale difference. Average height; For height variance; For the defect volume; The mean value of channel a in the Lab space; This represents the mean value of channel b in the Lab space.

[0036] The defect determination result is as follows:

[0037] Surface stains: ;

[0038] Scratches / Indentations: ;

[0039] pit: ;

[0040] Lumps of pulp: ;

[0041] Elimination of false positives: or It was determined to be an invalid defect.

[0042] Preferably, the method further includes:

[0043] The test result output and production line linkage control module is configured to convert the test results into traceable reports and actionable actions on the production line, realize the recording and analysis of defect data and intelligent linkage of the production line, and support fault alarms and manual intervention.

[0044] A machine vision-based method for detecting defects in printed paper includes the following steps:

[0045] Build a hardware platform adapted to the needs of the printing paper production line, and configure 2D high-definition color imaging, 3D shape contour measurement and industrial control integration equipment;

[0046] Spatiotemporal alignment acquisition of 2D color images and 3D point cloud data is performed, and accurate association of 2D / 3D information in the same area of ​​paper is achieved through time compensation and spatial calibration.

[0047] The system performs flat field correction, color conversion, and filtering noise reduction on 2D images, removes invalid points and filters 3D point clouds, and generates a 3D height map aligned with the 2D image by unifying coordinates.

[0048] Suspected defect areas are extracted from 2D images and 3D height maps respectively, and their features are calculated. These features are then fused to form a comprehensive feature vector, and the defect type is determined according to rules to eliminate false positives.

[0049] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0050] 1. This invention effectively solves the problems of missed detection and false detection in traditional single detection methods by deeply integrating 2D texture and 3D morphology data; 2D images can accurately capture color difference and texture defects, while 3D height maps can accurately identify concave and convex defects. After spatiotemporal alignment and pixel-level coordinate association, a comprehensive description of defect features in both planar and three-dimensional forms is formed; it achieves accurate differentiation of various types of defects such as stains, scratches, pits, and lumps, and can also eliminate invalid noise interference through quantitative features.

[0051] 2. This invention achieves real-time online detection on high-speed production lines through hardware selection optimization and algorithm process simplification. The hardware uses high frame rate cameras, high scanning frequency laser profilometers, and high-performance industrial control computers to ensure that data acquisition and transmission delays are controlled at the millisecond level. It has hierarchical linkage handling, fault alarm and data traceability functions, and can automatically trigger scrap removal, inkjet coding or manual confirmation according to the severity of defects. Combined with equipment fault alarm and defect rate anomaly analysis, it not only reduces the generation of invalid waste, but also provides data support for production line maintenance. Attached Figure Description

[0052] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which:

[0053] Figure 1 This is a system structure diagram of the present invention;

[0054] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0055] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0056] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0057] Example 1

[0058] Its specific implementation method is combined with the appendix Figure 1 and attached Figure 2 Please provide a detailed explanation.

[0059] Appendix Figure 1 The diagram shows the structural block diagram of a machine vision-based printing paper defect detection system provided in this embodiment of the invention. It illustrates the connection relationship between the system hardware architecture modules and the defect candidate extraction and fusion module, and marks the main functional interaction flow of each module.

[0060] Appendix Figure 2 The flowchart of the machine vision-based printing paper defect detection method provided in the embodiments of the present invention illustrates the complete steps from building a hardware platform adapted to the needs of the printing paper production line to outputting traceable detection reports and real-time visualized data.

[0061] In this embodiment, it includes:

[0062] The system hardware architecture module is configured to build a hardware platform adapted to the needs of the printing paper production line, providing hardware support for high-quality acquisition of 2D texture and 3D morphology data;

[0063] Specifically, it includes:

[0064] Based on the detection accuracy, the resolution is deduced to determine the camera required for 2D high-definition color imaging; the light source is selected according to the defect type: diffuse surface light source (uniformity ≥90%) is suitable for uneven fiber texture and slight color difference; coaxial light is suitable for transparent glue spots and light-colored stains, which can improve contrast by more than 30%; the color is selected as white (general), red (suitable for blue / black defects) or blue (suitable for red / orange defects); equipped with a constant current controller (current fluctuation ≤1%), the light intensity is automatically calibrated every hour to avoid attenuation causing the detection threshold to fail;

[0065] A rotary encoder with 1024 pulses / revolution is selected and installed on a 100mm diameter conveyor roller. The paper movement distance corresponding to each pulse is calculated using the formula: The trigger interval is set based on a single frame covering 0.1m (i.e., 100mm) of paper. The required number of pulses is: And ensure fixed-distance data acquisition;

[0066] Select a suitable line laser profilometer for 3D topographic contour measurement, for example: Z-axis resolution ≤ 2μm (suitable for 5μm scratch detection), scanning frequency ≥ 120kHz (ensuring adjacent profile intervals ≤ 0.1mm at 10m / s speed, with no data breaks), and the scanning frequency matching formula is: The reference distance is 50mm and the depth of field is ≥5mm (adapting to paper with ±2mm jump); each scan outputs 1024 three-dimensional points (X,Y,Z), which are transmitted via EtherCAT protocol (delay ≤1ms).

[0067] The distance between the 2D camera and the production line is ≤50mm, and the alignment deviation of the center axis is ≤0.1mm; the same encoder trigger signal is shared (delay ≤10μs) to reduce data misalignment caused by time and space differences.

[0068] The multi-source data perception module is configured to collect 2D color images and 3D point cloud data in a spatiotemporal alignment, ensuring that the 2D texture information and 3D shape information of the same paper area can be accurately correlated.

[0069] Specifically, it includes:

[0070] After the encoder generates a trigger pulse, the motion control card of the industrial computer receives the signal and synchronously triggers the 2D camera to expose, so that the camera transmits the image data (such as a 25-megapixel RGB image) to the industrial computer buffer through the GigEVision protocol to obtain 2D image data.

[0071] Sharing the same trigger pulse with the 2D camera, the laser profilometer emits a laser line to the paper surface, and the camera receives the reflected light and calculates the 3D point cloud (1024 points / scan). The point cloud data is transmitted to the industrial control computer via the EtherCAT protocol to obtain 3D point cloud data.

[0072] After acquiring the corresponding 2D image data and 3D point cloud data respectively, we will ensure that the 2D and 3D acquisition delays are less than the preset standard in order to achieve time synchronization of data acquisition.

[0073] Trigger signals are uniformly distributed through a motion control card to ensure that the 2D and 3D acquisition delays are ≤10μs; if there is a slight delay (such as... , This represents the number of 2D trigger pulses. (where 3D trigger pulse count is used), the Y-coordinate (paper movement direction) of the 3D data is compensated using the following formula to ensure that the same area is acquired at the same time: ;

[0074] System calibration was performed: A checkerboard-patterned, raised-dot 3D calibration plate was used to collect 2D / 3D data at four different angles (0°, 30°, 60°, and 90°) to calculate the 4×4 rigid body transformation matrix. The 3D point cloud is transformed from the sensor coordinate system to the 2D pixel coordinate system to achieve spatial synchronization of data acquisition. The transformation formula is as follows: Conversion error ≤ 0.5 pixels, achieving 2D pixel ( , ) and 3D points ( , , Spatial correspondence.

[0075] The data preprocessing and coordinate unification module is configured to eliminate noise and interference in 2D images and 3D point clouds, converting irregular 3D point clouds into height maps pixel-aligned with the 2D images, thus clearing obstacles for subsequent defect candidate region extraction. Specifically, this includes:

[0076] 2D image preprocessing includes the following:

[0077] Field correction: Eliminates lens vignetting and uneven lighting using the following formula, where... For pixel coordinates:

[0078] ;

[0079] Dark field images are completely black images obscured by the lens cap (average of 10 frames), and bright field images are standard white images (average of 10 frames).

[0080] Color space conversion: When detecting color difference defects, the RGB image is converted to Lab color space (using the a / b channel to highlight color differences); when detecting grayscale defects, it is converted to a grayscale image, and the grayscale value calculation formula is:

[0081] ;

[0082] After conversion, histogram equalization is performed to improve the contrast between defects and the background.

[0083] Filtering and noise reduction: Non-local mean denoising is used to smooth noise; for paper with complex textures, additional guided filtering is applied to preserve edge details;

[0084] 3D point cloud preprocessing includes the following:

[0085] Invalid point removal: Traverse the point cloud and remove laser-obstructed points with Z values ​​exceeding the range (Z<-10 or Z>+10mm) and edge points with X / Y values ​​exceeding the paper width; trigger a sensor calibration reminder when the proportion of invalid points exceeds 5%;

[0086] Point cloud filtering: First, statistical outlier removal is performed to eliminate discrete noise; then, radius filtering is performed to eliminate isolated points; the point cloud density fluctuation after filtering is ≤3%.

[0087] Coordinate unification and height map generation: applying rigid body transformation matrices The 3D point cloud is transformed to a 2D pixel coordinate system; Kriging interpolation is used to convert the irregular point cloud into a height map with the same size as the 2D image, with each pixel... Corresponding height value This achieves sequential pixel alignment;

[0088] After 2D preprocessing, ensure that the gray standard deviation of the defect-free area is ≤5, and the absolute value of the gray difference between the defect area and the background is ≥3 times the gray standard deviation of the background.

[0089] After 3D preprocessing, ensure that the height variance of the flat area is ≤ The absolute value of the height change of the raised / depressed defect is ≥ The elevation map shows no obvious faults or noise points.

[0090] The defect candidate extraction and fusion module is configured to quickly locate suspected defect areas from the preprocessed 2D image and 3D height map; after forming a comprehensive feature vector by associating 2D and 3D features, it performs defect classification and eliminates false positives.

[0091] Specifically, it includes:

[0092] The defect candidate region extraction process includes:

[0093] 2D Defect Screening: Constructing a Background Model: Collect 500 images of good products, and analyze each pixel. Calculate the mean of gray values with standard deviation The mean background image and standard deviation background image are obtained, and the model is updated every hour with 100 new good product images;

[0094] Defect detection: For each pixel of the current frame image Defective pixels are determined using the following formula: ;

[0095] in For the current frame image in The grayscale value at the location; (k is the threshold, usually taken as 3 to 5, adjusted according to the defect contrast);

[0096] The process is divided into 32×32 pixel blocks. Only blocks whose mean value differs from the mean value of the background block by more than a threshold are detected pixel by pixel to improve computational efficiency.

[0097] For complex defects, the following approach is adopted:

[0098] The model is a lightweight U-Net (with MobileNetV2 as the encoder and 2 million parameters), with a 512×512 pixel image as input (the original image is downsampled to reduce computation) and a binary defect map as output (white indicates suspected defects).

[0099] 10,000 labeled defect images (divided into training and validation sets in an 8:2 ratio) were used for training, and the Dice loss function was employed for training.

[0100] The model was quantized to FP16 precision using TensorRT.

[0101] For each suspected defect region (connected region), calculate its geometric features, center coordinates, grayscale features, and color features; then link these features to the 3D morphological anomaly region screening process.

[0102] 6. The machine vision-based printing paper defect detection system according to claim 5, characterized in that it further includes 3D morphological anomaly screening:

[0103] For each pixel of the heightmap Take a 3×3 neighborhood window and calculate the variance of the Z value. (The Z-value is the height coordinate perpendicular to the paper surface in three-dimensional space. It is used to quantify the degree of concavity or convexity of a point on the paper surface relative to a reference plane and is a core parameter characterizing the 3D morphology of the paper.)

[0104] If satisfied > ( If the variance is determined by 1.5 times the maximum variance of 500 defect-free height maps, then it is marked as an abnormal pixel;

[0105] Using a Gaussian high-pass filter to preserve regions of dramatic height changes, if the absolute value of the Z-value after filtering is... If so, it is marked as an abnormal pixel; The determination was made by combining the inherent undulation characteristics of normal paper surface, the minimum detectable size requirement of defects, and experimental verification. The specific process will not be elaborated here.

[0106] Calculate the height and volume features for each suspected anomaly region;

[0107] Height characteristics include: average height Maximum height Height variance ;

[0108] Volumetric features include: region volume ;

[0109] in This is the normal area height. The physical area of ​​a single pixel is represented by the volume of a pit, which is negative, while the volume of a bump is positive.

[0110] It also includes defect determination:

[0111] Based on pixel-aligned height maps, the pixel coordinates of suspected 2D defect areas are... Directly map to a 3D height map and extract the 3D features of the corresponding region;

[0112] Conversely, 3D suspected regions can also be associated with 2D features to form a combined 2D and 3D feature vector for each candidate region: ;

[0113] in,

[0114] The area of ​​the 2D defect region describes the planar size of the suspected defect region in the 2D image, reflecting the lateral coverage of the defect on the paper surface.

[0115] Circularity is a dimensionless property ranging from 0 to 1; it describes the regularity of the shape of a defect area. The closer the circularity is to 1, the closer the area is to a circle; the closer it is to 0, the closer it is to an irregular, elongated shape (such as a scratch).

[0116] The grayscale difference (dimensionless, grayscale value range 0 to 255) describes the grayscale contrast between the defective area and the surrounding normal background. The larger the difference, the more obvious the defect is in the 2D image, such as the difference between a dark spot and a light background.

[0117] The average height describes the overall convexity / concave trend of the corresponding defect area in the 3D height map. A positive value indicates an overall convexity, and a negative value indicates an overall depression.

[0118] The variance of height describes the dispersion of height values ​​within the defect area. The larger the variance, the more drastic the height fluctuations within the area (such as the uneven surface of a slurry lump); the smaller the variance, the more uniform the height of the area.

[0119] The defect volume describes the amount of space occupied by the defective area relative to the normal paper surface. Positive values ​​represent the volume of a raised area, while negative values ​​represent the volume of a sunken area, reflecting the three-dimensional severity of the defect.

[0120] The mean value of the a channel in the Lab space describes the red-green bias of the defect area (positive values ​​of the a channel are redder, and negative values ​​are greener), which is used to distinguish color-related defects (such as red spots and green color differences).

[0121] The mean value of the b channel in the Lab color space describes the yellow-blue tint of the defect area (positive values ​​in the b channel are more yellow, and negative values ​​are more blue), and... Together they constitute color characteristics, enhancing the distinguishability of color defects;

[0122] The defect determination result is as follows:

[0123] Surface stains: ;

[0124] Scratches / Indentations: ;

[0125] pit: ;

[0126] Lumps of pulp: ;

[0127] Elimination of false positives: or It was determined to be an invalid defect.

[0128] The inspection result output and production line linkage control module is configured to convert inspection results into traceable reports and executable actions on the production line, enabling defect data recording and analysis, and intelligent production line linkage (such as removing critical defects and marking minor defects). It also supports fault alarms and manual intervention, specifically including:

[0129] The industrial control computer interface displays a 2D image (with red boxes marking the defect locations), a 3D height map (false color display of height, red for protrusions and blue for depressions), a defect distribution heat map (arranged according to the production line time axis, with different colors marking defect types), real-time statistics (number of defects per hour, percentage of each type, such as percentage of lumps and percentage of scratches), and generates an inspection report.

[0130] Defect levels are constructed based on defects:

[0131] Level 1 defect (serious, such as...) (For lumps of pulp): Trigger the scrap kicker (pneumatic push rod), calculate the time it takes for the defect to reach the scrap kick position by the encoder pulse count, send a command 0.5 seconds in advance, and remove the paper to the waste bin;

[0132] Level 2 Defect (Minor): Trigger the UV inkjet printer to print a mark 10mm behind the defect;

[0133] Level 3 Defect (Pending Confirmation): Triggers an audible and visual alarm (flashing red light, buzzer), prompting the operator to view the real-time image and manually determine whether to handle it (manual marking of qualified / unqualified is supported).

[0134] Detection system malfunction (such as camera disconnection or laser failure): immediately trigger the yellow alarm light, display the cause of the malfunction on the interface, and simultaneously suspend production line operation to avoid misoperation;

[0135] Abnormal Defect Rate: Triggers an orange alarm, prompting an inspection of production line processes (such as ink concentration and roller wear), and automatically generates an abnormal defect rate analysis report.

[0136] Example 2

[0137] Please see Figure 2 A machine vision-based method for detecting defects in printed paper includes the following components:

[0138] A hardware platform adapted to the needs of the printing paper production line was built, and 2D high-definition color imaging, 3D shape contour measurement and industrial control integration equipment were configured to support high-quality acquisition of multi-source data.

[0139] Spatiotemporal alignment acquisition of 2D color images and 3D point cloud data is performed, and accurate association of 2D / 3D information in the same paper area is achieved through time compensation and spatial calibration.

[0140] The system performs flat field correction, color conversion, and filtering noise reduction on 2D images, removes invalid points and filters 3D point clouds, and generates a 3D height map that is pixel-level aligned with the 2D image by unifying coordinates.

[0141] Suspected defect areas are extracted from 2D images and 3D height maps respectively, and their features are calculated. These features are then fused to form a comprehensive feature vector, and the defect type is determined according to rules to eliminate false positives.

[0142] It outputs traceable inspection reports and real-time visualized data, enables graded linkage control of the production line based on defect levels, and triggers alarms for system faults and abnormal defect rates.

[0143] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0144] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0145] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely for distinguishing one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0146] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0147] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0148] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0149] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0150] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0151] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0152] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A machine vision based printed paper defect detection system characterized in that, Comprise: System hardware architecture module is configured to build the hardware platform required to adapt to the printing paper production line, to provide hardware support for high-quality collection of 2D texture and 3D topographic data; Multi-source data sensing module is configured to collect 2D color image and 3D point cloud data in space-time alignment, to ensure the association of 2D texture information and 3D topographic information of the same paper area; Defect candidate extraction and fusion module is configured to locate suspected defect area from preprocessed 2D image and 3D height map; through 2D and 3D feature association, form comprehensive feature vector for defect classification, and exclude false positives.

2. The machine vision-based printed paper defect detection system of claim 1, wherein, The system hardware architecture module specifically comprises: According to the detection accuracy to deduce the resolution, to determine the camera required for 2D high-definition color imaging; select light source according to defect type; Select a 1024 pulse / turn rotary encoder installed on a 100mm diameter conveying roller, and ensure fixed distance collection; Select the corresponding line laser profiler for 3D topographic contour measurement; The center axis of the 2D camera is aligned with an error of less than 0.1mm, and the distance between the 2D camera and the 3D camera along the production line is less than 50mm.

3. The machine vision-based printed paper defect detection system of claim 1, wherein, The multi-source data sensing module specifically comprises: After obtaining the corresponding 2D image data and 3D point cloud data respectively, ensure that the 2D and 3D collection delay is less than the preset standard, to realize time synchronization of data collection; Chessboard, convex dot stereoscopic calibration board is used to collect 4 groups of 2D / 3D data at different angles to solve 4×4 rigid transformation matrix The 3D point cloud is converted from the sensor coordinate system to the 2D pixel coordinate system to realize spatial synchronization of data acquisition.

4. The machine vision based printed paper defect detection system of claim 1, wherein, Further comprising: Data preprocessing and coordinate unification module is configured to eliminate noise and interference information in 2D image and 3D point cloud, and convert irregular 3D point cloud to height map aligned with 2D image pixel level, to clear obstacles for subsequent defect candidate area extraction.

5. The machine vision based printed paper defect detection system of claim 1, wherein, The defect candidate extraction and fusion module specifically comprises: 2D Defect Screening: Constructing a Background Model: Collect 500 images of good products, and analyze each pixel. Calculate the mean of gray values with standard deviation The mean background image and standard deviation background image are obtained, and the model is updated every hour with 100 new good product images; Defect detection: Defect pixel determination is performed on each pixel of the current frame image. , is performed on each pixel of the current frame image.

6. The machine vision-based printed paper defect detection system of claim 5, wherein, Further comprising 3D topographic anomaly screening: For each pixel of the height map , take a 3x3 neighborhood window, compute the variance of the Z values , if > , mark as outlier pixel; A Gaussian high-pass filter is used to retain the highly variable regions, and if the absolute value of the filtered Z value is greater than 0.5, the pixel is marked as an anomaly. Calculate the height feature and volume feature of each suspected abnormal area.

7. The machine vision-based printed paper defect detection system of claim 6, wherein, Further comprising defect judgment: Based on the pixel alignment height map, the pixel coordinates of the 2D suspected defect region Directly mapped to the 3D height map, extract the 3D features of the corresponding region; Conversely, 3D suspect regions can also be associated with 2D features, forming a 2D and 3D comprehensive feature vector for each candidate region: ; wherein, is the 2D defect area; is the circularity; is the gray level difference; is the average height; is the height variance; is the defect volume; is the Lab space a channel mean; is the Lab space b channel mean; The defect judgment result is: Surface blemish: ; Scratches / indentations: ; Pit: ; Pannus: ; Misjudgment exclusion: Or , determine as invalid defects.

8. The machine vision-based printed paper defect detection system of claim 1, wherein, Further comprising: The detection result output and production line linkage control module is configured to convert the detection result into traceable report and production line executable action, to realize defect data recording and analysis and intelligent linkage of production line, and to support fault alarm and manual intervention.

9. The method of detecting defects in printed paper based on machine vision according to any one of claims 1-8, characterized in that, Comprise the following steps: Build a hardware platform that meets the needs of the printing paper production line, configure 2D high-definition color imaging, 3D topographic contour measurement and industrial control integrated equipment; Collect 2D color image and 3D point cloud data in space-time alignment, realize accurate association of 2D / 3D information of the same paper area through time compensation and space calibration; Perform flat field correction, color conversion and filter denoising on 2D image, remove invalid points and filter on 3D point cloud, and generate 3D height map aligned with 2D image through coordinate unification; Extract suspected defect area from 2D image and 3D height map respectively and calculate features, form comprehensive feature vector, determine defect type according to rules and exclude false positives.

Images