A method for defect homologous identification and batch processing of multi-opening printed matter

Abstract

This invention discloses a method for identifying and batch processing defects in multi-page printed materials, relating to the technical field of printing quality inspection. The method first establishes a three-level coordinate system to achieve structured alignment of the entire page. Secondly, it collects multi-dimensional features to establish a statistical sample library. It then constructs a normal process drift model to dynamically compensate for the target features of each page size. The entropy weight method quantifies the weights of 12-dimensional features, including printing sequence correlation deviations. Combined with weighted distance, adjacent distance difference, and IQR constraint verification mechanisms, it achieves defect grouping and anomaly removal. Finally, a three-component Gaussian mixture model is introduced to complete the final determination of the source relationship, followed by batch labeling. This invention transforms the detection complexity from O(N) step-by-step judgment to batch processing based on position indexing, significantly improving quality control efficiency and providing accurate traceability. Through multi-dimensional feature cascade analysis and probabilistic model judgment, it effectively distinguishes between regular process fluctuations and real defect differences, eliminating environmental drift misjudgments.

Description

Technical Field

[0001] This invention relates to the technical field of printing quality inspection, and more specifically, to a method for identifying and batch processing defects from the same source in multi-page printed materials. Background Technology

[0002] Before mass production, printing companies typically need to conduct a proofing inspection of the first printed product to avoid large-scale scrapping due to layout errors. In the first inspection stage, current technology often uses a method of comparing the electronic file of the printing plate with the image of the large-format printed product for defect identification. However, this method only outputs the absolute location of the differences on the large sheet, requiring subsequent manual verification of each of these absolute locations to complete the quality judgment of the multi-page printed product. This type of method generally has the following drawbacks:

[0003] (1) The verification efficiency is extremely low and the manual burden is heavy: only the absolute position of the difference point is output. For multi-page printed materials (a large sheet of paper with N identical labels), even if the systemic defect caused by the same plate-making error is identified as N independent difference points, the manual verification must click, check and judge one by one. The verification efficiency decreases linearly with the increase of the number of pages N. The time taken for a single first inspection is often 20-30 minutes, which greatly restricts the production cycle.

[0004] (2) Lack of homology correlation analysis and inability to process in batches: Existing technical methods do not deeply mine the image features (such as shape, color, texture, etc.) of the differences, and lack defect homology analysis, which makes the system unable to summarize the homology problems in multiple images and perform batch operations. The review process involves a lot of repetitive work and low review efficiency.

[0005] (3) Insufficient accuracy in judgment under complex deformation: In the real printing environment, due to paper stretching, mechanical shaking or registration deviation, defects between multi-page products often have slight displacement or deformation (not completely matched). Existing technical methods rely too much on "completely consistent local coordinates" or simple "global feature matching" for defect judgment. When faced with targets that are not completely matched, it is very easy to misjudge defects that belong to the same source as targets that are not from the same source, resulting in inaccurate statistical analysis or even missed judgment.

[0006] (4) Limitations of existing patented technologies: For example, the pre-press matching detection method described in patent CN108182677A only focuses on pre-press comparison between design manuscripts, lacks the ability to detect "high-resolution real printed images", and does not involve multi-page statistical analysis, homology verification and subsequent index batch processing mechanism based on the image features of difference points. Summary of the Invention

[0007] The purpose of this invention is to provide a method for identifying and batch processing defects in multi-page printed materials, so as to solve the problem of low efficiency in multi-page printing and defect verification in the prior art, support batch processing of multi-page materials at one time, and reduce verification time and manual verification costs.

[0008] The technical solution of this invention is: a method for identifying and batch processing defects in multi-page printed materials, the method comprising:

[0009] Step 1: Create a standard single-mold small-format image of the target product and acquire a large-format multi-format printed image. Use the standard image to segment and extract the large-format multi-format printed image to obtain single-mold small-format images of multiple sizes to be tested. Establish a three-level coordinate system including the large-format absolute coordinate system, the size index, and the local coordinate system. Based on the three-level coordinate system, determine the coordinate registration relationship between the large-format multi-format printed image and each single-mold small-format image. Based on the coordinate registration relationship, determine the position of each target object in the large-format multi-format printed image in the local coordinate system. The target objects include product text and image elements and defects.

[0010] Step 2: Select single-mode small-size image samples from the large-format multi-page printed images in layers according to the printing time sequence, obtain the multidimensional features of all target objects in the samples, including geometric features, color features, texture features and position features, remove abnormal samples whose multidimensional feature values ​​deviate from the template mean, and use the remaining valid samples to establish a statistical sample library.

[0011] Step 3: For a single product text and image element, calculate the time series trend coefficient corresponding to each dimension feature based on the printing sequence number and multi-dimensional feature value of each valid sample in which it is located. Construct a normal process drift model using the multi-dimensional time series trend coefficients corresponding to the product text and image element. Compensate for the multi-dimensional features of each target object in the large-format multi-page printing image based on the normal process drift model.

[0012] Step 4: Obtain the multidimensional features of all target objects in the large-format multi-page printed image, and compensate for the multidimensional features of all product text and image elements based on the normal process drift model.

[0013] Step 5: Select one target object from all target objects in the large multi-page printed image as the target sample object, search for other target objects in its neighborhood, and filter these target objects based on positional features to form an initial group of common defects. Quantify the feature discrimination based on the information entropy of each dimension feature within the group and determine the weight of each dimension feature.

[0014] Step 6: Calculate the weighted total distance between other target objects and target sample objects in the initial same defect group based on the feature weights of each dimension, sort all weighted total distances in ascending order, determine the elbow point based on the difference of the adjacent distance, perform constraint verification on the elbow point distance, and update the same defect group based on the constraint verification results.

[0015] Step 7: Establish a three-component Gaussian mixture model, train the mixture model using the multi-dimensional features of the product text and image elements in the effective samples, calculate the homology probability between other target objects and target sample objects in the same defect group using the trained model, and update the same defect group again.

[0016] Step 8: Select target sample objects again from the remaining target objects of the multi-page printed image, and obtain the same defect group again in the manner of steps 5 to 7. Repeat this process until all target objects are grouped. Perform batch marking processing on each group of same defects. The marking results include qualified and unqualified.

[0017] The beneficial effects of this invention are:

[0018] First, the technical solution in this invention establishes a three-level coordinate system, including a large-format absolute coordinate system, an opening index, and a local coordinate system, to accurately align the structure of large-format multi-page printed images with the single-format small-page images within them. This establishes a unified spatial benchmark for subsequent defect source identification, location, and indexing. By breaking down the spatial isolation of different opening sizes through the unified coordinate system, it enables the subsequent identification of systemic defects using position association and feature clustering to switch from single-page detection to multi-page batch processing. This transforms the detection complexity from an O(N) mode of individual judgment to a batch processing mode based on position indexing. While ensuring identification accuracy, this significantly improves the efficiency of multi-page printing quality control. It also provides clear physical coordinate basis for subsequent defect tracing, which helps in adjusting printing equipment and troubleshooting.

[0019] Secondly, the technical solution of this invention constructs a defect homology determination mechanism based on multi-dimensional features. It selects 12 key features, including geometric features, color features, texture features, positional features, and process-related features, to construct a multi-dimensional feature vector. Information entropy is used to distinguish and quantify each feature, assigning initial weights and performing reasonable optimization. Then, the weighted total distance is calculated to initially group the sample objects in the graph. Combined with the neighbor distance difference and IQR constraint verification mechanism, numerical mutation points are identified, effectively identifying and eliminating heterogeneous interference terms to obtain updated homology defect groups. Furthermore, a three-component Gaussian mixture model is introduced, utilizing the homology relationship within the model group for probabilistic determination. This upgrades homology determination from traditional hard threshold rules to refined statistical probability determination, resulting in optimized homology defect groups and significantly improving determination accuracy.

[0020] The technical solution of this invention is based on multi-dimensional features for defect homology determination. It innovatively introduces printing sequence correlation deviation as a key process correlation feature, deeply integrating it into a 12-dimensional feature vector system. This system measures and reflects the evolution intensity of each feature with the production sequence in real time, avoiding misjudgments or omissions caused by single feature threshold judgments. It achieves multi-dimensional feature cascade analysis, objectively measuring the information contribution of each dimension through the entropy weight method, enabling the model to automatically adapt to the characteristic differences of different batches of printed materials. The mechanism of distance mutation plus statistical constraint verification effectively eliminates abnormal target objects, reduces outlier interference, and significantly improves the stability of homology grouping. Simultaneously, a three-component Gaussian mixture model is introduced to probabilistically model the multi-peak difference distribution under process fluctuations, ensuring that homology identification maintains extremely high accuracy even when facing printing environment drift. The technical solution of this invention has more comprehensive identification dimensions, enabling efficient and accurate determination of homology defects in large-format multi-page printed images.

[0021] Third, the technical solution of this invention addresses the slow process drift phenomenon that occurs with the printing sequence during multi-page printing. It first calculates the time series trend coefficients corresponding to each dimension of the feature, then uses these coefficients to construct a normal process drift model. This model is then used to compensate and correct the multidimensional features of all target objects before determining the origin of the defect, eliminating false differences in features caused by process fluctuations (such as ink consumption, paper deformation, and other time variables). The technical solution of this invention can effectively distinguish between normal process regularity fluctuations and real random differences in defects, avoiding the drawback of traditional algorithms misjudging process drift in long-cycle printing as non-originating defects. This significantly improves the discrimination stability of origin determination during long-term, high-intensity continuous printing. Furthermore, the technical solution of this invention incorporates the spatiotemporal fluctuation patterns of the printing process into the identification process, greatly enhancing the robustness of the system under complex working conditions and environmental drift, and aligning with actual production business logic. Attached Figure Description

[0022] The advantages of the above and additional aspects of the present invention will become apparent and readily understood in the description of the embodiments in conjunction with the following drawings, wherein:

[0023] Figure 1 This is a schematic flowchart of a method for identifying and batch processing defects in multi-page printed materials according to an embodiment of the present invention.

[0024] Figure 2 This is a schematic diagram of the coordinate registration relationship between a single-mold small-format standard image and a large-format multi-page printing image according to an embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram of batch marking processing of single-mode small-opening standard images corresponding to a single set of homogeneous defects according to an embodiment of the present invention;

[0026] Figure 4 A detection apparatus according to an embodiment of the present invention for performing a method for identifying the common source of defects and batch processing of multi-page printed materials;

[0027] Figure 5 It is a large-format, multi-page printed image of a certain type of printed material in Example 1;

[0028] Figure 6 This is a comparison image of a single single-mode small-open image in Example 1 after alignment with the standard single-mode small-open image;

[0029] Figure 7 This is a comparison image of a magnified portion of a single single-mode small-open image in Example 1, aligned with a magnified portion of a standard single-mode small-open image;

[0030] Figure 8 This is a comparison image of a magnified portion of a single-mode small-scale image without local pattern offset defects in Example 1, and a magnified portion of a small-scale image with local pattern position offset, after alignment.

[0031] Figure 9 This is a magnified view of a small-scale single-mold image without overprinting defects in Example 1;

[0032] Figure 10 This is a magnified view of two single-mold small-open images containing printing deviation defects in Example 1;

[0033] Figure 11 This is a magnified view of a small-scale image of a single mold containing printing deviation defects in Example 1;

[0034] Figure 12 This is a magnified view of two single-mold small-open images containing printing deviation defects in Example 1. Detailed Implementation

[0035] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other.

[0036] In the following description, many specific details are set forth in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0037] like Figure 1 As shown in the figure, this embodiment provides a method for identifying and batch processing defects from the same source in multi-page printed materials. The method includes:

[0038] Step 1: Create a standard single-mold small-format image of the target product and acquire a large-format multi-format printed image of the target product. Use the standard image to segment and extract the large-format multi-format printed image to obtain single-mold small-format images of multiple sizes to be tested. Establish a three-level coordinate system including the large-format absolute coordinate system, the size index, and the local coordinate system. Based on the three-level coordinate system, determine the coordinate registration relationship between the large-format multi-format printed image and each single-mold small-format image (to lay the spatial correspondence foundation for subsequent indexing and homology recognition of target objects at the same position across sizes). Based on the coordinate registration relationship, determine the position of each target object in the large-format multi-format printed image in the local coordinate system. The target object includes product text and image elements and defects.

[0039] It should be noted that in this embodiment, "open" represents a small page. In multi-page printing, the physical position of each open page on the large sheet of paper is relatively fixed. For example, in 44-page printing, there are 44 small page images, and each small page can be arranged in an 11×4 matrix. In order to achieve accurate indexing of the same position, a three-level coordinate system needs to be established to clarify the overall coordinates of the large multi-page printed image (large page), the relative position of each small page image (small page), and the pixel position inside each small page image in sequence.

[0040] Creating a standard small-scale single-mold design image of the target product, specifically including:

[0041] First, obtain the electronic file (.pdf) of the single-mold small-format printing corresponding to the target product from the existing database. In the electronic file, set the boundary of the small-format image corresponding to the target product (i.e., the boundary of the small-format layout) so that the pixel area where the target product is located is centered. Then, extract the pixel area where the small-format image is located. For the extracted small-format image, adjust the scale according to the actual photographed single-mold small-format image corresponding to the target product to ensure that its size ratio is consistent with the actual photographed image and meets the quality control standards required for defect detection, thus obtaining the standard single-mold small-format image of the target product.

[0042] Large-format multi-page printed images are segmented and extracted using standard images to obtain single-mode small-format images of multiple sizes to be tested. A three-level coordinate system is established, including the absolute coordinate system of the large-format, the size index, and the local coordinate system. Based on the three-level coordinate system, the coordinate registration relationship between the large-format multi-page printed images and each single-mode small-format image is determined, specifically including:

[0043] Based on the size of the single-mold small-format standard image and the position of the target product pixel area in it, set the boundary of the single-mold small-format image corresponding to each target product pixel area in the large-format multi-page printing image (i.e., the boundary of each small page in the large page), so that each segmented single-mold small-format image has the same size and proportion as the single-mold small-format standard image, and the position of the target product pixel area in the image corresponds.

[0044] like Figure 2 As shown, a first reference point (such as the upper left, lower left, upper right, or lower right corner) is selected for the large-format multi-page printing image. A large-format absolute coordinate system O-XY is established with the first reference point as the origin O(0,0), where the X-axis is along the paper length direction and the Y-axis is along the paper width direction, with units of physical dimensions (e.g., mm). The pixel coordinates in the large-format multi-page printing image are mapped to the physical coordinates in the large-format absolute coordinate system O-XY, ensuring a one-to-one correspondence between the image position and the actual printing position. A second reference point (such as the upper left corner, or other positions with no process deviations and clear, easily identifiable features among all small pages) is selected for each single-mold small-page image in the large-format multi-page printing image. A local coordinate system o-xy is established with the second reference point as the origin o(0,0), where the x and y axes are parallel to the corresponding X and Y axes of the large-format absolute coordinate system. Each single-mold small-page image (i.e., each page size) in the large-format multi-page printing image is numbered, and the coordinates (X, Y, Y) of the second reference point corresponding to each single-mold small-page image in the large-format absolute coordinate system are recorded. n ,Y n The coordinate registration relationship between the large-format multi-page printed image and each single-format small-format image is obtained, expressed as: x n =XX n y n =YY n Where n is the aperture index of each single-mode small-format image in the large-format multi-page printing image, and (X,Y) are the physical coordinates of the pixels in the large-format multi-page printing image. n ,Y n (x) represents the coordinates of the second reference point corresponding to the single-mode small opening image of the nth opening degree. n ,y n ) represents the physical coordinates of the pixel in the single-mode small-opening image of the nth opening degree.

[0045] In this embodiment, the above three-level coordinate system allows for indexing of defects or product graphic elements at the same location. The local coordinate systems for different aperture sizes within the large-format image are identical and correspond to the standard single-mode small-aperture image. The position (x, y) in a certain local coordinate system corresponds to the absolute coordinates (X, y) of the large-format image. n +x,Y n For example, if we have a 1*1 sheet (10mm, 20mm) and a 1*2 sheet (10mm, 20mm), and the reference point for the 1*2 sheet is (X2, Y2), then their absolute coordinates on the large board are (X1+10, Y1+20) and (X2+10, Y2+20) respectively. These two positions can be indexed by (x=10, y=20).

[0046] The position of each target object in a large-format multi-page printed image is determined in the local coordinate system based on the coordinate registration relationship. Specifically, the position of each target object in the large-format multi-page printed image is determined using the obtained coordinate registration relationship x. n =XX n y n =YY n Transform to a local coordinate system to obtain the distribution of each target object in the local coordinate system.

[0047] It should be noted that the large-format multi-page printed images in this embodiment have obtained the recognition results of each target object (including product graphic elements and existing defects) through image processing. Regarding the specific target object detection and segmentation methods, those skilled in the art can choose known technical means to implement them according to actual needs, such as, but not limited to, methods based on threshold segmentation, edge detection, region growing, traditional machine learning methods or deep learning models. This application mainly focuses on further realizing cross-small-format homology comparison, batch processing and defect tracing based on the target object recognition results already obtained. Therefore, the specific algorithm for the aforementioned target object recognition will not be elaborated in detail.

[0048] Step 2: Select single-pattern small-format image samples from the large-format multi-page printed images in chronological order of printing time, obtain the multidimensional features of all target objects in the samples, including geometric features, color features, texture features and position features, remove abnormal samples whose multidimensional feature values ​​deviate from the template mean, and use the remaining valid samples to establish a statistical sample library.

[0049] Specifically, according to the printing time sequence (large-format printing is carried out in a "top-to-bottom" order, and process fluctuations may drift with the printing sequence), the individual small-format images in the large-format multi-page printing images are divided into three groups: early stage (small-format at the top of the large-format, early printing stage), middle stage (small-format in the middle of the large-format, stable printing stage), and late stage (small-format at the bottom of the large-format, late printing stage). A predetermined number of individual small-format image samples (e.g., 20-30 images in total, covering the entire printing cycle and avoiding sample bias) are randomly selected from these three groups in proportion to form an initial statistical sample library. Any product text and image element is selected as a reference target (e.g., product logo). The multidimensional feature values ​​of the reference target in all individual small-format image samples are obtained. For a single dimension of feature, its feature mean in all individual small-format image samples is calculated as the template mean corresponding to that dimension of feature. Then, individual small-format image samples that exceed the template mean ±10% range are removed. This process is repeated for each product text and image element. The remaining valid samples are used to establish a local feature statistical sample library to improve statistical accuracy and avoid interference.

[0050] Obtain the multidimensional feature values ​​of the reference target across all single-mode small-format image samples. Specifically, for a single target object, obtain its geometric features, color features, texture features, and positional features, including:

[0051] The geometric features of a single target object include roundness C, edge smoothness E, and area S. The area S is the actual pixel area occupied by the target object (e.g., if it occupies 5 pixels, its area S = 5). Roundness C and edge smoothness E are expressed as follows:

[0052] ;

[0053] In the formula, L is the actual perimeter of the target object (the sum of edge pixel distances, with horizontally or vertically adjacent pixel distances counted as 1, and diagonally adjacent pixel distances counted as 1). ), Let S be the ideal perimeter of the target object (assuming the target object is a regular circle or rectangle, calculate the perimeter of the perfect circle or rectangle with the same area S, denoted as S). ); ); roundness reflects the compliance of the target shape, edge smoothness reflects the quality of the target edge, and area reflects the overall size consistency of the target.

[0054] Color characteristics of a single target object include mean color saturation. (The average color saturation of the target area is a core indicator of visual consistency), R channel average. G channel mean and B channel mean .

[0055] Texture features of a single target object include GLCM energy En and LBP mean. GLCM energy En is an energy value calculated based on the gray-level co-occurrence matrix, reflecting the texture coarseness and uniformity of the target surface. LBP mean It is the statistical mean of the Local Binary Pattern (LBP) features, reflecting local texture consistency.

[0056] The positional features of a single target object include relative distance deviation ΔD and relative angle deviation Δθ (core indicators for multi-opening alignment). Distance deviation ΔD is the straight-line distance between the actual position of the target object and its corresponding template reference position in the single-mold small-opening standard image. The specific calculation process is as follows: Select any product text / image element in the single-mold small-opening image where the current target object is located. Use the center coordinates of this product text / image element in the current single-mold small-opening image as the actual position. Simultaneously, extract the center coordinates of this product text / image element in the single-mold small-opening standard image as the template reference position. Calculate the actual position and the template reference position. The Euclidean distance between the positions is used as the relative distance deviation ΔD; the relative angle deviation Δθ is the deviation between the actual angle of the target object and the template reference angle corresponding to it in the single-mold small-scale standard image. The specific calculation process is as follows: select two feature points (such as the two edge corners of any product text image) from the single-mold small-scale image where the current target object is located, calculate the horizontal angle between the lines where these two feature points are located, read the coordinates of these two feature points in the single-mold small-scale standard image, and calculate the horizontal angle between the lines where these two feature points are located in the single-mold small-scale standard image. Subtract the two to obtain the relative angle deviation Δθ.

[0057] In this embodiment, the product graphic elements can be a brand circular logo, the product name "XX Repair Shampoo", the capacity label "500ml", and gradient color block decorations, etc., which can be manually set in actual operation.

[0058] In this embodiment, the calculation process of GLCM energy En is as follows: First, the target object region image is compressed in grayscale to construct a grayscale co-occurrence matrix reflecting the spatial correlation of pixels. Then, the sum of squares of all elements in the matrix is ​​calculated. This index is a core indicator in this field for characterizing image texture consistency. The more uniform the grayscale distribution and the simpler the texture on the target object surface, the higher the En value. LBP mean L m The calculation process is as follows: Within the target object region, for each pixel as the center, its grayscale value is compared with that of its eight neighboring pixels. If the grayscale value is greater than that of the center pixel, it is marked as 1; otherwise, it is marked as 0. This generates an 8-bit binary number, which is then converted into a decimal LBP operator. The average value of the global LBP operator is then calculated to obtain the result. Both are common indicators for micro-texture recognition in this field, and their calculation process belongs to mature technologies in the field of image processing, so they will not be elaborated here.

[0059] Step 3: For a single product text and image element, calculate the time series trend coefficient corresponding to each dimension feature based on the printing sequence number of each valid sample and the multi-dimensional feature value. Use the multi-dimensional time series trend coefficients corresponding to the product text and image element to construct a normal process drift model.

[0060] It should be noted that, since multi-page printing is a continuous process, the features corresponding to the text and graphic elements of the product will "slowly drift" with the printing sequence (such as a slight decrease in the average color value due to the gradual consumption of ink, which is a normal process fluctuation). Before verifying and determining the same source target, this drift trend should be used to compensate for the multi-dimensional features of each product text and graphic element in order to improve the accuracy of subsequent same source identification.

[0061] Specifically, the valid samples in the statistical sample library are sorted according to the printing time order, and the printing sequence number corresponding to each valid sample is obtained (e.g., in 44-page printing, the printing sequence number corresponding to the 5th sample is 5). A single product text and image element (which exists in the single-mode small-page image within a large-format multi-page printing image) is selected. Based on the printing sequence number t of each valid sample and the multi-dimensional feature values ​​(values ​​corresponding to geometric features, color features, texture features, and positional features), the time series trend coefficient corresponding to each dimension feature is calculated using linear regression. The time series trend coefficient corresponding to a single dimension feature is expressed as:

[0062] ;

[0063] ;

[0064] ;

[0065] In the formula, Let j be the j-th dimension feature value of a single product image element. This represents the time series trend coefficient corresponding to the j-th dimension of the text and image elements of a single product. The intercept of the j-th dimension feature of a single product's text and image element, where j = 1, 2, ..., 11. For a single product image element, the feature observation value in the i-th sample. This represents the average feature of a single product's text and image elements across all samples. This represents the printing sequence number of the i-th sample containing the text and image elements of a single product. The average value of the sample printing sequence number is used; the multi-dimensional time series trend coefficients corresponding to the text and image elements of the product are summarized to construct a normal process drift model.

[0066] It should be noted that, based on color characteristics For example, if the printing sequence number and Perform linear regression to obtain = -0.02t + 231, then the slope -0.02 means that for every 10 small prints, the color characteristics... The decrease of 0.2 is in line with the normal trend of ink consumption.

[0067] Step 4: Obtain the multidimensional features of all target objects in the large-format multi-page printing image, and compensate for the multidimensional features of each target object in the large-format multi-page printing image based on the normal process drift model.

[0068] Specifically, multi-dimensional features of all target objects (including product text and image elements and defects) in each single-mold small-scale image are obtained, including geometric features, color features, texture features, and positional features. For a single-dimensional feature of a product text and image element, the printing order correlation deviation is calculated based on its feature value in each sample. This printing order correlation deviation is used to reflect the degree of process drift in multiple printing runs, and its calculation formula is as follows:

[0069] ;

[0070] In the formula, T[j] represents the printing order correlation deviation corresponding to the j-th dimension feature of a single product's text and image element, and its value ranges from [-1, 1]. The Pearson correlation coefficient ranges from -1 to 1. The j-th dimension feature of a single product image element is the feature value of the i-th sample (including the feature values ​​corresponding to geometric features, color features, texture features, and position features).

[0071] Based on the calculated value of T[j], compensation is applied to the multidimensional features of all target objects. Specifically, for the j-th dimension feature of a single target object, when T[j] > 0, it means that its feature value is positively correlated with the printing sequence number, and the magnitude of the feature value increases with the printing sequence (e.g., the cumulative stretching of paper leads to an increase in the LOGO area S). When T[j] < 0, it means that its feature value is negatively correlated with the printing sequence number, and the magnitude of the feature value decreases with the printing sequence (e.g., ink consumption leads to an increase in R). m (attenuation), uniformly subtracting eigenvalues Compensation will be performed, and the value after compensation will be... When T[j]=0, it means that the correlation between its feature value and the printing sequence number is weaker, and the feature value has no obvious temporal drift (such as the edge smoothness E being less affected by the printing sequence, T≈0), so no compensation is needed.

[0072] It should also be noted that the method for obtaining the multidimensional features of all product text and image elements in each single-mode small-scale image is the same as the method for obtaining the multidimensional features of all target objects in the sample in step 2, and will not be repeated here.

[0073] Step 5: Select one target object from all target objects in the large-format multi-page printed image as the target sample object, search for other target objects in its neighborhood, and filter these target objects based on positional features to form an initial group of common defects. Quantify the feature discrimination based on the information entropy of each dimension feature within the group and determine the weight of each dimension feature.

[0074] It should be noted that, since different features have different importance in determining the same source, before identifying the same source for each defect, it is necessary to first quantify the feature discrimination based on the information entropy of each dimension feature, determine the weight of the 12-dimensional features, and achieve a preliminary weight allocation without subjective bias. Then, the subsequent comprehensive score and dynamic threshold calculation are carried out to improve the accuracy of the same source identification.

[0075] Specifically, one target object is randomly selected from all target objects in the large-format multi-page printed image as the target sample object. In the local coordinate system, all target objects that fall within its neighborhood and are located in other single-mode small-format images (i.e., other single-mode small-format images contained in the large-format multi-page printed image) are searched. For all the searched target objects, they are filtered based on the relative distance deviation ΔD and the relative angle deviation Δθ. The filtering process is as follows: for a single target object, if its ΔD value and Δθ value are both within the preset range, it is kept; if its ΔD value or Δθ value exceeds the preset range, it is removed. The filtered target objects are then classified into the initial homologous defect group.

[0076] The feature discrimination is quantified based on the information entropy of each dimension of features within the group, and the weights of each dimension of features are determined, specifically including:

[0077] Read the geometric, color, texture, and positional features of all target objects in the initial group of shared defects. For a single-dimensional feature, calculate the printing order correlation deviation based on its feature values ​​on each target object, resulting in 11-dimensional printing order correlation deviations. These 11-dimensional printing order correlation deviations are used as process correlation features and categorized into multi-dimensional features to form 12-dimensional features. The calculation method for the printing order correlation deviation is the same as that in step 4, and will not be repeated here. Read the 12-dimensional features of all target objects in the group. For the j-th dimension feature, use positive transformation to eliminate dimensional differences, expressed as:

[0078] ;

[0079] In the formula, For the j-th dimension feature of the r-th target object, After positive transformation value, , These are the minimum and maximum values ​​of the j-th feature within the group, respectively; based on the forward processing... The probability of the j-th feature on the r-th target object is calculated as follows:

[0080] ;

[0081] In the formula, Let R be the probability of the j-th feature on the r-th target object, and R be the number of target objects in the group; based on probability The information entropy of the j-th dimension feature is calculated as follows:

[0082] ;

[0083] In the formula, Let the information entropy of the j-th dimension feature be denoted as . The smaller the feature size, the stronger the distinguishability. The constant coefficients, Based on information entropy Calculate the difference coefficient of the j-th dimension feature Then based on the difference coefficient The preliminary weights of the j-th feature are calculated and expressed as follows:

[0084] ;

[0085] In the formula, The initial weights are for the j-th dimension feature; by iterating through the features of each dimension, the initial weights of the 12-dimensional features are obtained.

[0086] The initial weights of the 12-dimensional features are optimized as follows: if a feature is a core judgment indicator, its initial weight is increased by 10% to 20%; if a feature is a process-related indicator, its initial weight is increased by 5% to 10%; if a feature is affected by normal process fluctuations, its initial weight is decreased by 10% to 15%; if a feature does not belong to the above three categories, no adjustment is made. The adjusted weights are then normalized to obtain the final optimized weights of the 12-dimensional features. Features belonging to core judgment indicators include relative distance deviation ΔD, roundness, and average color saturation; features belonging to process-related indicators include printing sequence correlation deviation; and features affected by normal process fluctuations include relative angle deviation Δθ and area.

[0087] In this embodiment, different features have varying importance in determining homologous targets. This application assigns weights to different features based on the dispersion of feature data and their actual detection value, avoiding subjective experience bias. Specifically, the discriminative power of each feature is quantified by calculating its information entropy. The smaller the information entropy of each feature, the stronger its discriminative power in determining homologous targets and tracing process drift, and the greater its corresponding weight. In another specific embodiment, the average color saturation... If the entropy value is the smallest (0.28), then its weight is set to the largest. =0.15), the entropy value of the relative angle deviation Δθ is the largest (0.72), so its weight is set to the smallest (ω). Δθ=0.04); Regarding the adaptability of the weight assignment: It matches the color saturation of the target detection ( The top three features with the highest weights are (ω=0.15), (C, ω=0.12), and (ΔD, ω=0.11). In practice, these three features directly determine visual consistency and positional alignment, and can meet the multi-page alignment detection requirements of high consistency targets. They are the core focus items during manual review. The printing sequence correlation deviation (T, ω=0.09) supporting the process traceability requirements has a higher weight than features such as area (S) and angle deviation (Δθ). It can achieve the purpose of analyzing the root cause of defects through process correlation features (e.g., when T < -0.8, it can be known that insufficient ink volume leads to color drift in subsequent versions). Area (S, ω=0.05) and angle deviation (Δθ, ω=0.04) have the lowest weights because these two types of features are easily affected by normal process fluctuations such as slight paper stretching and minor equipment vibration. Over-reliance on them can lead to misjudgment. It is important to avoid judging normal process deviations as non-homogeneous features.

[0088] Step 6: Calculate the weighted total distance between other target objects and target sample objects in the initial same defect group based on the feature weights of each dimension. Sort the weighted total distances of other target objects in the group in ascending order. Determine the elbow point where the value changes abruptly based on the difference of the adjacent distances. Perform constraint verification on the elbow point distance. Update the same defect group based on the constraint verification results.

[0089] Specifically, for any other target object within the group, the weighted total distance between it and the target sample object is calculated based on the optimized weights of the 12-dimensional features, and is expressed as:

[0090] ;

[0091] ;

[0092] In the formula, Let F be the weighted total distance between the f-th other target object in the group and the target sample object, where F is the number of other target objects in the group besides the target sample object. For the optimized weights of the j-th dimension feature, Let be the distance value between the f-th other target object in the group and the target sample object on the j-th feature dimension.

[0093] The smaller the value, the higher the similarity between other target objects and the target sample. The larger the value, the lower the similarity between other target objects and the target sample; the higher the weight of a feature (such as...). =0.15), the greater the impact of the distance on the overall similarity, for example, the greater the impact of the distance on the overall similarity of samples of the same category. The distance must be small; otherwise, even if the distances of other features are small, the total distance will exceed the limit, which conforms to the judgment logic of prioritizing core industrial features and identifying defects of the same origin.

[0094] The calculated weighted total distance of each other target object within the group Sort the data in ascending order so that the transition from defects of the same category to defects of different categories within a group is transformed into a gradual change in the distance sequence from flat to steep, making it easier to identify abrupt change points and calculate the weighted total distance between adjacent groups. The difference , is represented as:

[0095] ;

[0096] In the formula, The difference is the s-th weighted total distance in the distance sequence, where s = 2, 3, ..., F; This reflects how much farther the s-th other target object is from the target sample object compared to the (s-1)-th other target object, and other target objects of the same category (i.e., of the same origin). Smaller (distance increases slowly), other target objects in different categories It will suddenly increase (distance increases dramatically).

[0097] In the difference Find the maximum value in the sequence and will The corresponding position is denoted as the elbow point. ,Right now The distance corresponding to the elbow point As the initial constraint threshold Setting the location of the maximum difference as the elbow point for subsequent constraint verification means that from the first... -1 other target objects to the The most dramatic decrease in similarity occurred among other target objects, and this position may be the dividing line between the same category and different categories.

[0098] Constraint verification is performed on the elbow point distance, and the same-origin defect group is updated based on the constraint verification results, specifically including:

[0099] The IQR method was used to count distances greater than or equal to the elbow. distance To obtain the verification criteria, specifically, calculate the quartiles Q1 and Q3, where Q1 is the lower quartile value (i.e., the 25th percentile of the distance to be counted). ≤ this value), Q3 is the upper quartile (i.e., the 75th percentile of the distances to be counted). (≤ this value), the IQR value (interquartile range) is calculated based on the interquartiles Q1 and Q3, expressed as IQR = Q3 - Q1, and the upper limit value is calculated based on Q3 and the IQR value. =Q3 + 1.5 × IQR; based on the upper limit value Determine the initial constraint threshold Whether it is effective, if Then the initial constraint threshold Effective, for groups with a distance less than or equal to the elbow point. Other target objects are classified as homologous defects of the target sample objects, and those with a distance greater than the elbow point within the group are removed. Other target objects, if Then use the elbow point corresponding difference Perform mutation verification, if This indicates that there are other target groups with smaller differences (other target groups with smaller differences refer to those with a distance less than the elbow point). distance If there are feature mutations in the data, determine the initial constraint threshold. Effective, for groups with a distance less than or equal to the elbow point. Other target objects are classified as homologous defects of the target sample objects, and those with a distance greater than the elbow point within the group are removed. Other target objects, if This indicates that even at the point of maximum difference within a group (where the degree of similarity decreases the most), there is insufficient difference to determine the initial constraint threshold. Invalid, adjust it to 1, and classify all other target objects in the group as homologous defects of the target sample object.

[0100] It should be noted that the constraint verification process, by performing correlation verification on the degree of absolute difference between other target objects, the location of the difference mutation, and the degree of difference mutation itself, can effectively determine whether other target objects (abnormality means different categories) exist for abnormal data.

[0101] It should be noted that this application is based on the similarity of the performance effects of defects from the same source. However, due to the characteristic of feature drift caused by process fluctuations, 12 key features in 5 categories (geometric features, color features, texture features, spatiotemporal correlation features, and process correlation features) are selected. Using these 12 features, the same source of defects is determined through multi-dimensional feature cascade analysis and entropy weight assignment. This can comprehensively evaluate the visual consistency of defects and achieve accurate determination of defects from the same source. This multi-feature fusion strategy not only conforms to the physical laws and visual perception of multi-page printing, but also balances recognition accuracy and computational efficiency (without the need for individual judgment and processing). It can avoid misjudgment or omission caused by a single feature and ensure reliable traceability of the source of defects in complex printing environments. Simultaneously, process-related features are introduced to fully consider the spatiotemporal correlation of multi-page printing, namely the slow fluctuations that occur over time and in sequence during the process (such as color decay due to ink consumption or area increase due to cumulative paper stretching). The large-format image is divided into three groups of samples—early, middle, and late—based on the printing sequence, covering the entire printing cycle. A normal process drift model (such as a linear regression trend equation) is constructed to compensate for these normal fluctuations, correcting them and avoiding misjudging process drift as non-homogeneous defects. This reduces the risk of error and improves the overall stability and practicality of the judgment. In this application, the use of multi-dimensional features for homogeneity identification and the application of "process-related features" to regress feature drift caused by "spatiotemporal fluctuations in the process flow" are more in line with actual business practices. This approach conforms to the physical laws of multi-page alignment detection, the visual perception logic of homogeneous defects, and the requirements of process correlation, while achieving a balance between accuracy and efficiency.

[0102] Step 7: Establish a three-component Gaussian mixture model. Train the mixture model using the multi-dimensional features of the product text and image elements in the effective samples, obtain the model parameters, and use the trained model to calculate the homology probability between other target objects and target sample objects in the updated homology defect group. Update the homology defect group based on the homology probability to obtain the final homology defect group.

[0103] Specifically, a three-component Gaussian mixture model is established, comprising three Gaussian components. These three components correspond to samples from the same period, continuous periods, and interval periods, respectively. Each Gaussian component includes three parameters: a mean vector, a covariance matrix, and component weights. The three-component Gaussian mixture model is represented as follows:

[0104] ;

[0105] ;

[0106] In the formula, P is the homology probability, and h is the index of the Gaussian component, h=1,2,3. Let h be the weight of the h-th Gaussian component. Let h be the mean vector of the h-th Gaussian component. Let h be the covariance matrix of the h-th Gaussian component. The absolute difference between the 12-dimensional feature vectors corresponding to target objects A and B. express The probability density value under the h-th Gaussian distribution, where the three Gaussian components correspond to... The sum is 1.

[0107] For a single product's text and image elements, based on The corresponding formula calculates all difference vectors. And construct a training sample set Then, the EM algorithm is used to train the model, and finally the parameter set is obtained. .

[0108] The trained model parameters are fed into a three-component Gaussian mixture model, and the homology probability between other target objects and target sample objects within the same defect group is calculated based on the three-component Gaussian mixture model. The absolute difference of the 12-dimensional feature vectors is then considered. In the formula, Features representing the target sample object. The characteristics of the c-th other target object in the group are represented. The homology probability P is then used to update the homology defect group. For a single other target object in the group, if its homology probability with the target sample object is greater than or equal to the preset probability threshold (e.g., set to 0.9), it is determined to be a homology defect. If its homology probability with the target sample object is less than the preset probability threshold, it is determined to be a non-homology defect and is removed from the group to obtain the final homology defect group.

[0109] It should be noted that using the EM algorithm (Expectation-Maximization Algorithm) to train a three-component Gaussian mixture model is a commonly used method in statistics and machine learning for fitting multimodal distributions. This involves training a known set of training samples. In this case, the method maximizes the expectation and determines the parameters through the following iterative logic: Based on the current model parameters (the initial model parameters are manually set initial values), the posterior probability of each sample belonging to one of the three Gaussian components—concurrent, continuous, or intermittent periods—is calculated, i.e., estimating which distribution produced each observation. The maximization step (M-step) uses the response calculated in the previous step as weights to recalculate and update the mean vector, covariance matrix, and component weights of each Gaussian component to improve the model's log-likelihood of the observed data. The above two steps are iteratively executed alternately, allowing the model parameters to gradually approach the local optimum. The resulting mixed probability distribution most realistically fits the process drift characteristics reflected in the sample set. This method effectively decomposes complex, multi-peaked feature difference data caused by printing time differences into three normally distributed components with clear physical meaning, thus providing an accurate mathematical basis for subsequent calculations of homogeneity probabilities.

[0110] Step 8: Select one more target object from the remaining target objects in the large multi-page printed image as the target sample object, and obtain the same defect group again in the manner of steps 5 to 7. Repeat this process until all target objects in the large multi-page printed image are grouped. Perform batch marking processing on each group of same defects. The marking results include qualified and unqualified.

[0111] Specifically, for a single group of defects originating from the same source, a unified label is assigned to all target objects (including product graphic elements and defects) within the group based on the coordinate position and defect type of the target sample object (i.e., pre-acquired during defect identification), such as a systematic ink spot defect at location (10,20). Then, manual review is performed, and each group of defects from the same source is batch-labeled according to the actual detection conditions set, such as... Figure 3 As shown.

[0112] It should be noted that the target objects removed from the group in steps 6 to 7 are returned to the remaining target objects to be selected; after the calculation of a single group of similar defects is completed, the target objects in the selected group of similar defects need to be excluded from the remaining target objects to be selected.

[0113] In this embodiment, the final obtained groups of defects from the same source can also be used for subsequent root cause tracing. The root causes of defects from the same source are usually related to the editing, modification, and imposition of single-mold small-format printing files, as well as the "repetitive working parts" of the printing equipment. By analyzing the continuity of the distribution of defects from the same source, the systematic causes of these defects can be inferred. For example, if in N-format printing, more than 80% of the opening sizes have defects at this position, it can be determined that there is a problem in the editing and modification process of the single-mold small-format printing files; if the Y-coordinate of the defect group is fixed (e.g., ... If the pressure is constant, it may be related to impurities or uneven pressure in the axial (Y direction) direction of the impression cylinder; if the absolute coordinates of the defect group corresponding to the large plate are... ,and The numbering increases linearly with the opening degree, such as This indicates that the defects appear periodically along the X direction, which may be related to the circumferential wear of the printing plate cylinder (the X direction is the direction of cylinder rotation). These tracing processes are completed manually, and the cause of the same source defect is obtained based on the root cause tracing results.

[0114] In this embodiment, the core of batch processing of defects in the same location based on the location index is to break the spatial isolation of different openings by using a unified coordinate system, and to identify systematic defects by using location association and feature clustering. Ultimately, it achieves an upgrade from "single-opening detection" to "batch processing of multiple openings". The method of this application greatly improves the quality control efficiency of multi-opening printing (such as reducing the detection time from O(N) to O(1), where N is the number of openings), and is especially suitable for multi-opening printing scenarios such as packaging and labels.

[0115] like Figure 4 As shown, this embodiment also provides a detection device for performing a method for identifying defects from the same source and batch processing of multi-page printed materials. The detection device includes: an image acquisition module, a display module, a control module, and a host computer.

[0116] The image acquisition module uses a high-resolution linear or area array sensor to acquire large-format, multi-page printed images of the product. The sensor supports dynamic adjustment of the shooting frequency and exposure time.

[0117] The control module serves as the hardware coordination hub of the device, precisely driving the image acquisition module to work according to the trigger commands issued by the host computer. At the same time, the control module also establishes communication with the printing equipment, enabling it to control the printing equipment to pause for subsequent correction when the host computer detects systemic defects, thus achieving closed-loop control by linking the printing equipment and the host computer.

[0118] The display module integrates a high-resolution display array and a human-computer interaction interface for real-time visualization of the current printed image (including overall and partial views), defect coordinate distribution map, etc. Through the interaction interface, users can manually trigger instructions for batch removal and batch marking of defects from the same source.

[0119] The host computer, as the core processing engine, is used to execute the above-mentioned defect homology identification algorithm process through its built-in high-speed processor, including data preparation (establishing a standard library), registration and detection, reverse localization (coordinate system establishment and indexing), batch processing, root cause tracing, etc. The host computer is also equipped with a storage module, which can use a large-capacity SSD array to store data such as images, feature data, and detection results.

[0120] Example 1:

[0121] Taking the processing of a certain type of printed matter as an example, its single-mold small-format standard image is as follows: Figure 6 As shown in the upper image, its large-format multi-page printed image is as follows: Figure 5 As shown, this is a 44-page print, comprising 44 small-format images arranged in an 11×4 matrix format. After coordinate registration using a three-level coordinate system, each single-format small-page image is aligned with the single-format small-page standard image, as shown below. Figure 6 As shown, to avoid involving specific product layout content and commercial copyright information, the single-mold small-format image and single-mold small-format standard image corresponding to a certain type of printed matter have been partially coded, retaining only the spatial location of the defect and the necessary local structural information around it, in order to illustrate the technical effect of the method of the present invention in coordinate registration and defect homology determination.

[0122] like Figure 7 As shown, during equipment operation, based on the established three-level coordinate system, the single-mode small-opening image and the single-mode small-opening standard image can be accurately indexed according to the specific coordinate position, and automatic positioning, magnified display and parallel comparison analysis of local areas are supported; after the coordinate registration and defect homology determination are completed by the method of this invention, multiple homology defect groups can be automatically output to realize cross-small-opening layout association clustering and batch determination.

[0123] The following explanation uses typical examples of local pattern misalignment defects (taking the hand as an example) and registration deviation defects; for example Figure 8 As shown, the image on the right is a magnified view of a single-mold small-scale image without local pattern offset defects, and its structural morphology is consistent with the corresponding structure in the standard single-mold small-scale image; the image on the left is a small-scale image with local pattern position offset. It can be observed that the "hand" pattern in the left image has a certain degree of displacement relative to the red background (which can be judged by the distance between it and the background boundary), but this offset is within the allowable positional deviation range and is not considered a defect in practice. Multiple single-mold small-scale images in this group of shared defects can be manually batch-marked as acceptable. Figure 9 The image shown is a magnified view of a small-scale single-mold image without any overprinting defects. Figure 10 to Figure 12 The image shown is a magnified view of a small-scale single-mold image containing a misregistration defect. It is clearly visible that... Figure 9 to Figure 12Gray, yellow, and red ink dots appeared around the hands in the image. These abnormalities exhibited spatial clustering and cross-page consistency (i.e., they were distributed in multiple small-format images, with consistent spatial positions and similar features). This invention grouped these defects into the same group of defects with the same origin. Tracing the source, it was found that these defects were all caused by overprinting deviations, which are typical batch defects with the same origin. For the multiple single-mode small-format images corresponding to this group of defects with the same origin, they can be batch-marked as unqualified, thereby avoiding manual judgment one by one. This application can significantly improve efficiency. Through the identification of the same origin and batch processing, the detection time is reduced from O(N) to O(1). Under the same detection task, the traditional method takes 20-30 minutes, while the technical solution of this application only takes 2-3 minutes, improving efficiency by 90%.

[0124] The steps in this invention can be adjusted, combined, or deleted according to actual needs.

[0125] The units in the device of the present invention can be merged, divided, or reduced according to actual needs.

[0126] In this invention, the terms "installation," "connection," "linking," and "fixing" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; "linking" can be a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of these terms in this invention according to the specific circumstances.

[0127] The shapes of the components in the accompanying drawings are schematic and may differ from their actual shapes. The drawings are only used to illustrate the principles of the present invention and are not intended to limit the present invention.

[0128] Although the invention has been disclosed in detail with reference to the accompanying drawings, it should be understood that these descriptions are merely exemplary and not intended to limit the application of the invention. The scope of protection of the invention is defined by the appended claims and may include various modifications, alterations, and equivalents made to the invention without departing from the scope and spirit of the invention.

Claims

1. A method for identifying and batch processing defects in multi-page printed materials, characterized in that, The method includes: Step 1: Create a standard single-mold small-format image of the target product and acquire a large-format multi-format printed image. Use the standard image to segment and extract the large-format multi-format printed image to obtain single-mold small-format images of multiple sizes to be tested. Establish a three-level coordinate system including the large-format absolute coordinate system, the size index, and the local coordinate system. Based on the three-level coordinate system, determine the coordinate registration relationship between the large-format multi-format printed image and each single-mold small-format image. Based on the coordinate registration relationship, determine the position of each target object in the large-format multi-format printed image in the local coordinate system. The target objects include product text and image elements and defects. Step 2: Select single-mode small-size image samples from the large-format multi-page printed images in layers according to the printing time sequence, obtain the multidimensional features of all target objects in the samples, including geometric features, color features, texture features and position features, remove abnormal samples whose multidimensional feature values ​​deviate from the template mean, and use the remaining valid samples to establish a statistical sample library. Step 3: For a single product text and image element, calculate the time series trend coefficient corresponding to each dimension feature based on the printing sequence number and multi-dimensional feature value of each valid sample in which it belongs, and use the multi-dimensional time series trend coefficient corresponding to the product text and image element to construct a normal process drift model. Step 4: Obtain the multidimensional features of all target objects in the large-format multi-page printing image, and compensate for the multidimensional features of all target objects based on the normal process drift model; Step 5: Select one target object from all target objects in the large multi-page printed image as the target sample object, search for other target objects in its neighborhood, and filter these target objects based on positional features to form an initial group of common defects. Quantify the feature discrimination based on the information entropy of each dimension feature within the group and determine the weight of each dimension feature. Step 6: Calculate the weighted total distance between other target objects and target sample objects in the initial same defect group based on the feature weights of each dimension, sort all weighted total distances in ascending order, determine the elbow point based on the difference of the adjacent distance, perform constraint verification on the elbow point distance, and update the same defect group based on the constraint verification results. Step 7: Establish a three-component Gaussian mixture model, train the mixture model using the multi-dimensional features of the product text and image elements in the effective samples, calculate the homology probability between other target objects and target sample objects in the same defect group using the trained model, and update the same defect group again. Step 8: Select target sample objects again from the remaining target objects of the multi-page printed image, and obtain the same defect group again in the manner of steps 5 to 7. Repeat this process until all target objects are grouped. Perform batch marking processing on each group of same defects. The marking results include qualified and unqualified.

2. The method for identifying and batch processing defects in multi-page printed materials as described in claim 1, characterized in that, Step 1 specifically includes: Based on the size of the single-mold small-format standard image and the position of the target product pixel area in it, the boundary of the single-mold small-format image corresponding to each target product pixel area in the large-format multi-format printing image is set so that each segmented single-mold small-format image has the same size and proportion as the single-mold small-format standard image, and the position of the target product pixel area in the image corresponds. A first reference point is selected for the large-format multi-page printing image. A large-format absolute coordinate system O-XY is established with the first reference point as the origin, where the X-axis is along the paper length direction and the Y-axis is along the paper width direction. The pixel coordinates in the large-format multi-page printing image are mapped to physical coordinates in the large-format absolute coordinate system O-XY. A second reference point is selected for each single-mold small-format image in the large-format multi-page printing image. A local coordinate system o-xy is established with the second reference point as the origin, where the X and Y axes are parallel to the corresponding X and Y axes of the large-format absolute coordinate system. Each single-mold small-format image is numbered, and the coordinates (X, Y, X, Y) of the second reference point corresponding to each single-mold small-format image in the large-format absolute coordinate system are recorded. n ,Y n The coordinate registration relationship between the large-format multi-page printed image and each single-format small-format image is obtained, expressed as: x n =XX n y n =YY n Where n is the aperture index of each single-mode small-format image in the large-format multi-page printing image, and (X,Y) are the physical coordinates of the pixels in the large-format multi-page printing image, (x,Y) n ,y n ) represents the physical coordinates of the pixel in the single-mode small-opening image of the nth opening degree.

3. The method for identifying and batch processing defects in multi-page printed materials as described in claim 1, characterized in that, Step 2 specifically includes: According to the printing time sequence, the single-mold small-format images in the large-sheet multi-page printing images are divided into three groups: early, middle and late. A predetermined number of single-mold small-format image samples are randomly selected from these three groups in equal proportions to form the initial statistical sample library. Any product text and image element is selected as the reference target, and the multidimensional feature values ​​of the reference target in all single-mold small-format image samples are obtained. For a single-dimensional feature, its feature mean in all single-mold small-format image samples is calculated as the template mean corresponding to that dimension feature. Then, single-mold small-format image samples that exceed the template mean ±10% range are removed. This process is repeated for each product text and image element, and the remaining valid samples are used to build a local feature statistical sample library.

4. The method for identifying and batch processing defects in multi-page printed materials as described in claim 3, characterized in that, Step 2 further includes: for a single target object, acquiring its geometric features, color features, texture features, and positional features, including: Geometric features include circularity C, edge smoothness E, and area S, where area S is the actual pixel area occupied by the target object, and circularity C and edge smoothness E are expressed as follows: ； In the formula, L is the actual perimeter of the target object. The ideal edge perimeter of the target object; Color characteristics include mean color saturation. R-channel mean G channel mean and B channel mean Texture features include GLCM energy (En) and LBP mean. ; Positional features include relative distance deviation ΔD and relative angle deviation Δθ. Distance deviation ΔD is the straight-line distance between the actual position of the target object and its corresponding template reference position in the single-mold small-scale standard image; relative angle deviation Δθ is the deviation between the actual angle of the target object and its corresponding template reference angle in the single-mold small-scale standard image.

5. The method for identifying and batch processing defects in multi-page printed materials as described in claim 4, characterized in that, Step 3 specifically includes: The valid samples in the statistical sample library are sorted according to the printing time order, and the printing sequence number corresponding to each valid sample is obtained. A single product text and image element is selected, and based on its printing sequence number 't' in each valid sample and its multi-dimensional feature values, the time series trend coefficient corresponding to each dimension feature is calculated using linear regression. The time series trend coefficient corresponding to a single dimension feature is expressed as: ； ； ； In the formula, Let j be the j-th dimension feature value of a single product image element. This represents the time series trend coefficient corresponding to the j-th dimension of the text and image elements of a single product. The intercept corresponding to the j-th dimension feature of a single product's text and image element. For a single product image element, the feature observation value in the i-th sample. This represents the average feature of a single product's text and image elements across all samples. This represents the printing sequence number of the i-th sample containing the text and image elements of a single product. The average value of the sample printing sequence number is used; the multi-dimensional time series trend coefficients corresponding to the text and image elements of the product are summarized to construct a normal process drift model.

6. The method for identifying and batch processing defects in multi-page printed materials as described in claim 5, characterized in that, Step 4 specifically includes: For a single-dimensional feature of a product's text and image element, the printing order correlation deviation is calculated based on its feature value in each sample. The calculation formula is as follows: ； In the formula, T[j] represents the printing order correlation deviation corresponding to the j-th dimension feature of a single product's text and image element, and its value ranges from [-1, 1]. The Pearson correlation coefficient ranges from -1 to 1. Let be the feature value of the j-th dimension feature of a single product text and image element in the i-th sample; The multidimensional features of all target objects are compensated based on the value of T[j]. For the j-th feature of a single target object, when T[j]>0, it means that its feature value is positively correlated with the printing sequence number; when T[j]<0, it means that its feature value is negatively correlated with the printing sequence number. The feature value is uniformly subtracted. Compensation will be performed, and the value after compensation will be... When T[j]=0, it means that the eigenvalues ​​have no obvious temporal drift and no compensation is needed.

7. The method for identifying and batch processing defects in multi-page printed materials as described in claim 4, characterized in that, Step 5 specifically includes: From all target objects in a large multi-page printed image, one target object is randomly selected as the target sample object. In the local coordinate system, all target objects that fall within its neighborhood and are located in other single-mode small-page images are searched. For all the target objects found, they are filtered based on the relative distance deviation ΔD and the relative angle deviation Δθ. The filtering process is as follows: for a single target object, if its ΔD value and Δθ value are both within the preset range, it is kept; if its ΔD value or Δθ value exceeds the preset range, it is removed. The filtered target objects are then classified into the initial homogeneous defect group.

8. The method for identifying and batch processing defects in multi-page printed materials as described in claim 7, characterized in that, Step 5 further includes: Read the geometric, color, texture, and positional features of all target objects in the initial group of shared defects. For a single-dimensional feature, calculate the printing order correlation deviation based on its feature values ​​on each target object, resulting in 11 dimensions of printing order correlation deviation. These 11 dimensions of printing order correlation deviation are used as process correlation features and categorized into multi-dimensional features to form 12-dimensional features. Read the 12-dimensional features of all target objects in the group. For the j-th dimension feature, use a positive transformation process to eliminate dimensional differences, expressed as: ； In the formula, For the j-th dimension feature of the r-th target object, After positive transformation value, , These are the minimum and maximum values ​​of the j-th feature within the group, respectively; based on the forward processing... The probability of the j-th feature on the r-th target object is calculated as follows: ； In the formula, Let R be the probability of the j-th feature on the r-th target object, and R be the number of target objects in the group; based on probability The information entropy of the j-th dimension feature is calculated as follows: ； In the formula, Let the information entropy of the j-th dimension feature be denoted as . , The constant coefficients, Based on information entropy Calculate the difference coefficient of the j-th dimension feature Then based on the difference coefficient The preliminary weights of the j-th feature are calculated and expressed as follows: ； In the formula, The initial weights are for the j-th feature; by iterating through the features of each dimension, the initial weights of the 12-dimensional features are obtained. The initial weights of the 12 features are optimized. If a feature is a core judgment indicator, its initial weight is increased by 10% to 20%. If a feature is a process-related indicator, its initial weight is increased by 5% to 10%. If a feature is affected by normal process fluctuations, its initial weight is decreased by 10% to 15%. The feature itself is not adjusted. The adjusted weights are then normalized to obtain the final optimized weights of the 12 features. Features belonging to core judgment indicators include relative distance deviation, roundness, and average color saturation. Features belonging to process-related indicators include printing sequence deviation. Features affected by normal process fluctuations include relative angle deviation and area.

9. The method for identifying and batch processing defects in multi-page printed materials as described in claim 8, characterized in that, Step 6 specifically includes: For any other target object within the group, the weighted total distance between it and the target sample object is calculated based on the optimized weights of the 12-dimensional features, and is expressed as: ； ； In the formula, Let F be the weighted total distance between the f-th other target object in the group and the target sample object, where F is the number of other target objects in the group besides the target sample object. For the optimized weights of the j-th dimension feature, Let f be the distance between the f-th other target object in the group and the target sample object on the j-th feature dimension. The weighted total distance of all other target objects within the group Sort the groups in ascending order and calculate the weighted total distance between adjacent groups. The difference , is represented as: ； In the formula, The s-th weighted total distance in the distance sequence; in the difference Find the maximum value in the sequence and will The corresponding position is denoted as the elbow point. ,Right now The distance corresponding to the elbow point As the initial constraint threshold ; Constraint verification of elbow distance: The IQR method is used to obtain verification conditions. Specifically, quartiles Q1 and Q3 are calculated, where Q1 is the lower quartile value and Q3 is the upper quartile value. The IQR value is calculated based on quartiles Q1 and Q3, expressed as IQR = Q3 - Q1. The upper limit value is calculated based on Q3 and the IQR value. =Q3 + 1.5 × IQR; based on the upper limit value Determine the initial constraint threshold Whether it is effective, if Then the initial constraint threshold Effective, for groups with a distance less than or equal to the elbow point. Other target objects are classified as homologous defects of the target sample objects, and those with a distance greater than the elbow point within the group are removed. Other target objects, if Then use the elbow point corresponding difference Perform mutation verification, if Then determine the initial constraint threshold. Effective, for groups with a distance less than or equal to the elbow point. Other target objects are classified as homologous defects of the target sample objects, and those with a distance greater than the elbow point within the group are removed. Other target objects, if Then determine the initial constraint threshold. Invalid, adjust it to 1, and classify all other target objects in the group as homologous defects of the target sample object.

10. The method for identifying and batch processing defects in multi-page printed materials as described in claim 8, characterized in that, Step 7 specifically includes: A three-component Gaussian mixture model is established, comprising three Gaussian components. These components correspond to samples from the same period, continuous periods, and interval periods, respectively. Each Gaussian component includes three parameters: a mean vector, a covariance matrix, and component weights. The three-component Gaussian mixture model is expressed as follows: ； ； In the formula, P is the homology probability, and h is the index of the Gaussian component, h=1,2,3. Let h be the weight of the h-th Gaussian component. Let h be the mean vector of the h-th Gaussian component. Let h be the covariance matrix of the h-th Gaussian component. The absolute difference between the 12-dimensional feature vectors. express The probability density value under the h-th Gaussian distribution; for a single product's text and image elements, calculate all difference vectors. And construct a training sample set Then, the model is trained using the training sample set to obtain the parameter set. ; The parameter set is fed into a three-component Gaussian mixture model, and the homology probability between other target objects and target sample objects within the updated homology defect group is calculated based on the three-component Gaussian mixture model. The absolute difference of the 12-dimensional feature vectors is also calculated. In the formula, Features representing the target sample object. The characteristics of the c-th other target object in the group are represented. The homology probability P is then used to update the homology defect group. For a single other target object in the group, if its homology probability with the target sample object is greater than or equal to the preset probability threshold, it is determined to be a homology defect. If its homology probability with the target sample object is less than the preset probability threshold, it is determined to be a non-homology defect and is removed from the group to obtain the final homology defect group.

Images