A multi-region color difference prediction method for packaging decoration prints

Abstract

The application discloses a multi-region color difference prediction method for packaging decoration prints and relates to the technical field of printing process control. The method comprises the following steps: constructing a region topology graph and generating an ink path space attenuation matrix; according to an ink superimposition layer level relationship, extracting spectral absorbance of ink of a bottom region, and calculating an optical coupling coefficient between superimposition layers; inputting initial chroma values of each region as node features into a graph convolution network; using the ink path space attenuation matrix and the optical coupling coefficient between superimposition layers, continuously modulating aggregation weights of the graph convolution network when messages are transmitted between adjacent nodes, and outputting a predicted color difference field of each region; according to the predicted color difference field, reversely solving a target compensation adjustment amount of each ink key, and sending the target compensation adjustment amount to an ink control system of a printing machine. The application realizes physical law driven prediction and closed-loop compensation of multi-region color difference, and improves the control precision of color consistency of prints.

Description

Technical Field

[0001] This invention relates to the field of printing process control technology, and more specifically to a multi-region color difference prediction method for packaging and decoration printing. Background Technology

[0002] In the production of packaging and decorative printing, color consistency is a core indicator for measuring product quality. Packaging printing typically includes multiple color-separated areas such as brand logo areas, main visual image areas, and solid background areas. Each area has different requirements for color accuracy, and color differences between areas can affect each other through ink path spatial coupling and ink overprinting effects. Traditional color difference control mainly relies on operator experience, manually adjusting the ink key opening after visual comparison or offline colorimeter sampling. This approach suffers from problems such as response lag, blind adjustment, and difficulty in achieving multi-area collaborative optimization.

[0003] Existing machine learning-based color difference prediction methods mainly utilize convolutional neural networks for image feature extraction and classification, or graph neural networks for connected component feature analysis and color group error propagation modeling, treating each color group as a graph node for message passing. However, these methods still have significant shortcomings when applied to color difference prediction in packaging and decoration printing: First, existing methods do not consider the physical attenuation relationship between ink bond spatial distribution and printing area, and the impact of ink bond adjustment on non-aligned areas cannot be accurately quantified; second, existing methods ignore the optical coupling effect during multi-layer ink overprinting, where the color difference of the underlying ink affects the surface color through transmission, but a model to quantitatively describe this transmission relationship is lacking; third, existing methods use a fixed graph topology for message passing, failing to dynamically adjust the aggregation weights between nodes based on the physical laws of ink path attenuation and overprinting coupling, resulting in limited prediction accuracy. Therefore, a multi-region color difference prediction method that can integrate the physical laws of ink path spatial attenuation and overprinting optical coupling is urgently needed. Summary of the Invention

[0004] This invention addresses the technical problems in existing technologies where multi-region color difference prediction for packaging and decoration printing does not consider the spatial attenuation of ink bonds and the optical coupling effect of overprinting, and where the aggregation weights of graph convolutional networks lack physical law-driven mechanisms. It provides a multi-region color difference prediction method for packaging and decoration printing.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: This invention provides a multi-region color difference prediction method for packaging and decoration printing, including: The prepress file is parsed to obtain the region topology map of the printed matter to be predicted. The nodes of the region topology map are the color separation regions on the printed matter, and the edges are the spatial adjacency relationships of adjacent regions. Obtain the current opening data of each ink key of the printing press, calculate the influence coefficient of each ink key fluctuation on the ink volume of each area based on the spatial mapping relationship between the ink key and each area, and generate the ink path spatial attenuation matrix. Based on the ink overprinting layer relationship between different areas in the prepress file, the spectral absorption rate of the ink in the bottom layer is extracted, and the optical coupling coefficient between the overprinting layers is calculated. The initial chromaticity values ​​of each region are used as node features input to the graph convolutional network. The aggregation weights of the graph convolutional network during message passing between adjacent nodes are continuously modulated using the ink path spatial attenuation matrix and the optical coupling coefficient between overprinting layers to output the predicted chromaticity field of each region. Based on the predicted color difference field, the target compensation adjustment amount of each ink key is calculated in reverse, and the target compensation adjustment amount is sent to the printing press ink control system.

[0006] The beneficial effects of this invention are: Compared to existing technologies, this invention first constructs a regional topology map by analyzing prepress files, transforming the multi-regional spatial structure of packaging and decoration printing materials into a graph structure, providing a structured data foundation for subsequent message passing in the graph convolutional network. Secondly, based on the spatial mapping relationship between ink keys and each region, and combined with ink roller travel and dot gain, the spatial diffusion scale of ink volume is determined, the ink volume influence coefficient is calculated, and an ink path spatial attenuation matrix is ​​generated. This allows the graph convolutional network to adaptively adjust weights based on the physical attenuation law of the ink path when aggregating neighbor node information. Thirdly, the spectral absorption rate of the bottom layer is extracted based on the ink overprinting layer relationship, and the optical coupling coefficient between overprinting layers is calculated, enabling the graph convolutional network to capture the optical mechanism of color difference transmission from the bottom layer to the surface layer. Finally, the predicted color difference field output by the graph convolutional network is inversely calculated into the target compensation adjustment amount for each ink key through a sensitivity matrix, realizing closed-loop control from color difference prediction to actuator compensation, solving the problem of prediction and execution disconnect in existing methods. Attached Figure Description

[0007] Figure 1 A flowchart illustrating a multi-region color difference prediction method for packaging and decoration printing provided by the present invention; Figure 2 A logical diagram illustrating a multi-region color difference prediction method for packaging and decoration printing provided by the present invention; Figure 3 This is a schematic diagram of the graph convolutional network module structure provided by the present invention. Detailed Implementation

[0008] Examples, such as Figure 1 , Figure 2 As shown, this embodiment of the invention provides a multi-region color difference prediction method for packaging and decoration printing, including: S10: Parse the prepress file to obtain the region topology map of the printed matter to be predicted. The nodes of the region topology map are the color separation regions on the printed matter, and the edges are the spatial adjacency relationships of adjacent regions. In packaging and decoration printing, prepress files contain crucial information such as the geometric outlines of each color separation region, layer stacking order, and color values. The spatial adjacency between different color separation regions determines the range of influence of ink path spatial attenuation and the propagation path of overprinting effects, and these spatial relationships are difficult to effectively express in traditional tabular data. Furthermore, each color separation region is not isolated; color differences between adjacent regions affect each other through ink path spatial attenuation. Specifically, a change in the opening of a certain ink key not only affects the area directly opposite it but also ripples to adjacent areas due to the axial movement of the ink roller and the lateral diffusion of ink. If each region is treated as an independent unit for color difference prediction, the coupling relationship between regions cannot be accurately depicted, leading to distorted prediction results.

[0009] Therefore, the first step is to transform each color separation region and its spatial adjacency into graph-structured data, where nodes represent each color separation region and edges represent the spatial adjacency of adjacent regions. By constructing a regional topology graph, structured node relationships and adjacency information can be provided for subsequent message passing in the graph convolutional network, enabling color difference information to propagate and aggregate between regions along the ink path spatial attenuation direction. The regional topology graph established in this step will serve as the common foundation for subsequent ink path spatial attenuation matrix mapping, overprinting optical coupling coefficient calculation, and graph convolutional network message passing.

[0010] Specifically, in this embodiment, the prepress file is parsed to obtain a region topology map of the printed matter to be predicted. The nodes of the region topology map are the color separation regions on the printed matter, and the edges are the spatial adjacency relationships of adjacent regions, including: Analyze the layer information and geometric contour data in the prepress file to extract the polygon boundaries of each color separation area; A two-dimensional coordinate system is established with the paper feed direction of the printing press and the axis of the cylinder as the reference, and the outer rectangle of the polygon of each color separation area is calculated. Traverse all bounding rectangles. When it is determined that the shortest distance between the bounding rectangles of two regions is less than the preset proximity tolerance, establish a spatial adjacency edge between the corresponding two regions. Construct a regional topology graph using each color-separated region as a node and spatially adjacent edges as edges.

[0011] First, the layer information and geometric contour data in the prepress file are parsed to extract the polygon boundaries of each color separation area. Prepress files are typically stored in PDF or EPS format, containing vector graphics data for each layer. Each layer corresponds to an ink color or an independent graphic element, and the vector graphics data is stored as path objects. For example, in an optional embodiment using PDF format, path objects in the page content stream can be read through a PDF parsing library. Each path object contains a series of path construction operators and corresponding coordinate parameters. Construction operators include moving to a specified starting point, connecting to the next point, and closing the path. When a path closure construction operator is encountered, all coordinate points along the path from the starting point to the ending point are extracted sequentially to form a closed polygon contour. Since a color separation area in an actual printed product may consist of multiple discrete closed paths—for example, text and graphics in a logo may be the same color but spatially separated—these discrete paths need to be merged when extracting the polygon boundaries. The merging rule is as follows: After extracting all closed paths, determine the shortest distance between the bounding rectangles of different paths. When the shortest distance is less than the preset merging tolerance, the corresponding multiple paths are merged into a polygon boundary of a single color-separated region. This is achieved by taking the union of the coordinate points of all paths and calculating their minimum convex or concave hull as the boundary of the merged polygon. The concave hull algorithm is preferred to preserve the concave features of the region boundary. When the shortest distance is greater than or equal to the preset merging tolerance, each path is processed as an independent color-separated region. The preset merging tolerance can be set according to the drafting precision of the prepress file, for example, it can be set to 0.2 mm.

[0012] The polygonal boundary of each color separation region is formed by connecting a series of vertex coordinates sequentially. Adjacent vertices are connected by straight line segments, each corresponding to an edge unit. Multiple edge units are connected to form a closed contour. The coordinates of each vertex are recorded in millimeters, describing the precise contour and position of the color separation region on the printed surface. Through the above extraction and merging operations, the complete polygonal boundary data of each color separation region can be obtained. Specifically, the polygonal boundary describes the precise contour and position of the color separation region on the printed surface.

[0013] Secondly, a two-dimensional coordinate system is established based on the paper feed direction and the cylinder axis to calculate the bounding rectangle of each color separation region polygon. The paper feed direction refers to the direction in which the paper moves during printing, and the cylinder axis refers to the axial direction of the printing cylinder; the two are perpendicular to each other. A two-dimensional coordinate system is established with the paper feed direction as the Y-axis and the cylinder axis as the X-axis. The bounding rectangle is the smallest rectangle that completely contains the polygon, with each side parallel to the coordinate axis. For example, in specific execution, the bounding rectangle of region i can be set to R. i =[x mini ,x max i ,y min i ,y max i ], where x min i and x max i These are the minimum and maximum coordinates of the region along the roller axis, y and y', respectively. min i and y max i These are the minimum and maximum coordinates of the region in the paper feed direction, respectively. The enclosing rectangle simplifies the determination of spatial relationships between polygons, reducing computational complexity.

[0014] Next, all bounding rectangles are traversed. When the shortest distance between the bounding rectangles of two regions is determined to be less than the preset proximity tolerance, a spatial adjacency edge is established between the corresponding two regions. The shortest distance between two rectangles refers to the minimum distance between any two points on the boundaries of the two rectangles.

[0015] For example, in actual execution, let the outer rectangles of the two regions be R. i and R j The overlap interval between the two along the x-axis is Δx = max(0, max(x)). min i ,x min j )-min(x max i ,x max j The overlap interval along the y-axis is Δy = max(0, max(y)). min i ,y min j )-min(y max i ,y max j If the shortest distance between the two bounding rectangles is given, then the shortest distance between them is given. .in, Let Δx be the distance between two points i and j, Δy be the difference in their x-coordinates, and Δy be the difference in their y-coordinates.

[0016] Specifically, when D ijWhen the distance is less than the preset proximity tolerance, two regions are determined to be spatially adjacent, and an adjacent edge is established. The preset proximity tolerance is set based on the following: During the pre-press production of packaging and decoration printing documents, there are often tiny gaps between different color separation regions due to drafting accuracy or design requirements. These gaps are typically between 0.1 mm and 0.3 mm. Simultaneously, mechanical positioning errors in the paper feed direction and roller axis during printing also cause slight offsets in region boundaries, generally within 0.2 mm. If the tolerance is set too small, for example, less than 0.1 mm, the aforementioned drafting gaps and mechanical positioning errors will cause regions that should be adjacent to be judged as non-adjacent. The missing edges in the region topology graph will prevent the graph convolutional network from transmitting color difference information between adjacent regions, reducing prediction accuracy. If the tolerance is set too large, for example, greater than 1 mm, regions that are spatially far apart and do not actually have ink path coupling relationships will be incorrectly judged as adjacent, introducing unnecessary edge connections. This causes the graph convolutional network to aggregate information from irrelevant regions during message transmission, thus introducing noise. In this embodiment, the preset proximity tolerance is set to 0.5 mm. This tolerance value is greater than typical drafting gaps and mechanical positioning errors, but less than significant spatial distances. It effectively identifies all areas where color difference coupling occurs due to ink path spatial attenuation, while avoiding interference from irrelevant areas. This value can be adjusted according to the actual precision requirements of the printed material. For high-precision packaging printing, it can be appropriately reduced to 0.3 mm; for ordinary printing, it can be appropriately increased to 0.8 mm. Through this tolerance setting, it can be ensured that the edge connection relationships of the region topology map accurately reflect the ink path coupling relationships between different regions of the printed material.

[0017] Specifically, a preset proximity tolerance is used to determine whether two regions are spatially close to each other. Even if there may be a tiny gap between them, as long as the gap width is less than the preset proximity tolerance, they are still considered adjacent. After establishing adjacent edges, the color difference information between two regions can be propagated through the message passing of the graph convolutional network via the edges.

[0018] Finally, a regional topology graph is constructed using each color separation region as a node and spatial adjacency edges as edges. In the regional topology graph, the set of nodes corresponds to all color separation regions, and the set of edges corresponds to all spatial adjacency relationships. Through this regional topology graph, the spatial positional relationships between each color separation region on the printed material are quantified into graph structure data, enabling subsequent graph convolutional networks to pass messages between adjacent regions, achieving spatial propagation and collaborative prediction of color difference information.

[0019] S20: Obtain the current opening data of each ink key of the printing press, calculate the influence coefficient of each ink key fluctuation on the ink volume of each area according to the spatial mapping relationship between the ink key and each area, and generate the ink path spatial attenuation matrix. Secondly, after obtaining the regional topology map, it is necessary to establish the physical mapping relationship between the ink key control space and the printing area. The ink keys of the printing press are arranged along the cylinder axis, with each ink key corresponding to a fixed control position axially. When the opening of an ink key changes, the change in ink output by that ink key will first affect the printing area directly opposite it. However, due to the axial reciprocating movement of the ink roller during printing, and the lateral diffusion of ink on the roller surface, the ink volume impact caused by the change in ink key opening is not strictly limited to the directly opposite area, but extends axially to both sides, causing adjacent areas to also experience a certain degree of ink volume disturbance. The degree of ink volume impact decreases with increasing axial distance from the ink key control center; areas closer to the center are more affected, and areas farther away are less affected. This spatial attenuation effect causes the contribution of any ink key fluctuation to the ink volume of all areas across the entire printing press to exhibit a continuously decaying distribution with a peak value at the ink key control center.

[0020] To accurately describe this physical process, this step establishes a one-dimensional spatial coordinate system, calculates the axial distance between the control center of each ink key and the geometric center of each color separation region, determines the spatial diffusion scale of ink volume by combining the ink roller traverse stroke and dot gain rate, and calculates the influence coefficient of each ink key on the ink volume of each region using a Gaussian decay function, generating an ink path spatial decay matrix. This ink path spatial decay matrix will serve as an important physical basis for the dynamic edge weight calculation in the subsequent graph convolutional network message passing process, enabling the network to transmit color difference information between adjacent nodes based on the actual ink path decay law. Simultaneously, the introduction of this ink path spatial decay matrix can solve the problem in traditional methods where the influence of ink key adjustment on non-aligned regions cannot be accurately quantified.

[0021] Specifically, in this embodiment, the current opening data of each ink key of the printing press is obtained, and the influence coefficient of ink volume of each ink key fluctuation on each region is calculated based on the spatial mapping relationship between the ink keys and each region, generating an ink path spatial attenuation matrix, including: A one-dimensional spatial coordinate system is established along the axial direction of the printing press cylinder to obtain the control center coordinates of each ink key and the geometric center coordinates of each color separation area. Calculate the axial distance from the control center of any ink key to the geometric center of any color separation region; Obtain the traverse stroke parameters of the ink rollers of the printing press and the dot gain of the current printed paper. Based on the traverse stroke parameters and the dot gain, determine the spatial diffusion scale of ink volume. Based on the correlation between axial distance and ink volume spatial diffusion scale, an index value characterizing the degree of ink volume attenuation is calculated as the ink bond's influence coefficient on the ink volume of the corresponding color separation region. Summarize the ink volume influence coefficients of all ink keys and all color separation regions, and construct the ink path space attenuation matrix.

[0022] First, a one-dimensional coordinate system is established along the axial direction of the printing press cylinder to obtain the control center coordinates of each ink key and the geometric center coordinates of each color separation region. The ink keys of the printing press are uniformly arranged along the cylinder axis, and each ink key corresponds to a fixed control position along the axis. The axial coordinate value of this position is used as the control center coordinate of the ink key. The geometric center coordinates of the color separation region are obtained by taking the arithmetic mean of the x-coordinates of all vertices of the polygon in that region; that is, the axial coordinate of the region's geometric center is the sum of the x-coordinates of all vertices of the polygon divided by the number of vertices. This geometric center coordinate reflects the average position of the color separation region along the axial direction.

[0023] Next, calculate the axial distance from the control center of any ink key to the geometric center of any color separation area. For example, in actual execution, let the axial coordinate of the control center of ink key k be X. k The geometric center axial coordinate of color separation region i is X. i Then the axial distance This axial distance is used to quantify the spatial interval between the ink key and the area in the axial direction, and is the basic variable for calculating the ink volume influence coefficient.

[0024] Furthermore, the traverse stroke parameters of the printing press's ink rollers and the dot gain of the current printed paper are obtained. Based on these parameters, the spatial diffusion scale of ink volume is determined. The ink roller traverse stroke refers to the amplitude of the ink roller's axial reciprocating motion. This parameter is determined by the printing press's mechanical structure and is measured in millimeters. The traverse stroke varies between different printing press models, typically ranging from 10 mm to 30 mm for commonly used presses. A larger traverse stroke results in a wider axial movement range for the ink roller and a more uniform axial ink diffusion distribution; a smaller traverse stroke makes the ink roller more sensitive to axial position, and the ink distribution is more concentrated in the area directly opposite the ink keys. Specifically, this traverse stroke parameter can be found in the printing press's technical manual or obtained by measuring the axial displacement of the ink roller's end.

[0025] Dot gain is the percentage difference between the actual dot area on the printed material and the corresponding dot area on the printing plate, minus 1. The causes of dot gain include: the mechanical expansion of ink towards the dot edges under pressure, the penetration and diffusion of ink through the capillary action of the paper, and uneven ink absorption due to paper surface roughness. Dot gain varies depending on paper type, ink properties, and printing pressure. For example, the dot gain of offset paper is typically between 12% and 18%, coated paper between 8% and 12%, and newsprint between 20% and 30%. A higher dot gain indicates more significant ink diffusion on the paper and a wider lateral transmission range of ink in the axial direction; a lower dot gain indicates a more concentrated spatial distribution of ink. Specifically, dot gain can be obtained by measuring the density of printed test samples. The specific method is as follows: print test samples under stable printing press conditions, use a densitometer to measure the solid density and dot area coverage of each tone dot area on the sample, read the dot gain value through the built-in calculation function of the densitometer, and take the average value of the measurement results of multiple test points, such as 10 points, as the dot gain of the current printing paper.

[0026] In this embodiment, the formula for calculating the ink volume spatial diffusion scale is: Ink volume spatial diffusion scale σ = traverse stroke parameter × (dot gain rate / reference dot gain rate) × preset proportional coefficient. The reference dot gain rate can be set according to the standard test conditions of the printing press, for example, it can be set to 15%. When the dot gain rate is greater than the reference value, the ink volume spatial diffusion scale increases, and the axial attenuation of ink volume slows down; when the dot gain rate is less than the reference value, the ink volume spatial diffusion scale decreases, and the axial attenuation of ink volume accelerates. Through this calculation method, the ink volume spatial diffusion scale can comprehensively reflect the combined influence of the mechanical traverse of the ink roller and the physical diffusion of ink on the spatial distribution of ink volume.

[0027] Specifically, the physical meaning of this formula can be explained as follows: the spatial diffusion scale of ink volume is directly proportional to the travel distance of the ink roller. The greater the travel distance, the wider the range of axial movement of the ink roller, and the greater the axial diffusion range of ink volume. The spatial diffusion scale of ink volume is also directly proportional to the dot gain. The higher the dot gain, the more significant the lateral diffusion of ink on the paper, and the wider the spatial distribution range of ink volume. Normalizing by dividing the dot gain by a reference dot gain eliminates the influence of different printing conditions on the absolute value of the dot gain, allowing the formula to adapt to variations in paper type and ink properties. A preset proportionality coefficient is used to convert the combined effect of travel distance and dot gain to the same millimeter scale as the axial distance. Its value can be determined experimentally, for example, by fitting the calculated value to the range of influence of ink bond fluctuations in actual printing. Through the above multiplicative relationship, the spatial diffusion scale of ink volume can adaptively adjust with changes in printing conditions, making the subsequent calculation of the ink volume influence coefficient closer to the actual physical process.

[0028] In the above formula, the preset scaling factor is used to convert the combined effect of traverse travel and dot gain into a spatial diffusion scale value of ink volume that matches the actual printing conditions. It represents a comprehensive scaling factor for the contribution of traverse travel and dot gain to ink volume diffusion. This preset scaling factor is determined through experimental calibration based on the printing press model and the surface characteristics of the ink rollers; for example, it can be set to 0.5. When the surface roughness of the ink roller is large, the adhesion and lateral transfer ability of ink on the ink roller surface are enhanced, and the ink diffusion range expands, so the preset scaling factor should be appropriately increased. When the surface smoothness of the ink roller is high, the axial diffusion of ink is suppressed, so the preset scaling factor should be appropriately decreased. For example, for a conventional printing press equipped with rubber ink rollers, the preset scaling factor can be set to 0.5; for a high-speed printing press equipped with metal ink rollers, the preset scaling factor can be set to 0.3. In practical applications, the preset scaling factor can be fine-tuned based on the ink distribution measurement results of printed test samples.

[0029] Next, based on the correlation between axial distance and the spatial diffusion scale of ink volume, an exponential value characterizing the degree of ink volume attenuation is calculated, serving as the ink volume influence coefficient of the inking bond on the corresponding color separation region. Specifically, let the spatial diffusion scale of ink volume be σ, where σ is determined by the ink roller travel distance and dot gain rate, then the ink volume influence coefficient... When the axial distance is zero, the ink volume influence coefficient is 1, indicating that the ink key has the greatest influence on the ink volume in its directly opposite area. As the axial distance increases, the ink volume influence coefficient decays according to a Gaussian curve, with the decay becoming more significant as the axial distance increases. When the axial distance is much larger than the spatial diffusion scale of ink volume, the ink volume influence coefficient approaches zero. This Gaussian decay function can accurately describe the physical law that the influence of ink volume caused by ink roller movement and ink diffusion decreases with increasing axial distance.

[0030] Finally, the ink volume influence coefficients of all ink keys and all color separation regions are summarized to construct the ink path space attenuation matrix. Specifically, assuming there are K ink keys and N color separation regions, the ink path space attenuation matrix W is a K-row N-column matrix, i.e. The element in the k-th row and i-th column of the matrix is ​​the influence coefficient of ink key k on the ink amount of region i. Each column of the matrix corresponds to a color separation region and records the distribution of the ink amount influence of all ink keys on that region; each row corresponds to an ink key and records the distribution of the ink amount influence of that ink key on all regions.

[0031] The resulting ink path spatial attenuation matrix will serve as an important physical basis for the dynamic edge weight calculation during subsequent graph convolutional network message passing, enabling the graph convolutional network to transmit color difference information between adjacent nodes based on the actual spatial mapping relationship between ink keys and regions.

[0032] S30: Based on the ink overprinting layer relationship between different areas in the prepress file, extract the spectral absorption rate of the ink in the bottom layer and calculate the optical coupling coefficient between the overprinting layers. Furthermore, in packaging and decoration printing, multiple color separation areas may have an overprinting relationship in space. For example, in the brand logo area, text or patterns on the surface may be overprinted on top of the underlying spot color ink. When the chromaticity value of the underlying ink shifts due to ink volume fluctuations, this shift is transmitted to the surface layer through ink layer transmission, causing a corresponding change in the surface color. This optical coupling effect means that the color difference of the underlying layer does not exist independently, but is partially superimposed on the color difference of the surface layer. If this interlayer coupling is ignored, the color difference prediction of the surface area will miss the contribution of the color difference from the underlying layer, thus causing the prediction result to deviate from the actual printing effect.

[0033] Therefore, this step identifies the underlying and surface regions with overprinting relationships by analyzing the layer stacking order in the prepress file. Based on the ink spectral database, it extracts the spectral absorption characteristics of the underlying ink in the visible light band, calculates the spectral change of color difference transmitted from the underlying layer to the surface layer, and then determines the optical coupling coefficient between the overprinting layers. This optical coupling coefficient will be used to modulate the aggregation weights between adjacent nodes in the subsequent message passing of the graph convolutional network, enabling the color difference information between the underlying and surface regions to propagate and aggregate according to the actual optical coupling rules.

[0034] Specifically, in this embodiment, based on the ink overprinting layer relationship between different areas in the pre-press file, the spectral absorption rate of the ink in the bottom layer is extracted, and the optical coupling coefficient between the overprinting layers is calculated, including: Analyze the layer stacking order in prepress files to identify the bottom and top layers with overlapping relationships; Based on the ink spectral database, the spectral absorption and scattering characteristics of the underlying ink in the visible light band are extracted; Based on the spectral absorption characteristics, the amount of spectral change transmitted to the surface ink is calculated when the chromaticity value of the bottom ink shifts. The spectral variation is integrated and mapped in the visible light band to convert it into the chromaticity shift of the surface color. The ratio of the chromaticity shift of the surface color to the chromaticity shift of the underlying ink is used as the optical coupling coefficient between the overprinting layers. The optical coupling coefficient between the overprinting layers is used to characterize the proportion of color difference transferred from the underlying layer to the surface layer.

[0035] First, the layer stacking order in the prepress file is analyzed to identify the bottom and top layers with overprinting relationships. The layers in the prepress file are arranged from top to bottom according to the printing order, with the first printed layer at the bottom and the later printed layers covering the top. By analyzing the layer order, pairs of layers that overlap spatially and have a top-bottom covering relationship can be identified, where the layer at the bottom is the bottom layer area, and the layer at the top is the top layer area. The color shift of the bottom ink will affect the color of the top ink through ink layer transmission; therefore, it is necessary to extract the overprinting relationship between the bottom and top areas.

[0036] Next, based on the ink spectral database, the spectral absorption and scattering characteristics of the underlying inks in the visible light band are extracted. The ink spectral database is a structured dataset that stores spectral data categorized by ink brand, model, and color number, with each type of ink corresponding to a complete spectral characteristic curve. The spectral characteristics are measured using a spectrophotometer, measuring the ink samples wavelength by wavelength within the visible light band to obtain the spectral absorbance and scattering coefficient at each wavelength. Common measurement wavelengths include 380nm and 390nm, with adjacent wavelengths typically spaced 10nm or 5nm apart. The measurement data for each type of ink is obtained through multiple measurements, such as 10 measurements, and the average value is stored in the database to form the spectral feature vector for that ink, where each element corresponds to the spectral absorbance or scattering coefficient value at a specific wavelength.

[0037] By querying an ink spectral database, using the brand, model, and color number of the underlying ink as an index in the pre-press file, the database is used to retrieve the corresponding ink's spectral absorption vector and scattering coefficient vector. If no perfectly matching ink model exists in the database, the ink model with the closest spectral characteristics is selected as an approximation, or the spectral characteristics of the target ink are estimated through interpolation. These spectral absorption and scattering characteristics will serve as the basic input data for subsequent calculations of the spectral changes in color difference transmitted from the underlying layer to the surface layer. Spectral absorption rate characterizes the degree to which the ink absorbs incident light energy, and scattering coefficient characterizes the ink's ability to scatter incident light.

[0038] Furthermore, based on the spectral absorption characteristics, the spectral change transmitted to the surface ink is calculated when the chromaticity value of the underlying ink shifts. Specifically, based on the Kubelka-Munk theory, the optical behavior of the ink layer can be described by the ratio of the absorption coefficient to the scattering coefficient. When the chromaticity of the underlying ink shifts, its spectral reflectance changes by an amount ΔR. base (λ). A portion of this change in light energy penetrates the underlying ink layer and is transmitted to the surface layer; the transmission ratio depends on the spectral absorption rate of the underlying ink.

[0039] Specifically, let the spectral absorption rate of the bottom ink be A. base (λ), the change in spectral reflectance caused by the underlying chromaticity shift is ΔR. base (λ), then the spectral change transmitted to the surface is calculated according to the following formula: .in, Let λ represent the spectral change transmitted to the surface, and λ be the spectral wavelength. The physical meaning of this formula is: the higher the spectral absorptivity, the more light energy is absorbed by the underlying layer, and the lower the proportion transmitted to the surface; conversely, the lower the spectral absorptivity, the higher the transmission ratio.

[0040] Next, the spectral change is integrated and mapped within the visible light band to convert it into a chromaticity shift of the surface color. The CIE chromaticity system defines a tristimulus matching function for a standard observer. , , These correspond to the sensitivity curves of the human eye to the three primary colors: red, green, and blue. The offset of the surface color from the tristimulus values ​​is: Similarly, ΔY can be calculated. top ΔZ top .in, This represents the spectral change transmitted to the surface. Here, λ represents the CIE standard chromaticity observer color matching function, and λ is the spectral wavelength. Then, the tristimulus value offsets are converted to chromaticity offsets ΔC in the CIE Lab color space. Top ΔC Top For ΔL * ,Δa * ,Δb * The vector magnitude.

[0041] Finally, the ratio of the chromaticity shift of the surface color to the chromaticity shift of the underlying ink is used as the optical coupling coefficient between the overprinting layers. .in, This represents the surface chromaticity offset in Lab color space. This represents the bottom layer chromaticity offset in Lab space. The calculation method for the bottom layer ink chromaticity offset is the same as that for the surface layer, obtained by applying the same CIE integral mapping and Lab transformation to the change in bottom layer spectral reflectance. This ratio reflects the proportion of color difference transfer from the bottom layer to the surface layer, and its value ranges from 0 to 1. When the spectral absorbance of the bottom layer ink is low, the transmittance is high, the optical coupling coefficient between the overprinting layers is large, and the influence of the bottom layer color difference on the surface layer color is significant. When the spectral absorbance of the bottom layer ink is high, the transmittance is low, the optical coupling coefficient between the overprinting layers is small, and the influence of the bottom layer color difference on the surface layer color is weakened. This optical coupling coefficient will be used to modulate the aggregation weights between the bottom layer and the surface layer regions in the subsequent message passing of the graph convolutional network.

[0042] S40: The initial chromaticity values ​​of each region are used as node features to input the graph convolutional network. The aggregation weights of the graph convolutional network during message passing between adjacent nodes are continuously modulated using the ink path spatial attenuation matrix and the optical coupling coefficient between overprinting layers to output the predicted chromaticity field of each region. Furthermore, after completing the construction of the regional topology map, the generation of the ink path spatial attenuation matrix, and the calculation of the optical coupling coefficient between overprinting layers, it is necessary to integrate the aforementioned physical information into the inference process of the graph convolutional network. The core mechanism of the graph convolutional network is message passing between adjacent nodes, that is, each node updates its own feature representation by exchanging feature information with its neighboring nodes. In standard graph convolutional networks, the aggregation weights between nodes are usually determined by a fixed symmetric adjacency matrix, which cannot reflect physical laws such as ink key spatial attenuation and overprinting optical coupling. If a standard graph convolutional network is directly applied for color difference prediction, the network cannot distinguish the actual physical influence of different neighboring nodes on the target node when aggregating neighboring node information, resulting in a mismatch between the propagation path of color difference information and the real physical process.

[0043] This step uses the initial chromaticity values ​​of each color-separated region as node features input to the graph convolutional network. In each message-passing iteration of the network, the ink amount influence coefficient between adjacent nodes is extracted from the ink path spatial attenuation matrix, and the optical coupling coefficient between adjacent nodes is extracted from the inter-layer optical coupling coefficient. After normalization, these two coefficients are multiplied to obtain dynamic edge weights, which replace the fixed symmetric adjacency matrix in the graph convolutional network. This allows the network to adaptively adjust the weights based on the actual physical attenuation law when aggregating neighbor node features. After multiple message-passing iterations, the graph convolutional network outputs the predicted color difference field for each region. This predicted color difference field integrates the physical mechanisms of ink path spatial attenuation and inter-layer optical coupling, providing accurate input for the subsequent inverse calculation of ink key compensation.

[0044] Among them, such as Figure 3 As shown, in this embodiment, the construction and training steps of the graph convolutional network include: Obtain a historical printing sample set, which includes the measured chromaticity values ​​of each color separation region and the corresponding standard chromaticity values, as training samples. The graph structure is constructed by using nodes of the regional topology graph to correspond to each color-coded region and edges to correspond to the spatial adjacency relationships of adjacent regions. The deviation between the measured chromaticity value and the standard chromaticity value of each node is used as the initial feature of the node, and the adjacency matrix of the region topology graph is used as the initial edge weight to construct the initial graph convolutional network. The training samples are input into the initial image convolutional network for forward propagation, and the difference between the predicted color difference and the measured color difference output by the network is calculated as the training loss. The network parameters of the graph convolutional network are updated by backpropagation based on the training loss. The training is iterated until the training loss converges, and the trained graph convolutional network is obtained.

[0045] The first step is to obtain a historical printing sample set, which includes the measured chromaticity values ​​and corresponding standard chromaticity values ​​for each color separation region, serving as training samples. This historical sample set originates from printing data accumulated from multiple batches during normal production. Each batch of printed materials is measured using an online or offline colorimeter to obtain the measured chromaticity values ​​for each color separation region. These measured chromaticity values ​​are represented by the L*, a*, and b* components of the CIE Lab color space. The standard chromaticity value is the target chromaticity value for that region in the prepress file, also represented by L*, a*, and b*. The difference between the measured chromaticity value and the standard chromaticity value is the color difference for that region, used to characterize the degree of deviation between the printed color and the target color. Each training sample set contains the measured chromaticity values ​​and corresponding standard chromaticity values ​​for each color separation region in one batch of printed materials.

[0046] The second step involves constructing a graph structure, where nodes in the regional topology graph correspond to each color-coded region, and edges correspond to the spatial adjacency relationships between adjacent regions. The number of nodes in the graph structure equals the total number of color-coded regions, and the number of edges equals the total number of adjacency relationships between adjacent regions. This graph structure is topologically consistent with the regional topology graph constructed in step S10 above, and serves as the data structure for graph convolutional networks.

[0047] The third step involves constructing an initial graph convolutional network using the deviation between the measured chromaticity values ​​and the standard chromaticity values ​​of each node as the initial node features and the adjacency matrix of the region topology graph as the initial edge weights. The initial node features are three-dimensional vectors, corresponding to the chromaticity difference values ​​of the L*, a*, and b* channels, respectively. ,in , Δa i and Δb i Similarly, the adjacency matrix A is an N×N square matrix, where N is the total number of nodes. When there is a spatial adjacency edge between node i and node j, A... ij =1, otherwise A ij=0. Graph convolutional networks employ a multi-layered structure, with each layer performing message passing and feature transformation operations: H (l) Let W be the feature matrix of the nodes in the l-th layer. (l) Let σ be the learnable weight matrix for the l-th layer, and σ be the activation function, such as ReLU. The initial number of layers in the graphical convolutional network can be set to 2 or 3.

[0048] The fourth step involves inputting the training samples into the initial image convolutional network for forward propagation, calculating the difference between the network's predicted color difference and the measured color difference, and using this difference as the training loss. For each batch of training samples, the network outputs the predicted color difference for each node after forward propagation. Actual color difference label The training loss uses mean squared error: , where ||·|2 is the Euclidean distance, which is the square root of the sum of the squares of the differences between the predicted and measured color difference vectors. This loss function measures the degree of deviation between the network output and the true label; the smaller the loss, the more accurate the network prediction.

[0049] The fifth step involves updating the network parameters of the graph convolutional network through backpropagation based on the training loss, iterating until the training loss converges, thus obtaining the trained graph convolutional network. Preferably, the Adam optimizer is used for backpropagation, with an initial learning rate set to 0.001, momentum parameters β1 set to 0.9, and β2 set to 0.999. In each training epoch, all or part of the training samples are used for forward and backpropagation to update the network weights W. (l) After each training round, the loss value on the validation set is calculated. Training stops when the validation set loss no longer decreases over several consecutive rounds. The model parameters with the minimum validation set loss are saved, resulting in the trained graph convolutional network. The number of early stopping rounds can be set based on the loss convergence characteristics during model training. For example, it can be set to 20 consecutive training rounds. That is, when the validation set loss does not fall below the previously recorded minimum validation loss for 20 consecutive training rounds, the model is considered converged, training stops, and the model parameters with the minimum validation set loss are saved, resulting in the trained graph convolutional network. This trained graph convolutional network can receive new node feature inputs during subsequent inference and output predicted color difference fields for each region.

[0050] Furthermore, in this embodiment, the initial chromaticity values ​​of each region are used as node features input to the graph convolutional network. Utilizing the ink path spatial attenuation matrix and the interlayer optical coupling coefficient, the aggregation weights of the graph convolutional network during message passing between adjacent nodes are continuously modulated to output the predicted chromatic difference field for each region, including: In each message passing iteration of the graph convolutional network, the ink volume influence coefficient and the optical coupling coefficient between the overprinting layers corresponding to the two adjacent nodes are extracted. After normalizing the ink volume influence coefficient and the optical coupling coefficient between overprinting layers, the product is calculated and used as the dynamic edge weight between adjacent nodes. By replacing the fixed symmetric adjacency matrix in the graph convolutional network with dynamic edge weights, the feature vectors of neighboring nodes are weighted, summed, and aggregated. After multiple layers of message passing iterations, the output is a predicted chromatic difference field for each node that integrates the spatial attenuation of ink path and the optical coupling relationship of overprinting.

[0051] Specifically, during each message-passing iteration of the graph convolutional network, the ink volume influence coefficient and the interlayer optical coupling coefficient between adjacent nodes are extracted. For any pair of adjacent nodes i and j, the ink volume influence coefficient W is extracted from the ink path spatial attenuation matrix. ij This coefficient reflects the degree to which changes in ink bond aperture affect the amount of ink produced jointly by adjacent regions i and j. γ is extracted from the optical coupling coefficient between the overprinting layers. ij This coefficient reflects the proportion of ink overprint optical coupling transmission between region i and region j. The ink volume influence coefficient and the optical coupling coefficient quantify the correlation strength between adjacent nodes from two dimensions: physical attenuation in ink path space and optical coupling of overprinting, respectively. The two are physically independent of each other and jointly determine the actual degree of propagation of color difference information between nodes.

[0052] Next, the ink volume influence coefficient and the optical coupling coefficient between overprinting layers are normalized and then multiplied. The resulting product is used as the dynamic edge weight between adjacent nodes. The specific normalization process is as follows: for all neighboring nodes k of the target node i, the ink volume influence coefficient W is calculated respectively. ik and optical coupling coefficient γ ik The relative proportion of this node among all its neighbors, i.e., W ik Divided by the sum of the W values ​​of all neighbors, γ ik Divide by the sum of the γ values ​​of all neighbors. Then take the product of the two normalized values ​​as the dynamic edge weight A. ij Here, all neighboring nodes k of target node i refer to all nodes in the regional topology graph that have spatially adjacent edges to node i, i.e., all nodes connected by the adjacent edges established by the shortest distance determination of the outer rectangle in step S10 above. Neighboring nodes are spatially adjacent to target node i, constituting the propagation objects of ink path spatial attenuation effect and ink overprinting optical coupling effect.

[0053] The physical meaning of this product method is as follows: when both the ink volume influence coefficient and the optical coupling coefficient are large, the propagation capability of color difference information between the two regions is strong, and the dynamic edge weight is large; when either coefficient is small, the dynamic edge weight decreases accordingly, and the propagation of color difference information is suppressed. This dynamic edge weight comprehensively reflects the synergistic effect of ink path spatial attenuation and overprinting optical coupling.

[0054] Furthermore, dynamic edge weights are used to replace the fixed symmetric adjacency matrix in the graph convolutional network, and the feature vectors of neighboring nodes are weighted and aggregated. In the standard message passing of the graph convolutional network, the new feature of node i is obtained by weighting and summing the features of all neighboring nodes j, with the weights determined by the dynamic edge weights A. ij Decision. The aggregation process can be represented as: H j (l) Let W be the feature vector of node j in the l-th layer. (l) Let B be the learnable weight matrix of the l-th layer. i σ represents the weights retained by each node, and σ is the activation function. This aggregation method ensures that neighboring nodes with strong ink path decay and overprint coupling contribute more to the feature updates of the target node, while neighboring nodes with weaker physical connections contribute less, thus achieving message passing driven by physical laws.

[0055] Finally, after multiple layers of message passing iterations, the output is a predicted color difference field for each node that integrates the spatial attenuation of ink paths and the optical coupling of overprinting. Each layer of message passing extends the color difference information outwards along the edges of the region topology graph. After multiple layers of transmission, the feature representation of each node incorporates the color difference information of its multi-order neighbors. Simultaneously, this information is physically constrained by ink path attenuation and overprinting coupling during propagation. In the final output predicted color difference field, the predicted color difference value of each node reflects the combined result of the region's own color difference and the contribution of color difference from adjacent regions propagating through ink path spatial attenuation and optical coupling of overprinting, making the prediction results closer to the actual color difference formation mechanism in the printing process.

[0056] Furthermore, the execution steps of this embodiment, before inputting the initial chromaticity values ​​of each region as node feature inputs into the graph convolutional network, also include: Identify the type attributes of each color separation area based on the prepress documents. The type attributes include brand logo area, main visual image area, background solid area, and explanatory text area. Based on the preset weights of different types of attributes in the visual evaluation of packaging and decoration, each node is assigned a visual salience score. By concatenating the visual saliency score with the initial chromaticity value to form a composite feature vector, the input graph convolutional network is made so that when highly saliency nodes aggregate neighbor messages, their own features are retained at a higher proportion in the final node representation.

[0057] Specifically, in a preferred embodiment, a preprocessing step is included before inputting the initial chromaticity values ​​of each region as node feature inputs into the convolutional network, which is explained in detail below:

[0058] First, different areas on packaging and printed materials have varying importance in visual evaluation. The brand logo area serves as the primary means of brand identification and is the area consumers notice first; its color accuracy directly impacts brand image and product recognizability. The main visual image area occupies the majority of the printed material's visual space, and color deviation directly affects the overall visual effect and consumers' perception of product quality. While the background solid area is larger, its visual attention is relatively lower. The explanatory text area primarily conveys information, and its tolerance for color accuracy is generally higher than the other two. If all areas are treated equally in color difference prediction, the color difference between the brand logo area and the main visual image area may be diluted by the large amount of data in the background area, leading to insufficient prediction accuracy for highly important areas.

[0059] Therefore, this step identifies the type attributes of each color separation area based on the prepress file. These type attributes include the brand logo area, the main visual image area, the solid background area, and the explanatory text area. Identification can be based on layer names, layer groups, or user-annotated information in the prepress file. Based on the preset weights of different type attributes in the visual evaluation of packaging design, each node is assigned a visual salience score. For example, the brand logo area is the most salience area, with a visual salience score set to 0.9; the main visual image area is second, with a score of 0.7; the solid background area is next, with a score of 0.4; and the explanatory text area is the least salience area, with a score of 0.2. The visual salience score of each area reflects the importance of color accuracy in the overall quality evaluation of that area.

[0060] Next, the visual saliency score is concatenated with the initial chromaticity value to form a composite feature vector, which is then input into the graph convolutional network. This ensures that when highly saliency nodes aggregate neighbor messages, their own features are retained at a higher proportion in the final node representation. Specifically, in each layer of the graph convolutional network's message passing, a retention weight B is set for the node itself. i (l) =Sigmoid(S i When the visual saliency score S of node i i When the value is high, Sigmoid(S i A larger value, close to 1, indicates that the node retains a high proportion of its original features during aggregation. This means that highly saliency nodes primarily rely on their own features for judgment, and the color difference information of neighboring nodes has a relatively small impact. When the visual saliency score S of node i is... i At lower levels, Sigmoid(S i A smaller value, close to 0.5, indicates that the node is relatively more influenced by the features of its neighboring nodes during aggregation. Low-saliency nodes are more likely to absorb color difference information propagated from adjacent areas. This mechanism ensures the dominance of high visual saliency regions in color difference prediction, avoiding the obscuring of color difference features in key areas such as brand logos due to the large-scale propagation of color differences in the background area.

[0061] S50: Based on the predicted color difference field, reversely calculate the target compensation adjustment amount for each ink key, and send the target compensation adjustment amount to the printing press ink control system.

[0062] Finally, the predicted color difference field output by the graph convolutional network provides the expected color difference distribution for each color separation region under the current ink key opening configuration. This prediction result itself only completes the forward deduction from the physical state to the color difference. To eliminate the color difference indicated in the predicted color difference field, the color difference information needs to be mapped from the region space back to the ink key control space, and the ink key opening adjustment amount required to make the color difference of each region approach zero needs to be solved. The physical basis of this inverse mapping is that there is a definite sensitivity relationship between the ink key opening and the region chromaticity during the printing process. When the opening of a certain ink key changes, the chromaticity of its directly opposite region and adjacent regions will change accordingly according to the ink path space attenuation law.

[0063] Specifically, this step first maps the color difference vectors of each region in the predicted color difference field to the control range of each ink key, obtaining the regional color difference vectors within each ink key control range. Then, a sensitivity relationship matrix is ​​established between the change in ink key opening and the change in regional color difference. The elements of this matrix are derived from the ink path space attenuation matrix combined with colorimetric principles, reflecting the degree of influence of a unit change in ink key opening on regional color difference. With the goal of eliminating predicted color difference, the change in ink key opening is calculated using a least-squares algorithm, and it is determined whether the compensation adjustment of each ink key exceeds the physical maximum adjustment stroke. Compensation amounts within the stroke are sent to the ink control system for closed-loop compensation, while alarm signals are generated for abnormal situations of exceeding the stroke. Through this reverse calculation and closed-loop execution mechanism, the color difference prediction results of the graph convolutional network can be transformed into physical adjustment commands that the printing press actuator can actually execute, thereby achieving complete control from color difference prediction to physical compensation.

[0064] Specifically, in this embodiment, the target compensation adjustment amount for each ink key is calculated in reverse based on the predicted color difference field, and the target compensation adjustment amount is sent to the printing press ink control system, including: The predicted color difference field is mapped to the printing press ink key control space to obtain the regional color difference vector within the control range of each ink key. Establish a sensitivity relationship matrix between the change in ink bond opening and the change in regional color difference; The ink bond opening change that makes the color difference value of each region in the predicted color difference field approach zero is calculated using the least squares solution algorithm, and is used as the target compensation adjustment amount. Determine whether the target compensation adjustment amount of each ink key exceeds its physical maximum adjustment stroke; When the target compensation adjustment is within limits, the target compensation adjustment amount is sent to the ink control system to perform closed-loop compensation. When the target is exceeded, the corresponding area is marked as an unrecoverable color difference abnormality and an alarm signal is generated.

[0065] First, the predicted color difference field is mapped to the printing press inking key control space to obtain the regional color difference vector within each inking key control range. Specifically, each color separation region in the predicted color difference field corresponds to a three-dimensional color difference vector, composed of the color differences of the L, a, and b channels. Each component of the color difference vector represents the deviation between the measured chromaticity value and the standard chromaticity value of that region under the current printing state in the L, a, and b channels. This color difference vector needs to be mapped from the region space to the inking key control space to determine the amount of color difference that needs to be compensated for by each inking key.

[0066] The ink keys of the printing press are uniformly arranged along the cylinder axis, with each ink key corresponding to a fixed width range in the cylinder axis. The control range of the ink key is determined by the boundary coordinates of this width range. The color difference vector of each color separation region in the predicted color difference field is associated with the position coordinates of that region in the cylinder axis. The position coordinates of the region in the cylinder axis are represented by the coordinates of the geometric center of the region in the axial direction, which have been calculated in step S20.

[0067] For each ink key, all color regions whose geometric center axial coordinates fall within the control range of that ink key are selected from the predicted color difference field, forming the associated region set for that ink key. The determination criterion is that the geometric center coordinates of the region are greater than or equal to the starting coordinates and less than or equal to the ending coordinates of the ink key's control range. Since the control range of an ink key is a continuous axial interval, each region is uniquely assigned to its corresponding ink key based on its geometric center coordinates. For each ink key, the color difference vector of the region within its control range is equal to the weighted average of the color difference vectors of all regions in its associated region set, with the weights determined by the ink quantity influence coefficient in the ink path space attenuation matrix. Regions directly opposite the ink key, i.e., regions whose geometric center coordinates are near the center of the ink key's control range, have a larger ink quantity influence coefficient and a higher weight; regions not directly opposite the ink key, i.e., regions whose geometric center coordinates are near the edge of the ink key's control range, have a smaller ink quantity influence coefficient and a lower weight. The specific calculation method is as follows: For a given ink key, the ink quantity influence coefficient of that ink key on each region is used as a weight, multiplied by the color difference vector of the corresponding region to obtain a weighted color difference vector for each region. Then, the weighted color difference vectors of all regions are summed, and the result is divided by the sum of the ink quantity influence coefficients of that ink key on all regions to obtain the regional color difference vector within the control range of that ink key. This weighted averaging method allows the regional color difference vector within the control range of the ink key to more accurately reflect the color difference situation of the region directly opposite the ink key, while also taking into account the color difference contribution of non-directly opposite regions. Through the above mapping, the predicted color difference field is transformed from the discrete color difference vectors of each color separation region into regional color difference vectors within the control range of each ink key.

[0068] Next, a sensitivity relationship matrix is ​​established between the change in ink bond opening and the change in regional color difference. Assume there are K ink bonds and N color separation regions. The change in ink bond opening ΔU is a K-dimensional vector, and the change in regional color difference ΔE is an N-dimensional vector. The sensitivity relationship matrix J is an N-row, K-column matrix, where the element in the i-th row and k-th column J... ik This represents the change in color difference in region i when the ink key k changes by a unit amount. The elements of this sensitivity matrix can be derived from the ink path space attenuation matrix combined with colorimetric principles: J ik =W ki ×c, where W ki Let be the influence coefficient of ink key k on the ink amount in region i, and c be the proportionality coefficient between ink amount change and color change, which can be determined through offline calibration experiments of the printing press.

[0069] Furthermore, using a least-squares algorithm, the change in ink bond opening that makes the color difference values ​​of each region in the predicted color difference field approach zero is calculated and used as the target compensation adjustment amount. Let E be the color difference vector of each region in the predicted color difference field. pred The objective equation requires J×ΔU=-E pred The change in color difference resulting from adjusting the ink key exactly offsets the current predicted color difference. The least squares solution is... Where λ is the regularization damping coefficient, used to prevent excessively large abnormal adjustment values ​​from being calculated, J is the Jacobian matrix of the ink key region color difference, and I is the identity matrix. Let be the target compensation adjustment amount for each ink key. This solution yields the target compensation adjustment amount for each ink key.

[0070] Next, it is determined whether the target compensation adjustment amount for each ink key exceeds its physical maximum adjustment stroke. The physical adjustment range of each ink key is determined by the mechanical structure of the printing press. The upper limit is the maximum percentage of the ink key's opening that can be opened, and the lower limit is the minimum percentage of the ink key's opening that can be closed. The upper limit is usually 100%, and the lower limit is usually 0%. The current ink key opening refers to the actual percentage of the ink key's opening at the current moment. The target compensation adjustment amount is the percentage change in the opening that the ink key needs to increase or decrease. The compensated opening is equal to the current opening plus the target compensation adjustment amount.

[0071] When the compensated ink opening is less than or equal to the upper limit and greater than or equal to the lower limit, it indicates that the target compensation adjustment amount has not exceeded the physical adjustment capability of the ink key. The target compensation adjustment amount is then sent to the ink control system to perform closed-loop compensation. The ink control system drives each ink key to change its opening according to this adjustment amount, thereby achieving automatic color difference compensation. When the compensated ink opening is greater than the upper limit or less than the lower limit, it indicates that the target compensation adjustment amount has exceeded the physical adjustment capability. Even if the ink key is opened to its maximum or minimum size, the color difference in the corresponding area cannot be eliminated. The area associated with this ink key is marked as an irreversible color difference abnormality, and an alarm signal is generated, prompting the operator that the color difference in this area has exceeded the adjustment capability of the ink key, and intervention is required by other means such as replacing the printing plate or adjusting the ink formula.

[0072] In addition, after the target compensation adjustment amount is sent to the ink control system to perform closed-loop compensation, the following steps are also included: Obtain the measured color difference data of the next batch of printed materials after the compensation adjustment is performed; The predicted color difference field after compensation adjustment is used as the theoretical compensation effect, and the residual deviation between the measured color difference data and the theoretical compensation effect is calculated. When the residual deviation of multiple consecutive regions along a certain axis is greater than the preset failure threshold, and the corresponding compensation adjustment has reached the stroke limit, it is determined that the ink roller of the printing press corresponding to the axial interval has local wear or ink accumulation. Extract the axial coordinate range of the area where the residual deviation exceeds the limit, generate a local maintenance early warning signal for the ink roller, and send it to the workshop terminal.

[0073] After the target compensation adjustment is sent to the ink control system to perform closed-loop compensation, this embodiment also includes the steps of diagnosing local wear and contamination of the ink roller, which are explained in detail below:

[0074] First, obtain the measured color difference data for the next batch of printed materials after the compensation adjustment. The next batch of printed materials refers to the subsequent batch produced after the ink control system adjusts the ink key opening according to the target compensation adjustment amount. Measure the actual chromaticity values ​​of each color separation region in this batch of printed materials using an online or offline colorimeter, calculate the difference from the standard chromaticity values, and obtain the measured color difference data.

[0075] Next, the predicted color difference field after compensation adjustment is used as the theoretical compensation effect. When the graph convolutional network outputs the predicted color difference field, this predicted value is the expected color difference distribution under the current ink key aperture configuration. After compensation adjustment, the ink key aperture has changed, and the theoretical compensation effect corresponding to the predicted color difference field should be recalculated based on the new ink key aperture, i.e., the compensated predicted color difference field. This theoretical compensation effect reflects the expected residual value of color difference in each region after ink key compensation.

[0076] Calculate the residual deviation between the measured color difference data and the theoretical compensation effect. The residual deviation is equal to the absolute difference between the measured color difference data and the theoretical compensation effect. When the residual deviation is small, it indicates that the actual effect of ink key compensation is consistent with the theoretical prediction, and the ink path system is operating normally. When the residual deviation is large, it indicates that the actual compensation effect deviates from the theoretical expectation, and there may be a physical abnormality in the ink path system.

[0077] Specifically, when the residual deviation of multiple consecutive regions along a certain axis exceeds a preset failure threshold, and the corresponding compensation adjustment has reached its travel limit, it is determined that the ink roller of the printing press corresponding to the axial section has local wear or ink accumulation. For example, the preset failure threshold can be set to 2 color difference units, and the number of consecutive axial regions can be set to 3. When the ink key compensation adjustment has reached its upper or lower limit, i.e., the ink key is at its maximum or minimum opening position, but the color difference in the corresponding region still exists and cannot be eliminated, and the residual deviation of this region and its adjacent regions both exceed the preset failure threshold, it indicates that the ink roller in this axial section may have mechanical local wear or ink accumulation. Local wear of the ink roller leads to uneven ink transfer efficiency, and ink accumulation leads to abnormal local ink layer thickness. Both situations will cause ink key adjustment to fail and cannot be corrected by adjusting the ink key opening.

[0078] Finally, the axial coordinate range of the residual deviation exceeding the limit is extracted, a local maintenance warning signal for the ink roller is generated, and sent to the workshop terminal. The axial coordinate range refers to the start and end positions of the continuous area of ​​residual deviation exceeding the limit along the roller axis. This coordinate range can be used to locate the specific location on the ink roller that needs maintenance. After receiving the warning signal, the workshop terminal can arrange a shutdown for inspection, clean or replace the corresponding position on the ink roller, restore the normal ink transfer capability of the ink roller, and thus avoid persistent color difference problems caused by local abnormalities of the ink roller. Through this diagnostic mechanism, adaptive identification of the cause of ink key compensation failure can be achieved, elevating the color difference problem from the control parameter adjustment level to the equipment maintenance level.

[0079] In addition, the execution steps of this embodiment also include: Calculate the mean and variance of the predicted residuals for each region in the current production batch; When the mean of the predicted residual is within the preset range but the variance of the predicted residual continues to increase, it is determined that the physical properties of the ink have drifted. Obtain the current workshop environment temperature and humidity parameters and ink viscosity parameters, and dynamically correct the ink volume spatial diffusion scale and the optical coupling coefficient between overprinting layers based on the environmental temperature and humidity parameters and ink viscosity parameters; By recalculating the dynamic edge weights using the corrected physical parameters, the aggregate weights of the graph convolutional network are updated online, while keeping the basic learnable parameters of the graph convolutional network unchanged.

[0080] Specifically, this embodiment also includes the step of online model anti-drift update: The first step is to statistically analyze the mean and variance of the predicted residuals for each region in the current production batch. The predicted residual refers to the difference between the measured color difference data and the theoretical compensation effect. After each production batch is completed, the predicted residuals for all regions are statistically analyzed. The arithmetic mean of the predicted residuals for all regions is calculated as the mean of the predicted residuals, and the average of the squared deviations of the predicted residuals from their mean is calculated as the variance of the predicted residuals. The absolute value of the mean of the predicted residuals reflects the overall deviation of the compensation system, while the variance of the predicted residuals reflects the consistency of the compensation effect across different regions.

[0081] The second step involves determining that the ink's physical properties have drifted when the mean of the predicted residuals is within a preset range but the variance of the predicted residuals continues to increase. A mean predicted residual within the preset range indicates that the overall color difference compensation capability of the system has not undergone a systematic shift, and the compensation amount is generally still at a normal level. However, a continuous increase in the variance of the predicted residuals indicates that the difference in compensation effect between different regions is widening; some regions have effectively compensated for color difference while others have insufficient compensation. This phenomenon is usually caused by changes in the physical properties of the ink, such as changes in ink viscosity with ambient temperature and humidity, or changes in the dispersion state of pigment particles in the ink. These changes affect the ink's transfer and optical properties, making the physical parameters originally calibrated based on historical data no longer applicable to the current operating conditions.

[0082] The third step involves acquiring the current ambient temperature and humidity parameters, as well as the ink viscosity parameters. Based on these parameters, the spatial diffusion scale of the ink volume and the optical coupling coefficient between the overprinting layers are dynamically adjusted. As the ambient temperature increases, the ink viscosity decreases, increasing the axial diffusion range of the ink, thus requiring a corresponding increase in the spatial diffusion scale. Conversely, as the ambient humidity increases, the paper's moisture content increases, altering the ink's absorption and drying characteristics, necessitating adjustments to the optical coupling coefficient between the overprinting layers. Ink viscosity directly determines the rheological behavior of the ink on the ink roller surface; changes in viscosity affect ink transfer efficiency and overprinting optical performance. Optionally, the correction method involves substituting the current ambient temperature and humidity parameters and the ink viscosity parameters into a preset correction function. This function can be obtained by fitting historical data, outputting the corrected spatial diffusion scale of the ink volume and the corrected optical coupling coefficient between the overprinting layers.

[0083] The fourth step involves recalculating the dynamic edge weights using the corrected physical parameters to update the aggregate weights of the graph convolutional network online, while keeping the network's fundamental learnable parameters unchanged. Recalculating the dynamic edge weights involves recalculating the ink volume influence coefficient using the corrected ink volume spatial diffusion scale, recalculating the optical coupling coefficient using the corrected interlayer optical coupling coefficient, and then recalculating the dynamic edge weights between adjacent nodes using the normalization and multiplication methods described in step S40. The updated dynamic edge weights are then substituted into the forward propagation process of the graph convolutional network, replacing the original aggregate weights. During this process, the fundamental learnable parameters of the graph convolutional network remain unchanged; only the dynamic edge weights driven by the physical parameters are adjusted. This design allows the graph convolutional network to quickly adapt to changes in ink physical properties without retraining the entire network, thus achieving online adaptive capability during long-term production processes.

[0084] In summary, the embodiments of this application have at least the following technical effects: This invention first extracts the polygonal boundaries of each color separation region by parsing the prepress file. Based on the shortest distance determination of the outer rectangle, a regional topology map is established, transforming the multi-regional spatial structure of packaging and decoration printing materials into graph-structured data. This provides a structured data foundation for subsequent message passing in the graph convolutional network. Secondly, based on the axial distance between the ink key control center and the geometric center of the color separation region, combined with the ink roller travel distance and dot gain rate, the spatial diffusion scale of ink volume is determined. The ink volume influence coefficient is calculated using a Gaussian decay function, and an ink path spatial decay matrix is ​​generated. This allows the graph convolutional network to adaptively adjust weights according to the physical decay law of the ink path when aggregating neighbor node information, solving the technical problem in existing methods where the influence of ink key adjustment on non-aligned regions cannot be accurately quantified.

[0085] Furthermore, based on the layer relationships of ink overprinting, the spectral absorption characteristics of the bottom layer are extracted. The spectral change in color difference transmitted from the bottom layer to the surface layer is calculated using Kubelka-Munk theory. The optical coupling coefficient between overprinting layers is obtained through CIE colorimetric system integral mapping, enabling the network to capture the optical mechanism of color difference transmission from the bottom layer to the surface layer. This solves the problem of quantitatively describing the impact of bottom layer color difference on surface color during multi-layer ink overprinting. Finally, the predicted color difference field output by the graph convolutional network is inversely calculated into the target compensation adjustment amount for each ink bond using a sensitivity relationship matrix. After stroke determination, this is sent to the printing press ink control system for closed-loop compensation. After compensation, residual deviation analysis is used to diagnose local wear and contamination of the ink rollers. Simultaneously, the change in predicted residual variance is used to determine ink physical property drift and update physical parameters online. This achieves complete closed-loop control from color difference prediction to actuator physical compensation, equipment status diagnosis, and model adaptive updating.

[0086] This invention solves the technical problems in the prior art, such as the disconnect between color difference prediction and execution, the inability to quantify the impact of ink key adjustment on non-aligned areas, the difficulty in quantitatively describing the optical coupling effect of overprinting, and the inability of the model to adapt to changes in ink properties online.

[0087] It should be noted that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting color difference in multiple areas of packaging and decoration printed materials, characterized in that, include: The prepress file is parsed to obtain the region topology map of the printed matter to be predicted. The nodes of the region topology map are the color separation regions on the printed matter, and the edges are the spatial adjacency relationships of adjacent regions. Obtain the current opening data of each ink key of the printing press, calculate the influence coefficient of each ink key fluctuation on the ink volume of each area based on the spatial mapping relationship between the ink key and each area, and generate the ink path spatial attenuation matrix. Based on the ink overprinting layer relationship between different areas in the prepress file, the spectral absorption rate of the ink in the bottom layer is extracted, and the optical coupling coefficient between the overprinting layers is calculated. The initial chromaticity values ​​of each region are used as node features and input into the graph convolutional network. The aggregation weights of the graph convolutional network during message passing between adjacent nodes are continuously modulated using the ink path spatial attenuation matrix and the optical coupling coefficient between overprinting layers, and the predicted chromaticity field of each region is output. Based on the predicted color difference field, the target compensation adjustment amount of each ink key is calculated in reverse, and the target compensation adjustment amount is sent to the printing press ink control system.

2. The method for predicting multi-region color difference in packaging and decoration printed materials according to claim 1, characterized in that, Parsing prepress documents yields a region topology map of the printed matter to be predicted. The nodes of this topology map represent the color separation regions on the printed matter, and the edges represent the spatial adjacency relationships between adjacent regions, including: Analyze the layer information and geometric contour data in the prepress file to extract the polygon boundaries of each color separation area; A two-dimensional coordinate system is established with the paper feed direction of the printing press and the axis of the cylinder as the reference, and the outer rectangle of the polygon of each color separation area is calculated. Traverse all bounding rectangles. When it is determined that the shortest distance between the bounding rectangles of two regions is less than the preset proximity tolerance, establish a spatial adjacency edge between the corresponding two regions. A regional topology graph is constructed using each color-separated region as a node and the adjacent edges in the space as edges.

3. The method for predicting multi-region color difference in packaging and decoration printed materials according to claim 1, characterized in that, Obtain the current opening data of each ink key on the printing press. Based on the spatial mapping relationship between the ink keys and each region, calculate the influence coefficient of ink volume on each region caused by the fluctuation of each ink key. Generate the ink path spatial attenuation matrix, including: A one-dimensional spatial coordinate system is established along the axial direction of the printing press cylinder to obtain the control center coordinates of each ink key and the geometric center coordinates of each color separation area. Calculate the axial distance from the control center of any ink key to the geometric center of any color separation region; Obtain the traverse stroke parameters of the ink roller of the printing press and the dot gain of the current printed paper, and determine the spatial diffusion scale of ink volume based on the traverse stroke parameters and the dot gain. Based on the correlation between the axial distance and the spatial diffusion scale of the ink volume, an index value characterizing the degree of ink volume attenuation is calculated as the ink bond's influence coefficient on the ink volume of the corresponding color separation region. Summarize the ink volume influence coefficients of all ink keys and all color separation regions, and construct the ink path space attenuation matrix.

4. The method for predicting multi-region color difference in packaging and decoration printed materials according to claim 1, characterized in that, Based on the ink overprinting layer relationships between different areas in the prepress file, the spectral absorption rate of the ink in the bottom layer is extracted, and the optical coupling coefficient between the overprinting layers is calculated, including: Analyze the layer stacking order in prepress files to identify the bottom and top layers with overlapping relationships; Based on the ink spectral database, the spectral absorption and scattering characteristics of the underlying ink in the visible light band are extracted; Based on the spectral absorption characteristics, calculate the amount of spectral change transmitted to the surface ink when the chromaticity value of the bottom ink shifts; The spectral change is integrated and mapped in the visible light band to convert it into a chromaticity shift of the surface color. The ratio of the chromaticity shift of the surface color to the chromaticity shift of the underlying ink is used as the optical coupling coefficient between the overprinting layers. The optical coupling coefficient between the overprinting layers is used to characterize the proportion of color difference transferred from the underlying layer to the surface layer.

5. The method for predicting multi-region color difference in packaging and decoration printed materials according to claim 1, characterized in that, The steps for constructing and training a graph convolutional network include: Obtain a historical printing sample set, which includes the measured chromaticity values ​​of each color separation region and the corresponding standard chromaticity values, as training samples; A graph structure is constructed by using the nodes of the aforementioned regional topology graph to correspond to each color-coded region and the edges to correspond to the spatial adjacency relationships of adjacent regions. The deviation between the measured chromaticity value and the standard chromaticity value of each node is used as the initial feature of the node, and the adjacency matrix of the region topology graph is used as the initial edge weight to construct an initial graph convolutional network. The training samples are input into the initial graph convolutional network for forward propagation, and the difference between the predicted color difference and the measured color difference output by the network is calculated as the training loss. The network parameters of the graph convolutional network are updated by backpropagation based on the training loss, and the training is iterated until the training loss converges, resulting in a trained graph convolutional network.

6. The method for predicting multi-region color difference in packaging and decoration printed materials according to claim 1, characterized in that, The initial chromaticity values ​​of each region are used as node features and input into the graph convolutional network. Utilizing the ink path spatial attenuation matrix and the interlayer optical coupling coefficient, the aggregation weights of the graph convolutional network during message passing between adjacent nodes are continuously modulated to output the predicted chromatic difference field for each region, including: In each message passing iteration of the graph convolutional network, the ink volume influence coefficient and the optical coupling coefficient between the overprinting layers corresponding to the two adjacent nodes are extracted. After normalizing the ink volume influence coefficient and the optical coupling coefficient between overprinting layers, the product is taken and the resulting product is used as the dynamic edge weight between adjacent nodes. The dynamic edge weights are used to replace the fixed symmetric adjacency matrix in the graph convolutional network, and the feature vectors of neighboring nodes are weighted, summed, and aggregated. After multiple layers of message passing iterations, the output is a predicted chromatic difference field for each node that integrates the spatial attenuation of ink path and the optical coupling relationship of overprinting.

7. The method for predicting multi-region color difference in packaging and decoration printing as described in claim 6, characterized in that, Before inputting the initial chromaticity values ​​of each region into the convolutional network of the node feature graph, the following steps are also included: Identify the type attributes of each color separation area based on the prepress documents. The type attributes include brand logo area, main visual image area, background solid area, and explanatory text area. Based on the preset weights of different types of attributes in the visual evaluation of packaging and decoration, each node is assigned a visual salience score. The visual saliency score is concatenated with the initial chromaticity value to form a composite feature vector, which is then input into the graph convolutional network. This results in a higher retention rate of the features of highly saliency nodes in the final node representation when aggregating neighbor messages.

8. The method for predicting multi-region color difference in packaging and decoration printed materials according to claim 1, characterized in that, Based on the predicted color difference field, the target compensation adjustment amount for each ink key is calculated in reverse, and the target compensation adjustment amount is sent to the printing press ink control system, including: The predicted color difference field is mapped to the printing press ink key control space to obtain the regional color difference vector within the control range of each ink key. Establish a sensitivity relationship matrix between the change in ink bond opening and the change in regional color difference; The ink bond opening change that makes the color difference value of each region in the predicted color difference field approach zero is calculated using the least squares solution algorithm and is used as the target compensation adjustment amount. Determine whether the target compensation adjustment amount of each ink key exceeds its physical maximum adjustment stroke; When the target compensation adjustment amount is within the limit, it is sent to the ink control system to perform closed-loop compensation. When the target is exceeded, the corresponding area is marked as an unrecoverable color difference abnormality and an alarm signal is generated.

9. A method for predicting multi-region color difference in packaging and decoration printed materials according to claim 8, characterized in that, After the target compensation adjustment amount is sent to the ink control system to perform closed-loop compensation, the following steps are also included: Obtain the measured color difference data of the next batch of printed materials after the compensation adjustment is performed; Using the predicted color difference field after compensation adjustment as the theoretical compensation effect, the residual deviation between the measured color difference data and the theoretical compensation effect is calculated. When the residual deviation of multiple consecutive regions along a certain axis is greater than the preset failure threshold, and the corresponding compensation adjustment has reached the stroke limit, it is determined that the ink roller of the printing press corresponding to the axial interval has local wear or ink accumulation. Extract the axial coordinate range of the area where the residual deviation exceeds the limit, generate a local maintenance early warning signal for the ink roller, and send it to the workshop terminal.

10. A multi-region color difference prediction method for packaging and decoration printing according to claim 9, characterized in that, Also includes: Calculate the mean and variance of the predicted residuals for each region in the current production batch; When the mean of the predicted residual is within a preset range but the variance of the predicted residual continues to increase, it is determined that the physical properties of the ink have drifted. Obtain the current workshop environment temperature and humidity parameters and ink viscosity parameters, and dynamically correct the ink volume spatial diffusion scale and the optical coupling coefficient between overprinting layers based on the environmental temperature and humidity parameters and ink viscosity parameters; By recalculating the dynamic edge weights using the corrected physical parameters, the aggregate weights of the graph convolutional network are updated online, while keeping the basic learnable parameters of the graph convolutional network unchanged.

Images