Deep learning-based color printing product color difference intelligent detection and correction processing system

Abstract

The invention relates to the technical field of printing automation control and image processing, and discloses a color printing product color difference intelligent detection and correction processing system based on deep learning, which comprises an image acquisition device, printing equipment and a data processing server. The system decomposes a source image into microcosmic and macroscopic blocks through an image preprocessing module. The microscopic lattice point topology analysis module extracts lattice point features, and the geometric constraint type spectral manifold reconstruction module reconstructs a spectral reflectivity vector in combination with macroscopic data. The self-adaptive feedback calibration module reversely optimizes a microscopic segmentation threshold value by using a spectrum confidence coefficient residual error, so that the physical consistency of feature extraction is ensured. And the heterogeneous feature tensor coding and reverse control module maps the spectrum and the dot features into a combined state tensor, decouples the ink chemical absorption and dot physical deformation effect, and generates an independent ink path and imprinting adjustment instruction based on a rheological coupling matrix. Accurate decoupling control over the ink amount and the pressure in the printing process is achieved, and the color consistency and the printing quality are improved.

Description

Technical Field

[0001] This invention relates to the field of printing automation control and image processing technology, specifically a deep learning-based intelligent detection and correction system for color difference in color printed materials. Background Technology

[0002] In modern high-speed multicolor offset printing production, in order to meet increasingly stringent requirements for color consistency and printing quality, machine vision-based online quality inspection systems have gradually replaced traditional manual sampling inspection methods. Current mainstream inspection solutions typically use industrial line scan cameras to acquire RGB images of the printed surface, and then calculate color difference or density deviation to adjust the ink bond opening in order to maintain the color stability of the printed product.

[0003] However, existing online detection and control technologies still face multi-dimensional technical bottlenecks in practical applications. First, relying solely on RGB three-channel data for color control has inherent limitations. Due to metamerism, the low-dimensional RGB response cannot correspond one-to-one with the high-dimensional spectral reflectance, making it difficult for the system to accurately distinguish between changes in the spectral absorption characteristics of ink and the influence of external light fluctuations. This means that data collected under non-standard lighting conditions often cannot truly reflect the physical properties of printed materials under standard light sources, resulting in inaccurate detection where the data is acceptable but the visual appearance is not.

[0004] More importantly, most existing control logics are based on a single chromaticity-ink volume mapping model, lacking in-depth analysis and decoupling capabilities for the physical causes of color difference. The printing process is a complex physicochemical coupling process. The resulting color deviation stems from both changes in ink layer thickness caused by abnormal ink supply (chemical dimension) and mechanical dot amplification or deformation caused by abnormal roller pressure and blanket relaxation (physical dimension). Existing technologies struggle to distinguish between these two distinct causes, often mistakenly attributing tonal darkening due to dot gain caused by excessive pressure to excessive ink volume, thus incorrectly issuing instructions to reduce ink volume. This attribution fallacy not only fails to fundamentally eliminate color difference but also leads to reduced image contrast, loss of detail, and even oscillations in the ink path and printing mechanism.

[0005] Furthermore, in the microscopic dot feature extraction stage, existing technologies typically employ image segmentation algorithms with fixed parameters. However, the printing production environment is complex and variable; dampening solution emulsification, plate wear, or paper dust accumulation can all cause nonlinear blurring or distortion of dot edge features. Under these non-ideal conditions, algorithms based on preset fixed thresholds cannot adaptively capture the true dot topology, resulting in significant errors in calculated key parameters such as dot coverage, further misleading subsequent control decisions. Therefore, there is an urgent need for an intelligent detection and correction system that can integrate microscopic physical morphology and macroscopic spectral characteristics, and achieve decoupled control of ink volume and pressure. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a deep learning-based intelligent detection and correction system for color difference in printed materials. This system solves the problem that existing online printing detection technologies cannot distinguish between the mixed effects of ink supply variations (chemical factors) and dot mechanical deformation caused by roller pressure (physical factors) on color difference, which leads to attribution errors in the control system, causing adjustment oscillations and unstable printing quality.

[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a deep learning-based intelligent color difference detection and correction system for color printed materials. The system includes an image acquisition device, printing equipment, and a data processing server. The image acquisition device is configured to acquire the source RGB image of the printed material output by the printing equipment. The printing equipment includes an ink key actuator for adjusting the ink supply and a pressure regulating mechanism for adjusting the roller contact pressure. The data processing server is communicatively connected to the image acquisition device and the printing equipment.

[0008] In the data processing server's operational logic, the source RGB image is first decomposed into two independent analysis channels by the image preprocessing module: microscopic patches containing halftone dot structures and macroscopic patches containing average color data. This splitting process aims to separate high-frequency spatial texture information from low-frequency color energy information.

[0009] Subsequently, the micro-dot topology analysis module extracts features from micro-patterns based on a segmentation threshold, outputting a dot topology feature vector. This process utilizes anisotropic differential operators to calculate the gradient vector of each pixel in the micro-pattern and constructs a local structure tensor matrix based on these gradient vectors. Eigenvalue decomposition is performed on the structure tensor matrix, and the anisotropy index is calculated using the obtained eigenvalues. This anisotropy index characterizes the degree of deformation (e.g., stretching or slippage) of the dot edges under mechanical action. The system combines the dot mask generated by the segmentation threshold with statistical analysis of the anisotropy index distribution characteristics, thereby generating a dot topology feature vector describing the micro-morphology of the dots.

[0010] This invention further introduces a geometrically constrained spectral manifold reconstruction module to address the ill-conditioned inversion problem of recovering high-dimensional spectral data from low-dimensional RGB data. This module pre-stores a printing gamut manifold basis matrix calibrated using standard test data. During runtime, the module uses the dot topological feature vector as input to the geometric physical mapping unit and the macroscopic tiles as input to the color response mapping unit, fusing their outputs to calculate the latent coordinate vector within the manifold space. Subsequently, the latent coordinate vector is projected onto the printing gamut manifold basis matrix to reconstruct a pixel-level spectral reflectance vector. This step utilizes microscopic geometric features as physical constraints, eliminating the ambiguity caused by metamerism.

[0011] To ensure the accuracy of feature extraction, this invention incorporates an adaptive feedback calibration module. This module identifies the target manifold cluster based on the color coordinates of the macroscopic patches and obtains the cluster's center vector and covariance matrix. The system calculates the Mahalanobis distance between the reconstructed spectral reflectance vector and the target manifold cluster, defining this distance as the spectral confidence residual. When the spectral confidence residual exceeds a preset tolerance, it indicates that the current microscopic segmentation parameters do not possess physical consistency with the macroscopic color representation. In this case, the module calculates the gradient of the spectral confidence residual with respect to the segmentation threshold and uses this gradient to update the segmentation threshold in reverse, triggering the preceding modules to re-execute calculations using the updated parameters until the system converges to a physically consistent state.

[0012] In the control decision-making phase, the heterogeneous feature tensor encoding and inverse control module is responsible for decoupling the multivariable control problem. This module includes a chemical feature encoder and a physical feature encoder, which respectively map the spectral reflectance vector to a chemical state vector and the dot topology feature vector to a physical state vector. The system performs an outer product operation on these two state vectors and applies orthogonalization constraints to generate a joint state tensor. This joint state tensor can mathematically separate the mixed effects of ink chemical absorption and dot physical amplification. Based on a pre-stored rheological coupling matrix (a transfer function describing the effect of ink bond opening changes and roller pressure changes on the spectrum and dot morphology), the module solves for the optimal control vector through a multi-objective optimization function and decomposes it into independent ink path adjustment commands and printing adjustment commands.

[0013] Finally, the ink key actuator adjusts the local opening of the ink fountains of each color group according to the ink path adjustment command to correct the spectral absorption deviation in the chemical dimension; the pressure adjustment mechanism adjusts the contact pressure between the blanket cylinder and the impression cylinder according to the impression adjustment command to correct the dot morphology distortion in the physical dimension.

[0014] Furthermore, in terms of image acquisition, the preferred image acquisition device is an industrial line scan camera mounted above the paper delivery unit of the printing equipment, equipped with a diffuse light source component, and its resolution is configured to resolve the edge structure of a single halftone dot. In the preprocessing stage, the system processes the source image through geometric registration and multi-scale Gaussian pyramid decomposition, and extracts image regions from the bottom of the pyramid for illumination homogenization correction to obtain microscopic patches. Simultaneously, it calculates a three-dimensional color vector without spatial resolution as a macroscopic patch through global pixel integration.

[0015] This invention provides a deep learning-based intelligent color difference detection and correction system for color-printed products. It offers the following advantages: 1. This invention establishes an adaptive feedback calibration module and uses the spectral confidence residual of the reconstructed spectrum to perform reverse gradient updates on the segmentation threshold of the micro-dot topology analysis module. This establishes a closed loop for verifying the macro-color physical reliability of micro-image processing parameters, solving the problem of inaccurate feature extraction by traditional fixed threshold algorithms under non-ideal conditions such as ink emulsification and plate wear that cause dot edge blurring. This ensures the physical consistency and robustness of dot feature extraction and spectral reconstruction results in complex printing production environments.

[0016] 2. This invention uses a geometrically constrained spectral manifold reconstruction module to introduce the microscopic dot topological feature vector as a geometric and physical constraint term into the mapping process from low-dimensional macroscopic tiles to high-dimensional spectral reflectance vectors. By using a pre-calibrated printing color gamut manifold basis matrix to constrain the solution space, it eliminates the metamerism ambiguity that exists when relying solely on RGB color data for spectral inversion. This enables the system to calculate pixel-level spectral data that combines color accuracy and physical state interpretation without relying on high-cost online spectrometers.

[0017] 3. This invention constructs a joint state tensor with orthogonalization constraints through heterogeneous feature tensor encoding and inverse control modules, and performs multi-objective inverse solving by combining rheological coupling matrices. This achieves decoupled analysis of ink chemical absorption characteristics (determined by ink volume) and dot physical and mechanical deformation characteristics (determined by pressure). It can accurately distinguish between two physical causes of the same color difference phenomenon: abnormal ink supply and abnormal roller pressure, and output independent ink path adjustment commands and printing pressure adjustment commands respectively. This avoids the adjustment oscillations caused by fuzzy attribution in traditional control methods and improves the accuracy and stability of printing process control. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall system structure of the present invention; Figure 2 This is a schematic diagram of the microscopic dot topology feature extraction process of the present invention; Figure 3This is a schematic diagram of the geometrically constrained spectral manifold reconstruction process of the present invention; Figure 4 This is a schematic diagram of the adaptive closed-loop feedback calibration process of the present invention; Figure 5 This is a schematic diagram of the system preset and calibration process of the present invention.

[0019] Among them, 110 is an image acquisition device; 120 is a printing equipment; 121 is an ink key actuator; 122 is a pressure regulating mechanism; 130 is a data processing server; 131 is an image preprocessing module; 132 is a microscopic dot topology analysis module; 133 is a geometrically constrained spectral manifold reconstruction module; 134 is an adaptive feedback calibration module; and 135 is a heterogeneous feature tensor encoding and inverse control module. Detailed Implementation

[0020] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] See attached document Figure 1 The present invention provides a printing color closed-loop control system based on multi-scale dot topology constraints and spectral manifold reconstruction. The system mainly includes: an image acquisition device 110, a printing equipment 120, and a data processing server 130.

[0022] Image acquisition device 110 is configured to acquire surface image data of printed matter output from printing equipment 120. In one embodiment, image acquisition device 110 is an industrial-grade high-speed linear CCD or CMOS camera, mounted above the delivery unit of printing equipment 120. Image acquisition device 110 is also equipped with a constant color temperature diffuse light source assembly to provide a uniform lighting environment and eliminate specular reflection interference. The resolution parameters of image acquisition device 110 are configured to resolve the edge structure of a single halftone dot, and its output data is defined as a source RGB image.

[0023] Printing apparatus 120 includes mechanical and ink supply components for performing printing operations. Specifically, printing apparatus 120 includes an ink key actuator 121 and a pressure regulating mechanism 122. The ink key actuator 121 is configured to adjust the local opening of the ink fountains of each color group, thereby controlling the ink supply amount in different areas along the width direction, and thus changing the ink layer thickness on the surface of the printed matter. The pressure regulating mechanism 122 is configured to adjust the contact pressure between the blanket cylinder and the impression cylinder, thereby changing the degree of mechanical extrusion during the ink transfer process.

[0024] The data processing server 130 is communicatively connected to the image acquisition device 110 and the printing equipment 120 via a high-speed industrial bus interface. The data processing server 130 receives the source RGB image from the image acquisition device 110, processes it, and then sends control commands to the printing equipment 120. These control commands include ink path adjustment commands for the ink key actuator 121 and printing pressure adjustment commands for the pressure regulating mechanism 122.

[0025] The data processing server 130 internally stores computer-executable instructions. When these instructions are executed by the processor, they perform processing and control calculations on the source RGB image. Logically, the data processing server 130 includes: an image preprocessing module 131, a microscopic dot topology analysis module 132, a geometrically constrained spectral manifold reconstruction module 133, an adaptive feedback calibration module 134, and a heterogeneous feature tensor encoding and inverse control module 135.

[0026] Image preprocessing module 131 is configured to receive a source RGB image and perform multi-scale pyramid decomposition and region of interest extraction on the source RGB image. Image preprocessing module 131 spatially divides the source RGB image into two types of data units: micro-patterns and macro-patterns. Micro-patterns contain the complete halftone dot periodic structure for subsequent geometric morphology analysis; macro-patterns contain the average color response value of the corresponding region for subsequent spectral data extrapolation. Image preprocessing module 131 is also configured to perform Gaussian smoothing filtering on the micro-patterns to suppress high-frequency noise generated by paper fiber texture.

[0027] The micro-dot topology analysis module 132 is configured to receive preprocessed micro-patterns and calculate the geometric morphological features of the dots. The module constructs a structure tensor matrix for a local region of the micro-pattern using an anisotropic differential operator and performs eigenvalue decomposition on this tensor matrix to calculate the anisotropy index. This anisotropy index characterizes the degree of physical deformation of the dots, including the stretching direction and edge roughness. The module also binarizes the micro-patterns according to a segmentation threshold to generate a dot contour mask and outputs a dot topological feature vector based on the anisotropy index.

[0028] The geometrically constrained spectral manifold reconstruction module 133 is configured to receive color data from macroscopic patches and dot topology feature vectors from the microscopic dot topology analysis module 132. The module 133 pre-stores a printing color gamut manifold basis, which describes the high-dimensional spectral distribution of the printing color space. Using the dot topology feature vectors as physical constraints, the module 133 maps the color data of the macroscopic patches onto the printing color gamut manifold basis, thereby reconstructing the high-dimensional spectral reflectance vector of the pixels in that region.

[0029] The adaptive feedback calibration module 134 is configured to construct a closed-loop feedback loop between microscopic feature extraction and macroscopic spectral reconstruction. The adaptive feedback calibration module 134 calculates the Mahalanobis distance between the spectral reflectance vector output by the geometrically constrained spectral manifold reconstruction module 133 and the preset manifold cluster center; this Mahalanobis distance is defined as the spectral confidence residual. The adaptive feedback calibration module 134 calculates a gradient value for a segmentation threshold based on this spectral confidence residual and uses this gradient value to adjust the segmentation threshold used in the microscopic dot topology analysis module 132 until the spectral confidence residual converges to a preset range.

[0030] The heterogeneous feature tensor encoding and inverse control module 135 is configured to receive calibrated dot topology feature vectors and spectral reflectance vectors. The module maps the spectral reflectance vector to chemical property features and the dot topology feature vector to physical property features, and synthesizes the chemical and physical property features into a joint state tensor in the latent space. This joint state tensor decouples ink layer thickness information from dot mechanical amplification information through orthogonalization constraints. The module then uses a pre-calibrated manifold coupling matrix to inversely solve the joint state tensor, generating a control vector containing ink path adjustment commands and printing adjustment commands, and sends it to the printing equipment 120.

[0031] The image preprocessing module 131 is configured to perform geometric correction, scale decomposition, and data splitting on the raw image data input from the image acquisition device 110 to meet the dual-channel processing requirements of subsequent microscopic topological analysis and macroscopic spectral reconstruction.

[0032] The image preprocessing module 131 first performs a geometric registration operation. During high-speed printing, the paper undergoes nonlinear stretching and deformation due to dampening solution penetration and mechanical traction, which directly results in a shift of the region of interest (ROI). The image preprocessing module 131 identifies crosshair markings or preset corner features at the edges of the original image, calculates the affine transformation matrix using a feature point matching algorithm, and maps the pixel coordinates of the original image to the coordinate system of the standard digital sample, thus eliminating geometric distortion.

[0033] After completing geometric registration, the image preprocessing module 131 constructs a multi-scale Gaussian pyramid data structure. The image at layer 0 is defined. For the original resolution image, the first Layer Image For the first The image is a layer image after low-pass filtering and downsampling. In this embodiment, the pyramid structure is used to separate the high-frequency texture information and low-frequency color information of the image.

[0034] Image preprocessing module 131, based on preset control bar coordinates or key image coordinates, starts from the 0th layer image. Extracting micro-patterns Microscopic tiles The physical size range is configured to cover at least A complete halftone dot cycle is required to meet the sample size requirements for statistical feature extraction. This is for micro-patterns. The image preprocessing module 131 performs illumination uniformity correction to eliminate brightness attenuation caused by vignetting effects from industrial camera lenses or uneven illumination. The corrected micro-pixel values... Calculate using the following formula: ; in, These are the original pixel values. This is the response function of the illumination distribution field obtained by photographing a standard whiteboard before calibration. This is the globally normalized brightness constant. After illumination correction, the system applies a small-scale Gaussian smoothing filter to the microscopic patches, with its kernel size set to be less than 1 / 10 of the diameter of a single dot, in order to suppress paper fiber texture noise while maintaining the sharpness of the dot edge gradient.

[0035] Image preprocessing module 131 synchronously generates macroscopic tiles Macroscopic plot Instead of directly capturing images of the pyramid's upper levels, it uses microscopic images... The macroscopic patch is obtained through global integration calculation. This process simulates the optical integration effect of a spectrophotometer measuring aperture, calculating the arithmetic or weighted average of all pixels within the microscopic patch region in each of the R, G, and B channels. It is a three-dimensional color vector without spatial resolution, which represents the average spectral response characteristics of the image region at a macroscopic observation scale. It removes the interference of the spatial arrangement structure of the dots on color interpretation and serves as the input reference for the subsequent geometrically constrained spectral manifold reconstruction module 133.

[0036] See attached document Figure 2The micro-dot topology analysis module 132 is configured to perform pixel-level differential geometric analysis on the input micro-patterns to quantify the physical deformation state of halftone dots during the printing transfer process. The processing logic of this module does not rely on a preset standard dot template, but is based on the anisotropic statistical characteristics of local image texture.

[0037] The micro-dot topology analysis module 132 first calculates the gradient vector of each pixel in the micro-dot matrix. The horizontal gradient is then calculated using either the Sobel or Scharr operator. and vertical gradient To avoid noise interference from single-pixel gradients and capture the overall direction of halftone dot edges, the module constructs a local structure tensor matrix. For any location in the image... Structure tensor matrix Calculated using the following formula: ; in, The integral smoothing kernel is set to a scale that covers the transition region at the edges of the dots. The structure tensor matrix describes the distribution of gradient directions within the local neighborhood.

[0038] The micro-dot topology analysis module 132 performs eigenvalue decomposition on the structure tensor matrix of each pixel to obtain two non-negative eigenvalues. and (set up These two eigenvalues ​​have clear physical meanings: It characterizes the energy along the direction of maximum gradient change (i.e., the direction perpendicular to the edge of the grid point), while It represents the energy along the direction of minimum gradient change (i.e., parallel to the edge of the grid).

[0039] Based on eigenvalues, the micro-network topology analysis module 132 calculates the anisotropy index. This index is used to determine whether there is directional mechanical stretching or slippage at the edges of the dots. Anisotropy Index Defined as: ; When printing is normal and the dots maintain an ideal geometric shape (such as circles or squares), the local anisotropy distribution at the edges exhibits a regular pattern. When mechanical slippage or ghosting occurs, the dots are elongated or deformed along a preset direction, resulting in anisotropy in that direction. Abnormal increase, which in turn changes The statistical distribution. Simultaneously, the module calculates the principal deformation direction angles. It is used to indicate the physical source direction of mechanical failure (e.g., axial drive error or circumferential roller speed difference).

[0040] The micro-dot topology analysis module 132 further incorporates an initial segmentation threshold. The microscopic patches are binarized to generate a dot mask. Within the area covered by the mask, the module calculates the anisotropy index. The mean and variance, and the principal direction angle The mode of the network and the coverage rate of the network area are calculated. Finally, the micro-network topology analysis module 132 outputs a network topology feature vector containing the aforementioned physical statistics. This vector not only describes the size of the dots, but also carries information about the microscopic morphological health of the dots, serving as a geometric constraint term for subsequent spectral reconstruction.

[0041] See attached document Figure 3 The geometrically constrained spectral manifold reconstruction module 133 is configured to solve the ill-posedness problem when mapping from a low-dimensional RGB space to a high-dimensional spectral space. Its core is to introduce microscopic geometric features as physical prior constraints, thereby eliminating the detection ambiguity caused by metamerism.

[0042] The geometrically constrained spectral manifold reconstruction module 133 first establishes a low-dimensional manifold representation of the printing color space. Based on the principles of printing color rendering, although the spectral reflectance vector usually has high dimensions (e.g., a 31-dimensional vector formed by sampling at 10nm in the 400nm-700nm visible light band), due to the spectral absorption characteristics of ink and the reflectivity of paper, the actual effective printing spectral dataset is distributed on a low-dimensional nonlinear manifold surface. A manifold basis matrix is ​​pre-stored within the module. This matrix is ​​obtained by training a high-dimensional spectral dataset in standard printed conditions using principal component analysis (PCA) or manifold learning algorithms, and contains a set of orthogonal feature spectral vectors. Any measured spectral vector... Both can be approximately represented as a linear combination of superimposed spectral mean vectors of manifold substrates. .

[0043] The geometrically constrained spectral manifold reconstruction module 133 includes a bi-branch feature map network. The first branch is a color response mapping unit, which receives macroscopic patches. The first branch extracts color intensity features; the second branch is a geometric physical mapping unit, which receives dot topology feature vectors from the micro dot topology analysis module 132. This module does not process the two inputs independently, but instead performs a feature fusion mechanism based on physical state.

[0044] Specifically, the module computes the latent coordinate vectors within the manifold space. The coordinate vector This determines the specific location of the reconstructed spectrum in the manifold space. The calculation formula is as follows: ; in, This represents a nonlinear mapping function from the RGB color space to the manifold coordinate space, used to fit the main spectral reflectance trends; This represents a geometric feature encoding function used to analyze the perturbation effect of dot morphology on the spectrum; It is a dynamic gating matrix or a weighted coefficient matrix.

[0045] In this step, the network topology feature vector It played a crucial role in physical correction. When the anisotropy index When the display shows that the network points are under mechanical tension. The function will output a bias vector, through Correcting the latent coordinate vector The physical significance of this correction lies in the subtle differences in spectral absorption characteristics between dot gain caused by mechanical deformation and optical dot gain caused by ink spreading. This difference is typically manifested as enhanced nonlinear spectral absorption due to photobleeding. By introducing topological constraints, the system forces the latent coordinate vector... Instead of blindly matching based solely on RGB values, the offset is directed to a manifold subspace that conforms to the current physical deformation state.

[0046] Finally, the geometrically constrained spectral manifold reconstruction module 133 reconstructs the spectral manifold based on the corrected latent coordinate vector. and manifold basis matrix Perform inverse projection to generate a full-band spectral reflectance vector. : ; Output It is a high-dimensional vector per pixel or per region that not only matches the source image in terms of chromaticity values, but also conforms to the current microscopic dot physical structure in terms of spectral distribution, providing accurate chemical fingerprint information for subsequent decoupling control of ink volume and pressure.

[0047] See attached document Figure 4 The adaptive feedback calibration module 134 is configured to establish a cross-scale error correction loop, utilizing macroscopic spectral physical consistency to verify and optimize microscopic image processing parameters. This module is based on a technical insight: under correct physical conditions, there should be a strict mapping correspondence between microscopic dot morphology and macroscopic spectral response; any reconstruction result deviating from the preset manifold distribution indicates parameter deviation in the microscopic feature extraction stage.

[0048] The adaptive feedback calibration module 134 first performs a spectral confidence assessment. The module receives the reconstructed spectral vector output by the geometrically constrained spectral manifold reconstruction module 133. Because printing standards (such as ISO 12647) specify the standard spectral response range for color zones (such as solid, 25%, 50%, and 75% tones), these responses appear as several discretely distributed manifold clusters in the multidimensional spectral space. The module first identifies the target manifold cluster to which the current pixel belongs based on the color coordinates of the macroscopic patch, and then retrieves preset statistical parameters for that cluster, including the cluster center vector. Covariance Matrix .

[0049] Subsequently, the module calculates and reconstructs the spectral vector. The Mahalanobis distance relative to the target manifold cluster is defined as the spectral confidence residual.

[0050] ; The residual The physical reliability of the current reconstruction result is quantified. If... Exceeding the preset tolerance value indicates that the topological features of the dots extracted based on the current segmentation threshold cannot reasonably explain the observed macroscopic color (for example, the dot area calculation value is artificially high due to the halo beams at the dot edges, which forces the reconstructed spectrum to shift to darker tones, thus deviating from the normal manifold distribution area of ​​the color region).

[0051] To correct this bias, the adaptive feedback calibration module 134 performs gradient backpropagation. Its goal is to find an optimal micro-segmentation threshold. This makes the spectral confidence residual Minimization. Since standard image binarization segmentation is mathematically a non-differentiable step function, the module introduces a continuously differentiable smooth approximation function (such as the sigmoid function form) in the backpropagation computation path to simulate the gradient response of the thresholding segmentation process, thereby constructing a complete gradient chain from the spectral residual layer to the microscopic segmentation threshold layer. The module calculates the residual. For the segmentation threshold The partial derivatives are used to dynamically update the threshold using the gradient descent algorithm: ; in, The threshold for the current iteration step. To adaptively adjust the step size.

[0052] In each threshold update Subsequently, the system triggers the micro-dot topology analysis module 132 and the geometrically constrained spectral manifold reconstruction module 133 to re-execute the feature extraction and spectral reconstruction steps using the new threshold. This closed-loop iterative process continues until the spectral confidence residual... The algorithm converges to a minimum value or reaches the preset maximum number of iterations. At this point, the adaptive feedback calibration module 134 locks the current optimal threshold and outputs the network topology feature vector and spectral reflectance vector after physical consistency verification, as precise input data for subsequent control calculations.

[0053] The heterogeneous feature tensor encoding and inverse control module 135, as the system's decision-making center, is configured to convert the feature data verified by the preceding modules into specific control instructions for the printing press's actuators. The core logic of this module lies in decoupling the mixed effects of chemical ink volume changes and physical dot amplification during the printing process through mathematical means, thereby solving the problem of multivariate coupled control.

[0054] The heterogeneous feature tensor encoding and inverse control module 135 first receives the optimal spectral reflectance vector from the adaptive feedback calibration module 134. and the optimal network topology feature vector The module contains two parallel feature encoders: a chemical feature encoder and a physical feature encoder.

[0055] The chemical feature encoder maps high-dimensional spectral vectors to low-dimensional chemical state vectors. This vector represents the absorption and scattering coefficients of the current pixel at different wavelengths, essentially reflecting the ink layer thickness and ink overprinting state on the paper surface. The physical feature encoder maps the dot topology feature vector into a low-dimensional physical state vector. This vector characterizes the edge spread rate and geometric deformation mode of the dots, essentially reflecting the contact force distribution and mechanical slippage state between the impression cylinder and the blanket.

[0056] To distinguish between the two states in the control solution, heterogeneous feature tensor encoding and inverse control module 135 construct a joint state tensor. The module performs outer product operations or high-dimensional tensor product operations on eigenvectors to generate a tensor space that can simultaneously capture the interaction between ink volume and pressure. During this process, the module introduces orthogonalization constraints to enforce... and The feature orientations within the potential space should be as perpendicular as possible.

[0057] The technical significance of this orthogonalization process lies in the fact that when printed materials exhibit color differences (e.g., an overall darker appearance), the system can use tensors to... The decomposition clearly determines whether the deviation is due to excessive ink (large component projected on the chemical axis) or excessive pressure causing dot spread (large component projected on the physical axis), thus avoiding system oscillation caused by the simultaneous adjustment of ink keys and pressure due to fuzzy attribution in traditional PID controllers.

[0058] Based on joint state tensor The module then enters the inverse control solution phase. This phase relies on the pre-calibrated and stored rheological coupling matrix. This matrix describes the complex nonlinear interference relationships between the various color groups of the printing press, including: Wet-on-wet effect: The nonlinear function of the adhesion rate of the subsequent ink on the undried preceding ink layer as a function of the thickness of the preceding ink layer.

[0059] Mechanical transmission effect: The coefficient of influence of adjusting the roller pressure of a certain color group on paper tension and the registration accuracy of subsequent color groups.

[0060] The heterogeneous feature tensor encoding and inverse control module 135 constructs a multi-objective optimization function to minimize color difference. The goal is to minimize the geometric deformation of the dots, using the rheological coupling matrix. Given the system transfer function, solve for the optimal control vector. The solution to the equation is expressed as: ; in, The prediction deviation of the network topology is... This is a regularization constraint term for the control quantity, used to limit the adjustment range of the actuator and prevent mechanical overload.

[0061] Finally, the heterogeneous feature tensor encoding and inverse control module 135 decomposes the solved control vector u into two independent sets of command signals: the first set is the ink path adjustment command, which includes the specific opening value of the ink key motors for each CMYK color group, and is directly sent to the ink key actuator 121 to correct the spectral absorption deviation in the chemical dimension. The second set is the printing adjustment command, which includes the pressure adjustment step size of the servo motors on both sides of each color group roller, and is directly sent to the pressure adjustment mechanism 122 to correct the dot morphology distortion in the physical dimension.

[0062] Through the above process, the system achieves a precise mapping from single image acquisition data to multi-dimensional physical / chemical control commands.

[0063] To ensure accurate spectral reconstruction and inverse control during online operation, parameter calibration must be performed beforehand for the printing press status, paper type, and ink combination. This stage mainly involves the construction of the spectral manifold space and parameter identification of the rheological coupling model.

[0064] In the spectral manifold construction phase, the system first controls the printing equipment 120 to output a standard test plate (e.g., EC12002 or IT8.7 / 4 standard color target) containing samples of the entire color gamut. This test plate contains thousands of color patches with different CMYK combinations, covering the complete tonal range from highlights to shadows. Subsequently, a high-precision offline spectrophotometer is used to measure each color patch on the test plate individually to obtain its standard spectral reflectance vector set. Meanwhile, the image acquisition device 110 acquires images of the test version under standard lighting conditions, and extracts the RGB response values ​​and microscopic dot topology features of the corresponding color blocks.

[0065] Data processing server 130 utilizes the acquired set of spectral reflectance vectors Manifold learning training is performed. The server executes principal component analysis (PCA) or locally linear embedding algorithms to extract the frontier components from the high-dimensional spectral data. The principal eigenvectors form an orthogonal manifold basis matrix. The basis matrix defines the physically existing spectral variation subspace under these printing conditions. Furthermore, the server divides the spectral data into several local manifold clusters (e.g., cyan, magenta, overprinting areas, etc.) based on chromaticity coordinates, and calculates the statistical distribution parameters for each cluster, including the spectral vector at the cluster center. and spectral covariance matrix These statistical parameters are stored in a database and used as a benchmark for the subsequent adaptive feedback calibration module 134 to calculate the spectral confidence residuals.

[0066] During the parameter identification phase of the rheological coupling model, the system performs a series of excitation and response experiments to establish the transfer function between the control and state variables. The system loads a dedicated rheological test plate, which includes control strips for detecting overprint rate and Siemens stars for detecting dot deformation.

[0067] The system first maintains constant pressure and applies a step signal excitation to the ink key actuator 121, that is, independently increasing or decreasing the ink key opening of each channel (C, M, Y, K) by a preset step size. The data processing server 130 synchronously records the resulting change in the spectral reflectance vector and calculates the ink key spectral sensitivity matrix through linear regression fitting.

[0068] Subsequently, the system maintains a constant ink key opening and applies a step signal excitation to the pressure regulating mechanism 122, thereby fine-tuning the roller gap pressure. The data processing server 130 records the resulting changes in the dot anisotropy index and area coverage, and calculates the pressure dot morphology sensitivity matrix.

[0069] Specifically, to calibrate the nonlinear interference of multi-color ink wet-on-wet printing, the system performs coupling tests in the multi-color overprinting area. For example, while keeping the amount of subsequent ink (e.g., magenta) constant, the amount of ink for the preceding ink (e.g., cyan) is varied, and the adhesion rate of magenta ink on the cyan ink layer surface is observed. Based on these observation data, the system identifies the coupling coefficients and integrates all the aforementioned sensitivity sub-matrices and coupling coefficients to construct the final rheological coupling matrix. This matrix quantifies the combined effects of ink path adjustment and printing pressure adjustment on the final print quality, and serves as the mathematical basis for subsequent inverse calculations.

[0070] See attached document Figure 5 After the system's preset parameters and calibration are completed, the data processing server 130 enters online operation mode to monitor and control the printing production process in real time. This stage not only involves unidirectional data flow, but also, more importantly, includes an adaptive iterative optimization process for micro-parameters.

[0071] The online detection and correction phase begins with the acquisition of a real-time image stream. The image acquisition device 110 continuously scans the surface of the high-speed moving substrate and outputs raw image data frame by frame. The data processing server 130 locks the key measurement and control area within each sampling cycle and, according to the logic of the aforementioned image preprocessing module 131, splits the data stream into a microscopic patch stream with spatial resolution and a macroscopic patch stream with color statistical characteristics.

[0072] Subsequently, the system enters the feature extraction and adaptive cross-calibration sub-process. This is a complete iterative calculation process, not a single forward propagation. First, the system uses the default or inherited segmentation threshold from the previous frame to perform anisotropic analysis on the microscopic tiles of the current frame, generating an initial dot topology feature vector. Then, the system uses this initial topology feature vector as a physical constraint term, substitutes it into the spectral manifold reconstruction model, and derives the current pseudo-spectral reflectance vector from the macroscopic tiles.

[0073] At this point, the system immediately initiates a confidence verification mechanism. It calculates the confidence residual of the pseudo-spectral vector relative to the target manifold cluster. If this residual value is lower than a preset convergence tolerance... If the system determines that the current microscopic segmentation parameters and macroscopic color representation are physically consistent, it will directly output the current feature vector.

[0074] Conversely, if the residual value exceeds the tolerance, the system determines that there is a parameter mismatch (for example, due to the emulsification of dampening solution causing blurred dot edges, the geometric segmentation based on a fixed threshold cannot accurately reflect the actual ink coverage). In this case, the system keeps the current input image data unchanged, uses the aforementioned gradient backpropagation algorithm to calculate the corrected gradient for the segmentation threshold based on the spectral residual, and updates the threshold. The system then uses the updated threshold to re-binarize, extract features, and reconstruct the spectrum of the same micro-patch. This extraction-reconstruction-verification-correction loop is repeated within a millisecond-level time window until the residual converges, thereby obtaining the optimal dot topology feature vector verified by physical and logical consistency. and the optimal spectral reflectance vector .

[0075] After acquiring high-precision feature data, the system enters the heterogeneous tensor decoupling and control decision sub-process. The data processing server 130 will then process the verified data... and Mapping to the latent space, a joint state tensor is constructed. In this step, the system utilizes the orthogonal decomposition property of the tensor to perform attribution analysis on the current quality deviation. For example, when a color difference in a certain area exceeds the standard, the system determines the cause through tensor projection: If the deviation is mainly projected onto the chemical characteristic axis, it indicates that the main problem is insufficient or excessive ink supply. If the deviation is mainly projected onto the physical feature axis, it indicates that the abnormal imprinting pressure is the main cause of mechanical amplification or dispersion of the dots.

[0076] Based on the above attribution results, the system calls the pre-calibrated rheological coupling matrix for inverse solution. Under the constraints of ink path response lag time and mechanical adjustment accuracy, the solver calculates the optimal control vector for the current control cycle.

[0077] The system decomposes the vector into discrete action commands: the calculated ink key opening increment signal is converted into a pulse sequence and sent to the ink key actuator 121 to drive the servo motor to adjust the ink fountain blade gap; the calculated pressure adjustment signal is sent to the pressure adjustment mechanism 122 to drive the roller clutch mechanism to fine-tune the printing distance.

[0078] Through the above process, the system completes a closed loop in each control cycle, from microscopic physical morphology perception to macroscopic chemical composition analysis, and then to precise adjustment of the mechanical actuator, achieving synchronous control of color consistency and dot integrity in a high-speed printing environment.

Claims

1. A deep learning-based intelligent detection and correction system for color difference in printed materials, characterized in that, include: An image acquisition device used to acquire the source RGB image of printed matter output by printing equipment; Printing equipment, comprising an ink key actuator for adjusting the ink supply and a pressure regulating mechanism for adjusting the roller contact pressure; A data processing server, communicatively connected to the image acquisition device and the printing equipment, comprises: An image preprocessing module is used to decompose the source RGB image into micro-patterns containing halftone dot structures and macro-patterns containing average color data. The micro-dot topology analysis module is used to extract features from the micro-dots based on a segmentation threshold and output a dot topology feature vector. A geometrically constrained spectral manifold reconstruction module is used to reconstruct pixel-level spectral reflectance vectors based on the dot topological feature vectors and the macroscopic tiles. An adaptive feedback calibration module is used to adjust the segmentation threshold used by the micro-dot topology analysis module in reverse based on the physical confidence of the spectral reflectance vector, thereby obtaining a calibrated dot topology feature vector; The heterogeneous feature tensor encoding and inverse control module is used to generate control commands based on the calibrated dot topology feature vector and the spectral reflectance vector. The control commands include ink path adjustment commands sent to the ink key actuator and imprint adjustment commands sent to the pressure regulating mechanism.

2. The intelligent detection and correction system for color difference in printed materials based on deep learning as described in claim 1, characterized in that, The image preprocessing module is specifically configured as follows: Geometric registration and multi-scale Gaussian pyramid decomposition are performed on the source RGB image; The micro-pattern is obtained by extracting an image region containing a complete dot period from the bottom image of the pyramid and performing illumination uniformity correction. The macroscopic patch is obtained by performing global pixel integration on the microscopic patch, and the macroscopic patch is a three-dimensional color vector without spatial resolution.

3. The intelligent detection and correction system for color difference in printed materials based on deep learning as described in claim 1, characterized in that, The specific configuration of the micro-network topology analysis module is as follows: The gradient vector of each pixel in the microscopic patch is calculated using an anisotropic differential operator; A local structure tensor matrix is ​​constructed based on the gradient vector, and eigenvalue decomposition is performed on the structure tensor matrix. The anisotropy index is calculated using the eigenvalues ​​obtained from the decomposition. The anisotropy index characterizes the degree of mechanical deformation at the edges of the dots. By combining the dot mask generated by the segmentation threshold, the distribution characteristics of the anisotropy index are statistically analyzed to generate the dot topology feature vector.

4. The intelligent detection and correction system for color difference in printed materials based on deep learning according to claim 3, characterized in that, The geometrically constrained spectral manifold reconstruction module pre-stores a printed color gamut manifold basis matrix. The specific configuration of the geometrically constrained spectral manifold reconstruction module is as follows: The topological feature vector of the network points is used as the input to the geometric-physical mapping unit, and the macroscopic tiles are used as the input to the color response mapping unit. The outputs of the geometric physical mapping unit and the color response mapping unit are combined, and the potential coordinate vectors in the manifold space are calculated. The potential coordinate vector is projected onto the printed color gamut manifold basis matrix to reconstruct the spectral reflectance vector.

5. The intelligent detection and correction system for color difference in color printed products based on deep learning according to claim 1, characterized in that, The adaptive feedback calibration module is specifically configured as follows: Identify the target manifold cluster based on the color coordinates of the macroscopic patch, and obtain the center vector and covariance matrix of the target manifold cluster; Calculate the Mahalanobis distance between the spectral reflectance vector and the target manifold cluster, and define the Mahalanobis distance as the spectral confidence residual; When the spectral confidence residual exceeds a preset tolerance, the gradient of the spectral confidence residual with respect to the segmentation threshold is calculated; The gradient is used to update the segmentation threshold, and the micro-dot topology analysis module and the geometrically constrained spectral manifold reconstruction module are triggered to re-execute the calculation using the updated segmentation threshold.

6. The intelligent detection and correction system for color difference in printed materials based on deep learning according to claim 1, characterized in that, The heterogeneous feature tensor encoding and inverse control module is specifically configured as follows: It includes chemical feature encoders and physical feature encoders; The chemical feature encoder is used to map the spectral reflectance vector to a chemical state vector, and the physical feature encoder is used to map the dot topology feature vector to a physical state vector. An outer product operation is performed on the chemical state vector and the physical state vector, and an orthogonalization constraint is applied to generate a joint state tensor. The joint state tensor is used to decouple the mixed effect of ink chemical absorption and dot physical amplification.

7. The intelligent detection and correction system for color difference in color printed products based on deep learning according to claim 6, characterized in that, The heterogeneous feature tensor encoding and inverse control module has a pre-stored rheological coupling matrix, which describes the transfer function of ink bond opening change and roller pressure change on spectrum and dot morphology. The heterogeneous feature tensor encoding and inverse control module solves for the optimal control vector based on the joint state tensor and the rheological coupling matrix through a multi-objective optimization function. The optimal control vector is decomposed into independent ink path adjustment commands and printing adjustment commands.

8. The intelligent detection and correction system for color difference in color printed products based on deep learning according to claim 1, characterized in that, The ink key actuator is used to adjust the local opening of each color group ink fountain according to the ink path adjustment command to correct the spectral absorption deviation in the chemical dimension. The pressure adjustment mechanism is used to adjust the contact pressure between the blanket cylinder and the impression cylinder according to the impression adjustment command to correct the dot morphology distortion in the physical dimension.

9. The intelligent detection and correction system for color difference in printed materials based on deep learning according to claim 4, characterized in that, The printing gamut manifold substrate matrix is ​​pre-calibrated in the following manner: Acquire the image data of the standard test version and the corresponding standard spectral measurement data; Principal component analysis or manifold learning algorithms are performed on the standard spectral measurement data to extract orthogonal feature spectral vectors to form the printing color gamut manifold basis matrix.

10. The intelligent detection and correction system for color difference in printed materials based on deep learning according to claim 1, characterized in that, The image acquisition device is an industrial line scan camera mounted above the paper receiving unit of the printing equipment, and is equipped with a diffuse reflection light source component; The industrial line scan camera has the resolution to resolve the edge structure of a single halftone dot.

Images