A computer vision-based printing defect detection method and system

Abstract

The application relates to the technical field of image detection, and discloses a printing defect detection method and system based on computer vision. The method collects a multi-channel fusion image of a printing surface through a multi-modal imaging device, corrects distortion through a dynamic segmentation module, extracts a detection area mask, and outputs a standard image. The standard image is input into a defect detection model, microscopic dot and macroscopic texture features are extracted, matched with a template library, and a probability distribution graph is generated. Multi-level discrimination is performed according to the graph, defects are quantitatively evaluated, and finally, defect coordinates and type labels with time and space annotations are output. The system comprises multi-source acquisition, geometric correction, feature matching, hierarchical decision and closed-loop control modules, realizes real-time detection and feedback control of printing defects, improves detection precision and efficiency, is suitable for different types of printed matter, and has a wide application prospect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image detection, in particular to a printing defect detection method and system based on computer vision. BACKGROUND

[0002] In the printing industry, the quality of printed matter is of great importance. However, traditional printing defect detection techniques often rely on manual visual inspection, which is not only time-consuming and labor-intensive, but also susceptible to human factors, resulting in low detection accuracy and efficiency. With the continuous development of computer vision technology, more and more researchers have begun to try to apply it to printing defect detection. However, existing computer vision-based printing defect detection techniques still have some problems, such as insufficient detection accuracy and weak adaptability to complex printed matter. Therefore, there is an urgent need for a more efficient and accurate printing defect detection method to meet the needs of the printing industry. SUMMARY

[0003] In view of the above technical deficiencies, the present application aims to provide a printing defect detection method and system based on computer vision, which solves the problems of low detection accuracy and weak adaptability to complex printed matter in the prior art.

[0004] To solve the above technical problems, the present application adopts the following technical solutions:

[0005] In a first aspect, the present application provides a printing defect detection method based on computer vision, which comprises:

[0006] A multi-modal imaging device is used to collect a multi-channel fusion image of the printed surface containing ink physical property parameters, and a multi-spectral composite light source is used for time sequence excitation;

[0007] The multi-channel fusion image of the printed surface is input into a dynamic segmentation module, a preset dot parameter and a registration line correction mechanical transmission distortion are corrected, a detection area mask is extracted through morphological closing operation, and a standard image after geometric alignment is output;

[0008] The standard image after geometric alignment is input into a defect detection model, micro dot features and macro texture features are extracted, cross-scale matching is performed with a template library, and a probability distribution graph containing abnormal area position and confidence information is generated;

[0009] According to the probability distribution graph, multi-level discrimination is performed, in which edge diffusion deviation is quantitatively evaluated at the dot level, and topological structure is analyzed at the pattern level, and finally printed defect coordinates and type labels with space-time labels are output;

[0010] The printing defect coordinates are mapped to a printing unit phase coordinate system, a sorting mechanism is triggered to perform a phase-synchronous rejection operation, and an ink pressure adjustment parameter and a temperature control parameter are associated to generate a closed-loop feedback instruction.

[0011] Preferably, in a possible implementation form of the first aspect, the multi-modal imaging device adopts a set of LED light sources distributed in a ring array, including a multi-spectrum composite light source of visible light band, near-infrared band and short-wave infrared band.

[0012] The imaging unit is composed of a high-resolution area CCD camera and a multi-channel filter wheel, and a multi-channel fusion image containing ink thickness, curing degree and adhesion physical property parameters is constructed by synchronously collecting reflectance, fluorescence intensity and thermal radiation response data.

[0013] Preferably, in a possible implementation form of the first aspect, the dynamic segmentation module specifically comprises:

[0014] A reference grid template is generated based on preset screen dot parameters, and the screen dot parameters include screen line angle, screen dot shape and screen line number;

[0015] The actual position of the overprint reference line is detected by Hough transform, the offset vector of the actual position from the theoretical position is calculated, and affine transformation matrix is used to correct the stretching and rotation distortion caused by mechanical transmission;

[0016] The corrected image is subjected to multi-scale morphological closing operation, a circular structural element is selected to perform inflation and corrosion operations with a radius of 3 pixels and 5 pixels in turn, local noise is eliminated and microscopic pores are filled, and a connected domain complete detection area mask is generated;

[0017] The mask is aligned with the original image by bit operation, and a standard image with eliminated position deviation is output, and a sub-pixel level mapping relationship is established between the pixel coordinate system of the standard image and the encoder signal of the printing unit.

[0018] Preferably, in a possible implementation form of the first aspect, the affine transformation matrix is constructed to satisfy the following relationship:

[0019]

[0020] Wherein 、 is the image pixel coordinate, is the angle deviation of the overprint reference line, 、 is the translation compensation amount;

[0021] The mathematical expression of the morphological closing operation is:

[0022]

[0023] Wherein The output image matrix after morphological closing operation. Given an input image matrix, As a structural element, This indicates an expansion operation. This indicates a corrosion operation.

[0024] Preferably, in one possible implementation of the first aspect, the workflow of the defect detection model includes:

[0025] A dual-branch deep convolutional neural network is used. The first branch extracts micro-dot morphology features through a 3×3 small kernel convolutional layer, capturing edge jaggedness, dot area ratio and ink diffusion coefficient.

[0026] The second branch extracts macroscopic texture features through dilated convolutional layers, analyzing color difference distribution, moiré patterns, and overprinting misalignment.

[0027] Multi-scale features are input into the cross-modal attention mechanism module and similarity matching is performed with standard samples in the template library. The template library is indexed hierarchically according to printing color plate and printing material type.

[0028] The probability distribution map of the outlier region is reconstructed using a deconvolution layer, and its confidence value is normalized by the Softmax function to... Interval.

[0029] Preferably, in one possible implementation of the first aspect, the similarity matching employs the following metric:

[0030] Microscopic feature similarity:

[0031]

[0032] Macro-feature similarity:

[0033]

[0034] in For the test sample Each micro-eigenvalue, For the first in the template library A standard microscopic eigenvalue, For microscopic features, For the macroscopic feature vector of the test sample, These are standard macroscopic feature vectors from the template library. The maximum Euclidean distance between the macroscopic feature vectors of all samples in the template library;

[0035] The final probability value is determined by the weighted fusion formula. Sure, and These are the weighting coefficients. ,when If the probability exceeds a preset threshold, it is considered an abnormal region.

[0036] Preferably, in one possible implementation of the first aspect, the multi-level discrimination specifically includes:

[0037] At the dot level, a contour tracking algorithm is executed on abnormal areas in the probability distribution map to extract curvature extreme points along the dot edges, calculate the deviation between the actual contour and the theoretical Gaussian diffusion curve, and mark it as an edge diffusion defect when the deviation exceeds the deviation threshold.

[0038] At the pattern level, a topological skeleton graph of the abnormal region is constructed. Graph theory algorithms are used to analyze the node connectivity, number of loops and subgraph isomorphism. When a critical path break or adjacency mismatch is detected, it is marked as a structural defect.

[0039] Spatiotemporal information is integrated to generate spatiotemporal annotation data that includes the phase angle of the printing press, defect size, and type code.

[0040] Preferably, in one possible implementation of the first aspect, the quantification formula for the edge diffusion deviation is:

[0041]

[0042] in Let be the curvature function of the actual contour. Let be the curvature function of the ideal Gaussian diffusion curve. For the first The arc length of each contour segment, This represents the total number of outline segments;

[0043] The topology analysis employs the adjacency matrix eigenvalue comparison method:

[0044] The eigenvalues ​​of the adjacency matrix of the standard pattern are The measured characteristic value of the abnormal area is Then the topological similarity is calculated by the following formula:

[0045]

[0046] in For the first Eigenvalues ​​of the adjacency matrix of a standard pattern , For the first Measured feature values ​​of anomaly regions , The number of eigenvalues ​​in the adjacency matrix, when If the similarity is less than the topological similarity threshold, it is considered a structural defect.

[0047] Preferably, in a possible implementation form of the first aspect, the mapping of the print defect coordinates to the printing unit phase coordinate system comprises resolving a current printing unit phase angle from the encoder signals converting the defect coordinates to a rejection trigger time in the phase coordinate system wherein is a current phase angle of the printing unit, is a cylinder angular velocity, is a delay compensation.

[0048] In a second aspect, the present application provides a computer vision-based print defect detection system, comprising:

[0049] a multi-source acquisition module for acquiring a multi-channel fusion image of a printed surface containing ink physical property parameters through a multi-modal imaging device, the multi-modal imaging device being configured to perform time-series excitation using a multi-spectral composite light source;

[0050] a geometric correction module for inputting the multi-channel fusion image of the printed surface into a dynamic segmentation module, correcting mechanical transmission distortion based on preset dot parameters and overprint reference lines, extracting a detection area mask through morphological closing operation, and outputting a geometrically aligned standard image;

[0051] a feature matching module for inputting the geometrically aligned standard image into a defect detection model, extracting microscopic dot features and macroscopic texture features, performing cross-scale matching with a template library, and generating a probability distribution map containing abnormal area positions and confidence information;

[0052] a hierarchical decision-making module for performing multi-level discrimination based on the probability distribution map, wherein edge diffusion deviation is quantitatively evaluated at the dot level, and topological structure is analyzed at the pattern level, and finally, print defect coordinates and type labels with spatiotemporal annotations are outputted;

[0053] a closed-loop control module for mapping the print defect coordinates to a printing unit phase coordinate system, triggering a sorting mechanism to perform a phase-synchronous rejection operation, and associating ink pressure adjustment parameters and temperature control parameters to generate a closed-loop feedback instruction.

[0054] The present application has the advantages that: the multi-channel fusion image of the printed surface is acquired through the multi-modal imaging device, and the dynamic segmentation module, the defect detection model, the hierarchical decision-making module, and the closed-loop control module are used to realize efficient and accurate detection of print defects.

[0055] The method not only improves the detection accuracy and efficiency, but also has strong adaptability and robustness, and can be applied to different types of printed matter and printing processes.

[0056] In addition, the method can also realize real-time image detection and feedback control of printing defects, further enhancing the application value of image detection technology in printing quality control and production optimization. BRIEF DESCRIPTION OF DRAWINGS

[0057] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0058] Figure 1 A computer vision-based printing defect detection method flowchart is provided for the present application.

[0059] Figure 2 A computer vision-based printing defect detection system structure diagram is provided for the present application.

[0060] The reference signs are explained as follows: 1-multisource acquisition module, 2-geometric correction module, 3-feature matching module, 4-hierarchical decision module, 5-closed-loop control module. DETAILED DESCRIPTION

[0061] The technical solutions in the embodiments of the present application will be described clearly and completely in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0062] Embodiment one: as shown, the present application provides a computer vision-based printing defect detection method, comprising: Figure 1

[0063] A multi-modal imaging device is used to acquire a multi-channel fusion image of a printing surface containing ink physical property parameters, and the multi-modal imaging device uses a multi-spectral composite light source for time sequence excitation.

[0064] In the present embodiment, the multi-modal imaging device uses a LED light source group distributed in a ring array, the spectral coverage range of which includes a visible light band (400-700 nm), a near-infrared band (700-1100 nm) and a short-wave infrared band (1100-2500 nm), and a preset time sequence switching of different band light source excitation modes is realized through a time-sharing trigger control module. The light source group is distributed in a ring array, ensuring that the light uniformity of each area of the printing surface reaches ​The irradiance deviation range of the light source is within 10%, and the intensity of each waveband light source is independently controlled by a pulse width modulation technique.

[0065] The imaging unit is composed of a high-resolution area array CCD camera and a multi-channel filter wheel, in which the pixel size of the CCD camera is equipped with a 16-bit analog-to-digital converter to realize high dynamic range imaging. The filter wheel is built-in eight sets of narrow-band filters, and the center wavelengths correspond to the key spectral feature points of ink property detection, including ink absorption peaks (such as 530 nm for magenta ink and 650 nm for cyan ink) and printing material fluorescence excitation wavelengths. During image acquisition, the imaging unit and the light source group are triggered synchronously by an encoder signal to capture reflectance data, fluorescence intensity data and thermal radiation response data with a time resolution of 5 ms. Among them, the reflectance data is used to calculate the optical density of the ink, the fluorescence intensity data is used to analyze the curing degree of the ink through the Stokes shift, and the thermal radiation response data is used to inverse the ink adhesion force parameter based on the Stefan-Boltzmann law.

[0066] Through multi-source data fusion, the three-channel physical property parameters are mapped to a unified pixel coordinate system to construct a multi-channel fusion image containing ink thickness, curing degree and adhesion. The fusion process adopts a weighted superposition method, in which the ink thickness is calculated by the near-infrared band reflectance and the Mie scattering model, the curing degree is quantified by the fluorescence lifetime imaging technology, and the adhesion is derived according to the empirical relationship between the thermal radiation relaxation time and the ink layer peeling strength.

[0067] The multi-channel fusion image of the printed surface is input into a dynamic segmentation module, the mechanical transmission distortion is corrected based on the preset dot parameters and the overprint reference line, the detection area mask is extracted by morphological closing operation, and the standard image after geometric alignment is output.

[0068] In this embodiment, the dynamic segmentation module establishes a geometric correction reference based on the preset printing process parameters, in which the dot parameters include the screen angle, the dot shape (circle / square / diamond) and the screen line number (such as 175 LPI). Based on these parameters, a reference grid template is generated, which is composed of periodically arranged dot units, and the size of each unit is determined by the screen line number, and the grid spacing is

[0069]

[0070] wherein , is the image pixel coordinate, is the overprint reference line angle deviation, , is the translation compensation, the matrix is used to correct the geometric distortion and rotation deviation of the input image caused by the mechanical transmission system through the bilinear interpolation algorithm. The corrected image enters the morphological processing stage, and a multi-scale closed operation is used to eliminate local noise and fill micro pores. The mathematical expression of the morphological closed operation is:

[0071]

[0072] wherein is the output image matrix after morphological closed operation, is the input image matrix, is the structure element, represents the dilation operation, represents the erosion operation. In this embodiment, a circular structure element with a radius of 3 pixels is first selected to perform the dilation operation to fill micro pores smaller than 6 pixels; then the same structure element is used for erosion operation to restore the main outline while eliminating isolated noise points. To further enhance the integrity of the connected domain, a second closed operation is performed using a structure element with a radius of 5 pixels, and a detection area mask is finally generated. The mask is accurately aligned with the original image through bit operation, so that the pixel coordinate system of the output standard image is mapped to the encoder signal of the printing unit, and the mapping accuracy is controlled within phase angle range by the phase locking algorithm.

[0073] In the specific implementation process, the system acquires the position information of the printing material by real-time acquisition of the encoder pulse signal, and dynamically calibrates the image coordinate system origin and the printing unit phase origin. In the calibrated coordinate system, each pixel point corresponds to a specific mechanical phase angle (wherein n is the encoder count value, and N is the total number of pulses corresponding to the circumference of the cylinder), realizing sub-pixel level synchronization between image space and mechanical space. The residual error of the geometric distortion of the standard image after dynamic segmentation processing is controlled within , meeting the accuracy requirements of subsequent defect detection. The standard image data output by this module is stored in 16-bit grayscale format, and is accompanied by metadata containing dot parameters, correction matrix coefficients and time stamps, providing complete geometric feature reference for subsequent processing.

[0074] The standard image after geometric alignment is input into the defect detection model, the micro dot features and macro texture features are extracted, and the probability distribution graph containing abnormal area position and confidence information is generated through cross-scale matching with the template library.

[0075] In this embodiment, the defect detection model adopts a dual-branch deep convolutional neural network architecture to realize parallel extraction and cross-scale fusion of microscopic features and macroscopic textures. The first branch is dedicated to capturing the microscopic morphological features of the printing dots. Its input layer receives the standard image after geometric correction, which is input into the feature extraction network composed of five cascaded 3x3 small kernel convolutional layers after normalization processing. Each convolutional layer uses a ReLU activation function and adds a batch normalization operation. The first layer has a step size of 1 to preserve detailed information, and the subsequent layers gradually increase the receptive field. Through this branch, three key features can be extracted: edge jaggedness, dot area ratio, and ink diffusion coefficient. The edge jaggedness is quantified by the kurtosis value of the contour gradient histogram, reflecting the smoothness of the dot edge; the dot area ratio is calculated by the effective printing area ratio after pixel binarization; and the ink diffusion coefficient is obtained based on the optical density decay curve fitting from the dot boundary to the center, with the mathematical expression being where is the radial light intensity distribution, is the diffusion attenuation factor.

[0076] The second branch is designed as a dilated convolutional network to analyze the macroscopic texture features of the printed pattern. This branch uses three groups of dilated convolutional layers with dilation rates of 2, 4, and 6 to build a multi-scale feature pyramid, and integrates semantic information from different levels through skip connections. Macroscopic feature extraction focuses on three dimensions: color difference distribution, Moire fringe, and overprint misregistration. The color difference distribution is represented by the standard deviation and skewness values of the a and b channels in the CIE-Lab color space; the Moire fringe detection uses two-dimensional Fourier transform to extract the frequency domain abnormal peak value, and its intensity is quantified by the energy ratio of the fundamental component ; and the overprint misregistration is calculated by cross-correlation matching of the four-color registration marks, and the offset vector corresponding to the maximum cross-correlation value is the misregistration.

[0077] The multi-scale features extracted by the two branches are fused and enhanced through a cross-modal attention mechanism. This module first concatenates the microscopic feature map and the macroscopic feature map at the channel level, and then inputs them into the multi-head self-attention layer after dimension reduction by 1x1 convolution. The attention weight calculation uses the scaled dot product formula , where the query matrix is generated by projecting the microscopic features, and the key matrix and the value matrix come from the macroscopic features. After parallel processing by eight attention heads, the output enhanced feature vector retains both microscopic details and macroscopic context information.

[0078] Feature matching stage, the system will be fused after the multi-modal features and template library in the standard sample similarity measurement. Template library according to the printing color plate type, printing materials (copper plate paper / offset paper / special paper) and printing process parameters to establish a three-dimensional hierarchical index structure, support based on hash code fast retrieval. For test sample The first micro feature value , with the similarity calculation of the first Standard sample In the template library:

[0079] Micro feature similarity calculation is based on cosine similarity:

[0080] Where Is the micro feature dimension.

[0081] Macro feature similarity uses normalized Euclidean distance measurement:

[0082] Where Is the test sample macro feature vector, Is the standard macro feature vector in the template library, Is the maximum Euclidean distance of all sample macro feature vectors in the template library.

[0083] The final matching probability value is determined by the weighted fusion formula , where the weight coefficient , Is determined by grid search optimization, which satisfies the constraint condition When the fusion probability value exceeds the preset probability threshold 0.85, it is determined that there is an anomaly in the region. The probability distribution graph reconstruction process is realized by using the deconvolution network, which contains three levels of transposed convolution layers, each equipped with LeakyReLU activation function and Dropout regularization, and finally the confidence is normalized to The interval by the Softmax function, generating a probability heat map consistent with the spatial resolution of the input image.

[0084] The model training adopts a two-stage optimization strategy: first, pre-training on the ImageNet dataset to obtain the basic feature extraction capability; then fine-tuning on the self-built printing defect dataset, which contains 35,000 labeled images of 12 common defects. The loss function is designed as a weighted combination of Focal Loss and Dice loss:

[0085]

[0086] Where , . The training process uses the Adam optimizer, and the initial learning rate is set to , decay 0.5 times per 10 epochs, and converge after 150 epochs of training.

[0087] A multi-level discrimination is performed according to the probability distribution map, in which the edge diffusion deviation is quantitatively evaluated at the dot level, and the topological structure is analyzed at the pattern level, and finally the printing defect coordinates and type labels with space-time labels are output.

[0088] In the present embodiment, the multi-level discrimination process realizes accurate defect classification through the cooperative analysis of spatial features and topological relationships. For the abnormal regions in the probability distribution map with a confidence exceeding a threshold value, the system adopts a hierarchical progressive detection strategy: focusing on the morphological distortion of a single dot at a microscopic scale, and analyzing the topological integrity of the pattern structure at a macroscopic scale. For the analysis at the dot level, the system first performs a contour tracking algorithm to extract the closed contour of the abnormal region, and uses the Douglas-Peucker algorithm to discretize the contour into a piecewise linear approximation curve, and samples 1000 points along the curve arc length parameter s. The curvature value of each sampling point is calculated by the three-point method, that is, the reciprocal of the local circular arc radius is fitted using the coordinates of the adjacent three points as the curvature estimation value, and the calculation formula is , wherein , is the difference in coordinates of adjacent points. The curvature distribution curve of the actual contour is compared with the ideal curvature curve generated by the theoretical Gaussian diffusion model point by point, wherein is the center position of the dot, which is determined by the printing process parameters. The edge diffusion deviation quantifies the morphological difference between the actual and ideal contours in the form of integration:

[0089]

[0090] , wherein is the number of segmented contour segments, is the arc length of the th contour segment. The system presets a dynamic deviation threshold ( , which is the theoretical radius of the dot), and when , it is determined as an edge diffusion defect, triggering the ink viscosity adjustment instruction.

[0091] In the topological structure analysis at the pattern level, the system first performs skeletonization processing on the abnormal region, and uses a thinning algorithm to iteratively remove boundary pixels until a single-pixel-width topological skeleton is generated. An adjacency matrix is constructed based on the skeleton graph, wherein the number of nodes is determined by the total number of skeleton branch points and end points, and the matrix element This indicates a direct connection between nodes i and j; otherwise, it is 0. This is achieved by solving the characteristic equation of the matrix. Obtain the set of eigenvalues This characteristic spectrum uniquely characterizes the connectivity, number of loops, and subgraph distribution of the topological structure. The set of characteristic values ​​for the measured anomaly region. With standard patterns in the template library Perform similarity matching and calculate their topological similarity:

[0092]

[0093] in For the first Eigenvalues ​​of the adjacency matrix of a standard pattern , For the first Measured feature values ​​of anomaly regions , The number of eigenvalues ​​in the adjacency matrix is ​​used. When Stopo is less than the topological similarity threshold of 0.82, a structural defect is identified. The system automatically retrieves defect cases with similar topological features from historical data and generates optimization suggestions based on the current printing press's ink key opening and roller pressure parameters. For detected critical path breaks (such as isolated nodes in the skeleton graph) or adjacency mismatches (eigenvalue deviations exceeding 3 standard deviations), the system classifies them as misregistration or mechanical damage defects, triggering a forced rejection instruction from the sorting mechanism.

[0094] The spatiotemporal annotation data includes phase angle, defect envelope rectangle size and type code, integrates the phase synchronization information of the printing press and the geometric features of the defect, and finally outputs the coordinates and type labels of the printing defect with spatiotemporal annotation.

[0095] The coordinates of printing defects are mapped to the phase coordinate system of the printing press, triggering the sorting mechanism to perform a phase-synchronous rejection operation, and generating a closed-loop feedback command by associating the ink pressure adjustment parameters and temperature control parameters.

[0096] In this embodiment, a precise spatiotemporal correlation model between the spatial location of printing defects and the mechanical motion of the printing press is constructed to achieve phase-synchronized control of defect location and rejection actions. The system acquires the angular displacement signal of the main drive shaft of the printing press in real time through a high-speed encoder and analyzes the current roller phase angle. For the coordinates of the detected defects The system maps the image coordinate system to the phase coordinate system of the printing press based on the pre-calibrated transformation relationship between the image coordinate system and the machine coordinate system. Specifically, through the sub-pixel level mapping model established by the geometric correction module, the defect coordinates... Corresponding mechanical phase angle Calculated by the following formula:

[0097]

[0098] wherein is the circumference of the printing cylinder, is the effective detection range in the printing width direction, is the phase angle increment of adjacent detection lines in the cylinder axial direction. The model is established by combining offline calibration and online calibration. In the calibration process, the reference mark points on the standard calibration plate are used for bidirectional coordinate system registration, and finally the phase angle mapping accuracy of is realized.

[0099] In the phase synchronization rejection control, the system dynamically calculates the trigger time t of the rejection action according to the current cylinder angular velocity and the mechanical delay characteristics of the sorting mechanism. The mathematical model of the rejection trigger time is:

[0100]

[0101] wherein is the current phase angle of the printing unit, is the cylinder angular velocity, is the delay compensation amount, , is the inherent mechanical delay of the sorting mechanism (including electromagnetic valve response, cylinder action time), which is determined by step response experiment; is the system processing delay (including image processing, communication transmission time), which is obtained by real-time performance test statistics.

[0102] The generation of closed-loop feedback instructions is based on multi-dimensional analysis of defect type and spatial distribution. By statistically analyzing the occurrence frequency and distribution pattern of each type of defect in the current production batch, combined with the real-time data of the ink area pressure sensor and temperature sensor, an adaptive adjustment model of ink pressure-temperature control parameters is constructed. Specifically, for edge diffusion type defects, the system updates the ink fountain roller pressure set value according to the deviation amount of the ink diffusion coefficient D of the screen dot level from the theoretical threshold value, using proportional-integral control algorithm:

[0103]

[0104] wherein , are the PID control parameters, which are determined by orthogonal experiment method. The adjustment of the temperature control parameter is based on the temperature difference between the ink layer solidification temperature T collected by the thermal imager and the target value , and the heater power adjustment instruction is generated using fuzzy control strategy, and the temperature control accuracy is .

[0105] The system sends adjustment instructions to the printing machine PLC control system through the OPC UA protocol, and records historical adjustment data to optimize the control model. Every 1000 meters of printed matter production, the system automatically performs self-checking and calibration of control parameters: by injecting a preset test pattern, evaluating ink layer uniformity, overprint accuracy and other indicators, dynamically correcting the phase compensation amount and mechanical delay in the mapping model, to ensure that the closed-loop control system maintains stable detection accuracy and feedback response speed throughout its life cycle.

[0106] Embodiment two: as shown, the present application provides a computer vision-based printing defect detection system, comprising: Figure 2

[0107] A multi-source acquisition module 1 acquires a multi-channel fusion image of a printed surface containing ink physical property parameters through a multi-modal imaging device, and the multi-modal imaging device uses a multi-spectral composite light source for time sequence excitation.

[0108] A geometric correction module 2 is used to input the multi-channel fusion image of the printed surface into a dynamic segmentation module, correct mechanical transmission distortion based on preset dot parameters and overprint reference lines, extract a detection area mask through morphological closing operation, and output a geometrically aligned standard image.

[0109] A feature matching module 3 is used to input the geometrically aligned standard image into a defect detection model, extract microscopic dot features and macroscopic texture features, and generate a probability distribution graph containing abnormal area positions and confidence information through cross-scale matching with a template library.

[0110] A hierarchical decision module 4 performs multi-level discrimination according to the probability distribution graph, where the edge diffusion deviation is quantitatively evaluated at the dot level, and the topological structure is analyzed at the pattern level, and finally the printing defect coordinates and type labels with spatiotemporal annotations are output.

[0111] A closed-loop control module 5 is used to map the printing defect coordinates to the phase coordinate system of the printing unit, trigger the sorting mechanism to perform phase-synchronous rejection operation, and associate the ink pressure adjustment parameters with the temperature control parameters to generate closed-loop feedback instructions.

[0112] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application also intends to include these modifications and variations.​

Claims

1. A computer vision-based printed defect detection method, characterized by, The method comprises: Collecting a multi-channel fusion image of a printed surface containing ink physical property parameters by a multi-modal imaging device, which uses a multi-spectral composite light source for time sequence excitation; Inputting the multi-channel fusion image of the printed surface into a dynamic segmentation module, correcting mechanical transmission distortion based on preset dot parameters and overprint reference line, extracting a detection area mask through morphological closing operation, and outputting a standard image after geometric alignment; Inputting the standard image after geometric alignment into a defect detection model, extracting micro dot features and macro texture features, and generating a probability distribution graph containing abnormal area positions and confidence information through cross-scale matching with a template library; The workflow of the defect detection model comprises: Using a double-branch deep convolutional neural network, the first branch extracts micro dot morphological features through a 3x3 small kernel convolution layer to capture edge sawtooth degree, dot area rate and ink diffusion coefficient; The second branch extracts macro texture features through a hollow convolution layer to analyze color difference distribution, moire and overprint misregistration; Inputting multi-scale features into a cross-modal attention mechanism module, and performing similarity matching with standard samples in a template library, which is indexed in layers according to printing color plates and types of printing materials; The similarity matching adopts the following measurement methods: Micro-feature similarity: Macro-feature similarity: wherein is the i-th micro-feature value of the test sample, is the i-th standard micro-feature value of the template library, is the micro-feature dimension, is the macro-feature vector of the test sample, is the macro-feature vector of the template library, is the maximum Euclidean distance of all the macro-feature vectors of the template library.​​ The final probability value is obtained by a weighted fusion formula determined, and is a weight coefficient, when greater than a preset probability threshold value, is determined as an abnormal region. The probability distribution map of the abnormal region is reconstructed by the deconvolution layer, and the confidence value is normalized to interval; Performing multi-level discrimination according to the probability distribution graph, wherein edge diffusion deviation is quantitatively evaluated at the dot level, and topological structure is analyzed at the pattern level, and finally printing defect coordinates and type labels with space-time labels are outputted; Mapping the printing defect coordinates to a printing unit phase coordinate system, triggering a sorting mechanism to perform a phase-synchronous rejection operation, and generating a closed-loop feedback instruction by associating ink pressure adjustment parameters and temperature control parameters.

2. The computer vision based printed defect detection method of claim 1, wherein, The multi-modal imaging device uses a ring array distributed LED light source group, which contains a multi-spectral composite light source of visible light waveband, near-infrared waveband and short-wave infrared waveband; An imaging unit is composed of a high-resolution area array CCD camera and a multi-channel filter wheel, which constructs a multi-channel fusion image containing ink thickness, curing degree and adhesion physical property parameters by synchronously collecting reflectivity, fluorescence intensity and thermal radiation response data.

3. The computer vision based printed defect detection method of claim 1, wherein, The dynamic segmentation module specifically comprises: Generating a reference grid template based on preset dot parameters, which include screen line angle, dot shape and screen line number; Detecting the actual position of the overprint reference line through Hough transform, calculating the offset vector of the actual position from the theoretical position, and correcting the stretching and rotation distortion caused by mechanical transmission by using an affine transformation matrix; Performing multi-scale morphological closing operation on the corrected image, sequentially performing inflation and corrosion operations with a radius of 3 pixels and 5 pixels by using a circular structural element, eliminating local noise and filling micro pores, and generating a complete detection area mask of connected domains; Aligning the mask and the original image through bit operation, and outputting a standard image with eliminated position deviation, which has a sub-pixel level mapping relationship with the encoder signal of the printing unit.

4. The computer vision based printed defect detection method of claim 3, wherein, The construction of the affine transformation matrix satisfies the following relationship: wherein 、 is an image pixel coordinate, is an overprint reference line angle deviation, 、 is a translation compensation amount; The mathematical expression of the morphological closing operation is: wherein is the output image matrix after a morphological closing operation, is the input image matrix, is the structuring element, denotes a dilation operation, denotes an erosion operation.

5. The computer vision based printed defect detection method of claim 1, wherein, The multi-level discrimination specifically comprises: At the dot level, an edge diffusion deviation is calculated by performing a contour tracking algorithm on the abnormal region in the probability distribution map, extracting curvature extreme points along the dot edge, and calculating the deviation between the actual contour and the theoretical Gaussian diffusion curve. When the deviation exceeds a deviation threshold, it is marked as an edge diffusion defect; At the pattern level, a topological skeleton map of the abnormal region is constructed, and the node connectivity, loop number, and subgraph isomorphism are analyzed by a graph theory algorithm. When a critical path is broken or the adjacency relationship is mismatched, it is marked as a structural defect; The spatial-temporal information is integrated to generate spatial-temporal annotation data containing the phase angle of the printing unit, the defect size, and the type code.

6. The computer vision based printed defect detection method of claim 5, wherein, The quantification formula of the edge diffusion deviation is: wherein is the curvature function of the actual profile, is the curvature function of the ideal Gaussian spread curve, is the arc length of the profile segment, is the total number of profile segments; The topological structure analysis adopts an adjacency matrix eigenvalue comparison method: The eigenvalue of the adjacency matrix of the standard pattern is The eigenvalue of the measured abnormal region is The topological similarity is calculated by the following formula: wherein is the eigenvalue of the adjacency matrix of the th standard pattern, , is the eigenvalue of the measured anomaly region of the th standard pattern, , is the number of eigenvalues of the adjacency matrix, and is determined as a structural defect when the topological similarity is less than a threshold value.

7. The computer vision based printed defect detection method of claim 1, wherein, Mapping of print defect coordinates to the printing unit phase coordinate system comprises resolving the current printing unit phase angle from the encoder signals Converting the defect coordinates to a rejection trigger time in the phase coordinate system wherein is the current phase angle of the printing unit, is the roll angular velocity, is the delay compensation.

8. A computer vision based printed defect detection system, characterized in that, The system is based on any one of claims 1 to 7, and the printing defect detection method based on computer vision is realized, comprising: A multi-source acquisition module acquires a multi-channel fusion image of a printed surface containing ink physical property parameters through a multi-modal imaging device, and the multi-modal imaging device uses a multi-spectral composite light source for time sequence excitation; A geometric correction module is used to input the multi-channel fusion image of the printed surface into a dynamic segmentation module, correct mechanical transmission distortion based on preset dot parameters and overprint reference lines, extract a detection area mask through morphological closing operation, and output a geometrically aligned standard image; A feature matching module is used to input the geometrically aligned standard image into a defect detection model, extract micro-dot features and macro-texture features, and generate a probability distribution map containing abnormal region positions and confidence information through cross-scale matching with a template library; A hierarchical decision-making module performs multi-level discrimination according to the probability distribution map, wherein the edge diffusion deviation is quantitatively evaluated at the dot level, and the topological structure is analyzed at the pattern level. Finally, the printing defect coordinates and type labels with spatial-temporal annotations are outputted. A closed-loop control module is used to map the printing defect coordinates to the phase coordinate system of the printing unit, trigger a sorting mechanism to perform a phase-synchronous rejection operation, and generate a closed-loop feedback instruction by associating ink pressure adjustment parameters and temperature control parameters.

Images