Printing process evaluation method and system based on visual analysis

Abstract

The invention relates to the technical field of printing evaluation, in particular to a printing process evaluation method and system based on visual analysis, and the method comprises the steps: extracting a reference visual feature, a quality constraint feature and a scene identification feature based on sample image data, and constructing a multi-dimensional visual reference; comparing the multispectral image data with the multidimensional visual reference to obtain a visual deviation feature set; and constructing a visual feature relation tree based on the visual deviation feature set and texture features in the multispectral image data. According to the method, correlation analysis is carried out on the to-be-selected image sample set and the visual deviation feature set through the defect evaluation model, potential quality problems can be found in time, conduction paths, quality risk levels and process influence degrees of the defects can be calculated, quality risks can be recognized and controlled in time in the production process, unqualified products are reduced, and the production efficiency is improved. And the risks of reworking and loss are reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of printing evaluation, in particular to a printing process evaluation method and system based on visual analysis. BACKGROUND

[0002] Printing is a technology that transfers text, images or other design patterns from a printing plate, such as a printing template or a printing plate, to the surface of paper, cloth, plastic or other materials; printing technology is widely used in publishing, packaging, advertising, art and product identification, etc.

[0003] At present, the traditional method generally does not have real-time dynamic quality correction and process adjustment mechanism in the printing process; usually, it is through the reverse adjustment after checking the finished product, or through the periodic inspection and batch feedback to adjust the production parameters; this leads to the quality fluctuation in the production process cannot be identified and corrected in time, affecting the production efficiency; and, the traditional method usually focuses on checking some specific deviations, such as color deviation, register problem, etc., and lacks systematic quality risk assessment and defect transmission analysis; the traditional quality control method is usually improved for single problem, without considering the potential quality risk from the whole process, so it is easy to ignore the mutual influence between defects.

[0004] In addition, the traditional method usually adjusts based on single process parameters, such as color regulation, register accuracy, etc., and cannot comprehensively consider the mutual relationship and influence between multiple parameters; this method is one-sided, and it is difficult to achieve optimal process adjustment in complex production environment; and, the process adjustment scheme in the traditional method is usually based on experience or static rules, lacking dynamic verification of the feasibility of different adjustment schemes; in some cases, process adjustment may cause new problems or side effects, and the traditional method often lacks effective mechanism to verify the final effect of process adjustment. SUMMARY

[0005] To achieve the above purpose, the present application provides the following technical scheme: a printing process evaluation method based on visual analysis, comprising: obtaining multi-spectral image data of the surface of the printed product and sample image data of the historical qualified printed product; extracting reference visual features, quality constraint features and scene identification features based on the sample image data, and constructing a multi-dimensional visual reference; comparing the multi-spectral image data with the multi-dimensional visual reference to obtain a set of visual deviation features; constructing a visual feature relationship tree based on the set of visual deviation features and the texture features in the multi-spectral image data; analyzing the visual feature path from the core node to the edge node in the visual feature relationship tree, and classifying the multi-spectral image data based on the visual feature path to obtain a plurality of feature analysis areas; For each feature analysis region, the texture coverage and cohesion within the region are calculated, and a multi-layer visual feature indexing framework is constructed by combining the deviation parameters of the corresponding regions in the visual deviation feature set. Based on the multi-layer visual feature indexing framework, the visual feature relationship tree is scanned. During the scanning process, a dynamic quality correction factor for each image region is calculated by combining real-time production parameter data. Based on the dynamic quality correction factor, the feature analysis region is weighted and sorted, and a set of candidate image samples that meet the preset quality standards is selected from the multi-layer visual feature indexing framework. The defect correlation assessment model is used to perform correlation analysis on the candidate image sample set and the corresponding visual deviation feature set to obtain the defect assessment results, wherein the defect assessment results include defect transmission path features, quality risk level parameters and process impact parameters. Based on the defect assessment results, a printing process optimization model is constructed, with the objectives of blocking defect transmission paths, reducing quality risk levels, and optimizing process impact. Combined with preset constraints, an initial set of process adjustment schemes is obtained. The feasibility of the initial set of process adjustment schemes is verified, and printing process adjustment instructions are obtained after verification.

[0006] Preferably, the sample image data includes standard color reference information, precise registration and positioning information, and complete image and text structure information; the multispectral image data includes surface texture information and color distribution information; the visual feature relationship tree includes primary visual nodes and secondary visual nodes that establish intensity correlations; and the texture information and deviation information of the corresponding feature analysis areas are associated at each level in the multi-layer visual feature index framework.

[0007] Preferably, a multidimensional visual benchmark is constructed by extracting benchmark visual features, quality constraint features, and scene identification features based on the sample image data; the multispectral image data is then compared with the multidimensional visual benchmark to obtain a visual deviation feature set, including: The sample image data is classified into multiple typical production scene data columns based on the characteristics of printing material type, printing speed range, and environmental temperature and humidity. The average color density feature, registration position deviation feature, image and text outline clarity feature, and ink layer uniformity feature are extracted from each typical production scenario data column and fused to obtain the baseline visual features. Constraint rules are extracted from the preset printing quality standards and converted into quality constraint features. The benchmark visual features, the quality constraint features, and the scene identification features of the corresponding typical production scenarios are fused to obtain a multi-dimensional visual benchmark. The benchmark visual features and quality constraint features are then matched with the production scene identification of the real-time printed matter. The multispectral image data is compared layer by layer with the matched baseline visual features and quality constraint features. Color consistency deviation, registration accuracy deviation and image integrity deviation are calculated respectively, and the result is a visual deviation feature set after fusion.

[0008] Preferably, a visual feature relationship tree is constructed based on the visual deviation feature set and the texture features in the multispectral image data; the visual feature paths from core nodes to edge nodes in the visual feature relationship tree are parsed, and the multispectral image data is classified into regions based on the visual feature paths to obtain multiple feature analysis regions, including: Color deviation features, registration deviation features, and image-text deviation features are extracted from the visual deviation feature set; texture distribution features, texture density features, and texture direction features are extracted from the multispectral image data. Using various deviation features and texture features as the basic attributes of nodes, we establish primary visual nodes and secondary visual nodes, and construct the strength relationship between nodes based on the degree of influence between features to obtain a visual feature relationship tree; Extract node attributes from the visual feature relationship tree, determine the continuous node groups and the strength of the relationship between nodes on each visual feature path, and calculate the path significance value of each visual feature path; Calculate the path correlation degree of visual feature paths corresponding to any two image regions, select the image regions with the highest sum of path significance values ​​as the initial clustering core, and establish the mapping relationship between image regions and paths. Based on the mapping relationship, the path correlation between the image region to be classified and the initial cluster core is calculated, and the image region to be classified is assigned to the cluster core with the highest path correlation, thus obtaining multiple feature analysis regions.

[0009] Preferably, for each feature analysis region, texture coverage and intra-region cohesion are calculated, and a multi-layer visual feature indexing framework is constructed by combining the deviation parameters of the corresponding regions in the visual deviation feature set, including: For each feature analysis region, the ratio of the sum of the intensity weights of directly related texture features to the sum of the intensity weights of all texture features in the visual feature relationship tree is used as the texture coverage. Calculate the correlation between any two image regions within the feature analysis area, and take the average of all correlations as the cohesion within the area; Extract color deviation parameters, registration deviation parameters, and image-text deviation parameters corresponding to each feature analysis area from the visual deviation feature set, and calculate the comprehensive deviation index of each feature analysis area; The texture coverage, the intra-region cohesion, and the comprehensive deviation index are weighted to determine the index level of each feature analysis region; the separation degree between adjacent index levels is calculated, and adjacent levels with a separation degree less than a preset threshold are fused to obtain a multi-layer visual feature index framework.

[0010] Preferably, based on the multi-layer visual feature indexing framework, the visual feature relationship tree is scanned, and during the scanning process, a dynamic quality correction factor for each image region is calculated in conjunction with real-time production parameter data; based on the dynamic quality correction factor, the feature analysis regions are weighted and sorted, and a candidate image sample set that meets the preset quality standard is selected from the multi-layer visual feature indexing framework, including: Based on the aforementioned multi-layer visual feature indexing framework, and by combining the node hierarchy depth, the number of branches, and the number of associated image regions, the scanning priority of each visual feature node is calculated. Based on the scanning priority, the visual feature relationship tree is scanned in a breadth-first manner to obtain a scanning sequence. From the scanning sequence, the set of image regions corresponding to each node and the associated real-time production parameter records are extracted. Based on the hierarchical structure of the scanning sequence, a set of visual feature paths for each image region is constructed. The basic quality coefficient is calculated based on the quality index of each feature on the path, and the process response coefficient is calculated based on the production parameter records. The basic quality coefficient and the process response coefficient are standardized and then weighted and fused to obtain the dynamic quality correction factor for each image region. The feature analysis regions are then ranked by weight based on the dynamic quality correction factor. Feature analysis regions are selected based on weighted ranking results, image samples are extracted from the selected feature analysis regions, and a set of candidate image samples that meet the requirements is selected by combining preset quality standards.

[0011] Preferably, a pre-trained defect correlation assessment model is used to perform correlation analysis on the candidate image sample set and the corresponding visual deviation feature set to obtain defect assessment results, including: The image features and visual deviation feature sets corresponding to the candidate image sample set are input into the feature preprocessing layer of the defect correlation assessment model to obtain the standard deviation feature matrix; the standard deviation feature matrix is ​​used as the node features and the causal relationship between different deviation features is used as the edge relationship to construct the defect propagation graph; Critical path mining is performed on the defect propagation graph to identify the transmission chain from the initial defect node to the final quality defect node, and the path node density, path dependency strength and path propagation rate are extracted as defect propagation path features. The defect superposition effect value is calculated based on the defect transmission path characteristics, and the quality risk level parameter is generated by combining the tolerance characteristics of the preset quality standard. The visual deviation feature set is associated and mapped with real-time printing process parameters. The sensitivity of process parameter adjustment to the deviation feature correction is analyzed to obtain the process influence parameter. The quality risk level parameter and the process influence parameter are integrated to obtain the defect assessment result.

[0012] Preferably, a printing process optimization model is constructed based on the defect assessment results, aiming to block defect transmission paths, reduce quality risk levels, and optimize process impact. Combined with preset constraints, an initial set of process adjustment schemes is obtained, including: Using the defect propagation path characteristics, quality risk level parameters, and process impact parameters from the defect assessment results as core inputs, a multi-objective function is constructed. Color tolerance constraints, registration accuracy constraints, image integrity constraints, and process parameter range constraints are extracted from the preset printing quality standards and used as constraints for the printing process optimization model. The path parameters and correlation strength parameters in the defect propagation path characteristics are input into the path blocking target, the quality risk level parameter is input into the risk reduction target, and the process influence parameter is input into the influence optimization target to determine the weight coefficient of each target; The multi-objective function is iteratively solved under constraints to obtain multiple sets of candidate process solutions; The multiple sets of candidate process solutions are subjected to multi-objective optimization processing to select candidate process solutions that simultaneously meet the requirements of each optimization objective, thereby obtaining an initial set of process adjustment schemes.

[0013] Preferably, the initial process adjustment scheme set is subjected to feasibility verification processing. After verification, a printing process adjustment instruction is obtained, including: Select one scheme from the initial process adjustment scheme set, and analyze the ink volume adjustment parameters, printing pressure adjustment parameters, and registration accuracy correction parameters therein; The ink volume adjustment parameters are input into the ink zone control simulation module to simulate the ink layer distribution state, and the ink volume adjustment simulation results are obtained. The printing pressure adjustment parameters are superimposed with the real-time pressure data to verify whether they meet the adaptability rules. The registration accuracy correction parameters are input into the registration system simulation module to simulate the correction effect, verify whether the registration accuracy meets the preset standard, and calculate the feasibility index by combining the three results. If the feasibility index reaches the preset threshold, it is marked as a feasible solution; otherwise, the next solution is selected for re-verification, and all feasible solutions are sorted in descending order of feasibility index. The top-ranked feasible solution is selected as the final process adjustment solution, and the final process adjustment solution is converted into a printing process adjustment instruction that can be recognized by the printing equipment.

[0014] A visual analysis-based printing process evaluation system, applicable to the aforementioned visual analysis-based printing process evaluation method, includes: The data acquisition unit is used to acquire multispectral image data of the surface of the printed product and sample image data of historical qualified printed products; extract benchmark visual features, quality constraint features and scene identification features based on the sample image data to construct a multidimensional visual benchmark; and compare the multispectral image data with the multidimensional visual benchmark to obtain a visual deviation feature set. The feature parsing unit is used to construct a visual feature relationship tree based on the visual deviation feature set and the texture features in the multispectral image data; parse the visual feature path from the core node to the edge node in the visual feature relationship tree; and perform region classification on the multispectral image data based on the visual feature path to obtain multiple feature analysis regions. The feature indexing unit is used to calculate the texture coverage and intra-region cohesion for each feature analysis region, and to construct a multi-layer visual feature indexing framework by combining the deviation parameters of the corresponding regions in the visual deviation feature set. The quality correction unit is used to scan the visual feature relationship tree based on the multi-layer visual feature index framework, calculate the dynamic quality correction factor for each image region in combination with real-time production parameter data during the scanning process, sort the feature analysis areas by weight based on the dynamic quality correction factor, and select a set of candidate image samples that meet the preset quality standards from the multi-layer visual feature index framework. The defect assessment unit is used to perform correlation analysis on the candidate image sample set and the corresponding visual deviation feature set based on a pre-trained defect correlation assessment model to obtain defect assessment results, wherein the defect assessment results include defect transmission path features, quality risk level parameters and process impact parameters. The process adjustment unit is used to construct a printing process optimization model based on the defect assessment results, with the objectives of blocking defect transmission paths, reducing quality risk levels, and optimizing process impact. Combined with preset constraints, an initial process adjustment scheme set is obtained. The feasibility of the initial process adjustment scheme set is verified, and a printing process adjustment instruction is obtained after verification.

[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention, by constructing a visual feature relationship tree and a multi-layer visual feature index framework, can classify, analyze, and dynamically correct various deviations in the printing process. This automated optimization mechanism avoids errors in manual inspection, improves production efficiency, and can automatically adjust and optimize the printing process in different production scenarios. This invention uses a defect assessment model to perform correlation analysis between the selected image sample set and the visual deviation feature set, which can promptly identify potential quality problems and calculate the transmission path of defects, quality risk level, and process impact. This helps to identify and control quality risks in a timely manner during the production process, reduce the occurrence of non-conforming products, and reduce the risk of rework and losses. This invention, through feasibility verification, can effectively verify the feasibility of each process adjustment scheme, ensuring that the adopted adjustment scheme can truly improve printing quality and adapt to the current production environment. This precise process adjustment not only optimizes the production process but also can quickly respond to possible changes in production. Furthermore, by constructing a multi-objective optimization model and a multi-level quality assessment framework, the printing process can be flexibly adjusted according to the actual production situation. Optimization objectives include blocking defect transmission paths, reducing quality risks, and optimizing process impact. By comprehensively considering multiple indicators, it can provide a more scientific and efficient adjustment scheme for production, avoiding the limitations of adjusting a single process parameter. Attached Figure Description

[0016] Figure 1 This is a schematic flowchart of the overall method in one embodiment of the present invention; Figure 2 This is a schematic diagram of the overall system architecture in one embodiment of the present invention.

[0017] In the diagram: 1. Data acquisition unit; 2. Feature parsing unit; 3. Feature indexing unit; 4. Quality correction unit; 5. Defect assessment unit; 6. Process adjustment unit. Detailed Implementation

[0018] The technical solutions of 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.

[0019] Example 1, please refer to Figure 1 This invention provides a technical solution: a printing process evaluation method based on visual analysis, comprising: S1. Obtain multispectral image data of the surface of the printed product and sample image data of historical qualified printed products; extract benchmark visual features, quality constraint features and scene identification features based on the sample image data to construct a multidimensional visual benchmark; compare the multispectral image data with the multidimensional visual benchmark to obtain a visual deviation feature set; S2. Construct a visual feature relationship tree based on the visual deviation feature set and texture features in multispectral image data; parse the visual feature path from the core node to the edge node in the visual feature relationship tree, and perform region classification on the multispectral image data based on the visual feature path to obtain multiple feature analysis areas; S3. Calculate the texture coverage and cohesion within each feature analysis region, and construct a multi-layer visual feature indexing framework by combining the deviation parameters of the corresponding regions in the visual deviation feature set. S4. Based on the multi-layer visual feature indexing framework, scan the visual feature relationship tree. During the scanning process, combine real-time production parameter data to calculate the dynamic quality correction factor for each image region. Based on the dynamic quality correction factor, sort the feature analysis areas by weight and select a set of candidate image samples that meet the preset quality standards from the multi-layer visual feature indexing framework. S5. Based on the pre-trained defect correlation assessment model, the selected image sample set and the corresponding visual deviation feature set are correlated to obtain the defect assessment results. The defect assessment results include defect transmission path characteristics, quality risk level parameters and process impact parameters. S6. Based on the defect assessment results, construct a printing process optimization model with the goals of blocking defect transmission paths, reducing quality risk levels, and optimizing process impact. Combined with preset constraints, obtain an initial set of process adjustment schemes. Perform feasibility verification on the initial set of process adjustment schemes, and obtain printing process adjustment instructions after verification.

[0020] In an optional embodiment, the sample image data includes standard color reference information, precise registration and positioning information, and complete image and text structure information; the multispectral image data includes surface texture information and color distribution information; the visual feature relationship tree includes primary visual nodes and secondary visual nodes that establish intensity correlation relationships; and the texture information and deviation information of the corresponding feature analysis areas are associated at each level in the multi-layer visual feature index framework.

[0021] In an optional embodiment, a multidimensional visual benchmark is constructed by extracting benchmark visual features, quality constraint features, and scene identification features based on sample image data; the multispectral image data is compared with the multidimensional visual benchmark to obtain a visual deviation feature set, including: The sample image data is classified into production scene categories based on the characteristics of printing material type, printing speed range, and environmental temperature and humidity. The average color density feature, registration position deviation feature, image and text outline clarity feature, and ink layer uniformity feature are extracted from each typical production scenario data column and fused to obtain the baseline visual features. Constraint rules are extracted from the preset printing quality standards and converted into quality constraint features. By fusing the baseline visual features, quality constraint features, and scene identification features corresponding to typical production scenarios, a multi-dimensional visual baseline is obtained. The corresponding baseline visual features and quality constraint features are then matched based on the production scene identification of real-time printed materials. Multispectral image data is compared layer by layer with the matched baseline visual features and quality constraint features. Color consistency deviation, registration accuracy deviation and image integrity deviation are calculated respectively. After fusion, a visual deviation feature set is obtained.

[0022] It's important to note the following: Printing material type: Different printing materials (such as paper, fabric, etc.) can affect the printing effect; for example, coating thickness and surface roughness will affect print quality. Printing speed range: The speed of the printing press also affects the final print result; high-speed printing may lead to uneven ink layers, while low-speed printing may affect production efficiency. Ambient temperature and humidity: Temperature and humidity affect ink drying speed, material deformation, etc., thus affecting print quality; for example: Scenario 1: Using high-quality paper with a thick coating, low temperature and low humidity environment, low-speed printing; Scenario 2: Using ordinary paper. High temperature and humidity environment, high speed printing; average color density characteristic refers to the color concentration level of the printed image, which affects the color performance of the printed matter; for example, colors that are too dark or too light will affect the visual effect; registration position deviation characteristic refers to the alignment error between multiple color layers during the printing process, which usually manifests as image and text offset or misalignment; image and text outline sharpness characteristic refers to the evaluation of image sharpness, and blurred image and text outlines may be due to factors such as uneven ink distribution or unstable pressure; ink layer uniformity characteristic refers to the analysis of the uniformity of ink distribution during the printing process, and uneven ink layer will lead to inconsistent color depth, affecting the final effect; For example, in scenario 1, the extracted visual features are assumed to be high color density (due to paper and low temperature environment), small registration position deviation, clear outline, and uniform ink layer; in scenario 2, due to high-speed printing and high humidity environment, the color density may be low, the registration position deviation may be large, the outline may be blurred, and the ink layer may be uneven. Quality standards usually include indicators such as color density, registration accuracy, outline clarity, and ink layer uniformity. These standards are converted into numerical features as constraints for subsequent evaluation. For example, quality constraint features may include: color density standard range of 0.8-1.2, registration position deviation not exceeding 0.2mm, image and text outline clarity exceeding 90%, and ink layer uniformity above 90%. By integrating baseline visual features, quality constraint features, and production scenario identifier features, a comprehensive visual baseline representing a specific production scenario can be formed. These baselines can help match real-time image data during the production process to determine whether it meets quality standards. For example, for scenario 1, the baseline visual features might be: high color density, low registration deviation, high outline sharpness, and uniform ink layer; the quality constraint features are set according to quality standards; the identifier features of the entire scenario might be paper printing in a low-speed, low-temperature environment. By comparing the real-time monitored image data with the previously extracted baseline visual features and quality constraint features, it can be determined whether the current printing meets the standards. For example, if the color density in the real-time image data is low, the registration deviation is large, and the outline is not clear, it can be determined that the current printing quality does not meet expectations and the printing conditions may need to be adjusted. Color consistency deviation: compares the consistency between the actual color in the image and the preset standard color; registration accuracy deviation: analyzes the deviation between the alignment of the image layers in the actual image and the standard alignment requirements; image integrity deviation: checks the integrity of the image and text, whether there are missing or blurred parts; ultimately, these deviations can help judge the printing quality, detect problems in time, and make adjustments; for example, in real-time monitoring, if a large color consistency deviation is detected, such as uneven color, or the registration accuracy deviation exceeds the standard value, an alarm can be issued in time, and the printing process can be adjusted.

[0023] In an optional embodiment, a visual feature relationship tree is constructed based on the visual deviation feature set and texture features in the multispectral image data; the visual feature paths from core nodes to edge nodes in the visual feature relationship tree are analyzed, and the multispectral image data is classified into regions based on the visual feature paths to obtain multiple feature analysis regions, including: Color deviation features, registration deviation features, and image-text deviation features are extracted from the visual deviation feature set; texture distribution features, texture density features, and texture direction features are extracted from multispectral image data. Using various deviation features and texture features as the basic attributes of nodes, we establish primary visual nodes and secondary visual nodes, and construct the strength relationship between nodes based on the degree of influence between features to obtain a visual feature relationship tree; Extract node attributes from the visual feature relationship tree, determine the continuous node groups and the strength of the relationship between nodes on each visual feature path, and calculate the path significance value of each visual feature path; Calculate the path correlation degree of visual feature paths corresponding to any two image regions, select the image regions with the highest sum of path significance values ​​as the initial clustering core, and establish the mapping relationship between image regions and paths. Based on the mapping relationship, the path correlation between the image region to be classified and the initial cluster core is calculated. The image region to be classified is then assigned to the cluster core with the highest path correlation, resulting in multiple feature analysis regions.

[0024] It should be noted that color deviation features assess the degree of color deviation in an image, such as the difference from a standard color; registration deviation features indicate the accuracy of alignment between different layers (such as color layers and text layers) in an image; the greater the registration deviation, the greater the alignment error between the image and text; image-text deviation features refer to the integrity and clarity of the image and text, which may include blurring, loss, or misalignment; texture distribution features refer to the spatial distribution pattern of textures in an image, such as whether the texture is uniform or repetitive; texture density features refer to the density of the texture; the higher the density, the more complex or delicate the texture in the image; texture direction features refer to the directionality of the texture; when the texture direction is regular, it can help determine the structure of the image; for example: for In a printed image, color deviation features can reflect whether there is a significant color deviation, registration deviation features can indicate whether text and patterns are aligned, and texture distribution features can analyze whether the distribution of printed patterns is uniform. Each visual feature (such as color deviation, registration deviation, texture features, etc.) serves as an attribute of a node, forming nodes in a relationship tree. The primary visual node is likely to be the most important feature node, such as color deviation. Depending on the actual needs, secondary nodes may be texture direction, registration deviation, etc. The strength of the correlation between nodes is constructed based on the mutual influence between features, establishing the connection between these nodes. For example, color deviation may affect registration accuracy, or texture distribution may have some relationship with registration deviation. For example, a large color deviation in an image may affect the clarity of the image and text (image-text deviation), thus leading to a strong correlation between these two features. In this way, a feature tree can help understand the interaction and influence between different features. A visual feature path refers to the path taken through the feature tree, starting from a node and following the branches to the terminal node. Each path represents a transition from one visual feature to another. Calculating the saliency of each path allows us to assess its importance based on the strength of the association between nodes. Paths with high saliency values ​​may represent important combinations of visual features. For example, if the path from the "color deviation" node to the "registration deviation" node has high saliency, it indicates that color deviation has a significant impact on registration deviation, and therefore this path has a high saliency value. Path correlation is used to measure the similarity between two image regions, that is, the degree of similarity between them on visual feature paths. If the feature paths of two image regions are similar, then their path correlation is high. By calculating the sum of path saliency values, the image regions with the highest sums are selected as the core regions for clustering. For example, assuming there are multiple image regions, the sum of the saliency values ​​of the visual feature paths of each region is calculated, and the regions with the highest saliency values ​​are selected as the initial cluster cores, such as the regions with the most similar color deviation and texture distribution. The image regions to be classified are then assigned to the most relevant cluster cores based on their path correlation with the initial cluster cores. Through this process, image regions can be divided into different feature analysis regions, each representing a group of regions that are similar in visual features. For example, if the color deviation and texture direction of a region to be classified are similar to a cluster core, it will be classified under that core, forming a feature analysis region. In this way, several feature analysis regions can be formed based on the relationship of visual feature paths. The image regions within each analysis region are highly similar in features, allowing for further analysis.

[0025] In an optional embodiment, texture coverage and intra-region cohesion are calculated for each feature analysis region, and a multi-layer visual feature indexing framework is constructed by combining the deviation parameters of the corresponding regions in the visual deviation feature set, including: For each feature analysis region, the ratio of the sum of the intensity weights of directly related texture features to the sum of the intensity weights of all texture features in the visual feature relationship tree is used as the texture coverage. Calculate the correlation between any two image regions within the feature analysis area, and take the average of all correlations as the cohesion within the area; Extract color deviation parameters, registration deviation parameters, and image-text deviation parameters corresponding to each feature analysis area from the visual deviation feature set, and calculate the comprehensive deviation index of each feature analysis area; The index hierarchy of each feature analysis area is determined by weighting texture coverage, intra-region cohesion, and comprehensive deviation index; the separation degree between adjacent index levels is calculated, and adjacent levels with a separation degree less than a preset threshold are fused to obtain a multi-layer visual feature index framework.

[0026] It should be noted that texture coverage describes the ratio of texture features within a feature analysis area to all texture features, reflecting the "representativeness" or "coverage" of texture features in the feature analysis area; the sum of the intensity weights of directly related texture features refers to the sum of the weight values ​​of texture features directly related to a certain feature analysis area; for example, if a region has "texture distribution features" and "texture direction features," then the weights of these two features are summed; the sum of the intensity weights of all texture features refers to the total weight of all possible texture features in the visual feature relationship tree; for example, all texture features include texture distribution, texture density, texture direction, etc. Intra-region cohesion measures the correlation between different image regions within the same feature analysis region. High cohesion indicates that the image regions within the region are highly similar in features, while low cohesion means that there are significant differences between the regions within the region. For all image regions within the same feature analysis region, the correlation between them is calculated. The correlation can be measured based on visual feature paths, color, texture, etc. For example, the similarity between features such as color deviation, texture distribution, and registration deviation between two image regions can be calculated. The average correlation between all image regions is used to obtain the intra-region cohesion. The comprehensive deviation index evaluates the overall deviation of a feature analysis region by considering multiple visual deviation features (such as color deviation, registration deviation, and image-text deviation). Extract color deviation, registration deviation, and image-text deviation parameters for each feature analysis region: Extract deviation parameters related to the feature analysis region from the visual deviation feature set; The comprehensive deviation index can be calculated using a weighted average, considering the weight of different deviations; for example, color deviation may contribute more to the comprehensive deviation index; By weighting texture coverage, intra-region cohesion, and comprehensive deviation index, the index level of each feature analysis region can be determined, thus providing a basis for image region classification or priority ranking; Separation is used to measure the difference between adjacent index levels; If the separation between two levels is less than a preset threshold, it indicates that their difference is not significant, and the two levels can be merged; The greater the difference between adjacent index levels, the higher the separation; When the separation is less than the threshold, it indicates that the difference between the two levels is not significant, and fusion can be performed; For example, suppose the index levels of two feature analysis regions are 0.4808 and 0.4850, respectively, and the preset separation threshold is 0.05. Since their difference is less than 0.05, these two levels can be fused. By sorting and fusing the index levels of each feature analysis region, a multi-level visual feature index framework is finally obtained. This framework helps to perform more effective image analysis and classification in practical applications. For example, suppose that after fusion processing, multiple visual feature index levels are obtained, each corresponding to different image region features. Finally, image classification, anomaly detection, and other tasks can be performed based on these levels.

[0027] In an optional embodiment, based on a multi-layer visual feature indexing framework, a visual feature relationship tree is scanned. During the scanning process, a dynamic quality correction factor for each image region is calculated in conjunction with real-time production parameter data. Based on the dynamic quality correction factor, the feature analysis regions are weighted and sorted, and a candidate image sample set that meets the preset quality standards is selected from the multi-layer visual feature indexing framework, including: Based on a multi-layer visual feature indexing framework, the scanning priority of each visual feature node is calculated by combining the node hierarchy depth, the number of branches, and the number of associated image regions. Based on the scanning priority, a breadth-first scan is performed on the visual feature relationship tree to obtain a scan sequence. From the scan sequence, the set of image regions corresponding to each node and the associated real-time production parameter records are extracted. A set of visual feature paths for each image region is constructed based on the hierarchical structure of the scanning sequence. The basic quality coefficient is calculated based on the quality index of each feature on the path, and the process response coefficient is calculated based on the production parameter records. After standardizing and weighting the basic quality coefficient and process response coefficient, a dynamic quality correction factor for each image region is obtained. The feature analysis regions are then ranked by weight based on the dynamic quality correction factor. Feature analysis regions are selected based on weighted ranking results, image samples are extracted from the selected feature analysis regions, and a set of candidate image samples that meet the requirements is selected by combining preset quality standards.

[0028] It should be noted that the scanning priority is calculated based on the hierarchical depth, number of branches, and number of associated image regions of each visual feature node, and is used to determine which visual feature nodes should be scanned first. The deeper a node is in the feature relationship tree, the more complex or difficult the features involved in the node are to be calculated, so its scanning priority may be lower. Conversely, shallower nodes may be simpler and have higher priority. The more branches a node has, the more complex the feature relationships it connects to, so it may take more time to process; in this case, its scanning priority may be lower. The more image regions associated with a feature node, the wider the distribution of that node's features in the image, and therefore its scanning priority is likely to be higher. For example, suppose a visual feature tree has two nodes A and B: node A is in the shallow layer of the tree, with 3 branches and 100 associated image regions; node B is in the deep layer of the tree, with 1 branch and 20 associated image regions. Because node A has more associated regions and is in a shallower layer, its scanning priority is higher, while node B's scanning priority is lower. Based on the scanning priority of each node, the visual feature tree is scanned in a breadth-first manner to obtain the scanning order, and the set of image regions associated with each node and real-time production parameter records are extracted from the scanning sequence. Breadth-first scanning means scanning the shallower nodes of the tree first, and then scanning downwards layer by layer; this allows for the rapid processing of features with a wide impact; each visual feature node corresponds to a certain set of image regions, and the features of these regions are associated with that node; the image regions of each node are usually associated with real-time production parameters (such as temperature, pressure, speed, etc.), which have a significant impact on product quality; for example: scanning order: assuming the breadth-first scanning order of the visual feature relationship tree is nodes A, B, C, D, node A is associated with 100 image regions, node B is associated with 20 image regions, and each node also has real-time production parameter records, such as production speed, equipment temperature, etc. Based on the hierarchical structure of the scanning sequence, a set of visual feature paths is constructed for each image region. A basic quality coefficient is calculated using the quality indices of each feature along the path, and then the process response coefficient is calculated by combining these with production parameter records. For each image region, a set of paths is constructed based on the nodes in the scanning sequence. The path set consists of multiple feature nodes, each representing a visual feature. The basic quality coefficient is calculated based on the quality indices (such as sharpness, contrast, and texture) of each feature along the path; these indices reflect the visual quality of the image region. The process response coefficient is calculated based on the impact of real-time production parameters (such as temperature, humidity, and pressure) on product quality. For example, high temperatures may degrade image quality, thus potentially lowering the score of a certain region in the process response coefficient. For example: Suppose the visual feature path of image region A is A→B→C→D, where A, B, C, and D correspond to features such as image sharpness, contrast, texture quality, and color distribution, respectively; each feature has a quality index, such as image A having a sharpness score of 0.8, a contrast score of 0.6, a texture quality score of 0.7, and a color distribution score of 0.5; meanwhile, production parameters show that the current production temperature is high, which may affect image quality, resulting in a process response coefficient of 0.9; To avoid the influence of different dimensions, the basic quality coefficient and process response coefficient need to be standardized to ensure they fall within the same dimension. Based on actual needs, different weights are assigned to the basic quality coefficient and process response coefficient, and a weighted fusion is performed to obtain the dynamic quality correction factor for each image region. All image regions are then ranked according to the dynamic quality correction factor, with regions scoring higher having higher priority. Based on the ranking, regions with higher quality are selected as the focus for further processing. For example, assuming there are multiple feature analysis regions, the calculated dynamic quality correction factors are 0.83, 0.75, 0.91, and 0.68 respectively. According to the ranking, the feature analysis region with the highest weight is the region with a dynamic quality correction factor of 0.91, followed by the region with 0.83. Based on the weighted ranking results, feature analysis regions with higher dynamic quality correction factors are selected; relevant image samples are extracted from these high-priority feature analysis regions; according to preset quality standards, such as sharpness and contrast, a sample set that meets the requirements is selected from the extracted image samples; for example, assuming that the selected feature analysis regions are regions with dynamic quality correction factors of 0.91 and 0.83, image samples from these regions are extracted, and a candidate image sample set that meets the requirements is obtained through quality standard screening.

[0029] In an optional embodiment, a pre-trained defect correlation assessment model is used to perform correlation analysis on the selected image sample set and the corresponding visual deviation feature set to obtain defect assessment results, including: The image features and visual deviation feature sets corresponding to the candidate image sample set are input into the feature preprocessing layer of the defect correlation assessment model to obtain the standard deviation feature matrix; the standard deviation feature matrix is ​​used as the node features and the causal relationship between different deviation features is used as the edge relationship to construct the defect propagation graph; Critical path mining is performed on the defect propagation graph to identify the transmission chain from the initial defect node to the final quality defect node, and the path node density, path dependency strength and path propagation rate are extracted as defect propagation path features. The defect superposition effect value is calculated based on the defect transmission path characteristics, and the quality risk level parameters are generated by combining the tolerance characteristics of the preset quality standards. By mapping the visual deviation feature set to real-time printing process parameters, analyzing the sensitivity of process parameter adjustments to the correction of deviation features, and obtaining process influence parameters, the quality risk level parameters and process influence parameters are integrated to obtain defect assessment results.

[0030] It should be noted that each image in the candidate image sample set has features such as sharpness, contrast, and texture. These features may be associated with visual deviations, such as color distortion and defocus. By inputting these image features along with visual deviation features into the defect correlation assessment model, a standard deviation feature matrix can be generated. The features of each image reflect the visual quality of the image; for example, the features of image A may include sharpness, contrast, and texture. Visual deviation features are those related to visual deviations that may occur during the production process, such as color deviation and blurriness. By standardizing these features, a matrix can be obtained, in which the deviation of each feature is marked, helping to identify which image regions have abnormal features. For example, suppose the image sample set includes multiple images, some of which show obvious color distortion, while others show insufficient contrast. Through the defect correlation assessment model, the standard deviation matrix of these features can be obtained, reflecting the feature deviation of each image region. By treating each feature in the standard deviation feature matrix as a node in the graph and the causal relationships between nodes as edges, a defect propagation graph is constructed to illustrate the influence relationships between features. Each feature in the standard deviation feature matrix, such as image sharpness and contrast, can be considered a node. The causal relationships between nodes, i.e., the deviation of one feature may affect another feature (for example, blurriness may lead to a decrease in contrast), establish an edge relationship between these two to represent causal influence. For example, suppose the sharpness deviation of image A affects the image contrast, and the color deviation of image B may lead to a decrease in its sharpness; then, in the defect propagation graph, sharpness, contrast, and color deviation will be connected together through edge relationships. By analyzing the defect propagation graph, the transmission chain from the initial defect node to the final quality defect node is identified. Critical path mining helps to understand how defects propagate from one feature to another, ultimately affecting overall quality. The critical path refers to the path taken from an initial defect (such as sharpness deviation) to the node that ultimately affects product quality (such as image non-conformity). Defect propagation path characteristics include node density (the number of features involved in the path), path dependency strength (the strength of the influence between features), and path propagation rate (the speed at which the defect's impact propagates). For example, if it starts with sharpness deviation, affects contrast through blurring, and ultimately leads to image non-conformity, then the path from sharpness deviation to the non-conformity node is the critical path. By analyzing the node density and propagation rate in the path, the contribution of each node on the path to the quality defect can be evaluated. Based on the characteristics of the defect propagation path, such as the node density and dependency strength of the path, the defect superposition effect value is calculated, and combined with preset quality standards, quality risk level parameters are generated. The final defect superposition effect is calculated through the influence of multiple nodes on the path. Deviations at each node on the path may superimpose, leading to more severe quality problems. Combined with preset quality standards, such as allowable sharpness deviation range and contrast range, the risk level is assessed to determine which defects may cause product quality to fail. For example, assuming the path leads from sharpness deviation to blurriness, and then to contrast deviation, the superposition effect of these nodes may cause the final product's image quality to be far below the acceptable standard. If the sharpness deviation exceeds the allowable range, the superposition effect may also make contrast and color deviations severe, ultimately leading to quality failure and an assessment of a high-risk level. By analyzing the changes in various process parameters (such as temperature and pressure) during production, the effectiveness of these parameters in correcting image deviation characteristics (such as color deviation and sharpness deviation) can be evaluated. For example, changing the printing temperature may help reduce color distortion. The effect of adjusting process parameters can be measured by a "process influence degree," which reflects the effectiveness of process adjustments in reducing defects. For example, assuming that excessively high temperatures during printing cause image color distortion, this deviation can be corrected by adjusting the temperature. If the color deviation is significantly reduced after adjusting the temperature, then the process influence degree parameter of temperature will be high, indicating that temperature has a strong ability to correct color deviation. The calculated quality risk level parameters and process impact parameters are integrated to obtain the final defect assessment result. This result reflects the overall assessment of product quality and the risks that can be reduced by adjusting process parameters. By combining the quality risk level and process impact, it is possible to determine which defects are most severe and which process adjustments are most effective, thus providing decision support for quality control. For example, suppose that after analysis, a certain image has a high quality risk level, and adjusting the process temperature can significantly improve the deviation of the image. By integrating the risk level and process impact, an optimization plan is derived, namely, adjusting the temperature in future production processes to reduce risks and improve the final product quality.

[0031] In an optional embodiment, a printing process optimization model is constructed based on the defect assessment results. With the objectives of blocking defect propagation paths, reducing quality risk levels, and optimizing process impact, and combined with preset constraints, an initial set of process adjustment schemes is obtained, including: Using the defect propagation path characteristics, quality risk level parameters, and process impact parameters from the defect assessment results as core inputs, a multi-objective function is constructed. Color tolerance constraints, registration accuracy constraints, image integrity constraints, and process parameter range constraints are extracted from the preset printing quality standards and used as constraints for the printing process optimization model. Input the path parameters and correlation strength parameters from the defect propagation path characteristics into the path blocking target, input the quality risk level parameter into the risk reduction target, input the process impact parameter into the impact optimization target, and determine the weight coefficient of each target; The multi-objective function is iteratively solved under constraints to obtain multiple sets of candidate process solutions; Multi-objective optimization is performed on multiple sets of candidate process solutions to select candidate process solutions that simultaneously meet the requirements of each optimization objective, thus obtaining an initial set of process adjustment schemes.

[0032] It should be noted that the three core inputs in the defect assessment results (defect propagation path features, quality risk level parameters, and process influence parameters) are used to construct a multi-objective optimization function. These parameters reflect image quality, potential defects, and the effectiveness of process adjustments. The defect propagation path features reflect the process by which defects propagate from one feature to another; for example, from image blurriness to contrast, and finally to quality defects. The quality risk level parameter quantifies the current image's quality risk, indicating how far it is from the preset quality standard. The process influence parameter quantifies the degree of influence of different printing process parameters (such as temperature and pressure) on image quality. Combining these three inputs to construct a multi-objective function means simultaneously optimizing multiple objectives, such as reducing defects, improving quality, and optimizing the process. For example, assuming the current printed image has a high quality risk level, it indicates a high probability of quality problems and significant blurriness and contrast deviations. The process influence parameter indicates that adjusting printing temperature or pressure can improve sharpness and contrast, thereby reducing quality risk. The multi-objective function will attempt to optimize these aspects simultaneously. In the optimization model, certain constraints need to be set to ensure that process adjustments do not violate preset quality standards and process parameter ranges. These constraints can be derived from the following rules extracted from preset printing quality standards: Color tolerance constraint rules: specify the allowable range of color deviation during printing; exceeding this range may lead to color distortion or color difference problems; Registration accuracy constraint rules: control the precise alignment of printed patterns, ensuring the correct position of patterns and text, and avoiding misalignment; Image and text integrity constraint rules: ensure that images and text are not missing or blurred during printing, ensuring clear and complete output; Process parameter range constraint rules: such as printing temperature, pressure, speed, etc., must be within a certain range to avoid affecting image quality or equipment safety; for example, during printing, the color tolerance rule may specify that the deviation of a certain color should not exceed ±3%; if the deviation exceeds this range, it will be considered unqualified; other rules, such as registration accuracy requirements, require that images and text must be aligned, and process parameters such as temperature cannot be too high, otherwise it may lead to thermal distortion of the image; Next, different objective functions need to be input based on the different nature of the optimization objectives: Path blocking objectives aim to prevent defects from propagating from one feature to another, especially along the defect propagation path; for example, if the sharpness deviation is too large, it may affect the normal presentation of contrast, so such paths need to be "blocked"; Risk reduction objectives refer to reducing quality risk based on the quality risk level parameter; for example, if an image has a high quality risk level, the risk can be reduced by adjusting the process, such as improving printing quality or adjusting the temperature to avoid image blurring; Impact optimization objectives aim to maximize the process effect by adjusting process parameters (such as printing pressure, temperature, etc.) to achieve the optimization goal; Each objective will have a corresponding weight coefficient to determine which objective needs to be prioritized during the optimization process; the weight coefficient reflects the importance of the objective and is usually set by experience or specific needs; for example, in a real-world scenario, the path blocking objective may need a higher weight because if defect propagation is not effectively prevented, it may lead to a significant drop in overall quality; while the risk reduction objective may have a lower weight because it focuses on ensuring that the risk is at a safe level; the impact optimization objective aims to improve process efficiency, and although it has a lower weight, it is still important. After establishing the multi-objective function and constraint rules, the system finds multiple candidate process solutions through iterative solutions. Each iteration adjusts the printing process parameters according to different objective functions and calculates their optimization effect on each objective. For example, after the first iteration, the system may adjust the printing temperature and pressure to try to optimize color deviation and image alignment. After multiple iterations, the system may propose multiple process schemes, each performing well on different objectives. After obtaining multiple candidate process solutions, the next step is multi-objective optimization. This step selects the best scheme based on whether each candidate scheme can simultaneously meet the requirements of all objectives. For each candidate scheme, it checks whether it can meet all constraint rules, such as color tolerance and registration accuracy, and performs a comprehensive evaluation of different objectives. From the multiple candidate process solutions, one or more process schemes that best meet the current production needs are selected. For example, suppose that after multiple iterations, two candidate schemes are obtained: Scheme A and Scheme B. Scheme A performs well in terms of color deviation but has poor registration accuracy; Scheme B is more balanced in terms of color and registration accuracy. Through optimization, the system selects Scheme B as the final process adjustment scheme. Finally, after the above steps, a preliminary set of process adjustment schemes is obtained. These schemes meet all quality requirements and constraints, and achieve a reasonable balance between defect risk, quality improvement, and process impact. For example, in the final step, the system may recommend several adjustment schemes, such as: Scheme 1: Increase the printing temperature and moderately increase the pressure to reduce color deviation, but careful handling of image alignment is required; Scheme 2: Adjust the printing speed to ensure registration accuracy and optimize color contrast; Scheme 3: Slightly adjust the pressure and enhance image clarity based on the current machine conditions, without changing the temperature. These schemes will serve as a preliminary set of process adjustment schemes for further production trials and evaluation.

[0033] In an optional embodiment, the initial set of process adjustment schemes undergoes feasibility verification. Upon successful verification, a printing process adjustment instruction is obtained, including: Select one scheme from the initial process adjustment scheme set and analyze its ink volume adjustment parameters, printing pressure adjustment parameters, and registration accuracy correction parameters; Input the ink volume adjustment parameters into the ink zone control simulation module to simulate the ink layer distribution state and obtain the ink volume adjustment simulation results. Then, superimpose the printing pressure adjustment parameters with the real-time pressure data to verify whether it conforms to the adaptive rules. Input the registration accuracy correction parameters into the registration system simulation module to simulate the correction effect, verify whether the registration accuracy meets the preset standard, and calculate the feasibility index by combining the three results. If the feasibility index reaches the preset threshold, it is marked as a feasible solution; otherwise, the next solution is selected for re-verification, and all feasible solutions are sorted in descending order of feasibility index. Select the top-ranked feasible solution as the final process adjustment solution, and convert the final process adjustment solution into a printing process adjustment instruction that the printing equipment can recognize.

[0034] It should be noted that, firstly, among all the initial process adjustment schemes, one scheme is selected for detailed analysis. Each scheme includes multiple parameters, such as ink volume adjustment, printing pressure adjustment, and registration accuracy correction, which are all key factors affecting the final printing quality and efficiency. The ink volume adjustment parameter determines the amount of ink in each ink zone during the printing process; too much or too little ink can lead to poor printing results. The printing pressure adjustment parameter affects the degree of contact between the paper and the printing cylinder, directly affecting ink transfer and the clarity of the pattern. The registration accuracy correction parameter is used to correct the alignment accuracy of the pattern or text, ensuring the positional matching between different printing layers and avoiding misalignment or blurring. For example, suppose that in the selected initial scheme, the ink volume adjustment parameter is set to increase the ink volume by 10%, the printing pressure is adjusted to increase the pressure by 5%, and the registration accuracy correction is to increase the automatic calibration frequency. Next, the ink volume adjustment parameters are input into the ink zone control simulation module. This module simulates how ink is distributed during the printing process, thereby predicting the uniformity of the ink layer and the actual effect of the ink. This process helps determine whether the ink volume adjustment is reasonable and predicts the final printing quality. For example, in the simulation, an increase in ink volume may lead to an excessively thick ink layer, affecting the display of details or color saturation. The simulation results may show uneven ink layer distribution, causing the pattern to be blurry. The printing pressure adjustment parameters are then superimposed with real-time pressure data for verification to confirm whether the adjusted pressure conforms to the adaptive rules. The adaptive rules refer to the fact that in actual production, the pressure should not be too high or too low. Too high a pressure may cause paper damage or image blurring, while too low a pressure may lead to incomplete ink transfer. For example, assuming the adjusted printing pressure is 105% of the original, if the real-time pressure data shows that the actual pressure fluctuation is within a reasonable range, such as 102%-108%, which conforms to the adaptive rules, then the pressure adjustment is reasonable. Registration accuracy correction parameters are input into the registration system simulation module to simulate the image correction effect. Through simulation, the system can evaluate the positional accuracy between different printing layers and verify whether it meets the preset registration accuracy standard. For example, assuming that the corrected pattern shows perfect alignment in the simulation, meeting the registration accuracy standard, this step ensures that there is no misalignment of the text and images, and that all color and pattern layers are correctly aligned. Based on the simulation and verification results of ink volume adjustment, printing pressure verification, and registration accuracy correction, the system calculates a "feasibility index." This index measures the overall feasibility of a process solution. The higher the feasibility index, the more the solution meets quality and process requirements. For example, assuming that the ink volume adjustment simulation results show minor problems, the pressure verification results are good, and the registration accuracy is perfect, the final calculated feasibility index may be 80%. If the preset feasibility threshold is 75%, the solution can be considered feasible. If the feasibility index of the currently selected solution fails to reach the preset threshold (e.g., below 75%), the solution will be marked as infeasible, and the next solution will be selected for verification. This process will continue until a feasible solution that meets all requirements is found. For example, if the feasibility index of the current solution is only 65%, the system will automatically select the next process solution and repeat the above verification process until a feasible solution is found. Once multiple solutions have been verified and the feasibility of each solution has been determined, the system will sort all feasible solutions in descending order of feasibility index. The solution with the highest feasibility index is the optimal solution. For example, assuming there are three feasible solutions, solution A has a feasibility index of 85%, solution B has 80%, and solution C has 78%, the system will ultimately select solution A as the final process adjustment solution. Finally, the selected optimal solution, i.e., the solution ranked first in the feasibility index, is transformed into adjustment instructions that the printing equipment can recognize and execute. These instructions will guide the equipment on how to adjust parameters such as ink volume, pressure, and registration accuracy to optimize print quality. For example, assuming that solution A is selected as the final solution, it may include the following instructions: Ink volume adjustment: increase ink volume by 10%; Printing pressure: increase pressure by 5%; Registration accuracy: increase the automatic calibration frequency to once every 10 seconds. These instructions will be transmitted to the printing equipment, which will adjust the relevant parameters according to the instructions and execute the final printing process.

[0035] Example 2, please refer to Figure 2 This invention provides a technical solution: a printing process evaluation system based on visual analysis, applicable to the aforementioned printing process evaluation method based on visual analysis, comprising: Data acquisition unit 1 is used to acquire multispectral image data of the surface of the printed product and sample image data of historical qualified printed products; extract benchmark visual features, quality constraint features and scene identification features based on the sample image data to construct a multidimensional visual benchmark; compare the multispectral image data with the multidimensional visual benchmark to obtain a visual deviation feature set; Feature parsing unit 2 is used to construct a visual feature relationship tree based on the visual deviation feature set and texture features in multispectral image data; parse the visual feature path from core node to edge node in the visual feature relationship tree; perform region classification on multispectral image data based on the visual feature path; and obtain multiple feature analysis regions. Feature indexing unit 3 is used to calculate the texture coverage and cohesion within each feature analysis region, and to construct a multi-layer visual feature indexing framework by combining the deviation parameters of the corresponding region in the visual deviation feature set. The quality correction unit 4 is used to scan the visual feature relationship tree based on the multi-layer visual feature index framework, and calculate the dynamic quality correction factor for each image region in combination with real-time production parameter data during the scanning process; and to sort the feature analysis area by weight based on the dynamic quality correction factor, and select the candidate image sample set that meets the preset quality standard from the multi-layer visual feature index framework. Defect assessment unit 5 is used to perform correlation analysis on the selected image sample set and the corresponding visual deviation feature set based on the pre-trained defect correlation assessment model to obtain defect assessment results. The defect assessment results include defect transmission path features, quality risk level parameters and process impact parameters. The process adjustment unit 6 is used to construct a printing process optimization model based on the defect assessment results. With the goals of blocking defect transmission paths, reducing quality risk levels, and optimizing process impact, it obtains an initial set of process adjustment schemes by combining preset constraints. The initial set of process adjustment schemes is then subjected to feasibility verification, and printing process adjustment instructions are obtained after verification.

[0036] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A method for print process evaluation based on visual analysis, characterized in that, The method comprises the following steps: acquiring multi-spectral image data of a printed product surface and sample image data of a historical qualified printed product; extracting reference visual features, quality constraint features and scene identification features based on the sample image data, constructing a multi-dimensional visual reference, comparing the multi-spectral image data with the multi-dimensional visual reference to obtain a visual deviation feature set; constructing a visual feature relationship tree based on the visual deviation feature set and texture features in the multi-spectral image data; analyzing visual feature paths from core nodes to edge nodes in the visual feature relationship tree, regionally classifying the multi-spectral image data based on the visual feature paths to obtain a plurality of feature analysis regions; calculating texture coverage and intra-regional cohesion for each feature analysis region, and constructing a multi-layer visual feature index framework in combination with deviation parameters of corresponding regions in the visual deviation feature set; based on the multi-layer visual feature index framework, scanning the visual feature relationship tree, and calculating a dynamic quality correction factor for each image region in combination with real-time production parameter data during the scanning process; based on the dynamic quality correction factor, weight ordering the feature analysis regions, and screening a set of selected image samples meeting a preset quality standard from the multi-layer visual feature index framework; based on a pre-trained defect correlation evaluation model, performing correlation analysis on the set of selected image samples and corresponding visual deviation feature sets to obtain a defect evaluation result, wherein the defect evaluation result includes defect transmission path features, quality risk level parameters and process influence degree parameters; based on the defect evaluation result, constructing a printing process optimization model, taking defect transmission path blocking, quality risk level reduction and process influence degree optimization as targets, and combining preset constraint conditions to obtain an initial process adjustment scheme set; performing feasibility verification processing on the initial process adjustment scheme set, and obtaining a printing process adjustment instruction after verification.

2. A method of print process evaluation based on visual analysis according to claim 1, characterized in that, The sample image data includes standard color reference information, accurate registration positioning information and complete graphic structure information; The multi-spectral image data includes surface texture information and color distribution information; the visual feature relationship tree includes primary visual nodes and secondary visual nodes establishing intensity correlation; and the texture information and deviation information of each level of associated feature analysis regions in the multi-layer visual feature index framework.

3. A method of print process evaluation based on visual analysis according to claim 2, characterized in that, extracting reference visual features, quality constraint features and scene identification features based on the sample image data, and constructing a multi-dimensional visual reference; comparing the multi-spectral image data with the multi-dimensional visual reference to obtain a visual deviation feature set, including: performing production scene classification processing on the sample image data, dividing the sample image data into a plurality of typical production scene data columns based on printed material type features, printing speed range features and environmental temperature and humidity features; extracting average color density features, registration position deviation features, graphic contour sharpness features and ink layer uniformity features from each typical production scene data column to obtain reference visual features, and extracting constraint rules from a preset printing quality standard to convert into quality constraint features; Fusing the reference visual features, the quality constraint features and the scene identification features of the corresponding typical production scene to obtain a multi-dimensional visual reference, and matching the corresponding reference visual features and quality constraint features based on the production scene identification of the real-time printed matter; Comparing the multi-spectrum image data with the matched reference visual features and quality constraint features layer by layer, respectively calculating the color consistency deviation, the register accuracy deviation and the graphic integrity deviation, and fusing to obtain a visual deviation feature set.

4. A method of print process evaluation based on visual analysis according to claim 3, characterized in that, Constructing a visual feature relationship tree based on the visual deviation feature set and the texture features in the multi-spectrum image data; Analyzing the visual feature paths from the core nodes to the edge nodes in the visual feature relationship tree, and classifying the multi-spectrum image data based on the visual feature paths to obtain a plurality of feature analysis regions, including: Extracting the color deviation features, the register deviation features and the graphic deviation features from the visual deviation feature set, and extracting the texture distribution features, the texture density features and the texture direction features from the multi-spectrum image data; Taking each type of deviation feature and texture feature as a node basic attribute, establishing a primary visual node and a secondary visual node, constructing an intensity correlation relationship between the nodes based on the influence degree between the features, and obtaining a visual feature relationship tree; Extracting the node attributes in the visual feature relationship tree, determining the continuous node groups and the intensity correlation relationship between the nodes on each visual feature path, and calculating the path saliency value of each visual feature path; Calculating the path correlation degree of the visual feature paths corresponding to any two image regions, selecting a plurality of image regions with the highest path saliency value sum as initial clustering cores, and establishing a mapping relationship between the image regions and the paths; Based on the mapping relationship, calculating the path correlation degree of the to-be-classified image region and the initial clustering core, and classifying the to-be-classified image region to the clustering core with the highest path correlation degree to obtain a plurality of feature analysis regions.

5. A method of print process evaluation based on visual analysis according to claim 4, characterized in that, For each feature analysis region, calculating the texture coverage rate and the intra-region cohesion degree, and constructing a multi-layer visual feature index framework in combination with the deviation parameters of the corresponding region in the visual deviation feature set, including: For each feature analysis region, taking the ratio of the sum of the directly associated texture feature intensity weight and the sum of all texture feature intensity weights in the visual feature relationship tree as the texture coverage rate; Calculating the correlation degree of any two image regions in the feature analysis region, and taking the average value of all correlation degrees as the intra-region cohesion degree; Extracting the color deviation parameters, the register deviation parameters and the graphic deviation parameters corresponding to each feature analysis region from the visual deviation feature set, and calculating the comprehensive deviation index of each feature analysis region; Weighting the texture coverage rate, the intra-region cohesion degree and the comprehensive deviation index to determine the index level of each feature analysis region, calculating the separation degree between adjacent index levels, fusing the adjacent levels with a separation degree less than a preset threshold, and obtaining a multi-layer visual feature index framework.

6. A print process evaluation method based on visual analysis according to claim 5, characterized in that, Based on the multi-layer visual feature index framework, the visual feature relationship tree is scanned, and a dynamic quality correction factor of each image area is calculated in combination with real-time production parameter data during the scanning process; based on the dynamic quality correction factor, the feature analysis area is weighted and ordered, and a selected image sample set meeting a preset quality standard is screened from the multi-layer visual feature index framework, including: Based on the multi-layer visual feature index framework, and in combination with node hierarchical depth, branch quantity and associated image area quantity, a scanning priority of each visual feature node is calculated; Based on the scanning priority, the visual feature relationship tree is scanned in breadth priority, a scanning sequence is obtained, and an image area set corresponding to each node and associated real-time production parameter records are extracted from the scanning sequence; Based on the hierarchical structure of the scanning sequence, a visual feature path set of each image area is constructed, a basic quality coefficient is calculated based on the quality indicators of the features on the path, and a process response coefficient is calculated based on the production parameter records; After the basic quality coefficient and the process response coefficient are standardized and weighted fusion, a dynamic quality correction factor of each image area is obtained, and the feature analysis area is weighted and ordered based on the dynamic quality correction factor; Based on the weight ordering result, the feature analysis area is screened, and image samples are extracted from the screened feature analysis area, and a selected image sample set meeting the requirements is screened in combination with a preset quality standard.

7. A method of print process evaluation based on visual analysis according to claim 6, characterized in that, Based on the pre-trained defect correlation evaluation model, the selected image sample set and the corresponding visual deviation feature set are analyzed, and a defect evaluation result is obtained, including: The image features corresponding to the selected image sample set and the visual deviation feature set are input into the feature preprocessing layer of the defect correlation evaluation model to obtain a standard deviation feature matrix; the standard deviation feature matrix is taken as a node feature, and a causal correlation between different deviation features is taken as an edge relationship to construct a defect propagation graph; The defect propagation graph is mined for a critical path to identify a conduction chain from an initial defect node to a final quality defect node, and path node density, path dependence strength and path propagation rate are extracted as defect conduction path features; Based on the defect conduction path features, a defect superposition effect value is calculated, and a quality risk level parameter is generated in combination with a tolerance feature of a preset quality standard; The visual deviation feature set is associated and mapped with real-time printing process parameters, the correction sensitivity of process parameter adjustment to deviation features is analyzed, a process influence degree parameter is obtained, and the quality risk level parameter and the process influence degree parameter are integrated to obtain a defect evaluation result.

8. A print process evaluation method based on visual analysis according to claim 7, characterized in that, Based on the defect evaluation result, a printing process optimization model is constructed, taking defect conduction path blocking, quality risk level reduction and process influence degree optimization as the target, and combining a preset constraint condition, an initial process adjustment scheme set is obtained, including: Taking the defect conduction path features, quality risk level parameters and process influence degree parameters in the defect evaluation result as core inputs, a multi-objective function is constructed; Extract color tolerance constraint rules, register accuracy constraint rules, graphic integrity constraint rules and process parameter range constraint rules from preset printing quality standards as constraint conditions of the printing process optimization model; Input the path parameters and the correlation strength parameters in the defect conduction path features into the path blocking target, input the quality risk level parameters into the risk reduction target, and input the process influence degree parameters into the influence degree optimization target to determine the weight coefficients of each target; Iteratively solve the multi-objective function under the constraint conditions to obtain a plurality of groups of candidate process solutions; Perform multi-objective optimization processing on the plurality of groups of candidate process solutions to screen out candidate process solutions that simultaneously meet the requirements of each optimization target and obtain an initial process adjustment scheme set.

9. A print process evaluation method based on visual analysis according to claim 8, characterized in that, Perform feasibility verification processing on the initial process adjustment scheme set to obtain printing process adjustment instructions after verification, including: Selecting a scheme from the initial process adjustment scheme set, analyzing the ink amount adjustment parameters, printing pressure adjustment parameters and register accuracy correction parameters therein; Inputting the ink amount adjustment parameters into an ink area control simulation module to simulate the ink layer distribution state and obtaining an ink amount adjustment simulation result, and superimposing the printing pressure adjustment parameters and real-time pressure data to verify whether they conform to the adaptability rules; Inputting the register accuracy correction parameters into a register system simulation module to simulate the correction effect, verifying whether the register accuracy reaches the preset standard, and comprehensively calculating the feasibility index based on the three results; If the feasibility index reaches the preset threshold, it is marked as a feasible scheme, otherwise the next scheme is selected for re-verification, and all feasible schemes are arranged in descending order of the feasibility index; Selecting the top-ranked feasible scheme as the final process adjustment scheme and converting the final process adjustment scheme into a printing process adjustment instruction recognizable by the printing equipment.

10. A visual analysis based print process evaluation system adapted to a visual analysis based print process evaluation method according to any one of claims 1 to 9, characterized in that Including: A data acquisition unit configured to acquire multi-spectral image data of a printed product surface and sample image data of a historical qualified printed product; Extracting reference visual features, quality constraint features and scene identification features based on the sample image data to construct a multi-dimensional visual reference, and comparing the multi-spectral image data with the multi-dimensional visual reference to obtain a visual deviation feature set; A feature analysis unit configured to construct a visual feature relationship tree based on the visual deviation feature set and texture features in the multi-spectral image data; Analyzing a visual feature path from a core node to an edge node in the visual feature relationship tree, regionally classifying the multi-spectral image data based on the visual feature path, and obtaining a plurality of feature analysis regions; A feature indexing unit configured to calculate texture coverage and intra-region cohesion for each feature analysis region, and construct a multi-layer visual feature indexing framework in combination with deviation parameters of corresponding regions in the visual deviation feature set; A quality correction unit configured to scan the visual feature relationship tree based on the multi-layer visual feature indexing framework, calculate a dynamic quality correction factor for each image region in a scanning process in combination with real-time production parameter data, and perform weight ordering on the feature analysis regions based on the dynamic quality correction factor, and screen a to-be-selected image sample set that meets a preset quality standard from the multi-layer visual feature indexing framework. The defect evaluation unit is configured to perform correlation analysis on the candidate image sample set and the corresponding visual deviation feature set based on a pre-trained defect correlation evaluation model to obtain a defect evaluation result, wherein the defect evaluation result includes a defect conduction path feature, a quality risk level parameter, and a process influence degree parameter. The process adjustment unit is configured to construct a printing process optimization model based on the defect evaluation result, to obtain an initial process adjustment scheme set in combination with a preset constraint condition, with the goal of blocking a defect conduction path, reducing a quality risk level, and optimizing a process influence degree, and to perform feasibility verification processing on the initial process adjustment scheme set to obtain a printing process adjustment instruction after verification.

Images