A method and system for detecting defects in a printing screen based on image analysis

Abstract

The present application relates to the technical field of defect detection, and discloses a printing screen defect detection method and system based on image analysis, which comprises the following steps: dividing a printing screen to be detected into a plurality of detection units, obtaining feature source data corresponding to each detection unit according to detection unit description information; mapping each detection unit into a unit behavior portrait vector, and performing clustering analysis and benchmark calculation to obtain a plurality of screen pattern sets and corresponding set benchmarks; calculating the toughness index of the detection unit, combining the function area weight and the historical defect weight in the detection unit description information, generating a monitoring priority, combining a context multi-combination reward to generate a hierarchical collection task set; re-collecting the image of a target detection unit, and using the Mahalanobis distance of a new behavior portrait vector relative to the set benchmark of the screen pattern set to which the new behavior portrait vector belongs and a defect type discrimination rule to output an abnormal defect detection result, thereby realizing continuous monitoring and abnormal identification of printing screen defects.

Description

Technical Field

[0001] This invention relates to the field of defect detection technology, and more specifically, to a method and system for detecting defects in printing screens based on image analysis. Background Technology

[0002] The printing screen uses a high-precision frame of 450mm*450mm with a flatness of ≤0.20mm, combined with high-strength polyester mesh and ultra-high mesh steel wire mesh, achieving a perfect combination of fine filtration and stable support. Simultaneously, the printing screen undergoes processes such as mesh fabric lamination, PI film bonding, laser-laid opening, and wire drawing, resulting in significant improvements in frame flatness, mesh strength, polyester mesh tension, steel wire mesh diameter, and line width / opening. At the same time, printing screen defects have shifted from macroscopic damage to earlier forms such as minute dimensional deviations, localized adhesion anomalies, and slight contamination particles. These defects are weaker and more fragmented, placing higher demands on detection resolution, imaging consistency, and algorithm robustness.

[0003] Current printing screen defect detection methods largely rely on manual visual inspection or machine vision solutions with fixed parameters. Manual visual inspection is affected by human experience, fatigue, and ambient light, making it difficult to consistently identify minor defects such as small opening size deviations, broken / bridged fine lines, slight warping of PI film adhesion, and mild foreign matter contamination. While machine vision inspection with fixed thresholds or single templates can be used for some significant defects, the normal texture itself varies greatly across different functional areas, process batches, and imaging conditions, easily leading to problems such as difficulty in standardizing thresholds and the coexistence of false positives and false negatives. In addition, most solutions tend to make judgments after each acquisition, lacking utilization of short-term fluctuations and the consistency of repeated sampling, making it difficult to distinguish between occasional imaging noise and the evolution of real defects. At the same time, the lack of adaptive scheduling of inspection resources results in either excessive resource consumption affecting cycle time or insufficient resources leading to missed detection of critical areas.

[0004] No effective solutions have yet been proposed to address the problems in the relevant technologies. Summary of the Invention

[0005] In view of the problems in related technologies, this invention proposes a method and system for detecting defects in printing screens based on image analysis, so as to overcome the above-mentioned technical problems existing in the existing related technologies.

[0006] Therefore, the specific technical solution adopted by the present invention is as follows: According to one aspect of the present invention, a method for detecting defects in printing screens based on image analysis is provided, comprising: The printing screen to be inspected is divided into several inspection units, and inspection unit description information is configured for each inspection unit; Based on the description information of the detection unit, personalized image acquisition and preprocessing are performed on each detection unit to obtain the feature source data corresponding to each detection unit; Based on the feature source data, each detection unit is mapped to a unit behavior profile vector, and cluster analysis and benchmark calculation are performed on the unit behavior profile vector to obtain several sets of network usage patterns and corresponding set benchmarks. Based on the network usage pattern set and behavior profile resampling mechanism, the short-term stability index and the group universality index of the detection unit are calculated, and the resilience index is obtained by combining the fusion rules. The resilience index is combined with the functional area weight and historical defect weight in the detection unit description information to generate a monitoring priority; at the same time, based on the monitoring priority, a hierarchical collection task set is generated by combining multiple rewards in context. According to the hierarchical acquisition task set, the images of the target detection units are re-acquired, and new behavioral profile vectors are obtained. Using the Mahalanobis distance of the new behavioral profile vectors relative to the set reference of the network pattern set to which they belong, as well as the defect type discrimination rules, the abnormal defect detection results are output.

[0007] Furthermore, the printing screen to be inspected is divided into several inspection units, and each inspection unit is configured with inspection unit description information, including: The printing screen to be inspected is divided into multiple inspection units according to the effective detection area boundary and the preset spatial resolution. Generate detection unit description information for each detection unit, including detection unit identifier, two-dimensional coordinate position, functional area type, and process identifier.

[0008] Furthermore, based on the detection unit description information, personalized image acquisition and preprocessing are performed on each detection unit to obtain the feature source data corresponding to each detection unit, including: Based on the detection unit description information, the corresponding imaging mode is selected and the detection unit image is acquired. Then, imaging consistency correction, noise reduction, contrast enhancement and background texture suppression are performed to generate feature source data. Feature source data includes candidate masks for opening regions, candidate masks for line structures, candidate masks for foreign objects, and grayscale statistics.

[0009] Furthermore, based on the feature source data, each detection unit is mapped to a unit behavior profile vector, and cluster analysis and benchmark calculation are performed on the unit behavior profile vectors to obtain several sets of network usage patterns and corresponding set benchmarks, including: Calculate the aperture ratio based on the candidate masks for the opening regions in the feature source data; calculate the linewidth deviation based on the candidate masks for the line structures in the feature source data; Calculate the foreign object area distribution based on the foreign object candidate mask in the feature source data; calculate the transmittance statistic using the grayscale statistical image; After standardizing the aperture ratio, linewidth deviation, foreign object area distribution, and transmittance statistics, they are spliced ​​together in a fixed order to form a unit behavior profile vector. Cluster analysis is performed on the unit behavior profile vectors to divide the detection units into multiple network usage pattern sets; the set benchmark is determined by using the statistics of all detection unit behavior profile vectors within the network usage pattern set.

[0010] Furthermore, based on the network usage pattern set and behavioral profile resampling mechanism, the short-term stability index and group universality index of the detection unit are calculated, and the resilience index is obtained by combining the fusion rules, including: Within a preset time window, each detection unit is subjected to at least two repeated image acquisitions, and a corresponding unit behavior profile vector is constructed based on the feature source data obtained from each acquisition. The unit behavior profile vector is compared with the set benchmark of the network pattern set to which the corresponding detection unit belongs, and the Mahalanobis distance corresponding to each sampling is calculated. Short-term stability indices are calculated based on Mahalanobis distance sequences, and population universality indices are calculated based on the mean of Mahalanobis distance. After limiting and normalizing the short-term stability indicators and the general population indicators, a weighted sum or multiplicative fusion is performed to obtain the resilience index.

[0011] Furthermore, by combining the resilience index with the functional area weights and historical defect weights in the detection unit description information, a monitoring priority is generated, including: The risk quantity is obtained by inverse mapping of the resilience index, and the risk quantity is then weighted and fused or interactively enhanced with the functional area weight and historical defect weight of the detection unit to obtain the monitoring priority.

[0012] Furthermore, based on monitoring priority and combined with contextual multiple reward combinations, a tiered collection task set is generated, including: Obtain all detection units with monitoring priority and configure context information vectors for each detection unit, including resilience index, functional area type, functional area weight, and historical defect weight; A set of preset acquisition actions is provided. The acquisition parameters for each acquisition action include acquisition frequency, acquisition resolution, exposure gain, number of fields of view, scanning path, number of re-inspections, and the imaging channel used. A cost function is preset for each acquisition action, where the cost represents the resources consumed by the action in the runtime environment corresponding to the context information vector, including time, computing power, bandwidth, device usage, and cycle time impact; a total budget constraint is set to limit the total cost upper limit of the acquisition tasks that can be executed within the preset scheduling period; A basic reward is constructed for the detection unit and the acquisition action, and the basic reward is monotonically correlated with the monitoring priority; at least two context combination reward items are constructed based on the context information vector. The context combination reward items are used to characterize the gain or suppression of the acquisition benefit by the context factors. According to the preset multi-combination fusion rules, the basic reward and the context combination reward items are combined to form the total reward of the candidate collection action, and the unit cost benefit is calculated in combination with the collection action cost to represent the collection benefit under unit resource consumption. For each detection unit, enumerate the subset of actions that satisfy the feasibility constraints, and select the action with the highest unit cost benefit from the subset of actions to generate candidate tasks for the detection unit. Configure task scores and task costs for all candidate tasks, with task scores being monotonically correlated with total rewards and task costs being monotonically correlated with the cost function; Select the subset of tasks whose task costs satisfy the total budget constraint and whose total task score is the largest, and form the selected task set. The selected task set is divided into at least two levels according to the preset grading threshold, and different sets of collection parameters are bound to different levels to output the graded collection task set.

[0013] Furthermore, according to the hierarchical acquisition task set, images of the target detection units are reacquired, and new behavioral profile vectors are obtained, including: Receive a set of hierarchical acquisition tasks, determine the target detection unit identifier, acquisition action identifier, and acquisition parameter identifier corresponding to any acquisition task, and determine the acquisition level; Under the condition of confirmed executability, the acquisition configuration and acquisition action are executed, and after acquiring new images and special resources, a new behavioral profile vector is obtained.

[0014] Furthermore, using the Mahalanobis distance of the new behavioral profile vector relative to the set benchmark of its corresponding network pattern set, and the defect type discrimination rule, the abnormal defect detection results are output, including: For each detection unit, calculate the Mahalanobis distance of the new behavior profile vector relative to the set reference of the set of network patterns to which it belongs, and use it as the deviation set; The rules for determining defect types include the rules for opening size deviation, line width deviation, broken line bridging, abnormal polyimide film bonding, and foreign matter contamination. By comparing the deviation set with the defect type discrimination rule and combining it with the conflict resolution rule, the abnormal defect detection result is obtained.

[0015] According to another aspect of the present invention, a printing screen defect detection system based on image analysis is provided, comprising: The unit configuration module is used to divide the printing screen to be inspected into several inspection units and configure inspection unit description information for each inspection unit; The acquisition and preprocessing module is used to perform personalized image acquisition and preprocessing for each detection unit based on the detection unit description information, so as to obtain the feature source data corresponding to each detection unit; The profile benchmark module is used to map each detection unit into a unit behavior profile vector based on feature source data, and to perform cluster analysis and benchmark calculation on the unit behavior profile vector to obtain several sets of network patterns and corresponding set benchmarks. The resilience assessment module is used to calculate the short-term stability index and the generality index of the detection unit based on the network usage pattern set and the behavior profile repeated sampling mechanism, and to obtain the resilience index by combining the fusion rules. The priority task module is used to combine the resilience index with the functional area weight and historical defect weight in the detection unit description information to generate a monitoring priority; at the same time, based on the monitoring priority, combined with contextual multiple combination rewards, a hierarchical collection task set is generated. The defect detection module is used to re-acquire images of target detection units according to the hierarchical acquisition task set and obtain new behavioral profile vectors; using the Mahalanobis distance of the new behavioral profile vectors relative to the set reference of the network pattern set to which they belong and the defect type discrimination rules, the module outputs abnormal defect detection results.

[0016] The beneficial effects of this invention are as follows: 1. This invention uses the detection unit as the smallest granularity to achieve adaptive, interpretable, and resource-controlled continuous monitoring and anomaly identification of printing screen defects. Compared with fixed parameters, single acquisition, or single threshold discrimination methods, this invention can maintain detection consistency and stability in different functional areas, different processes, and different screen usage modes, and reduces the impact of random noise and misjudgment through a secondary sampling mechanism.

[0017] 2. Pattern Adaptation and Cross-Regional Consistency. Clustering of behavioral profile vectors yields several mesh pattern sets and set benchmarks, ensuring that the normal baseline of each detection unit no longer depends on a globally unified threshold but rather on the statistical benchmark of its respective pattern set. Mahalanobis distance is used to characterize deviation, enabling statistical absorption of differences across different functional areas, textured backgrounds, and imaging conditions. This improves the consistency and generalization ability of cross-regional detection and reduces systematic false alarms caused by differences in manufacturing processes or structures.

[0018] 3. Noise Resistance and Early Risk Detection. Through a resampling mechanism within a preset time window, the profile vector of each sample is compared with the ensemble benchmark to obtain the Mahalanobis distance sequence. From this, short-term stability indicators and population universality indicators are constructed, and then fused to obtain the resilience index. This achieves the separation and characterization of two types of risk signals: volatility / instability and deviation from the population norm. It can capture statistical offsets and increased volatility caused by fit changes, contamination accumulation, and operating condition drift before defects develop into significant morphological anomalies, thus achieving earlier anomaly warnings.

[0019] 4. Resource-controlled proactive monitoring and interpretable localization. The resilience index is inversely mapped to a risk level and integrated with functional area weights and historical defect weights to generate monitoring priorities. Then, under the constraint of the total budget, the collection action with the highest unit cost benefit is selected through basic rewards, contextual multi-combination rewards, and cost functions to form a hierarchical collection task set. This enables the preferential allocation of detection resources to high-risk, high-value areas, improving defect discovery efficiency and production line cycle time friendliness per unit of resource input. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of a printing screen defect detection method based on image analysis according to an embodiment of the present invention; Figure 2 This is a schematic diagram of a printing screen defect detection system based on image analysis according to an embodiment of the present invention. Detailed Implementation

[0022] To further illustrate the various embodiments, the present invention provides accompanying drawings, which are part of the disclosure of the present invention. These drawings are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these drawings, those skilled in the art should be able to understand other possible implementation methods and the advantages of the present invention. The components in the drawings are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0023] According to embodiments of the present invention, a method and system for detecting defects in printing screens based on image analysis are provided.

[0024] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1As shown, according to an embodiment of the present invention, a method for detecting defects in printing screens based on image analysis is provided, comprising: S1. Divide the printing screen to be inspected into several inspection units, and configure inspection unit description information for each inspection unit.

[0025] In one embodiment, dividing the printing screen to be inspected into several inspection units and configuring inspection unit description information for each inspection unit includes: dividing the printing screen to be inspected into multiple inspection units according to the effective inspection area boundary and a preset spatial resolution; and generating inspection unit description information for each inspection unit, including inspection unit identifier, two-dimensional coordinate position, type of functional area, and process identifier information.

[0026] The effective detection area boundary of the screen to be inspected is determined by removing undetectable or meaningless areas such as screen frame occlusion, clamping areas, and areas without graphics. The effective detection area is then divided into multiple detection units according to a preset spatial resolution, ensuring that each unit's physical size matches the imaging field of view and feature statistical scale. Simultaneously, descriptive information is bound to each detection unit. Through unitization and descriptive information configuration, spatial positioning and defect traceability are achieved, with defects accurately falling to the unit ID and coordinates. Furthermore, differentiated detection strategies are implemented. The preset spatial resolution is determined based on the following factors: the camera's single-shot imaging field of view; the minimum target defect size; the screen pattern period or linewidth; the minimum number of pixels required for feature statistics; and the motion positioning accuracy of the detection equipment. For example, it can be specified that the physical size of the detection unit is no larger than the single-shot imaging field of view, and that the number of pixels corresponding to the minimum target defect within each detection unit is no less than a preset pixel threshold.

[0027] The functional area type can be obtained from design documents, process formula documents, historical test templates, or manually labeled information, and is bound to the test unit through coordinate mapping. Process identification information includes screen batch, printing material type, mesh count, emulsion thickness, exposure process parameters, target linewidth, target aperture size, etc.

[0028] S2. Based on the description information of the detection unit, perform personalized image acquisition and preprocessing for each detection unit to obtain the feature source data corresponding to each detection unit.

[0029] In one embodiment, based on the detection unit description information, personalized image acquisition and preprocessing are performed on each detection unit to obtain the feature source data corresponding to each detection unit. This includes: selecting the corresponding imaging mode and acquiring the detection unit image based on the detection unit description information, and performing imaging consistency correction, noise reduction, contrast enhancement, and background texture suppression to generate feature source data; the feature source data includes candidate masks for opening regions, candidate masks for line structures, candidate masks for foreign objects, and grayscale statistics.

[0030] Specifically, using the functional area type and process identifier in the detection unit description information as conditional variables, the imaging method that best highlights the defect characterization of the unit is selected, such as bright field / dark field, coaxial / ring, backlight transmission, different bands or polarization channels, to maximize the imaging contrast between targets such as openings, line structures, bonding interfaces, and foreign particles and the background. Subsequently, imaging consistency correction is performed, including flat field / dark field correction, geometric distortion and brightness drift compensation, to ensure comparability across fields of view and across time. Through denoising, contrast enhancement, and background texture suppression, structural targets are separated from periodic textures / material textures / uniform illumination. Finally, the original image is transformed into an intermediate representation closer to subsequent feature calculations: candidate masks for opening regions, candidate masks for line structures, candidate masks for foreign objects, and a grayscale statistical map for transmittance / grayscale distribution modeling. Candidate masks for opening regions are obtained by performing grayscale thresholding on the backlit image, combined with morphological closing operations and area filtering. Candidate masks for line structures are obtained by performing edge enhancement, binarization, skeleton extraction, and orientation consistency filtering on the bright field image. Foreign object candidate masks are obtained through background modeling subtraction, dark field highlight target extraction, connected component filtering, and area threshold filtering.

[0031] Upgrading from passive acquisition with uniform parameters to active acquisition customized to individual units directly improves the visibility of weak defects and reduces imaging differences between different regions. Consistency correction and texture suppression make subsequent calculations of aperture ratio, linewidth deviation, foreign object area distribution, and transmittance statistics more comparable, thereby reducing false alarms caused by uneven illumination, device drift, and background texture. Candidate masks reduce the dimensionality of subsequent analysis from full image processing to structured target processing, improving computational efficiency and providing directly usable geometric and statistical support for defect type rules.

[0032] S3. Based on the feature source data, each detection unit is mapped to a unit behavior profile vector, and cluster analysis and benchmark calculation are performed on the unit behavior profile vector to obtain several sets of network usage patterns and corresponding set benchmarks.

[0033] In one embodiment, based on feature source data, each detection unit is mapped to a unit behavior profile vector, and cluster analysis and benchmark calculation are performed on the unit behavior profile vectors to obtain several sets of screen usage patterns and corresponding set benchmarks, including: calculating the aperture ratio based on candidate masks for opening areas in the feature source data; calculating the line width deviation based on candidate masks for line structures in the feature source data; calculating the foreign object area distribution based on candidate masks for foreign objects in the feature source data; calculating the transmittance statistic using a grayscale statistical graph; standardizing the aperture ratio, line width deviation, foreign object area distribution, and transmittance statistic, and then splicing them into a unit behavior profile vector in a fixed order; performing cluster analysis on the unit behavior profile vectors to divide the detection units into multiple sets of screen usage patterns, that is, statistical groups of normal screen usage behavior characteristics exhibited by the printing screen under the combined effects of specific functional area types, structural forms, and process conditions; and determining the set benchmark using the statistics of all detection unit behavior profile vectors within the set of screen usage patterns.

[0034] To transform structural features into computable and comparable numerical vectors, for example, in candidate masks for opening regions, the number of pixels with a value of 1 in the mask is counted and compared with the total number of pixels in the unit to obtain the aperture ratio. A binary mask for the line structure and the nominal linewidth of the unit are obtained. Several sampling positions perpendicular to the line direction are selected within the unit. For each sampling line, the length of consecutive pixels with a value of 1 in the mask is extracted as the actual linewidth. The average actual linewidth of the unit is obtained by averaging all sampled actual linewidths. The linewidth deviation is equal to the average actual linewidth minus the nominal linewidth. For candidate masks for foreign objects, a cluster of connected pixels with a value of 1 is considered a foreign object. For each foreign object, the number of pixels it contains is counted to obtain the area of ​​the foreign object. These areas are then grouped and counted by interval to obtain the area distribution. For grayscale statistics, grayscale values, standard deviation, etc., are calculated as transmittance statistics. These values ​​are standardized and concatenated in a fixed order to obtain the unit behavior profile vector. Clustering employs methods such as K-means, GMM, and hierarchical clustering to automatically group units with similar feature distributions into the same network usage pattern set. For each set, a set benchmark is obtained through statistical analysis of vectors, including a mean vector representing the center of the network usage pattern set in the feature space; a covariance matrix representing the dispersion of each feature dimension and the correlation between dimensions within the set; and also including the standard deviation of features, sample size, and label information to define the normal center and fluctuation range of the pattern. Specifically, the mean vector and covariance matrix of the set are calculated by statistically analyzing all detection unit behavior profile vectors within a given network usage pattern set, and these two are used together as the set benchmark for subsequent Mahalanobis distance deviation calculations.

[0035] In this invention, K-means clustering combined with Euclidean distance is used to perform clustering analysis on the unit behavior profile vectors. For any two d-dimensional unit behavior profile vectors a and b, the Euclidean distance is defined as: ; K-means achieves clustering by minimizing the sum of squared errors within clusters. Its objective function is: ; In the formula, For Euclidean distance, It is a norm 2. For the unit behavior profile vectors a and b, the first... One component; The number of detection units. For the standardized first u The behavioral profile vector of each unit, such as by performing zero-mean unit variance standardization on each feature dimension. For the first The center vector of the cluster to which each unit belongs.

[0036] During K-means iteration, for each cell, it is assigned to the cluster center with the nearest Euclidean distance. For each cluster, the cluster center is updated with the mean of all samples within that cluster. The minimum value of the number of clusters K is 2, and the maximum value is: ,in, The number of detection units participating in clustering. To round down, K is chosen as the final number of clusters, maximizing the average silhouette coefficient. The maximum number of clustering iterations is set to [value missing]. The convergence condition is that when the objective function changes, it satisfies: ; in, , , Let be the number of iterations. To reduce the impact of random initialization, for each candidate... Repeated run For example ), take the objective function The smallest one as the The clustering results are then used to calculate the silhouette coefficient. The cluster affiliation obtained based on the Euclidean distance principle is used as the set number of the mesh pattern for each detection unit.

[0037] For each detection unit, first calculate the average Euclidean distance between it and other units in the same cluster as the intra-cluster dissimilarity. Then, calculate the average Euclidean distance between it and all units in each other cluster, and take the smallest one as the nearest neighbor cluster dissimilarity. The ratio of the difference between the nearest neighbor cluster dissimilarity and the intra-cluster dissimilarity to the larger of the two is taken as the silhouette coefficient of the unit. The arithmetic mean of the silhouette coefficients of all detection units is taken to obtain the average silhouette coefficient, and the number of clusters with the largest average silhouette coefficient is selected as the final number of clusters.

[0038] This invention unifies complex image differences into a single feature space, forming a comparison mechanism between similar features: different functional areas, such as thin lines, large openings, and mixed textures, even if all are in a normal state, often have different aperture ratios, linewidth statistics, and transmittance baselines. After clustering, each has its own independent ensemble benchmark, and subsequent anomaly detection no longer uses a global threshold, thereby reducing false alarms caused by regional differences and process batch differences. The ensemble benchmark provides an interpretable normal profile, facilitating the tracing of typical aperture ratio ranges, linewidth fluctuations, foreign object background levels, and transmittance distributions for a particular pattern, providing data support for process monitoring and parameter backtracking.

[0039] S4. Based on the network usage pattern set and behavior profile resampling mechanism, calculate the short-term stability index and the generality index of the detection unit, and obtain the resilience index by combining the fusion rules.

[0040] In one embodiment, based on the network usage pattern set and the behavior profile resampling mechanism, the short-term stability index and the group universality index of the detection unit are calculated, and the resilience index is obtained by combining the fusion rules. This includes: performing at least two repeated image acquisitions on each detection unit within a preset time window, constructing a corresponding unit behavior profile vector based on the feature source data obtained from each acquisition; comparing the unit behavior profile vector with the set benchmark of the network usage pattern set to which the corresponding detection unit belongs, and calculating the Mahalanobis distance corresponding to each sampling; calculating the short-term stability index based on the Mahalanobis distance sequence, and calculating the group universality index based on the mean of the Mahalanobis distance; and performing weighted summation or multiplicative fusion on the short-term stability index and the group universality index after amplitude limiting normalization to obtain the resilience index.

[0041] Repeated sampling of the same detection unit within a preset time window can distinguish the true structural state from imaging noise / transient disturbances, such as vibration, reflection, dust settling, and short-term contamination. After obtaining the behavioral profile vector for each sampling, it is compared with the statistical benchmark of its corresponding network pattern set. Mahalanobis distance is used to unify the degree of deviation from the normal group center to a comparable scale, while considering the dimensions and correlations of various features. Subsequently, the fluctuation of the Mahalanobis distance sequence is used to measure whether the unit is stable in the short term; stability = consistent repeated measurement results. The mean of the Mahalanobis distance is used to measure whether its relative group is generally normal; the smaller the mean, the closer it is to the normal center of the same type. For example: A weighted summation method is used for fusion: ; In the formula, For detection unit u The resilience index, ∈[0,1], obtained by limiting and normalizing the short-term stability index and the population universality index; This serves as a short-term stability indicator after the amplitude limit is applied. This is a general indicator for the group after the amplitude limit is applied; The weighting coefficient (non-negative) for the stability index. This is the weighting coefficient (non-negative) for general indicators. The sum of the two weighting coefficients is equal, and the specific value can be set according to actual needs.

[0042] The short-term stability index is used to characterize the fluctuation of Mahalanobis distance within a short time window. A short-term window is formed by a number of recent Mahalanobis distances. The mean and fluctuation of the Mahalanobis distance within this window are calculated, and the ratio of the fluctuation to the mean is used as the short-term stability index. When the short-term stability index does not exceed a preset threshold, short-term stability is determined; if it exceeds the threshold, short-term instability is determined, and corresponding processing is triggered. This preset threshold can be determined by statistical analysis of historical stable period data or set based on process experience.

[0043] Mahalanobis distance measures the standardized deviation of each detection unit from the baseline (mean vector and covariance matrix) of its network pattern set. Within the same set, group statistics on Mahalanobis distances, such as taking the mean, reflect the prevalence of deviations at the group level: a larger mean indicates more units deviating more significantly and anomalies being more widespread; a smaller mean indicates closer alignment to the baseline and stronger group consistency.

[0044] For example, a detection unit in a thin line region sampled three times within a 10-minute window, and the Mahalanobis distance sequence compared with the baseline of its corresponding thin line pattern set was [0.9, 1.1, 1.0]. The mean was close to 1 and the fluctuation was small, indicating that it was close to the normal center of the group and stable, with a high resilience index. Another unit had a distance sequence of [1.2, 3.8, 2.9], with a significantly larger mean and greater fluctuation. This may be due to local bridging of the line width, contamination causing continuous abnormalities in features and fluctuations affected by imaging conditions. After normalization, both the universality and stability deteriorated, and the resilience index was significantly lower, thus it was preferentially marked as a high-risk re-examination target.

[0045] The formula for calculating Mahalanobis distance is: ; In the formula, These are the mean vector and covariance matrix of the respective network usage pattern set. This refers to the behavioral profile vector obtained from a single sampling of the detection unit to be evaluated.

[0046] S5. Combine the resilience index with the functional area weight and historical defect weight in the detection unit description information to generate a monitoring priority; at the same time, based on the monitoring priority, combine the context with multiple reward combinations to generate a hierarchical collection task set.

[0047] In one embodiment, generating a monitoring priority by combining the resilience index with the functional area weights and historical defect weights in the detection unit description information includes: inversely mapping the resilience index to obtain a risk quantity, and then weighting and fusioning the risk quantity with the functional area weights and historical defect weights of the detection unit or performing interactive enhancement fusion to obtain the monitoring priority.

[0048] In one embodiment, generating a hierarchical collection task set based on monitoring priority and contextual multi-combination rewards includes: acquiring all detection units with monitoring priority and configuring contextual information vectors for each detection unit, including resilience index, functional area type, functional area weight, and historical defect weight; pre-setting a collection action set, where the collection parameters for each action include collection frequency, collection resolution, exposure gain, number of fields of view, scanning path, number of re-inspections, and the imaging channel used; pre-setting a cost function for each collection action, where the cost represents the resources consumed by the action in the operating environment corresponding to the contextual information vector, including time, computing power, bandwidth, equipment usage, and cycle time impact; setting a total budget constraint to limit the total cost ceiling of the collection tasks that can be executed within a preset scheduling period; constructing basic rewards for detection units and collection actions, where the basic rewards are monotonically correlated with monitoring priority; and constructing at least two contextual combination reward items based on the contextual information vector, whereby the contextual combination reward items characterize contextual factors. The system aims to improve or suppress the acquisition gains. According to preset multi-combination fusion rules, such as weighted summation fusion, multiplicative fusion, or gated fusion, the basic reward and context-based combined reward items are combined to form the total reward of candidate acquisition actions. The unit cost benefit is calculated based on the acquisition action cost to characterize the acquisition gains per unit of resource consumption. For each detection unit, a subset of acquisition actions satisfying feasibility constraints is enumerated, such as process allowance constraints, equipment availability constraints, field-of-view reachability or scan path executable constraints, and time window constraints. The acquisition action with the highest unit cost benefit is selected from this subset to generate candidate tasks for the detection unit. Task scores and task costs are configured for all candidate tasks, with task scores monotonically correlated with the total reward and task costs monotonically correlated with the cost function. The subset of tasks whose task costs satisfy the total budget constraint and have the highest total task score is selected to form the selected task set. The selected task set is divided into at least two levels according to preset grading thresholds, and different acquisition parameter sets are bound to different levels, outputting a tiered acquisition task set. Higher levels correspond to higher acquisition frequency, higher acquisition resolution, more fields of view, or more re-inspection times.

[0049] The resilience index of this invention reflects whether a unit is stable in the short term and consistent with the group. The lower the index, the more likely it is to be on the verge of deterioration or anomaly. However, monitoring resources should be prioritized for locations with greater impact and a higher probability of recurrence. Therefore, the resilience index is inversely mapped to a risk quantity, and then integrated with functional area weights and historical defect weights. Functional area weights include business impact / criticality, and historical defect weights include prior risk / recurrence probability. The resulting monitoring priority includes: current state risk, structural criticality, and prior knowledge of past defects, realizing a shift from average resource allocation to risk-driven allocation. For example, if a unit has a low resilience index, but is located in a critical conductive thin-line functional area (i.e., has a high functional area weight), and has a history of multiple bridging defects (i.e., has a high historical defect weight), then the inverse risk quantity is high, and the integrated monitoring priority is the highest. The scheduling system will allocate actions with higher frequency, higher resolution, and more re-inspection times.

[0050] This invention employs a contextual multi-combination reward mechanism to adaptively select acquisition actions for detection units under total resource budget constraints. This mechanism is a contextual multi-armed gambling machine algorithm, treating each candidate acquisition action as an arm, the current state information of the detection unit as the context, and the defect discovery gain or risk reduction resulting from acquisition as the reward. By learning to select the action with the greatest benefit under different contexts, resources are allocated towards high-risk, high-value areas. This invention treats the acquisition action for each unit as a budget-constrained decision optimization: each unit has a contextual information vector, such as resilience, functional area, and history; each action has benefits, such as defect discovery and uncertainty reduction; and costs, such as time, computing power, bandwidth, equipment usage, and cycle time impact. By combining the basic reward (monotonically proportional to monitoring priority) with the contextual combination reward, the gain / suppression of benefits by the context is described, resulting in a total reward. This total reward is then divided by the cost to obtain the unit cost benefit. The most cost-effective action is selected for each unit from the available actions. Finally, under the global budget, the subset of tasks with the largest total score is selected, and different parameter sets are bound according to threshold levels to achieve risk-driven, resource-controlled, and interpretable scheduling.

[0051] Specifically, the resilience index is inversely mapped to a risk level: ; The monitoring priority is obtained by combining the functional area weight and historical defect weight: ; In the formula, For detection unit u The amount of risk, As a functional area weight, the importance / influence of the functional area where the unit is located is a priori. For historical defect weighting, the historical defect frequency / severity of this unit or its neighborhood is used as a priori measure. These are the fusion coefficients corresponding to risk level, functional area weight, and historical defect weight, respectively. All of them are greater than or equal to zero and their sum is equal to one. The specific values ​​can be set according to actual needs.

[0052] In context-based multi-combination rewards, the context vector structure includes: resilience index or risk level, functional area type, functional area weight, and historical defect weight. The action space structure includes: acquisition frequency, resolution, exposure gain, number of fields of view, scan path, number of re-inspections, and imaging channels. The cost function structure is as follows: ; in, The total cost represents the overall resource consumption of a data acquisition action within a scheduling cycle or a single task execution. For time consumption items, This is a computing power consumption item. This is a bandwidth consumption item. For equipment occupancy items, For the beat effect, The time cost weighting coefficient, This is the weighting coefficient for computing power cost. This is a bandwidth cost weighting coefficient. This is the weighting coefficient for equipment occupancy costs. The cost weighting coefficients for cycle time are essentially shadow prices and opportunity costs that discount different resource consumptions to the same cost scale. They are monotonically consistent with resource scarcity, the strength of hard constraints, and sensitivity to production line targets: the scarcer the resource, the more likely it is to become a bottleneck, and the more severe the consequences of exceeding limits, the greater its corresponding weight should be; for example... Equals 1, Equals 2, Equals 1, It equals 1.5. It equals 6.

[0053] The time consumption term represents the actual time required to perform a specific acquisition action. During calculation, the number of fields of view to be acquired, the exposure time for each field of view, the camera readout time, the platform movement and positioning time, the light source or imaging channel switching time, and the number of re-inspections are determined. The exposure, readout, movement and positioning, and channel switching times required for each field of view during a single acquisition are summed to obtain the time required for a complete acquisition. This time is then amplified based on the acquisition frequency or number of re-inspections to obtain the total time consumption of the action within a scheduling cycle. The original unit of the time consumption term is usually seconds or milliseconds, which needs to be normalized according to the preset maximum allowable time, the historical maximum time consumption, or the current scheduling cycle time window to obtain a dimensionless time consumption score.

[0054] The computational cost item represents the computational resources required to process and analyze images after an acquisition action generates them. During calculation, the number of images generated by the action, the resolution of each image, and the total number of pixels are counted. The processing complexity is determined based on the image processing workflow used in the action, such as whether flattening correction, denoising, enhancement, background texture suppression, candidate mask generation, feature extraction, clustering benchmark comparison, Mahalanobis distance calculation, and defect rule discrimination are performed. A corresponding computational cost coefficient is configured for each processing module based on its complexity level or historical average processing time. The number of images, pixel scale, and processing module complexity are then combined and accumulated to obtain the computational cost item. This computational cost item can be represented by processing time, floating-point operations, CPU / GPU utilization integral, or computational task score. Since the computational power expression scale differs across devices, it should be converted to a uniform computational cost score before participating in the total cost calculation.

[0055] The bandwidth consumption term represents the data transmission resources used by an acquisition action during image transmission, feature uploading, result feedback, or storage. During calculation, the original image data volume is estimated based on image resolution, pixel depth, number of image channels, number of acquisition frames, and number of re-inspections. The data volume is adjusted based on whether image compression is enabled, whether only candidate masks or behavioral profile vectors are uploaded, and whether the complete original image and multi-channel images need to be uploaded. Finally, the data volume of intermediate or result data such as grayscale statistics, candidate masks, defect annotation results, and log information is added to obtain the total transmitted data volume generated by the acquisition action within one scheduling cycle. The original unit of the bandwidth consumption term can be bytes, megabytes, or megabits. To avoid the data volume value being significantly larger than other terms, it needs to be normalized based on the network bandwidth limit, historical maximum transmission volume, or allowable transmission budget per unit cycle.

[0056] The equipment occupancy item represents the degree to which a data acquisition action occupies the camera, light source, motion platform, controller, workstation, or detection equipment. During calculation, it determines which hardware resources are required during the action's execution and whether these resources are shared or exclusive. Based on the action's duration, equipment exclusivity, current equipment load, resource conflict level, and the action's requirements for equipment state switching, the equipment occupancy intensity is determined. The action duration and equipment occupancy intensity are then combined to obtain the equipment occupancy item. The equipment occupancy item is directly defined as an equipment occupancy score, or it can be obtained by multiplying the occupancy time by the equipment scarcity coefficient, exclusivity coefficient, and conflict coefficient, and then normalized to a dimensionless score.

[0057] The cycle time impact term represents the effect of data acquisition actions on production line cycle time, inspection cycle time, or process flow efficiency. During calculation, the time window or cycle time margin allowed for the current process inspection task is determined; the relationship between the actual execution time of the data acquisition action and the allowed time window is compared; if the action can be completed within the allowed time without causing subsequent processes to wait, the cycle time impact term can take a lower value or be zero; if the action execution time exceeds the allowed time window, or causes production line waiting, workpiece delays, equipment downtime, or changeover delays, the cycle time impact is calculated based on the excess time, waiting time, downtime, or degree of delay. The cycle time impact term differs from the simple time consumption term; the time consumption term reflects the time consumed by the action itself, while the cycle time impact term reflects whether this time consumption disrupts the production rhythm. The cycle time impact term can be represented by overtime, waiting time, downtime, or cycle time penalty score, and usually needs to be normalized based on the allowed cycle time margin or the maximum acceptable delay.

[0058] It should be noted that before calculating the total cost, the time consumption, computing power consumption, bandwidth consumption, equipment occupancy, and cycle time impact items are normalized to convert each item into a dimensionless resource consumption score. Based on the resource scarcity and process constraint strength, corresponding weight coefficients are configured, and the normalized items are weighted and fused to obtain the total cost of the candidate acquisition action.

[0059] S6. According to the hierarchical acquisition task set, re-acquire the image of the target detection unit and obtain the new behavior profile vector; use the Mahalanobis distance of the new behavior profile vector relative to the set reference of the network pattern set to which it belongs and the defect type discrimination rule to output the abnormal defect detection result.

[0060] In one embodiment, re-acquiring images of target detection units according to a hierarchical acquisition task set and obtaining a new behavioral profile vector includes: receiving a hierarchical acquisition task set, determining the target detection unit identifier, acquisition action identifier, and acquisition parameter identifier corresponding to any acquisition task, and determining the acquisition level; under the condition of executability confirmation, executing the acquisition configuration and acquisition action, and obtaining a new behavioral profile vector after acquiring new images and special resources.

[0061] In one embodiment, the abnormal defect detection result is output by using the Mahalanobis distance of the new behavior profile vector relative to the set reference of the set of network patterns to which it belongs, and the defect type discrimination rule. This includes: for each detection unit, calculating the Mahalanobis distance of the new behavior profile vector relative to the set reference of the set of network patterns to which it belongs, as the deviation set; using the opening size deviation rule, line width deviation rule, broken line bridging rule, polyimide film bonding anomaly rule, and foreign matter contamination rule as the defect type discrimination rule; and obtaining the abnormal defect detection result by comparing the deviation set with the defect type discrimination rule and combining it with the conflict resolution rule.

[0062] The hierarchical acquisition task set binds each target detection unit to a specific acquisition action and its acquisition parameters, including frequency, resolution, exposure gain, number of fields of view, scanning path, number of re-inspections, and imaging channels. When the executability constraints are met, the target unit is re-acquired based on risk matching. The new images obtained from the re-acquisition are preprocessed, segmented, measured, and statistically analyzed to extract features, forming a new behavioral profile vector. Anomaly detection results are obtained by comparing deviation and rules. For example, rules such as opening size deviation, line width deviation, broken wire bridging, abnormal polyimide film bonding, and foreign matter contamination are used as defect type discrimination rules. Deviation triggering is combined with morphology and measurement threshold triggering. When multiple rules are hit simultaneously, the final anomaly detection result is output through conflict resolution rules, such as priority based on severity, priority based on evidence strength, and priority based on mutually exclusive categories.

[0063] Opening size deviation: Measure the opening diameter, length, width, and area, and compare them with the standard value; deviations exceeding tolerances or significant shape deformation indicate an anomaly. Line width deviation: Measure the line width at multiple points along the conductor; overall or sustained deviations from tolerances in width or narrowness within a segment indicate an anomaly. Broken lines / bridging: Broken lines are those that should be connected but are not; bridging is those that should be separated but are not. Determine by broken connectivity or excessively small spacing. Polyimide film adhesion anomalies: Abnormalities are indicated by typical morphologies such as bubbles, wrinkles, curling edges, detachment, or offset coverage in the polyimide film area. Foreign matter contamination: The presence of particles, fibers, or stains not native to the material; if the area and contrast reach the threshold and are close to critical areas, an anomaly is indicated.

[0064] Figure 2 As shown, according to another embodiment of the present invention, a printing screen defect detection system based on image analysis is also provided, comprising: Unit configuration module 1 is used to divide the printing screen to be inspected into several inspection units and configure inspection unit description information for each inspection unit.

[0065] The acquisition and preprocessing module 2 is used to perform personalized image acquisition and preprocessing for each detection unit based on the detection unit description information, so as to obtain the feature source data corresponding to each detection unit.

[0066] The profile benchmark module 3 is used to map each detection unit into a unit behavior profile vector based on feature source data, and to perform cluster analysis and benchmark calculation on the unit behavior profile vector to obtain several network pattern sets and corresponding set benchmarks.

[0067] Resilience assessment module 4 is used to calculate the short-term stability index and the generality index of the detection unit based on the network usage pattern set and the behavior profile resampling mechanism, and to obtain the resilience index by combining the fusion rules.

[0068] Priority task module 5 is used to combine the resilience index with the functional area weight and historical defect weight in the detection unit description information to generate monitoring priorities; at the same time, based on the monitoring priorities, combined with contextual multi-combination rewards, a hierarchical collection task set is generated.

[0069] The defect detection module 6 is used to re-acquire images of the target detection unit according to the hierarchical acquisition task set and obtain a new behavior profile vector; using the Mahalanobis distance of the new behavior profile vector relative to the set reference of the network pattern set to which it belongs and the defect type discrimination rule, the abnormal defect detection result is output.

[0070] 10,000 inspection units were collected under the same production line, batch, and machine conditions. After manual review, electrical testing, and microscopic verification, 1,240 units were identified as defective, and 1,876 units as good. Technology A: Fixed parameters, single-shot acquisition, globally unified threshold, and single-model discrimination; it does not distinguish between network pattern sets, does not perform Mahalanobis distance deviation sequences, and does not perform resilience indices or active task scheduling. This invention: Using inspection units as the granularity, it uses behavioral profile vector clustering to form network pattern sets and set benchmarks, uses Mahalanobis distance to represent deviations and form time series, constructs a resilience index, generates a hierarchical acquisition task set under total budget constraints and triggers secondary sampling, and outputs interpretable evidence.

[0071] Table 1 Comparison of indicators between the present invention and technology A Therefore, the overall recall and overall precision of the proposed solution are improved, indicating that the detection unit granularity, pattern benchmark, and deviation sequence can stably capture anomalies. The standard deviation of the recall rate in each functional area of ​​the proposed solution is also lower than that of technology A, indicating that the adaptive verification pattern scheme using a network pattern set combined with Mahalanobis distance can reduce systematic fluctuations. Furthermore, the proposed solution provides early warning; while achieving the above improvements, the average cycle time only increases by 2.9%, meeting the goals of proactive monitoring and hierarchical data collection under budget constraints.

[0072] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting defects in printing screens based on image analysis, characterized in that, include: The printing screen to be inspected is divided into several inspection units, and inspection unit description information is configured for each inspection unit; Based on the description information of the detection unit, personalized image acquisition and preprocessing are performed on each detection unit to obtain the feature source data corresponding to each detection unit; Based on the feature source data, each detection unit is mapped to a unit behavior profile vector, and cluster analysis and benchmark calculation are performed on the unit behavior profile vector to obtain several sets of network usage patterns and corresponding set benchmarks. Based on the network usage pattern set and behavior profile resampling mechanism, the short-term stability index and the group universality index of the detection unit are calculated, and the resilience index is obtained by combining the fusion rules. The resilience index is combined with the functional area weights and historical defect weights in the detection unit description information to generate monitoring priorities. Simultaneously, based on monitoring priority and combined with contextual multiple reward combinations, a tiered collection task set is generated; According to the hierarchical acquisition task set, the images of the target detection units are re-acquired, and new behavioral profile vectors are obtained. Using the Mahalanobis distance of the new behavioral profile vectors relative to the set reference of the network pattern set to which they belong, as well as the defect type discrimination rules, the abnormal defect detection results are output.

2. The method for detecting defects in printing screens based on image analysis according to claim 1, characterized in that, The process of dividing the printing screen to be inspected into several inspection units and configuring inspection unit description information for each inspection unit includes: The printing screen to be inspected is divided into multiple inspection units according to the effective detection area boundary and the preset spatial resolution. Generate detection unit description information for each detection unit, including detection unit identifier, two-dimensional coordinate position, functional area type, and process identifier.

3. The method for detecting defects in printing screens based on image analysis according to claim 1, characterized in that, The step of performing personalized image acquisition and preprocessing for each detection unit based on the detection unit description information to obtain the feature source data corresponding to each detection unit includes: Based on the detection unit description information, the corresponding imaging mode is selected and the detection unit image is acquired. Then, imaging consistency correction, noise reduction, contrast enhancement and background texture suppression are performed to generate feature source data. Feature source data includes candidate masks for opening regions, candidate masks for line structures, candidate masks for foreign objects, and grayscale statistics.

4. The method for detecting defects in printing screens based on image analysis according to claim 3, characterized in that, Based on the feature source data, each detection unit is mapped to a unit behavior profile vector, and cluster analysis and benchmark calculation are performed on the unit behavior profile vectors to obtain several sets of network usage patterns and corresponding set benchmarks, including: Calculate the aperture ratio based on the candidate masks for the opening regions in the feature source data; calculate the linewidth deviation based on the candidate masks for the line structures in the feature source data; Calculate the foreign object area distribution based on the foreign object candidate mask in the feature source data; calculate the transmittance statistic using the grayscale statistical image; After standardizing the aperture ratio, linewidth deviation, foreign object area distribution, and transmittance statistics, they are spliced ​​together in a fixed order to form a unit behavior profile vector. Cluster analysis is performed on the unit behavior profile vectors to divide the detection units into multiple network usage pattern sets; the set benchmark is determined by using the statistics of all detection unit behavior profile vectors within the network usage pattern set.

5. The method for detecting defects in printing screens based on image analysis according to claim 1, characterized in that, The resampling mechanism based on network usage pattern sets and behavioral profiles calculates the short-term stability index and the general population index of the detection unit, and combines them with fusion rules to obtain the resilience index, including: Within a preset time window, each detection unit is subjected to at least two repeated image acquisitions, and a corresponding unit behavior profile vector is constructed based on the feature source data obtained from each acquisition. The unit behavior profile vector is compared with the set benchmark of the network pattern set to which the corresponding detection unit belongs, and the Mahalanobis distance corresponding to each sampling is calculated. Short-term stability indices are calculated based on Mahalanobis distance sequences, and population universality indices are calculated based on the mean of Mahalanobis distance. After limiting and normalizing the short-term stability indicators and the general population indicators, a weighted sum or multiplicative fusion is performed to obtain the resilience index.

6. The method for detecting defects in printing screens based on image analysis according to claim 1, characterized in that, The step of combining the toughness index with the functional area weights and historical defect weights in the detection unit description information to generate monitoring priorities includes: The risk quantity is obtained by inverse mapping of the resilience index, and the risk quantity is then weighted and fused or interactively enhanced with the functional area weight and historical defect weight of the detection unit to obtain the monitoring priority.

7. The method for detecting defects in printing screens based on image analysis according to claim 1, characterized in that, The process of generating a tiered collection task set based on monitoring priority and contextual multiple reward combinations includes: Obtain all detection units with monitoring priority and configure context information vectors for each detection unit, including resilience index, functional area type, functional area weight, and historical defect weight; A set of preset acquisition actions is provided. The acquisition parameters for each acquisition action include acquisition frequency, acquisition resolution, exposure gain, number of fields of view, scanning path, number of re-inspections, and the imaging channel used. A cost function is preset for each acquisition action, where the cost represents the resources consumed by the action in the runtime environment corresponding to the context information vector, including time, computing power, bandwidth, device usage, and cycle time impact; a total budget constraint is set to limit the total cost upper limit of the acquisition tasks that can be executed within the preset scheduling period; A basic reward is constructed for the detection unit and the acquisition action, and the basic reward is monotonically correlated with the monitoring priority; at least two context combination reward items are constructed based on the context information vector. The context combination reward items are used to characterize the gain or suppression of the acquisition benefit by the context factors. According to the preset multi-combination fusion rules, the basic reward and the context combination reward items are combined to form the total reward of the candidate collection action, and the unit cost benefit is calculated in combination with the collection action cost to represent the collection benefit under unit resource consumption. For each detection unit, enumerate the subset of actions that satisfy the feasibility constraints, and select the action with the highest unit cost benefit from the subset of actions to generate candidate tasks for the detection unit. Configure task scores and task costs for all candidate tasks, with task scores being monotonically correlated with total rewards and task costs being monotonically correlated with the cost function; Select the subset of tasks whose task costs satisfy the total budget constraint and whose total task score is the largest, and form the selected task set. The selected task set is divided into at least two levels according to the preset grading threshold, and different sets of collection parameters are bound to different levels to output the graded collection task set.

8. The method for detecting defects in printing screens based on image analysis according to claim 1, characterized in that, The step of re-acquiring images of the target detection unit according to the hierarchical acquisition task set and obtaining a new behavioral profile vector includes: Receive a set of hierarchical acquisition tasks, determine the target detection unit identifier, acquisition action identifier, and acquisition parameter identifier corresponding to any acquisition task, and determine the acquisition level; Under the condition of confirmed executability, the acquisition configuration and acquisition action are executed, and after acquiring new images and special resources, a new behavioral profile vector is obtained.

9. The method for detecting defects in printing screens based on image analysis according to claim 1, characterized in that, The method of using the Mahalanobis distance of the new behavior profile vector relative to the set reference of the network pattern set and the defect type discrimination rule to output abnormal defect detection results includes: For each detection unit, calculate the Mahalanobis distance of the new behavior profile vector relative to the set reference of the set of network patterns to which it belongs, and use it as the deviation set; The rules for determining defect types include the rules for opening size deviation, line width deviation, broken line bridging, abnormal polyimide film bonding, and foreign matter contamination. By comparing the deviation set with the defect type discrimination rule and combining it with the conflict resolution rule, the abnormal defect detection result is obtained.

10. A printing screen defect detection system based on image analysis, used to implement the printing screen defect detection method based on image analysis according to any one of claims 1-9, characterized in that, include: The unit configuration module is used to divide the printing screen to be inspected into several inspection units and configure inspection unit description information for each inspection unit; The acquisition and preprocessing module is used to perform personalized image acquisition and preprocessing for each detection unit based on the detection unit description information, so as to obtain the feature source data corresponding to each detection unit; The profile benchmark module is used to map each detection unit into a unit behavior profile vector based on feature source data, and to perform cluster analysis and benchmark calculation on the unit behavior profile vector to obtain several sets of network patterns and corresponding set benchmarks. The resilience assessment module is used to calculate the short-term stability index and the generality index of the detection unit based on the network usage pattern set and the behavior profile repeated sampling mechanism, and to obtain the resilience index by combining the fusion rules. The priority task module is used to combine the resilience index with the functional area weights and historical defect weights in the detection unit description information to generate monitoring priorities. Simultaneously, based on monitoring priority and combined with contextual multiple reward combinations, a tiered collection task set is generated; The defect detection module is used to re-acquire images of target detection units according to the hierarchical acquisition task set and obtain new behavioral profile vectors; using the Mahalanobis distance of the new behavioral profile vectors relative to the set reference of the network pattern set to which they belong and the defect type discrimination rules, the module outputs abnormal defect detection results.

Images