Intelligent printing quality detection method based on image recognition

Abstract

The invention relates to the technical field of intelligent printing quality detection, in particular to an intelligent printing quality detection method based on image recognition. The system comprises an image preprocessing module, a physical model construction module, a deformation iterative calculation module, an energy decoupling analysis module and a closed-loop feedback control module. The method comprises the steps of generating a feature point matrix and a to-be-detected feature field by acquiring standard and real-time image data; the core of the method is to construct a virtual tensor spring network model, map image differences into an external acting force field, and solve deformation potential energy distribution by using finite element iteration; a Laplace operator is adopted to execute energy decoupling, low-frequency isotropic stress and high-frequency topological potential energy singular points are separated, and a residual thermodynamic diagram is generated according to the singular points to position defects and adjust transmission tension; according to the method, the transformation from pure image comparison to physical potential energy evolution is realized, the mutual exclusion limitation of deformation tolerance and detection sensitivity is broken through, and the nonlinear deformation interference in the high-speed transmission of the flexible material can be effectively filtered.

Description

Technical Field

[0001] This invention relates to the field of intelligent printing quality inspection technology, specifically to an intelligent printing quality inspection method based on image recognition. Background Technology

[0002] In modern high-speed gravure and flexographic printing production environments, online quality inspection of flexible substrates such as plastic film and paper is a core element in ensuring production efficiency. To achieve automated quality inspection, existing solutions generally employ visual inspection technologies based on pixel grayscale contrast or traditional pattern recognition. While these solutions possess some processing capabilities in low-speed scenarios, the nonlinear physical deformation and random tension fluctuations that occur in flexible materials during high-speed transmission cause significant displacement deviations between the image under test and the reference master. Existing algorithms struggle to effectively decouple normal physical stretching from actual printing defects at the mathematical level, often resulting in a technical contradiction between high robustness and high sensitivity. Increasing deformation tolerance leads to missed defects, while enhancing the capture of minute features generates numerous false alarms. Furthermore, existing technologies primarily focus on discrete image analysis and lack a real-time feedback mechanism to convert detection results into physical control commands, making it difficult to dynamically correct the transmission state at the source to suppress deformation.

[0003] Therefore, how to achieve intelligent tolerance to complex physical deformations and accurately locate topological defects, while improving the closed-loop collaborative control capability of the detection system and the physical transmission mechanism, has become an urgent technical problem to be solved. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an intelligent printing quality inspection method based on image recognition. Specifically, the technical solution of this invention includes: S1. Acquire standard reference image data and real-time image data to be detected, and preprocess the standard reference image data and the real-time image data to be detected to generate a standard feature point matrix and feature field data to be tested; S2. Construct a virtual tensor spring network model based on the standard feature point matrix, define the virtual connection relationship and virtual spring coefficient between each mass node in the virtual tensor spring network model, and form an initial steady-state grid; S3. Map the data of the feature field to be measured as an external force field to the virtual tensor spring network model, perform finite element iterative calculation, solve the deformation potential energy distribution generated by the virtual tensor spring network model to match the data of the feature field to be measured, and generate global potential energy field data. S4. Perform energy decoupling processing on the global potential energy field data, use the Laplace operator to separate the low-frequency isotropic stress component and the high-frequency topological potential energy singularity component, and generate a potential energy residual thermogram based on the high-frequency topological potential energy singularity component. S5. Based on the potential energy residual thermogram, locate the coordinates of local anomalies and output the quality detection results. At the same time, generate tension closed-loop feedback control commands based on the low-frequency isotropic stress components to adjust the parameters of the physical transmission mechanism in real time.

[0005] Preferably, step S1 includes the following steps: S11. Acquire the original image signal of the continuous medium surface through a high-speed linear array acquisition device; S12. The original image signal is preprocessed at the hardware level using a field-programmable gate array to generate digitized standard reference image data and real-time image data to be detected; S13. The standard reference image data is converted into the standard feature point matrix by the data conversion and encoding module running in the processing unit, and the real-time image data to be detected is converted into the test feature field data with force field attributes. The test feature field data is used to characterize the real-time physical deformation state.

[0006] Preferably, step S2 includes the following steps: S21. Extract the key feature points from the standard feature point matrix and define them as quality nodes to construct a gridded topology; S22. The virtual spring coefficient between the quality nodes is dynamically determined based on the local texture complexity of the standard reference image data, and a larger virtual spring coefficient is assigned to regions with texture complexity higher than a preset complexity threshold. S23. Establish the virtual tensor spring network model, which is configured to simulate the elastic deformation response of a continuous medium under external force, rather than directly calculating pixel grayscale differences.

[0007] Preferably, step S3 includes the following steps: S31. The measured feature field data is regarded as a continuous external force field applied to the initial steady-state grid; S32. Run the finite element analysis algorithm in the graphics processor core, perform multiple fast iterations on the virtual tensor spring network model, and calculate the total potential energy required for the initial steady-state mesh to reach the matching state under the action of the external force field; S33. Record the stress distribution state of each node inside the mesh to generate global potential energy field data containing global potential energy distribution information. The global potential energy field data reflects the physical energy required to stretch the standard reference image data to match the real-time image data to be detected.

[0008] Preferably, step S4 includes the following steps: S41. Spatial filtering is performed on the global potential energy field data using the Laplace operator; S42. Extract the low-frequency energy component from the filtering result as the low-frequency isotropic stress component, wherein the low-frequency isotropic stress component characterizes the overall physical deformation of the continuous medium. S43. Extract the high-frequency residual part from the filtering result as the high-frequency topological potential singularity component, the high-frequency topological potential singularity component characterizing topological tearing or local potential abrupt change that cannot be explained by continuous physical stretching. S44. Map the high-frequency topological potential singularity component to a visualized potential energy residual thermogram, where the thermogram values ​​represent the severity of local defects.

[0009] Preferably, step S5 includes the following steps: S51. Set an anomaly detection threshold and identify regions in the potential energy residual thermal map where the thermal value is greater than the anomaly detection threshold. S52. Perform inverse mapping on the identified region, calculate its position in the original image coordinate system, mark it as defect coordinates, and output the rejection signal to the controller; S53. Analyze the distribution characteristics of the low-frequency isotropic stress components and calculate the tension deviation value in the current physical transmission process; S54. The tension deviation value is converted into a servo motor adjustment signal, which serves as the tension closed-loop feedback control command to dynamically adjust the rotational speed or pressure of the rollers on both sides of the physical transmission mechanism to eliminate nonlinear deformation.

[0010] Preferably, the anomaly determination threshold in S51 is dynamically adjusted according to the following rules: If the mean value of the low-frequency isotropic stress component is greater than the preset stress reference value, it indicates that the overall deformation is large, and the anomaly judgment threshold is automatically increased to reduce the false alarm rate. If the mean value of the low-frequency isotropic stress component is less than or equal to the stress reference value, it indicates that the overall operation is stable and the preset reference anomaly judgment threshold is maintained.

[0011] Preferably, the method is applied to the inspection of flexible substrates with continuous physical surfaces and is configured to execute the following decision logic: Deformation tolerance determination: For the difference identified as the low-frequency isotropic stress component, even if it causes pixel-level mismatch, it is determined to be normal physical deformation and no defect alarm is triggered; Topological tear determination: For differences identified as high-frequency topological potential energy singularity components, regardless of the magnitude of pixel grayscale change, as long as the mesh topological continuity is disrupted, it is determined to be a printing quality defect. Real-time control: The processing and feedback control of a single frame image are limited to a preset time window to match the operating rhythm of the physical transmission mechanism.

[0012] Compared with the prior art, the present invention has the following beneficial effects: 1. This method creatively transforms image differences into a physical potential energy field by constructing a virtual tensor spring network model, and uses mathematical operators to achieve orthogonal separation of normal physical deformation and real printing defects at the mathematical level. This approach breaks the limitation of the traditional detection scheme where deformation tolerance and detection sensitivity are mutually restrictive, and can effectively filter out nonlinear physical deformation interference caused by high-speed transmission of flexible materials. While ensuring high tolerance to material stretching, it accurately captures minute topological defects.

[0013] 2. This method not only identifies defects, but also uses the decoupled low-frequency isotropic stress components to analyze the tension deviation in the physical transmission process in real time, and generates closed-loop feedback control commands to dynamically adjust the transmission mechanism. This mechanism realizes deep collaboration between the detection system and the production mechanism, and can dynamically correct the transmission state and suppress deformation from the source, thereby shifting from simply rejecting defective products to reducing defective product generation, and significantly improving the stability of the entire production process.

[0014] 3. By introducing a dynamic spring coefficient based on texture complexity and an adaptive decision threshold logic, this method can simulate the cognitive characteristics of human observers, strictly monitor sensitive areas with rich details, and show intelligent tolerance for flat areas. This adaptive mechanism enables the system to adjust its sensitivity in real time according to fluctuations in production line speed or changes in material batches, and maintains an extremely low false alarm rate and an extremely high detection rate under complex and ever-changing production conditions.

[0015] 4. This method adopts hardware-level preprocessing and high-performance computing architecture. Through the parallel collaboration of field-programmable gate arrays and graphics processors, it ensures that massive image data can complete the entire process from acquisition, physical simulation to feedback control within a millisecond-level time window. This efficient computing chain enables complex finite element analysis to perfectly match the operating cycle of modern high-speed printing presses, solving the problem of high-precision algorithms being difficult to implement in real time in large-scale industrial production. Attached Figure Description

[0016] The present invention will be further explained below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0018] Example 1: Please see Figure 1 A smart printing quality inspection method based on image recognition includes the following steps: S1. Acquire standard reference image data and real-time image data to be detected, and preprocess the standard reference image data and real-time image data to be detected to generate standard feature point matrix and feature field data to be tested; S2. Construct a virtual tensor spring network model based on the standard feature point matrix, define the virtual connection relationship and virtual spring coefficient between each mass node in the virtual tensor spring network model, and form an initial steady-state grid; S3. Map the data of the feature field to be measured as an external force field to the virtual tensor spring network model, perform finite element iterative calculation, solve the deformation potential energy distribution generated by the virtual tensor spring network model to match the data of the feature field to be measured, and generate global potential energy field data. S4. Perform energy decoupling processing on the global potential energy field data, use the Laplace operator to separate the low-frequency isotropic stress component and the high-frequency topological potential energy singularity component, and generate a potential energy residual thermogram based on the high-frequency topological potential energy singularity component. S5. Based on the potential energy residual thermogram, locate the coordinates of local anomalies and output the quality detection results. At the same time, generate tension closed-loop feedback control commands based on low-frequency isotropic stress components to adjust the parameters of the physical transmission mechanism in real time.

[0019] This embodiment discloses the overall process of an intelligent printing quality inspection method, which is applied to the online inspection system of a high-speed gravure printing machine or flexographic printing machine. S1. Data Acquisition and Preprocessing: The system acquires standard reference image data and real-time image data to be inspected on the current production line in parallel through an industrial camera. The standard reference image data refers to a pre-stored digital master that has been manually confirmed to be defect-free, and its function is to serve as a static benchmark for physical simulation. The real-time image data to be inspected refers to the streaming data that is collected in real time during the production process and contains potential deformations and defects. The preprocessing module aligns the two and converts them into a data format suitable for physical calculation, generating a standard feature point matrix and the feature field data to be measured. S2. Construct a virtual physical model. Based on the standard feature point matrix, the system constructs a virtual tensor spring network model. The virtual tensor spring network model refers to a computational architecture that does not directly compare pixel gray values ​​but simulates the physical properties of materials. In this model, the feature points of the image are defined as nodes with mass, and the nodes are connected by virtual springs. The key to this step is to define the virtual connection relationship and virtual spring coefficient between each mass node, thereby forming an initial steady-state grid in the lowest energy state. S3. Finite Element Potential Energy Calculation: The system treats the target feature field data as an external force field and forcibly maps it onto the initial steady-state mesh. Using the finite element analysis algorithm, it calculates the energy accumulated in the entire network when the mesh deforms to match the target data, generating global potential energy field data. Global potential energy field data refers to a matrix where the value of each element represents the physical deformation potential energy that is forced to be generated at that location to match the real-time image. S4. Energy decoupling and feature separation are the core steps of this invention. The system performs energy decoupling processing on the global potential energy field data. Using the Laplace operator as a spatial filter, the total energy field is separated into two orthogonal components: the low-frequency isotropic stress component represents a large-scale, smooth stretching or twisting; the high-frequency topological potential singularity component represents a local, abrupt energy spike. Based on the latter, the system generates a potential energy residual heatmap, which intuitively reflects the distribution of non-physical differences, i.e., defects. S5. Result Output and Closed-Loop Control: The system locates local anomaly coordinates based on the potential energy residual thermogram. If the thermogram value exceeds the standard, an alarm or rejection signal is output. At the same time, the system creatively uses the separated low-frequency isotropic stress components to generate tension closed-loop feedback control commands, which directly adjust physical transmission mechanisms such as rollers to suppress deformation from the source. By converting image differences into potential energy fields, this invention successfully achieves mathematical orthogonal separation between normal deformation manifested as low-frequency stress and printing defects manifested as high-frequency potential energy singularities. This enables the system to accurately capture minute defects while tolerating nonlinear deformation of soft materials, thus resolving the technical contradiction in existing technologies where high robustness inevitably leads to low sensitivity.

[0020] Example 2: S1 includes the following steps: S11. Acquire the original image signal of the continuous medium surface through a high-speed linear array acquisition device; S12. The original image signal is preprocessed at the hardware level using a field-programmable gate array to generate digitized standard reference image data and real-time image data to be detected. S13. The standard reference image data is converted into a standard feature point matrix by the data conversion and encoding module running in the processing unit, and the real-time image data to be detected is converted into the feature field data to be measured with force field attributes. The feature field data to be measured is used to characterize the real-time physical deformation state.

[0021] This embodiment is a concretization of data acquisition and preprocessing, focusing on hardware acceleration and data stream construction; S11. Image Acquisition: In this embodiment, a high-speed linear array acquisition device is used to acquire the original image signal of the continuous medium surface; specifically, a high-speed linear array acquisition device with a line frequency of [missing information] is selected. of Line scan camera, through or The interface transmits data to adapt to a frequency of one minute. High-speed production lines; S12. FPGA hardware preprocessing: To reduce latency, this embodiment utilizes a field-programmable gate array (FPGA). Hardware-level preprocessing of the raw image signal; Internally configured with a parallel pipeline, it performs illumination correction, noise reduction, and distortion correction, directly generating digitized standard reference image data and real-time image data to be detected at the bitstream level, without requiring a host computer. resource; S13. Force field attribute conversion: The image data is converted into a physical model input through a preset feature extraction and encoding software algorithm module. Specifically, the algorithm module converts the texture gradient of the standard image into a standard feature point matrix as grid node coordinates, and converts the real-time image data to be detected into the feature field data to be measured. In this embodiment, the feature field data to be measured is defined as a vector field, the value of which represents the optical flow vector or gray-level gradient vector of the corresponding pixel in the real-time image, used to characterize the real-time physical deformation state. Based on this, the vector field is converted into a physical model input through a preset conversion operator. Mapped as external force The mapping logic follows the formula ,in, These are force field mapping coefficients used to convert pixel displacement intensity into mechanical amplitudes in the physical simulation domain, within this virtual physical model. The unit is defined as a virtual energy unit; pass Hardware anchoring with a high-speed camera ensures that the throughput of data acquisition and preprocessing meets the above industrial requirements, providing a low-latency real-time data stream for subsequent physics field calculations.

[0022] Example 3: S2 includes the following steps: S21. Extract key feature points from the standard feature point matrix and define them as quality nodes to construct a gridded topology; S22. Based on the local texture complexity of the standard reference image data, dynamically determine the virtual spring coefficient between quality nodes, and assign a larger virtual spring coefficient to regions with texture complexity higher than the preset complexity threshold; S23. Establish a virtual tensor spring network model. The virtual tensor spring network model is configured to simulate the elastic deformation response of a continuous medium under external force, rather than directly calculating pixel grayscale differences.

[0023] This embodiment is a concretization of the construction of a virtual physical model, and details the virtual tensor spring network. The construction logic is key to achieving intelligent deformation tolerance; S21. Mesh Topology Construction; The system extracts key feature points from the standard feature point matrix, such as edge corners and texture centers, and defines them as quality nodes, based on... Triangulation algorithm constructs a meshed topology; S22. Dynamic definition of virtual spring coefficient; To accurately simulate the physical properties of printed materials, this embodiment introduces a texture-adaptive stiffness model; For any connected edge in the mesh, its virtual spring coefficient... The specific calculation formula is as follows:

[0024] in, For texture weighting factors, It is the basic stiffness constant; Among them, nodes Local texture complexity at the location The calculation method is as follows:

[0025] in, For pixels grayscale value, This represents the average gray level within that neighborhood. Areas with complex textures, such as text and intricate patterns, are extremely sensitive to deformation; even the slightest misalignment should be considered a defect, and therefore require greater precision. The value represents high stiffness; while flat textured areas, such as large color blocks, are not sensitive to slight stretching and are assigned a smaller value. The value represents low stiffness; S23. Physical simulation configuration; Establish a virtual tensor spring network model; This model is configured to simulate the elastic deformation response of a continuous medium under external force; Unlike the traditional algorithm that directly subtracts pixel values, this model calculates the internal forces required to maintain the integrity of the topology. The spring coefficient is dynamically determined by the texture complexity. The model can understand images like a human observer: it is strict (high stiffness) in areas with rich detail and lenient (low stiffness) in areas with background. This biomimetic design significantly reduces the false alarm rate caused by the natural stretching of materials.

[0026] Example 4: S3 includes the following steps: S31. Treat the data of the feature field to be measured as a continuous external force field applied to the initial steady-state grid; S32. Run the finite element analysis algorithm in the graphics processor core, perform multiple fast iterations on the virtual tensor spring network model, and calculate the total potential energy required for the initial steady-state mesh to reach the matching state under the action of the external force field. S33. Record the stress distribution state of each node inside the mesh to generate global potential energy field data containing global potential energy distribution information. The global potential energy field data reflects the physical energy required to stretch the standard reference image data to match the real-time image data to be detected.

[0027] This embodiment is a concretization of finite element potential energy calculation, illustrating how to... The potential energy field is solved by finite element analysis. S31. External force field application: The system treats the feature field data to be measured as a continuous external force field applied to the initial steady-state grid; that is, it is assumed that the image to be measured is the result of a standard image after being subjected to some kind of mechanical action. This step aims to infer the magnitude of this force. S32 and GPU-accelerated iterations; in graphics processing units of The core runs a simplified finite element analysis algorithm; the system performs analysis on the virtual tensor spring network model. The rapid iteration aims to solve the problem of minimizing the total potential energy of the system. If the preset energy convergence threshold is not reached within the preset number of iterations, the state with the lowest total potential energy in the current iteration cycle is taken as the approximate matching solution to prioritize ensuring the real-time control requirements of the high-speed production line. In each iteration, the node coordinates The update follows the force equilibrium equation:

[0028] in, The preset virtual node quality, The values ​​of both are virtual damping coefficients, and are preset based on the unit consistency of the virtual physical domain to adjust the smoothness of the simulation evolution. and Representing nodes respectively The first-order velocity vector and the second-order acceleration vector in the virtual physics simulation evolution process; For external forces; the calculation of the global potential field data is based on the following generalized formula of Hooke's law:

[0029] in, For nodes The accumulated potential energy at the point is expressed in virtual energy units and is obtained through iterative calculation. For nodes The set of neighboring nodes; For the node after iteration and Spatial coordinates; For nodes and Initial Euclidean distance in the standard reference image; Through The system can perform parallel computation of the above formulas, enabling it to... The evolution simulation of the physical field is completed within the level; the global potential energy field data quantitatively reflects the physical energy required to stretch the standard image into a real-time image, providing a unique physical benchmark for distinguishing between normal stretching and abnormal defects.

[0030] Example 5: S4 includes the following steps: S41. Spatial filtering is performed on the global potential energy field data using the Laplace operator; S42. Extract the low-frequency energy component from the filtering result as the low-frequency isotropic stress component. The low-frequency isotropic stress component characterizes the overall physical deformation of the continuous medium. S43. Extract the high-frequency residual part from the filtering result as the high-frequency topological potential singularity component. The high-frequency topological potential singularity component represents the topological tear or local potential abrupt change that cannot be explained by continuous physical stretching. S44. Map the high-frequency topological potential singularity components to a visualized potential residual thermogram, where the thermogram values ​​represent the severity of local defects.

[0031] This embodiment is a concretization of energy decoupling and feature separation, describing how to use mathematical operators to achieve automatic separation of defects and deformations; S41. Laplace spatial filtering: Applying the discrete Laplace operator to global potential energy field data. Perform spatial filtering; Because this embodiment uses an irregular mesh topology based on Delaunay triangulation, the discrete Laplace operator is defined in graph Laplace form; for nodes... Potential energy value at Its spatially filtered output The calculation is as follows:

[0032] in, For nodes The degree, i.e., the number of neighboring nodes; For nodes The set of neighboring nodes; this operation achieves orthogonal separation of high and low frequency energy components by extracting the potential energy difference between each node and its neighborhood; S42. Extract low-frequency isotropic stress and low-frequency components from the filtering results; since the physical stretching of paper or film is usually a smooth change over the entire domain, it manifests as a low-frequency signal in the potential energy field; this part is extracted as low-frequency isotropic stress components, characterizing the overall physical deformation of the continuous medium, such as overall elongation. ; S43. Extract high-frequency topological potential energy singularities and extract the high-frequency residuals from the filtering results. When printing defects such as ink spots or missing prints occur, the standard image cannot be transformed into the image under test through continuous physical stretching. This causes the mesh to tear or distort greatly at the defect, thus forming singularities with abrupt energy changes in the potential energy field. This part was extracted as the high-frequency topological potential singularity component. S44. Heatmap generation: Maps the high-frequency topological potential energy singularity components into a visualized potential energy residual heatmap; the darker the color in the heatmap, the more severe the topological tearing at that location, i.e., the greater the possibility of defects. By leveraging the natural separation capability of the Laplacian operator for high-frequency and low-frequency signals, this invention achieves automated decoupling of deformation and defects; without the need for complex deep learning training, it can accurately identify image differences that cannot be explained by physical stretching, i.e., true defects, solely based on mathematical and physical properties.

[0033] Example 6: S5 includes the following steps: S51. Set an anomaly detection threshold and identify areas in the potential energy residual thermal map where the thermal value is greater than the anomaly detection threshold. S52. Perform inverse mapping on the identified region, calculate its position in the original image coordinate system, mark it as defect coordinates, and output the rejection signal to the controller; S53. Analyze the distribution characteristics of low-frequency isotropic stress components and calculate the tension deviation value in the current physical transmission process; S54. Convert the tension deviation value into a servo motor adjustment signal, which serves as a tension closed-loop feedback control command to dynamically adjust the speed or pressure of the rollers on both sides of the physical transmission mechanism to eliminate nonlinear deformation.

[0034] This embodiment is a concretization of the result output and closed-loop control, detailing the application of the detection results and the tension feedback control mechanism; S61-S62, Defect Judgment and Removal: Setting anomaly judgment thresholds; In the potential energy residual thermogram, all thermodynamic values... The regions are all locked; the system performs inverse mapping on these regions, calculates their pixel positions in the original image coordinate system, marks them as defect coordinates, and outputs signals to drive... The controller performs defective product rejection at the end of the production line; S63. Tension deviation calculation: Using the separated low-frequency isotropic stress components, the system calculates its transverse stress deviation. and longitudinal The average value is used to determine the tension deviation during the current physical transmission process; for example, if the low-frequency stress in the left area is generally higher than that in the right area, it indicates that the tension of the left roller is too large. S64, Closed-loop feedback control, converts the tension deviation value into a servo motor adjustment signal as a tension closed-loop feedback control command; this command is sent to the servo driver of the printing press in real time to dynamically fine-tune the speed difference or pressure of the rollers on both sides of the physical transmission mechanism to eliminate nonlinear deformation. Output speed compensation value of tension closed-loop feedback control command The calculation formula uses the discrete PID algorithm:

[0035] in, Indicates the current sampling time. Historical sampling time The calculated tension deviation value; These are the proportional, integral, and differential coefficients, respectively. This embodiment is not only a quality inspection system, but also a process optimization system. It uses low-frequency stress data, a byproduct generated during the inspection process, to optimize the production process in reverse, achieving a technological leap from passively rejecting waste to actively reducing waste generation.

[0036] Example 7: The anomaly detection threshold in S51 is dynamically adjusted according to the following rules: If the mean value of the low-frequency isotropic stress component is greater than the preset stress reference value, it indicates that the overall deformation is large, and the anomaly judgment threshold is automatically increased to reduce the false alarm rate. If the mean value of the low-frequency isotropic stress component is less than or equal to the stress reference value, it indicates that the overall operation is stable and the preset reference anomaly judgment threshold is maintained.

[0037] This embodiment further defines the dynamic adjustment logic of the anomaly detection threshold to adapt to different working conditions; In this embodiment, the anomaly detection threshold is not a fixed value, but is dynamically adjusted according to the following adaptive threshold formula:

[0038] in, The dynamic anomaly detection threshold used for the current frame; This is a preset baseline threshold. This is the mean value of the low-frequency isotropic stress components in the current frame; This is the preset stress reference value; This is the adjustment coefficient; Among them, stress reference value The acquisition method is as follows: Under no-load idling conditions on the production line, 100 standard images are continuously acquired and the average value of the global potential energy field is calculated. 1.2 times the maximum value is taken as the safety benchmark; adjustment coefficient... This is an empirical constant obtained by fitting the response curves of known defect samples under different tensile stresses using the least squares method. In this embodiment, its value range is [value range missing]. ; like If the average low-frequency stress is greater than the baseline value, it indicates that the material is currently in a state of large deformation, such as an accelerated phase or batch variation. In this case, the system automatically raises the judgment threshold; the purpose is to prevent false alarms caused by severe global stretching and reduce the false positive rate. This indicates that the operation is stable, maintaining the preset benchmark threshold and high sensitivity; This dynamic threshold mechanism gives the system environmental adaptability, ensuring that the system will not generate a large number of false alarms when the production line speed fluctuates or the material shakes violently, while maintaining an extremely high detection rate of minor defects during stable operation.

[0039] Example 8: The method is applied to the inspection of flexible substrates with continuous physical surfaces and is configured to execute the following decision logic: Deformation tolerance judgment: For differences identified as low-frequency isotropic stress components, even if they cause pixel-level mismatch, they are judged as normal physical deformation and no defect alarm is triggered. Topological tearing determination: For differences identified as high-frequency topological potential energy singularity components, regardless of the magnitude of the pixel grayscale change, as long as the grid topological continuity is disrupted, it is determined to be a printing quality defect. Real-time control: The processing and feedback control of a single frame image are limited to a preset time window to match the operating rhythm of the physical transmission mechanism.

[0040] This embodiment defines the applicable scenarios and core judgment logic of this method; This method is specifically designed for testing flexible substrates with continuous physical surfaces, such as plastic films, paper, and aluminum foil. Deformation tolerance criterion: For image differences separated into low-frequency isotropic stress components, even if they cause visible pixel displacement such as overall pattern shift, the deformation tolerance criterion is determined. For individual pixels, the system determines it as normal physical deformation and does not trigger a defect alarm; Topological tear judgment: For the difference that is separated into high-frequency topological potential energy singularity components, regardless of the magnitude of the pixel grayscale change, even if it is just a faint ink dot, as long as it destroys the topological continuity of the grid, i.e. the energy gradient is extremely large, the system determines it as a printing quality defect. The system forces the processing of single-frame images, i.e. With feedback control Limited to a preset time window, for example Completed within minutes, with strict matching every minute. The operating cycle time of physical transmission mechanisms exceeding 1 meter; Through clear logical definition, this embodiment ensures the usability of the algorithm in industrial settings, avoids misjudging harmless physical stretching as a quality problem, and guarantees the real-time response capability of high-speed production lines.

[0041] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for intelligent printing quality inspection based on image recognition, characterized in that, Includes the following steps: S1. Acquire standard reference image data and real-time image data to be detected, and preprocess the standard reference image data and the real-time image data to be detected to generate a standard feature point matrix and feature field data to be tested; S2. Construct a virtual tensor spring network model based on the standard feature point matrix, define the virtual connection relationship and virtual spring coefficient between each mass node in the virtual tensor spring network model, and form an initial steady-state grid; S3. Map the data of the feature field to be measured as an external force field to the virtual tensor spring network model, perform finite element iterative calculation, solve the deformation potential energy distribution generated by the virtual tensor spring network model to match the data of the feature field to be measured, and generate global potential energy field data. S4. Perform energy decoupling processing on the global potential energy field data, use the Laplace operator to separate the low-frequency isotropic stress component and the high-frequency topological potential energy singularity component, and generate a potential energy residual thermogram based on the high-frequency topological potential energy singularity component. S5. Based on the potential energy residual thermogram, locate the coordinates of local anomalies and output the quality detection results. At the same time, generate tension closed-loop feedback control commands based on the low-frequency isotropic stress components to adjust the parameters of the physical transmission mechanism in real time.

2. The intelligent printing quality inspection method based on image recognition according to claim 1, characterized in that, S1 includes the following steps: S11. Acquire the original image signal of the continuous medium surface through a high-speed linear array acquisition device; S12. The original image signal is preprocessed at the hardware level using a field-programmable gate array to generate digitized standard reference image data and real-time image data to be detected; S13. The standard reference image data is converted into the standard feature point matrix by the data conversion and encoding module running in the processing unit, and the real-time image data to be detected is converted into the test feature field data with force field attributes. The test feature field data is used to characterize the real-time physical deformation state.

3. The intelligent printing quality inspection method based on image recognition according to claim 2, characterized in that, S2 includes the following steps: S21. Extract the key feature points from the standard feature point matrix and define them as quality nodes to construct a gridded topology; S22. The virtual spring coefficient between the quality nodes is dynamically determined based on the local texture complexity of the standard reference image data, and a larger virtual spring coefficient is assigned to regions with texture complexity higher than a preset complexity threshold. S23. Establish the virtual tensor spring network model, which is configured to simulate the elastic deformation response of a continuous medium under external force, rather than directly calculating pixel grayscale differences.

4. The intelligent printing quality inspection method based on image recognition according to claim 3, characterized in that, S3 includes the following steps: S31. The measured feature field data is regarded as a continuous external force field applied to the initial steady-state grid; S32. Run the finite element analysis algorithm in the graphics processor core, perform multiple fast iterations on the virtual tensor spring network model, and calculate the total potential energy required for the initial steady-state mesh to reach the matching state under the action of the external force field; S33. Record the stress distribution state of each node inside the mesh to generate global potential energy field data containing global potential energy distribution information. The global potential energy field data reflects the physical energy required to stretch the standard reference image data to match the real-time image data to be detected.

5. The intelligent printing quality inspection method based on image recognition according to claim 4, characterized in that, S4 includes the following steps: S41. Spatial filtering is performed on the global potential energy field data using the Laplace operator; S42. Extract the low-frequency energy component from the filtering result as the low-frequency isotropic stress component, wherein the low-frequency isotropic stress component characterizes the overall physical deformation of the continuous medium. S43. Extract the high-frequency residual part from the filtering result as the high-frequency topological potential singularity component, the high-frequency topological potential singularity component characterizing topological tearing or local potential abrupt change that cannot be explained by continuous physical stretching. S44. Map the high-frequency topological potential singularity component to a visualized potential energy residual thermogram, where the thermogram values ​​represent the severity of local defects.

6. The intelligent printing quality inspection method based on image recognition according to claim 5, characterized in that, S5 includes the following steps: S51. Set an anomaly detection threshold and identify regions in the potential energy residual thermal map where the thermal value is greater than the anomaly detection threshold. S52. Perform inverse mapping on the identified region, calculate its position in the original image coordinate system, mark it as defect coordinates, and output the rejection signal to the controller; S53. Analyze the distribution characteristics of the low-frequency isotropic stress components and calculate the tension deviation value in the current physical transmission process; S54. The tension deviation value is converted into a servo motor adjustment signal, which serves as the tension closed-loop feedback control command to dynamically adjust the rotational speed or pressure of the rollers on both sides of the physical transmission mechanism to eliminate nonlinear deformation.

7. The intelligent printing quality inspection method based on image recognition according to claim 6, characterized in that, The anomaly detection threshold in S51 is dynamically adjusted according to the following rules: If the mean value of the low-frequency isotropic stress component is greater than the preset stress reference value, it indicates that the overall deformation is large, and the anomaly judgment threshold is automatically increased to reduce the false alarm rate. If the mean value of the low-frequency isotropic stress component is less than or equal to the stress reference value, it indicates that the overall operation is stable and the preset reference anomaly judgment threshold is maintained.

8. The intelligent printing quality inspection method based on image recognition according to claim 1, characterized in that, The method is applied to the inspection of flexible substrates with continuous physical surfaces and is configured to execute the following decision logic: Deformation tolerance determination: For the difference identified as the low-frequency isotropic stress component, even if it causes pixel-level mismatch, it is determined to be normal physical deformation and no defect alarm is triggered; Topological tear determination: For differences identified as high-frequency topological potential energy singularity components, regardless of the magnitude of pixel grayscale change, as long as the mesh topological continuity is disrupted, it is determined to be a printing quality defect. Real-time control: The processing and feedback control of a single frame image are limited to a preset time window to match the operating rhythm of the physical transmission mechanism.

Images