A digital printing quality detection method based on image analysis

Abstract

The application discloses a kind of based on image analysis digital printing quality detection method, it is related to digital printing quality intelligent detection technical field, including, obtain the surface texture feature of visible light channel in HDR image group and the substrate penetration feature of infrared channel, cross-modal correlation is carried out through space-time attention mechanism, generates three-dimensional feature cube;Material physical characteristics of printed matter are obtained using reverse Monte Carlo ray tracing algorithm, and enhanced feature tensor is generated by three-dimensional convolution network splicing, and defect detection and positioning are carried out through cross-scale feature aggregation;Improved DBSCAN clustering algorithm is used to generate complete defect area by spatial aggregation of defect coordinates, and defect detection report is output by dynamic threshold classification.The application realizes the accurate inversion of ink thickness, roughness and other material characteristics by constructing three-dimensional feature cube containing surface texture and substrate penetration feature, and improves the accuracy and reliability of digital printing quality detection.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection technology for digital printing quality, and in particular to a digital printing quality inspection method based on image analysis. Background Technology

[0002] In the field of digital printing quality inspection, existing technologies generally employ multi-band imaging and high dynamic range (HDR) fusion methods, combined with machine vision algorithms, to achieve defect detection. Conventional methods acquire reflected light data from the printed surface using multi-band light sources (such as visible light and short-wave infrared), utilize HDR technology to compensate for the dynamic range of highly reflective areas, and then extract surface texture features and substrate penetration characteristics. For example, multiple images are acquired through short-exposure, medium-exposure, and long-exposure sequences, fused using the Mertens algorithm to generate an HDR image set, and combined with convolutional neural networks (such as MobileNet and ResNet) to extract multi-scale features. For defect detection, traditional methods typically use clustering algorithms such as DBSCAN to spatially aggregate abnormal regions, combining geometric parameters (such as area and aspect ratio) and energy parameters (such as average intensity) to achieve defect classification. Furthermore, existing technologies integrate visible light and infrared features through cross-modal feature association (such as spatiotemporal attention mechanisms) to improve the comprehensiveness of defect detection. These methods have achieved certain results in the detection of ordinary printed areas (such as paper and plastic substrates), effectively identifying common defects such as misregistration and missing ink dots.

[0003] However, existing methods still have two limitations: First, when processing highly reflective areas such as metallic inks, traditional HDR fusion algorithms (such as the Mertens algorithm) use a uniform weight calculation strategy, which makes it difficult to optimize the highlight details of metallic areas and the color transition of ordinary areas synchronously. For example, the specular reflection characteristics of metallic inks can easily cause local overexposure, while the low reflectivity features of ordinary areas may produce color shift due to uneven distribution of fusion weights. Second, in the process of cross-modal feature association, the dynamic registration accuracy of visible light and infrared features is insufficient. Especially in the case of mechanical vibration or substrate deformation, the spatial alignment error of multi-scale feature maps will significantly reduce the representation ability of the three-dimensional feature cube, thereby affecting the inversion accuracy of material physical properties (such as ink thickness and surface roughness). Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a digital printing quality inspection method based on image analysis to solve the problem of reduced accuracy in material property inversion caused by loss of detail in HDR fusion of metallic inks and visible light-infrared feature registration errors.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a digital printing quality inspection method based on image analysis, comprising: acquiring surface reflected light data and ambient light data of printed matter; performing high reflectivity compensation on metallic ink areas through dynamic adjustment of multi-band light sources to generate an HDR image set; acquiring surface texture features of the visible light channel and substrate penetration features of the infrared channel in the HDR image set; performing cross-modal association through a spatiotemporal attention mechanism to generate a three-dimensional feature cube; employing a reverse Monte Carlo ray tracing algorithm to acquire the material physical properties of printed matter; generating an enhanced feature tensor through three-dimensional convolutional network splicing; and performing defect detection and localization through cross-scale feature aggregation; employing an improved DBSCAN clustering algorithm to spatially aggregate defect coordinates to generate a complete defect region; and outputting a defect detection report through dynamic threshold classification.

[0008] As a preferred embodiment of the image analysis-based digital printing quality inspection method of the present invention, the step of dynamically adjusting the multi-band light source to perform high reflectivity compensation on the metallic ink area and generating an HDR image set includes the following steps.

[0009] A mask for highly reflective areas on the surface of printed materials is generated by combining edge gradient analysis with reflectivity differences.

[0010] The data of reflected light on the surface of printed materials are dynamically adjusted. Images of metallic ink areas and ordinary areas on the surface of printed materials are acquired in short exposure, medium exposure and long exposure sequences, and then HDR image groups are generated by fusing them through an improved Mertens algorithm.

[0011] As a preferred embodiment of the image analysis-based digital printing quality detection method of the present invention, the surface texture features of the visible light channel and the substrate penetration features of the infrared channel in the HDR image group are obtained through the MobileNetV3 network and the ConvNeXt network, respectively.

[0012] As a preferred embodiment of the image analysis-based digital printing quality inspection method of the present invention, the step of generating a three-dimensional feature cube through cross-modal association using a spatiotemporal attention mechanism is as follows:

[0013] Multi-scale feature maps for the visible light band are generated by fusing multi-level feature pyramids.

[0014] Multi-scale feature maps in the short-wave infrared band are generated by cross-scale attention aggregation.

[0015] By employing a spatiotemporal cross-attention mechanism, cross-modal feature association is performed on multi-scale feature maps in the visible light band and the short-wave infrared band to generate a three-dimensional feature cube.

[0016] As a preferred embodiment of the image analysis-based digital printing quality inspection method of the present invention, the steps for obtaining the material physical properties of the printed matter using the reverse Monte Carlo ray tracing algorithm are as follows:

[0017] By inverting the infrared characteristic response values ​​at each spatial location in the multi-scale characteristic map of the short-wave infrared band using the ink absorbance-thickness relationship curve, the ink layer thickness distribution is generated.

[0018] By comparing the varnish curing degree calibration curve with the dual-band characteristic ratio analysis method, and using nonlinear least squares fitting, a varnish curing degree thermogram was generated.

[0019] By using the micro-surface BRDF fitting algorithm, surface normal distribution modeling and scattering energy analysis are performed on the multi-scale feature map of visible light band in a three-dimensional feature cube to obtain the surface roughness parameters of the substrate.

[0020] As a preferred embodiment of the image analysis-based digital printing quality detection method of the present invention, the step of generating an enhanced feature tensor by splicing three-dimensional convolutional networks refers to the following: the first layer of multi-scale convolution identifies shallow joint features; the second layer of dilated convolution expands the shallow joint features to generate a deep feature tensor; and the third layer of cross-modal fusion convolution performs cross-modal skip splicing of HDR image groups and deep feature tensors, and combines a cross-modal attention mechanism to generate an enhanced feature tensor.

[0021] As a preferred embodiment of the image analysis-based digital printing quality inspection method of the present invention, the defect detection and localization through cross-scale feature aggregation comprises the following steps.

[0022] Based on the enhanced feature tensor, channel joint normalization calculation is performed through cross-scale feature aggregation to predict the surface anomaly score of printed materials and compare it with the defect threshold to identify defects on the surface of printed materials.

[0023] Subpixel-level coordinate boundaries of surface defects in printed materials are obtained through nonmaximum suppression and connected component analysis.

[0024] As a preferred embodiment of the image analysis-based digital printing quality inspection method of the present invention, the steps are as follows: The improved DBSCAN clustering algorithm is used to spatially aggregate defect coordinates to generate a complete defect region, and a defect detection report is output through dynamic threshold classification.

[0025] A regional attribute analysis algorithm is used to obtain the geometric parameters and defect energy parameters of each defect coordinate cluster. At the same time, a decision tree classifier is used to segment the data, and a dynamic process threshold is defined as the decision boundary of the decision tree classifier to generate defect level labels.

[0026] A defect detection report is generated based on the defect level label, subpixel coordinate bounding box, defect geometric parameters, and defect energy parameters.

[0027] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the image analysis-based digital print quality inspection method described in the first aspect of the present invention.

[0028] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the image analysis-based digital print quality inspection method described in the first aspect of the present invention.

[0029] The beneficial effects of this invention are as follows: By using the improved Mertens algorithm, the fusion weights of the metallic ink region and the ordinary region are dynamically adjusted using a high-reflectivity region mask, achieving synchronous optimization of highlight details and color transitions during the image pyramid fusion stage; at the same time, by using a spatiotemporal cross-attention mechanism to perform cross-modal correlation of visible light and infrared multi-scale feature maps, a three-dimensional feature cube containing surface texture and substrate penetration features is constructed, realizing accurate inversion of material properties such as ink thickness and roughness, and improving the accuracy and reliability of digital printing quality inspection. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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.

[0031] Figure 1 This is a flowchart of a digital printing quality inspection method based on image analysis.

[0032] Figure 2 This is a flowchart for generating HDR image groups based on multi-exposure fusion.

[0033] Figure 3 A flowchart for generating a 3D feature cube based on a spatiotemporal cross-attention mechanism.

[0034] Figure 4 This is a flowchart of material property inversion and defect detection based on reverse Monte Carlo ray tracing. Detailed Implementation

[0035] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0036] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0037] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0038] Reference Figures 1-4 This is one embodiment of the present invention, which provides a digital printing quality inspection method based on image analysis, including the following steps:

[0039] S1. Collect data on reflected light from the surface of printed materials and ambient light data. Use multi-band light source dynamic adjustment to compensate for high reflectivity in the metallic ink area and generate HDR image group.

[0040] Data on reflected light from printed surfaces includes reflected light signals in the visible light band and reflected light signals in the short-wave infrared band;

[0041] Ambient light data includes ambient light intensity and ambient light angle;

[0042] It should be noted that the reflected light signals in the visible light band and the reflected light signals in the short-wave infrared band are acquired by a high-precision industrial camera, the ambient light intensity is measured by an illuminance meter, and the ambient light illumination angle is recorded by a three-axis gyroscope and a photoelectric encoder.

[0043] Based on the reflected light data of the printed surface and ambient light data, a mask for the highly reflective area of ​​the printed surface is generated by combining edge gradient analysis with reflectivity differences.

[0044] Furthermore, the Sobel edge detection algorithm is used to process the visible light band reflected light signals acquired by a high-precision industrial camera, identifying high-reflectivity pixels with reflectivity exceeding the reflectivity feature threshold. Simultaneously, a region growing algorithm is used to process the short-wave infrared band reflected light signals, marking low-reflectivity pixels with reflectivity below the reflectivity feature threshold. The high-reflectivity pixels identified by the Sobel edge detection algorithm in the visible light band reflected light signals are spatially matched with the low-reflectivity pixels marked by the region growing algorithm in the short-wave infrared band reflected light signals. A morphological closing operation algorithm is then used to fuse successfully matched adjacent pixels to form a continuous region. The final output is a high-reflectivity region mask containing both metallic ink regions and ordinary regions.

[0045] It should be noted that the reflectance characteristic threshold is set based on a comparison of the standard reflectance spectral curves of metallic ink and ordinary ink in the visible light and short-wave infrared bands. The value range is [0.1, 0.3].

[0046] Based on the high reflectivity mask on the surface of the printed material, the reflected light data of the printed material surface is dynamically adjusted, and images of the metallic ink area and the ordinary area on the surface of the printed material are acquired in short exposure, medium exposure and long exposure sequence.

[0047] Furthermore, a mask representing the highly reflective area of ​​the printed surface is used to guide the selection of exposure parameters. The metallic ink area is captured using a short exposure time, while the ordinary area is captured using medium and long exposure times respectively. The exposure parameters are dynamically determined based on the spatial distribution of the high-reflectivity mask. The short exposure time is set for the high reflectivity characteristics of the metallic ink area, the medium exposure time is set for the main reflective features of the ordinary area, and the long exposure time is set for the low-reflectivity details of the ordinary area. This exposure sequence is used to acquire images of the metallic ink area and the ordinary area, respectively. The image of the metallic ink area retains high-reflectivity details, while the image of the ordinary area includes images with two different brightness levels: medium and long exposures.

[0048] An improved Mertens algorithm is used to fuse images of metallic ink regions and ordinary regions to generate an HDR image set.

[0049] The improved Mertens algorithm and the process of generating HDR image sets are as follows: In the existing Mertens algorithm's multi-exposure fusion framework based on weighted graphs, considering the difference in reflectivity between metallic ink areas and ordinary areas on the printed surface, a mask of the highly reflective area on the printed surface is used as a dynamic adjustment factor. A weight calculation strategy based on local contrast enhancement is adopted for the metallic ink area image, while the brightness and saturation weight calculation method of the traditional Mertens algorithm is retained for the ordinary area image. In the image pyramid fusion stage, the highly reflective area mask on the printed surface guides the metallic ink area image and the ordinary area image to undergo adaptive weighted mixing at each layer of the Laplacian pyramid, ensuring that the highlight details of the metallic ink area and the color transition of the ordinary area are synchronously optimized and presented in the HDR image set.

[0050] S2. Obtain the surface texture features of the visible light channel and the substrate penetration features of the infrared channel in the HDR image group, and perform cross-modal correlation through the spatiotemporal attention mechanism to generate a three-dimensional feature cube;

[0051] Subpixel-level registration correction is performed on the surface reflection data of printed materials in the HDR image group to eliminate the offset caused by mechanical vibration.

[0052] Furthermore, the surface reflection data of printed materials in the HDR image set were used for sub-pixel registration and correction using a phase correlation algorithm. Fourier transform was employed to calculate the inter-image displacement vector, and a pyramid layering strategy was combined to achieve displacement estimation, performing coarse registration at low resolution followed by finer registration at high resolution. Bilinear interpolation was used to complete displacement compensation, preserving high-frequency details of the image. Adaptive thresholding was used during the registration process to handle motion differences in different regions, ultimately eliminating offsets caused by mechanical vibration.

[0053] Based on the corrected visible light band reflected light signal, the surface texture features (including overprinting error and ink spot missing features) of the visible light channel are extracted by the MobileNetV3 network, and multi-scale feature maps of the visible light band are generated by multi-level feature pyramid fusion.

[0054] Furthermore, the corrected visible light band reflected light signal is processed through the depthwise separable convolutional structure of the MobileNetV3 network, and surface texture features at different scales are extracted progressively using the inverse residual method. Shallow processing extracts basic texture features such as edges and corners, mid-level processing captures overprinting deviation features, and deep processing identifies subtle defects such as missing ink dots. The multi-level feature pyramid fusion process adopts a top-down and horizontally connected structure, aligning the feature maps at different scales in terms of resolution and stitching together channels to form a multi-scale visible light band feature map with a rich hierarchical structure. The fusion process uses an attention mechanism to dynamically adjust the weights of texture features at different levels, highlighting channels important for defect detection. The final generated multi-scale visible light band feature map contains complete texture information from microscopic to macroscopic levels.

[0055] Based on the corrected short-wave infrared band reflected light signal, the substrate penetration characteristics (including uncured ink and fiber breakage characteristics) of the infrared channel are extracted by ConvNeXt network, and multi-scale feature maps of the short-wave infrared band are generated by cross-scale attention aggregation.

[0056] Furthermore, the corrected short-wave infrared reflected light signal is processed through a hierarchical convolutional structure of the ConvNeXt network, employing large-kernel depthwise convolution to capture long-range dependencies. The feature extraction process focuses on substrate penetration characteristics; shallow processing identifies uncured ink features, while deep processing detects fiber breakage features. A cross-scale attention aggregation mechanism establishes feature associations through both spatial and channel dimensions, employing a progressive upsampling strategy to maintain the spatial continuity of the feature map. The resulting multi-scale feature map of the short-wave infrared band comprehensively characterizes the substrate penetration features, including both uncured ink and fiber breakage characteristics.

[0057] By employing a spatiotemporal cross-attention mechanism, cross-modal feature association is performed on multi-scale feature maps in the visible light band and the short-wave infrared band to generate a three-dimensional feature cube.

[0058] Furthermore, when processing multi-scale feature maps in the visible light and short-wave infrared bands using the spatiotemporal cross-attention mechanism, a cross-modal feature correlation matrix is ​​first established. Bidirectional attention calculations are performed on the spatial dimension of the visible light multi-scale feature map and the channel dimension of the short-wave infrared multi-scale feature map to generate spatial-channel interaction weights. These interaction weights guide the feature recombination process, deeply fusing the surface texture features of the visible light multi-scale feature map with the substrate penetration features of the short-wave infrared multi-scale feature map. Feature recombination employs pointwise convolution operations, expanding the feature representation along the channel dimension while maintaining spatial resolution. The fused features undergo layer normalization and residual connections to form a three-dimensional feature cube with spatiotemporal consistency. The height and width dimensions of the three-dimensional feature cube retain the original image spatial information, while the depth dimension integrates cross-modal features, achieving a unified representation of overprinting deviation features, ink dot loss features, uncured ink features, and fiber breakage features.

[0059] S3. The reverse Monte Carlo ray tracing algorithm is used to obtain the material physical properties of printed materials. An enhanced feature tensor is generated by splicing three-dimensional convolutional networks, and defect detection and localization are performed by cross-scale feature aggregation.

[0060] The physical properties of printed materials include ink layer thickness distribution, substrate surface roughness parameters, and varnish curing temperature profile.

[0061] By inverting the infrared characteristic response values ​​at each spatial location in the multi-scale characteristic map of the short-wave infrared band using the ink absorbance-thickness relationship curve, the ink layer thickness distribution is generated.

[0062] Furthermore, the infrared feature response values ​​at each spatial location in the short-wave infrared multi-scale feature map are matched with the ink absorbance-thickness relationship curve to obtain the preliminary thickness value of each pixel in the multi-scale feature map. The infrared feature response values ​​are then compared with the ink absorbance-thickness relationship curve using an interpolation algorithm to convert the preliminary thickness estimate into a thickness value with sub-pixel accuracy, while maintaining the spatial correspondence during the conversion process. This ultimately generates the ink layer thickness distribution.

[0063] By comparing the varnish curing degree calibration curve with the dual-band characteristic ratio analysis method, and using nonlinear least squares fitting, a varnish curing degree thermogram was generated.

[0064] Furthermore, based on the spectral absorption characteristics of the substrate's permeation features, two characteristic absorption bands were selected, and the center wavelength of each band was determined using the peak detection method. The ratio of the response values ​​of the two bands at each spatial location was compared point-by-point with the varnish curing degree calibration curve. The varnish curing degree calibration curve was established using differential scanning calorimetry (DSC) measured data, characterizing the nonlinear mapping relationship between the characteristic ratio and the degree of curing. The comparison process optimized the correspondence between the characteristic ratio and the degree of curing through nonlinear least squares fitting, maintaining spatial consistency during the fitting process, ultimately generating a varnish curing degree thermogram.

[0065] By using the micro-surface BRDF fitting algorithm, surface normal distribution modeling and scattering energy analysis are performed on the multi-scale feature map of visible light band in a three-dimensional feature cube to obtain the surface roughness parameters of the substrate.

[0066] Furthermore, a framework for describing reflection characteristics is established based on microfacet theory. Spatial gradient information from the multi-scale feature map in the visible light band is used to construct the surface normal distribution, and an anisotropic Gaussian distribution is employed to characterize the micro-surface orientation properties. Reflection behavior under multi-angle illumination conditions is simulated using the Monte Carlo ray tracing method. The simulation results (anisotropic Gaussian distribution parameters output by the micro-surface BRDF fitting algorithm that match the measured reflection characteristics of the multi-scale feature map in the visible light band) are iteratively optimized with the measured reflection characteristics of the multi-scale feature map in the visible light band. During the optimization process, the micro-surface height field parameters are adjusted to achieve the best match between the reflection characteristic description and the reflection characteristics of the multi-scale feature map in the visible light band, thereby obtaining the substrate surface roughness parameters.

[0067] It should be noted that the ink absorbance-thickness relationship curve is obtained through the following process: under standard temperature and humidity conditions, the absorbance values ​​of ink samples of different known thicknesses in the short-wave infrared band are measured using a spectrometer, and the ink absorbance-thickness relationship curve is established by fitting using the least squares method.

[0068] The process of obtaining the curing degree calibration curve of varnish: Based on the reflected light signal of the short-wave infrared band, the actual curing degree is measured by differential scanning calorimetry as the ordinate. The corresponding relationship curve between the curing degree of varnish and the ratio of characteristic absorption peaks is established by polynomial regression fitting, and finally the curing degree calibration curve of varnish is output.

[0069] A multimodal affine transformation registration algorithm is used to spatially register a 3D feature cube and material physical properties to generate a multi-source number set after registration.

[0070] Furthermore, the multimodal affine transformation registration algorithm establishes geometric transformation relationships based on translation, rotation, and scaling parameters, and optimizes parameters by matching multi-scale feature points of the 3D feature cube with spatial key points of material physical properties. The matching process employs a hierarchical strategy, first performing low-resolution coarse registration followed by high-resolution fine registration, ensuring consistency between the visible light and short-wave infrared multi-scale feature maps of the 3D feature cube and the spatial representation of material physical properties. Finally, a registered multi-source data set is generated.

[0071] By using hierarchical feature recombination and zero-filling expansion, the registered multi-source fusion data set is structured and organized according to the channel dimension, depth dimension and spatial dimension to construct a standardized three-dimensional tensor.

[0072] Furthermore, the quasi-resolved multi-source data set is structured through hierarchical feature recombination and zero-filling expansion. Channel dimension recombination arranges the visible light multi-scale feature maps, short-wave infrared multi-scale feature maps, ink layer thickness distribution, substrate surface roughness parameters, and varnish curing thermal maps in categorized order; the depth dimension is uniformly expanded to the same depth using zero-filling; and the spatial dimension uses bilinear interpolation to align the resolution. This ultimately forms a standardized three-dimensional tensor.

[0073] Based on standardized 3D tensors, shallow joint features are identified through the first layer of multi-scale convolution (7×7×3 convolution kernel);

[0074] Furthermore, the normalized 3D tensor is processed through a first-layer multi-scale convolution, using a 7×7×3 convolution kernel that slides synchronously across the height, width, and depth dimensions. The convolution kernel captures local texture details from the visible light multi-scale feature map, substrate penetration features from the short-wave infrared multi-scale feature map, and shallow correlation patterns from material physical properties such as ink layer thickness distribution, substrate surface roughness parameters, and varnish curing degree thermal maps. The output feature map of the multi-scale convolution retains the original spatial resolution, and the channel dimension integrates primary joint features from multiple data sources to generate shallow joint features including edge response, thickness gradient, and curing degree differences.

[0075] The shallow joint features are expanded in the three dimensions of height, width and depth by a second dilated convolution (3×3×3 convolution kernel) to generate a deep feature tensor;

[0076] Furthermore, the shallow joint features are processed through a second layer of dilated convolution, using a 3×3×3 convolution kernel to expand the height, width, and depth dimensions. Dilated convolution expands the receptive field through interval sampling, simultaneously associating distant ink thickness variation areas, surface roughness anomaly areas, and varnish curing abrupt change areas within the shallow joint features. The convolution kernel learns the collaborative variation patterns of multimodal features across channels in the depth dimension, linking local texture features with global material properties. The output is a deep feature tensor.

[0077] The HDR image set and the deep feature tensor are spliced ​​together by a third-layer cross-modal fusion convolution (1×1×1 pointwise convolution kernel), and an enhanced feature tensor is generated by combining a cross-modal attention mechanism.

[0078] Furthermore, the deep feature tensor and the HDR image set are processed through a third-layer cross-modal fusion convolution, using a 1×1×1 pointwise convolution kernel for channel-dimensional reweighting. Cross-modal skip stitching concatenates the complete dynamic range reflection information of the HDR image set with the abstract semantic features of the deep feature tensor along the channel dimension. The stitched features are dynamically weighted through a cross-modal attention mechanism: spatial attention focuses on the boundary between metallic ink regions and ordinary regions, while channel attention enhances the response intensity of uncured ink features and fiber fracture features. The fused enhanced feature tensor integrates multi-exposure details and material property relationships, preserving cross-modal complementary information in the depth dimension, providing a basis for defect classification.

[0079] Based on the enhanced feature tensor, channel joint normalization is performed through cross-scale feature aggregation to predict the surface anomaly score of printed materials. The expression is as follows:

[0080] ;

[0081] in, It is a score for surface abnormalities of printed materials. It is the vertical dimension of the multi-scale feature map in the visible light band. It is the horizontal dimension of the multi-scale feature map in the shortwave infrared band. It is the channel dimension of the three-dimensional feature cube. It is to enhance the feature tensor at position The The value of each channel.

[0082] It should be noted that, firstly, a position-by-position squaring operation is performed on the spatial feature map of each channel along the channel dimension. The channel joint normalization calculation then places the squared value of each channel in the vertical dimension of the multi-scale feature map in the visible light band. The horizontal dimension of the multi-scale feature map in the shortwave infrared band Spatial averaging is performed to obtain the average energy of each channel. The average energy of each channel is then expressed in the channel dimension of the three-dimensional feature cube. A global average is performed to generate a surface anomaly score for the printed matter. The calculation process strictly maintains the spatial correspondence of the enhanced feature tensor, and the surface anomaly scoring of printed materials is performed. The value reflects the consistency of multimodal characteristic energy distribution at various locations on the surface of the printed material; the higher the value, the greater the probability of surface anomalies.

[0083] Based on the surface anomaly scores of historical qualified printed products, the defect threshold S1 is defined using the 6σ criterion.

[0084] when If the value is ≥S1, then the surface of the printed material is considered to have a defect.

[0085] when When <S1, the printed matter is considered to meet the printing standards;

[0086] It should be noted that when defining the defect threshold S1 based on historical qualified printed matter surface anomaly score data using the 6σ criterion, the historical qualified printed matter surface anomaly score data is used to establish a statistical distribution benchmark. The defect threshold S1 is jointly determined by the distribution characteristics of the historical qualified printed matter surface anomaly score dataset and the 6σ criterion. The value range of the defect threshold S1 is from four to six standard deviations of the mean of the historical qualified printed matter surface anomaly scores.

[0087] Printing standards refer to the acceptable range of surface quality parameters of printed materials, including specific indicators such as registration deviation tolerance (e.g., ≤0.05mm), ink dot integrity (missing diameter <80μm), ink layer thickness tolerance (±2μm), substrate roughness threshold (Ra≤0.8μm), and varnish curing degree compliance value (≥95%). These standards are formulated based on the statistical distribution of measured data such as ink layer thickness distribution and surface roughness parameters of historically qualified printed materials.

[0088] Based on the spatial distribution of the enhanced feature tensor, the sub-pixel level coordinate boundaries of surface defects of printed materials are obtained through nonmaximum suppression and connected component analysis.

[0089] Furthermore, when enhancing the spatial distribution of the feature tensor through non-maximum suppression, the neighborhood response values ​​are first compared pixel by pixel along the height and width dimensions to suppress candidate points of non-local maxima and retain the peak regions of the defect feature response. The binarized mask output by non-maximum suppression marks independent defect regions through connected component analysis, and cubic spline interpolation is used to fit the gradient changes of the defect edges to identify the sub-pixel level coordinate boundaries of defects on the printed surface with sub-pixel accuracy.

[0090] Spatial distribution of enhanced feature tensors The expression is:

[0091] ;

[0092] S4. An improved DBSCAN clustering algorithm is used to spatially aggregate the defect coordinates to generate a complete defect region, and a defect detection report is output through dynamic threshold classification.

[0093] The improvement process of the DBSCAN clustering algorithm is as follows: First, based on the fixed neighborhood density calculation of the traditional DBSCAN, a density gradient weighting method is combined with a Gaussian kernel function to enhance the spatial distribution of the feature tensor. Local weighting is applied to give high-energy defect regions a greater density weight, thereby improving the detection sensitivity of weak defects. Secondly, to address the cluster boundary missegmentation problem that is prone to occur in traditional algorithms, a boundary point reclassification mechanism is added (specifically, when a boundary point connects two clusters and meets the density ratio condition, it is reclassified into a higher-density neighboring cluster). The density ratio between the boundary point and the adjacent cluster is obtained through density ratio comparison analysis, and based on the statistical analysis of the density ratio between the boundary point and the cluster in the historical qualified printed surface defect data, an empirical threshold of α=1.2 is set. Boundary points that meet the empirical threshold are reclassified into the neighboring high-density cluster, effectively avoiding oversegmentation.

[0094] Based on the sub-pixel level coordinate boundaries of surface defects in printed materials, a set of defect coordinate clusters is generated by density gradient weighting and boundary point reclassification.

[0095] Furthermore, a Gaussian kernel function is used to assign local density weights to the coordinate neighborhood, enhancing the spatial locations with higher response values ​​in the feature tensor to receive greater density weights. The weighted coordinate set is then processed through a boundary point reclassification mechanism. When a boundary point simultaneously connects two core clusters and the density ratio of adjacent clusters exceeds an empirical threshold, the boundary point is reclassified into a higher-density cluster. The density ratio is determined by comparing the ratio of the number of defect points within a cluster to the area of ​​its neighborhood. This reclassification process eliminates false boundary segmentation caused by local density fluctuations, and the final generated defect coordinate cluster set represents the complete defect region.

[0096] The geometric parameters and defect energy parameters of each defect coordinate cluster are obtained through a regional attribute analysis algorithm.

[0097] Furthermore, the defect coordinate cluster set is used to obtain geometric parameters and defect energy parameters in two steps through a region attribute analysis algorithm:

[0098] Geometric parameters are obtained by using a sub-pixel level coordinate boundary point set based on defect coordinate clustering. A regional attribute analysis algorithm constructs a grid topology for the defect region through Delaunay triangulation, and the area of ​​the defect region is generated by accumulating and adding grid cell faces. The perimeter is generated by accumulating the Euclidean distances between adjacent points in the coordinate boundary point sequence. Principal component analysis extracts the variance ratio of the coordinate point set to generate the aspect ratio. A rotating caliper algorithm generates the minimum bounding rectangle size. Finally, the geometric parameters are obtained.

[0099] The defect energy parameters are obtained by traversing the sub-pixel coordinates within the defect coordinate cluster using a regional attribute analysis algorithm, extracting the channel response values ​​at the corresponding positions of the enhanced feature tensor, and aggregating the mean and standard deviation of the response values ​​of all coordinate points to generate the defect energy parameters.

[0100] The geometric parameters and defect energy parameters are segmented using a decision tree classifier, and a dynamic process threshold is defined as the decision boundary of the decision tree classifier to generate defect level labels.

[0101] Furthermore, a judgment rule is first established based on a dynamic process threshold, which is set according to printing standards such as ink layer thickness tolerance and registration deviation tolerance. The defect area is used as the primary splitting feature, and the defect energy parameter is used as the secondary judgment criterion. The defect is divided into different subsets through recursive binary division. Finally, the leaf node generates a defect level label. The dynamic process threshold is adjusted in real time according to the ink layer thickness distribution, ambient temperature and humidity, printing pressure, and ink viscosity. For example, the dynamic process threshold is automatically increased when the ink layer thickness increases, and the dynamic process threshold is decreased accordingly when the ambient humidity increases.

[0102] It should be noted that the dynamic process threshold is determined by the statistical distribution of defect parameters of historical qualified printed products and the printing process requirements, and the value range is [0.1, 0.5].

[0103] A defect detection report is generated based on the defect level label, subpixel coordinate bounding box, defect geometric parameters, and defect energy parameters.

[0104] This embodiment also provides a computer device applicable to the digital printing quality inspection method based on image analysis, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the digital printing quality inspection method based on image analysis as proposed in the above embodiment.

[0105] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0106] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the image analysis-based digital printing quality inspection method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0107] In summary, this invention achieves simultaneous optimization of highlight details and color transitions during the image pyramid fusion stage by using an improved Mertens algorithm and dynamically adjusting the fusion weights of metallic ink regions and ordinary regions using a high-reflectivity area mask. Simultaneously, it constructs a three-dimensional feature cube containing surface texture and substrate penetration features by performing cross-modal correlation of visible light and infrared multi-scale feature maps through a spatiotemporal cross-attention mechanism. This enables accurate inversion of material properties such as ink thickness and roughness, improving the accuracy and reliability of digital printing quality inspection.

[0108] 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, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A digital printing quality inspection method based on image analysis, characterized in that: include, Data on reflected light from the printed surface and ambient light are collected. High reflectivity compensation is applied to the metallic ink areas using dynamic adjustment of multi-band light sources to generate an HDR image set. The steps are as follows: The data on reflected light from the surface of the printed material includes reflected light signals in the visible light band and reflected light signals in the short-wave infrared band. The ambient light data includes ambient light intensity and ambient light angle. A mask for highly reflective areas on the surface of printed materials is generated by combining edge gradient analysis with reflectivity differences. The data of reflected light on the surface of printed matter are dynamically adjusted. Images of metallic ink areas and ordinary areas on the surface of printed matter are acquired in short exposure, medium exposure and long exposure sequences, and HDR image groups are generated by fusing them through an improved Mertens algorithm. The surface texture features of the visible light channel and the substrate penetration features of the infrared channel in the HDR image group were obtained through the MobileNetV3 network and the ConvNeXt network, respectively. The surface texture features of the visible light channel and the substrate penetration features of the infrared channel in the HDR image set are obtained. Cross-modal correlation is performed through a spatiotemporal attention mechanism to generate a 3D feature cube. The steps are as follows. Multi-scale feature maps for the visible light band are generated by fusing multi-level feature pyramids. Multi-scale feature maps in the short-wave infrared band are generated by cross-scale attention aggregation. By using a spatiotemporal cross-attention mechanism, cross-modal feature association is performed on multi-scale feature maps of the visible light band and the short-wave infrared band to generate a three-dimensional feature cube. The reverse Monte Carlo ray tracing algorithm is used to obtain the material physical properties of printed materials. An enhanced feature tensor is generated by stitching together 3D convolutional networks, and defect detection and localization are performed through cross-scale feature aggregation. The steps are as follows: By inverting the infrared characteristic response values ​​at each spatial location in the multi-scale characteristic map of the short-wave infrared band using the ink absorbance-thickness relationship curve, the ink layer thickness distribution is generated. By comparing the varnish curing degree calibration curve with the dual-band characteristic ratio analysis method, and using nonlinear least squares fitting, a varnish curing degree thermogram was generated. By using the micro-surface BRDF fitting algorithm, surface normal distribution modeling and scattering energy analysis are performed on the multi-scale feature map of visible light band in a three-dimensional feature cube to obtain the surface roughness parameters of the substrate. The process of generating enhanced feature tensors by stitching together three-dimensional convolutional networks involves the first layer of multi-scale convolution recognizing shallow joint features, the second layer of dilated convolution expanding the shallow joint features to generate deep feature tensors, and the third layer of cross-modal fusion convolution performing cross-modal skip stitching of HDR image groups and deep feature tensors, combined with a cross-modal attention mechanism to generate enhanced feature tensors. Based on the enhanced feature tensor, channel joint normalization calculation is performed through cross-scale feature aggregation to predict the surface anomaly score of printed materials and compare it with the defect threshold to identify defects on the surface of printed materials. The expression for predicting surface anomaly scores on printed materials is: ； in, It is a score for surface abnormalities of printed materials. It is the vertical dimension of the multi-scale feature map in the visible light band. It is the horizontal dimension of the multi-scale feature map in the shortwave infrared band. It is the channel dimension of the three-dimensional feature cube. It is to enhance the feature tensor at position The The value of each channel; Subpixel-level coordinate boundaries of surface defects in printed materials are obtained through nonmaximum suppression and connected component analysis. An improved DBSCAN clustering algorithm is used to spatially aggregate defect coordinates to generate a complete defect region, and a defect detection report is output through dynamic threshold classification. The steps are as follows. A regional attribute analysis algorithm is used to obtain the geometric parameters and defect energy parameters of each defect coordinate cluster. At the same time, a decision tree classifier is used to segment the data, and a dynamic process threshold is defined as the decision boundary of the decision tree classifier to generate defect level labels. A defect detection report is generated based on the defect level label, subpixel coordinate bounding box, defect geometric parameters, and defect energy parameters.

2. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the digital printing quality inspection method based on image analysis as described in claim 1.

3. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the digital printing quality inspection method based on image analysis as described in claim 1.

Images