Detection method for printing effect verification

Abstract

The invention discloses a detection method for printing effect verification, and the method comprises the steps: collecting scanning image information and environment sensing information, employing an improved OTSU algorithm and an illumination compensation module to process the scanning image information, and extracting a standardized image feature matrix containing the characteristics of character integrity, edge sharpness and the like; identifying a defect mode by using an EAST text detection algorithm and morphological operation, and generating a quantitative evaluation vector; the vector and environment sensing information are subjected to space-time alignment, a printing quality degradation model is constructed through a density peak value clustering algorithm, and a multi-dimensional quality index set is established; and finally, based on the graph attention network and the time sequence convolutional network, analyzing the incidence relation between the printing quality and the equipment state, and outputting a verification report containing a quality score, a defect positioning graph and a life prediction curve. According to the method, accurate quantitative evaluation of the printing quality is realized, a dynamic association model of the quality, the environmental parameters and the equipment aging is established, and an intelligent decision basis is provided for printing quality maintenance.

Description

Technical Field

[0001] This invention relates to the field of thermal printing technology, and in particular to a detection method for verifying printing quality. Background Technology

[0002] Thermal printing technology is a common printing method for documents, labels, and other fields. The stability of its print quality directly affects information readability and equipment reliability. Currently, the industry mainly relies on two types of technologies to inspect print quality: one is based on sensors to directly monitor hardware parameters such as printhead temperature, judging the equipment status through limited key parameters, but it is difficult to quantify the trend of print quality degradation; the other uses the principle of optical reflection to detect color development, but it is easily affected by ambient light interference and cannot identify progressive defects. These methods generally suffer from the problem of single detection dimensions, only able to judge the presence or absence of color development or the basic status of the equipment, unable to quantitatively evaluate key quality indicators such as character integrity and edge sharpness, and even more difficult to establish a correlation model between print quality and environmental parameters and equipment aging.

[0003] The core limitation of existing technologies lies in the lack of multi-dimensional dynamic assessment capabilities for print quality. Traditional detection methods cannot accurately distinguish the coupled effects of printhead performance degradation and environmental interference, nor can they predict progressive failure risks through image feature analysis. This makes it difficult for maintenance personnel to detect potential quality defects in a timely manner, and even more difficult to develop adaptive maintenance strategies for different environmental conditions. How to achieve intelligent verification of quantitative assessment of print quality, precise defect location, and lifespan prediction has become a key challenge in improving the reliability of thermal printing. Summary of the Invention

[0004] In view of this, the purpose of this invention is to propose a detection method for verifying printing effect, which realizes quantitative evaluation of printing quality and life prediction through multimodal feature fusion and dynamic degradation modeling, and solves the problem that existing detection methods cannot quantitatively evaluate the degradation trend of printing quality and the environmental coupling effect.

[0005] To achieve the above-mentioned technical objectives, the technical solution adopted by this invention is: a detection method for verifying printing effects, comprising: Collect scanned image information and environmental sensing information of thermal printing media. The scanned image information includes the grayscale distribution characteristics of the printed character area and the background area, and the environmental sensing information includes real-time temperature and humidity parameters and printer module working status parameters. An improved OTSU algorithm combined with an illumination compensation module is used to generate an adaptive segmentation threshold for the scanned image information. Geometric correction is used to eliminate medium deformation interference and extract a standardized image feature matrix containing character integrity features, edge sharpness features, and background noise distribution features. Based on the EAST text detection algorithm, the standardized image feature matrix is ​​used to locate tilted text regions. Combined with morphological operations, defect patterns of broken characters, blurred characters, and missing characters are identified, and a quantitative evaluation vector containing character loss rate, structural similarity index, and gradient change rate is generated. By aligning the quantitative evaluation vector with environmental sensing information in time and space, a print quality degradation model is constructed. The density peak clustering algorithm is used to identify the coupling relationship between the aging characteristics of the print core and the interference characteristics of temperature and humidity, and a multi-dimensional quality index set including color concentration gradient, lifetime decay coefficient and environmental compensation factor is generated. A graph attention network is established based on a multidimensional quality index set to model the spatial correlation between printhead temperature fluctuations, media coating characteristics, and character color rendering quality. A temporal convolutional network is used to capture progressive failure modes, and an verification report is output that includes an instant quality score, a defect location map, and a life prediction curve. The verification report includes a dynamic threshold early warning module and an optimal maintenance strategy recommendation module.

[0006] In some embodiments, an improved OTSU algorithm combined with an illumination compensation module is used to generate an adaptive segmentation threshold for the scanned image information. Geometric correction is used to eliminate medium deformation interference, and a standardized image feature matrix containing character integrity features, edge sharpness features, and background noise distribution features is extracted, including: The reflectance characteristic spectrum of the thermosensitive medium is obtained by multispectral imaging, and the reflectance characteristic spectrum is input into the local adaptive weighting module to generate dynamic weighting coefficients for different medium regions. The inter-class variance calculation of the OTSU algorithm is optimized by using dynamic weight coefficients, and the initial segmentation threshold that adapts to the inhomogeneity of the medium is output. The initial segmentation threshold is input into the illumination compensation module, and the dark areas of the scanned image information are gamma corrected by the nonlinear brightness mapping function, while the overexposure of the highlight areas is suppressed, generating a pre-processed image with balanced illumination. The preprocessed image is subjected to medium edge detection based on Hough transform. A perspective transformation matrix is ​​constructed based on the detected edge features. The medium deformation is corrected by the perspective transformation matrix to obtain a geometrically regular image to be analyzed. Within the ROI region of the image to be analyzed, the improved Sobel-Feldman operator is applied to extract the gradient magnitude of character edges, while the noise power spectral density of the background region is calculated based on Fourier transform, forming an intermediate feature set containing spatial and frequency domain features. The intermediate feature set is input into the pre-trained character skeleton topology network to extract the skeleton connectivity features of the characters. Combined with the breakpoint distribution map generated by morphological erosion operation, the character defect rate index is calculated. At the same time, a multi-scale structural similarity algorithm is used to evaluate the character deformation distortion. By fusing edge gradient magnitude, noise power spectral density, skeleton connectivity features, and deformation distortion index, a feature descriptor with uniform dimensions is generated through a feature standardization layer. After dimensionality reduction processing by principal component analysis, a traceable standardized image feature matrix is ​​output.

[0007] In some embodiments, an initial segmentation threshold is input into the illumination compensation module, and gamma correction is performed on the dark areas of the scanned image information through a nonlinear brightness mapping function, while suppressing overexposure in the highlight areas, to generate a pre-processed image with balanced illumination, including: A brightness perception conversion module based on a retinal model is constructed to convert the RGB color space of the scanned image information to the CIELAB uniform color space. The brightness component is adjusted for perceptual consistency through the human eye's visual sensitivity function, so that the converted scanned image information includes dark areas, medium brightness areas and highlight areas. A piecewise gamma correction function is constructed. In the dark area, dynamic gamma coefficients are used to enhance contrast. In the medium brightness area, linear mapping is maintained. In the highlight area, an inverse S-curve is applied to compress the dynamic range, and the corrected scan image information is obtained. The compensation intensity of the dark, medium and bright areas of the corrected scanned image information is dynamically adjusted based on the reflectance value of the thermosensitive medium in the 550nm band. The double exposure fusion algorithm is enabled in areas with reflectance below 40%. Using the gradient magnitude map of the original scanned image information as the guide image, edge-aware smoothing processing is performed on the scanned image information after illumination compensation, so as to eliminate uneven illumination while preserving the high-frequency details of character strokes. The color constancy correction is performed on the scanned image information after elimination. By calculating the color coordinate offset of the white points in the image, the hue component H is compensated and rotated in the HSV color space to eliminate the color deviation caused by the difference in light color temperature, and the preprocessed image is obtained.

[0008] In some embodiments, the standardized image feature matrix is ​​used to locate tilted text regions based on the EAST text detection algorithm, and morphological operations are combined to identify defect patterns of broken characters, blurred characters, and missing characters, generating a quantitative evaluation vector containing character loss rate, structural similarity index, and gradient change rate, including: The EAST text detection network, which is improved based on depthwise separable convolution, performs multi-scale feature extraction on the standardized image feature matrix and outputs a text location feature map containing the geometric center coordinates and tilt angle of the text region. The text location feature map is input into the attention-based affine transformation module, and the optimal text correction parameters are learned through the spatial transformation network to generate geometrically regular character analysis regions, while preserving the pixel mapping relationship between the original image and the corrected image. A multi-level morphological processing pipeline is constructed on the character analysis region. The parallel morphological processing pipeline is executed using a sequence of dynamic-sized structuring elements to extract character breakage features, halo diffusion features, and missing region features. A quantitative evaluation sub-model is constructed for character breakage features, smearing and diffusion features, and missing region features. The breakage feature sub-model calculates the cumulative breakage amount through the density distribution of skeleton endpoints, the smearing and diffusion feature sub-model uses the entropy value of the directional gradient histogram to measure the degree of edge blurring, and the missing region feature sub-model evaluates the proportion of missing area based on the adaptive region growth algorithm. The output features of each sub-model are input into the spatiotemporal feature fusion module, and the temporal evolution law of the defect features is learned through the long short-term memory network to generate a quantitative evaluation vector with spatiotemporal consistency.

[0009] In some embodiments, the EAST text detection network, based on depthwise separable convolution, performs multi-scale feature extraction on the normalized image feature matrix, outputting a text location feature map containing the geometric center coordinates and tilt angle of the text region, including: Hierarchical feature extraction is performed on the input normalized image feature matrix using a deep separable convolutional network. The high-resolution feature map output by the shallow network is used to capture character edge details, while the low-resolution feature map output by the deep network is used to identify semantic information of text regions, resulting in multi-scale feature maps. The multi-scale feature map is input into the feature pyramid fusion module. Semantic information is propagated through the top-down upsampling path, and geometric details are enhanced by the bottom-up lateral connection path to obtain the fused multi-scale feature map. The fused multi-scale feature map is subjected to bi-branch prediction, which includes a first branch and a second branch. The first branch generates a text region probability map through convolutional kernel sliding detection, which is denoted as the first output result. The second branch regresses the center coordinates, height, width and rotation angle parameters of each effective text region, which is denoted as the second output result. A rotated text detection box is constructed based on the first output result and the second output result. The first output result is used to filter candidate regions, and the second output result is used to determine the spatial pose of the text box. The rotated text detection box is subjected to non-maximum suppression based on rotation IoU, and the output text position feature map with rotation angle is generated.

[0010] In some embodiments, the multi-level morphological processing pipeline includes a primary processing pipeline, a secondary processing pipeline, and a high-level processing pipeline. A multi-level morphological processing pipeline is constructed on the character analysis region. The parallel morphological processing pipeline is executed using a sequence of dynamically sized structuring elements to extract character breakage features, halo diffusion features, and missing region features, including: The primary processing pipeline uses circular structural elements for basic morphological operations to eliminate noise interference; the intermediate processing pipeline uses elliptical structural elements for directional morphological operations to enhance stroke continuity; and the advanced processing pipeline uses rectangular structural elements for overall morphological analysis to extract character structural features, resulting in multi-level processing results. The size of the structural elements at each level is dynamically adjusted according to the resolution of the character analysis region to generate a sequence of structural elements of different sizes. Morphological operations of multiple structural element sizes are performed in parallel on the same processing level. The results of the parallel operations are then fused to obtain the final morphological processing result. Based on the final morphological processing results, character breakage features are extracted using skeleton extraction and endpoint detection algorithms, halo diffusion features are extracted using edge gradient analysis and region diffusion detection algorithms, and missing region features are extracted using connected component analysis and region comparison algorithms.

[0011] In some embodiments, the quantitative evaluation vector is spatiotemporally aligned with environmental sensing information to construct a print quality degradation model. A density peak clustering algorithm is used to identify the coupling relationship between print core aging characteristics and temperature and humidity interference characteristics, generating a multidimensional quality index set including color density gradient, lifetime decay coefficient, and environmental compensation factor, including: Establish a temporal alignment relationship between the quantized evaluation vector and the environmental sensing information flow, and use a bidirectional LSTM network to extract the temporal correlation of cross-modal features to construct a spatiotemporal feature tensor that includes the device's operating status and environmental conditions. An improved density peak clustering algorithm is applied to the spatiotemporal feature tensor. The Mahalanobis distance metric is introduced and combined with the Gaussian mixture model to fit the data distribution under different working conditions, and the inherent aging feature mode and environmental disturbance feature mode of the printing core are separated. Degradation assessment sub-models are established for the inherent aging characteristic mode and the environmental disturbance characteristic mode, respectively. The degradation assessment sub-models include the core aging sub-model, the environmental disturbance sub-model, and the dynamic compensation sub-model. The mechanism aging sub-model constructs a nonlinear decay function based on the amount of developer consumed to obtain the first output parameter. The environmental interference sub-model establishes a mapping relationship between temperature and humidity and print quality using support vector regression to obtain the second output parameter. The dynamic compensation sub-model achieves real-time error correction based on Kalman filtering to obtain the third output parameter. The first, second, and third output parameters are input into the quality index fusion module. The importance weights of different feature dimensions are dynamically adjusted through an attention mechanism to generate a multi-dimensional quality index set that includes color concentration gradient, lifetime decay coefficient, and environmental compensation factor. The color concentration gradient is generated by comparing multispectral imaging data with a reference color chart, the lifetime decay coefficient is calculated by the change rate of the printhead drive current, and the environmental compensation factor is a combination of the deviation between the real-time readings of the temperature and humidity sensor and the historical best operating conditions.

[0012] In some embodiments, an improved density peak clustering algorithm is applied to the spatiotemporal feature tensor, and a Mahalanobis distance metric is introduced combined with a Gaussian mixture model to fit the data distribution under different working conditions, separating the inherent aging characteristic mode and the environmental disturbance characteristic mode of the printing core, including: Multi-dimensional feature decoupling is performed on the spatiotemporal feature tensor, and the correlation between feature dimensions is established through covariance matrix analysis to form a feature space; A density peak calculation model based on Mahalanobis distance is constructed to calculate the local density and minimum distance of each sample point in the feature space, forming a density-distance distribution map; By setting dynamic threshold boundaries on the density-distance distribution map, high-density core points and low-density edge points are automatically identified, and the samples are divided into multiple candidate clusters. For each candidate cluster, a Gaussian distribution is fitted, the mean vector and covariance matrix of each cluster are calculated, and a probability density distribution model is established. By analyzing the overlap of probability density distributions, adjacent clusters with similarity exceeding a preset threshold are merged, and heterogeneous clusters with dissimilarity exceeding a preset threshold are separated. Modal labeling is performed on the final clustering results. Clusters with stable temporal characteristics are labeled as the inherent aging characteristic modes of the printing core, and clusters with environmentally related fluctuation characteristics are labeled as environmental disturbance characteristic modes.

[0013] In some embodiments, a graph attention network is established based on a multidimensional quality index set to model the spatial correlation between printhead temperature fluctuations, media coating characteristics, and character color rendering quality. A temporal convolutional network is used to capture progressive failure modes, outputting a verification report containing an instantaneous quality score, defect location map, and lifespan prediction curve. The verification report includes a dynamic threshold early warning module and an optimal maintenance strategy recommendation module, including: A spatial association model based on graph attention network is constructed. The printhead temperature distribution characteristics, media coating spectral characteristics, and character color development quality characteristics are used as nodes of a heterogeneous graph. The association strength matrix between nodes is calculated through spatial attention mechanism. The temperature-coating association edge is constrained by the heat conduction equation, and the coating-color development association edge is mapped through spectral reflectance. The spatial association graph is output. The spatial correlation graph is input into the spatiotemporal feature fusion module. While preserving the spatial correlation characteristics of the nodes, the dynamic evolution characteristics of each node are extracted through a temporal convolutional network. Gated temporal convolutional units are used to extract features from the printhead temperature fluctuation sequence. Causal convolution is used to capture the gradual change process of the medium coating characteristics, generating a composite feature map that simultaneously contains spatial correlation and dynamic evolution characteristics. A dual-channel evaluation network is constructed based on composite feature maps. The dual-channel evaluation network includes a quality evaluation channel and a lifetime prediction channel, including: The quality assessment channel aggregates global features through graph pooling operations and calculates real-time quality scores using a weighted self-attention mechanism. The life prediction channel inputs node features into a neural differential equation network to simulate the thermal fatigue accumulation process of the printhead and outputs a life prediction curve with confidence intervals. Key abnormal nodes are located using the graph attention weight backpropagation algorithm. Combined with the spatial coordinate mapping relationship of the nodes, a defect location map with thermal value annotations is generated in the original printed area. The color abnormality area is quantified by spectral difference, and the structural defect area is evaluated by topological feature similarity. An inspection report is generated based on real-time quality scores, life prediction curves, and defect location maps.

[0014] In some embodiments, the quality assessment channel aggregates global features through graph pooling operations and calculates an instantaneous quality score using a weighted self-attention mechanism, including: The peak heat load feature is extracted by maximum pooling at the printhead temperature node, the coating uniformity index is generated by average pooling at the media coating node, and the comprehensive color development score is obtained by weighted average pooling at the color development quality node. The weight of each color development quality node is calculated by the self-attention mechanism. The peak heat load feature, coating uniformity index and comprehensive color development score are integrated into the graph pooling output feature. The correlation strength between the printhead temperature node and the color development quality node is calculated through the self-attention weight calculation module. The self-attention weight calculation module contains a learnable bias parameter matrix, which modulates the initial weights through the spectral characteristics of the media coating node and outputs the final self-attention weight distribution. The actual color development quality data of historical printed samples is collected as a true reference. The graph pooling output features are compared with the true data. The bias parameters in the self-attention weight calculation module are dynamically updated through backpropagation of error gradient. The graph pooling output features and the dynamically evolved features are concatenated using tensors, and the concatenated feature vector is mapped to an instantaneous quality score in the range of 0-100 through a fully connected layer.

[0015] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art: It collects scanned image information and environmental sensing information, processes the scanned image information using an improved OTSU algorithm and an illumination compensation module, and extracts a standardized image feature matrix containing features such as character integrity and edge sharpness; it uses the EAST text detection algorithm and morphological operations to identify defect patterns and generate a quantitative evaluation vector; it aligns this vector spatiotemporally with the environmental sensing information, constructs a print quality degradation model using a density peak clustering algorithm, and establishes a multi-dimensional quality index set; finally, it analyzes the correlation between print quality and equipment status based on graph attention networks and temporal convolutional networks, and outputs a verification report containing quality scores, defect location maps, and lifespan prediction curves. The present invention achieves accurate quantitative evaluation of print quality, establishes a dynamic correlation model between quality and environmental parameters and equipment aging, and provides intelligent decision-making basis for print quality maintenance. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart illustrating steps S101 to S105 of the detection method described in the specific implementation embodiment; Figure 2 This is a flowchart illustrating steps S201 to S207 of the detection method described in the specific implementation. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This embodiment provides a detection method for verifying printing quality, including: S101. Collect scanned image information and environmental sensing information of thermal printing media. The scanned image information includes the grayscale distribution characteristics of the printed character area and the background area. The environmental sensing information includes real-time temperature and humidity parameters and printer module working status parameters. S102. An improved OTSU algorithm combined with an illumination compensation module is used to generate an adaptive segmentation threshold for the scanned image information. Geometric correction is used to eliminate medium deformation interference and extract a standardized image feature matrix containing character integrity features, edge sharpness features, and background noise distribution features. S103. Based on the EAST text detection algorithm, locate the tilted text region of the standardized image feature matrix, and combine morphological operations to identify the defect patterns of broken characters, blurred characters and missing characters, and generate a quantitative evaluation vector containing character loss rate, structural similarity index and gradient change rate. S104. The quantitative evaluation vector is spatiotemporally aligned with the environmental sensing information to construct a print quality degradation model. The density peak clustering algorithm is used to identify the coupling relationship between the aging characteristics of the print core and the temperature and humidity interference characteristics, and to generate a multi-dimensional quality index set including the color development concentration gradient, the life decay coefficient and the environmental compensation factor. S105. Based on a multi-dimensional quality index set, establish a graph attention network to model the spatial correlation between printhead temperature fluctuations, media coating characteristics, and character color rendering quality. Capture progressive failure modes through a temporal convolutional network and output a verification report that includes an instantaneous quality score, defect location map, and life prediction curve. The verification report includes a dynamic threshold early warning module and an optimal maintenance strategy suggestion module.

[0020] In step S101, the scanned image information refers to the surface image data of the thermal printing medium acquired by a high-resolution scanner, including the grayscale value distribution characteristics of the printed character area and the background area. The grayscale value distribution characteristics are used to distinguish the valid printed content from the background color of the medium. The environmental sensing information includes real-time temperature and humidity parameters and printer module operating status parameters. The temperature and humidity parameters can be collected by digital sensors, and the printer module operating status parameters include operating indicators such as printhead temperature and operating voltage. The collection of these basic data provides the raw input for subsequent analysis.

[0021] In step S102, the improved OTSU algorithm refers to an adaptive threshold segmentation method that integrates the gray-level statistical characteristics of local regions, dynamically adjusting the segmentation threshold by analyzing the local contrast of the image. Preferably, the illumination compensation module employs a piecewise gamma correction function, implementing differentiated processing for different brightness regions to eliminate the impact of uneven illumination on image quality. Geometric correction is achieved through medium edge detection and perspective transformation, used to correct image distortion caused by medium deformation. These processes collectively generate a standardized image feature matrix, providing standardized input data for subsequent defect identification.

[0022] In step S103, the EAST text detection algorithm refers to a deep learning-based text region localization method that can accurately identify tilted and curved text regions. Morphological operations use multi-scale structuring element analysis to analyze character morphological features, and skeleton extraction and connected component analysis are used to identify various printing defects. This step transforms image features into quantifiable evaluation indicators, enabling an objective evaluation of print quality.

[0023] In step S104, spatiotemporal alignment refers to establishing a time-series correlation between the quantized evaluation vector and environmental sensing information. Preferably, a bidirectional LSTM network is used to capture the temporal dependencies between features. Density peak clustering algorithm is used to separate equipment aging features from environmental interference features, and different influencing factors are identified by analyzing feature distribution. The print quality degradation model constructed in this step can distinguish the effects of inherent aging and environmental interference.

[0024] In step S105, the graph attention network establishes a spatial correlation model between printhead temperature, media properties, and color development quality, quantifying the mutual influence of each factor through an attention mechanism; the temporal convolutional network is used to analyze the gradual changes in printing performance and predict equipment lifespan trends. The final verification report not only includes the current quality status but also provides early warnings and maintenance recommendations.

[0025] In this embodiment, regarding threshold determination, an improved OTSU algorithm is used combined with local region characteristics to dynamically adjust the segmentation threshold, thereby improving segmentation accuracy in complex backgrounds. For defect identification, multi-scale morphological operations can adapt to the detection of character defects of different sizes. In quality assessment, the combination of spatiotemporal alignment and density peak clustering achieves effective separation of influencing factors. This embodiment constructs a complete print quality assessment system, solving the assessment blind spot problem in existing technologies and realizing comprehensive monitoring and prediction of print status.

[0026] Please see Figure 2 In some embodiments, an improved OTSU algorithm combined with an illumination compensation module is used to generate an adaptive segmentation threshold for the scanned image information. Geometric correction is used to eliminate medium deformation interference, and a standardized image feature matrix containing character integrity features, edge sharpness features, and background noise distribution features is extracted, including: S201. Obtain the reflectance characteristic spectrum of the thermosensitive medium based on multispectral imaging, input the reflectance characteristic spectrum into the local adaptive weighting module, and generate dynamic weighting coefficients for different medium regions. S202. Optimize the inter-class variance calculation of the OTSU algorithm using dynamic weight coefficients, and output an initial segmentation threshold that adapts to the inhomogeneity of the medium. S203. Input the initial segmentation threshold into the illumination compensation module, perform gamma correction on the dark areas of the scanned image information through a nonlinear brightness mapping function, and suppress overexposure in the highlight areas to generate a pre-processed image with balanced illumination. S204. Perform medium edge detection based on Hough transform on the preprocessed image, construct a perspective transformation matrix based on the detected edge features, correct medium deformation through the perspective transformation matrix, and obtain a geometrically regular image to be analyzed. S205. Within the ROI region of the image to be analyzed, the improved Sobel-Feldman operator is applied to extract the gradient magnitude of the character edges. At the same time, the noise power spectral density of the background region is calculated based on the Fourier transform, forming an intermediate feature set containing spatial and frequency domain features. S206. Input the intermediate feature set into the pre-trained character skeleton topology network, extract the skeleton connectivity features of the characters, combine the breakpoint distribution map generated by morphological erosion operation, calculate the character defect rate index, and at the same time use a multi-scale structural similarity algorithm to evaluate the character deformation distortion. S207 integrates edge gradient magnitude, noise power spectral density, skeleton connectivity features, and deformation distortion index. It generates a feature descriptor with uniform dimensions through a feature standardization layer. After dimensionality reduction processing by principal component analysis, it outputs a traceable standardized image feature matrix.

[0027] In step S201, multispectral imaging refers to acquiring the reflection characteristics of the thermosensitive medium at different wavelengths using a multi-band optical sensor. The reflectance characteristic spectrum characterizes the response characteristics of the medium surface to different spectra. The local adaptive weighting module generates dynamic weighting coefficients by analyzing the reflectance differences in each region, which are used for subsequent partitioning processing. This step solves the problem that traditional single-spectral imaging has difficulty distinguishing the background color of the medium from the printed content.

[0028] In step S202, the dynamic weighting coefficients optimize the inter-class variance calculation process of the OTSU algorithm, enabling threshold segmentation to adapt to the differences in optical properties of different regions of the medium. The initial segmentation threshold is determined by maximizing the weighted inter-class variance, and its calculation process takes into account the spatial distribution characteristics of the medium's reflectivity. This step achieves high-precision segmentation of non-uniform media.

[0029] In step S203, the illumination compensation module processes the scanned image information, using the image region brightness distribution information contained in the initial segmentation threshold to guide the parameter selection of the piecewise gamma correction function, thereby performing illumination equalization processing on the original image. The illumination compensation module employs a piecewise nonlinear brightness mapping function, where dark area gamma correction enhances details through exponential transformation, and highlight areas use an S-curve to suppress overexposure; the pre-processed image eliminates color cast through adaptive color temperature adjustment. This processing preserves character edge details while equalizing overall brightness.

[0030] In step S204, the Hough transform is used to detect straight line features at the edges of the medium, and the perspective transformation matrix is ​​constructed by solving for the intersection points of the edges; the geometrically regular image to be analyzed eliminates deformation distortion caused by medium wrinkles or tilts. This step provides a standardized image spatial benchmark for subsequent feature extraction.

[0031] In step S205, the improved Sobel-Feldman operator introduces a directional weighting mechanism based on the traditional Sobel operator. It enhances edge responses at different angles by constructing an 8-directional gradient template, while employing an adaptive threshold to suppress pseudo-edge interference, effectively extracting the fine contour features of the characters. Noise power spectral density analysis involves performing a two-dimensional Fourier transform on the background region and then calculating the energy distribution of each frequency band through integration over a ring region, characterizing the inherent texture of the thermal medium and the periodic noise introduced by the printhead. This joint spatial-frequency domain analysis method can capture both the local geometric features of the characters and identify the global statistical characteristics of the background region, providing multi-dimensional feature data for subsequent quality assessment.

[0032] In step S206, the character skeleton topology network extracts stroke connectivity features using a pre-trained deep learning model; a multi-scale structural similarity algorithm evaluates the degree of character deformation at different resolutions. This step enables a quantitative description of character defects.

[0033] In step S207, the feature normalization layer unifies the dimensions of each feature through z-score normalization; principal component analysis reduces dimensionality and preserves the main feature components. The resulting normalized image feature matrix is ​​traceable and comparable.

[0034] This embodiment solves the recognition difficulties caused by uneven background color in thermally sensitive media by combining multispectral imaging with local adaptive processing; it improves the robustness of feature extraction by employing joint spatial and frequency domain analysis; and it introduces deep learning-assisted skeleton analysis to achieve accurate quantification of character defects. This embodiment constitutes a complete image feature extraction process, providing a reliable data foundation for subsequent quality assessment.

[0035] In some embodiments, an initial segmentation threshold is input into the illumination compensation module, and gamma correction is performed on the dark areas of the scanned image information through a nonlinear brightness mapping function, while suppressing overexposure in the highlight areas, to generate a pre-processed image with balanced illumination, including: A brightness perception conversion module based on a retinal model is constructed to convert the RGB color space of the scanned image information to the CIELAB uniform color space. The brightness component is adjusted for perceptual consistency through the human eye's visual sensitivity function, so that the converted scanned image information includes dark areas, medium brightness areas and highlight areas. A piecewise gamma correction function is constructed. In the dark area, dynamic gamma coefficients are used to enhance contrast. In the medium brightness area, linear mapping is maintained. In the highlight area, an inverse S-curve is applied to compress the dynamic range, and the corrected scan image information is obtained. The compensation intensity of the dark, medium and bright areas of the corrected scanned image information is dynamically adjusted based on the reflectance value of the thermosensitive medium in the 550nm band. The double exposure fusion algorithm is enabled in areas with reflectance below 40%. Using the gradient magnitude map of the original scanned image information as the guide image, edge-aware smoothing processing is performed on the scanned image information after illumination compensation, so as to eliminate uneven illumination while preserving the high-frequency details of character strokes. The color constancy correction is performed on the scanned image information after elimination. By calculating the color coordinate offset of the white points in the image, the hue component H is compensated and rotated in the HSV color space to eliminate the color deviation caused by the difference in light color temperature, and the preprocessed image is obtained.

[0036] In this embodiment, the luminance perception conversion module based on the retinal model is a color space conversion device that simulates the visual characteristics of the human eye. It converts the RGB tristimulus values ​​into luminance and chromaticity components of the CIELAB color space using the CIE standard observer function. The luminance component undergoes weighting processing using a visual sensitivity function to ensure the conversion result conforms to the Weber-Fechner law. This conversion process divides the scanned image information into dark areas (<30 cd / m²) that conform to the perceptual characteristics of the human eye. 2 ), medium brightness area (30-80 cd / m 2 ) and highlight areas (>80cd / m 2 This lays the foundation for subsequent partitioning processes.

[0037] The piecewise gamma correction function is a nonlinear transformation function that employs different mapping strategies based on the brightness range. The dynamic gamma coefficient in the dark area is determined adaptively through local contrast, ranging from 1.8 to 2.4; the mid-brightness area maintains a linear mapping with γ=1; and the inverse S-curve in the highlight area uses cubic polynomial fitting to achieve dynamic range compression. This zonal processing method enhances details in the dark areas while avoiding information loss in the highlight areas.

[0038] The double exposure fusion algorithm is designed for areas with reflectivity below 40%. It uses image synthesis techniques simulating different exposure times to fuse the highlight details preserved in a short exposure with the shadow information extracted in a long exposure in the wavelet domain. This algorithm is specifically enabled in low-reflectivity areas of thermally sensitive media, effectively solving the problem that traditional single exposures struggle to capture both bright and dark details.

[0039] Edge-aware smoothing is an edge-preserving filtering process guided by the gradient magnitude map of the original image. By constructing an anisotropic diffusion equation, Gaussian smoothing is applied to areas with small gradients to eliminate uneven illumination, while maintaining character edge sharpness in areas with large gradients. When performing chromaticity constancy correction in the HSV color space, this process calculates the difference in chromaticity coordinates between a reference white point under a D65 standard light source and the actual white point in the image to determine the rotation compensation angle of the hue component, eliminating color rendering deviations caused by differences in ambient light color temperature.

[0040] This embodiment uses a retinal model to achieve brightness zoning that conforms to the perceptual characteristics of the human eye, making the processing results more consistent with visual evaluation standards. It employs reflectivity-adaptive double exposure fusion technology to specifically address the loss of detail in low-reflectivity areas of thermally sensitive media. Combined with edge preservation and chromaticity correction, it ensures accurate color reproduction of characters while eliminating illumination interference. This embodiment constitutes a complete illumination compensation solution, providing high-quality pre-processed images for subsequent image analysis.

[0041] In some embodiments, the standardized image feature matrix is ​​used to locate tilted text regions based on the EAST text detection algorithm, and morphological operations are combined to identify defect patterns of broken characters, blurred characters, and missing characters, generating a quantitative evaluation vector containing character loss rate, structural similarity index, and gradient change rate, including: The EAST text detection network, which is improved based on depthwise separable convolution, performs multi-scale feature extraction on the standardized image feature matrix and outputs a text location feature map containing the geometric center coordinates and tilt angle of the text region. The text location feature map is input into the attention-based affine transformation module, and the optimal text correction parameters are learned through the spatial transformation network to generate geometrically regular character analysis regions, while preserving the pixel mapping relationship between the original image and the corrected image. A multi-level morphological processing pipeline is constructed on the character analysis region. The parallel morphological processing pipeline is executed using a sequence of dynamic-sized structuring elements to extract character breakage features, halo diffusion features, and missing region features. A quantitative evaluation sub-model is constructed for character breakage features, smearing and diffusion features, and missing region features. The breakage feature sub-model calculates the cumulative breakage amount through the density distribution of skeleton endpoints, the smearing and diffusion feature sub-model uses the entropy value of the directional gradient histogram to measure the degree of edge blurring, and the missing region feature sub-model evaluates the proportion of missing area based on the adaptive region growth algorithm. The output features of each sub-model are input into the spatiotemporal feature fusion module, and the temporal evolution law of the defect features is learned through the long short-term memory network to generate a quantitative evaluation vector with spatiotemporal consistency.

[0042] In this embodiment, the EAST text detection network based on depthwise separable convolution refers to a lightweight neural network structure that replaces the standard convolution in the traditional EAST algorithm with depthwise separable convolution. By separating the computation of channel-wise convolution and pointwise convolution, the computational cost is reduced while maintaining text detection accuracy. This network achieves multi-scale feature extraction through a feature pyramid structure, and the output text location feature map contains the geometric center coordinates and tilt angle of each character region, where the tilt angle is calculated through the parameters of the rotated bounding box predicted by regression.

[0043] The affine transformation module based on the attention mechanism is a differential image transformation component that integrates a spatial attention mechanism. It establishes a pixel-level mapping relationship between the original image and the corrected image through a learnable sampling grid generator. This module dynamically adjusts the spatial transformation intensity in different regions using attention weights, maintaining high-fidelity transformation in the center of the text region and appropriately reducing computational precision in the background region, thereby optimizing the allocation of computational resources.

[0044] A multi-level morphological processing pipeline refers to a parallel processing architecture composed of basic morphological operations such as erosion, dilation, opening, and closing. The dynamically sized structuring element sequence adaptively adjusts the kernel size based on the character stroke width, with a small kernel (3×3) used for fine structure detection and a large kernel (7×7) used for macroscopic defect identification. This pipeline utilizes multi-threading technology to achieve parallel execution of various morphological operations, significantly improving processing efficiency.

[0045] In the quantitative evaluation sub-model, the skeleton endpoint density distribution refers to the statistical characteristic of the number of skeleton endpoints per unit area, which is used to distinguish between natural endpoints and broken endpoints through density clustering algorithms. The entropy value of the directional gradient histogram is used to quantify the degree of disorder in the gradient direction distribution, reflecting the edge blurring characteristics caused by smudging. The adaptive region growing algorithm controls the growth process through dynamic thresholds to accurately segment character missing regions. These feature quantification methods together construct an objective index system for defect evaluation.

[0046] The spatiotemporal feature fusion module learns the temporal evolution of defect features through an LSTM network. The time dimension corresponds to consecutive batches in the printing process, while the spatial dimension corresponds to the defect distribution in different parts of the character. The quantized evaluation vector output by this module has spatiotemporal consistency, reflecting both the static characteristics of the current print quality and the dynamic trend of printhead performance degradation.

[0047] This embodiment detects skewed text using a lightweight EAST network, performs geometric correction using an attention mechanism, identifies character defects using multi-level morphological processing, constructs a quantitative evaluation model to analyze breakage, smearing, and missing features, and finally outputs a comprehensive evaluation vector through a spatiotemporal fusion module, thus constructing a complete print quality evaluation chain and providing a reliable quantitative basis for predicting print cartridge life.

[0048] In some embodiments, the EAST text detection network, based on depthwise separable convolution, performs multi-scale feature extraction on the normalized image feature matrix, outputting a text location feature map containing the geometric center coordinates and tilt angle of the text region, including: Hierarchical feature extraction is performed on the input normalized image feature matrix using a deep separable convolutional network. The high-resolution feature map output by the shallow network is used to capture character edge details, while the low-resolution feature map output by the deep network is used to identify semantic information of text regions, resulting in multi-scale feature maps. The multi-scale feature map is input into the feature pyramid fusion module. Semantic information is propagated through the top-down upsampling path, and geometric details are enhanced by the bottom-up lateral connection path to obtain the fused multi-scale feature map. The fused multi-scale feature map is subjected to bi-branch prediction, which includes a first branch and a second branch. The first branch generates a text region probability map through convolutional kernel sliding detection, which is denoted as the first output result. The second branch regresses the center coordinates, height, width and rotation angle parameters of each effective text region, which is denoted as the second output result. A rotated text detection box is constructed based on the first output result and the second output result. The first output result is used to filter candidate regions, and the second output result is used to determine the spatial pose of the text box. The rotated text detection box is subjected to non-maximum suppression based on rotation IoU, and the output text position feature map with rotation angle is generated.

[0049] In this embodiment, a depthwise separable convolutional network refers to a lightweight network structure that decomposes standard convolution into two steps: depthwise convolution and pointwise convolution. Depthwise convolution performs spatial convolution independently on each input channel, while pointwise convolution combines channel features through 1×1 convolutions. This structure significantly reduces the number of parameters while maintaining feature extraction capabilities by separating the processing of spatial correlation and channel correlation. The high-resolution feature maps output by shallow networks contain rich edge and texture information, suitable for capturing the fine structure of characters; the low-resolution feature maps output by deep networks have stronger semantic representation capabilities, facilitating the recognition of the overall distribution of text regions.

[0050] The feature pyramid fusion module is a multi-scale feature integration structure that combines top-down and bottom-up approaches. The top-down approach upsamples high-level semantic features to lower levels, while the bottom-up approach supplements the high-level geometric details of the lower levels with lateral connections. This bidirectional fusion mechanism preserves the semantic information of the high levels while enhancing the localization accuracy of the low levels, enabling the fused multi-scale feature map to possess both accurate semantic understanding and refined spatial representation capabilities.

[0051] The first branch of the dual-branch prediction mechanism calculates the probability that each position belongs to a text region by sliding a 3×3 convolution kernel, generating a text region probability map. The second branch uses a 1×1 convolution to regress the geometric parameters of the text region, including center coordinates, height, width, and rotation angle. This separate prediction strategy allows the network to focus on tasks of different natures, ensuring both the recall rate of text detection and improving the accuracy of geometric parameter regression.

[0052] Rotation IoU nonmaximum suppression is a deduplication algorithm based on the intersection-union ratio (IoU) of rotated rectangular boxes. It accurately assesses the similarity between detected boxes by calculating the ratio of the overlap area to the union area between rotated text boxes. Compared to traditional horizontal box IoU, rotation IoU more accurately reflects the spatial relationships of tilted text, effectively avoiding false suppression caused by angle differences. This processing ensures that the final output text location feature map completely covers all text regions while avoiding duplicate detections.

[0053] This embodiment extracts multi-scale features through depthwise separable convolution, combines feature pyramid fusion and a bi-branch prediction mechanism to achieve accurate localization of tilted text, and finally uses rotation IoU non-maximum suppression to output the text position feature map. This significantly improves computational efficiency while ensuring detection accuracy, making it particularly suitable for quality assessment scenarios of thermal printed text.

[0054] In some embodiments, the multi-level morphological processing pipeline includes a primary processing pipeline, a secondary processing pipeline, and a high-level processing pipeline. A multi-level morphological processing pipeline is constructed on the character analysis region. The parallel morphological processing pipeline is executed using a sequence of dynamically sized structuring elements to extract character breakage features, halo diffusion features, and missing region features, including: The primary processing pipeline uses circular structural elements for basic morphological operations to eliminate noise interference; the intermediate processing pipeline uses elliptical structural elements for directional morphological operations to enhance stroke continuity; and the advanced processing pipeline uses rectangular structural elements for overall morphological analysis to extract character structural features, resulting in multi-level processing results. The size of the structural elements at each level is dynamically adjusted according to the resolution of the character analysis region to generate a sequence of structural elements of different sizes. Morphological operations of multiple structural element sizes are performed in parallel on the same processing level. The results of the parallel operations are then fused to obtain the final morphological processing result. Based on the final morphological processing results, character breakage features are extracted using skeleton extraction and endpoint detection algorithms, halo diffusion features are extracted using edge gradient analysis and region diffusion detection algorithms, and missing region features are extracted using connected component analysis and region comparison algorithms.

[0055] In this embodiment, the multi-level morphological processing pipeline consists of a parallel processing architecture comprising three processing levels: primary, intermediate, and advanced. Each level performs targeted processing on character features at different scales. Preferably, the primary processing pipeline uses a 3×3 circular structural element to perform basic erosion and dilation operations, mainly used to eliminate dot noise and isolated interference commonly found in thermal printing. The intermediate processing pipeline uses a 5×5 elliptical structural element to perform directional opening and closing operations, with its major axis aligned with the direction of the character strokes, effectively repairing broken strokes and preserving the original shape of the character. The advanced processing pipeline uses a 7×7 rectangular structural element for overall morphological analysis, focusing on extracting the macroscopic structural features of the character.

[0056] The dynamic-size structuring element sequence is a set of kernel sizes that is automatically adjusted based on the actual resolution of the character analysis region. Its adjustment strategy is determined by the ratio between character stroke width and image resolution. Larger structuring elements are used for high-resolution regions to ensure processing quality, while smaller elements are used for low-resolution regions to avoid over-erosion. Multiple morphological operations of different sizes are performed in parallel within the same processing level, and the results of these operations are weighted and fused to balance processing precision and robustness.

[0057] Preferably, the skeleton extraction and endpoint detection algorithms employ the Zhang-Suen thinning algorithm to obtain character skeletons with a single pixel width, distinguishing natural endpoints from broken endpoints by analyzing the spatial distribution density of skeleton endpoints; edge gradient analysis uses the Sobel operator to calculate the second derivative in the edge direction, assessing the degree of smudging diffusion through the consistency of gradient directions; connected component analysis labels connected regions based on the four-neighbor or eight-neighbor criteria, identifying missing regions by comparing the region coverage with a standard template. These feature extraction methods collectively constitute a complete quantitative evaluation system for printing defects, providing an objective basis for subsequent printing quality analysis.

[0058] This embodiment achieves accurate extraction of broken, blurred, and missing features of printed characters through parallel processing that dynamically adjusts the size of structural elements, combined with skeleton extraction, edge analysis, and connected component detection. This method significantly improves the accuracy and adaptability of thermal printing quality inspection, providing a reliable basis for the quantitative evaluation of printing effects.

[0059] In some embodiments, the quantitative evaluation vector is spatiotemporally aligned with environmental sensing information to construct a print quality degradation model. A density peak clustering algorithm is used to identify the coupling relationship between print core aging characteristics and temperature and humidity interference characteristics, generating a multidimensional quality index set including color density gradient, lifetime decay coefficient, and environmental compensation factor, including: Establish a temporal alignment relationship between the quantized evaluation vector and the environmental sensing information flow, and use a bidirectional LSTM network to extract the temporal correlation of cross-modal features to construct a spatiotemporal feature tensor that includes the device's operating status and environmental conditions. An improved density peak clustering algorithm is applied to the spatiotemporal feature tensor. The Mahalanobis distance metric is introduced and combined with the Gaussian mixture model to fit the data distribution under different working conditions, and the inherent aging feature mode and environmental disturbance feature mode of the printing core are separated. Degradation assessment sub-models are established for the inherent aging characteristic mode and the environmental disturbance characteristic mode, respectively. The degradation assessment sub-models include the core aging sub-model, the environmental disturbance sub-model, and the dynamic compensation sub-model. The mechanism aging sub-model constructs a nonlinear decay function based on the amount of developer consumed to obtain the first output parameter. The environmental interference sub-model establishes a mapping relationship between temperature and humidity and print quality using support vector regression to obtain the second output parameter. The dynamic compensation sub-model achieves real-time error correction based on Kalman filtering to obtain the third output parameter. The first, second, and third output parameters are input into the quality index fusion module. The importance weights of different feature dimensions are dynamically adjusted through an attention mechanism to generate a multi-dimensional quality index set that includes color concentration gradient, lifetime decay coefficient, and environmental compensation factor. The color concentration gradient is generated by comparing multispectral imaging data with a reference color chart, the lifetime decay coefficient is calculated by the change rate of the printhead drive current, and the environmental compensation factor is a combination of the deviation between the real-time readings of the temperature and humidity sensor and the historical best operating conditions.

[0060] In this embodiment, the spatiotemporal alignment of the quantized evaluation vector and environmental sensing information is achieved through timestamp synchronization and spatial coordinate mapping, establishing a precise correspondence between the print quality parameters obtained from image analysis and the temperature and humidity data collected by environmental sensors. A bidirectional LSTM network is used to capture the temporal correlation of cross-modal features. The number of hidden layer nodes is automatically adjusted according to the sensor sampling frequency and image processing cycle to ensure effective modeling of the dynamic characteristics of print quality changes over time and in the environment. The spatiotemporal feature tensor is a multidimensional data structure formed by splicing and fusing the quantized evaluation vector and environmental sensing information in the time and feature dimensions. Its dimensionality is determined by the number of feature channels and the length of the time window.

[0061] The improved density peak clustering algorithm introduces Mahalanobis distance metric on the basis of traditional algorithms. It eliminates the influence between features of different dimensions by calculating the covariance matrix between feature points. At the same time, it combines a Gaussian mixture model to probabilistically model the data distribution, which can accurately identify the cluster centers of printhead aging characteristics and environmental disturbance characteristics in the feature space. The inherent aging characteristic mode characterizes the irreversible degradation process of printhead material performance, and its change trend is mainly driven by the printing workload; the environmental disturbance characteristic mode reflects the instantaneous impact of external conditions such as temperature and humidity on printing effect, and has periodic fluctuation characteristics.

[0062] The printhead aging sub-model establishes a nonlinear mapping between developer consumption and print quality, fitting the printhead performance degradation curve using an exponential decay function. The environmental interference sub-model utilizes the kernel function technique of support vector regression to construct a complex nonlinear relationship between temperature and humidity parameters and print quality indicators in a high-dimensional feature space. The dynamic compensation sub-model uses a Kalman filter algorithm to estimate the system state in real time and correct prediction errors, with its process noise and observation noise parameters adaptively adjusted based on historical data. The attention mechanism in the quality indicator fusion module adopts a query-key-value architecture, dynamically allocating fusion weights by calculating the relevance scores of different feature dimensions to ensure that key quality indicators receive higher attention. The generation process of the multi-dimensional quality indicator set comprehensively considers the inherent degradation law of the printing system and external environmental interference factors, providing a comprehensive and objective quantitative basis for print quality assessment.

[0063] This embodiment establishes a print quality degradation model through spatiotemporal alignment and density peak clustering algorithms, effectively separating print cartridge aging characteristics from environmental interference characteristics. Employing multi-model fusion and attention mechanisms, it generates a multi-dimensional quality index set including color gradient, lifetime decay, and environmental compensation factors, achieving accurate print quality assessment and lifetime prediction, significantly improving detection accuracy and environmental adaptability.

[0064] In some embodiments, an improved density peak clustering algorithm is applied to the spatiotemporal feature tensor, and a Mahalanobis distance metric is introduced combined with a Gaussian mixture model to fit the data distribution under different working conditions, separating the inherent aging characteristic mode and the environmental disturbance characteristic mode of the printing core, including: Multi-dimensional feature decoupling is performed on the spatiotemporal feature tensor, and the correlation between feature dimensions is established through covariance matrix analysis to form a feature space; A density peak calculation model based on Mahalanobis distance is constructed to calculate the local density and minimum distance of each sample point in the feature space, forming a density-distance distribution map; By setting dynamic threshold boundaries on the density-distance distribution map, high-density core points and low-density edge points are automatically identified, and the samples are divided into multiple candidate clusters. For each candidate cluster, a Gaussian distribution is fitted, the mean vector and covariance matrix of each cluster are calculated, and a probability density distribution model is established. By analyzing the overlap of probability density distributions, adjacent clusters with similarity exceeding a preset threshold are merged, and heterogeneous clusters with dissimilarity exceeding a preset threshold are separated. Modal labeling is performed on the final clustering results. Clusters with stable temporal characteristics are labeled as the inherent aging characteristic modes of the printing core, and clusters with environmentally related fluctuation characteristics are labeled as environmental disturbance characteristic modes.

[0065] In this embodiment, multi-dimensional feature decoupling refers to performing principal component analysis on the spatiotemporal feature tensor. Eigenvalue decomposition projects the original high-dimensional features onto mutually orthogonal low-dimensional subspaces, eliminating redundant information between features. Covariance matrix analysis is used to quantify the linear correlation between feature dimensions; its diagonal elements represent feature autocorrelation, while off-diagonal elements reflect cross-correlation. The feature space is a low-dimensional vector space formed after feature decoupling. Its dimensionality is automatically determined based on the cumulative contribution rate of the eigenvalues, typically retaining principal components with a contribution rate exceeding 85%.

[0066] The density peak calculation model based on Mahalanobis distance, by considering the covariance structure of the feature distribution, more accurately reflects the true distribution of samples in a non-uniform feature space compared to Euclidean distance. Local density calculation employs a truncated kernel density estimation method, using the number of neighbors within a fixed radius around a sample point as the density metric; the minimum distance is defined as the minimum Mahalanobis distance from a sample point to a higher density point. The density-distance distribution map visualizes the local density and minimum distance of sample points in two dimensions, forming a decision map for cluster center identification.

[0067] The dynamic threshold boundary employs an adaptive dual-threshold strategy. The threshold for high-density core points is automatically calculated based on the statistical characteristics of samples in the upper right quadrant of the distribution map, while the threshold for low-density edge points is determined based on the percentiles of the overall distribution. Candidate cluster partitioning follows the density reachability principle, ensuring that samples within the same cluster exhibit continuous density distribution characteristics. Gaussian distribution fitting utilizes an iterative optimization algorithm based on expectation maximization, with the number of mixture components automatically selected using the Bayesian information criterion to avoid overfitting. Preferably, the overlap analysis of probability density distributions is based on KL divergence calculation; when the KL divergence of two distributions is below a preset threshold, they are merged to reflect the similarity between clusters.

[0068] The modal annotation process combines time-series analysis and environmental correlation testing. The criteria for determining inherent aging characteristic modes include long-term monotonic change trends and strong correlation with printing workload; environmental disturbance characteristic modes must satisfy synchronous fluctuation characteristics with environmental parameters and short-term reversibility. Through this refined modal separation method, the intrinsic mechanisms and external causes of print quality changes can be accurately distinguished.

[0069] This embodiment accurately separates the inherent aging characteristics of the printer cartridge from environmental disturbance characteristics through multidimensional feature decoupling and Mahalanobis distance density peak clustering. Adaptive thresholding and probability distribution analysis are used to achieve precise modal segmentation, and combined with temporal characteristic annotation, it effectively distinguishes between intrinsic factors and external interferences that cause print quality degradation, providing a more accurate diagnostic basis for printing system maintenance.

[0070] In some embodiments, a graph attention network is established based on a multidimensional quality index set to model the spatial correlation between printhead temperature fluctuations, media coating characteristics, and character color rendering quality. A temporal convolutional network is used to capture progressive failure modes, outputting a verification report containing an instantaneous quality score, defect location map, and lifespan prediction curve. The verification report includes a dynamic threshold early warning module and an optimal maintenance strategy recommendation module, including: A spatial association model based on graph attention network is constructed. The printhead temperature distribution characteristics, media coating spectral characteristics, and character color development quality characteristics are used as nodes of a heterogeneous graph. The association strength matrix between nodes is calculated through spatial attention mechanism. The temperature-coating association edge is constrained by the heat conduction equation, and the coating-color development association edge is mapped through spectral reflectance. The spatial association graph is output. The spatial correlation graph is input into the spatiotemporal feature fusion module. While preserving the spatial correlation characteristics of the nodes, the dynamic evolution characteristics of each node are extracted through a temporal convolutional network. Gated temporal convolutional units are used to extract features from the printhead temperature fluctuation sequence. Causal convolution is used to capture the gradual change process of the medium coating characteristics, generating a composite feature map that simultaneously contains spatial correlation and dynamic evolution characteristics. A dual-channel evaluation network is constructed based on composite feature maps. The dual-channel evaluation network includes a quality evaluation channel and a lifetime prediction channel, including: The quality assessment channel aggregates global features through graph pooling operations and calculates real-time quality scores using a weighted self-attention mechanism. The life prediction channel inputs node features into a neural differential equation network to simulate the thermal fatigue accumulation process of the printhead and outputs a life prediction curve with confidence intervals. Key abnormal nodes are located using the graph attention weight backpropagation algorithm. Combined with the spatial coordinate mapping relationship of the nodes, a defect location map with thermal value annotations is generated in the original printed area. The color abnormality area is quantified by spectral difference, and the structural defect area is evaluated by topological feature similarity. An inspection report is generated based on real-time quality scores, life prediction curves, and defect location maps.

[0071] In this embodiment, the spatial association model based on graph attention networks treats printhead temperature distribution characteristics, media coating spectral characteristics, and character color development quality characteristics as three different attribute nodes in the graph structure, representing their respective physical properties through node feature vectors. The spatial attention mechanism employs a multi-head attention computation approach, calculating the association strength between nodes using a learnable weight matrix. The temperature-coating association edge is physically constrained by Fourier's law in the heat conduction equation, while the coating-color development association edge establishes a mapping relationship between spectral reflectance and color development concentration based on the Lambert-Beer law. The association strength matrix is ​​normalized using a softmax function to ensure the interpretability of the attention weights.

[0072] The spatiotemporal feature fusion module employs a residual connection structure. While preserving the topology of the spatial correlation graph, it extracts the temporal dynamic features of each node through a temporal convolutional network. Gated temporal convolutional units control information flow through a gating mechanism, effectively capturing the non-stationary fluctuations in printhead temperature. Causal convolution ensures that the modeling of the gradual change process of the media coating properties conforms to temporal causality, preventing future information leakage. The composite feature map fuses spatial correlation and temporal evolution features through feature stitching, forming a unified representation containing multi-dimensional information.

[0073] In the dual-channel evaluation network, the quality evaluation channel employs global attention pooling, aggregating features from all nodes in the graph using learnable query vectors. It then dynamically assigns importance weights to different feature dimensions using a self-attention mechanism, ultimately outputting a standardized quality score through a fully connected layer. The lifetime prediction channel inputs node features into a neural network of ordinary differential equations, simulating the continuous accumulation of thermal fatigue in the printhead through numerical integration. It outputs a lifetime prediction curve with a confidence interval, the width of which is automatically adjusted based on prediction uncertainty.

[0074] The graph attention weight backpropagation algorithm identifies key nodes with the greatest impact on quality assessment by calculating the gradient of the output layer with respect to the attention weights. Combining the mapping relationship between node spatial coordinates and the printed area, it generates a defect location map on the original image. Color anomaly regions are quantified and labeled by calculating the Euclidean distance between the actual spectrum and the standard spectrum, while structural defect regions are evaluated using the cosine similarity of graph topological features. Finally, a visualized defect distribution map with thermal value annotations is generated. The inspection report integrates all assessment results, monitors quality status changes in real time through a dynamic threshold early warning module, and provides personalized maintenance solutions based on the life prediction curve and defect distribution characteristics through an optimal maintenance strategy suggestion module.

[0075] This embodiment models the complex relationships between printing elements by constructing a graph attention network, and combines temporal feature fusion and a dual-channel evaluation mechanism to achieve accurate assessment of print quality and lifespan prediction. This embodiment embeds physical laws into the network structure, providing interpretable defect localization and personalized maintenance solutions, significantly improving the status monitoring and early warning capabilities of the printing system.

[0076] In some embodiments, the quality assessment channel aggregates global features through graph pooling operations and calculates an instantaneous quality score using a weighted self-attention mechanism, including: The peak heat load feature is extracted by maximum pooling at the printhead temperature node, the coating uniformity index is generated by average pooling at the media coating node, and the comprehensive color development score is obtained by weighted average pooling at the color development quality node. The weight of each color development quality node is calculated by the self-attention mechanism. The peak heat load feature, coating uniformity index and comprehensive color development score are integrated into the graph pooling output feature. The correlation strength between the printhead temperature node and the color development quality node is calculated through the self-attention weight calculation module. The self-attention weight calculation module contains a learnable bias parameter matrix, which modulates the initial weights through the spectral characteristics of the media coating node and outputs the final self-attention weight distribution. The actual color development quality data of historical printed samples is collected as a true reference. The graph pooling output features are compared with the true data. The bias parameters in the self-attention weight calculation module are dynamically updated through backpropagation of error gradient. The graph pooling output features and the dynamically evolved features are concatenated using tensors, and the concatenated feature vector is mapped to an instantaneous quality score in the range of 0-100 through a fully connected layer.

[0077] In this embodiment, graph pooling employs differentiated feature aggregation strategies for different types of nodes in a heterogeneous graph structure. Max pooling is used to extract the peak thermal load characteristics of the printhead temperature node, representing the most severe heating condition by retaining the highest temperature value in the region; average pooling processes the media coating node to generate a coating uniformity index, reflecting the overall distribution characteristics of the coating thickness; weighted average pooling is applied to the color development quality node, with its node weights dynamically calculated through a self-attention mechanism to ensure that important color development areas receive higher weights, ultimately outputting a comprehensive color development score. This differentiated pooling strategy ensures that the physical characteristics of various nodes are not homogenized.

[0078] The self-attention weight calculation module modulates the initial attention weights using a learnable bias parameter matrix. The spectral characteristics of the medium coating nodes serve as a modulation factor, influencing the calculation of the correlation strength between temperature nodes and color development quality nodes through matrix multiplication. The bias parameter matrix is ​​randomly initialized during the initial training phase and continuously optimized through subsequent backpropagation. This modulation mechanism ensures that the attention weight distribution considers both the direct correlation between nodes and the influence of the optical properties of the medium coating, improving the physical plausibility of the correlation strength calculation.

[0079] The error gradient backpropagation process employs the mean squared error loss function, comparing the graph pooling output features with the actual color rendering quality values ​​of historical printed samples. The gradient of the loss function with respect to the bias parameters is calculated using the chain rule, and the parameters are updated using stochastic gradient descent. Dynamic evolution features refer to the temporal variation features of nodes extracted through a temporal convolutional network. These features are concatenated with the graph pooling output features to form a comprehensive feature vector. Finally, this vector is mapped to a standardized real-time quality score in the 0-100 range through a fully connected layer and a sigmoid activation function, achieving accurate quantitative evaluation of print quality.

[0080] This embodiment uses a differentiated graph pooling strategy and a modulated self-attention mechanism to accurately aggregate features from multiple sources and quantify the correlation strength. Combined with supervised optimization based on historical data, it finally outputs a standardized quality score, achieving refined modeling of the complex correlation between temperature, coating, and color development during the printing process. This significantly improves the accuracy and physical interpretability of quality assessment.

[0081] By adopting the above technical solutions, this invention differs from existing technologies and possesses the following beneficial effects: By constructing a complete print quality assessment system, it effectively solves the problem that existing detection methods cannot quantitatively assess the trend of print quality degradation and the coupled effects of the environment. Using an improved OTSU algorithm combined with multispectral imaging technology, high-precision segmentation of the uneven background color of thermal media is achieved; through EAST text detection and multi-level morphological processing, defect patterns such as breakpoints, smudging, and missing characters are accurately identified; using spatiotemporal alignment and density peak clustering algorithms, the inherent aging characteristics of the printhead and environmental disturbance characteristics are separated; finally, based on graph attention networks and temporal convolutional networks, a spatial correlation model of printhead temperature, media characteristics, and color development quality is established, outputting a verification report including real-time quality scores, defect location maps, and lifespan prediction curves. These technical solutions overcome the limitations of traditional binary assessment, establish a complete concentration gradient quantitative evaluation system, can provide early warning of printhead performance degradation, and eliminate the interference of environmental temperature and humidity on the assessment results, significantly improving the accuracy, reliability, and environmental adaptability of print quality detection.

[0082] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0083] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0084] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A detection method for verifying printing quality, characterized in that, include: Collect scanned image information and environmental sensing information of thermal printing media. The scanned image information includes the grayscale distribution characteristics of the printed character area and the background area. The environmental sensing information includes real-time temperature and humidity parameters and printer module working status parameters. An improved OTSU algorithm combined with an illumination compensation module is used to generate an adaptive segmentation threshold for the scanned image information. Geometric correction is used to eliminate medium deformation interference, and a standardized image feature matrix containing character integrity features, edge sharpness features, and background noise distribution features is extracted. The standardized image feature matrix is ​​used to locate tilted text regions based on the EAST text detection algorithm. Morphological operations are combined to identify defect patterns of broken characters, blurred characters, and missing characters, and a quantitative evaluation vector containing character loss rate, structural similarity index, and gradient change rate is generated. The quantitative evaluation vector is spatiotemporally aligned with environmental sensing information to construct a print quality degradation model. Density peak clustering algorithm is used to identify the coupling relationship between print core aging characteristics and temperature and humidity interference characteristics, generating a multidimensional quality index set including color concentration gradient, lifetime decay coefficient and environmental compensation factor. A graph attention network is established based on the multidimensional quality index set to model the spatial correlation between printhead temperature fluctuations, media coating characteristics, and character color rendering quality. A temporal convolutional network is used to capture progressive failure modes and output a verification report that includes an instant quality score, defect location map, and life prediction curve. The verification report includes a dynamic threshold early warning module and an optimal maintenance strategy suggestion module.

2. The detection method for verifying printing effect according to claim 1, characterized in that, An improved OTSU algorithm combined with an illumination compensation module is used to generate an adaptive segmentation threshold for the scanned image information. Geometric correction is used to eliminate medium deformation interference, and a standardized image feature matrix containing character integrity features, edge sharpness features, and background noise distribution features is extracted, including: The reflectance characteristic spectrum of the thermosensitive medium is obtained by multispectral imaging, and the reflectance characteristic spectrum is input into the local adaptive weighting module to generate dynamic weighting coefficients for different medium regions. The inter-class variance calculation of the OTSU algorithm is optimized using the dynamic weight coefficients, and an initial segmentation threshold that adapts to the inhomogeneity of the medium is output. The initial segmentation threshold is input into the illumination compensation module, and the dark areas of the scanned image information are gamma-corrected through a nonlinear brightness mapping function, while suppressing overexposure in the highlight areas, to generate a pre-processed image with balanced illumination. The preprocessed image is subjected to medium edge detection based on Hough transform. A perspective transformation matrix is ​​constructed based on the detected edge features. The medium deformation is corrected by the perspective transformation matrix to obtain a geometrically regular image to be analyzed. Within the ROI region of the image to be analyzed, the improved Sobel-Feldman operator is applied to extract the gradient magnitude of the character edges, while the noise power spectral density of the background region is calculated based on the Fourier transform, forming an intermediate feature set containing spatial and frequency domain features. The intermediate feature set is input into a pre-trained character skeleton topology network to extract the skeleton connectivity features of the characters. Combined with the breakpoint distribution map generated by morphological erosion operation, the character defect rate index is calculated. At the same time, a multi-scale structural similarity algorithm is used to evaluate the character deformation distortion. By fusing the edge gradient magnitude, noise power spectral density, skeleton connectivity features, and deformation distortion index, a feature descriptor with uniform dimensions is generated through a feature standardization layer. After dimensionality reduction processing by principal component analysis, a traceable standardized image feature matrix is ​​output.

3. The detection method for verifying printing effect according to claim 2, characterized in that, The initial segmentation threshold is input into the illumination compensation module, and gamma correction is performed on the dark areas of the scanned image information through a nonlinear brightness mapping function, while suppressing overexposure in the highlight areas, generating a pre-processed image with balanced illumination, including: A brightness perception conversion module based on a retinal model is constructed to convert the RGB color space of the scanned image information to the CIELAB uniform color space. The brightness component is adjusted for perceptual consistency through the human eye's visual sensitivity function, so that the converted scanned image information includes dark areas, medium brightness areas and highlight areas. A piecewise gamma correction function is constructed. In the dark area, dynamic gamma coefficients are used to enhance contrast. In the medium brightness area, linear mapping is maintained. In the highlight area, an inverse S-curve is applied to compress the dynamic range, and the corrected scan image information is obtained. The compensation intensity of the dark, medium and bright areas of the corrected scanned image information is dynamically adjusted based on the reflectance value of the thermosensitive medium in the 550nm band. The double exposure fusion algorithm is enabled in areas with reflectance below 40%. Using the gradient magnitude map of the original scanned image information as the guide image, edge-aware smoothing processing is performed on the scanned image information after illumination compensation, so as to eliminate uneven illumination while preserving the high-frequency details of character strokes. The color constancy correction is performed on the scanned image information after elimination. By calculating the color coordinate offset of the white points in the image, the hue component H is compensated and rotated in the HSV color space to eliminate the color deviation caused by the difference in light color temperature, and the preprocessed image is obtained.

4. The detection method for verifying printing effect according to claim 1, characterized in that, The standardized image feature matrix is ​​used to locate tilted text regions based on the EAST text detection algorithm. Morphological operations are then combined to identify defect patterns such as broken characters, blurred characters, and missing characters. A quantitative evaluation vector is generated, including character loss rate, structural similarity index, and gradient change rate. The EAST text detection network, which is improved based on depthwise separable convolution, performs multi-scale feature extraction on the standardized image feature matrix and outputs a text location feature map containing the geometric center coordinates and tilt angle of the text region. The text location feature map is input into an attention-based affine transformation module, and the optimal text correction parameters are learned through a spatial transformation network to generate a geometrically regular character analysis region, while preserving the pixel mapping relationship between the original image and the corrected image. A multi-level morphological processing pipeline is constructed on the character analysis region. A parallel morphological processing pipeline is executed using a dynamic size structuring element sequence to extract character breakage features, halo diffusion features, and missing region features. A quantitative evaluation sub-model is constructed for the character breakage feature, smear diffusion feature and missing region feature respectively. The breakage feature sub-model calculates the cumulative breakage amount by the density distribution of the skeleton endpoints. The smear diffusion feature sub-model uses the entropy value of the directional gradient histogram to measure the degree of edge blurring. The missing region feature sub-model evaluates the proportion of missing area based on the adaptive region growth algorithm. The output features of each sub-model are input into the spatiotemporal feature fusion module, and the temporal evolution law of the defect features is learned through the long short-term memory network to generate a quantitative evaluation vector with spatiotemporal consistency.

5. The detection method for verifying printing effect according to claim 4, characterized in that, The EAST text detection network, based on depthwise separable convolution, performs multi-scale feature extraction on the normalized image feature matrix, outputting a text location feature map containing the geometric center coordinates and tilt angle of the text region, including: Hierarchical feature extraction is performed on the input normalized image feature matrix using a deep separable convolutional network. The high-resolution feature map output by the shallow network is used to capture character edge details, while the low-resolution feature map output by the deep network is used to identify semantic information of text regions, resulting in multi-scale feature maps. The multi-scale feature map is input into the feature pyramid fusion module, and semantic information is propagated through the top-down upsampling path. Geometric details are enhanced by the bottom-up lateral connection path to obtain the fused multi-scale feature map. The fused multi-scale feature map is subjected to bi-branch prediction, which includes a first branch and a second branch. The first branch generates a text region probability map through convolutional kernel sliding detection, which is denoted as the first output result. The second branch regresses the center coordinates, height, width and rotation angle parameters of each effective text region, which is denoted as the second output result. A rotated text detection box is constructed based on the first output result and the second output result. The first output result is used to filter candidate regions, and the second output result is used to determine the spatial pose of the text box. The rotated text detection box is subjected to non-maximum suppression based on rotation IoU, and the output text position feature map with rotation angle is generated.

6. The detection method for verifying printing effect according to claim 4, characterized in that, The multi-level morphological processing pipeline includes a primary processing pipeline, a mid-level processing pipeline, and a high-level processing pipeline. A multi-level morphological processing pipeline is constructed on the character analysis region. A parallel morphological processing pipeline is executed using a dynamic-size structuring element sequence to extract character breakage features, halo diffusion features, and missing region features, including: The primary processing pipeline uses circular structural elements for basic morphological operations to eliminate noise interference; the intermediate processing pipeline uses elliptical structural elements for directional morphological operations to enhance stroke continuity; and the advanced processing pipeline uses rectangular structural elements for overall morphological analysis to extract character structural features, resulting in multi-level processing results. The size of the structural elements at each level is dynamically adjusted according to the resolution of the character analysis region to generate a sequence of structural elements of different sizes. Morphological operations of multiple structural element sizes are performed in parallel on the same processing level. The results of the parallel operations are then fused to obtain the final morphological processing result. Based on the final morphological processing results, character breakage features are extracted using skeleton extraction and endpoint detection algorithms, halo diffusion features are extracted using edge gradient analysis and region diffusion detection algorithms, and missing region features are extracted using connected component analysis and region comparison algorithms.

7. The detection method for verifying printing effect according to claim 1, characterized in that, The quantitative evaluation vector is spatiotemporally aligned with environmental sensing information to construct a print quality degradation model. A density peak clustering algorithm is used to identify the coupling relationship between print core aging characteristics and temperature and humidity interference characteristics, generating a multidimensional quality index set including color density gradient, lifetime decay coefficient, and environmental compensation factor, including: Establish a temporal alignment relationship between the quantized evaluation vector and the environmental sensing information flow, and use a bidirectional LSTM network to extract the temporal correlation of cross-modal features to construct a spatiotemporal feature tensor that includes the device's operating status and environmental conditions. An improved density peak clustering algorithm is applied to the spatiotemporal feature tensor. A Mahalanobis distance metric is introduced and combined with a Gaussian mixture model to fit the data distribution under different working conditions, thereby separating the inherent aging feature mode and the environmental disturbance feature mode of the printing core. Degradation assessment sub-models are established for the inherent aging characteristic mode and the environmental disturbance characteristic mode, respectively. The degradation assessment sub-models include the core aging sub-model, the environmental disturbance sub-model, and the dynamic compensation sub-model. The mechanism aging sub-model constructs a nonlinear decay function based on the amount of developer consumed to obtain the first output parameter. The environmental interference sub-model establishes a mapping relationship between temperature and humidity and print quality using support vector regression to obtain the second output parameter. The dynamic compensation sub-model achieves real-time error correction based on Kalman filtering to obtain the third output parameter. The first, second, and third output parameters are input into the quality index fusion module. The importance weights of different feature dimensions are dynamically adjusted through an attention mechanism to generate a multi-dimensional quality index set that includes color concentration gradient, lifetime decay coefficient, and environmental compensation factor. The color concentration gradient is generated by comparing multispectral imaging data with a reference color chart, the lifetime decay coefficient is calculated by the change rate of the printhead drive current, and the environmental compensation factor is a combination of the deviation between the real-time readings of the temperature and humidity sensor and the historical best operating conditions.

8. The detection method for verifying printing effect according to claim 7, characterized in that, An improved density peak clustering algorithm is applied to the spatiotemporal feature tensor. A Mahalanobis distance metric is introduced, combined with a Gaussian mixture model, to fit the data distribution under different operating conditions, separating the inherent aging characteristic modes and environmental disturbance characteristic modes of the printing core, including: Multi-dimensional feature decoupling is performed on the spatiotemporal feature tensor, and the correlation between feature dimensions is established through covariance matrix analysis to form a feature space; A density peak calculation model based on Mahalanobis distance is constructed to calculate the local density and minimum distance of each sample point in the feature space, forming a density-distance distribution map; By setting dynamic threshold boundaries on the density-distance distribution map, high-density core points and low-density edge points are automatically identified, and the samples are divided into multiple candidate clusters. For each candidate cluster, a Gaussian distribution is fitted, the mean vector and covariance matrix of each cluster are calculated, and a probability density distribution model is established. By analyzing the overlap of probability density distributions, adjacent clusters with similarity exceeding a preset threshold are merged, and heterogeneous clusters with dissimilarity exceeding a preset threshold are separated. Modal labeling is performed on the final clustering results. Clusters with stable temporal characteristics are labeled as the inherent aging characteristic modes of the printing core, and clusters with environmentally related fluctuation characteristics are labeled as environmental disturbance characteristic modes.

9. The detection method for verifying printing effect according to claim 1, characterized in that, A graph attention network is established based on the aforementioned multidimensional quality index set to model the spatial correlation between printhead temperature fluctuations, media coating characteristics, and character color rendering quality. A temporal convolutional network is used to capture progressive failure modes, outputting a verification report that includes an instantaneous quality score, defect location map, and lifespan prediction curve. This verification report includes a dynamic threshold early warning module and an optimal maintenance strategy suggestion module, comprising: A spatial association model based on graph attention network is constructed. The printhead temperature distribution characteristics, media coating spectral characteristics, and character color development quality characteristics are used as nodes of a heterogeneous graph. The association strength matrix between nodes is calculated through spatial attention mechanism. The temperature-coating association edge is constrained by the heat conduction equation, and the coating-color development association edge is mapped through spectral reflectance. The spatial association graph is output. The spatial correlation graph is input into the spatiotemporal feature fusion module. While preserving the spatial correlation characteristics of the nodes, the dynamic evolution characteristics of each node are extracted through a temporal convolutional network. Specifically, gated temporal convolutional units are used to extract features from the printhead temperature fluctuation sequence, and causal convolution is used to capture the gradual change process of the medium coating characteristics, generating a composite feature graph that simultaneously contains spatial correlation and dynamic evolution characteristics. A dual-channel evaluation network is constructed based on the composite feature map. The dual-channel evaluation network includes a quality evaluation channel and a lifetime prediction channel, including: The quality assessment channel aggregates global features through graph pooling operations and calculates real-time quality scores using a weighted self-attention mechanism. The life prediction channel inputs node features into a neural differential equation network to simulate the thermal fatigue accumulation process of the printhead and outputs a life prediction curve with confidence intervals. Key abnormal nodes are located using the graph attention weight backpropagation algorithm. Combined with the spatial coordinate mapping relationship of the nodes, a defect location map with thermal value annotations is generated in the original printed area. The color abnormality area is quantified by spectral difference, and the structural defect area is evaluated by topological feature similarity. The inspection report is generated based on the real-time quality score, life prediction curve, and defect location map.

10. The detection method for verifying printing effect according to claim 9, characterized in that, The quality assessment channel aggregates global features through graph pooling and calculates an instantaneous quality score using a weighted self-attention mechanism, including: The peak heat load feature is extracted by maximum pooling at the printhead temperature node, the coating uniformity index is generated by average pooling at the media coating node, and the comprehensive color development score is obtained by weighted average pooling at the color development quality node. The weight of each color development quality node is calculated by a self-attention mechanism. The peak heat load feature, coating uniformity index and comprehensive color development score are integrated into a graph pooling output feature. The correlation strength between the printhead temperature node and the color development quality node is calculated through the self-attention weight calculation module. The self-attention weight calculation module contains a learnable bias parameter matrix, which modulates the initial weights through the spectral characteristics of the media coating node and outputs the final self-attention weight distribution. The actual color development quality data of historical printed samples is collected as a true reference. The graph pooling output features are compared with the true data. The bias parameters in the self-attention weight calculation module are dynamically updated through backpropagation of error gradient. The graph pooling output features and the dynamically evolved features are concatenated using tensors, and the concatenated feature vector is mapped to an instantaneous quality score in the range of 0-100 through a fully connected layer.

Images