Industrial vision product surface printing quality detection method

Abstract

The invention provides a product surface printing quality detection method based on industrial vision, which comprises the following steps: acquiring paper type and ink type information used by a printed matter, and extracting material spectral characteristic parameters of a paper and ink combination from a pre-established material spectral characteristic database, comprising scattering reflection characteristics of the paper surface and absorption reflection characteristics of ink; for the area with the quality defect, analyzing the spectral data of the area, extracting refined spectral characteristics, accurately classifying defect types through a deep learning model, and analyzing the influence of paper and ink material characteristics on a defect forming mechanism; according to the defect type and the position information, similar cases are retrieved from a pre-established defect knowledge base, a defect cause analysis result and an improvement suggestion are obtained, the material spectral characteristic parameters are matched with defect characteristics, and an evaluation result of the influence degree of the material characteristics on defect formation is obtained.

Description

Technical Field

[0001] The invention relates to the field of information technology, and in particular to an industrial visual product surface printing quality detection method. Background Art

[0002] In the field of printing quality inspection, spectral analysis technology is facing a key technical challenge: different types of paper and ink will show different spectral characteristics, which makes it a difficult task to establish a universal and sensitive detection model. Due to the diversity of printed products, how to ensure that the detection model can adapt to a variety of different printing materials and ensure the reliability and consistency of the detection results under complex conditions has become a major challenge faced by the entire technical solution. Different types of paper and ink exhibit different characteristics when absorbing and reflecting light, and these differences directly affect the accuracy and stability of spectral analysis. For example, some papers may contain fluorescent brighteners while others do not; differences in the chemical composition, color concentration and coverage of inks can also lead to changes in spectral response. These factors make it difficult for a single spectral detection model to be applicable to all types of printed products. Summary of the invention

[0003] The present invention provides a method for detecting product surface printing quality by industrial vision, which mainly includes:

[0004] Performing a multi-band spectral scan on the printed matter according to preset spectral acquisition parameters to acquire spectral data, using a spectral camera array to synchronously acquire spectral images of multiple bands, and performing real-time preprocessing on the acquired spectral data through a parallel processing unit to obtain preprocessed spectral data;

[0005] Acquire information about the type of paper and ink used in the printed matter, and extract material spectrum characteristic parameters of the paper-ink combination from a pre-established material spectrum characteristic database, including the scattering reflection characteristics of the paper surface and the absorption reflection characteristics of the ink;

[0006] The pre-trained convolutional neural network model is used to extract features from the pre-processed spectral data. The material spectral characteristic parameters of paper and ink are combined to comprehensively analyze the modulation effect of the printed surface microstructure on the spectral reflection signal, extract the key spectral features that characterize the quality of the printed product, and convert the extracted features into feature vectors of fixed dimensions through the feature mapping layer.

[0007] The characteristic vector is analyzed according to the pre-established printed product quality assessment model, the influencing factors of paper and ink material characteristics are introduced, the judgment threshold of the quality assessment model is dynamically adjusted, and the scores of various quality indicators of the printed product are calculated. If the score of a certain indicator is lower than the adjusted threshold, it is determined that there is a quality defect in the area.

[0008] For areas judged to have quality defects, the spectral data of the area is analyzed, and refined spectral features are extracted. The defect types are accurately classified through deep learning models, and the influence of paper and ink material properties on the defect formation mechanism is analyzed;

[0009] According to the defect type and location information, similar cases are retrieved from the pre-established defect knowledge base to obtain defect cause analysis results and improvement suggestions, and the material spectral characteristic parameters are matched with the defect characteristics to obtain the evaluation results of the influence of material characteristics on defect formation;

[0010] Based on the impact assessment results, the defect cause analysis and improvement suggestions in the quality report are optimized and adjusted, including providing targeted process optimization solutions based on the characteristics of paper and ink materials, and generating a refined quality report containing detailed defect information and material improvement suggestions.

[0011] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0012] The present invention discloses a method for detecting the surface printing quality of products using industrial vision. The method collects multi-band spectral data of printed products through a high-speed spectral camera, combines the material spectral characteristics of paper and ink, and uses a convolutional neural network to extract key spectral features. The present invention introduces material property influencing factors, dynamically adjusts the quality assessment model, and realizes accurate defect detection and classification. For the detected defects, the present invention retrieves similar cases from the knowledge base, analyzes the influence of material properties on the formation of defects, and provides targeted process optimization solutions. Through the transfer learning method, the present invention can quickly adapt to new printing materials and realize universal detection of a variety of printed products. The method significantly improves the accuracy and efficiency of printed product quality detection and provides strong support for printing process optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 The present invention is a flowchart of a method for detecting product surface printing quality using industrial vision.

[0014] Figure 2 It is a schematic diagram of an industrial vision product surface printing quality detection method of the present invention.

[0015] Figure 3 It is another schematic diagram of an industrial vision product surface printing quality detection method of the present invention. DETAILED DESCRIPTION

[0016] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.

[0017] like Figure 1 -3. In this embodiment, a method for detecting product surface printing quality by industrial vision may specifically include:

[0018] Step S101, performing a multi-band spectral scan on a printed matter according to preset spectral acquisition parameters to acquire spectral data, using a spectral camera array to synchronously acquire spectral images of multiple bands, and performing real-time preprocessing on the acquired spectral data by a parallel processing unit to obtain preprocessed spectral data.

[0019] A high-speed spectral camera array is used to synchronously collect spectral images of multiple bands, and the spectral image of each band is processed by a grating spectrometry method to obtain spectral curve data. The spectral curve data is normalized, and the normalized spectral curve data is denoised by Daubechies wavelet to obtain denoised spectral data. The denoised spectral data is grouped by cluster analysis, and feature templates of different printed product categories are established to construct a spectral feature library. A support vector machine is used to train the data in the spectral feature library, and the kernel function is set to a radial basis function to obtain a classification model. The spectral feature vector of an unknown printed product is classified and identified according to the classification model to obtain the printed product category.

[0020] Exemplarily, multi-band spectral scanning data is obtained from a printed matter according to preset spectral acquisition parameters, a high-speed spectral camera array is used to synchronously acquire spectral images of multiple bands, and the spectral images of each band are processed by a grating spectrophotometry method to obtain spectral curve data. GPU is used to accelerate the real-time processing of spectral data, and multiple band data are calculated in parallel to improve data processing efficiency. The acquired spectral curve data is normalized, and the normalized spectral curve data is denoised by 5-layer decomposition using Daubechies wavelet, and the first three principal components are selected for dimensionality reduction to obtain a feature vector. The feature vectors are grouped by cluster analysis, feature templates of different printed matter categories are established, and a spectral feature library is constructed. The data in the spectral feature library are trained by a support vector machine, the kernel function is set to a radial basis function, and the penalty parameter C is set to 1.0 to obtain a classification model. The spectral feature vectors of unknown printed matter are classified and identified by the classification model to obtain the printed matter category. According to the identified printed product category, the corresponding quality assessment standard is obtained from the pre-established printed product quality assessment database, and a fuzzy relationship matrix is ​​established. The comprehensive evaluation value is calculated by the weighted average method, and the spectral feature vector of the printed product is evaluated to obtain the printed product quality grade. The quality grade is compared with the preset threshold to determine whether the printed product is qualified. If it is unqualified, a warning message is output. In the process of spectral data acquisition, the preset spectral acquisition parameters include a wavelength range of 400-700nm, a sampling interval of 2nm, and an exposure time of 50ms. The high-speed spectral camera array consists of 5 linear array CCDs, each of which is responsible for collecting spectral data in the 60nm wavelength range. Grating spectrometry technology is adopted, and a 600-line / mm diffraction grating is used to disperse the incident light into spectra of different wavelengths. The GPU parallel processing unit adopts the CUDA architecture, is equipped with 2048 CUDA cores, and has a main frequency of 1.5GHz, to realize parallel processing of data collected simultaneously by 5 CCDs. The spectral curve data normalization adopts the maximum-minimum value normalization method to map the data to the range of 0-1. Daubechies wavelet uses DB4 wavelet, and performs 5-layer wavelet decomposition, retaining the approximate coefficients and the first 3 layers of detail coefficients. The principal component analysis selects the first 3 principal components with a cumulative contribution rate of 95% as feature vectors. Cluster analysis uses the K-means algorithm, sets the number of cluster centers to 10, the number of iterations to 100, and constructs a spectral feature library. The support vector machine classification model uses the radial basis function as the kernel function, the penalty parameter C is set to 1.0, the gamma parameter is set to 0.1, the training data set contains 1000 samples, and the cross validation uses 5-fold cross validation. In the fuzzy comprehensive evaluation, a 5×5 fuzzy relationship matrix is ​​established, and the evaluation factors include color reproduction, hierarchical expression, detail fidelity, color uniformity and overall effect, with weights of 0.3, 0.2, 0.2, 0.15, and 0.15 respectively. The weighted average method is used to calculate the comprehensive evaluation value, and the score range is 0-100 points.Set the quality level threshold to 80 points. Points below this threshold are considered unqualified and a warning message is output.

[0021] Step S102, obtaining information on the paper type and ink type used for the printed product, and extracting material spectrum characteristic parameters of the paper-ink combination from a pre-established material spectrum characteristic database, including the scattering reflection characteristics of the paper surface and the absorption reflection characteristics of the ink.

[0022] The grayscale co-occurrence matrix of the printed product surface is obtained, and the contrast, correlation, energy and homogeneity are calculated according to the grayscale co-occurrence matrix to obtain the texture feature vector of the printed product surface; the texture feature vector is matched with the pre-stored sample data by using the nearest neighbor algorithm to determine the paper type and ink type of the printed product; for the paper type and ink type, the spectral characteristic parameters of similar paper and ink combinations are obtained from a pre-established material spectral characteristic database, and the spectral characteristic parameters are interpolated and calculated by using the cubic spline interpolation method to obtain a complete spectral curve of the printed product; according to the complete spectral curve, the scattering and reflection process of light on the printed product surface is simulated by using a ray tracing algorithm to calculate the bidirectional reflection distribution function of the printed product surface; the scattering coefficient and absorption coefficient of the ink type are calculated by using the Kubecka-Munk equation, and a multi-layer structure optical model of the printed product is constructed in combination with the scattering and reflection characteristics of the paper type.

[0023] Exemplarily, the paper type and ink type information are obtained from the image recognition database according to the surface characteristics of the printed matter, the printed matter surface is scanned through the grayscale co-occurrence matrix, and the statistics such as contrast, correlation, energy and homogeneity are calculated to obtain the texture feature vector, and the specific type is determined by matching the pre-stored paper and ink sample data using the nearest neighbor algorithm. The detailed spectral characteristic parameters of the paper and ink combination are extracted from the pre-established material spectral characteristic database, and the cubic spline interpolation method is used to interpolate the similar paper and ink combination data in the database, and the complete spectral curve is fitted according to the known spectral parameter points to obtain the spectral parameters of the target combination. According to the extracted spectral characteristic parameters, the ray tracing algorithm is used to simulate the scattering and reflection process of light on the paper surface, and the number of rays is set to 10^6, the incident angle range is 0-90 degrees, the wavelength range is 400-700nm, and the step size is 10nm, and the bidirectional reflection distribution function of the paper surface is calculated. The absorption and reflection characteristics of ink are modeled, the scattering coefficient S and absorption coefficient K of the ink layer are calculated, and the reflectivity R and transmittance T are solved by substituting them into the Kubeka-Munk equation. Combined with the scattering and reflection characteristics of paper, the optical model of the multilayer structure is constructed using the superposition principle to obtain a complete paper and ink combination optical model. In the surface feature analysis of printed products, a high-resolution scanner is used to obtain the surface image of the printed product at a resolution of 600dpi. The grayscale co-occurrence matrix method is used, and the pixel distance is set to 1 to calculate the 0°, 45°, and 90°

[0024] The texture features in four directions of 10° and 135° were obtained. Four statistics, namely contrast, correlation, energy and homogeneity, were calculated for each direction to obtain a 16-dimensional feature vector. Using the KNN algorithm, the K value was set to 5, and the paper and ink types were determined by matching in a pre-established database containing 1000 paper and 500 ink samples. The spectral parameters of the target paper-ink combination were extracted from the material spectral characteristic database, which contains spectral reflectance data with an interval of 5nm in the wavelength range of 400-700nm. The cubic spline interpolation method was used to increase the number of data points from 61 to 301, and a complete spectral curve with an interval of 1nm was obtained. In the ray tracing process, the number of incident rays was set to 10^6, the incident angle was from 0° to 90°, the step size was 1°, the wavelength range was 400-700nm, and the step size was 10nm. The calculated bidirectional reflectance distribution function was stored with an angular resolution of 5°×5°. When modeling the ink layer, the Kubeka-Munk theory is used to calculate the scattering coefficient S and the absorption coefficient K, where the S value range is 0.1-10mm^-1 and the K value range is 0.01-1mm^-1. The calculated bidirectional reflection distribution function of the paper is combined with the reflectivity and transmittance data of the ink layer to construct an optical model of a three-layer structure, with the top layer being the ink layer, the middle layer being the paper surface scattering layer, and the bottom layer being the paper internal scattering layer.

[0025] Step S103, using a pre-trained convolutional neural network model to extract features from the preprocessed spectral data, combining the material spectral characteristic parameters of paper and ink, comprehensively analyzing the modulation effect of the printed surface microstructure on the spectral reflection signal, extracting the key spectral features that characterize the quality of the printed product, and converting the extracted features into feature vectors of fixed dimensions through a feature mapping layer.

[0026] Obtain preprocessed spectral data and material spectral characteristic parameters, and construct an input data matrix based on the spectral data and material parameters; wherein the input data matrix includes normalized spectral data and material parameters. Use the pre-trained convolutional neural network to extract features from the input data matrix; wherein the convolutional neural network includes a convolution layer, a pooling layer, an activation function layer, and a batch normalization layer. Weight the feature map according to the correlation score between the feature map and the material parameter; wherein the correlation score is calculated by a self-attention mechanism. Input the weighted feature map into a fully connected layer to obtain a fixed-dimensional feature vector; wherein the fixed-dimensional feature vector contains modulation information of the spectral reflectance signal by the microstructure of the printed surface. Compare the feature vectors of samples with different surface roughness to extract a feature subset that characterizes the influence of the microstructure; wherein the feature subset is used to characterize the key spectral features of the quality of the printed product.

[0027] Exemplarily, an input data matrix is ​​constructed based on the preprocessed spectral data and material spectral characteristic parameters. After the spectral data and material parameters are normalized, they are spliced ​​into a two-dimensional matrix. The spectral data occupies the first 301 columns of the matrix, and the material parameters occupy the last 2 columns. A pre-trained convolutional neural network is used to extract features from the input data matrix. The network contains 5 convolutional layers and 3 pooling layers. Each convolutional layer uses a 3x3 convolution kernel. The pooling layer uses a 2x2 maximum pooling. Each convolutional layer is followed by a ReLU activation function and a batch normalization layer. The last convolutional layer outputs 128 feature maps. The feature map is obtained through the output of the last convolutional layer of the network. Combined with the material spectral characteristic parameters of paper and ink, the self-attention mechanism is used to calculate the correlation score between the feature map and the material parameters. The feature map is weighted according to the score to highlight the key spectral features that characterize the quality of printed products. The weighted feature map is input into the fully connected layer. Two layers of fully connected layers are used. The first layer outputs a 512-dimensional vector, and the second layer outputs a 256-dimensional vector. The ReLU activation function is used in the middle. The feature is converted into a 256-dimensional fixed-dimensional feature vector through dimensionality reduction operation. The feature vector contains the modulation information of the microstructure of the printed surface on the spectral reflectance signal. By comparing the feature vectors of samples with different surface roughness, the feature subset that characterizes the influence of the microstructure is extracted. In the implementation process, the spectral data in the wavelength range of 400-700nm is first preprocessed, and the reflectivity value is mapped to the range of 0-1 using the minimum-maximum normalization method. At the same time, the paper scattering coefficient and the ink absorption coefficient are normalized to the range of 0-1 respectively. Construct an input data matrix with 303 columns, the first 301 columns are normalized spectral data, and the last 2 columns are normalized material parameters. A pre-trained convolutional neural network was used for feature extraction. The network structure consisted of five convolutional layers, with a convolution kernel size of 3x3 and the number of channels of 32, 64, 128, 128, and 128, respectively; three maximum pooling layers with a size of 2x2; and each convolutional layer was followed by a ReLU activation function and a batch normalization layer. After the input data matrix was processed by the network, 128 feature maps of size 16x16 were obtained. The self-attention mechanism was used to weight the feature maps, and the correlation score between each feature map and the material parameters was calculated, with a score range of 0-1. The score was used as the weight and multiplied with the original feature map to obtain a weighted feature map. The weighted feature map was reduced in dimension through two fully connected layers. The first layer outputted a 512-dimensional vector, and the second layer outputted a 256-dimensional vector. The ReLU activation function was used between the two layers. Finally, a 256-dimensional fixed-dimensional feature vector was obtained. In order to analyze the influence of surface microstructure, five paper samples with different roughness, including roughness Ra values ​​of 0.2μm, 0.5μm, 1.0μm, 2.0μm and 5.0μm, were selected. Their eigenvectors were compared, and the top 20 eigencomponents with the largest fluctuations were extracted as the feature subset characterizing the influence of microstructure.

[0028] Step S104, analyzing the feature vector according to the pre-established printed product quality assessment model, introducing the influencing factors of paper and ink material characteristics, dynamically adjusting the judgment threshold of the quality assessment model, and calculating the scores of various quality indicators of the printed product. If the score of a certain indicator is lower than the adjusted threshold, it is determined that there is a quality defect in the area.

[0029] Receive a multidimensional feature vector, and input the multidimensional feature vector into a pre-established print quality assessment model; process the multidimensional feature vector through three fully connected layers in the print quality assessment model to obtain a preliminary quality index score; obtain material characteristic parameters of current paper and ink in a material database and construct a material characteristic vector; process the material characteristic vector using a multilayer perceptron to obtain an influencing factor weight; multiply the influencing factor weight by the preliminary quality index score element by element to obtain an adjusted quality index score; dynamically segment the historical score distribution of each quality indicator using the OTSU algorithm to calculate an optimal judgment threshold; if the adjusted quality index score is lower than the optimal judgment threshold, it is determined that the area has a quality defect; after processing a preset sample, recalculate the influencing factor weight of the material characteristic and update the judgment threshold of the print quality assessment model.

[0030] Exemplarily, according to the pre-established print quality assessment model, the 256-dimensional feature vector is input into the model, processed through three layers of fully connected layers, with the number of nodes in each layer being 128, 64, and 5 respectively, and the activation function using ReLU, and the preliminary quality index score is calculated, including the scores of color reproduction, layer expression, detail fidelity, color uniformity, and overall effect. The material characteristic parameters of the current paper and ink are obtained from the material database, including the smoothness of the paper 0-100%, the ink absorption 0-10mL / m2, the whiteness 0-100%, and the viscosity of the ink 1-100Pa·s, the drying speed 1-10 levels, and the covering power 0-100%, to construct the material characteristic vector. A three-layer multilayer perceptron is used to process the material characteristic vector, with the number of nodes in the input layer being the same as the number of material characteristic parameters, 50 nodes in the hidden layer, and 5 nodes in the output layer, corresponding to the weights of the influencing factors of the 5 quality indicators, and the influencing factor weights are multiplied element by element with the preliminary quality index score to obtain the adjusted quality index score. According to the adjusted quality index scores, the OTSU algorithm is used to dynamically segment the historical score distribution of each quality index, and the optimal judgment threshold is calculated. If the score of a certain index is lower than the corresponding threshold, it is determined that there is a quality defect in the area, and the defect type and location information are recorded. After processing every 100 samples, the weights of the influencing factors of material properties are recalculated, and the judgment threshold of the quality assessment model is updated to achieve dynamic adjustment. In the actual implementation process, the 256-dimensional feature vector is first input into the quality assessment model. The model consists of a three-layer fully connected neural network with an input dimension of 256-dimensional feature vector, 128 and 64 nodes between hidden layers 1 and 2, respectively, 5 nodes in the output layer, and ReLU activation function. After model processing, 5 preliminary quality index scores are obtained, corresponding to color reproduction, layer expression, detail fidelity, color uniformity and overall effect, with a score range of 0-100. Subsequently, the characteristic parameters of the current paper and ink are extracted from the material database, such as paper smoothness 75%, ink absorbency 5.5mL / m2, whiteness 92%, ink viscosity 35Pa·s, drying speed 7, hiding power 85%, etc., to form a 6-dimensional material characteristic vector. This vector is input into a three-layer perceptron, the input dimension of the input layer is a 6-dimensional feature vector, the number of hidden layer nodes is 50, the number of output layer nodes is 5, and the Tanh activation function is used to output 5 influencing factor weights. These 5 weights are multiplied by the preliminary quality index score to obtain the adjusted quality index score. Then, the OTSU algorithm is used to dynamically segment the most recent 1000 historical scores of each quality index to calculate the optimal threshold. For example, the threshold of color reproduction may be set to 75 points. If the color reproduction score of a certain area is lower than 75 points, it is determined that the area has color quality defects.After processing every 100 new samples, the multi-layer perceptron will be retrained, the influencing factor weights will be updated, and the OTSU threshold will be recalculated. In this way, the quality assessment model can dynamically adjust the judgment criteria according to the different characteristics of paper and ink materials, thereby improving the accuracy and adaptability of quality defect detection.

[0031] Step S105, for the area determined to have quality defects, the spectral data of the area is analyzed, and refined spectral features are extracted. The defect type is accurately classified through a deep learning model, and the influence of paper and ink material properties on the defect formation mechanism is analyzed.

[0032] Perform multi-layer decomposition on the spectral data of the quality defect area, and perform adaptive threshold enhancement on the high-frequency coefficients to obtain enhanced spectral data. Calculate the covariance matrix of the enhanced spectral data, solve the eigenvalues ​​and eigenvectors, and select the top k principal components whose cumulative contribution rate reaches the preset threshold as eigenvectors. Input the eigenvectors into a pre-trained convolutional neural network, which includes multiple convolutional layers and fully connected layers, where the last layer uses a softmax activation function to output the probability distribution of defect types. According to the material characteristic parameters of paper and ink, establish the association rules between material characteristics and defect types, and construct a defect formation mechanism model. Use the output of the defect formation mechanism model as an additional feature, splice it with the eigenvector of the penultimate layer of the convolutional neural network, and obtain the defect classification result combined with the influence of material characteristics through the fully connected layer.

[0033] Exemplarily, the area where quality defects are determined is locally enhanced, the spectral data of the area is decomposed into three layers using Daubechies wavelet, the high-frequency coefficients are adaptively threshold enhanced, and the enhanced spectral data is reconstructed by inverse transformation. More refined spectral features are extracted from the enhanced spectral data, the covariance matrix of the spectral data is calculated, the eigenvalues ​​and eigenvectors are solved, and the first k principal components with a cumulative contribution rate of 95% are selected as eigenvectors. The extracted refined spectral features are input into a pre-trained convolutional neural network, which contains 5 convolutional layers and 3 fully connected layers. The convolution kernel size of the convolutional layer is 3x3, the number of channels is 32, 64, 128, 256, and 512 respectively, the number of nodes in the fully connected layer is 1024, 512, and the number of defect categories respectively. The last layer uses a softmax activation function to output the probability distribution of different defect types. According to the material characteristic parameters of paper and ink, the decision tree algorithm is used to establish the association rules between material characteristics and common defect types, and the defect formation mechanism model is constructed. The output of the model is used as an additional feature and concatenated with the feature vector of the penultimate layer of the convolutional neural network. Then, a fully connected layer is used to obtain the final defect classification result considering the influence of material characteristics. In the actual implementation process, the area where quality defects are determined to exist is first locally enhanced. The spectral data of the area is decomposed into 3 layers using Daubechies-4 wavelet to obtain 7 sub-bands. For each high-frequency sub-band, its standard deviation σ is calculated, and the adaptive threshold T = 3σ is set. The coefficients greater than T are multiplied by 1.5 for enhancement, and the coefficients less than T remain unchanged. The enhanced spectral data is reconstructed by inverse wavelet transform. Subsequently, the 301×301 covariance matrix of the enhanced spectral data is calculated, and its eigenvalues ​​and eigenvectors are solved. The first 20 principal components with a cumulative contribution rate of 95% are selected as eigenvectors, and the original 301-dimensional spectral data is reduced to 20 dimensions. These 20-dimensional eigenvectors are used as input and sent to the pre-trained convolutional neural network. The network contains 5 convolutional layers, with the convolution kernel size of 3×3 and the number of channels of 32, 64, 128, 256, and 512, respectively. Each layer is followed by a ReLU activation function and batch normalization. Then there are 3 fully connected layers with the number of nodes of 1024, 512, and 10, respectively. Assuming there are 10 types of defects, the last layer uses a softmax activation function to output the probability of the defect type. At the same time, using the C4.5 decision tree algorithm, with paper smoothness, ink absorption, whiteness, ink viscosity, drying speed, and hiding power as features, and defect types as labels, a decision tree with a maximum depth of 5 is trained as a defect formation mechanism model. The 10-dimensional output probability vector of the model is concatenated with the 512-dimensional feature vector of the penultimate layer of the convolutional neural network to obtain a 522-dimensional vector. Finally, through a fully connected layer with an output of 10, the final defect classification result that integrates the influence of material properties is obtained.

[0034] Step S106, based on the defect type and location information, similar cases are retrieved from a pre-established defect knowledge base to obtain defect cause analysis results and improvement suggestions, and the material spectral characteristic parameters are matched with the defect characteristics to obtain an evaluation result of the degree of influence of material characteristics on defect formation.

[0035] According to the defect type, a one-hot encoding vector is obtained, and the position information is normalized to obtain a two-dimensional coordinate in the range of 0 to 1. The query vector is formed by splicing the one-hot encoding vector and the two-dimensional coordinate. A vector space model based on cosine similarity is used to retrieve similar cases from a pre-established defect knowledge base, and the top N cases with the highest similarity are selected as reference samples. Text mining is performed on the reference samples to extract defect cause keywords and improvement suggestion keywords, and the frequency of each word in the document and the inverse document frequency in the entire document set are calculated; an n-dimensional material spectral characteristic parameter vector and an m-dimensional defect feature vector are obtained, and an n+m-dimensional vector is formed by direct splicing to construct an input data matrix. The Apriori algorithm is used to mine the association rules between material characteristics and defect characteristics. If the support is greater than the first threshold and the confidence is greater than the second threshold, it is determined to be a valid association rule. According to the association rule, the influence of material characteristics on defect formation is calculated, and the confidence of the association rule is used as the initial weight. The weight is optimized by fitting historical data by the least squares method to obtain the evaluation result of the influence of material characteristics on defect formation.

[0036] Exemplarily, according to the defect type and location information, the defect type is encoded as a one-hot vector, the location information is normalized to a two-dimensional coordinate in the range of 0-1, and the query vector is formed by splicing. The vector space model based on cosine similarity is used to retrieve similar cases from the pre-established defect knowledge base, and the top 5 cases with the highest similarity are selected as reference samples. Text mining is performed on the selected reference samples to extract defect cause keywords and improvement suggestion keywords, and the frequency TF of each word in the document and the inverse document frequency IDF in the entire document set are calculated. The two are multiplied to obtain the TF-IDF value. According to the keywords sorted by the TF-IDF value, a predefined template is used to generate defect cause analysis sentences and improvement suggestion summaries. The n-dimensional material spectral characteristic parameter vector and the m-dimensional defect feature vector are directly spliced ​​into an n+m-dimensional vector, and the input data matrix is ​​constructed. The Apriori algorithm is used to mine the association rules between material characteristics and defect characteristics, and the minimum support is set to 0.1 and the minimum confidence is set to 0.5. According to the association rules obtained by mining, the influence of material characteristics on defect formation is calculated. The confidence of the association rules is used as the initial weight. The weight is optimized by fitting historical data through the least squares method. The weighted summation method is used to comprehensively evaluate the influence weights of various material characteristics, and the evaluation results of the influence of material characteristics on defect formation are obtained. In the actual implementation process, the defect types, such as color difference, ink spots, ghosting, etc., are first encoded into 10-dimensional one-hot vectors, and the position information, such as x=0.3, y=0.7, is normalized to two-dimensional coordinates in the range of 0-1, and spliced ​​to form a 12-dimensional query vector. The cosine similarity is used to calculate the similarity between the query vector and the 1,000 historical cases in the knowledge base, and the top 5 cases with the highest similarity are selected. The text content of these 5 cases is segmented and the TF-IDF value is calculated. For example, for the word "ink viscosity is too high", assuming that it appears 3 times in the current document, the total number of words in the document is 100, and the TF is 3 / 100=0.03; the word appears 4 times in 5 documents, and the IDF is

[0037] log(5 / 4)=0.09, then the TF-IDF value is 0.03*0.09=0.0027. According to the TF-IDF value sorting, the first 10 keywords are selected and substituted into the predefined template such as "Due to {keyword 1} and {keyword 2}, {defect type} is generated" to generate defect cause analysis. The 20-dimensional material spectral characteristic parameter vector, such as reflectivity and scattering coefficient, and the 30-dimensional defect feature vector, such as area, shape, and color, are spliced ​​into a 50-dimensional vector. The Apriori algorithm is used to mine association rules on 100 samples to obtain rules such as "high ink viscosity => ghosting" with a confidence level of 0.8. Using the confidence level of 0.8 as the initial weight, 1000 historical samples are fitted by the least squares method to obtain the optimized weight of 0.75. Finally, the weighted sum of the 20 material characteristic parameters is performed to obtain a score for the comprehensive influence of material characteristics on defect formation, such as 85 points, with a full score of 100 points.

[0038] Step S107, based on the impact assessment results, optimize and adjust the defect cause analysis and improvement suggestions in the quality report, including providing targeted process optimization solutions based on the paper ink material properties, and generating a refined quality report containing defect detailed information and material improvement suggestions.

[0039] Obtain the percentage value of the degree of influence of material characteristics, determine the weight according to the percentage value, and use the weighted summation method to calculate the comprehensive score of each defect cause. If the comprehensive score ranking reaches the preset ranking, it is determined to be the main cause. According to the main cause, retrieve the process optimization plan from the process parameter database. Use cosine similarity to calculate the similarity between the process parameter vector and the sample in the database. If the similarity ranking reaches the preset ranking, it is determined to be a process parameter adjustment suggestion. Use the C4.5 algorithm to build a decision tree, with the information gain ratio as the splitting criterion. Obtain the performance evaluation result of the decision tree through 10-fold cross validation. Obtain the current material parameters, input the decision tree, and obtain material improvement suggestions. The material improvement suggestions include paper selection solutions and ink formula adjustment solutions. Generate a structured and detailed quality report based on defect detailed information, main cause analysis results, process optimization solutions and material improvement suggestions. The quality report uses JSON format and contains defect type, location, area, cause analysis, process suggestion and material suggestion fields.

[0040] Exemplarily, according to the impact assessment results, the percentage value of the material property impact is directly set as the weight, such as an impact of 85% corresponding to a weight of 0.85, and the defect cause analysis in the quality report is prioritized, and the weighted sum method is used to calculate the comprehensive score of each cause, and the top three with the highest scores are retained as the main causes. Based on the main causes and the impact of material characteristics, the relevant process optimization solutions are retrieved from the process parameter database, and the cosine similarity is used to calculate the similarity between the process parameter vector and the samples in the database. The top three solutions with the highest similarity are selected as process parameter adjustment suggestions to generate targeted process optimization solutions. Combined with the material characteristic parameters of paper and ink, the C4.5 algorithm is used to construct a decision tree, with the information gain ratio as the splitting criterion, and a 10-fold cross validation is used to evaluate the model performance. The current material parameters are input, and specific material improvement suggestions are output, including paper selection and ink formula adjustment solutions. Integrate defect details, main cause analysis, process optimization plan and material improvement suggestions, use template-based method to generate structured and refined quality report, predefine report structure template, fill each part into corresponding position, use synonym replacement and sentence transformation to increase language diversity, report adopts JSON format, contains defect type, location, area, cause analysis, process suggestion and material suggestion and other fields, convenient for subsequent processing and display. In the actual implementation process, first convert the influence degree of material characteristics 85% into weight 0.85, and perform weighted calculation on 5 defect causes. For example, if cause A scores 90 and the weight is 0.85, the weighted score is 76.5. Select the top three high-scoring factors, such as "ink viscosity is too high", "paper ink absorption is insufficient" and "printing pressure is low". Then extract the parameter vectors related to these three causes from the process parameter database containing 1000 records, such as [ink viscosity, paper ink absorption, printing pressure] = [35Pa·s, 5mL / m2, 0.5MPa]. The cosine similarity was used to calculate the similarity between the vector and the samples in the database, and the top three solutions with the highest similarity were selected, such as "lowering the ink viscosity to 25 Pa·s", "selecting paper with an ink absorption of 8 mL / m2", and "adjusting the printing pressure to 0.7 MPa". Then, the C4.5 algorithm was used to construct a decision tree, and the current material parameters [paper smoothness = 75%, ink drying speed = level 7] were input. After the decision tree classification, the output material improvement suggestion was "selecting paper with a smoothness of 85%, and adjusting the ink drying speed to level 5".Finally, a quality report is generated using a predefined JSON template, such as {"Defect type": "Ghosting", "Position": [0.3, 0.7], "Area": ​​5cm2, "Cause analysis": ["Ink viscosity is too high", "Paper ink absorption is insufficient", "Printing pressure is too low"], "Process suggestions": ["Reduce ink viscosity to 25Pa·s", "Select paper with ink absorption of 8mL / m2", "Adjust printing pressure to 0.7MPa"], "Material suggestions": "Select paper with a smoothness of 85%, and adjust the ink drying speed to level 5"}, to form a complete and refined quality report.

[0041] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for detecting product surface printing quality using industrial vision, characterized in that: The method comprises: Performing a multi-band spectral scan on the printed matter according to preset spectral acquisition parameters to acquire spectral data, using a spectral camera array to synchronously acquire spectral images of multiple bands, and performing real-time preprocessing on the acquired spectral data through a parallel processing unit to obtain preprocessed spectral data; Acquire information about the type of paper and ink used in the printed matter, and extract material spectrum characteristic parameters of the paper-ink combination from a pre-established material spectrum characteristic database, including the scattering reflection characteristics of the paper surface and the absorption reflection characteristics of the ink; The pre-trained convolutional neural network model is used to extract features from the pre-processed spectral data. The material spectral characteristic parameters of paper and ink are combined to comprehensively analyze the modulation effect of the printed surface microstructure on the spectral reflection signal, extract the key spectral features that characterize the quality of the printed product, and convert the extracted features into feature vectors of fixed dimensions through the feature mapping layer. The characteristic vector is analyzed according to the pre-established printed product quality assessment model, the influencing factors of paper and ink material characteristics are introduced, the judgment threshold of the quality assessment model is dynamically adjusted, and the scores of various quality indicators of the printed product are calculated. If the score of a certain indicator is lower than the adjusted threshold, it is determined that there is a quality defect in the area. For areas judged to have quality defects, the spectral data of the area is analyzed, and refined spectral features are extracted. The defect types are accurately classified through deep learning models, and the influence of paper and ink material properties on the defect formation mechanism is analyzed; According to the defect type and location information, similar cases are retrieved from the pre-established defect knowledge base to obtain defect cause analysis results and improvement suggestions, and the material spectral characteristic parameters are matched with the defect characteristics to obtain the evaluation results of the influence of material characteristics on defect formation; Based on the impact assessment results, the defect cause analysis and improvement suggestions in the quality report are optimized and adjusted, including providing targeted process optimization solutions based on the characteristics of paper and ink materials, and generating a refined quality report containing detailed defect information and material improvement suggestions.

2. The method according to claim 1, wherein: The method includes performing multi-band spectral scanning on a printed product according to preset spectral acquisition parameters to acquire spectral data, using a spectral camera array to synchronously acquire spectral images of multiple bands, and performing real-time preprocessing on the acquired spectral data by a parallel processing unit to obtain preprocessed spectral data, including: A high-speed spectral camera array is used to synchronously collect spectral images of multiple bands, and the spectral image of each band is processed by the grating spectrometry method to obtain spectral curve data; The spectral curve data is normalized, and the normalized spectral curve data is denoised using Daubechies wavelet to obtain denoised spectral data; The denoised spectral data are grouped by cluster analysis to establish feature templates for different types of printed matter and construct a spectral feature library; A support vector machine is used to train the data in the spectral feature library, and a kernel function is set as a radial basis function to obtain a classification model; The spectral feature vector of the unknown printed matter is classified and identified according to the classification model to obtain the printed matter category.

3. The method according to claim 1, wherein: The method of obtaining information on the type of paper and ink used in the printed product and extracting material spectrum characteristic parameters of the paper-ink combination from a pre-established material spectrum characteristic database, including the scattering reflection characteristics of the paper surface and the absorption reflection characteristics of the ink, includes: Obtaining a gray level co-occurrence matrix of a printed product surface, calculating contrast, correlation, energy and homogeneity according to the gray level co-occurrence matrix, and obtaining a texture feature vector of the printed product surface; Matching the texture feature vector with pre-stored sample data using a nearest neighbor algorithm to determine the paper type and ink type of the printed product; According to the paper type and ink type, the spectral characteristic parameters of similar paper and ink combinations are obtained from the pre-established material spectral characteristic database, and the spectral characteristic parameters are interpolated and calculated using the cubic spline interpolation method to obtain the complete spectral curve of the printed product; According to the complete spectrum curve, a ray tracing algorithm is used to simulate the scattering and reflection process of light on the surface of the printed matter, and a bidirectional reflection distribution function of the surface of the printed matter is calculated; The scattering coefficient and absorption coefficient of the ink type are calculated by the Kubeka-Munk equation, and a multi-layer structure optical model of the printed product is constructed in combination with the scattering and reflection characteristics of the paper type.

4. The method according to claim 1, wherein: The method uses a pre-trained convolutional neural network model to extract features from the pre-processed spectral data, combines the material spectral characteristic parameters of paper and ink, comprehensively analyzes the modulation effect of the printed surface microstructure on the spectral reflection signal, extracts the key spectral features that characterize the quality of the printed product, and converts the extracted features into feature vectors of fixed dimensions through a feature mapping layer, including: Acquire preprocessed spectral data and material spectral characteristic parameters, and construct an input data matrix according to the spectral data and material parameters; Wherein, the input data matrix includes normalized spectral data and material parameters; Using the pre-trained convolutional neural network to perform feature extraction on the input data matrix; Wherein, the convolutional neural network includes a convolution layer, a pooling layer, an activation function layer and a batch normalization layer; weighting the feature map according to a correlation score between the feature map and a material parameter; Wherein, the relevance score is calculated by a self-attention mechanism; Input the weighted feature map into the fully connected layer to obtain a fixed-dimensional feature vector; Wherein, the fixed-dimensional feature vector contains modulation information of the printed product surface microstructure on the spectral reflection signal; Compare the feature vectors of samples with different surface roughness and extract the feature subset that characterizes the microstructural influence; The feature subset is used to characterize key spectral features of printed product quality.

5. The method according to claim 1, wherein: The characteristic vector is analyzed according to the pre-established printed product quality assessment model, the influencing factors of paper and ink material characteristics are introduced, the judgment threshold of the quality assessment model is dynamically adjusted, and the scores of various quality indicators of the printed product are calculated. If the score of a certain indicator is lower than the adjusted threshold, it is determined that the area has quality defects, including: receiving a multidimensional feature vector, and inputting the multidimensional feature vector into a pre-established print quality assessment model; Processing the multi-dimensional feature vector through three fully connected layers in the print quality assessment model to obtain a preliminary quality index score; Obtain the material characteristic parameters of the current paper and ink in the material database and construct the material characteristic vector; A multi-layer perceptron is used to process the material characteristic vector to obtain the weight of the influencing factor; Multiplying the impact factor weight by the preliminary quality indicator score element by element to obtain an adjusted quality indicator score; The OTSU algorithm is used to dynamically segment the historical score distribution of each quality indicator and calculate the optimal judgment threshold; If the adjusted quality index score is lower than the optimal judgment threshold, it is determined that the area has quality defects; After processing the preset samples, the weights of the influencing factors of the material properties are recalculated, and the judgment threshold of the printed product quality assessment model is updated.

6. The method according to claim 1, wherein: For the area determined to have quality defects, the spectral data of the area is analyzed, refined spectral features are extracted, the defect type is accurately classified through a deep learning model, and the influence of paper and ink material properties on the defect formation mechanism is analyzed, including: Perform multi-layer decomposition on the spectral data of the quality defect area, and perform adaptive threshold enhancement on the high-frequency coefficients to obtain enhanced spectral data; Calculate the covariance matrix of the enhanced spectral data, solve the eigenvalues ​​and eigenvectors, and select the first k principal components whose cumulative contribution rate reaches a preset threshold as eigenvectors; Input the feature vector into a pre-trained convolutional neural network, wherein the convolutional neural network comprises a plurality of convolutional layers and fully connected layers, wherein the last layer uses a softmax activation function to output a probability distribution of defect types; According to the material characteristic parameters of paper and ink, the association rules between material characteristics and defect types are established, and the defect formation mechanism model is constructed; The output of the defect formation mechanism model is used as an additional feature and concatenated with the feature vector of the penultimate layer of the convolutional neural network, and a defect classification result combined with the influence of material properties is obtained through a fully connected layer.

7. The method according to claim 1, wherein: According to the defect type and location information, similar cases are retrieved from the pre-established defect knowledge base to obtain defect cause analysis results and improvement suggestions, and the material spectral characteristic parameters are matched with the defect characteristics to obtain the impact of material characteristics on defect formation. The evaluation results include: Obtain a one-hot encoding vector according to the defect type, normalize the position information to obtain a two-dimensional coordinate in the range of 0 to 1, and form a query vector by concatenating the one-hot encoding vector and the two-dimensional coordinate; A vector space model based on cosine similarity is used to retrieve similar cases from the pre-established defect knowledge base, and the top N cases with the highest similarity are selected as reference samples; Performing text mining on the reference sample, extracting defect cause keywords and improvement suggestion keywords, and calculating the frequency of each word in the document and the inverse document frequency in the entire document set; Obtain n-dimensional material spectral characteristic parameter vector and m-dimensional defect feature vector, form n+m-dimensional vector by direct splicing, and construct input data matrix; The Apriori algorithm is used to mine the association rules between material properties and defect features. If the support is greater than the first threshold and the confidence is greater than the second threshold, it is determined to be a valid association rule; The influence of material characteristics on defect formation is calculated according to the association rules, the confidence of the association rules is used as the initial weight, the weight is optimized by fitting historical data through the least squares method, and the evaluation result of the influence of material characteristics on defect formation is obtained.

8. The method according to claim 1, wherein: According to the impact assessment results, the defect cause analysis and improvement suggestions in the quality report are optimized and adjusted, including providing targeted process optimization solutions based on the characteristics of paper and ink materials, and generating a refined quality report containing detailed defect information and material improvement suggestions, including: Obtaining a percentage value of the degree of influence of material characteristics, determining a weight according to the percentage value, and calculating a comprehensive score of each defect cause by a weighted summation method; If the comprehensive score ranking reaches the preset ranking, it is determined to be the main cause; According to the main causes, a process optimization solution is retrieved from a process parameter database; The cosine similarity is used to calculate the similarity between the process parameter vector and the samples in the database; If the similarity ranking reaches the preset ranking, it is determined as a process parameter adjustment suggestion; The C4.5 algorithm was used to construct the decision tree, with the information gain ratio as the splitting criterion; The performance evaluation results of the decision tree were obtained through 10-fold cross validation; Obtain current material parameters, input the decision tree, and obtain material improvement suggestions; The material improvement suggestions include paper selection solutions and ink formulation adjustment solutions; Generate structured and detailed quality reports based on defect details, main cause analysis results, process optimization solutions and material improvement suggestions; The quality report is in JSON format and includes fields such as defect type, location, area, cause analysis, process recommendation, and material recommendation.

Images