An intelligent coated paper coating defect identification method and system

Abstract

The application discloses a kind of intelligent coating paper coating defect identification method and system, method includes multidimensional data acquisition, data optimization processing, moire difference amplification, identification model construction and defect identification output.The original data is obtained by data acquisition;Adopt the data optimization processing method of image denoising filter, ROI cutting, size normalization, gray standardization and data set division;Ideal defect-free template is constructed and is carried out optical interference with the image to be measured sample, the difference amplification moire image is obtained, the abstract digital difference is converted into visual texture interference pattern, the lower limit and stability of defect identification are improved;Dual-path deep network model is used as identification model, by combining local and global feature extraction and introducing orthogonal decoupling attention mechanism, feature space is decomposed into defect main signal and mixed factor signal, can be stripped out from complex surface anomaly mode Pure representation belonging to coating defect.

Description

Technical Field

[0001] This invention relates to the field of intelligent identification technology for coating defects in coated paper, specifically to an intelligent method and system for identifying coating defects in coated paper. Background Technology

[0002] Coated paper coating defect identification is a technology that identifies and classifies coating defects by analyzing the image features of the coated paper surface. It utilizes computer vision and deep learning algorithms to process acquired images of the coated paper surface and extract key indicators to determine coating quality and defect types. This enables non-contact, real-time online detection, helping manufacturers promptly eliminate substandard products and improve product quality.

[0003] However, traditional methods for identifying coating defects in coated paper suffer from technical problems. They rely directly on original grayscale images or manually designed texture features for feature extraction and modeling. The image variation caused by coating defects is often small and easily obscured by the inherent fiber texture of the paper itself, coating unevenness, and changes in illumination, making the extraction of key defect signals extremely difficult. Furthermore, traditional methods for identifying coating defects in coated paper lack explicit decoupling designs for confounding factors. Surface anomalies in coated paper are complex manifestations of multiple coexisting factors. Confounding factors such as paper type, coating process parameters, and ambient illumination can become entangled with the real defect signals in the feature space, often resulting in identification results mixed with interference from non-defect factors. Summary of the Invention

[0004] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an intelligent method and system for identifying coating defects in coated paper. Traditional methods for identifying coating defects in coated paper suffer from the technical problem of directly extracting and modeling features based on original grayscale images or manually designed texture features. Image variations caused by coating defects are often small and easily masked by the inherent fiber texture of the paper, coating unevenness, and changes in illumination, making the extraction of key defect signals extremely difficult. This solution creatively employs a method of constructing an ideal defect-free template and performing optical interference with the image of the sample to be tested to obtain a difference-magnified moiré pattern. This transforms abstract digital differences into visual texture interference patterns, enabling the identification of coating defect signals. Significantly enhanced in spatial dimension, improving the lower limit and stability of defect identification; addressing the technical problem of traditional coated paper coating defect identification methods lacking explicit decoupling design for confounding factors. Coated paper surface anomalies are complex manifestations of multiple coexisting factors. Confounding factors such as paper type, coating process parameters, and ambient lighting can become entangled with the real defect signal in the feature space, and the identification results are often mixed with interference from non-defect factors. This solution creatively adopts a dual-path deep network model as the identification model. By combining local and global feature extraction and introducing an orthogonal decoupling attention mechanism, the feature space is decomposed into the main defect signal and confounding factor signals, which can extract the pure characterization belonging to the coating defect from the complex surface anomaly pattern.

[0005] The technical solution adopted by this invention is as follows: This invention provides an intelligent method for identifying coating defects in coated paper, which includes the following steps:

[0006] Step S1: Multidimensional data acquisition;

[0007] Step S2: Data optimization processing;

[0008] Step S3: Magnify the difference in moiré patterns;

[0009] Step S4: Recognition model construction;

[0010] Step S5: Defect identification output.

[0011] Further, in step S1, the multidimensional data acquisition is used to provide raw multidimensional data for the training and application of the subsequent model. Specifically, it involves obtaining a raw defect identification dataset through data collection. The raw defect identification dataset includes a historical raw defect identification dataset and a current raw defect identification dataset, both of which contain sample image data and alignment information data. The historical raw defect identification dataset also contains defect-free baseline image data, defect type annotation data, and confounding factor labels.

[0012] Further, in step S2, the data optimization processing is used to optimize the collected raw multidimensional data, transforming the raw multidimensional data into a regular and model-readable format, specifically including the following steps:

[0013] Step S21: Image denoising filtering, used to eliminate random noise introduced by the camera sensor and lighting, specifically by smoothing the grayscale image of the coated paper surface through Gaussian filtering to obtain a denoised image;

[0014] Step S22: ROI cropping, used to extract the effective detection area and remove useless background edges. Specifically, the coated paper area is determined by the edge detection method, and the region of interest containing only the surface of the coated paper is cropped to obtain the cropped image.

[0015] Step S23: Size normalization, used to eliminate size differences caused by different batches and different camera resolutions. Specifically, all cropped images are scaled to a fixed width and height pixel size through bilinear interpolation to obtain images with uniform size.

[0016] Step S24: Gray-scale normalization is used to eliminate the overall brightness deviation caused by changes in illumination intensity. Specifically, it involves processing the pixel values ​​of each image through histogram matching and Z-score normalization to obtain a normalized image.

[0017] Step S25: Dataset partitioning, used to obtain training data and validation data, specifically, the original historical defect identification dataset is partitioned to obtain a preliminary training set and a preliminary test set;

[0018] The current defect identification raw dataset is optimized by image denoising filtering, ROI cropping, size normalization, and grayscale standardization to obtain a preliminary dataset to be identified. The historical defect identification raw dataset is optimized by image denoising filtering, ROI cropping, size normalization, grayscale standardization, and dataset partitioning to obtain a preliminary training set and a preliminary test set.

[0019] Further, in step S3, the moiré pattern difference amplification is used to enhance minute coating defect differences into significant visual texture patterns. Specifically, it is achieved by constructing an ideal defect-free template and performing optical interference with the image of the sample to be tested to obtain a difference-amplified moiré pattern image.

[0020] The amplification of moiré pattern differences specifically includes the following steps:

[0021] Step S31: Defect-free benchmark raster animation generation, used to create a dynamic visual template representing the surface of the defect-free standard coated paper as a benchmark for difference comparison, includes the following steps:

[0022] Step S311: Calculate the group average defect-free image features to eliminate individual differences and define a common defect-free standard. Specifically, by statistically calculating the mean and standard deviation of grayscale at each pixel position of all defect-free benchmark image samples, the ideal defect-free template image and its normal fluctuation image are obtained.

[0023] Step S312: Generate multi-scale Gaussian dot matrix parameters to encode the speed information of grayscale changes in visual representation to enhance dynamic expression capabilities. Specifically, by calculating the gradient magnitude of each pixel on the ideal defect-free template image and dynamically adjusting the size and transparency of the rendering spot accordingly, dot matrix rendering parameters reflecting the speed of local changes are obtained.

[0024] Step S313: Render standard defect-free dot matrix animation, which is used to transform the digitized ideal defect-free template image into a dynamic visual template. Specifically, each pixel position on the plane of the ideal defect-free template image is mapped to a point on the canvas, and a Gaussian spot is drawn based on its coordinates and gray values ​​to obtain a standard defect-free dot matrix image sequence as the ideal defect-free template.

[0025] Step S32: Generating dot matrix animation of the sample to be tested, which is used to process the image of the sample to be tested with visual rules that are completely consistent with the ideal defect-free template. Specifically, the same mapping relationship as the ideal defect-free template is used, and based on the same steps, the image of the sample to be tested is rendered into a dot matrix image sequence to obtain the template of the sample to be tested.

[0026] Step S33: Dynamic moiré pattern synthesis, used to generate a core, amplified dynamic texture pattern, includes the following steps:

[0027] Step S331: Calculate the pixel-by-pixel difference intensity heatmap to quantify and locate the degree of defect anomaly at each location. Specifically, the difference intensity heatmap is obtained by calculating the absolute difference between the image of the sample to be tested and the ideal defect-free template image at each pixel location and normalizing it.

[0028] Step S332: Adaptive overlay blending, used to generate a moiré pattern with visual focus concentrated in the difference region. Specifically, it involves overlaying an ideal defect-free template with a template of the sample to be tested under a preset spatial offset, and using the calculated difference intensity heat map as a weight map to adaptively control the blending ratio of each pixel to obtain a synthetic moiré image.

[0029] Step S34: Standardized image generation, used to generate standardized image data suitable for the input of the deep learning model. Specifically, the synthesized moiré image is subjected to size scaling, grayscale normalization and contrast-limited adaptive histogram equalization to obtain a difference-enlarged moiré image with uniform size and contrast.

[0030] The preliminary dataset to be identified, the preliminary training set, and the preliminary test set are used as inputs to the moiré pattern difference amplification step. Defect-free reference dot matrix animation generation, test sample dot matrix animation generation, dynamic moiré pattern synthesis, and standardized image generation are performed sequentially to obtain the dataset to be identified, the difference amplification training set, and the difference amplification test set.

[0031] Further, in step S4, the identification model construction is used to construct the model required to decouple pure defect-related features from the moiré image and perform defect identification. Specifically, it is to construct a dual-path deep network model as the identification model. The dual-path deep network model outputs a defect confidence index, defect type label and defect region heatmap by combining local and global feature extraction and introducing an orthogonal decoupling attention mechanism.

[0032] The construction of the recognition model specifically includes the following steps:

[0033] Step S41: Backbone feature extraction, used to extract feature representations rich in spatial texture information from the moiré image, including the following steps:

[0034] Step S411: Input preprocessing, used to normalize the model input for stable training, specifically by performing channel-level Z-score normalization on the differential magnified moiré image of the input to obtain the model input tensor;

[0035] Step S412: Local texture path design, used to capture the local details and spatial variations of the moiré pattern. Specifically, the input tensor of the model is processed through a residual convolutional network. After four stages of downsampling and residual blocks, a global average pooling layer is used to output the local pattern features.

[0036] Step S413: Global structural path design, used to model the overall layout and long-distance dependencies of the moiré pattern. Specifically, the model input tensor is segmented into non-overlapping image blocks and transformed into a token sequence through linear projection. The token sequence is then input into a Transformer encoder for processing, global average pooling is used, and the output is the global structural features.

[0037] Step S414: Dual-path feature fusion, used to comprehensively utilize local details and global context information. Specifically, the features extracted from the two paths are concatenated and fused and dimensionality reduced through a fully connected layer to obtain comprehensive backbone features.

[0038] Step S42: Orthogonal decoupling attention design, used to explicitly separate the main defect signal from confounding factor signals in the feature space, including the following steps:

[0039] Step S421: Construction of learnable orthogonal direction basis, used to define a set of mutually independent semantic directions in the feature space. Specifically, by initializing and constraining the column vectors of a learnable matrix to maintain unit orthogonality, an orthogonal basis for eigenvalue decomposition is obtained. The first basis vector of the orthogonal basis is specified as the defect principal direction vector.

[0040] Step S422: Orthogonal projection and reconstruction, used to decompose the comprehensive backbone features into different orthogonal directions. Specifically, the comprehensive backbone features are projected onto the orthogonal basis to obtain projection coefficients, and the defect pure features are reconstructed separately based on the projection coefficients of the defect principal direction. At the same time, the difference between the reconstruction and the total features is calculated to obtain the projection coefficient vector, defect pure features and mixed difference features.

[0041] Step S423: Adversarial training of multi-head orthogonal discriminators, used to force the heterogeneous features not to contain information related to defects. Specifically, multiple discriminator heads are trained to try to predict heterogeneous factor labels from heterogeneous features, and gradient inversion is applied for adversarial training, forcing the network to encode heterogeneous factors into a subspace orthogonal to the main direction of defects.

[0042] Step S43: Model classification prediction, used to map the defect-pure features to the required output results. Specifically, it outputs the defect confidence index by performing a learnable affine transformation on the projection coefficients obtained based on the defect principal direction, and outputs the defect type label by processing the defect-pure features with a softmax classification head.

[0043] Step S44: Interpretable attention generation, which provides visual clues for model decision-making. Specifically, the defect confidence index is used as the query to perform attention weighting on local pattern features and restores the original input size through bilinear interpolation. Finally, a defect area heatmap with highlighted abnormal areas related to defects is obtained.

[0044] Step S45: Construct and train the model. Specifically, the dual-path deep network model is constructed by integrating the backbone feature extraction, the orthogonal decoupled attention design, the model classification prediction, and the interpretable attention generation. The model is trained and its performance is verified based on the difference amplification training set and the difference amplification test set to obtain the dual-path deep network model as the recognition model.

[0045] Further, in step S5, the defect identification output specifically involves using the dataset to be identified as input to the identification model to identify coating defects, and the defect confidence index, defect type label, and defect area heat map output by the identification model as the final defect identification result.

[0046] The present invention provides an intelligent coating paper coating defect identification system, including a multi-dimensional data acquisition module, a data optimization and processing module, a moiré pattern difference amplification module, an identification model construction module, and a defect identification output module;

[0047] The multidimensional data acquisition module is used to collect raw data, obtain a defect identification raw dataset by collecting raw data, and send the defect identification raw dataset to the data optimization processing module.

[0048] The data optimization processing module is used for data optimization processing. Through data optimization processing, a preliminary dataset to be identified, a preliminary training set, and a preliminary test set are obtained, and the preliminary dataset to be identified, the preliminary training set, and the preliminary test set are sent to the moiré pattern difference amplification module.

[0049] The moiré pattern difference amplification module is used for moiré pattern difference amplification. It constructs an ideal defect-free template and performs optical interference with the image of the sample to be tested. It processes the preliminary dataset to be identified, the preliminary training set, and the preliminary test set to obtain the dataset to be identified, the difference amplification training set, and the difference amplification test set. The dataset to be identified is sent to the defect identification output module, and the difference amplification training set and the difference amplification test set are sent to the identification model construction module.

[0050] The identification model construction module is used to construct the identification model. By constructing a dual-path deep network model, the identification model is obtained and then sent to the defect identification output module.

[0051] The defect identification output module is used for defect identification output. By using the identification model, it identifies coating defects in coated paper based on the dataset to be identified, and obtains the final defect identification result.

[0052] The beneficial effects achieved by the present invention using the above solution are as follows:

[0053] (1) Traditional methods for identifying coating defects in coated paper have the technical problem of directly extracting and modeling features based on the original grayscale image or manually designed texture features. The image variation caused by coating defects is often small and easily overwhelmed by the inherent fiber texture, coating unevenness and light changes of the paper itself, making it extremely difficult to extract key defect signals. This solution creatively adopts the method of constructing an ideal defect-free template and performing optical interference with the image of the sample to be tested to obtain a difference magnified moiré pattern. This transforms the abstract digital difference into a visual texture interference pattern, which can significantly enhance the coating defect signal in the spatial dimension and improve the lower limit and stability of defect identification.

[0054] (2) In view of the technical problem that traditional coating paper coating defect identification methods lack explicit decoupling design for mixed factors, the surface anomalies of coated paper are complex manifestations of multiple factors coexisting. Mixed factors such as paper type, coating process parameters, and ambient light will become entangled with the real defect signal in the feature space, and the identification results are often mixed with interference from non-defect factors. This scheme creatively adopts a dual-path deep network model as the identification model. By combining local and global feature extraction and introducing an orthogonal decoupling attention mechanism, the feature space is decomposed into the main defect signal and mixed factor signal, which can peel off the pure characterization belonging to the coating defect from the complex surface anomaly pattern. Attached Figure Description

[0055] Figure 1 A flowchart illustrating an intelligent method for identifying coating defects in coated paper provided by this invention;

[0056] Figure 2 A schematic diagram of a module for an intelligent coating paper coating defect identification system provided by the present invention;

[0057] Figure 3 A flowchart illustrating the data optimization process in step S2;

[0058] Figure 4 This is a flowchart illustrating the process of amplifying the moiré pattern difference in step S3.

[0059] Figure 5 A flowchart illustrating the process of model construction for step S4.

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0061] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0062] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0063] Example 1, see Figure 1 The present invention provides an intelligent method for identifying coating defects in coated paper, the method comprising the following steps:

[0064] Step S1: Multidimensional data acquisition;

[0065] Step S2: Data optimization processing;

[0066] Step S3: Magnify the difference in moiré patterns;

[0067] Step S4: Recognition model construction;

[0068] Step S5: Defect identification output.

[0069] Example 2, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In step S1, the multidimensional data acquisition is used to provide original multidimensional data for the training and application of the subsequent model. Specifically, it obtains the original defect identification dataset through data collection. The original defect identification dataset specifically includes the historical original defect identification dataset and the current original defect identification dataset. Both of them contain sample image data and alignment information data. The historical original defect identification dataset also contains defect-free baseline image data, defect type annotation data, and confounding factor labels.

[0070] The sample image data specifically refers to grayscale images of the coated paper surface, acquired by an area array camera. The alignment information data specifically includes camera parameters, light source angle, and image acquisition time. The defect-free baseline image data specifically refers to grayscale images of standard coated paper surfaces that have been manually confirmed to be defect-free under the same batch and coating process. The defect type labeling data includes, but is not limited to, scratches, bubbles, and pinholes. The confounding factor labels include, but are not limited to, paper type labels, coating process labels, and light intensity level labels.

[0071] Example 3, see Figure 1 , Figure 2 and Figure 3 This embodiment is based on the above embodiment. In step S2, the data optimization processing is used to optimize the collected raw multidimensional data and transform it into a regular and model-readable format. Specifically, it includes the following steps:

[0072] Step S21: Image denoising filtering, used to eliminate random noise introduced by the camera sensor and lighting, specifically by smoothing the grayscale image of the coated paper surface through Gaussian filtering to obtain a denoised image;

[0073] Step S22: ROI cropping, used to extract the effective detection area and remove useless background edges. Specifically, the coated paper area is determined by the edge detection method, and the region of interest containing only the surface of the coated paper is cropped to obtain the cropped image.

[0074] Step S23: Size normalization, used to eliminate size differences caused by different batches and different camera resolutions. Specifically, all cropped images are scaled to a fixed width and height pixel size through bilinear interpolation to obtain images with uniform size.

[0075] Step S24: Gray-scale normalization is used to eliminate the overall brightness deviation caused by changes in illumination intensity. Specifically, it involves processing the pixel values ​​of each image through histogram matching and Z-score normalization to obtain a normalized image.

[0076] Step S25: Dataset partitioning, used to obtain training data and validation data, specifically, the original historical defect identification dataset is partitioned to obtain a preliminary training set and a preliminary test set;

[0077] The current defect identification raw dataset is optimized by image denoising filtering, ROI cropping, size normalization, and grayscale standardization to obtain a preliminary dataset to be identified. The historical defect identification raw dataset is optimized by image denoising filtering, ROI cropping, size normalization, grayscale standardization, and dataset partitioning to obtain a preliminary training set and a preliminary test set.

[0078] Example 4, see Figure 1 , Figure 2 and Figure 4 This embodiment is based on the above embodiment. In step S3, the moiré pattern difference amplification is used to enhance the minute coating defect differences into significant visual texture patterns. Specifically, it is obtained by constructing an ideal defect-free template and performing optical interference with the image of the sample to be tested to obtain a difference-amplified moiré pattern image.

[0079] The amplification of moiré pattern differences specifically includes the following steps:

[0080] Step S31: Defect-free benchmark raster animation generation, used to create a dynamic visual template representing the surface of the defect-free standard coated paper as a benchmark for difference comparison, includes the following steps:

[0081] Step S311: Calculate the group average defect-free image features to eliminate individual differences and define a common defect-free standard. Specifically, this involves statistically calculating the mean and standard deviation of the grayscale values ​​at each pixel location of all defect-free baseline image samples to obtain the ideal defect-free template image and its normal fluctuation image. The formula used is as follows:

[0082] ;

[0083] In the formula, The pixel mean function represents the ideal, defect-free template image, used to obtain the ideal, defect-free template image. The pixel standard deviation function representing an ideal, defect-free template image is used to obtain images with normal fluctuations. This represents the total number of samples in the defect-free baseline image sample. The variable represents the pixel location, and num represents the index of the defect-free baseline image sample. A function representing the pixel value of the num-th defect-free reference image sample;

[0084] Step S312: Generate multi-scale Gaussian dot matrix parameters to encode the speed information of grayscale changes in visual representation to enhance dynamic expressiveness. Specifically, this involves calculating the gradient magnitude of each pixel on an ideal, defect-free template image and dynamically adjusting the size and transparency of the rendered spot accordingly to obtain dot matrix rendering parameters that reflect the speed of local changes. The formula used is as follows:

[0085] ;

[0086] In the formula, This represents the gradient magnitude function of an ideal, defect-free template image. This represents the Gaussian spot radius function based on an ideal, defect-free template image. This represents the Gaussian spot transparency function based on an ideal, defect-free template image. This represents the minimum preset Gaussian spot radius. This indicates the maximum preset Gaussian spot radius. This indicates the minimum preset Gaussian spot transparency. This indicates the maximum preset Gaussian spot transparency. This represents the minimum gradient magnitude. This indicates the maximum gradient magnitude;

[0087] Step S313: Render standard defect-free dot matrix animation, which is used to transform the digitized ideal defect-free template image into a dynamic visual template. Specifically, each pixel position on the plane of the ideal defect-free template image is mapped to a point on the canvas, and a Gaussian spot is drawn based on its coordinates and gray values ​​to obtain a standard defect-free dot matrix image sequence as the ideal defect-free template.

[0088] Step S32: Generating dot matrix animation of the sample to be tested, which is used to process the image of the sample to be tested with visual rules that are completely consistent with the ideal defect-free template. Specifically, the same mapping relationship as the ideal defect-free template is used, and based on the same steps, the image of the sample to be tested is rendered into a dot matrix image sequence to obtain the template of the sample to be tested.

[0089] Step S33: Dynamic moiré pattern synthesis, used to generate a core, amplified dynamic texture pattern, includes the following steps:

[0090] Step S331: Calculate the pixel-by-pixel difference intensity heatmap to quantify and locate the degree of defect anomaly at each position. Specifically, this is done by calculating and normalizing the absolute difference between the image of the sample to be tested and the ideal defect-free template image at each pixel position to obtain the difference intensity heatmap. The formula used is as follows:

[0091] ;

[0092] In the formula, This represents the function for calculating the absolute difference. A function representing the pixel values ​​of the image of the sample to be tested. This represents the minimum absolute difference. This represents the maximum absolute difference. Function representing pixel values ​​in a heatmap of intensity differences;

[0093] Step S332: Adaptive overlay blending, used to generate moiré patterns with visual focus concentrated in the difference regions. Specifically, it involves overlaying an ideal defect-free template with a template of the sample to be tested under a preset spatial offset, and using a calculated difference intensity heatmap as a weight map to adaptively control the blending ratio of each pixel to obtain a synthesized moiré image. The formula used is as follows:

[0094] ;

[0095] In the formula, The function representing the pixel values ​​of a composite moiré image. This represents the function that takes the minimum value. This represents the hyperparameter of mixing intensity. A function representing the pixel values ​​of an ideal, defect-free template. A function representing the pixel values ​​of the sample template after a preset spatial offset;

[0096] Step S34: Standardized image generation, used to generate standardized image data suitable for the input of the deep learning model. Specifically, the synthesized moiré image is subjected to size scaling, grayscale normalization and contrast-limited adaptive histogram equalization to obtain a difference-enlarged moiré image with uniform size and contrast.

[0097] The preliminary dataset to be identified, the preliminary training set, and the preliminary test set are used as inputs to the moiré pattern difference amplification step. Defect-free reference dot matrix animation generation, test sample dot matrix animation generation, dynamic moiré pattern synthesis, and standardized image generation are performed sequentially to obtain the dataset to be identified, the difference amplification training set, and the difference amplification test set.

[0098] By performing the above operations, this solution addresses the technical problem of traditional coated paper coating defect identification methods that rely directly on original grayscale images or manually designed texture features for feature extraction and modeling. The image variation caused by coating defects is often small and easily obscured by the inherent fiber texture of the paper itself, coating unevenness, and lighting changes, making the extraction of key defect signals extremely difficult. This solution creatively adopts a method of constructing an ideal defect-free template and performing optical interference with the image of the sample to be tested to obtain a difference-enhanced moiré pattern. This transforms abstract digital differences into visual texture interference patterns, which can significantly enhance the coating defect signal in the spatial dimension, thereby improving the lower limit and stability of defect identification.

[0099] Example 5, see Figure 1 , Figure 2 and Figure 5 This embodiment is based on the above embodiment. In step S4, the recognition model construction is used to construct the model required to decouple the pure defect-related features from the moiré image and perform defect recognition. Specifically, it is constructed by building a dual-path deep network model as the recognition model. The dual-path deep network model outputs the defect confidence index, defect type label and defect area heat map by combining local and global feature extraction and introducing an orthogonal decoupling attention mechanism.

[0100] The construction of the recognition model specifically includes the following steps:

[0101] Step S41: Backbone feature extraction, used to extract feature representations rich in spatial texture information from the moiré image, including the following steps:

[0102] Step S411: Input preprocessing, used to normalize the model input for stable training, specifically by performing channel-level Z-score normalization on the differential magnified moiré image of the input to obtain the model input tensor;

[0103] Step S412: Local texture path design, used to capture the local details and spatial variations of the moiré pattern. Specifically, the input tensor of the model is processed through a residual convolutional network. After four stages of downsampling and residual blocks, a global average pooling layer is used to output the local pattern features.

[0104] Step S413: Global structural path design, used to model the overall layout and long-distance dependencies of the moiré pattern. Specifically, the model input tensor is segmented into non-overlapping image blocks and transformed into a token sequence through linear projection. The token sequence is then input into a Transformer encoder for processing, global average pooling is used, and the output is the global structural features.

[0105] Step S414: Dual-path feature fusion, used to comprehensively utilize local details and global contextual information. Specifically, the features extracted from the two paths are concatenated and fused and dimensionality reduced through a fully connected layer to obtain comprehensive backbone features. The formula used is as follows:

[0106] ;

[0107] In the formula, Indicates comprehensive backbone characteristics, This represents the weights of the fully connected layer used for dual-path feature fusion. This represents the bias term of the fully connected layer used for dual-path feature fusion. Representing local pattern features, Represents global structural features;

[0108] Step S42: Orthogonal decoupling attention design, used to explicitly separate the main defect signal from confounding factor signals in the feature space, including the following steps:

[0109] Step S421: Construction of learnable orthogonal direction basis, used to define a set of mutually independent semantic directions in the feature space. Specifically, by initializing and constraining the column vectors of a learnable matrix to maintain unit orthogonality, an orthogonal basis for eigenvalue decomposition is obtained. The first basis vector of the orthogonal basis is specified as the defect principal direction vector.

[0110] Step S422: Orthogonal projection and reconstruction, used to decompose the comprehensive backbone features into different orthogonal directions. Specifically, the comprehensive backbone features are projected onto a normal orthogonal basis to obtain projection coefficients, and the defect-clean features are reconstructed separately based on the projection coefficients of the defect principal direction. At the same time, the difference between the projection coefficients and the reconstructed total features is calculated to obtain the projection coefficient vector, defect-clean features, and hybrid difference features. The formulas used are as follows:

[0111] ;

[0112] In the formula, pc represents the projection coefficient vector, and U represents the orthonormal basis. Indicates the purity characteristics of defects. This represents the value of the first element of the projection coefficient vector, i.e., the projection coefficient obtained based on the principal direction vector of the defect. This represents the principal direction vector of the defect. Indicates the overall characteristics of reconstruction. Tr represents the mixed differential characteristics, and Tr represents the transpose operation;

[0113] Step S423: Adversarial training of multi-head orthogonal discriminators, used to force the profanity features to not contain information related to defects. Specifically, multiple discriminator heads are trained to attempt to predict profanity labels from the profanity features, and gradient inversion is applied for adversarial training, forcing the network to encode profanity factors into a subspace orthogonal to the main direction of the defects. The total adversarial loss is calculated as follows:

[0114] ;

[0115] In the formula, This represents the total adversarial loss value, where I represents the total number of discriminator heads. This represents the function for calculating cross-entropy loss. This represents the running function of the i-th discriminator head. Indicates the label of the i-th confounding factor;

[0116] Step S43: Model classification prediction, used to map the defect-pure features to the required output results. Specifically, it outputs the defect confidence index by performing a learnable affine transformation on the projection coefficients obtained based on the defect principal direction, and outputs the defect type label by processing the defect-pure features with a softmax classification head.

[0117] Step S44: Interpretable attention generation, which provides visual clues for model decision-making. Specifically, the defect confidence index is used as the query to perform attention weighting on local pattern features and restores the original input size through bilinear interpolation. Finally, a defect area heatmap with highlighted abnormal areas related to defects is obtained.

[0118] Step S45: Construct and train the model. Specifically, the dual-path deep network model is constructed by integrating the backbone feature extraction, the orthogonal decoupled attention design, the model classification prediction, and the interpretable attention generation. The model is trained and its performance is verified based on the difference amplification training set and the difference amplification test set to obtain the dual-path deep network model as the recognition model.

[0119] By performing the above operations, this solution addresses the technical problem of traditional coated paper coating defect identification methods lacking explicit decoupling design for confounding factors. Coated paper surface anomalies are complex manifestations of multiple coexisting factors. Confounding factors such as paper type, coating process parameters, and ambient lighting can become entangled with the real defect signal in the feature space, and the identification results are often mixed with interference from non-defect factors. This solution creatively adopts a dual-path deep network model as the identification model. By combining local and global feature extraction and introducing an orthogonal decoupling attention mechanism, the feature space is decomposed into the main defect signal and the confounding factor signal, which can extract the pure characterization belonging to the coating defect from the complex surface anomaly pattern.

[0120] Example 6, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. Specifically, the defect identification output is to use the dataset to be identified as the input of the identification model to identify coating defects, and the defect confidence index, defect type label and defect area heat map output by the identification model are used as the final defect identification result.

[0121] Example 7, see Figure 1 and Figure 2 Based on the above embodiments, this embodiment provides an intelligent coating paper coating defect identification system, including a multi-dimensional data acquisition module, a data optimization processing module, a moiré pattern difference amplification module, an identification model construction module, and a defect identification output module;

[0122] The multidimensional data acquisition module is used to collect raw data, obtain a defect identification raw dataset by collecting raw data, and send the defect identification raw dataset to the data optimization processing module.

[0123] The data optimization processing module is used for data optimization processing. Through data optimization processing, a preliminary dataset to be identified, a preliminary training set, and a preliminary test set are obtained, and the preliminary dataset to be identified, the preliminary training set, and the preliminary test set are sent to the moiré pattern difference amplification module.

[0124] The moiré pattern difference amplification module is used for moiré pattern difference amplification. It constructs an ideal defect-free template and performs optical interference with the image of the sample to be tested. It processes the preliminary dataset to be identified, the preliminary training set, and the preliminary test set to obtain the dataset to be identified, the difference amplification training set, and the difference amplification test set. The dataset to be identified is sent to the defect identification output module, and the difference amplification training set and the difference amplification test set are sent to the identification model construction module.

[0125] The identification model construction module is used to construct the identification model. By constructing a dual-path deep network model, the identification model is obtained and then sent to the defect identification output module.

[0126] The defect identification output module is used for defect identification output. By using the identification model, it identifies coating defects in coated paper based on the dataset to be identified, and obtains the final defect identification result.

[0127] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0128] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0129] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An intelligent method for identifying coating defects in coated paper, characterized in that: The method includes the following steps: S1: Multidimensional data acquisition: By collecting data, the original dataset for defect identification is obtained. Specifically, the original dataset for defect identification includes historical original datasets for defect identification and current original datasets for defect identification. S2: Data optimization processing, which optimizes the collected raw multidimensional data to obtain a preliminary dataset to be identified, a preliminary training set, and a preliminary test set; S3: Moiré pattern difference amplification, used to enhance minute coating defect differences into significant visual texture patterns. Specifically, it involves constructing an ideal defect-free template and optically interfering it with the image of the sample to be tested to obtain a difference-amplified moiré pattern image. After the initial dataset to be identified, the initial training set, and the initial test set are processed by the moiré pattern difference amplification step, the dataset to be identified, the difference amplification training set, and the difference amplification test set are obtained. S4: Recognition Model Construction, used to construct the model required to decouple pure defect-related features from moiré images and perform defect recognition. Specifically, a dual-path deep network model is constructed as the recognition model. The dual-path deep network model outputs a defect confidence index, defect type label, and defect region heatmap by combining local and global feature extraction and introducing an orthogonal decoupling attention mechanism. S5: Defect identification output, specifically, the dataset to be identified is used as the input of the identification model to identify coating defects, and the defect confidence index, defect type label and defect area heat map output by the identification model are used as the final defect identification result.

2. The intelligent method for identifying coating defects in coated paper according to claim 1, characterized in that: The amplification of moiré pattern differences specifically includes the following steps: Step S31: Defect-free benchmark raster animation generation, used to create a dynamic visual template representing the surface of the defect-free standard coated paper as a benchmark for difference comparison, includes the following steps: Step S311: Calculate the group average defect-free image features to eliminate individual differences and define a common defect-free standard. Specifically, by statistically calculating the mean and standard deviation of grayscale at each pixel position of all defect-free benchmark image samples, the ideal defect-free template image and its normal fluctuation image are obtained. Step S312: Generate multi-scale Gaussian dot matrix parameters to encode the speed information of grayscale changes in visual representation to enhance dynamic expression capabilities. Specifically, by calculating the gradient magnitude of each pixel on the ideal defect-free template image and dynamically adjusting the size and transparency of the rendering spot accordingly, dot matrix rendering parameters reflecting the speed of local changes are obtained. Step S313: Render standard defect-free dot matrix animation, which is used to transform the digitized ideal defect-free template image into a dynamic visual template. Specifically, each pixel position on the plane of the ideal defect-free template image is mapped to a point on the canvas, and a Gaussian spot is drawn based on its coordinates and gray values ​​to obtain a standard defect-free dot matrix image sequence as the ideal defect-free template. Step S32: Generating dot matrix animation of the sample to be tested, which is used to process the image of the sample to be tested with visual rules that are completely consistent with the ideal defect-free template. Specifically, the same mapping relationship as the ideal defect-free template is used, and based on the same steps, the image of the sample to be tested is rendered into a dot matrix image sequence to obtain the template of the sample to be tested. Step S33: Dynamic moiré pattern synthesis, used to generate a core, amplified dynamic texture pattern, includes the following steps: Step S331: Calculate the pixel-by-pixel difference intensity heatmap to quantify and locate the degree of defect anomaly at each location. Specifically, the difference intensity heatmap is obtained by calculating the absolute difference between the image of the sample to be tested and the ideal defect-free template image at each pixel location and normalizing it. Step S332: Adaptive overlay blending, used to generate a moiré pattern with visual focus concentrated in the difference region. Specifically, it involves overlaying an ideal defect-free template with a template of the sample to be tested under a preset spatial offset, and using the calculated difference intensity heat map as a weight map to adaptively control the blending ratio of each pixel to obtain a synthetic moiré image. Step S34: Standardized image generation, used to generate standardized image data suitable for the input of the deep learning model. Specifically, the synthesized moiré image is subjected to size scaling, grayscale normalization and contrast-limited adaptive histogram equalization to obtain a difference-enlarged moiré image with uniform size and contrast. The preliminary dataset to be identified, the preliminary training set, and the preliminary test set are used as inputs to the moiré pattern difference amplification step. Defect-free reference dot matrix animation generation, test sample dot matrix animation generation, dynamic moiré pattern synthesis, and standardized image generation are performed sequentially to obtain the dataset to be identified, the difference amplification training set, and the difference amplification test set.

3. The intelligent method for identifying coating defects in coated paper according to claim 1, characterized in that: The construction of the recognition model specifically includes the following steps: Step S41: Backbone feature extraction, used to extract feature representations rich in spatial texture information from the moiré image, including the following steps: Step S411: Input preprocessing, used to normalize the model input for stable training, specifically by performing channel-level Z-score normalization on the differential magnified moiré image of the input to obtain the model input tensor; Step S412: Local texture path design, used to capture the local details and spatial variations of the moiré pattern. Specifically, the input tensor of the model is processed through a residual convolutional network. After four stages of downsampling and residual blocks, a global average pooling layer is used to output the local pattern features. Step S413: Global structural path design, used to model the overall layout and long-distance dependencies of the moiré pattern. Specifically, the model input tensor is segmented into non-overlapping image blocks and transformed into a token sequence through linear projection. The token sequence is then input into a Transformer encoder for processing, global average pooling is used, and the output is the global structural features. Step S414: Dual-path feature fusion, used to comprehensively utilize local details and global context information. Specifically, the features extracted from the two paths are concatenated and fused and dimensionality reduced through a fully connected layer to obtain comprehensive backbone features. Step S42: Orthogonal decoupling attention design, used to explicitly separate the main defect signal from confounding factor signals in the feature space, including the following steps: Step S421: Construction of learnable orthogonal direction basis, used to define a set of mutually independent semantic directions in the feature space. Specifically, by initializing and constraining the column vectors of a learnable matrix to maintain unit orthogonality, an orthogonal basis for eigenvalue decomposition is obtained. The first basis vector of the orthogonal basis is specified as the defect principal direction vector. Step S422: Orthogonal projection and reconstruction, used to decompose the comprehensive backbone features into different orthogonal directions. Specifically, the comprehensive backbone features are projected onto the orthogonal basis to obtain projection coefficients, and the defect pure features are reconstructed separately based on the projection coefficients of the defect principal direction. At the same time, the difference between the reconstruction and the total features is calculated to obtain the projection coefficient vector, defect pure features and mixed difference features. Step S423: Adversarial training of multi-head orthogonal discriminators, used to force the heterogeneous features not to contain information related to defects. Specifically, multiple discriminator heads are trained to try to predict heterogeneous factor labels from heterogeneous features, and gradient inversion is applied for adversarial training, forcing the network to encode heterogeneous factors into a subspace orthogonal to the main direction of defects. Step S43: Model classification prediction, used to map the defect-pure features to the required output results. Specifically, it outputs the defect confidence index by performing a learnable affine transformation on the projection coefficients obtained based on the defect principal direction, and outputs the defect type label by processing the defect-pure features with a softmax classification head. Step S44: Interpretable attention generation, which provides visual clues for model decision-making. Specifically, the defect confidence index is used as the query to perform attention weighting on local pattern features and restores the original input size through bilinear interpolation. Finally, a defect area heatmap with highlighted abnormal areas related to defects is obtained. Step S45: Construct and train the model. Specifically, the dual-path deep network model is constructed by integrating the backbone feature extraction, the orthogonal decoupled attention design, the model classification prediction, and the interpretable attention generation. The model is trained and its performance is verified based on the difference amplification training set and the difference amplification test set to obtain the dual-path deep network model as the recognition model.

4. The intelligent method for identifying coating defects in coated paper according to claim 1, characterized in that: Both the historical defect identification raw dataset and the current defect identification raw dataset contain sample image data and alignment information data. The historical defect identification raw dataset also contains defect-free baseline image data, defect type annotation data, and confounding factor labels.

5. The intelligent method for identifying coating defects in coated paper according to claim 1, characterized in that: The data optimization process specifically includes the following steps: Step S21: Image denoising filtering, used to eliminate random noise introduced by the camera sensor and lighting, specifically by smoothing the grayscale image of the coated paper surface through Gaussian filtering to obtain a denoised image; Step S22: ROI cropping, used to extract the effective detection area and remove useless background edges. Specifically, the coated paper area is determined by the edge detection method, and the region of interest containing only the surface of the coated paper is cropped to obtain the cropped image. Step S23: Size normalization, used to eliminate size differences caused by different batches and different camera resolutions. Specifically, all cropped images are scaled to a fixed width and height pixel size through bilinear interpolation to obtain images with uniform size. Step S24: Gray-scale normalization is used to eliminate the overall brightness deviation caused by changes in illumination intensity. Specifically, it involves processing the pixel values ​​of each image through histogram matching and Z-score normalization to obtain a normalized image. Step S25: Dataset partitioning, used to obtain training data and validation data, specifically, the original historical defect identification dataset is partitioned to obtain a preliminary training set and a preliminary test set; The current defect identification raw dataset is optimized by image denoising filtering, ROI cropping, size normalization, and grayscale standardization to obtain a preliminary dataset to be identified. The historical defect identification raw dataset is optimized by image denoising filtering, ROI cropping, size normalization, grayscale standardization, and dataset partitioning to obtain a preliminary training set and a preliminary test set.

6. An intelligent coating paper coating defect identification system, used to implement the intelligent coating paper coating defect identification method as described in any one of claims 1-5, characterized in that: It includes a multi-dimensional data acquisition module, a data optimization and processing module, a moiré pattern difference amplification module, a recognition model construction module, and a defect recognition output module.

7. The intelligent coating defect identification system for coated paper according to claim 6, characterized in that: The multidimensional data acquisition module is used to collect raw data, obtain a defect identification raw dataset by collecting raw data, and send the defect identification raw dataset to the data optimization processing module. The data optimization processing module is used for data optimization processing. Through data optimization processing, a preliminary dataset to be identified, a preliminary training set, and a preliminary test set are obtained, and the preliminary dataset to be identified, the preliminary training set, and the preliminary test set are sent to the moiré pattern difference amplification module. The moiré pattern difference amplification module is used for moiré pattern difference amplification. It constructs an ideal defect-free template and performs optical interference with the image of the sample to be tested. It processes the preliminary dataset to be identified, the preliminary training set, and the preliminary test set to obtain the dataset to be identified, the difference amplification training set, and the difference amplification test set. The dataset to be identified is sent to the defect identification output module, and the difference amplification training set and the difference amplification test set are sent to the identification model construction module. The identification model construction module is used to construct the identification model. By constructing a dual-path deep network model, the identification model is obtained and then sent to the defect identification output module. The defect identification output module is used for defect identification output. By using the identification model, it identifies coating defects in coated paper based on the dataset to be identified, and obtains the final defect identification result.

Images