Online Quality Discrimination Algorithm and System for Decorative Paper

Abstract

The present invention relates to the technical field of decorative paper quality detection, and discloses an online quality discrimination algorithm and system for decorative paper. The algorithm first collects high-definition texture images of the decorative paper samples to be detected, presets three-level scale decomposition thresholds, performs three-level scale decomposition and feature extraction on the high-definition texture images to obtain a multi-scale texture image group and a multi-scale texture feature set. Then, a texture quality attention model is constructed to obtain a multi-scale attention weight vector set; based on the multi-scale attention weight vector set, the multi-scale texture features are weighted and fused to obtain a multi-scale fusion feature set; finally, the quality of the decorative paper samples to be detected is determined according to the multi-scale fusion features, and the quality determination level of the decorative paper samples to be detected is obtained. The present invention can accurately reconstruct the three-dimensional morphology of the reticulated surface, quantitatively extract the dot morphology parameters, and significantly improve the printing quality and efficiency of intaglio printing of decorative paper.

Description

Technical Field

[0001] The present invention relates to the technical field of decorative paper quality detection, and more particularly to an algorithm and system for online quality determination of decorative paper. Background Art

[0002] With the continuous advancement of technology and the improvement of people's living standards, gravure printing of decorative paper has become an indispensable process in the manufacture of decorative materials such as furniture and flooring. This printing process places high demands on printing quality. The key challenge facing the decorative paper printing industry is how to measure the printing quality online and adjust printing parameters in a timely manner.

[0003] Chinese patent publication number CN104034424A discloses a method for detecting and analyzing the printing quality of decorative paper. This method uses a spectrophotometer to scan the measurement and control strips on a standard sample and the current print, acquiring spectral data for each color block. The method then calculates the chromaticity and dot gain values for each color block. By comparing the chromaticity and dot gain values of the standard sample and the current print, the printing quality is evaluated and the necessary printing process adjustments are indicated. This patent addresses the problem of decorative paper printing quality detection and control relying solely on worker experience and lacking quantitative detection and control technology. However, this patent fails to automatically locate and accurately quantify printing quality issues, and subsequent optimization and adjustment of the printing process still requires manual experience.

[0004] A Chinese patent with authorization announcement number CN113610850B discloses a method for detecting texture anomalies in decorative paper based on image processing. The method first captures a surface image of the wood-grained paper to be tested, extracts the morphological features of the edge to be tested and the true edge, calculates the feature similarity between the two, and obtains multiple target true edges similar to the edge to be tested; then, the trend vectors of each pixel point of the edge to be tested and the true edge are obtained, and the trend similarity between the set of trend vectors to be tested and the set of true trend vectors is calculated; finally, the texture smoothness of the edge to be tested is determined from the feature similarity and trend similarity to determine whether the wood-grained paper is abnormal. This patent uses feature similarity and trend similarity to evaluate the smoothness of edge texture, which can detect subtle edge defects and improve the accuracy of anomaly detection. However, this patent mainly focuses on the detection of edge texture of wood-grained paper and has limited ability to evaluate the overall quality of decorative paper surface texture. In addition, this patent does not consider the fusion analysis of texture features at different scales, and the comprehensiveness and accuracy of texture quality evaluation needs to be further improved.

[0005] In summary, the existing decorative paper printing quality detection methods lack the extraction and fusion analysis of multi-scale texture features, making it difficult to comprehensively and accurately characterize the surface texture quality of decorative paper. The printing quality assessment results are not interpretable enough, and there is a lack of reliable quantitative basis for subsequent printing parameter optimization. Summary of the Invention

[0006] To overcome the above defects of the prior art, the present invention provides an online quality discrimination algorithm and system for decorative paper, which realizes multi-scale feature extraction, fusion and intelligent analysis of the surface texture of decorative paper, and establishes a quantitative evaluation and optimization adjustment mechanism for printing quality.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] An online quality discrimination algorithm for decorative paper, comprising:

[0009] Collect high-definition texture images of the decorative paper samples to be detected, preset three-level scale decomposition thresholds, and perform three-level scale decomposition and feature extraction on the high-definition texture images according to the preset three-level scale decomposition thresholds to obtain a multi-scale texture image group and a multi-scale texture feature set; evaluate the effect of the three-level scale decomposition; the multi-scale texture feature set includes first-scale texture features, second-scale texture features and third-scale texture features;

[0010] Construct a texture quality attention model, and obtain a multi-scale attention weight vector set according to the multi-scale texture feature set and the texture quality attention model; based on the multi-scale attention weight vector set, perform weighted fusion on the first-scale texture features, second-scale texture features and third-scale texture features to obtain a multi-scale fusion feature set;

[0011] According to the multi-scale fusion features, perform quality determination on the decorative paper samples to be detected to obtain the quality determination level of the decorative paper samples to be detected.

[0012] Further, the preset three-level scale decomposition thresholds include:

[0013] Introduce a scale factor, and establish a three-level scale decomposition threshold calculation formula according to the size, resolution Res and scale factor of the high-definition texture image, and calculate the three-level scale decomposition thresholds, which are respectively the fine-scale decomposition threshold T1, the medium-scale decomposition threshold T2 and the large-scale decomposition threshold T3; the size of the high-definition texture image includes the image width Wid and the image height Height.

[0014] Further, the three-level scale decomposition of the high-definition texture image includes:

[0015] Perform Gaussian smoothing on the high-definition texture image, and then perform adaptive downsampling according to the fine-scale decomposition threshold T1 to generate a first-scale texture image;

[0016] Perform Gaussian smoothing on the first-scale texture image, perform adaptive downsampling according to the medium-scale decomposition threshold T2, and fuse the high-frequency information of the Laplacian pyramid decomposition to generate a second-scale texture image;

[0017] Perform Gaussian smoothing on the second-scale texture image, perform adaptive downsampling according to the large-scale decomposition threshold T3, and fuse the high-frequency information of the Laplacian pyramid decomposition to generate the third-scale texture image;

[0018] Construct a multi-scale texture image group from the first-scale texture image, the second-scale texture image, and the third-scale texture image.

[0019] Furthermore, the feature extraction of the high-definition texture image includes:

[0020] Extract first-scale texture features from the first-scale texture image; the first-scale texture features include first-scale Gabor texture features, first-scale LBP texture features, and first-scale GLCM texture features;

[0021] Extract second-scale texture features from the second-scale texture image; the second-scale texture features include second-scale Gabor texture features, second-scale LBP texture features, and second-scale GLCM texture features;

[0022] Extract third-scale texture features from the third-scale texture image; the third-scale texture features include third-scale Gabor texture features, third-scale LBP texture features, and third-scale GLCM texture features;

[0023] Construct a multi-scale texture feature set from the first-scale texture features, the second-scale texture features, and the third-scale texture features.

[0024] Furthermore, the evaluation of the effect of the three-level scale decomposition includes:

[0025] Calculate the structural similarity SSIM index and the scale-space feature stability SSFS index of the first-scale texture image, the second-scale texture image, and the third-scale texture image with the high-definition texture image respectively to obtain an SSIM index set and an SSFS index set;

[0026] The SSIM index set includes SSIM indexes of three scales, and the SSIM indexes of the three scales are the first structural similarity SSIM index SSIM1, the second structural similarity SSIM index SSIM2, and the third structural similarity SSIM index SSIM3 respectively;

[0027] The SSFS index set includes SSFS indexes of three scales, and the SSFS indexes of the three scales are the first scale-space feature stability SSFS index SSFS1, the second scale-space feature stability SSFS index SSFS2, and the third scale-space feature stability SSFS index SSFS3 respectively;

[0028] Judge whether the SSIM index of all three scales is greater than the SSIM similarity threshold Ts, and whether the SSFS index of all three scales is greater than the SSFS stability threshold Tf; if not, adaptively adjust the scale factor in the formula for calculating the three-level scale decomposition threshold until the multi-scale decomposition effect is ideal.

[0029] Further, the multi-scale attention weight vector set includes a first-scale attention weight vector, a second-scale attention weight vector, and a third-scale attention weight vector;

[0030] The obtaining of the multi-scale attention weight vector set according to the multi-scale texture feature set and the texture quality attention model includes:

[0031] Concatenate the first-scale Gabor texture feature, the first-scale LBP texture feature, and the first-scale GLCM texture feature to form a first-scale comprehensive feature; concatenate the second-scale Gabor texture feature, the second-scale LBP texture feature, and the second-scale GLCM texture feature to form a second-scale comprehensive feature; concatenate the third-scale Gabor texture feature, the third-scale LBP texture feature, and the third-scale GLCM texture feature to form a third-scale comprehensive feature.

[0032] Input the first-scale comprehensive feature into the texture quality attention model to obtain a first-scale attention heat map; input the second-scale comprehensive feature into the texture quality attention model to obtain a second-scale attention heat map; input the third-scale comprehensive feature into the texture quality attention model to obtain a third-scale attention heat map;

[0033] Perform global average pooling operations on the first-scale attention heat map, the second-scale attention heat map, and the third-scale attention heat map to obtain a first-scale attention weight vector, a second-scale attention weight vector, and a third-scale attention weight vector.

[0034] Further, the obtaining of the multi-scale fusion feature includes:

[0035] Obtain a first-scale fusion feature according to the first-scale attention weight vector and the first-scale texture feature;

[0036] Obtain a second-scale fusion feature according to the second-scale attention weight vector and the second-scale texture feature;

[0037] Obtain a third-scale fusion feature according to the third-scale attention weight vector and the third-scale texture feature;

[0038] Concatenate the first-scale fusion feature, the second-scale fusion feature, and the third-scale fusion feature to generate a multi-scale fusion feature.

[0039] Further, the obtaining of the first-scale fusion feature includes:

[0040] Multiply the first-scale attention weight vector by the first-scale Gabor texture feature to obtain the first-scale Gabor weighted feature;

[0041] Multiply the first-scale attention weight vector by the first-scale LBP texture feature to obtain the first-scale LBP weighted feature;

[0042] Multiply the first-scale attention weight vector by the first-scale GLCM texture feature to obtain the first-scale GLCM weighted feature;

[0043] Fuse the first-scale Gabor weighted feature, the first-scale LBP weighted feature, and the first-scale GLCM weighted feature to obtain the first-scale fusion feature.

[0044] Further, the quality determination of the to-be-detected decorative paper sample according to the multi-scale fusion feature includes:

[0045] Construct a lightweight texture feature extraction network, input the multi-scale fusion feature into the lightweight texture feature extraction network to obtain the multi-scale lightweight fusion feature;

[0046] According to the multi-scale lightweight fusion feature and the pre-constructed quality discrimination classifier, obtain the quality determination level and the quality determination score corresponding to the quality determination level;

[0047] Judge whether the quality determination score corresponding to the quality determination level meets the preset quality level threshold. If it meets, output the final quality determination level. If it does not meet, output a warning signal.

[0048] An on-line quality discrimination system for decorative paper, which is used to implement the above-mentioned on-line quality discrimination algorithm for decorative paper. The system includes:

[0049] An image decomposition module: used to collect the high-definition texture image of the to-be-detected decorative paper sample, preset a three-level scale decomposition threshold, and perform three-level scale decomposition and feature extraction on the high-definition texture image according to the preset three-level scale decomposition threshold to obtain a multi-scale texture image group and a multi-scale texture feature set; the multi-scale texture feature set includes a first-scale texture feature, a second-scale texture feature, and a third-scale texture feature;

[0050] A decomposition effect evaluation module: used to evaluate the effect of the three-level scale decomposition;

[0051] Feature fusion module: used to construct a texture quality attention model, and obtain a multi-scale attention weight vector set according to the multi-scale texture feature set and the texture quality attention model; based on the multi-scale attention weight vector set, perform weighted fusion on the first-scale texture feature, the second-scale texture feature, and the third-scale texture feature to obtain a multi-scale fusion feature set;

[0052] Quality determination module: according to the multi-scale fusion features, perform quality determination on the decorative paper sample to be detected, and obtain the quality determination level of the decorative paper sample to be detected.

[0053] An electronic device includes a memory, a central processing unit, and a computer program stored on the memory and executable on the central processing unit. When the central processing unit executes the computer program, the above-mentioned online quality discrimination algorithm for decorative paper is implemented.

[0054] A computer-readable storage medium stores a computer program, and when the computer program is executed, the above-mentioned online quality discrimination algorithm for decorative paper is implemented.

[0055] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0056] By using a high-resolution industrial camera to collect high-definition texture images of the decorative paper surface, the integrity and accuracy of texture details are ensured, providing a high-quality data basis for subsequent multi-scale texture analysis. The three-level scale decomposition strategy is adopted, which can comprehensively extract texture features at different scales, including both fine local texture details and global texture distribution rules, enhancing the characterization ability of texture features. The texture quality attention mechanism is incorporated, and the importance of different-scale texture features is adaptively adjusted through attention weights, highlighting the key texture regions that have a greater impact on printing quality and improving the accuracy of quality determination. The multi-scale texture features are weighted and fused, making full use of the complementarity of different-scale textures to form a more comprehensive and robust texture feature set, enhancing the reliability of quality determination. A lightweight texture feature extraction network and a quality discrimination classifier are built to achieve end-to-end online detection of the printing quality of decorative paper, and the quality determination level and score can be quantitatively output, providing an intuitive and quantitative basis for quality monitoring and improvement in the production process. The entire algorithm process adopts a modular design, and each module can be flexibly combined, facilitating function expansion and performance optimization. At the same time, the cooperation of the software and hardware systems is realized, forming a complete online detection solution for the quality of decorative paper, which can significantly improve the quality management level and production efficiency of intaglio printing of decorative paper. Description of the Drawings

[0057] To more clearly illustrate the technical solutions in the embodiments of the present invention or in the prior art, the following will briefly introduce the drawings required in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0058] Figure 1 It is the principle flowchart of the online quality discrimination algorithm for decorative paper in the present invention;

[0059] Figure 2 It is the method flowchart for obtaining the multi-scale texture image group and the multi-scale texture feature set in the online quality discrimination algorithm for decorative paper of the present invention;

[0060] Figure 3 It is the method flowchart for obtaining the multi-scale attention weight vector set in the online quality discrimination algorithm for decorative paper of the present invention;

[0061] Figure 4 It is the method flowchart for obtaining the multi-scale fusion feature in the online quality discrimination algorithm for decorative paper of the present invention;

[0062] Figure 5 It is the method flowchart for obtaining the first-scale fusion feature in the online quality discrimination algorithm for decorative paper of the present invention;

[0063] Figure 6 It is the method flowchart for obtaining the second-scale fusion feature in the online quality discrimination algorithm for decorative paper of the present invention;

[0064] Figure 7 It is the method flowchart for obtaining the third-scale fusion feature in the online quality discrimination algorithm for decorative paper of the present invention;

[0065] Figure 8 It is the method flowchart for performing quality determination on the decorative paper sample to be detected in the online quality discrimination algorithm for decorative paper of the present invention;

[0066] Figure 9 It is the functional module diagram of the online quality discrimination system for decorative paper in the present invention. Specific Embodiments

[0067] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0068] Embodiment 1

[0069] Please refer to Figure 1 As shown, this embodiment provides an online quality discrimination algorithm for decorative paper, including:

[0070] Step S1000, collect high-definition texture images of the decorative paper samples to be detected, preset three-level scale decomposition thresholds, and perform three-level scale decomposition and feature extraction on the high-definition texture images according to the preset three-level scale decomposition thresholds to obtain a multi-scale texture image group and a multi-scale texture feature set; evaluate the effect of the three-level scale decomposition; the multi-scale texture feature set includes first-scale texture features, second-scale texture features, and third-scale texture features;

[0071] Furthermore, step S1000 includes:

[0072] Step S1100, collect high-definition texture images of the decorative paper samples to be detected;

[0073] Specifically, use a high-resolution industrial camera to scan and image the surface of the decorative paper samples to obtain high-definition texture images with a resolution of not less than 1200 dpi; the high resolution of the camera ensures that the collected texture images contain rich texture detail information, providing a high-quality data basis for subsequent multi-scale decomposition and feature extraction. At the same time, the scanning and imaging method can quickly obtain complete texture images of a large area of decorative paper, improving the detection efficiency. Perform preprocessing on the collected high-definition texture images, including image denoising, image enhancement, and image correction; image denoising can suppress the noise interference mixed in the imaging process, and common denoising methods include median filtering, wavelet denoising, etc. Image enhancement can improve the contrast and clarity of the image, such as histogram equalization, Retinex and other algorithms. Image correction can eliminate the geometric distortion generated during the imaging process, ensuring the geometric consistency of the texture images. These preprocessing operations clear the obstacles for the accurate extraction of texture features. Store the processed high-definition texture images in the form of a data set and add necessary sample identification information, providing a unified and standardized data source for subsequent training and testing, and improving the systematicness of data management.

[0074] Step S1100 starts from the selection of high-standard hardware devices, uses a high-resolution industrial camera and scanning imaging method to collect the surface texture of the decorative paper with ultra-high resolution, obtaining a large texture image with rich details, integrity, and accuracy. And through a series of preprocessing operations, remove the noise interference in the image, correct the distortion, improve the image quality, and finally centrally manage the high-quality texture image data, making full preparations for subsequent texture analysis. High-quality texture images are the fundamental guarantee for texture feature extraction and have a decisive impact on the final detection performance. This step lays a solid data foundation for texture feature extraction.

[0075] Step S1200, preset three-level scale decomposition thresholds;

[0076] The preset three - level scale decomposition threshold includes:

[0077] Introduce a scale factor. According to the size, resolution Res, and scale factor of the high - definition texture image, establish a calculation formula for the three - level scale decomposition threshold, and calculate the three - level scale decomposition threshold. The three - level scale decomposition thresholds are respectively the fine - scale decomposition threshold T1, the medium - scale decomposition threshold T2, and the large - scale decomposition threshold T3; the size of the high - definition texture image includes the image width Wid and the image height Height.

[0078] The calculation formula for the three - level scale decomposition threshold is:

[0079]

[0080] where represents the scale level, and the values are 1, 2, 3; is the scale factor, k1 < k2 < k3, the initial value of k1 is 1, the initial value of k2 is 4, and the initial value of k3 is 16.

[0081] Specifically, the preset three - level scale decomposition threshold is to set a suitable scale range for the multi - scale decomposition of the decorative paper texture image. The three levels of scale respectively correspond to the fine scale, the medium scale, and the large scale, and can capture the morphological characteristics of the decorative paper surface texture at different observation distances. The fine scale focuses on the fine texture of the paper, such as fiber arrangement, dot pattern, etc.; the medium scale pays attention to the medium - sized patterns on the paper surface, such as patterns, wood grains, etc.; the large scale reflects the overall texture of the paper, such as roughness, gloss, etc. Reasonably setting the decomposition thresholds of the three scales is crucial for comprehensively extracting the multi - scale texture information of the decorative paper.

[0082] Considering the differences in the size and resolution of the originally collected high - definition texture images, it is necessary to adaptively calculate the three - level scale decomposition threshold. The calculation of the threshold is based on the image width Wid, height Height, and resolution Res, and at the same time introduces a scale factor , which gives flexibility to the threshold calculation. The image width Wid and image height Height are in pixels. The initial values of the scale factors are k1 = 1, k2 = 4, k3 = 16, which means that the medium scale downsamples the original image by 4 times in both the width and height directions, and the large scale downsamples by 16 times. This initial setting can obtain a relatively balanced decomposition effect for the three scales. However, in practical applications, the scale factors can be optimized and adjusted according to the texture characteristics of the decorative paper and the requirements of quality inspection.

[0083] Downsampling the original high-definition texture image with thresholds at three scales can obtain texture images at three scales, reflecting the changes in the paper surface texture features at different scales. The size of the threshold determines the degree of texture information retention: the smaller the threshold, the fewer texture details are retained and the coarser the pattern; conversely, finer textures can be depicted. Therefore, by adjusting the scale factor, the fineness of texture decomposition can be flexibly controlled at different scales to meet the detection requirements of different types of decorative papers.

[0084] The initial value of k1 is 1, the initial value of k2 is 4, and the initial value of k3 is 16. This means that the fine scale retains all the information of the original texture image, the medium scale retains about 1 / 4 of the information volume, and the large scale retains about 1 / 16. This decomposition can better reflect the hierarchical changes in the texture of the decorative paper, facilitating the subsequent extraction of discriminative multi-scale texture features.

[0085] By presetting the three-level scale decomposition threshold, the multi-scale decomposition process can be quantified to make it more objective and reproducible. At the same time, calculating the threshold adaptively based on the image size and resolution can ensure that the decomposition method is applicable to texture images of different specifications, with strong robustness and practicality. In the subsequent feature extraction and quality assessment, the texture images at these three scales will play an important role.

[0086] Step S1300, perform three-level scale decomposition and feature extraction on the high-definition texture image according to the preset three-level scale decomposition threshold to obtain a multi-scale texture image group and a multi-scale texture feature set;

[0087] Furthermore, as Figure 2 shown, step S1300 includes:

[0088] Step S1310, perform Gaussian smoothing on the high-definition texture image, and then perform adaptive downsampling according to the fine scale decomposition threshold T1 to generate a first-scale texture image; and extract first-scale texture features from the first-scale texture image; the first-scale texture features include first-scale Gabor texture features, first-scale LBP texture features, and first-scale GLCM texture features;

[0089] Step S1320, perform Gaussian smoothing on the first-scale texture image, perform adaptive downsampling according to the medium scale decomposition threshold T2, and fuse the high-frequency information of the Laplacian pyramid decomposition to generate a second-scale texture image; and extract second-scale texture features from the second-scale texture image; the second-scale texture features include second-scale Gabor texture features, second-scale LBP texture features, and second-scale GLCM texture features;

[0090] Step S1330: Perform Gaussian smoothing on the second-scale texture image, perform adaptive downsampling according to the large-scale decomposition threshold T3, and fuse the high-frequency information of the Laplacian pyramid decomposition to generate the third-scale texture image; and extract the third-scale texture features from the third-scale texture image; the third-scale texture features include the third-scale Gabor texture features, the third-scale LBP texture features, and the third-scale GLCM texture features.

[0091] Step S1340: Combine the first-scale texture image, the second-scale texture image, and the third-scale texture image to form a multi-scale texture image group; combine the first-scale texture features, the second-scale texture features, and the third-scale texture features to form a multi-scale texture feature set.

[0092] Specifically, the Gaussian smoothing in step S1310 is an image preprocessing technique. It uses a Gaussian function to smooth the image, which can remove the noise and details in the image, thereby highlighting the main features of the image. Gaussian smoothing is equivalent to a low-pass filter, which can smooth and blur the image.

[0093] After Gaussian smoothing, perform adaptive downsampling on the image according to the fine-scale decomposition threshold T1. The so-called adaptive downsampling means dynamically adjusting the sampling interval according to T1 instead of a fixed sampling. When T1 is larger, the sampling interval is smaller, retaining more image details; when T1 is smaller, the sampling interval is larger, and the image details are simplified. Through adaptive downsampling, the texture image of the first scale can be obtained. This image retains the main features of the original texture image, but the details are simplified. This is beneficial for extracting the fine-scale features of the texture.

[0094] Steps S1320 and S1330 are similar to S1310, but the thresholds used are the medium-scale decomposition threshold T2 and the large-scale decomposition threshold T3 respectively. Compared with T1, the values of T2 and T3 are smaller. This means that the sampling interval is larger and the degree of image simplification is higher. The obtained second-scale and third-scale texture images only retain the rough contours of the original texture, and the details are greatly simplified. They are respectively used to extract the medium-scale and large-scale features of the texture.

[0095] In the downsampling of the second scale and the third scale, the high-frequency information of the Laplacian pyramid decomposition is also fused. The Laplacian pyramid is a multi-resolution image representation method. It recursively downsamples and smooths the image to obtain a sequence of low-pass images (Gaussian pyramid). Subtracting the adjacent two-layer Gaussian images can obtain a difference image containing high-frequency information, that is, the Laplacian pyramid. The high-frequency information corresponds to the detailed features of the texture. Incorporating it into the second and third scale images can compensate for the details lost in downsampling, making the final multi-scale features more comprehensive.

[0096] The settings of T1, T2, and T3 are not directly related to the final sampling interval. They more play a role in adjusting the granularity of texture features at different scales. T1 is the largest, obtaining fine-scale texture features corresponding to the tiny structures in the texture image; T2 is the second largest, obtaining medium-scale texture features; T3 is the smallest, obtaining large-scale texture features corresponding to the large patterns in the texture image. The determination of the sampling interval also needs to consider factors such as the resolution of the original image and can be optimized through experiments.

[0097] Multi-scale analysis can characterize texture features at different granularities, capturing both the details and the overall situation of the texture. This provides richer and more reliable information for subsequent texture feature fusion and its quality evaluation, helping to comprehensively and accurately judge the texture quality. At the same time, adaptive downsampling avoids feature loss caused by improper selection of the sampling interval, improving the robustness of the algorithm. Laplacian pyramid decomposition can compensate for the high-frequency details lost during downsampling, making the final multi-scale features more complete.

[0098] Gabor texture features use Gabor wavelet transform to perform multi-scale and multi-directional texture analysis on images at different scales and directions, capturing the scale and direction information of image textures. Gabor wavelets are a class of wavelet functions that simulate the human visual system and are sensitive to the directionality and scale changes of textures. By setting Gabor wavelet kernels with different frequencies and directions, texture information with different granularities and orientations can be extracted. Gabor features can be represented as statistics of Gabor wavelet coefficients at different scales and directions, such as mean, standard deviation, energy, etc.

[0099] LBP (Local Binary Pattern) texture features are an effective local texture descriptor. It encodes the neighborhood information into a binary pattern by comparing the size relationship between a pixel and its neighboring pixels, and statistically analyzes the histogram of these binary patterns as texture features. LBP features have strong robustness to changes such as illumination and rotation and can capture the local texture information of images. By changing the radius and the number of sampling points of the LBP operator, LBP features at different scales can be extracted.

[0100] GLCM (Gray-Level Co-occurrence Matrix) is a classic texture analysis method for describing the spatial correlation of image gray levels. It statistically analyzes the frequency of simultaneous occurrence of different gray-level pixel pairs in a certain direction and distance in the image, forming a gray-level co-occurrence matrix. By extracting second-order statistics such as energy, entropy, and contrast from the GLCM, characteristics such as the roughness and regularity of image textures can be quantified. By changing the direction and step size of the GLCM, the spatial correlation of textures at different scales can be analyzed.

[0101] By integrating three classic texture analysis methods, namely Gabor, LBP, and GLCM, the multi-scale texture features of decorative paper can be comprehensively characterized from the perspectives of frequency domain, spatial domain, and statistics, providing rich discriminant information for subsequent quality assessment. By extracting these features in three scale spaces respectively, the variation law of the texture of decorative paper at different observation distances can be analyzed, and multi-granularity texture quality evaluation can be realized. The advantage of this method is that it integrates multiple texture description methods, can analyze the texture organization form on the paper surface in all directions, is sensitive to texture defects and differences, and provides a reliable feature representation for fine-grained quality inspection of decorative paper.

[0102] In step S1340, the texture images and texture features of the three scales are respectively summarized into a multi-scale texture image group and a multi-scale texture feature set to prepare for subsequent quality assessment. Multi-scale decomposition can simulate the process of the human eye observing the paper texture at different distances. The multi-scale texture image group intuitively shows the morphological changes of the paper surface pattern at different scales, and the multi-scale texture feature set quantifies these changes numerically, facilitating intelligent discrimination by the computer.

[0103] Step S1400: Evaluate the effect of the three-level scale decomposition and determine whether the multi-scale texture image group meets the feature extraction requirements;

[0104] Furthermore, step S1400 includes:

[0105] Step S1410: Calculate the structural similarity SSIM index and the scale space feature stability SSFS index between the first-scale texture image, the second-scale texture image, and the third-scale texture image and the high-definition texture image respectively, to obtain an SSIM index set and an SSFS index set;

[0106] The SSIM index set includes the SSIM indexes of the three scales. The SSIM indexes of the three scales are the first structural similarity SSIM index SSIM1, the second structural similarity SSIM index SSIM2, and the third structural similarity SSIM index SSIM3 respectively. Among them, SSIM1 is the structural similarity SSIM index between the first-scale texture image and the high-definition texture image, SSIM2 is the structural similarity SSIM index between the second-scale texture image and the high-definition texture image, and SSIM3 is the structural similarity SSIM index between the third-scale texture image and the high-definition texture image.

[0107] The SSFS index set includes SSFS indices at three scales, namely the first-scale spatial feature stability SSFS index SSFS1, the second-scale spatial feature stability SSFS index SSFS2, and the third-scale spatial feature stability SSFS index SSFS3; among them, SSFS1 is the scale-space feature stability SSFS index of the first-scale texture image and the high-definition texture image, SSFS2 is the scale-space feature stability SSFS index of the second-scale texture image and the high-definition texture image, and SSFS3 is the scale-space feature stability SSFS index of the third-scale texture image and the high-definition texture image.

[0108] The method for calculating the structural similarity SSIM index SSIM1 between the first-scale texture image and the high-definition texture image includes:

[0109]

[0110] Where:

[0111] represents the first-scale texture image;

[0112] represents the original high-definition texture image;

[0113] represents the mean value of;

[0114] represents the mean value of, and can be obtained by sliding a window over the image and calculating the average pixel value within each window;

[0115] represents the standard deviation of, represents the standard deviation of, which can also be calculated within the sliding window;

[0116] represents and the covariance of, reflecting the structural correlation between the two images, which can be obtained by calculating the mean value of the difference between the corresponding pixels of and within the sliding window;

[0117] and are small constants used to avoid the denominator being zero, usually taking the value of , , where is the dynamic range of pixel values;

[0118] represents the mean gradient magnitude of the first-scale texture image , represents the mean gradient magnitude of the original high-definition texture image , which reflects the sharpness of the texture and can be obtained by calculating the gradient magnitude using methods such as the Sobel operator and taking the average;

[0119] is an adjustment factor that controls the impact of gradient differences on the overall similarity and can be set according to experience.

[0120] Based on the original SSIM calculation of the three components of brightness, contrast, and structure, this formula introduces the gradient information of the texture image to form a sigmoid function correction term. When the texture sharpness of two images is close, the sigmoid function approaches 1 and has little impact; when the gradient difference is large, the sigmoid function will significantly reduce the value, indicating that there are obvious differences in the texture details of the two images.

[0121] Therefore, this formula comprehensively considers the similarity degree of images in terms of brightness, contrast, structural correlation, and texture details, and more comprehensively depicts the similarity between texture images of different scales and the original image. ranges from 0 to 1, and the closer it is to 1, the higher the structural similarity between the first-scale texture image and the original high-definition image. If exceeds the preset similarity threshold, it can be considered that the decomposition effect of the first scale is good and can better retain the structural information of the original texture.

[0122] This formula enhances the sensitivity of the SSIM index to texture detail changes, which helps to more finely evaluate the quality of multi-scale decomposition. At the same time, through the non-linear transformation of the sigmoid function, SSIM is more sensitive to large texture differences and responds more gently to subtle differences, which conforms to the non-linear characteristics of subjective visual evaluation. Therefore, the use of this formula can provide more targeted guidance for optimizing the effect of multi-scale texture decomposition.

[0123] The method for calculating the scale space feature stability SSFS index SSFS1 of the first-scale texture image and the high-definition texture image includes:

[0124]

[0125] Where:

[0126] Indicates the number of rows of the image, Indicates the number of columns of the image;

[0127] Represents the first scale texture image At pixel position The gradient amplitude at , Represents a high-definition texture image At pixel position The gradient amplitude at ; the gradient amplitude can be calculated using methods such as the Sobel operator;

[0128] Represents pixels The weight coefficient at is used to highlight the contribution of the texture salient area; It can be obtained through texture saliency detection algorithms, such as spectral residual method;

[0129] express The histogram distribution of express The histogram distribution of express and The correlation coefficient measures the similarity of their shapes.

[0130] To control the parameters, control the steepness of the sigmoid function;

[0131] is the correlation coefficient threshold, and It is set by those skilled in the art based on experience or a large number of experiments.

[0132] The first term in the formula is a weighted gradient similarity that measures the consistency of the two images in terms of texture edges and details. Pixels with smaller gradient magnitudes contribute more to the similarity.

[0133] The second sigmoid function describes the correlation between the grayscale distributions of the two images. The larger the correlation coefficient, the closer the sigmoid function output is to 1, indicating that the grayscale distributions of the two images are closer and the global appearance of the textures is more similar.

[0134] along with and The difference in gradient and grayscale distribution increases, The value of will drop significantly. Therefore, It can comprehensively evaluate the degree of approximation of the first-scale texture image to the original image in terms of local texture details and overall grayscale distribution, and provide a more fine-grained reference for optimizing the multi-scale decomposition strategy. Texture details and When they are basically the same and the gray-scale distributions are highly correlated, it approaches 1. Therefore, it can be set that the threshold value of (such as 0.85) is used as the standard for evaluating the scale decomposition effect.

[0135] This formula characterizes the feature stability of the scale space from two perspectives of local gradient information and global statistical features, can more comprehensively guide the optimization of multi-scale decomposition parameters, obtain a multi-scale representation highly consistent with the original texture, and provide more reliable texture features for subsequent quality assessment.

[0136] Step S1420: Determine whether the SSIM indices of the three scales are all greater than the SSIM similarity threshold Ts, and whether the SSFS indices of the three scales are all greater than the SSFS stability threshold Tf; if both are satisfied, it is considered that the multi-scale decomposition effect is ideal, and proceed to step S2000; if not, return to step S1200, adaptively adjust the scale factor in the three-level scale decomposition threshold calculation formula until the multi-scale decomposition effect is ideal.

[0137] Specifically, the SSIM index measures the similarity degree of two images in terms of brightness, contrast, and structure. Introducing it into multi-scale texture analysis can evaluate the fidelity of texture images at different scales. The higher the SSIM index, the better the structural similarity between the texture image and the original image at that scale, and the more successfully the multi-scale decomposition retains the original texture information.

[0138] The SSFS index is an evaluation index specifically designed for multi-scale texture analysis, used to measure the stability of texture features in the scale space. By calculating the weighted similarity of the gradient distributions of texture images at different scales and the correlation of gray-scale histograms, the consistency of texture features with scale changes is evaluated. The higher the SSFS index, the better the scale invariance of texture features, and the more stable the texture form is maintained during the process of magnification or reduction. This helps to enhance the scale adaptability of subsequent discriminant models.

[0139] Step S1420 performs threshold determination on the SSIM and SSFS index sets to determine whether the multi-scale decomposition effect is ideal. If the SSIM indices of the three scales all exceed the given threshold Ts, and the SSFS indices all exceed the threshold Tf, it is considered that the multi-scale texture image has a high structural similarity with the original image, the texture features have good scale stability, and the multi-scale decomposition achieves the expected effect. Otherwise, it is necessary to return to step S1200, adjust the calculation formula of the scale decomposition threshold, and through adaptive optimization of the scale factor

[0140] When the SSIM index of the three scales is not all greater than the SSIM similarity threshold Ts, or the SSFS index of the three scales is not all greater than the SSFS stability threshold Tf, it is necessary to adaptively adjust the scale factor in the formula for calculating the three-level scale decomposition threshold until the multi-scale decomposition effect is ideal. The advantage of this adaptive adjustment mechanism is that the texture characteristics of different decorative paper samples vary greatly, and fixed scale decomposition parameters are difficult to adapt to all situations. By evaluating the decomposition effect and dynamically optimizing the decomposition strategy, the optimal scale space representation can be automatically searched, which will neither lose texture details due to being too rough nor introduce redundant information due to being too fine. After optimized multi-scale decomposition, it can present the multi-level texture structure of decorative paper with appropriate spatial granularity, uncover the quality problems hidden on the paper surface, and lay a foundation for accurate discrimination.

[0141] For example, for decorative paper with fine patterns and complex textures, using a larger scale decomposition threshold can obtain a more detailed multi-scale texture image, capture more local texture details, and avoid overly rough texture features that cannot accurately locate defects. For decorative paper with rough textures and simple patterns, using a smaller scale decomposition threshold can meet the requirements and avoid increasing the computational burden due to overly trivial texture representations.

[0142] Step S2000, construct a texture quality attention model. According to the multi-scale texture feature set and the texture quality attention model, obtain a multi-scale attention weight vector set; based on the multi-scale attention weight vector set, perform weighted fusion on the first-scale texture feature, the second-scale texture feature, and the third-scale texture feature to obtain a multi-scale fusion feature set.

[0143] Furthermore, step S2000 includes:

[0144] Step S2100, construct a texture quality attention model;

[0145] Specifically, the texture quality attention model is a visual attention mechanism model based on deep learning. Its purpose is to learn the regions in the decorative paper texture image that are highly correlated with quality and assign higher weights to these regions to improve the accuracy of quality assessment. Through end-to-end training, the model automatically learns to extract discriminative texture features from a large amount of labeled data and optimizes the attention weight distribution to achieve fine-grained evaluation of the decorative paper texture quality.

[0146] First, collect a large number of decorative paper samples covering different quality grades and different texture types to form an original dataset. Invite industry experts to score the texture quality of each sample and mark the key texture areas affecting the quality to obtain quality scores and pixel-level attention labels. Then, perform data augmentation on the sample images to simulate deformations, stains, lighting changes, etc. that may occur during the actual production and use of the paper, improving the generalization ability of the model. And divide the training set and test set according to a certain ratio.

[0147] Next, design the network structure of the texture quality attention model. Using a classic convolutional neural network (CNN) as the backbone, two sub-networks are connected in parallel at the end of the network: one regresses the global texture quality score through a fully connected layer as the basis for quality evaluation; the other restores the feature map to the original resolution through upsampling and applies the Sigmoid activation function to output a pixel-level attention heat map, representing the influence weights of different regions on the quality.

[0148] In the training stage, input the sample images and their corresponding quality labels and attention labels into the model, continuously optimize the model parameters through backpropagation, minimize the quality regression error and attention prediction error, so that the model learns to automatically extract texture quality-related features and assign reasonable regional weights. After multiple rounds of iterative training, once the model converges, it can be used to evaluate the texture quality of decorative paper. Input a texture image of any size, the model can evaluate the quality score of the sample, and at the same time generate the corresponding attention heat map, intuitively presenting the contribution degree of different texture regions to the overall quality. This evaluation result based on the attention mechanism is more transparent and interpretable, which is conducive to discovering specific quality defects and has important guiding significance for the production process optimization and quality control of decorative paper.

[0149] In summary, the texture quality attention model mimics the human visual attention mechanism through deep learning algorithms and adaptively focuses on the regions closely related to quality in the decorative paper texture. Combining texture feature extraction with attention weight learning not only improves the accuracy of quality evaluation but also enhances the interpretability of the results. The construction of this model requires the support of rich decorative paper big data and industry expert knowledge, which has a certain threshold for industrial applications. However, in the long run, intelligent and fine-grained texture quality evaluation technology will surely become an important development direction in the field of decorative paper production and quality control. The texture quality attention model provides a feasible technical path to achieve this goal.

[0150] Step S2200, obtain a multi-scale attention weight vector set according to the multi-scale texture feature set and the texture quality attention model; the multi-scale attention weight vector set includes a first-scale attention weight vector, a second-scale attention weight vector, and a third-scale attention weight vector;

[0151] Furthermore, asFigure 3 As shown, step S2200 includes:

[0152] Step S2210, splicing the first-scale Gabor texture feature, the first-scale LBP texture feature, and the first-scale GLCM texture feature to form a first-scale comprehensive feature;

[0153] Step S2220, splicing the second-scale Gabor texture feature, the second-scale LBP texture feature, and the second-scale GLCM texture feature to form a second-scale comprehensive feature;

[0154] Step S2230, splicing the third-scale Gabor texture feature, the third-scale LBP texture feature, and the third-scale GLCM texture feature to form a third-scale comprehensive feature;

[0155] Step S2240, inputting the first-scale comprehensive feature into the texture quality attention model to obtain a first-scale attention heat map;

[0156] Step S2250, inputting the second-scale comprehensive feature into the texture quality attention model to obtain a second-scale attention heat map;

[0157] Step S2260, inputting the third-scale comprehensive feature into the texture quality attention model to obtain a third-scale attention heat map;

[0158] Step S2270, performing global average pooling operations on the first-scale attention heat map, the second-scale attention heat map, and the third-scale attention heat map to obtain a first-scale attention weight vector, a second-scale attention weight vector, and a third-scale attention weight vector.

[0159] Specifically, in step S1300, the texture features of the decorative paper sample at three scales have been extracted to form a multi-scale texture feature set. These features depict the texture tissue morphology of the paper surface from different angles and contain rich quality discrimination information. However, not all feature regions are crucial for quality assessment, and the importance of different regions varies. Therefore, it is necessary to further explore the inherent quality correlation based on the multi-scale texture features, highlight the features that have a greater impact on the discrimination process, and weaken redundant and interfering information. This is where the attention mechanism comes into play.

[0160] The texture quality attention model is an intelligent algorithm that can automatically learn the importance weights of features. By inputting multi-scale texture features into the trained attention model, the contribution of different texture regions at each scale to quality discrimination can be calculated, and the corresponding attention weight vector can be obtained. The term "attention" here borrows the concept of human visual attention, representing the selective attention and processing of different information by the visual system. In the task of evaluating the texture quality of decorative paper, the attention weight values obtained by texture features reflect the influence of the texture features in this region on the evaluation results.

[0161] From the perspective of information processing, the process of obtaining the multi-scale attention weight vector set can be understood as a process of feature selection and importance weighting. Through the "quality evaluation expert" - the attention model, the most discriminative key features are intelligently selected from a large number of multi-scale texture features, and weights are assigned to them to make them play a greater role in subsequent quality discrimination. This mechanism makes the process of evaluating the texture quality of decorative paper more efficient and accurate, overcoming the limitations of manual feature selection in traditional methods.

[0162] Gabor texture features, LBP texture features, and GLCM texture features describe the texture morphology from the frequency domain, spatial domain, and statistical perspectives at the same scale. The three complement each other and can more comprehensively depict the complex and variable textures of decorative paper. By concatenating the three features in the channel dimension, they are organically integrated into a comprehensive feature, which is equivalent to achieving integration at the feature level and can further improve the representation ability of features and their correlation with quality. The length of the comprehensive feature is equal to the sum of the lengths of the three texture features. This fusion method is simple and effective, and is of great benefit to downstream attention learning and quality discrimination.

[0163] For example, assume that the Gabor, LBP, and GLCM texture features at the first scale are 128-dimensional, 256-dimensional, and 32-dimensional vectors respectively. Then the comprehensive feature at the first scale after concatenation is a 416-dimensional vector. This 416-dimensional vector covers the feature information extracted by the three types of texture descriptors and more meticulously depicts the texture texture of the decorative paper at this scale. The subsequent attention model will also assign different weights to the 416 feature dimensions to make full use of the discriminative power of the comprehensive feature.

[0164] Steps S2240 - S2260 generate attention heatmaps at corresponding scales by respectively inputting the comprehensive features at three scales into the attention model. The attention heatmap is actually a grayscale image of the same size as the original texture image. The grayscale value at each pixel position in the image represents the importance of the texture features at that position for quality discrimination. The higher (brighter) the grayscale value, the greater the impact of the texture in that area on quality, and a higher weight should be assigned to it during the evaluation process; conversely, the lower (darker) the grayscale value, the less important the texture in that area is and the smaller its impact on the evaluation result.

[0165] From a visual perspective, the attention heatmap is equivalent to a distribution map of regional importance. It intuitively presents the quality correlation of different texture regions, making it clear at a glance. Through the heatmap, we can quickly locate the key textures in the decorative paper sample that have the greatest impact on quality, such as texture continuity, consistency, clarity, etc., providing a basis for judging quality defects. The attention model is trained with a large amount of labeled data, so the heatmap it generates can reflect the quality judgment ideas and focus of human experts, which is an intelligent way to learn expert experience from data and has a certain degree of interpretability.

[0166] It should be noted that since the attention heatmap is at the pixel level, its resolution is the same as that of the input texture image. However, the length of the texture feature vector is much smaller than the number of image pixels. This requires the attention model to upsample and spatially map the features when generating the heatmap, expanding the attention weights in the feature dimension to the pixel dimension. A common approach is to first map the feature vector to a two-dimensional matrix with the same height and width as the image through a fully connected layer, and then apply the Sigmoid function to normalize the attention weights to between 0 and 1 to obtain a heatmap with the same size as the original image.

[0167] Functionally, the process of generating attention heatmaps by steps S2240 - S2260 through the comprehensive features at three scales realizes the mapping from multi-scale texture features to regional quality importance. This importance-guided feature enhancement method makes the originally equally treated texture features have discriminative differences, providing a more explicit basis for subsequent quality assessment decisions. The visualization results of the multi-scale attention heatmaps also provide an intuitive and quantitative reference for quality analysis and process improvement in decorative paper production.

[0168] Step S2270 is to convert the pixel-level two-dimensional attention heat map into a one-dimensional attention weight vector with the same length as the texture feature vector, preparing for subsequent feature weighting. Global average pooling operation is a commonly used method for feature compression and information aggregation. It calculates the average value of each channel of the input feature map and outputs a one-dimensional vector with the same number of channels as the input. The optimized result is equivalent to statistically calculating the overall importance of each region in terms of channels, which not only reduces the subsequent computational complexity but also avoids the interference of regional outliers.

[0169] For example, the size of the first-scale attention heat map is 1024×1024 with a total of 416 channels. Performing global average pooling on it means averaging the attention weights of 1024×1024 pixels within each channel, and finally obtaining a 416-dimensional first-scale attention weight vector. Each element of the vector represents the average quality importance of the texture region corresponding to that dimension in the first-scale comprehensive feature. This vector is exactly the same length as the first-scale comprehensive feature, and the corresponding weights can be assigned to each texture feature dimension by element-wise multiplication, highlighting the significant quality-related regions and weakening the relatively less important regions. This weighted feature vector contains more refined quality discrimination information.

[0170] The generation processes of the second and third-scale attention weight vectors are similar. The three finally obtained weight vectors together form a multi-scale attention weight vector set. This set comprehensively depicts the quality importance distribution of the texture regions of the decorative paper at three scales, providing a quantitative weight reference for subsequent multi-scale quality feature fusion. Using the attention weight vector as a bridge, the texture features are mapped from the thousands-dimensional pixel space to the semantic space highly related to quality, realizing the sublimation and optimization of feature representation, which is the key to improving the quality assessment performance.

[0171] In summary, step S2200 and its sub-steps use the attention model to represent the quality correlation learning of multi-scale texture features as attention weight vectors, and use global average pooling to achieve feature compression from the pixel-level attention heat map to the regional-level weight vector. This series of operations makes the originally homogeneous massive texture features have discriminative differences, providing an importance-guided feature selection mechanism for quality assessment. Automatically learning the quality correlation weights from texture features overcomes the limitations of traditional manual design and is a typical case of applying computer vision technology to the quality assessment of decorative paper. Introducing the visual attention mechanism not only improves the objectivity and accuracy of the decorative paper quality assessment but also provides a new idea for the interpretability of the assessment results. The online inspection of the quality of decorative paper for intelligent manufacturing will surely benefit from this attention-guided multi-scale texture analysis.

[0172] Step S2300, based on the multi-scale attention weight vector set, perform weighted fusion on the first-scale texture feature, the second-scale texture feature, and the third-scale texture feature to obtain a multi-scale fusion feature;

[0173] Further, as Figure 4 shown, step S2300 includes:

[0174] Step S2310, according to the first-scale attention weight vector and the first-scale texture feature, obtain a first-scale fusion feature;

[0175] Further, as Figure 5 shown, step S2310 includes:

[0176] Step S2311, multiply the first-scale attention weight vector by the first-scale Gabor texture feature to obtain a first-scale Gabor weighted feature;

[0177] Step S2312, multiply the first-scale attention weight vector by the first-scale LBP texture feature to obtain a first-scale LBP weighted feature;

[0178] Step S2313, multiply the first-scale attention weight vector by the first-scale GLCM texture feature to obtain a first-scale GLCM weighted feature;

[0179] Step S2314, fuse the first-scale Gabor weighted feature, the first-scale LBP weighted feature, and the first-scale GLCM weighted feature to obtain a first-scale fusion feature.

[0180] Step S2320, according to the second-scale attention weight vector and the second-scale texture feature, obtain a second-scale fusion feature;

[0181] Further, as Figure 6 shown, step S2320 includes:

[0182] Step S2321, multiply the second-scale attention weight vector by the second-scale Gabor texture feature to obtain a second-scale Gabor weighted feature;

[0183] Step S2322, multiply the second-scale attention weight vector by the second-scale LBP texture feature to obtain a second-scale LBP weighted feature;

[0184] Step S2323, multiply the second-scale attention weight vector by the second-scale GLCM texture feature to obtain a second-scale GLCM weighted feature;

[0185] Step S2324: Fuse the second-scale Gabor weighted feature, the second-scale LBP weighted feature, and the second-scale GLCM weighted feature to obtain the second-scale fused feature.

[0186] Step S2330: Obtain the third-scale fused feature based on the third-scale attention weight vector and the third-scale texture feature;

[0187] Further, as Figure 7 shown, Step S2330 includes:

[0188] Step S2331: Multiply the third-scale attention weight vector by the third-scale Gabor texture feature to obtain the third-scale Gabor weighted feature;

[0189] Step S2332: Multiply the third-scale attention weight vector by the third-scale LBP texture feature to obtain the third-scale LBP weighted feature;

[0190] Step S2333: Multiply the third-scale attention weight vector by the third-scale GLCM texture feature to obtain the third-scale GLCM weighted feature;

[0191] Step S2334: Fuse the third-scale Gabor weighted feature, the third-scale LBP weighted feature, and the third-scale GLCM weighted feature to obtain the third-scale fused feature.

[0192] Step S2340: Concatenate the first-scale fused feature, the second-scale fused feature, and the third-scale fused feature to generate the multi-scale fused feature.

[0193] Specifically, Step S2300 introduces an attention mechanism to adaptively weight and fuse the multi-scale texture features, generating discriminative multi-scale fused features. This step comprehensively considers the importance differences of texture features at different scales and adjusts their contributions to the final quality assessment result by assigning weights.

[0194] In step S2310, the fine-scale Gabor, LBP, and GLCM texture features are weighted using the attention weight vector of the first scale. Specifically, in step S2311, the first-scale attention weight vector is multiplied element-wise with the first-scale Gabor texture features to obtain the first-scale Gabor weighted features. Here, the Gabor features are multi-scale and multi-directional texture descriptors extracted through Gabor wavelet transform, which are sensitive to the frequency-domain characteristics of the texture. Multiplying by the attention weight is equivalent to adjusting the importance of the Gabor features at different frequencies and directions. Steps S2312 and S2313 multiply the first-scale attention weight vector with the LBP features and GLCM features in a similar manner to obtain the corresponding weighted features. The LBP features characterize the spatial patterns of the pixel neighborhood, and the GLCM features describe the second-order statistical characteristics of the gray-level co-occurrence matrix. By assigning different weights to these features, certain texture patterns can be selectively emphasized or suppressed, improving the representational ability of the features. Step S2314 concatenates the three weighted features to form the first-scale fusion feature. This feature synthesizes the frequency-domain, spatial-domain, and statistical characteristics of the texture at the fine scale and has rich texture discriminative information.

[0195] Steps S2320 and S2330, in a similar manner to S2310, respectively use the attention weight vectors of the second scale and the third scale to weight and fuse the medium-scale and large-scale texture features. Among them, steps S2321 - S2324 generate the second-scale fusion feature, and steps S2331 - S2334 generate the third-scale fusion feature. The fusion features at these two scales respectively focus on the performance of the paper texture at medium and long observation distances, capturing the hierarchical changes in the texture morphology.

[0196] Finally, step S2340 concatenates the fusion features of the three scales to generate the final multi-scale fusion feature. This feature covers the key information of the surface texture of the decorative paper in different scale spaces and has the ability to analyze the texture differences at multiple granularities. By organically combining the texture characteristics at the fine, medium, and large scales, an information-rich and discriminative feature representation is constructed, laying a solid foundation for subsequent quality assessment.

[0197] The advantage of introducing the attention mechanism for multi-scale feature fusion is that it can adaptively adjust the importance of features at different scales, highlight the texture information beneficial to quality discrimination, and suppress the influence of interference factors. Traditional multi-scale feature fusion often uses simple concatenation or averaging, ignoring the importance differences between features. The attention-weighted fusion, on the other hand, quantitatively describes the discriminative contributions of features at each scale through the learned weight vector, making the fusion process more intelligent. The magnitude of the weight reflects the relevance of the texture features at that scale to the quality assessment task, providing a useful reference for feature selection. At the same time, since the attention weights are adaptively generated based on the texture images, the fusion method has a certain degree of robustness and can handle decorative paper samples of different types and qualities.

[0198] For example, for a high-grade decorative paper with delicate texture and clear pattern, the texture features at the fine scale may contain more information helpful for quality discrimination. Therefore, the corresponding element value in the first-scale attention weight vector will be larger, highlighting the role of this part of the features in the fused features. For ordinary decorative paper with rough texture and bold pattern, the features at the medium and large scales may be more discriminative, and the corresponding weight values will increase. This adaptive weighted fusion method can flexibly adjust the contributions of features at different levels according to the characteristics of the decorative paper texture, improving the accuracy and generalization ability of quality assessment.

[0199] In summary, step S2300 realizes the adaptive weighted fusion of multi-scale texture features through the attention mechanism, constructing a fused feature that takes into account both fine and macroscopic texture characteristics. This fusion method fully utilizes the discriminative information at different observation scales, enhancing the feature representation ability. At the same time, by quantifying the importance weights of features at each scale, the fusion process is more intelligent and flexible, and can adapt to decorative paper samples of different quality levels. The multi-scale fused feature provides a reliable basis for subsequent quality assessment, helping to improve the accuracy and efficiency of decorative paper quality inspection. In practical applications, the calculation method of the attention weights can be optimized according to factors such as the type and quality requirements of the decorative paper to further enhance the discriminative power and applicability of the fused feature.

[0200] Step S3000, based on the multi-scale fused feature, determines the quality of the decorative paper sample to be detected, and obtains the quality determination level of the decorative paper sample to be detected.

[0201] Furthermore, as Figure 8 shown, step S3000 includes:

[0202] Step S3100, construct a lightweight texture feature extraction network, input the multi-scale fused feature into the lightweight texture feature extraction network, and obtain a multi-scale lightweight fused feature;

[0203] Specifically, the lightweight texture feature extraction network is a convolutional neural network model specifically designed to extract multi-scale texture features of decorative paper. Compared with traditional deep convolutional networks, this network significantly reduces the number of model parameters and computational complexity while ensuring the feature extraction performance, and is more suitable for deployment on quality inspection equipment of the decorative paper production line for real-time detection.

[0204] The construction process of this network is as follows: First, based on classic lightweight convolutional neural network architectures (such as MobileNet, ShuffleNet, etc.), customized improvements are made according to the texture characteristics of decorative paper and quality inspection requirements. The network mainly consists of several convolutional layers, pooling layers, and fully connected layers. The settings of the convolutional kernel size and number comprehensively consider factors such as texture scale, directionality, and computational efficiency. Secondly, a multi-scale feature fusion module is introduced in the middle layer of the network. Through operations such as cross-layer connection and feature splicing, feature maps with different receptive fields are aggregated to capture the diverse features of the paper surface texture at different scales. Finally, several 1×1 convolutional layers are set at the end of the network for dimensionality reduction and feature compression to obtain compact multi-scale lightweight fusion features.

[0205] The training process of the network adopts an end-to-end supervised learning method. First, the high-definition original texture images of decorative paper are used as inputs, and at the same time, the quality grade labels manually annotated are used as supervision signals. Then, the original images are used to generate a multi-scale texture image group through a preprocessing module, and the corresponding multi-scale texture feature set is extracted. Next, the multi-scale texture features are input into the lightweight texture feature extraction network. After multiple convolutional, pooling, and feature fusion operations, high-level semantic multi-scale lightweight fusion features are extracted. Finally, the fusion features are input into the quality classifier, mapped to the quality grade categories through the softmax layer, and with the cross-entropy loss function as the optimization target, the network weights are updated using the backpropagation algorithm to achieve the joint optimization of the end-to-end feature extractor and classifier.

[0206] A large number of decorative paper samples of different types and quality grades are used as training data. The training set is augmented through data augmentation techniques (such as rotation, flipping, noise addition, etc.) to improve the generalization performance of the model. During the training process, with the multi-scale texture image group as the input and the corresponding quality grade labels as the output, the difference between the predicted quality grade and the true quality grade is minimized through iterative optimization, enabling the network to learn a robust multi-scale texture feature representation and establish a mapping relationship between the features and the quality grades.

[0207] The advantages of using a lightweight texture feature extraction network are as follows: First, through customized network design and multi-scale feature fusion, it can efficiently and accurately extract the key features of the texture on the surface of decorative paper, providing a reliable basis for quality discrimination. Second, by using the end-to-end learning method, the cumbersome process of manually designing features is avoided, and the feature extractor can automatically learn discriminative texture features from a large amount of data. Moreover, the lightweight network structure can significantly reduce the consumption of computing resources, speed up the detection speed, and is more conducive to industrial deployment applications.

[0208] For example, for a wood grain decorative paper, texture images at different granularities can be obtained through multi-scale decomposition, from the overall wood grain distribution to local wood grain details. Inputting these images into the lightweight texture feature extraction network, the network can automatically learn the key features of the wood grain, such as the density, arrangement direction, and thickness variation of the texture. Through feature fusion, the network can integrate the feature information at different scales, forming a more comprehensive and refined feature description of the wood grain texture, thus providing a more reliable basis for subsequent quality discrimination.

[0209] The multi-scale lightweight fusion features extracted by the lightweight texture feature extraction network not only contain the structural information of the original texture image at different scales but also incorporate the abstract semantic information learned by the network, providing a more comprehensive characterization of the factors affecting the quality of decorative paper. These fusion features can be used as the input of the quality discrimination classifier to make an overall determination of the quality level of decorative paper by integrating texture features from multiple aspects, improving the accuracy and reliability of quality inspection.

[0210] Step S3200: Obtain the quality determination level and the quality determination score corresponding to the quality determination level according to the multi-scale lightweight fusion features and the pre-constructed quality discrimination classifier;

[0211] Specifically, the quality discrimination classifier is an intelligent model based on machine learning, used to discriminate the quality level of decorative paper according to the input texture features. Common classifier models include Support Vector Machine (SVM), Random Forest, Neural Network, etc. Here, SVM is selected as the discrimination classifier, mainly for the following considerations:

[0212] SVM is a binary classification model based on statistical learning theory. It separates samples of different classes by finding the optimal classification hyperplane, with a good theoretical basis and extensive application practice.

[0213] SVM uses the Kernel Trick to transform the non-linear problem into a linear problem for solution. By introducing the kernel function, it can handle high-dimensional features and has good applicability in the field of texture analysis.

[0214] SVM only depends on support vectors (a small number of samples near the classification boundary), is less interfered by other samples, has a lower training complexity, and is easy to implement quickly.

[0215] Compared with neural networks, SVM has less dependence on parameter tuning, stronger generalization ability, and is more suitable for training with medium and small samples.

[0216] When constructing the SVM quality discrimination classifier, the multi-scale lightweight fusion features are used as the input, and the pre-calibrated quality grades of decorative papers are used as the output. The classifier is trained using training samples. The training samples include multi-scale texture images of several decorative papers and the corresponding manually calibrated quality grades. The multi-scale lightweight fusion feature vectors of each sample image are obtained through feature extraction and form training data pairs with the corresponding quality grade labels. The training data pairs are randomly divided into a training set and a validation set. The training set is used for model training, and the validation set is used to evaluate the model performance.

[0217] The training process adopts the method of cross-validation. The training set is divided into K subsets again. Each time, K - 1 of these subsets are selected for training, and the remaining 1 subset is used for validation. This process is repeated K times, and the average value of the K validation results is taken as the model performance index. This cross-validation can effectively utilize the limited training samples and reduce the risk of overfitting. The goal of model training is to find the optimal classification hyperplane parameters to minimize the classification error of the training samples while taking into account the maximum margin principle of the classification boundary.

[0218] The key parameters of SVM (such as kernel function type, penalty coefficient, kernel function parameters, etc.) are tuned through methods such as grid search to select the parameter combination with the optimal classification performance. The performance of the trained model is evaluated using the validation set, and indicators such as the confusion matrix, precision, recall, and F1 score are calculated to comprehensively measure the accuracy and robustness of the classifier. If the performance meets the expectations, it is determined as the final quality discrimination classification model; if the performance is not ideal, the features need to be adjusted or the model needs to be optimized until the requirements are met.

[0219] When applying the trained SVM quality discrimination classifier to the quality inspection of decorative papers, the multi-scale lightweight fusion features of the to-be-inspected decorative paper samples are used as the input. After the classifier inference, the corresponding quality determination grades (such as first-class products, qualified products, defective products, etc.) and quality determination scores are output. The quality determination score is a real number between 0 and 1, indicating the confidence or possibility that the sample belongs to this quality grade. The higher the score, the more likely the sample belongs to this grade; the lower the score, the less likely the sample belongs to this grade.

[0220] For example, assume that the quality discrimination classifier classifies decorative papers into three grades: A, B, and C, where A is the best grade, B is the qualified grade, and C is the unqualified grade. For a sample of decorative paper to be inspected, its lightweight fusion feature vector is extracted through multi-scale texture analysis and input into the trained SVM classifier. The output results are as follows: grade A with a score of 0.8, grade B with a score of 0.15, and grade C with a score of 0.05. This indicates that the sample is most likely to belong to grade A with a determination score of 0.8; secondly, it may belong to grade B with a determination score of 0.15; and it is least likely to belong to grade C with a determination score of only 0.05. Therefore, it can be determined that the quality grade of this decorative paper sample is A, and the quality is relatively good.

[0221] This quality determination method based on multi-scale texture features and machine learning classifiers can quickly and accurately discriminate and score the quality grades of decorative papers, reduce the subjectivity and instability of manual evaluation, and improve the objectivity and repeatability of quality determination. At the same time, the results of the grading determination are also more detailed and quantitative. It not only gives the quality grade but also gives the grade confidence score, which is convenient for subsequent quality analysis and improvement. For samples with unqualified quality, the reasons for quality defects can also be traced by analyzing the differences in their multi-scale texture features from standard samples, guiding the optimization of the production process.

[0222] In summary, this method makes full use of the advantages of multi-scale texture analysis and machine learning in the quality inspection of decorative papers, forming an intelligent, efficient, and accurate quality determination system, providing important technical support for improving the quality of decorative paper products.

[0223] Step S3300: Determine whether the quality determination score corresponding to the quality determination grade meets the preset quality grade threshold. If it meets, output the final quality determination grade; if it does not meet, output a warning signal and trigger the fault diagnosis process.

[0224] Specifically, the quality grade threshold refers to the acceptable minimum quality determination score set by the enterprise for decorative papers of different quality grades according to its own quality standards and customer requirements. When the quality determination score of a sample is higher than or equal to the quality threshold of this grade, it is considered that the sample meets the quality requirements of this grade; conversely, if the determination score is lower than the quality threshold, it means that the quality of the sample does not meet the standard and warnings and diagnoses are required.

[0225] Taking the three quality grades A, B, and C in the above example as an example, assume that the quality grade thresholds set by the enterprise are: grade A is 0.9, grade B is 0.7, and grade C is 0.6. This means that for a sample judged to be grade A, its quality judgment score needs to reach 0.9 or above to be considered a qualified grade A product; for a sample judged to be grade B, the quality judgment score needs to be between 0.7 and 0.9; and for a sample judged to be grade C, if the quality judgment score is lower than 0.6, it is a non-conforming product and needs to be rejected.

[0226] In actual production, the quality grade thresholds in the enterprise quality standard database can be input or retrieved through the human-machine interaction interface, and compared with the quality judgment score output in step S3200 in real time. If the judgment score of the sample meets the quality threshold of its judgment grade, the discrimination result is output, and the sample is classified into the corresponding quality grade; if the judgment score is lower than the quality threshold, an early warning signal is output to indicate that the quality of the sample is unqualified and further inspection and diagnosis are required.

[0227] The output discrimination result can be presented in a visual way. For example, samples of different quality grades are marked with different colors on the detection interface, and their judgment scores and grade labels are displayed, which is convenient for operators to intuitively understand the quality distribution of the samples. For unqualified samples that trigger an early warning, the system can automatically generate a quality report, record parameters such as the multi-scale texture features, judgment scores, and thresholds of the samples, and mark possible defect causes, such as uneven texture density, texture breakage, color abnormality, etc., providing a basis for quality diagnosis.

[0228] When the early warning signal is triggered, the fault diagnosis process can be automatically started. The purpose of this process is to analyze the reasons for the unqualified quality of the sample and propose corrective measures. The diagnosis process may include the following steps:

[0229] Isolate and label the unqualified samples to prevent them from flowing into the subsequent processes.

[0230] Collect high-definition images of the unqualified samples through industrial cameras or high-resolution scanners for fine-grained multi-scale texture analysis.

[0231] Compare the multi-scale texture features of the unqualified samples with standard samples and historical unqualified samples to find abnormal points in the feature space and infer possible defect patterns.

[0232] Combined with the technological process and production parameters of the decorative paper, analyze the links and reasons for introducing defects, such as raw material selection, slurry ratio, weaving and embossing, drying and shaping, etc.

[0233] According to the defect reasons, propose corrective and preventive measures, such as adjusting process parameters, replacing or maintaining equipment, improving raw material quality control, etc.

[0234] Feedback the diagnosis results and improvement measures to the production management system, and monitor the key process parameters and equipment status to prevent the recurrence of similar defects.

[0235] Through the fault diagnosis process, quality problems in decorative paper production can be quickly discovered and located, and corresponding measures can be taken in a timely manner to reduce the generation of unqualified products and improve the consistency and stability of product quality. At the same time, by accumulating the characteristic data of unqualified samples and diagnosis cases, a knowledge base of decorative paper quality defects can be established to provide empirical support for future quality determination and diagnosis.

[0236] In summary, step S3300 realizes the automatic determination, early warning and improvement closed-loop of decorative paper quality by introducing the quality grade threshold and the fault diagnosis process, forming a complete quality control system. This method gives full play to the advantages of multi-scale texture analysis and machine learning in quality detection, combines the actual quality standards of the enterprise, and conducts refined, digital and intelligent management of product quality, which helps to improve the production efficiency and quality level of decorative paper and enhance the comprehensive strength of the enterprise in the market competition.

[0237] Embodiment 2

[0238] Based on Embodiment 1, this embodiment provides an on-line quality discrimination system for decorative paper, as Figure 9 shown, including:

[0239] Image decomposition module: used to collect high-definition texture images of the decorative paper samples to be detected, preset three-level scale decomposition thresholds, and perform three-level scale decomposition and feature extraction on the high-definition texture images according to the preset three-level scale decomposition thresholds to obtain a multi-scale texture image group and a multi-scale texture feature set; the multi-scale texture feature set includes first-scale texture features, second-scale texture features and third-scale texture features;

[0240] Decomposition effect evaluation module: used to evaluate the effect of three-level scale decomposition;

[0241] Feature fusion module: used to construct a texture quality attention model, and obtain a multi-scale attention weight vector set according to the multi-scale texture feature set and the texture quality attention model; based on the multi-scale attention weight vector set, perform weighted fusion on the first-scale texture features, second-scale texture features and third-scale texture features to obtain a multi-scale fusion feature set;

[0242] Quality determination module: perform quality determination on the decorative paper samples to be detected according to the multi-scale fusion features to obtain the quality determination grade of the decorative paper samples to be detected.

[0243] In the image decomposition module, the preset three-level scale decomposition thresholds include:

[0244] Introduce a scale factor, and establish a calculation formula for the three-level scale decomposition threshold according to the size, resolution Res, and scale factor of the high-definition texture image, and calculate the three-level scale decomposition threshold, where the three-level scale decomposition thresholds are the fine-scale decomposition threshold T1, the medium-scale decomposition threshold T2, and the large-scale decomposition threshold T3; the size of the high-definition texture image includes the image width Wid and the image height Height.

[0245] The calculation formula for the three-level scale decomposition threshold is as follows:

[0246]

[0247] where represents the scale level, and its value is 1, 2, 3; is the scale factor, the initial value of k1 is 1, the initial value of k2 is 4, and the initial value of k3 is 16.

[0248] In the image decomposition module, the three-level scale decomposition and feature extraction of the high-definition texture image include:

[0249] Step S1310: Perform Gaussian smoothing on the high-definition texture image, and then perform adaptive downsampling according to the fine-scale decomposition threshold T1 to generate the first-scale texture image; and extract the first-scale texture features from the first-scale texture image; the first-scale texture features include the first-scale Gabor texture feature, the first-scale LBP texture feature, and the first-scale GLCM texture feature;

[0250] Step S1320: Perform Gaussian smoothing on the first-scale texture image, perform adaptive downsampling according to the medium-scale decomposition threshold T2, and fuse the high-frequency information of the Laplacian pyramid decomposition to generate the second-scale texture image; and extract the second-scale texture features from the second-scale texture image; the second-scale texture features include the second-scale Gabor texture feature, the second-scale LBP texture feature, and the second-scale GLCM texture feature;

[0251] Step S1330: Perform Gaussian smoothing on the second-scale texture image, perform adaptive downsampling according to the large-scale decomposition threshold T3, and fuse the high-frequency information of the Laplacian pyramid decomposition to generate the third-scale texture image; and extract the third-scale texture features from the third-scale texture image; the third-scale texture features include the third-scale Gabor texture feature, the third-scale LBP texture feature, and the third-scale GLCM texture feature;

[0252] Step S1340: Combine the first-scale texture image, the second-scale texture image, and the third-scale texture image to form a multi-scale texture image group; combine the first-scale texture feature, the second-scale texture feature, and the third-scale texture feature to form a multi-scale texture feature set.

[0253] In the decomposition effect evaluation module, the evaluation of the three-level scale decomposition effect includes:

[0254] Step S1410: Calculate the structural similarity SSIM index and the scale space feature stability SSFS index between the first-scale texture image, the second-scale texture image, and the third-scale texture image and the high-definition texture image respectively, to obtain an SSIM index set and an SSFS index set.

[0255] The SSIM index set includes SSIM indexes of three scales. The SSIM indexes of the three scales are the first structural similarity SSIM index SSIM1, the second structural similarity SSIM index SSIM2, and the third structural similarity SSIM index SSIM3 respectively. Among them, SSIM1 is the structural similarity SSIM index between the first-scale texture image and the high-definition texture image, SSIM2 is the structural similarity SSIM index between the second-scale texture image and the high-definition texture image, and SSIM3 is the structural similarity SSIM index between the third-scale texture image and the high-definition texture image.

[0256] The SSFS index set includes SSFS indexes of three scales. The SSFS indexes of the three scales are the first scale space feature stability SSFS index SSFS1, the second scale space feature stability SSFS index SSFS2, and the third scale space feature stability SSFS index SSFS3 respectively. Among them, SSFS1 is the scale space feature stability SSFS index between the first-scale texture image and the high-definition texture image, SSFS2 is the scale space feature stability SSFS index between the second-scale texture image and the high-definition texture image, and SSFS3 is the scale space feature stability SSFS index between the third-scale texture image and the high-definition texture image.

[0257] Step S1420: Judge whether the SSIM indexes of the three scales are all greater than the SSIM similarity threshold Ts, and whether the SSFS indexes of the three scales are all greater than the SSFS stability threshold Tf. If both are satisfied, it is considered that the multi-scale decomposition effect is ideal, and enter the subsequent steps; if not, adaptively adjust the scale factor in the three-level scale decomposition threshold calculation formula until the multi-scale decomposition effect is ideal.

[0258] In the feature fusion module, the multi-scale attention weight vector set includes a first-scale attention weight vector, a second-scale attention weight vector, and a third-scale attention weight vector;

[0259] The obtaining of the multi-scale attention weight vector set includes:

[0260] Step S2210: Concatenate the first-scale Gabor texture feature, the first-scale LBP texture feature, and the first-scale GLCM texture feature to form a first-scale comprehensive feature;

[0261] Step S2220: Concatenate the second-scale Gabor texture feature, the second-scale LBP texture feature, and the second-scale GLCM texture feature to form a second-scale comprehensive feature;

[0262] Step S2230: Concatenate the third-scale Gabor texture feature, the third-scale LBP texture feature, and the third-scale GLCM texture feature to form a third-scale comprehensive feature;

[0263] Step S2240: Input the first-scale comprehensive feature into the texture quality attention model to obtain a first-scale attention heat map;

[0264] Step S2250: Input the second-scale comprehensive feature into the texture quality attention model to obtain a second-scale attention heat map;

[0265] Step S2260: Input the third-scale comprehensive feature into the texture quality attention model to obtain a third-scale attention heat map;

[0266] Step S2270: Perform global average pooling operations on the first-scale attention heat map, the second-scale attention heat map, and the third-scale attention heat map to obtain a first-scale attention weight vector, a second-scale attention weight vector, and a third-scale attention weight vector.

[0267] In the feature fusion module, the obtaining of the multi-scale fusion feature includes:

[0268] Step S2310: Obtain a first-scale fusion feature according to the first-scale attention weight vector and the first-scale texture feature;

[0269] Step S2320: Obtain a second-scale fusion feature according to the second-scale attention weight vector and the second-scale texture feature;

[0270] Step S2330: Obtain a third-scale fusion feature according to the third-scale attention weight vector and the third-scale texture feature;

[0271] Step S2340: Concatenate the first-scale fusion feature, the second-scale fusion feature, and the third-scale fusion feature to generate a multi-scale fusion feature.

[0272] The said step S2310 includes:

[0273] Step S2311: Multiply the first-scale attention weight vector by the first-scale Gabor texture feature to obtain the first-scale Gabor weighted feature;

[0274] Step S2312: Multiply the first-scale attention weight vector by the first-scale LBP texture feature to obtain the first-scale LBP weighted feature;

[0275] Step S2313: Multiply the first-scale attention weight vector by the first-scale GLCM texture feature to obtain the first-scale GLCM weighted feature;

[0276] Step S2314: Fuse the first-scale Gabor weighted feature, the first-scale LBP weighted feature, and the first-scale GLCM weighted feature to obtain the first-scale fused feature.

[0277] The said step S2320 includes:

[0278] Step S2321: Multiply the second-scale attention weight vector by the second-scale Gabor texture feature to obtain the second-scale Gabor weighted feature;

[0279] Step S2322: Multiply the second-scale attention weight vector by the second-scale LBP texture feature to obtain the second-scale LBP weighted feature;

[0280] Step S2323: Multiply the second-scale attention weight vector by the second-scale GLCM texture feature to obtain the second-scale GLCM weighted feature;

[0281] Step S2324: Fuse the second-scale Gabor weighted feature, the second-scale LBP weighted feature, and the second-scale GLCM weighted feature to obtain the second-scale fused feature.

[0282] The said step S2330 includes:

[0283] Step S2331: Multiply the third-scale attention weight vector by the third-scale Gabor texture feature to obtain the third-scale Gabor weighted feature;

[0284] Step S2332: Multiply the third-scale attention weight vector by the third-scale LBP texture feature to obtain the third-scale LBP weighted feature;

[0285] Step S2333: Multiply the third-scale attention weight vector by the third-scale GLCM texture feature to obtain the third-scale GLCM weighted feature;

[0286] Step S2334: Fuse the third-scale Gabor weighted features, the third-scale LBP weighted features, and the third-scale GLCM weighted features to obtain the third-scale fused features.

[0287] In the quality determination module, the quality determination of the to-be-detected decorative paper sample includes:

[0288] Step S3100: Construct a lightweight texture feature extraction network, and input the multi-scale fused features into the lightweight texture feature extraction network to obtain multi-scale lightweight fused features;

[0289] Step S3200: According to the multi-scale lightweight fused features and the pre-constructed quality discrimination classifier, obtain the quality determination level and the quality determination score corresponding to the quality determination level;

[0290] Step S3300: Determine whether the quality determination score corresponding to the quality determination level meets the preset quality level threshold. If it meets, output the final quality determination level. If it does not meet, output a warning signal and trigger a fault diagnosis process.

[0291] Embodiment 3

[0292] This embodiment discloses an electronic device, which may include one or more processors and one or more memories. Among them, computer-readable code is stored in the memory, and when the computer-readable code is run by one or more processors, it can execute the above-mentioned online quality discrimination algorithm for decorative paper.

[0293] The method or system according to the embodiments of the present application can also be implemented by means of the architecture of an electronic device. The electronic device may include a bus, one or more CPUs, a read-only memory (ROM), a random access memory (RAM), a communication port connected to a network, input / output components, a hard disk, etc. The storage device in the electronic device, such as ROM or hard disk, can store the online quality discrimination algorithm for decorative paper provided by the present application. The online quality discrimination algorithm for decorative paper may include, for example: collecting high-definition texture images of the decorative paper samples to be detected, presetting three-level scale decomposition thresholds, and performing three-level scale decomposition and feature extraction on the high-definition texture images according to the preset three-level scale decomposition thresholds to obtain a multi-scale texture image group and a multi-scale texture feature set; evaluating the effect of the three-level scale decomposition; the multi-scale texture feature set includes first-scale texture features, second-scale texture features, and third-scale texture features; constructing a texture quality attention model, and obtaining a multi-scale attention weight vector set according to the multi-scale texture feature set and the texture quality attention model; based on the multi-scale attention weight vector set, performing weighted fusion on the first-scale texture features, second-scale texture features, and third-scale texture features to obtain a multi-scale fusion feature set; and performing quality determination on the decorative paper samples to be detected according to the multi-scale fusion features to obtain the quality determination level of the decorative paper samples to be detected.

[0294] Further, the electronic device may further include a user interface. Of course, the architecture disclosed in the present invention is only exemplary. When implementing different devices, one or more components in the electronic device disclosed in the present invention may be omitted according to actual needs.

[0295] Embodiment 4

[0296] This embodiment discloses a computer-readable storage medium, on which computer-readable instructions are stored. When the computer-readable instructions are run by a processor, the online quality discrimination algorithm for decorative paper according to the embodiments of the present application can be executed. The storage medium includes, but is not limited to, for example, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory, etc. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0297] In addition, according to the embodiments of the present application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, the present application provides a non-transitory machine-readable storage medium storing machine-readable instructions that can be run by a processor to execute instructions corresponding to the method steps provided by the present application. For example: acquiring a high-definition texture image of a decorative paper sample to be detected, presetting a three-level scale decomposition threshold, and performing three-level scale decomposition and feature extraction on the high-definition texture image according to the preset three-level scale decomposition threshold to obtain a multi-scale texture image group and a multi-scale texture feature set; evaluating the effect of the three-level scale decomposition; the multi-scale texture feature set includes first-scale texture features, second-scale texture features, and third-scale texture features; constructing a texture quality attention model, and obtaining a multi-scale attention weight vector set according to the multi-scale texture feature set and the texture quality attention model; based on the multi-scale attention weight vector set, performing weighted fusion on the first-scale texture features, second-scale texture features, and third-scale texture features to obtain a multi-scale fusion feature set; determining the quality of the decorative paper sample to be detected according to the multi-scale fusion features to obtain the quality determination level of the decorative paper sample to be detected. When the computer program is executed by a central processing unit (CPU), the above functions defined in the method of the present application are executed.

[0298] The methods, systems, and devices of the present application can be implemented in many ways. For example, the methods, systems, and devices of the present application can be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is for illustration only. The steps of the method of the present application are not limited to the specific order described above, unless otherwise specifically stated. In addition, in some embodiments, the present application can also be implemented as a program recorded in a recording medium, and these programs include machine-readable instructions for implementing the method according to the present application. Therefore, the present application also covers a recording medium storing a program for executing the method according to the present application.

[0299] In addition, parts of the above technical solutions provided in the embodiments of the present application that are consistent with the corresponding technical solutions in the prior art in terms of implementation principles are not described in detail to avoid excessive elaboration.

[0300] As described above in the specific embodiments, the purpose, technical solutions, and beneficial effects of the present invention have been further described in detail. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. Online quality discrimination algorithm for decorative paper, characterized in that The algorithm includes: Collecting a high-definition texture image of a decorative paper sample to be detected, presetting three-level scale decomposition thresholds, and performing three-level scale decomposition and feature extraction on the high-definition texture image according to the preset three-level scale decomposition thresholds to obtain a multi-scale texture image group and a multi-scale texture feature set; evaluating the effect of the three-level scale decomposition and determining whether the multi-scale texture image group meets the feature extraction requirements; the multi-scale texture feature set includes first-scale texture features, second-scale texture features, and third-scale texture features; The preset three-level scale decomposition thresholds include: Introducing a scale factor, establishing a calculation formula for the three-level scale decomposition thresholds according to the size, resolution Res, and scale factor of the high-definition texture image, and calculating the three-level scale decomposition thresholds, which are respectively a fine-scale decomposition threshold T1, a medium-scale decomposition threshold T2, and a large-scale decomposition threshold T3; the size of the high-definition texture image includes the image width Wid and the image height Height; The calculation formula for the three-level scale decomposition thresholds is: Among them, represents the scale level, and the values are 1, 2, 3; is the scale factor, k1 < k2 < k3, the initial value of k1 is 1, the initial value of k2 is 4, and the initial value of k3 is 16; the image width Wid and the image height Height are in pixels; The method for performing three-level scale decomposition and feature extraction on the high-definition texture image according to the preset three-level scale decomposition thresholds to obtain a multi-scale texture image group and a multi-scale texture feature set includes: Performing Gaussian smoothing on the high-definition texture image, and then performing adaptive downsampling according to the fine-scale decomposition threshold T1 to generate a first-scale texture image; performing Gaussian smoothing on the first-scale texture image, performing adaptive downsampling according to the medium-scale decomposition threshold T2, and fusing the high-frequency information of the Laplacian pyramid decomposition to generate a second-scale texture image; performing Gaussian smoothing on the second-scale texture image, performing adaptive downsampling according to the large-scale decomposition threshold T3, and fusing the high-frequency information of the Laplacian pyramid decomposition to generate a third-scale texture image; constructing a multi-scale texture image group from the first-scale texture image, the second-scale texture image, and the third-scale texture image; Extracting first-scale texture features from the first-scale texture image; the first-scale texture features include first-scale Gabor texture features, first-scale LBP texture features, and first-scale GLCM texture features; extracting second-scale texture features from the second-scale texture image; the second-scale texture features include second-scale Gabor texture features, second-scale LBP texture features, and second-scale GLCM texture features; extracting third-scale texture features from the third-scale texture image; the third-scale texture features include third-scale Gabor texture features, third-scale LBP texture features, and third-scale GLCM texture features; constructing a multi-scale texture feature set from the first-scale texture features, the second-scale texture features, and the third-scale texture features; Constructing a texture quality attention model, obtaining a multi-scale attention weight vector set according to the multi-scale texture feature set and the texture quality attention model; based on the multi-scale attention weight vector set, performing weighted fusion on the first-scale texture features, the second-scale texture features, and the third-scale texture features to obtain a multi-scale fusion feature set; Based on the multi-scale fusion features, the quality of the decorative paper sample to be detected is judged to obtain the quality judgment level of the decorative paper sample to be detected; The judging the quality of the decorative paper sample to be detected based on the multi-scale fusion features includes: Construct a lightweight texture feature extraction network, input the multi-scale fusion features into the lightweight texture feature extraction network to obtain multi-scale lightweight fusion features; according to the multi-scale lightweight fusion features and the pre-constructed quality discrimination classifier, obtain the quality judgment level and the quality judgment score corresponding to the quality judgment level; judge whether the quality judgment score corresponding to the quality judgment level meets the preset quality level threshold. If it meets, output the final quality judgment level. If it does not meet, output a warning signal.

2. The online quality discrimination algorithm for decorative paper according to claim 1, wherein The evaluating the effect of the three-level scale decomposition includes: Calculate the structural similarity SSIM index and the scale space feature stability SSFS index between the first-scale texture image, the second-scale texture image, and the third-scale texture image and the high-definition texture image respectively to obtain the SSIM index set and the SSFS index set; The SSIM index set includes SSIM indexes of three scales, and the SSIM indexes of the three scales are the first structural similarity SSIM index SSIM1, the second structural similarity SSIM index SSIM2, and the third structural similarity SSIM index SSIM3 respectively; The SSFS index set includes SSFS indexes of three scales, and the SSFS indexes of the three scales are the first-scale space feature stability SSFS index SSFS1, the second-scale space feature stability SSFS index SSFS2, and the third-scale space feature stability SSFS index SSFS3 respectively; Judge whether the SSIM indexes of the three scales are all greater than the SSIM similarity threshold Ts, and whether the SSFS indexes of the three scales are all greater than the SSFS stability threshold Tf; if not, adaptively adjust the scale factor in the three-level scale decomposition threshold calculation formula until the multi-scale decomposition effect is ideal.

3. The online quality discrimination algorithm for decorative paper according to claim 1, wherein The multi-scale attention weight vector set includes a first-scale attention weight vector, a second-scale attention weight vector, and a third-scale attention weight vector; The obtaining the multi-scale attention weight vector set according to the multi-scale texture feature set and the texture quality attention model includes: Stitch the first-scale Gabor texture feature, the first-scale LBP texture feature, and the first-scale GLCM texture feature to form a first-scale comprehensive feature; stitch the second-scale Gabor texture feature, the second-scale LBP texture feature, and the second-scale GLCM texture feature to form a second-scale comprehensive feature; stitch the third-scale Gabor texture feature, the third-scale LBP texture feature, and the third-scale GLCM texture feature to form a third-scale comprehensive feature; Input the first-scale comprehensive feature into the texture quality attention model to obtain a first-scale attention heat map; input the second-scale comprehensive feature into the texture quality attention model to obtain a second-scale attention heat map; input the third-scale comprehensive feature into the texture quality attention model to obtain a third-scale attention heat map; Perform global average pooling operations on the first-scale attention heatmap, the second-scale attention heatmap, and the third-scale attention heatmap to obtain the first-scale attention weight vector, the second-scale attention weight vector, and the third-scale attention weight vector.

4. The online quality discrimination algorithm for decorative paper according to claim 3, characterized in that, The obtaining of the multi-scale fusion features includes: Obtain the first-scale fusion feature according to the first-scale attention weight vector and the first-scale texture feature; Obtain the second-scale fusion feature according to the second-scale attention weight vector and the second-scale texture feature; Obtain the third-scale fusion feature according to the third-scale attention weight vector and the third-scale texture feature; Concatenate the first-scale fusion feature, the second-scale fusion feature, and the third-scale fusion feature to generate the multi-scale fusion feature.

5. The online quality discrimination algorithm for decorative paper according to claim 4, characterized in that The obtaining of the first-scale fusion feature includes: Multiply the first-scale attention weight vector by the first-scale Gabor texture feature to obtain the first-scale Gabor weighted feature; Multiply the first-scale attention weight vector by the first-scale LBP texture feature to obtain the first-scale LBP weighted feature; Multiply the first-scale attention weight vector by the first-scale GLCM texture feature to obtain the first-scale GLCM weighted feature; Fuse the first-scale Gabor weighted feature, the first-scale LBP weighted feature, and the first-scale GLCM weighted feature to obtain the first-scale fusion feature.

6. An online quality discrimination system for decorative paper, which is used to implement the online quality discrimination algorithm for decorative paper described in any one of claims 1-5, characterized in that, The system includes: Image decomposition module: used to collect the high-definition texture image of the decorative paper sample to be detected, preset the three-level scale decomposition threshold, and perform three-level scale decomposition and feature extraction on the high-definition texture image according to the preset three-level scale decomposition threshold to obtain a multi-scale texture image group and a multi-scale texture feature set; the multi-scale texture feature set includes the first-scale texture feature, the second-scale texture feature, and the third-scale texture feature; Decomposition effect evaluation module: used to evaluate the effect of the three-level scale decomposition; Feature fusion module: used to construct a texture quality attention model, obtain a multi-scale attention weight vector set according to the multi-scale texture feature set and the texture quality attention model; based on the multi-scale attention weight vector set, perform weighted fusion on the first-scale texture feature, the second-scale texture feature, and the third-scale texture feature to obtain a multi-scale fusion feature set; Quality determination module: perform quality determination on the decorative paper sample to be detected according to the multi-scale fusion features to obtain the quality determination level of the decorative paper sample to be detected.

Images