A method for evaluating the uniformity of a ceramic glaze
By constructing a multimodal data acquisition system and using deep learning technology, the problem of rapid, objective, and comprehensive evaluation of ceramic glaze uniformity has been solved, and high-precision automated evaluation of glaze quality has been achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JIANGXI CIMIC CERAMICS
- Filing Date
- 2026-01-22
- Publication Date
- 2026-06-23
AI Technical Summary
Existing technologies cannot achieve rapid, objective, comprehensive, and accurate automated evaluation of the uniformity of ceramic glaze surfaces, and suffer from problems such as strong subjectivity, susceptibility to interference, and low efficiency.
A multimodal data acquisition system was constructed to simultaneously acquire spectral, three-dimensional morphology, and multi-angle optical image data of the glaze surface. Deep learning technology was used for feature extraction and fusion, including hyperspectral imaging, structured light three-dimensional scanning, and multi-angle visible light imaging. Feature fusion was performed by combining a multi-branch deep feature extraction network and an attention mechanism, and finally, a comprehensive evaluation result was output.
It achieves a comprehensive, quantitative, and reliable evaluation of glaze uniformity, overcomes the limitations of a single data source, eliminates subjective human factors, and improves the accuracy and repeatability of the evaluation.
Smart Images
Figure CN122265137A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of ceramic testing technology, and in particular to a method for evaluating the uniformity of ceramic glaze. Background Technology
[0002] For ceramic products, especially high-end architectural and daily-use ceramics, the quality of the glaze is a key factor determining the product's grade and value. Glaze uniformity is a comprehensive evaluation indicator, encompassing not only color uniformity but also gloss, smoothness, texture consistency, and the presence of microscopic defects such as pinholes, bubbles, orange peel texture, and color spots. Currently, the evaluation of ceramic glaze quality both domestically and internationally primarily relies on the following methods: Manual visual inspection: This is the most traditional and common method. Experienced quality inspectors observe the glaze surface from different angles under a standard light source, judging whether its color, gloss, defects, etc., are up to standard based on their personal experience. This method is highly subjective, easily affected by factors such as personnel fatigue, emotions, and differences in experience, resulting in low efficiency. Furthermore, it cannot provide quantitative and standardized evaluation, making it difficult to meet the needs of modern large-scale production.
[0003] Single-image-based machine vision methods: These methods typically use industrial cameras to acquire two-dimensional images of the glaze surface under specific lighting conditions, and then extract features using image processing algorithms (such as grayscale statistics, texture analysis, edge detection, Fourier transform, etc.) to evaluate uniformity. For example, color uniformity is evaluated by calculating the grayscale variance of the image, or orange peel effect is evaluated by analyzing the spectrum of the texture. However, these methods have significant limitations: Limited information dimension: Two-dimensional images lose the three-dimensional morphological and spectral information of the surface. For areas with similar colors but different gloss levels, or for slight bumps and undulations (such as orange peel texture), two-dimensional images have difficulty distinguishing them effectively.
[0004] Highly affected by the environment: Even small changes in factors such as lighting conditions and camera angle can severely affect image quality, leading to unstable evaluation results.
[0005] Limited feature representation capability: Features extracted by traditional image processing algorithms are often manually designed, and their feature representation capability and robustness are insufficient for complex and subtle uniformity changes.
[0006] Contact / point measurement methods based on physical instruments: such as using a colorimeter to measure the color value of a specific point, using a gloss meter to measure the gloss of a specific point, and using a profilometer to measure the smoothness of lines. Although these methods can provide quantitative data, they are all based on sampling measurements of "points" or "lines", which cannot fully reflect the uniformity of the entire glaze surface. In addition, the measurement speed is slow and cannot achieve full inspection.
[0007] In conclusion, existing technologies cannot achieve a rapid, objective, comprehensive, and accurate automated evaluation of the uniformity of ceramic glazes. Therefore, there is an urgent need for a new evaluation method that can integrate multi-dimensional information and simulate or even surpass the comprehensive judgment capabilities of human experts. Summary of the Invention
[0008] This invention addresses the technical problems existing in the prior art by providing a method for evaluating the uniformity of ceramic glaze. It aims to simultaneously collect multimodal data such as the spectrum, three-dimensional morphology, and multi-angle optical images of the glaze surface, and use advanced deep learning technology for feature extraction and fusion to finally output a comprehensive, quantitative, and reliable uniformity evaluation result. This solves the problems of existing methods, such as single evaluation dimension, strong subjectivity, susceptibility to interference, and low efficiency.
[0009] The technical solution of this invention to solve the above-mentioned technical problems is as follows: a method for evaluating the uniformity of ceramic glaze; comprising the following steps: S1: Construct a multimodal data acquisition system, which includes a hyperspectral imaging unit, a structured light three-dimensional scanning unit, a multi-angle visible light imaging unit, and a central control and synchronization unit; S2: Using the multimodal data acquisition system, synchronous data acquisition is performed on the ceramic glaze sample to be tested, and hyperspectral image data cube, three-dimensional topographic point cloud data, and a set of high-resolution visible light images from at least three different illumination angles are obtained respectively. S3: Preprocess the raw data collected in S2, including dark current correction and reflectivity calibration of hyperspectral data, denoising and mesh reconstruction of 3D point cloud data, and brightness normalization and registration of multi-angle visible light images. S4: Construct and train a multi-branch deep feature extraction network, the multi-branch deep feature extraction network comprising: S4.1: Spectral-spatial feature extraction branch, which uses a three-dimensional convolutional neural network to extract the spectral distribution features of the glaze and its spatial variation features from the preprocessed hyperspectral data; S4.2: Shape feature extraction branch, which uses a point cloud processing network based on PointNet++ to extract the macroscopic smoothness, microscopic waviness and geometric features of surface defects of the glaze from the three-dimensional shape data. S4.3: Texture-Gloss Feature Extraction Branch, which uses a two-stream convolutional neural network to extract view-independent texture features and view-dependent gloss distribution features from the registered multi-angle visible light images. S5: Input the high-dimensional features extracted from the three branches in S4 into a multimodal feature fusion module based on an attention mechanism. This module calculates the attention weight of each feature channel and performs weighted fusion accordingly to obtain a unified fusion feature vector rich in multimodal information. S6: Input the fused feature vector into a comprehensive rating network. This network is a fully connected neural network, and its output layer simultaneously outputs a continuous value glaze uniformity index. A discrete glaze quality grade ; S7: Based on the output of S6, generate a ceramic glaze uniformity evaluation report.
[0010] Furthermore, the hyperspectral imaging unit has a spectral range covering 400nm to 1000nm and a spectral resolution better than 5nm; the structured light three-dimensional scanning unit has a longitudinal resolution better than 10μm; the multi-angle visible light imaging unit includes three cameras arranged in a circle and LED light sources registered with them, with incident angles of 15°, 45° and 75°, respectively, to capture specular reflection and diffuse reflection components.
[0011] Furthermore, in step S4.1, the spectral-spatial feature extraction branch first uses principal component analysis to reduce the dimensionality of the hyperspectral data, retaining the first k principal components, and then uses a network structure containing two 3D convolutional layers, a 3D pooling layer and a 3D batch normalization layer for feature extraction.
[0012] Furthermore, in step S4.2, the topography feature extraction branch calculates the surface normal field and curvature distribution map of the meshed three-dimensional topography data, and uses these geometric information and height information together as input features of the PointNet++ network to enhance its ability to perceive micro-topography undulations.
[0013] Furthermore, in step S4.3, the dual-stream structure of the texture-gloss feature extraction branch includes: Texture Stream: The average image from multiple angles is used as input, and texture features are extracted using a pre-trained ResNet50 backbone. Gloss Flow: Taking the pixel intensity ratio map between images at different angles as input, a lightweight CNN network is used to extract features that characterize the spatial uniformity of surface gloss.
[0014] Furthermore, in step S5, the specific operation of the attention-based multimodal feature fusion module is as follows: Let the eigenvectors of the three branches be: , , ; First, they are mapped to the same feature dimension through a fully connected layer, resulting in C: ; The mapped features are concatenated as follows: Attention weights are calculated using an attention network: ; The final fusion features are: .
[0015] Furthermore, the comprehensive rating network is trained using a multi-task learning loss function, wherein the loss function... Defined as the mean squared error loss for the uniform exponential regression task. Cross-entropy loss for quality rating classification tasks Weighted sum: ; in, and For hyperparameters, and This is a real label.
[0016] Furthermore, in step S7, the evaluation report includes not only the final UI and Class, but also a feature visualization map, which is used to show the spatial location of the defects or non-uniform regions that contribute the most to the final evaluation result on the glaze sample, thereby achieving interpretability of the evaluation process.
[0017] The beneficial effects of this invention are: For the first time, hyperspectral, three-dimensional morphology, and multi-angle optical information are integrated for glaze evaluation, and a comprehensive analysis is conducted from multiple dimensions such as color, gloss, texture, and flatness, overcoming the limitations of a single data source; it is entirely based on data and algorithms, eliminating human subjective factors; the deep learning model can learn more complex and essential uniformity patterns from massive amounts of data than manual features, resulting in high evaluation accuracy and good repeatability. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of a method for evaluating the uniformity of ceramic glaze. Detailed Implementation
[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0020] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.
[0021] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.
[0022] like Figure 1 As shown, in one embodiment, a method for evaluating the uniformity of ceramic glaze includes the following steps: S1: Construct a multimodal data acquisition system, which includes a hyperspectral imaging unit, a structured light three-dimensional scanning unit, a multi-angle visible light imaging unit, and a central control and synchronization unit; S2: Using the multimodal data acquisition system, synchronous data acquisition is performed on the ceramic glaze sample to be tested, and hyperspectral image data cube, three-dimensional topographic point cloud data, and a set of high-resolution visible light images from at least three different illumination angles are obtained respectively. S3: Preprocess the raw data collected in S2, including dark current correction and reflectivity calibration of hyperspectral data, denoising and mesh reconstruction of 3D point cloud data, and brightness normalization and registration of multi-angle visible light images. S4: Construct and train a multi-branch deep feature extraction network, the multi-branch deep feature extraction network comprising: S4.1: Spectral-spatial feature extraction branch, which uses a three-dimensional convolutional neural network to extract the spectral distribution features of the glaze and its spatial variation features from the preprocessed hyperspectral data; S4.2: Shape feature extraction branch, which uses a point cloud processing network based on PointNet++ to extract the macroscopic smoothness, microscopic waviness and geometric features of surface defects of the glaze from the three-dimensional shape data. S4.3: Texture-Gloss Feature Extraction Branch, which uses a two-stream convolutional neural network to extract view-independent texture features and view-dependent gloss distribution features from the registered multi-angle visible light images. S5: Input the high-dimensional features extracted from the three branches in S4 into a multimodal feature fusion module based on an attention mechanism. This module calculates the attention weight of each feature channel and performs weighted fusion accordingly to obtain a unified fusion feature vector rich in multimodal information. S6: Input the fused feature vector into a comprehensive rating network, which is a fully connected neural network whose output layer simultaneously outputs a continuous glaze uniformity index and a discrete glaze quality grade. S7: Based on the output of S6, generate a ceramic glaze uniformity evaluation report.
[0023] Specifically, the core of a method for evaluating the uniformity of ceramic glaze lies in constructing an automated pipeline for "data acquisition - feature extraction - feature fusion - comprehensive evaluation." The specific steps are as follows: Multimodal data acquisition system construction: Design and integrate a data acquisition platform including a hyperspectral imager, a structured light 3D scanner, and a multi-angle visible light camera. A central control unit ensures that all sensors are triggered synchronously in time and space to guarantee data consistency.
[0024] Multimodal data synchronous acquisition: Place the ceramic sample to be tested at the measurement station, and the system automatically completes one scan, acquiring data synchronously. Hyperspectral image data: Acquire the spectral curve of each pixel in the visible to near-infrared band, accurately reflecting the spatial distribution of the color and chemical composition of the glaze.
[0025] Three-dimensional morphology data: Obtain three-dimensional point clouds of the glaze surface to accurately characterize its macroscopic smoothness and microscopic ripples (such as orange peel effect).
[0026] Multi-angle visible light images: images are acquired from different illumination angles to separate surface texture information (angle-independent diffuse reflection) and gloss information (angle-dependent specular reflection).
[0027] Data preprocessing: The raw data is standardized to eliminate sensor noise and environmental interference, and all data are unified to the same coordinate system, laying the foundation for subsequent feature fusion.
[0028] Multi-branch deep feature extraction: This is one of the key innovations of this invention. We designed three parallel deep neural network branches, each optimized for different types of data, to extract the most discriminative features: Spectral-spatial branch: Using 3D-CNN to process hyperspectral data cubes, it captures information in both the spectral and spatial dimensions, effectively identifying defects such as color difference and color spots.
[0029] Topography Branch: Utilizes advanced point cloud processing networks (such as PointNet++) to analyze 3D topography, directly process 3D point clouds, and accurately quantify surface undulations and geometric defects.
[0030] Texture-Gloss Branch: A two-stream CNN structure is adopted. One stream focuses on fusing viewpoint-independent texture features from multi-angle images, while the other stream is dedicated to analyzing the differences between images from different angles to extract gloss uniformity features.
[0031] Attention-based multimodal feature fusion: An attention fusion module was designed that, instead of simply concatenating or adding features, allows the network to automatically learn the importance weights of features from different modalities. For example, spectral features are given higher weights when evaluating color uniformity, while morphological features are given higher weights when evaluating orange peel texture. This adaptive fusion mechanism greatly improves the efficiency and robustness of feature representation.
[0032] The integrated rating network inputs the fused high-level features into a fully connected network. This network employs a multi-task learning framework, simultaneously performing regression (outputting a uniformity index from 0 to 1) and classification (outputting excellent, good, average, and poor grades). This design makes the evaluation results both accurate and intuitive.
[0033] Results generation and visualization: Output a detailed evaluation report, and use techniques such as gradient weighted class activation mapping to highlight the key areas of interest of the network decision on the original image, making the evaluation process transparent and interpretable.
[0034] Specifically, the hyperspectral imaging unit has a spectral range covering 400nm to 1000nm and a spectral resolution better than 5nm; the structured light three-dimensional scanning unit has a longitudinal resolution better than 10μm; the multi-angle visible light imaging unit includes three cameras arranged in a circle and LED light sources registered with them, with incident angles of 15°, 45° and 75°, respectively, to capture specular reflection and diffuse reflection components.
[0035] Specifically, in step S4.1, the spectral-spatial feature extraction branch first uses principal component analysis to reduce the dimensionality of the hyperspectral data, retaining the first k principal components, and then uses a network structure containing two 3D convolutional layers, a 3D pooling layer and a 3D batch normalization layer for feature extraction.
[0036] Specifically, in step S4.2, the topography feature extraction branch calculates the surface normal vector field and curvature distribution map of the meshed three-dimensional topography data, and uses these geometric information and height information together as input features of the PointNet++ network to enhance its ability to perceive micro-topography undulations.
[0037] Specifically, in step S4.3, the two-stream structure of the texture-gloss feature extraction branch includes: Texture Stream: The average image from multiple angles is used as input, and texture features are extracted using a pre-trained ResNet50 backbone. Gloss Flow: Taking the pixel intensity ratio map between images at different angles as input, a lightweight CNN network is used to extract features that characterize the spatial uniformity of surface gloss.
[0038] Specifically, in step S5, the multimodal feature fusion module based on the attention mechanism operates as follows: Let the feature vectors of the three branches be: , , ; First, they are mapped to the same feature dimension through a fully connected layer, resulting in C: ; The mapped features are concatenated as follows: Attention weights are calculated using an attention network: ; The final fusion features are: .
[0039] Specifically, the comprehensive rating network is trained using a multi-task learning loss function, wherein the loss function... Defined as the mean squared error loss for the uniform exponential regression task. Cross-entropy loss for quality rating classification tasks Weighted sum: ; in, and For hyperparameters, and This is a real label.
[0040] Specifically, in step S7, the evaluation report includes not only the final UI and Class, but also a feature visualization map, which is used to show the spatial location of the defects or non-uniform areas that contribute the most to the final evaluation results on the glaze sample, thereby achieving interpretability of the evaluation process.
[0041] Example 1: This example uses a 600mm x 600mm fully polished glazed ceramic tile that has been identified as "superior quality" by human experts as a sample.
[0042] Step S1: System Construction and Calibration A multimodal data acquisition system was built, using a hyperspectral imager (model: HySpex VNIR-1800) with a spectral range of 400-1000nm and a resolution of 3nm. A structured light scanner (self-developed) employs a blue LED light source, achieving a measurement accuracy of ±5μm. The multi-angle visible light imaging unit consists of three 5-megapixel industrial cameras (Basler acA2500-14gc) and three programmable LED strip light sources. The cameras and light sources are arranged in pairs, with incident angles of 15° (grazing incidence), 45° (standard incidence), and 75° (high-angle incidence). All equipment is synchronously controlled via a PLC and industrial computer.
[0043] Step S2: Data Acquisition; The tile sample was fixed on a motorized translation stage. After system startup, the stage moved the sample at a constant speed through the measurement area. Under the control of a trigger signal, the hyperspectral imager operated in line scanning mode, the structured light scanner projected a grating and acquired phase data, and three visible light cameras simultaneously exposed at preset positions. The final result was: The hyperspectral data cube I_{hs} has a size of (2000 x 2000 x 186) (space x space x spectrum).
[0044] The 3D point cloud data P_{3d} contains approximately 4 million points.
[0045] Three high-resolution visible light images at Ivis15°, Ivis45°, and Ivis75°. Step S3: Data Preprocessing Dark current subtraction was performed on I_{hs}, and reflectivity conversion was performed using a standard whiteboard.
[0046] Outlier removal and smoothing are performed on P_{3d}, and then reconstructed into a 2.5D mesh height map H_{3d}(x, y); Using I_{vis}^{45°} as the reference, affine transformation registration is performed on I_{vis}^{15°} and I_{vis}^{75°} to ensure pixel-level alignment. Simultaneously, brightness normalization is applied to the three images to eliminate the influence of absolute light intensity.
[0047] Step S4: Multi-branch deep feature extraction; The preprocessed data is then input into the pre-trained multi-branch network.
[0048] Spectral-spatial branch: First, PCA is used to reduce the dimensionality of I_{hs}, retaining the first 30 principal components (contribution rate > 99.9%). Then, the dimensionality-reduced data is fed into a 3D-CNN. This network contains two 3D convolutional layers (kernel size 3x3x3), outputting 32 and 64 feature maps respectively. Each convolutional layer is followed by ReLU activation and 3D max pooling. Finally, it is flattened into a 1024-dimensional feature vector. .
[0049] Topology Branch: The mesh height map H_{3d} is converted into a point cloud, and the normal vector and mean curvature are calculated for each point. This enhanced point cloud is fed into an improved PointNet++ network. This network learns geometric features at different scales through multi-level sampling and grouping, and finally outputs a 1024-dimensional feature vector. ; Texture-Gloss Branches: Texture Flow: Calculate the mean image I_{mean} of the three registered images and input it into a pre-trained ResNet50 network (remove the last classification layer) to extract a 1024-dimensional texture feature vector.
[0050] Gloss flow: Calculate I_{vis}^{75°} / I_{vis}^{45°}, which effectively characterizes the spatial distribution of specular reflection intensity. Input the ratio map into a lightweight CNN (4 convolutional layers), and the output is a 512-dimensional gloss feature vector.
[0051] The texture and gloss feature vectors are concatenated and then compressed into a 1024-dimensional feature vector through a fully connected layer. .
[0052] Step S5: Attention-based feature fusion; Will , and A fully connected layer maps the vector to 512 dimensions. Then, these dimensions are concatenated into a 1536-dimensional vector, which is then input into the attention network. The attention network calculates the weights as follows: This indicates that for this sample, the network considers morphological smoothness and spectral color to be equally important, while gloss is slightly less so; by weighting and summing according to these weights, the final 512-dimensional fused feature is obtained. ; Step S6: Overall Rating; Will Input a comprehensive rating network. This network contains two hidden layers (256 and 128 neurons respectively), and the output layer has two heads: one uses the sigmoid activation function to output the UI, and the other uses the softmax function to output the class probability. Network output: Uniformity index The probability distribution of quality grades is as follows: Therefore, the final rating is "Excellent". Step S7: Report generation and visualization; The system generates an evaluation report; the feature visualization (e.g., generated based on Grad-CAM technology) shows that the network focuses on the overall features of the entire glaze when making decisions, and there are no obvious high-response abnormal areas, which is consistent with the expectation of "superior grade".
[0053] Example 2: Evaluation of glaze uniformity of a ceramic tile with slight orange peel defects; This example uses a tile that was found to have a slight "orange peel" effect during manual inspection as a sample.
[0054] Steps S1 and S2: Same as in Example 1, complete data acquisition.
[0055] Step S3: The preprocessing process is the same. Fine, periodic ripples can already be observed in the three-dimensional topography data.
[0056] Step S4: Feature extraction; Spectral-spatial branching: extraction The difference from Example 1 is not significant, as the color uniformity is acceptable.
[0057] Morphology branch: This branch has the strongest response; as can be seen from the feature map of the middle layer of the network, it has a clear activation of the micro-ripple structure of the surface.
[0058] Texture-Gloss Branch: Due to the orange peel effect, diffuse light is scattered, and the features extracted by gloss flow show a certain degree of irregularity.
[0059] Step S5: Feature fusion; The weights calculated by the attention network are Compared with Example 1, the weight of morphological features is significantly increased, indicating that the network automatically identifies that uneven morphology is the main contradiction leading to the decrease in uniformity of the sample.
[0060] Step S6: Overall Rating; Network output: Uniformity index The probability distribution of quality grades is as follows: Therefore, the final grade is determined to be "medium".
[0061] Step S7: Report generation and visualization; The evaluation report clearly pointed out that the problem mainly stemmed from the surface morphology; the feature visualization map successfully highlighted the area where the orange peel texture was located on the three-dimensional morphology data, providing clear guidance for adjusting the production process.
[0062] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0063] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.
Claims
1. A method for evaluating the uniformity of ceramic glaze, characterized in that, Includes the following steps: S1: Construct a multimodal data acquisition system, which includes a hyperspectral imaging unit, a structured light three-dimensional scanning unit, a multi-angle visible light imaging unit, and a central control and synchronization unit; S2: Using the multimodal data acquisition system, synchronous data acquisition is performed on the ceramic glaze sample to be tested, and hyperspectral image data cube, three-dimensional topographic point cloud data, and a set of high-resolution visible light images from at least three different illumination angles are obtained respectively. S3: Preprocess the raw data collected in S2, including dark current correction and reflectivity calibration of hyperspectral data, denoising and mesh reconstruction of 3D point cloud data, and brightness normalization and registration of multi-angle visible light images. S4: Construct and train a multi-branch deep feature extraction network, the multi-branch deep feature extraction network comprising: S4.1: Spectral-spatial feature extraction branch, which uses a three-dimensional convolutional neural network to extract the spectral distribution features of the glaze and its spatial variation features from the preprocessed hyperspectral data; S4.2: Shape feature extraction branch, which uses a point cloud processing network based on Point Net++ to extract the macroscopic smoothness, microscopic waviness and geometric features of surface defects of the glaze from the three-dimensional shape data; S4.3: Texture-Gloss Feature Extraction Branch, which uses a two-stream convolutional neural network to extract view-independent texture features and view-dependent gloss distribution features from the registered multi-angle visible light images. S5: Input the high-dimensional features extracted from the three branches in S4 into a multimodal feature fusion module based on an attention mechanism. This module calculates the attention weight of each feature channel and performs weighted fusion accordingly to obtain a unified fusion feature vector rich in multimodal information. S6: Input the fused feature vector into a comprehensive rating network. This network is a fully connected neural network, and its output layer simultaneously outputs a continuous value glaze uniformity index. A discrete glaze quality grade ; S7: Based on the output of S6, generate a ceramic glaze uniformity evaluation report.
2. The method for evaluating the uniformity of ceramic glaze according to claim 1, characterized in that, The hyperspectral imaging unit has a spectral range covering 400nm to 1000nm and a spectral resolution better than 5nm; the structured light three-dimensional scanning unit has a longitudinal resolution better than 10μm; the multi-angle visible light imaging unit includes three cameras arranged in a circle and LED light sources registered with them, with incident angles of 15°, 45° and 75°, respectively, to capture specular reflection and diffuse reflection components.
3. The method for evaluating the uniformity of ceramic glaze according to claim 1, characterized in that, In step S4.1, the spectral-spatial feature extraction branch first uses principal component analysis to reduce the dimensionality of the hyperspectral data, retaining the first k principal components, and then uses a network structure containing two 3D convolutional layers, a 3D pooling layer and a 3D batch normalization layer for feature extraction.
4. The method for evaluating the uniformity of ceramic glaze according to claim 1, characterized in that, In step S4.2, the topography feature extraction branch calculates the surface normal vector field and curvature distribution map of the meshed three-dimensional topography data, and uses these geometric information and height information together as input features of the PointNet++ network to enhance its ability to perceive micro-topography undulations.
5. The method for evaluating the uniformity of ceramic glaze according to claim 1, characterized in that, In step S4.3, the two-stream structure of the texture-gloss feature extraction branch includes: Texture Stream: The average image of the multi-angle images is used as input, and texture features are extracted using a pre-trained ResNet50 backbone; Gloss Flow: Taking the pixel intensity ratio map between images at different angles as input, a lightweight CNN network is used to extract features that characterize the spatial uniformity of surface gloss.
6. The method for evaluating the uniformity of ceramic glaze according to claim 1, characterized in that, In step S5, the specific operation of the attention-based multimodal feature fusion module is as follows: Let the eigenvectors of the three branches be: , , ; First, they are mapped to the same feature dimension through a fully connected layer, resulting in C: ; The mapped features are concatenated as follows: Attention weights are calculated using an attention network. ; The final fusion features are: .
7. The method for evaluating the uniformity of ceramic glaze according to claim 1, characterized in that, The comprehensive rating network is trained using a multi-task learning loss function. Defined as the mean squared error loss for the uniform exponential regression task. Cross-entropy loss for quality rating classification tasks Weighted sum: ; in, and For hyperparameters, and This is a real label.
8. The method for evaluating the uniformity of ceramic glaze according to claim 1, characterized in that, In step S7, the evaluation report includes not only the final UI and Class, but also a feature visualization map, which is used to show the spatial location of the defects or non-uniform areas that contribute the most to the final evaluation results on the glaze sample, thereby making the evaluation process interpretable.