Artificial intelligence-based phase segmentation method, computer equipment, and computer program products for cement-based materials based on high-dimensional multimodal dynamics.

Abstract

This application relates to an artificial intelligence phase segmentation method, computer equipment, and computer program product for cement-based materials based on high-dimensional multimodal imaging. It pertains to the fields of image analysis and phase recognition. The method includes: acquiring BSE images and EDS multi-element surface scan images from the same field of view; constructing a high-dimensional multimodal feature tensor that integrates composition, chemical ratios, BSE image features, and spatial coordinates; performing superpixel segmentation in the high-dimensional space and projecting it back onto a two-dimensional plane; extracting core representative points far from the boundary within the superpixel region; calculating elemental ratios and stoichiometric consistency descriptors; screening high-confidence phase seeds using preset chemical judgment rules; obtaining a complete phase label map through superpixel adjacency graph label propagation; interactively verifying and correcting regions with propagation confidence below a preset threshold and updating the label map; and finally outputting a spatially continuous, manually verified phase segmentation result, effectively reducing manual annotation time.

Description

Technical Field

[0001] This application relates to the field of image analysis and phase recognition technology, specifically to an artificial intelligence phase segmentation method for cement-based materials based on high-dimensional multimodal methods, computer equipment, and computer program products. Background Technology

[0002] During the hydration process, cement-based materials form a complex multiphase microstructure composed of unhydrated clinker particles, hydrated calcium silicate gel, calcium hydroxide, ettringite, carbonates, pores, and slag reaction products. The phase composition, spatial distribution, and degree of hydration of this microstructure directly affect the material's macroscopic properties such as strength and durability. Therefore, accurately identifying and quantitatively characterizing the microphases of cement-based materials is a crucial foundation for understanding the evolution of material properties and optimizing material design.

[0003] Backscattered electron (BSE) images and energy dispersive spectroscopy (EDS) elemental surface scan images are commonly used techniques in the microscopic characterization of cement-based materials. BSE images can reflect the differences in average atomic number and morphological characteristics of different micro-regions; EDS images can provide spatial distribution information of elements such as Ca, Si, Al, S, Fe, and O. Combining the two can provide richer information for phase identification than a single grayscale image.

[0004] However, existing BSE-EDS phase identification methods still have the following shortcomings: First, the traditional BSE grayscale threshold method mainly relies on grayscale differences, making it difficult to distinguish hydration products, carbonates, and mixed gel regions with overlapping grayscale ranges.

[0005] Second, existing identification methods based on chemical element ratios typically use low-dimensional ratio diagrams such as Si / Ca and Al / Ca for judgment, which are difficult to fully express the differences in multi-element composition and have insufficient accuracy in composite cementitious material systems.

[0006] Third, existing superpixel segmentation methods are usually performed on two-dimensional pseudo-color images composed of a few element channels, which cannot fully utilize multimodal information such as total element composition, BSE grayscale, local texture and spatial coordinates.

[0007] Fourth, existing manual ROI selection methods are inefficient and have poor repeatability.

[0008] Fifth, the classification results of unsupervised clustering or machine learning methods may not have clear chemical significance, and purely data-driven methods are prone to producing fragmented or uninterpretable classification results.

[0009] Sixth, pixels located at phase boundaries or in mixed regions may have EDS signals that simultaneously contain contributions from multiple neighboring phases. Classifying them directly on a single-pixel basis is easily affected by noise and boundary mixing.

[0010] Therefore, there is an urgent need for a method that can fully utilize BSE-EDS multimodal information, reduce pixel-level noise and phase boundary mixing effects, reduce human intervention, maintain chemical interpretability, and output spatially continuous phase segmentation maps. Summary of the Invention

[0011] This application provides an artificial intelligence phase segmentation method, computer equipment, and computer program product for cement-based materials based on high-dimensional multimodal methods, which can solve the technical problems of existing BSE-EDS phase recognition features being singular, having large boundary mixing interference, and having poor chemical interpretability.

[0012] In a first aspect, embodiments of this application provide an artificial intelligence-based phase segmentation method for cement-based materials based on high-dimensional multimodal dynamics. The method is executed by a computer device and includes the following steps: Step S1: Acquire BSE grayscale image and EDS multi-element surface scan image of cement-based material under the same field of view; Step S2: Preprocess the BSE grayscale image and EDS multi-element surface scan image to extract the effective solid phase region; Steps S3-S7: Within the effective solid phase region, construct composition feature blocks based on the EDS multi-element surface scan image, and construct chemical ratio feature blocks, BSE image feature blocks, and spatial coordinate feature blocks. Normalize the composition feature blocks, chemical ratio feature blocks, BSE image feature blocks, and spatial coordinate feature blocks respectively, and perform weighted fusion of the normalization results of each feature block according to preset weights to obtain a high-dimensional multimodal feature tensor. Steps S8-S9: In the high-dimensional feature space represented by the high-dimensional multimodal feature tensor, superpixel segmentation is performed in combination with two-dimensional spatial coordinate constraints to obtain multiple superpixel regions, and the label corresponding to each superpixel region is projected back onto the two-dimensional image plane. Steps S10-S11: For each superpixel region, extract a core pixel set based on the distance of each pixel in the region to the region boundary, calculate the core average descriptor of the superpixel region based on the core pixel set, and select the real pixel closest to the core average descriptor from the core pixel set as the core representative point. Steps S12-S13: Calculate the chemical descriptor based on the core representative point of each superpixel region. The chemical descriptor includes chemical element ratios and stoichiometric consistency descriptors. Step S14: According to the preset chemical determination rules, the core representative points of each superpixel region are screened, including: using the chemical descriptor to screen out superpixel regions that meet the corresponding phase chemical constraints, as high-confidence phase seeds; Step S15: Using each superpixel region as a graph node, construct a superpixel adjacency graph based on the spatial adjacency relationship in the two-dimensional image plane; Step S16: Take the high-confidence phase seed as the labeled node, take the superpixel region in the superpixel adjacency graph that is not labeled as a high-confidence phase seed as the node to be identified, perform label propagation on the superpixel adjacency graph, assign phase labels to the node to be identified according to the feature similarity and spatial proximity between adjacent nodes, and obtain a complete phase label map covering the entire superpixel region. Step S17: In response to the user's interactive review and correction of the superpixel region in the complete phase label image where the propagation confidence is lower than a preset confidence threshold, update the complete phase label image to obtain the updated complete phase label image; Step S18: Based on the updated complete phase label map, generate and output the phase segmentation result of cement-based material, the phase segmentation result including the phase label of each superpixel region and its spatial distribution in the two-dimensional image plane.

[0013] The technical advantages of this application are as follows: By acquiring BSE images and EDS multi-element surface scan images from the same field of view, a high-dimensional multimodal feature tensor is constructed that integrates composition, chemical ratios, BSE image features, and spatial coordinates. Superpixel segmentation is performed in the high-dimensional space and projected back onto a two-dimensional plane. Core representative points far from the boundary within the superpixel region are extracted. Elemental ratios and oxidative stoichiometry consistency descriptors are calculated. High-confidence phase seeds are selected using preset chemical judgment rules. Label propagation is performed through a superpixel adjacency graph, and finally, spatially continuous phase segmentation results are output. Through the above process and method, automated, region-level, and chemically interpretable phase segmentation can be achieved, reducing the influence of phase boundary mixing signals. This overcomes the shortcomings of traditional methods, such as single features, large boundary mixing interference, high human intervention, and poor chemical interpretability.

[0014] In one embodiment, the component feature block in step S3 is constructed in the following manner: Extract a multi-element composition vector from each pixel in the EDS multi-element area scan image; wherein, the multi-element composition vector includes multiple element value vectors; The element value vectors in the multi-element vector that are equal to 0 or less than a preset threshold are replaced with preset minimum positive numbers to obtain the replaced vector; The replaced vector is subjected to closure normalization, and then the result of closure normalization is subjected to central logarithmic ratio transformation or equidistant logarithmic ratio transformation to obtain the component feature block.

[0015] The technical effect of this embodiment is that it converts the constituent data into a feature representation suitable for Euclidean distance calculation, thereby improving the physical rationality of similarity measurement.

[0016] In one embodiment, the chemical ratio feature block is constructed in step S4 in the following manner: Based on the elemental composition in the EDS multi-element area scan image, at least one logarithmic chemical element ratio feature is calculated. The logarithmic chemical element ratio feature includes at least one of ln(Si / Ca), ln(Al / Ca), ln(S / Ca), and ln(Fe / Ca). The calculated logarithmic ratio is used as a chemical ratio feature block.

[0017] The technical effect of this embodiment is that by reducing the distribution skewness and long-tail effect through logarithmic transformation, the chemical ratio characteristics are more suitable for participating in high-dimensional distance metrics.

[0018] In one embodiment, step S8, which involves performing superpixel segmentation in the high-dimensional feature space represented by the high-dimensional multimodal feature tensor, combined with two-dimensional spatial coordinate constraints, to obtain multiple superpixel regions, includes: In the high-dimensional feature space represented by the high-dimensional multimodal feature tensor, the total distance between each pixel and each superpixel cluster center is calculated, and the total distance is determined by both the high-dimensional feature distance and the two-dimensional spatial distance. Each pixel is assigned to a corresponding cluster center based on the total distance, resulting in multiple superpixel regions; The high-dimensional feature distance represents the distance between the feature vectors corresponding to two pixels in the high-dimensional feature space; the high-dimensional feature distance is used to measure the similarity between two pixels on the component feature block, the chemical ratio feature block, and the BSE image feature block. The two-dimensional spatial distance represents the geometric distance between two pixels in the two-dimensional image plane; the two-dimensional spatial distance is used to constrain the spatial compactness and connectivity of the superpixel region.

[0019] The technical advantage of this embodiment is that, through the superpixel region, the true chemical composition similarity and morphological similarity can be reflected more effectively.

[0020] In one embodiment, in step S10, the core pixel set is extracted in the following manner: For each superpixel region, calculate the Euclidean distance from each pixel in the region to the boundary of the region and generate a distance transformation map; select pixels whose distance value is greater than a preset distance threshold or pixels whose distance value ranks in the top preset proportion as the core pixel set based on the distance transformation map; Accordingly, in step 11, the core representative point is determined in the following manner: The mean of the feature vectors of all pixels in the core pixel set is calculated as the core average descriptor; the pixel with the smallest Euclidean distance between the feature vector and the core average descriptor is selected from the core pixel set as the core representative point of the superpixel region.

[0021] The technical effect of this embodiment is that by determining the core representative point, the core representative point can be far away from the phase boundary and the edge of the aperture, thereby reducing the influence of mixed signals caused by the volume of electron beam interaction.

[0022] In one embodiment, in step S13, the oxidative stoichiometry consistency descriptor is constructed in the following manner: Obtain the measured oxygen-calcium ratio at the core representative point of each superpixel region. The measured oxygen-calcium ratio is the ratio of the oxygen element signal to the calcium element signal plus a preset small constant at the core representative point. The theoretical oxygen-to-calcium ratio of a candidate phase is calculated based on its chemical formula. The theoretical oxygen-to-calcium ratio is the ratio of the number of oxygen atoms to the number of calcium atoms in the chemical formula of the candidate phase. Calculate the absolute value of the difference between the measured oxygen-calcium ratio and the theoretical oxygen-calcium ratio; When the absolute value of the difference is less than a preset tolerance, the superpixel region is determined to satisfy the oxidative stoichiometric consistency constraint of the candidate phase.

[0023] The technical advantage of this embodiment is that it directly identifies the calcium phase by using the theoretical oxygen / calcium ratio in the chemical formula, thereby improving the chemical intuitiveness of phase differentiation.

[0024] In one embodiment, the preset chemical determination rule includes at least one of the following screening rules: If the Si / Ca ratio of the core representative point is within the first preset range, the Al / Ca ratio is less than the second preset threshold, and the BSE grayscale of the core representative point is within the preset BSE grayscale threshold range, then the candidate phase of the superpixel region is selected as a high-confidence phase seed of tricalcium silicate phase. If the Si / Ca ratio of the core representative point is within the third preset range, the Al / Ca ratio is less than the second preset threshold, and the BSE grayscale of the core representative point is within the preset BSE grayscale threshold range, then the candidate phase of the superpixel region is selected as a high-confidence phase seed of the dicalcium silicate phase. If the Si / Ca ratio of the core representative point is less than the fourth preset threshold, the Al / Ca ratio is less than the fifth preset threshold, and the difference between the measured oxygen-calcium ratio and the theoretical oxygen-calcium ratio of calcium hydroxide at the core representative point is less than the first preset tolerance, then the candidate phase of the superpixel region is selected as a high-confidence phase seed of the calcium hydroxide phase. If the Si / Ca ratio of the core representative point is less than the fourth preset threshold, the Al / Ca ratio is less than the fifth preset threshold, and the difference between the measured oxygen-calcium ratio of the core representative point and the theoretical oxygen-calcium ratio of calcite is less than the second preset tolerance, then the candidate phase of the superpixel region is selected as a high-confidence phase seed of calcite phase. If the Si / Ca ratio of the core representative point is greater than the sixth preset threshold and the Al / Ca ratio is less than the seventh preset threshold, then the candidate phase of the superpixel region is selected as a high-confidence phase seed of the quartz phase.

[0025] The technical advantages of this embodiment are: by automatically generating high-confidence phase seeds, manual intervention is reduced, and the correspondence between phase labels and cement chemical rules is maintained; it can overcome the traditional method's reliance on manual ROI selection, which has a low degree of automation and poor repeatability.

[0026] In one embodiment, step S17 responds to an interactive review and correction performed by a user on superpixel regions in the complete phase labeling image where the propagation confidence level is lower than a preset confidence threshold, to update the complete phase labeling image and obtain an updated complete phase labeling image, including: The superpixel regions in the complete phase label whose propagation confidence is lower than a preset confidence threshold are marked as regions to be reviewed. The region to be verified is visualized in a BSE grayscale image or a chemical element ratio chart. The system receives manually corrected tags from external input, uses the corrected tags as new seeds to re-execute tag propagation, updates the complete phase tag map, and continues until the propagation confidence of all regions to be verified meets the preset quality requirements.

[0027] The technical advantage of this embodiment is that, while maintaining automation efficiency, it allows for targeted manual correction of key areas, thereby improving the reliability of complex area identification.

[0028] In a second aspect, this application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the artificial intelligence phase segmentation method for cement-based materials based on high-dimensional multimodality as described in the first aspect above.

[0029] Thirdly, this application provides a computer program product that stores a computer program, which, when executed by a computer device, implements the artificial intelligence phase segmentation method for cement-based materials based on high-dimensional multimodality as described in the first aspect above.

[0030] It is understood that the beneficial effects of the second and third aspects mentioned above can be found in the relevant descriptions in the first aspect above, and will not be repeated here. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 The diagram shows the main process flow of an artificial intelligence phase segmentation method for cement-based materials based on high-dimensional multimodality provided in this application. Figure 2 A schematic diagram of the structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0033] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0034] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0035] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0036] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0037] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0038] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0039] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), unless otherwise expressly and specifically defined.

[0040] Understandably, to address the technical problems of existing BSE-EDS phase identification methods, such as single phase recognition features, large boundary mixing interference, and poor chemical interpretability, this application proposes, in one aspect, an artificial intelligence phase segmentation method for cement-based materials based on high-dimensional multimodal features. Through high-dimensional feature fusion, regional superpixel organization, core representative point extraction, high-confidence seed screening using regular inequalities, and propagational label expansion, it achieves automated, interpretable, and spatially continuous identification of multiple phase types in BSE-EDS images of cement-based materials. (Refer to...) Figure 1 The method includes the following steps: Step S1: Acquire BSE grayscale image and EDS multi-element surface scan image of cement-based material under the same field of view; In the specific implementation, BSE grayscale images and EDS multi-element surface scan images are acquired under the same microscopic field of view. The BSE grayscale images are used to characterize the grayscale differences, particle morphology, pore distribution, and interface boundaries in the microstructure of cement-based materials. The EDS multi-element surface scan images are used to characterize the spatial distribution of different elements within the microscopic region of the cement-based material.

[0041] For example, the EDS multi-element surface scan image includes at least a variety of element maps selected from Ca, Si, Al, S, Fe, Mg, Na, K, Cl, and O; wherein Ca, Si, Al, S, Fe, Mg, Na, K, and Cl are used to construct the main component features and chemical ratio features; the O element can be used to subsequently construct an oxygen-enhanced oxidative stoichiometry consistency descriptor to help distinguish low Si / Ca calcium phases such as Portlandite and Calcite.

[0042] Step S2: Preprocess the BSE grayscale image and EDS multi-element surface scan image to extract the effective solid phase region; It is understood that this step first performs denoising processing on the EDS element area scan image to reduce high-frequency random noise in the element image; as a specific implementation, the denoising processing includes at least one of total variation denoising, Gaussian smoothing, bilateral filtering, and BSE guided filtering; Then, the BSE grayscale image is normalized to convert its grayscale values ​​to a uniform range. In this embodiment, the threshold for the porous region or resin region can be determined based on the grayscale histogram of the BSE grayscale image. A pore mask is then extracted, and small connected component removal and small hole filling are performed on the pore mask to obtain the cleaned pore region. It should be noted that the grayscale histogram refers to the threshold division using the histogram "tangent method," which can divide the region into solid regions and porous regions; the effective solid phase region refers to the non-porous, non-resin background region defined by the mask of the effective solid phase region.

[0043] It is understood that this step involves extracting the pore region from the BSE grayscale image, and then using this pore region as a mask to remove it from the subsequent two-dimensional image domain involved in the analysis, so that the subsequent EDS features, BSE features and spatial coordinate features are constructed only within the effective solid phase region.

[0044] The effective solid phase region extracted in this step is used to limit subsequent feature construction (including component feature block construction, chemical ratio feature block construction, BSE image feature block construction, and spatial coordinate feature block construction in steps S3 to S7; and also includes step S8 high-dimensional SLIC superpixel segmentation, S10 region core pixel extraction, S14 high-confidence phase seed generation, and S16 label propagation or semi-supervised expansion. All subsequent segmentation is performed in this region), superpixel segmentation, representative point extraction, and phase identification processes, so that pores and resin background do not participate in solid phase classification.

[0045] Steps S3-S7: Within the effective solid region, construct composition feature blocks, chemical ratio feature blocks, BSE image feature blocks, and spatial coordinate feature blocks based on the EDS multi-element surface scan image. Normalize the composition feature blocks, chemical ratio feature blocks, BSE image feature blocks, and spatial coordinate feature blocks respectively. Then, perform weighted fusion of the normalization results of each feature block according to preset weights to obtain a high-dimensional multimodal feature tensor. It is understood that, in this embodiment, the following feature blocks need to be constructed: Step S3: Constructing a component feature block, including: extracting a multi-element composition vector from each pixel in the EDS multi-element area scan image; wherein the multi-element composition vector includes multiple element value vectors; replacing the element value vectors in the multi-element composition vector that are equal to 0 or less than a preset threshold with a preset minimum positive number to obtain a replaced vector; performing closure normalization processing on the replaced vector, and then performing a central logarithmic ratio transformation or an equidistant logarithmic ratio transformation on the result of closure normalization to obtain a component feature block.

[0046] As a specific implementation method, the composition vector of EDS multi-element corresponding to each pixel in the effective solid phase region can be extracted. As an example, the multi-element vector can be represented as:

[0047] in, The signal intensity, mass fraction, atomic fraction, or normalized count value of the i-th element in EDS multi-element analysis is represented by , and D represents the number of elements participating in the principal component analysis.

[0048] Zero-value substitution is performed on the multi-element composition vector to avoid numerical errors during subsequent logarithmic transformation. Zero-value substitution refers to replacing element value vectors in the multi-element composition vector x=(x1,x2,…,xD) that are equal to 0 or less than a preset threshold with a preset minimum positive number ε to avoid undefined situations during subsequent logarithmic transformation. ε can be a certain proportion of the minimum positive value of the corresponding element value vector within the effective solid-phase region, or a certain proportion of the minimum positive value in the set of all element vector positive values, or a preset constant.

[0049] Subsequently, the vector composed of elements after zero-value substitution is subjected to closure normalization (i.e., closure normalization processing):

[0050] in, This represents the value of the i-th element after normalization of the closure.

[0051] Furthermore, the constituent vectors after closure are subjected to CLR or ILR transformation; taking CLR transformation as an example:

[0052] in: It is the geometric mean of the constituent vectors after closure; Through the above processing, the element composition data affected by constant sum constraints is transformed into a feature representation suitable for Euclidean distance calculation, thereby reducing the impact of the composition data closure effect on subsequent high-dimensional SLIC segmentation.

[0053] In some alternative embodiments, principal component analysis can also be performed on the component features after CLR or ILR transformation to reduce feature redundancy and improve computational efficiency.

[0054] Step S4: Construct the chemical ratio feature block, including: Based on the elemental composition within the effective solid phase region, chemical ratio characteristics with phase interpretation significance are constructed; In one example, the chemical ratio characteristic includes at least one of the following: , , , ; in, Used to distinguish between silicate clinker phase, CSH gel, and silica-rich regions; Used to distinguish aluminate phases, AFm-type phases, and aluminum-containing hydration products; Used to identify sulfate-related phases or sulfur-rich regions of AFt / AFm; Used to assist in the identification of C4AF or iron-containing phases; It is understandable that this example uses logarithmic ratios to reduce the skewness and long-tail effects of the original ratio distribution, making chemical ratio characteristics more suitable for distance calculation and region delineation.

[0055] Step S5: Construct BSE image feature blocks, including: In a specific implementation, BSE image features can be extracted within the effective solid-phase region based on the BSE grayscale image; wherein the BSE image features include at least one of the following: the original grayscale or normalized grayscale in the BSE grayscale image; local grayscale mean; local grayscale standard deviation; gradient magnitude; LoG response; local texture features.

[0056] Among them, the local gray mean is used to characterize the average gray level of the region; the local gray standard deviation is used to characterize the texture changes within the region; and the gradient amplitude is used to characterize the strength of phase boundaries, pore edges, or particle boundaries.

[0057] By introducing BSE image features, we can supplement the morphological and boundary information that is difficult to be directly reflected in the elemental composition features, making the subsequent superpixel regions more consistent with the real microstructure.

[0058] Step S6: Construct spatial coordinate feature blocks, including: In practical implementation, the two-dimensional image coordinates of each pixel within the effective solid region can be normalized to obtain spatial coordinate features. ); in, and These represent the normalized coordinates of the pixel in the horizontal and vertical directions, respectively. It should be noted that the spatial coordinate features are given a low weight in the subsequent high-dimensional feature fusion, in order to maintain the spatial continuity of the superpixel segmentation results and prevent pixels that are spatially far apart but have similar chemical compositions from being classified into the same region.

[0059] Step S7: Normalize the component feature block, the chemical ratio feature block, the BSE image feature block and the spatial coordinate feature block respectively, and perform weighted fusion of the normalization results of each feature block according to preset weights to obtain a high-dimensional multimodal feature tensor; As a specific implementation method, z-score normalization can be applied to the component feature blocks and BSE image feature blocks; robust normalization based on median and interquartile range can be applied to the chemical ratio feature blocks; and range normalization can be applied to the spatial coordinate feature blocks. Subsequently, the feature blocks are concatenated according to preset weights to obtain a high-dimensional multimodal feature vector:

[0060] in: Represents component feature blocks; Characteristic blocks representing chemical ratios; Represents BSE image feature blocks; Represents a spatial coordinate feature block; These represent the weights of the corresponding feature blocks; As a preferred example, the weights of the component feature blocks and chemical ratio feature blocks are higher than those of the spatial coordinate feature blocks, so that the segmentation results are mainly controlled by chemical composition and image structure, while maintaining a moderate degree of spatial continuity.

[0061] It should be noted that this step is not a simple stacking of the original element graphs, but rather a block-based processing of features with different physical meanings, followed by normalization and weighted fusion, thereby improving the physical rationality of the high-dimensional distance metric and the stability of the segmentation. It is understandable that two-dimensional features (such as BSE grayscale) are not distinguishable enough, and grayscale thresholds alone cannot distinguish them well. This step constructs a high-dimensional multimodal feature tensor by executing steps S3-S7, which integrates multi-element composition, chemical ratios, BSE image features, etc., so that each pixel is mapped from the original low-dimensional representation (a single grayscale value + several element channels) to a high-dimensional feature vector, that is, the original two-dimensional features are mapped to a high-dimensional feature space, making the features more obvious and easier to distinguish.

[0062] Steps S8-S9: In the high-dimensional feature space represented by the high-dimensional multimodal feature tensor, superpixel segmentation is performed in combination with two-dimensional spatial coordinate constraints to obtain multiple superpixel regions, and the label corresponding to each superpixel region is projected back onto the two-dimensional image plane: In this embodiment, SLIC superpixel segmentation with spatial constraints is performed in the high-dimensional multimodal feature space by executing step S8. It can be understood that the features mapped in step S7 are used for segmentation in the high-dimensional feature space, specifically based on the high-dimensional multimodal feature tensor obtained in step S7. This step performs SLIC superpixel segmentation within the effective solid-phase region. During the segmentation process, the similarity between pixels (i.e., between each pixel to be segmented and each superpixel cluster center) is determined by both the high-dimensional feature distance and the two-dimensional spatial distance. It should be noted that the "high-dimensional feature distance" refers to the distance between the feature vectors corresponding to two pixels in the high-dimensional feature space; the "two-dimensional spatial distance" refers to the geometric distance between two pixels in the original two-dimensional image plane.

[0063] Among them, the high-dimensional feature distance is used to measure the similarity of pixels in terms of the composition of the component feature block, the chemical ratio of the chemical ratio feature block, the features of the BSE grayscale image, and the local texture; the two-dimensional spatial distance is used to constrain the spatial compactness and connectivity of the superpixel region. Specifically, in the high-dimensional feature space represented by the high-dimensional multimodal feature tensor, superpixel segmentation is performed in combination with two-dimensional spatial coordinate constraints. The total distance between each pixel and each superpixel cluster center is calculated. The total distance is determined by the high-dimensional feature distance and the two-dimensional spatial distance. Each pixel is assigned to the corresponding cluster center according to the total distance to obtain multiple superpixel regions. For example, the superpixel segmentation in this embodiment specifically includes: initializing multiple superpixel cluster centers; for each pixel, calculating the total distance between the pixel and each superpixel cluster center, wherein the total distance is the sum of the high-dimensional feature distance and the weighted two-dimensional spatial distance, where the high-dimensional feature distance is the Euclidean distance between the feature vector of the pixel and the feature vector of the cluster center, and the two-dimensional spatial distance is the Euclidean distance between the image coordinates of the pixel and the image coordinates of the cluster center; assigning each pixel to the nearest cluster center according to the total distance; updating the cluster centers and iterating until convergence. This total distance can characterize the similarity between pixels (i.e., between each pixel to be segmented and each superpixel cluster center); After the steps described above in this embodiment, multiple superpixel regions on a two-dimensional image plane are obtained. Each superpixel region consists of spatially adjacent pixels with similar multimodal features. Unlike the traditional method of performing SLIC on a two-dimensional composite image, this embodiment performs SLIC in a high-dimensional feature space that integrates multi-elemental composition, chemical ratios, BSE grayscale, local texture, and spatial coordinates, making the superpixel regions more reflective of true chemical and morphological similarity.

[0064] Step S9: Project the high-dimensional SLIC segmentation results back onto the two-dimensional image plane; It is understandable that after the aforementioned steps map the two-dimensional features to the high-dimensional feature space to make the features easier to distinguish, since the high-dimensional space (usually tens of dimensions or even higher) cannot be directly displayed on the two-dimensional screen (the high-dimensional feature space cannot be directly displayed or visualized), it is necessary to perform this step S9 "project the high-dimensional SLIC segmentation result back onto the two-dimensional image plane". Specifically, the superpixel label results obtained in step S8 are mapped back to the two-dimensional image plane containing the original BSE grayscale image and the EDS multi-element surface scan image. Each superpixel region corresponds to a spatially connected region in the two-dimensional image. Through the above-described re-projection process in this embodiment, the position, shape, boundary, and corresponding BSE grayscale features of each superpixel region can be visually viewed in the two-dimensional image, while preserving its high-dimensional multimodal feature description.

[0065] Steps S10-S11: For each superpixel region, extract a core pixel set based on the distance of each pixel in the region to the region boundary, calculate the core average descriptor of the superpixel region based on the core pixel set, and select the real pixel closest to the core average descriptor from the core pixel set as the core representative point. It is understandable that this step first extracts the core pixel set of the region based on distance transformation, and then extracts the core representative points and the core average descriptor. As a specific implementation method, this embodiment can be implemented in the following way: Step S10: For each superpixel region, calculate the Euclidean distance from each pixel in the region to the boundary of the region, and generate a distance transformation map; select pixels whose distance value is greater than a preset distance threshold or pixels whose distance value ranks in the top preset proportion as the core pixel set according to the distance transformation map; For example, pixels with a distance value ranking of 10%–30% can be selected as the core pixel set of the region; It is understandable that the core pixel set of a region is farther away from the phase boundary, aperture edge, and adjacent phase than the edge pixels of the region, and therefore is less affected by the volume mixing of electron beam interactions. This step S10 can solve the defect that the traditional geometric center representative point may be close to the boundary and cannot effectively represent the pure phase region.

[0066] Step S11: Calculate the mean of the feature vectors of all pixels in the core pixel set as the core average descriptor; select the pixel with the smallest Euclidean distance between the feature vector and the core average descriptor in the core pixel set as the core representative point of the superpixel region. For example, the core average descriptor includes at least one of the following: core average elemental composition; core average CLR or ILR component features; core average chemical ratio; core average BSE grayscale; core average image texture features; and core average spatial coordinates.

[0067] Furthermore, within the set of core pixels in the region, the real pixel closest to the core average descriptor is selected as the most representative core point. As a preferred implementation, this core point possesses two characteristics: first, it is located inside the region in the two-dimensional image, far from the boundary; second, it is close to the region's core average feature in the high-dimensional feature space. Therefore, this core point is more representative of the phase composition characteristics of the region than the traditional geometric center point, and it can reduce the influence of boundary mixing signals on subsequent phase determination.

[0068] Steps S12-S13: Based on the core representative point of each superpixel region, calculate the chemical descriptor, which includes chemical element ratios and oxidation stoichiometry consistency descriptors, specifically including: Step S12: Calculate the chemical element ratios based on core representative points: For each superpixel region, a core representative point or core average descriptor is calculated to form a chemical descriptor for phase rule screening; as an example, the chemical descriptor includes at least one of the following: Si / Ca, Al / Ca, S / Ca, Fe / Ca, O / Ca, and an oxidation stoichiometric consistency descriptor. Among them, Si / Ca, Al / Ca, S / Ca, and Fe / Ca are used to characterize the compositional relationship of the main cations in the region, and O / Ca is used to characterize the stoichiometric relationship between oxygen and calcium in the calcium-containing phase. In some embodiments, for calcium phases such as Portlandite (calcium hydroxide) and Calcite (calcite) that have low Si / Ca and Al / Ca ratios but different oxygen-calcium stoichiometry, the embodiments of this application can further utilize the theoretical O / Ca atomic ratio in the chemical formula of the candidate phases for discrimination, thereby improving the phase differentiation capability of low Si / Ca calcium regions.

[0069] Step S13: Construct an oxidative stoichiometric consistency descriptor based on the theoretical O / Ca ratio of the chemical formula: First, the measured oxygen-to-calcium ratio at the core representative point of each superpixel region is obtained. This measured oxygen-to-calcium ratio is the ratio of the oxygen signal to the calcium signal plus a preset small constant at the core representative point. Specifically, for EDS area scan data containing O element information, the measured oxygen-to-calcium ratio of each superpixel core region is calculated:

[0070] in, This indicates the oxygen signal within the core region. This indicates the calcium signal within the core region. To prevent the use of a pre-defined small constant with a denominator of zero; Then, based on the candidate phase The theoretical oxygen-to-calcium ratio of the candidate phase is calculated using the chemical formula of the candidate phase. This theoretical oxygen-to-calcium ratio is the ratio of the number of oxygen atoms to the number of calcium atoms in the candidate phase's chemical formula. Specifically, this embodiment uses the following candidate phases. The theoretical oxygen-calcium ratio can be calculated using the chemical formula:

[0071] in, Indicates the number of oxygen atoms in the chemical formula of the candidate phase. Indicates the number of calcium atoms in the chemical formula of the candidate phase; Furthermore, it is also necessary to calculate the absolute value of the difference between the measured oxygen-calcium ratio and the theoretical oxygen-calcium ratio. Specifically, this embodiment uses the following formula for calculation:

[0072] Finally, by comparing the absolute values ​​of the above differences... Compared with the preset tolerance, in Less than the preset tolerance At that time, it was considered that the region met the requirements for candidate phase. Consistency constraints on oxidative stoichiometry.

[0073] For example, the chemical formula for Portlandite (calcium hydroxide) is... One calcium atom corresponds to two oxygen atoms, therefore: .

[0074] The chemical formula for Calcite is: One calcium atom corresponds to three oxygen atoms, therefore: ; Therefore, in calcium-containing regions with low Si / Ca and low Al / Ca ratios, if the formula is satisfied... If the formula is satisfied, then the region is more likely to be classified as a similar region to Portlandite (calcium hydroxide); If so, the area is more likely to be classified as a similar area to Calcite.

[0075] This embodiment, through the aforementioned oxidative stoichiometric consistency descriptor, can directly constrain candidate phases using the theoretical O / Ca ratio in the chemical formula, avoiding reliance on empirical linear mapping parameters and improving the chemical intuitiveness and interpretability of phase rule screening.

[0076] Step S14: According to the preset chemical determination rules, the core representative points of each superpixel region are screened, including: using the chemical descriptor to screen out superpixel regions that meet the corresponding phase chemical constraints, as high-confidence phase seeds; Based on the aforementioned steps, the core representative points of each superpixel region and their corresponding chemical descriptors have been obtained, including the Si / Ca ratio, Al / Ca ratio, Fe / Ca ratio, and measured oxygen-calcium ratio. And the backscattered electron grayscale value. This step, based on preset chemical determination rules, performs conditional judgment on the core representative points of each superpixel region, and automatically selects high-confidence phase seeds that meet the corresponding phase chemical constraints.

[0077] It should be noted that the preset chemical determination rules in this step are not intended to complete a mandatory classification of the entire map in one go, but rather to screen a small number of high-confidence, strongly constrained, and physically interpretable initial phase seeds. For regions that do not meet the strict rules, no definite phase label will be assigned for the time being; instead, they will be retained as unlabeled regions, mixed regions, or regions awaiting propagation, to be processed in subsequent label propagation steps.

[0078] The following section explains in detail the judgment logic of the preset chemical determination rules, taking into account specific phase types: S14.1, High-confidence seed screening for tricalcium silicate (C3S-like) phase: The following rules apply to the screening of high-confidence seeds for tricalcium silicate (C3S-like) phase: If the Si / Ca ratio of the core representative point of the superpixel region is within a first preset range, the Al / Ca ratio is less than a second preset threshold, and the backscattered electron grayscale of the core representative point is within a preset backscattered electron grayscale threshold range, then the candidate phase of the superpixel region is selected as a high-confidence phase seed of the tricalcium silicate phase.

[0079] In this embodiment, the first preset range is determined based on the theoretical chemical formula of tricalcium silicate (C3S), which is Ca3SiO5, where the theoretical Si / Ca ratio is approximately 1 / 3 ≈ 0.33. Considering the actual degree of hydration, elemental solid solution, and EDS quantitative error, the first preset range is preferably set to 0.30–0.36. The second preset threshold is used to exclude aluminate phases with high aluminum content and is preferably set to 0.10. The preset backscattered electron grayscale threshold range is the grayscale interval corresponding to the clinker phase, which can be specifically determined by statistically analyzing the BSE grayscale histogram of unhydrated clinker particles, for example, set to a BSE grayscale value greater than 80 (normalized range of 0–255). When the core representative point simultaneously meets the above three conditions, the superpixel region is determined to be a high-confidence seed of the tricalcium silicate phase.

[0080] S14.2. High-confidence seed screening for dicalcium silicate (C2S-like) phase: The following rules apply to the screening of high-confidence seeds for dicalcium silicate (C2S-like) phase: If the Si / Ca ratio of the core representative point is within the third preset range, the Al / Ca ratio is less than the second preset threshold, and the backscattered electron grayscale of the core representative point is within the preset backscattered electron grayscale threshold range, then the candidate phase of the superpixel region is selected as a high-confidence phase seed of the dicalcium silicate phase.

[0081] In this embodiment, the chemical formula of dicalcium silicate (C2S) is Ca2SiO4, where the theoretical Si / Ca ratio is 1 / 2 = 0.50. Considering actual deviations, the third preset range is preferably set to 0.45–0.55. The second preset threshold is also set to 0.10 to exclude interference from aluminum-containing phases. The backscattered electron grayscale threshold range is the same as in S14.1, using the grayscale range corresponding to the clinker. When the core representative point simultaneously meets the above three conditions, the superpixel region is determined to be a high-confidence seed of the dicalcium silicate phase.

[0082] S14.3. High-confidence seed screening for Portlandite-like calcium hydroxide phase: The following rules apply to the screening of high-confidence seeds for the Portlandite-like calcium hydroxide phase: If the Si / Ca ratio of the core representative point is less than the fourth preset threshold, the Al / Ca ratio is less than the fifth preset threshold, and the difference between the measured oxygen-calcium ratio and the theoretical oxygen-calcium ratio of calcium hydroxide at the core representative point is less than the first preset tolerance, then the candidate phase of the superpixel region is selected as a high-confidence phase seed of the calcium hydroxide phase.

[0083] In this embodiment, the chemical formula of calcium hydroxide (Portlandite) is Ca(OH)₂, where both the Si / Ca and Al / Ca ratios are low. The fourth preset threshold is used to limit the silicon content, preferably set to 0.20. The fifth preset threshold is used to limit the aluminum content, preferably set to 0.20. The theoretical oxygen-calcium ratio of the calcium hydroxide is calculated based on its chemical formula: each Ca atom corresponds to 2 O atoms, therefore... Measured oxygen-calcium ratio The ratio of the O element signal to the Ca element signal at the core representative point is used for calculation, i.e.: ; in, The signal strength of element O. The signal intensity of Ca element, To prevent extremely small positive numbers with a denominator of zero (e.g.) =10-6). The first preset tolerance is used to measure the closeness of the measured oxygen-calcium ratio to the theoretical oxygen-calcium ratio, and is preferably set to 0.3. When Si / Ca < 0.20, Al / Ca < 0.20, and at the core representative point At that time, the superpixel region was identified as a high-confidence seed of the calcium hydroxide phase.

[0084] This embodiment effectively distinguishes between the calcium hydroxide phase and the calcite phase (the theoretical oxygen-calcium ratio of calcite is 3) with low Si / Ca and Al / Ca ratios using an oxidative stoichiometric consistency descriptor, thus overcoming the shortcomings of traditional methods that rely solely on the Si / Ca-Al / Ca ratio to distinguish between these two calcareous hydrated phases.

[0085] S14.4, High-confidence seed screening for calcite-like phases: The following rules can be used for high-confidence seed screening for calcite-like phases: If the Si / Ca ratio of the core representative point is less than the fourth preset threshold, the Al / Ca ratio is less than the fifth preset threshold, and the difference between the measured oxygen-calcium ratio of the core representative point and the theoretical oxygen-calcium ratio of calcite is less than the second preset tolerance, then the candidate phase of the superpixel region is selected as a high-confidence phase seed of the calcite phase.

[0086] In this embodiment, the chemical formula of calcite is CaCO3, where the Si / Ca and Al / Ca ratios are also relatively low. The fourth and fifth preset thresholds are the same as in S14.3, set to 0.20 and 0.20 respectively. The theoretical oxygen-calcium ratio of calcite is calculated based on its chemical formula: each Ca atom corresponds to 3 O atoms, therefore... The second preset tolerance is used to measure the closeness of the measured oxygen-calcium ratio to the theoretical oxygen-calcium ratio of calcite, and is preferably set to 0.3. When Si / Ca < 0.20, Al / Ca < 0.20, and at the core representative point... At that time, the superpixel region was identified as a high-confidence seed of the calcite phase.

[0087] A comparison of S14.3 and S14.4 reveals that the calcium hydroxide and calcite phases exhibit low Si / Ca and Al / Ca ratios, making them difficult to distinguish. However, their theoretical oxygen-calcium ratios differ. This embodiment introduces an oxidation stoichiometric consistency descriptor, utilizing the difference in O / Ca information to effectively differentiate the two phases.

[0088] S14.5, Quartz-like high-confidence seed screening: The following rules apply to quartz-like high-confidence seed screening: If the Si / Ca ratio of the core representative point is greater than the sixth preset threshold and the Al / Ca ratio is less than the seventh preset threshold, then the candidate phase of the superpixel region is selected as a high-confidence phase seed of the quartz phase. In this embodiment, the main component of quartz is SiO2, exhibiting silicon enrichment and extremely low calcium content. The sixth preset threshold is used to identify silicon-rich regions, preferably set to 1.0. The seventh preset threshold is used to exclude aluminum-containing phase interference, preferably set to 0.20. When the Si / Ca ratio of the core representative point is greater than 1.0 and the Al / Ca ratio is less than 0.20, the superpixel region is determined to be a high-confidence seed of the quartz phase.

[0089] S14.6, Other Phase Types: Those skilled in the art should understand that the above-mentioned preset chemical determination rules can be extended according to the actual cement-based material system. For example, for calcium aluminate phase (C3A-like) or calcium aluminoferrite phase (C4AF-like), rules such as Al / Ca ratio greater than 0.5 and Fe / Ca ratio as an auxiliary judgment can be set; for AFm-type phases, a comprehensive judgment can be made by combining S / Ca, Cl / Ca and stoichiometric consistency descriptors; for CSH gel matrix, a comprehensive judgment can be made based on the Si / Ca and Al / Ca ratio ranges and the results of the exclusion of the above-mentioned identified phases. These extended rules are all within the protection scope of this invention.

[0090] S14.7, High-confidence phase seeds output: After being filtered by the rules in S14.1 to S14.5 above, superpixel regions that satisfy any of the rule conditions are marked as high-confidence phase seeds of the corresponding phase type. These seeds have the following characteristics: It is generated based on clearly defined pre-defined chemical determination rules for cement and has good chemical interpretability; Originating from core representative points in the superpixel region, far from phase boundaries and aperture edges, this reduces interference from mixed signals; it only covers sparse regions that meet strict constraints, without forcing classification of all regions. These high-confidence phase seeds will serve as initial labeling nodes for subsequent label propagation steps, expanding phase labels to the entire image region through the superpixel adjacency graph.

[0091] Step S15: Using each superpixel region as a graph node, construct a superpixel adjacency graph based on the spatial adjacency relationship in the two-dimensional image plane; In the specific implementation, each superpixel region is used as a graph node, and adjacency edges between nodes are established based on the spatial contact relationship in the two-dimensional image. If two superpixel regions share a boundary in the two-dimensional image plane, or the spatial distance is less than a preset threshold, they are considered to be adjacent, and a graph connection is established between them.

[0092] In one embodiment, the edge weight can be determined based on the high-dimensional feature distance, chemical ratio difference, BSE grayscale difference, and spatial distance between two superpixel regions; As a preferred example, the more similar two regions are, the larger their edge weights; the greater the difference between two regions, the smaller their edge weights.

[0093] Step S16: Take the high-confidence phase seed as the labeled node, take the superpixel region in the superpixel adjacency graph that is not labeled as a high-confidence phase seed as the node to be identified, perform label propagation on the superpixel adjacency graph, assign phase labels to the node to be identified according to the feature similarity and spatial proximity between adjacent nodes, and obtain a complete phase label map covering the entire superpixel region. In the specific implementation, the high-confidence phase seeds obtained in step S14 are used as labeled nodes, and the unlabeled superpixel regions are used as nodes to be identified. Label propagation, random walk, semi-supervised learning, graph propagation or similarity constraint clustering are performed on the superpixel adjacency graph constructed in step S15. During the propagation process, the labels of high-confidence seeds can remain fixed or be given higher weight, while unlabeled areas gradually acquire phase labels based on their characteristic similarity and spatial proximity with neighboring labeled areas. For regions where the propagation confidence level is lower than the preset threshold, they are retained as uncertain or mixture regions, or they are entered into a manual review process.

[0094] This step expands sparse but reliable high-confidence seeds into a continuous and complete regional phase map, while avoiding over-forced classification of mixed phases and boundary regions.

[0095] Step S17: In response to the user's interactive review and correction of the superpixel region in the complete phase label image where the propagation confidence is lower than a preset confidence threshold, update the complete phase label image to obtain the updated complete phase label image; It is understandable that the output of step 16 above is a complete phase label map covering the entire superpixel area. There is some uncertainty in the solid phase area or the mixed area, which requires manual correction by the user. Specifically, when the computer device receives a review instruction and / or correction instruction input by the user through the human-computer interaction interface, it marks the superpixel regions in the complete phase label with a propagation confidence level lower than a preset confidence threshold as regions to be reviewed; it visualizes the regions to be reviewed in a BSE grayscale image or a chemical element ratio map; it receives manually corrected labels from external input, uses the corrected labels as new seeds to re-execute label propagation, and updates the complete phase label map until the propagation confidence level of all regions to be reviewed meets the preset quality requirements.

[0096] It is understandable that for regions that remain in a low-confidence state after propagation, or superpixels located in critical phase boundaries, mixed hydration product regions, or slag reaction boundary regions, their positions in BSE grayscale images, EDS multi-element surface scan images, ratio maps, and phase diagrams can be displayed.

[0097] Operators can manually correct a small number of key areas based on chemical ratios, BSE grayscale, elemental distribution, and adjacent area labels; The corrected region labels are used again as seeds in label propagation until phase segmentation results that meet the preset quality requirements are obtained.

[0098] It should be noted that, compared with existing conventional technologies, the complete phase label map of the entire superpixel area output in step 16 is mostly correct and reliable. A small portion of the output requires manual adjustment. The manual review in step S17 does not involve manually labeling all pixels, but rather performing targeted corrections on a few key areas. This improves the reliability of complex area identification while maintaining automation efficiency, thus effectively reducing the time required for manual labeling.

[0099] Step S18: Based on the updated complete phase label map, generate and output the phase segmentation result of cement-based material, the phase segmentation result including the phase label of each superpixel region and its spatial distribution in the two-dimensional image plane.

[0100] For example, the phase segmentation results of cement-based materials output in this embodiment may include at least one of the following: final phase segmentation map; high-confidence seed map; superpixel region map; core representative point distribution map; area fraction of each phase; number of regions of each phase; average chemical composition of each phase; propagation confidence map of each region; uncertain or mixture region map; phase map superimposed with BSE grayscale image; and phase label result superimposed with EDS ratio map.

[0101] The final phase segmentation map can be used for quantitative analysis of the microstructure of cement-based materials, phase area statistics, hydration degree estimation, cross-validation with XRD / TGA data, and subsequent training of three-dimensional phase recognition models.

[0102] The technical advantage of this embodiment is that: First, this embodiment overcomes the problem of insufficient feature expression in traditional BSE grayscale threshold and low-dimensional ratio map methods by constructing a high-dimensional multimodal feature representation that integrates multi-element composition, chemical ratio, BSE grayscale image features and spatial coordinates, enabling the phase identification process to simultaneously utilize chemical information, morphological information and spatial continuity information. Secondly, this embodiment performs zero-value substitution, closure normalization, and CLR or ILR transformation on the element composition data, so that the EDS composition data is properly processed before entering the distance calculation and SLIC segmentation, reducing the distance distortion caused by the composition data constants and constraints, and improving the physical rationality of similarity measurement in high-dimensional feature space.

[0103] Furthermore, this embodiment does not perform superpixel segmentation on a two-dimensional composite pseudocolor image, but rather performs spatially constrained SLIC segmentation in a high-dimensional multimodal feature space. This allows the superpixel regions to better reflect the true chemical composition similarity, BSE morphological similarity, and spatial proximity relationships, thereby improving the region division quality of complex cement-based material microstructures.

[0104] Furthermore, this embodiment extracts the core pixel set within the superpixel region through distance transformation and calculates representative descriptors or medoid-like representative points based on the core region. Compared with traditional geometric center representative points, it can more effectively move away from phase boundaries, aperture edges and adjacent phases, reducing mixed signal interference caused by electron beam interaction volume.

[0105] In addition, this embodiment uses preset chemical judgment rules to automatically screen high-confidence phase seeds, which reduces the workload of traditional manual ROI selection, improves the automation and repeatability of the algorithm, and maintains the correspondence between phase labels and cement chemical rules.

[0106] Furthermore, this embodiment introduces an oxygen-enhanced stoichiometric consistency descriptor, which uses O / Ca information to help distinguish low Si / Ca and low Al / Ca regions such as Portlandite and Calcite, thus overcoming the deficiency that it is difficult to distinguish some calcareous hydrated phases by relying solely on the Si / Ca–Al / Ca ratio.

[0107] Furthermore, this embodiment uses superpixel regions as propagation nodes instead of directly using individual pixels as classification objects, which can reduce the impact of pixel-level noise, isolated outliers, and boundary mixed pixels on the final phase diagram, and improve the spatial continuity and regional consistency of the phase diagram.

[0108] Furthermore, this embodiment expands sparse high-confidence seeds into a complete phase diagram through tag propagation, semi-supervised expansion, or graph propagation methods, realizing the identification pattern of "a small number of strongly constrained seeds - regional-level propagation - complete phase diagram output", which takes into account both automation efficiency and chemical interpretability.

[0109] Furthermore, this embodiment allows for manual verification and interactive correction of low-confidence regions, mixed regions, or special interface regions, avoiding the forced classification of all regions into a specific phase and improving the identification reliability in complex hydration systems.

[0110] Finally, this embodiment is applicable to ordinary silicate cement, slag cement, composite cementitious materials, limestone calcined clay cement and other multiphase cement-based material systems, and can provide a reliable image data basis for quantitative analysis of microphase, evaluation of hydration degree, material design optimization and three-dimensional phase reconstruction.

[0111] Furthermore, the technical solutions of this application differ from the prior art in the following ways: The difference between this invention and the traditional BSE grayscale thresholding method is that the traditional BSE grayscale thresholding method mainly relies on grayscale differences for segmentation, while this invention integrates EDS multi-element composition, chemical ratio, BSE grayscale image features and spatial coordinates to achieve multimodal phase identification.

[0112] The difference between this invention and the low-dimensional ratio method is that the traditional low-dimensional ratio method mainly performs phase determination in a two-dimensional ratio plane such as Si / Ca–Al / Ca, while this invention first performs closure and CLR or ILR transformation on the multi-element composition, and constructs a high-dimensional multimodal feature space together with other features.

[0113] The difference between this invention and the two-dimensional composite image superpixel method is that traditional methods usually perform SLIC segmentation on two-dimensional composite images composed of a few element channels, while this invention performs SLIC segmentation in a high-dimensional multimodal feature space that integrates multi-element composition, chemical ratios, BSE grayscale image features and spatial coordinates.

[0114] The difference between this invention and manual ROI selection methods is that traditional methods require manual selection of representative regions or seed points, while this invention automatically generates high-confidence phase seeds through preset chemical judgment rules, reducing manual intervention and improving repeatability.

[0115] The difference between this invention and traditional superpixel representative point extraction methods is that traditional methods usually use the geometric center of the superpixel region as the representative point, while this invention extracts the core pixel set of the region based on distance transformation and selects the core mean or medoid-like core representative point to reduce the influence of phase boundary mixing signals.

[0116] The difference between this invention and ordinary unsupervised clustering methods is that ordinary unsupervised clustering results may not have a clear chemical meaning, while this invention uses chemical rules to generate high-confidence seeds and propagates them based on superpixel adjacency graphs, making the final phase labels more interpretable in cement chemistry.

[0117] Secondly, this application also proposes a computer device, please refer to... Figure 2 , Figure 2 This is a schematic diagram of the structure of an embodiment of the computer device provided in this application. The computer device may be a desktop computer, a laptop computer, or a smartphone device, etc. This embodiment does not limit it.

[0118] like Figure 2 As shown, the computer device 1 of this embodiment includes: at least one processor 10, a memory 11, and a computer program 12 stored in the memory 11 and executable on the at least one processor 10. When the processor 10 executes the computer program 12, it implements the steps in the embodiment of the artificial intelligence phase segmentation method for cement-based materials based on high-dimensional multimodality described in this application.

[0119] Figure 2 The computer device shown may include, but is not limited to, processor 10 and memory 11. Those skilled in the art will understand that... Figure 2 The examples of computer devices are merely examples and do not constitute a limitation on computer devices. They may include more or fewer components than shown in the illustration, or combinations of certain components, or different components. For example, they may also include input / output devices, network access devices, etc.

[0120] The processor 10 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0121] In some embodiments, the memory 11 may be an internal storage unit of the computer device, such as a hard drive or memory. In other embodiments, the memory 11 may be an external storage device of the computer device, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the memory 11 may include both internal and external storage units of the computer device. The memory 11 is used to store the operating system, applications, bootloader, data, and other programs, such as the program code of the editing program for the presentation file. The memory 11 can also be used to temporarily store data that has been output or will be output.

[0122] This application also provides a computer program product containing instructions that, when run on a computer, cause the computer to execute the method provided in the above-described embodiment of the artificial intelligence phase segmentation method for cement-based materials based on high-dimensional multimodality.

[0123] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A high-dimensional multimodal artificial intelligence-based phase segmentation method for cement-based materials, characterized in that, The method is executed by a computer device and includes the following steps: Acquire BSE grayscale images and EDS multi-element surface scan images of cement-based materials under the same field of view; The BSE grayscale image and EDS multi-element surface scan image are preprocessed to extract the effective solid phase region; Within the effective solid region, a component feature block is constructed based on the EDS multi-element surface scan image, and a chemical ratio feature block, a BSE image feature block, and a spatial coordinate feature block are also constructed. The component feature block, the chemical ratio feature block, the BSE image feature block, and the spatial coordinate feature block are normalized respectively, and the normalization results of each feature block are weighted and fused according to a preset weight to obtain a high-dimensional multimodal feature tensor. In the high-dimensional feature space represented by the high-dimensional multimodal feature tensor, superpixel segmentation is performed in combination with two-dimensional spatial coordinate constraints to obtain multiple superpixel regions, and the label corresponding to each superpixel region is projected back onto the two-dimensional image plane. For each superpixel region, a core pixel set is extracted based on the distance of each pixel in the region to the region boundary, and the core average descriptor of the superpixel region is calculated based on the core pixel set. The real pixel that is closest to the core average descriptor from the core pixel set is selected as the core representative point. Based on the core representative point of each superpixel region, a chemical descriptor is calculated, which includes chemical element ratios and stoichiometric consistency descriptors. According to the preset chemical determination rules, the core representative points of each superpixel region are screened, including: using the chemical descriptor to screen out superpixel regions that meet the corresponding phase chemical constraints, as high-confidence phase seeds; Using each superpixel region as a graph node, a superpixel adjacency graph is constructed based on the spatial adjacency relationship in the two-dimensional image plane; The high-confidence phase seeds are used as labeled nodes, and the superpixel regions in the superpixel adjacency graph that are not labeled as high-confidence phase seeds are used as nodes to be identified. Label propagation is performed on the superpixel adjacency graph, and phase labels are assigned to the nodes to be identified based on the feature similarity and spatial proximity between adjacent nodes, so as to obtain a complete phase label map covering the entire superpixel region. In response to the user's interactive review and correction of superpixel regions in the complete phase labeling image where the propagation confidence is lower than a preset confidence threshold, the complete phase labeling image is updated to obtain an updated complete phase labeling image; Based on the updated complete phase label map, the phase segmentation result of cement-based material is generated and output. The phase segmentation result includes the phase label of each superpixel region and its spatial distribution in the two-dimensional image plane.

2. The method as described in claim 1, characterized in that, The component feature block is constructed in the following manner: Extract a multi-element composition vector from each pixel in the EDS multi-element area scan image; wherein, the multi-element composition vector includes multiple element value vectors; The element value vectors in the multi-element vector that are equal to 0 or less than a preset threshold are replaced with preset minimum positive numbers to obtain the replaced vector; The replaced vector is subjected to closure normalization, and then the result of closure normalization is subjected to central logarithmic ratio transformation or equidistant logarithmic ratio transformation to obtain the component feature block.

3. The method as described in claim 1, characterized in that, The chemical ratio feature block is constructed in the following manner: Based on the elemental composition in the EDS multi-element area scan image, at least one logarithmic chemical element ratio feature is calculated. The logarithmic chemical element ratio feature includes at least one of ln(Si / Ca), ln(Al / Ca), ln(S / Ca), and ln(Fe / Ca). The calculated logarithmic ratio is used as a chemical ratio feature block.

4. The method as described in claim 1, characterized in that, The step of performing superpixel segmentation in the high-dimensional feature space represented by the high-dimensional multimodal feature tensor, combined with two-dimensional spatial coordinate constraints, to obtain multiple superpixel regions includes: In the high-dimensional feature space represented by the high-dimensional multimodal feature tensor, the total distance between each pixel and each superpixel cluster center is calculated, and the total distance is determined by both the high-dimensional feature distance and the two-dimensional spatial distance. Each pixel is assigned to a corresponding cluster center based on the total distance, resulting in multiple superpixel regions; The high-dimensional feature distance represents the distance between the feature vectors corresponding to two pixels in the high-dimensional feature space; the high-dimensional feature distance is used to measure the similarity between two pixels on the component feature block, the chemical ratio feature block, and the BSE image feature block. The two-dimensional spatial distance represents the geometric distance between two pixels in the two-dimensional image plane; the two-dimensional spatial distance is used to constrain the spatial compactness and connectivity of the superpixel region.

5. The method as described in claim 1, characterized in that, The core pixel set is extracted using the following method: For each superpixel region, calculate the Euclidean distance from each pixel in the region to the boundary of the region and generate a distance transformation map; select pixels whose distance value is greater than a preset distance threshold or pixels whose distance value ranks in the top preset proportion as the core pixel set based on the distance transformation map; Accordingly, the core representative points are determined in the following ways: The mean of the feature vectors of all pixels in the core pixel set is calculated as the core average descriptor; the pixel with the smallest Euclidean distance between the feature vector and the core average descriptor is selected from the core pixel set as the core representative point of the superpixel region.

6. The method as described in claim 1, characterized in that, The oxidative stoichiometry consistency descriptor is constructed in the following manner: Obtain the measured oxygen-calcium ratio at the core representative point of each superpixel region. The measured oxygen-calcium ratio is the ratio of the oxygen element signal to the calcium element signal plus a preset small constant at the core representative point. The theoretical oxygen-to-calcium ratio of a candidate phase is calculated based on its chemical formula. The theoretical oxygen-to-calcium ratio is the ratio of the number of oxygen atoms to the number of calcium atoms in the chemical formula of the candidate phase. Calculate the absolute value of the difference between the measured oxygen-calcium ratio and the theoretical oxygen-calcium ratio; When the absolute value of the difference is less than a preset tolerance, the superpixel region is determined to satisfy the oxidative stoichiometric consistency constraint of the candidate phase.

7. The method as described in claim 6, characterized in that, The preset chemical determination rules include at least one of the following screening rules: If the Si / Ca ratio of the core representative point is within the first preset range, the Al / Ca ratio is less than the second preset threshold, and the BSE grayscale of the core representative point is within the preset BSE grayscale threshold range, then the candidate phase of the superpixel region is selected as a high-confidence phase seed of tricalcium silicate phase. If the Si / Ca ratio of the core representative point is within the third preset range, the Al / Ca ratio is less than the second preset threshold, and the BSE grayscale of the core representative point is within the preset BSE grayscale threshold range, then the candidate phase of the superpixel region is selected as a high-confidence phase seed of the dicalcium silicate phase. If the Si / Ca ratio of the core representative point is less than the fourth preset threshold, the Al / Ca ratio is less than the fifth preset threshold, and the difference between the measured oxygen-calcium ratio and the theoretical oxygen-calcium ratio of calcium hydroxide at the core representative point is less than the first preset tolerance, then the candidate phase of the superpixel region is selected as a high-confidence phase seed of the calcium hydroxide phase. If the Si / Ca ratio of the core representative point is less than the fourth preset threshold, the Al / Ca ratio is less than the fifth preset threshold, and the difference between the measured oxygen-calcium ratio of the core representative point and the theoretical oxygen-calcium ratio of calcite is less than the second preset tolerance, then the candidate phase of the superpixel region is selected as a high-confidence phase seed of calcite phase. If the Si / Ca ratio of the core representative point is greater than the sixth preset threshold and the Al / Ca ratio is less than the seventh preset threshold, then the candidate phase of the superpixel region is selected as a high-confidence phase seed of the quartz phase.

8. The method as described in claim 6, characterized in that, The step of responding to the user's interactive review and correction of superpixel regions in the complete phase labeling image where the propagation confidence level is lower than a preset confidence threshold, to update the complete phase labeling image and obtain an updated complete phase labeling image, includes: The superpixel regions in the complete phase label whose propagation confidence is lower than a preset confidence threshold are marked as regions to be reviewed. The region to be verified is visualized in a BSE grayscale image or a chemical element ratio chart. The system receives manually corrected tags from external input, uses the corrected tags as new seeds to re-execute tag propagation, updates the complete phase tag map, and continues until the propagation confidence of all regions to be verified meets the preset quality requirements.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the artificial intelligence phase segmentation method for cement-based materials based on high-dimensional multimodality as described in any one of claims 1 to 8.

10. A computer program product, said computer program product storing a computer program, characterized in that, When the computer program is executed by a computer device, it implements the steps of the artificial intelligence phase segmentation method for cement-based materials based on high-dimensional multimodality as described in any one of claims 1 to 8.

Images