Cement-based material micro-phase segmentation method based on bse-eds multi-modal fusion

Abstract

The application provides a cement-based material micro-phase segmentation method based on BSE-EDS multi-modal fusion, comprising: providing a cement-based material to be analyzed; collecting a BSE image and an EDS element surface distribution map for the same field of view; constructing a BSE input tensor and an EDS input tensor; inputting the BSE and EDS input tensors into a pre-trained double-branch multi-modal fusion network; respectively enhancing the features of the two branches at each encoding stage, and obtaining BSE intra-modal enhanced features and EDS intra-modal enhanced features by combining channel attention; adaptively fusing the intra-modal enhanced features of the same encoding stage to obtain multi-scale fusion features; processing the multi-scale fusion features through a decoding output module to obtain a pixel-level phase segmentation map; and statistically calculating the pixel area fraction of each phase, and converting the pixel area fraction into a three-dimensional volume fraction based on the principle of stereology to obtain a micro-phase content parameter of the cement-based material. The application can accurately obtain the micro-phase content parameter, and provides reliable data support for explaining the performance evolution of the material and predicting long-term durability.

Description

Technical Field

[0001] This application relates to the field of microstructure characterization of cement-based materials, and more specifically, to a method for microphase segmentation of cement-based materials based on BSE-EDS multimodal fusion. Background Technology

[0002] The mechanical properties, durability, and service life of cement-based materials are closely related to their microstructure. For complex cement-based materials such as recycled aggregate concrete, residual mortar, interfacial transition zones, impurities, and pore structure can significantly affect macroscopic properties. Therefore, accurately identifying and quantitatively characterizing microphases is of great significance for explaining the evolution of material properties, optimizing mix proportions, and predicting long-term durability.

[0003] Backscattered electron (BSE) images provide high-resolution grayscale and morphological information. Their pixel grayscale values ​​are correlated with the average atomic number of the material, making them commonly used to identify phases such as pores, hydration products, and unhydrated clinker. However, in cement-based materials, the grayscale ranges of different hydration products, unhydrated clinker phases, aggregates, and impurities often overlap. This is particularly true in complex cement-based materials (such as admixture concrete and recycled aggregate concrete), where the boundaries of aggregates, residual mortar, and impurities are inconsistent, and internal grayscale fluctuations are significant. Relying solely on BSE grayscale thresholding or edge detection makes stable and accurate multi-phase segmentation difficult. Energy-dispersive X-ray spectroscopy (EDS) elemental distribution maps can reflect the spatial distribution of elements such as Ca, Si, Al, Mg, Fe, O, Na, and K, providing chemical composition information for phase identification. However, EDS imaging suffers from low spatial resolution, large characteristic X-ray interaction volumes, blurred boundaries, low count rate noise, and false detections of elemental signals in pore regions. Phase identification based solely on EDS images is prone to boundary deviations and noise interference.

[0004] Existing BSE-EDS joint analysis methods mostly employ image overlay, element ratio, threshold segmentation, clustering models, or simple feature stitching, which typically rely on human experience and prior rules, making it difficult to fully utilize the deep complementary relationship between BSE morphological boundaries and EDS element distribution. Furthermore, existing deep learning methods largely focus on BSE or SEM single-modal image segmentation, failing to simultaneously address both BSE grayscale overlap and EDS boundary blurring issues.

[0005] The search revealed: Patent application CN118691825A discloses a machine vision-based method and electronic device for phase segmentation of composite cement systems. This prior art mainly solves the problem of automatic phase segmentation in BSE images of composite cement systems. It mainly uses BSE images as image datasets and forms feature stacks through different feature extraction methods. It focuses on extracting machine vision features from BSE images and training a random forest algorithm.

[0006] Patent application CN116363161A discloses a method and apparatus for multi-class segmentation of cement hydration images. This prior art is aimed at microscopic X-CT images of cement hydration. It obtains multi-class image masks through multi-valued segmentation and uses a deep learning network with a double pyramid structure to perform multi-class segmentation of X-CT images, mainly to obtain multi-class volume data of cement hydration images.

[0007] Patent application CN118067577A discloses a method for quantitative analysis of cement clinker minerals that can automatically segment mineral phases. This prior art is mainly aimed at quantitative analysis of cement clinker mineral phases, relying on petrographic images and UNet models for mineral segmentation, and focusing on identifying clinker mineral phases such as C2S, C3S and intermediate phases. Summary of the Invention

[0008] To address one of the shortcomings of existing technologies, the purpose of this application is to provide a method for micro-phase segmentation of cement-based materials based on BSE-EDS multimodal fusion.

[0009] This application provides a method for microphase segmentation of cement-based materials based on BSE-EDS multimodal fusion, including: Provide the cement-based material to be analyzed; BSE images and EDS elemental surface distribution maps were acquired from the same field of view for the cement-based material. Construct the BSE input tensor and the EDS input tensor based on the BSE image and the EDS element surface distribution map; The BSE input tensor and the EDS input tensor are input into a pre-trained bi-branch multimodal fusion network; In each encoding stage of the dual-branch multimodal fusion network, feature enhancement is performed on the BSE encoding branch and the EDS encoding branch respectively, and the channel attention mechanism is combined to obtain the BSE intramodal enhancement features and the EDS intramodal enhancement features. The BSE intramodal enhancement features and the EDS intramodal enhancement features from the same encoding stage are input into the cross-modal adaptive fusion module of the dual-branch multimodal fusion network for adaptive fusion to obtain multi-scale fusion features; The multi-scale fusion features are processed by the decoding output module of the dual-branch multimodal fusion network to obtain a pixel-level phase segmentation map; Based on the phase segmentation diagram, the pixel area fraction of each phase is statistically analyzed, and the area fraction is converted into a three-dimensional volume fraction based on the principle of stereochemistry to obtain the microscopic phase content parameters of the cement-based material.

[0010] Optionally, in the step of acquiring the BSE image and the EDS elemental surface distribution map, the BSE image is a single-channel grayscale image, and the EDS elemental surface distribution map includes at least one elemental channel selected from Ca, Si, Al, Mg, Fe, O, Na, and K.

[0011] Optionally, constructing a BSE input tensor and an EDS input tensor based on the BSE image and the EDS element surface distribution map includes: registering, unifying the size, normalizing the grayscale, and stacking the channels of the BSE image and the EDS element surface distribution map to construct a BSE input tensor B and an EDS input tensor E, where B is a 1×H×W single-channel tensor, E is a CE×H×W multi-channel tensor, and CE is the number of channels of the selected EDS element.

[0012] Optionally, the training process of the dual-branch multimodal fusion network includes: Obtain BSE image samples and EDS element surface distribution map samples; A training dataset is constructed based on the BSE image samples, the EDS elemental surface distribution map samples, and the corresponding phase labels. The phase labels include at least porosity, magnesium oxide, aggregate, impurities, dicalcium silicate / tricalcium silicate, tricalcium aluminate / tetracalcium aluminoferrite, and hydration products. Construct the dual-branch multimodal fusion network and train it based on the training dataset.

[0013] Optionally, the method for obtaining the phase label includes: first obtaining an initial pre-segmentation result based on the grayscale histogram of the BSE image sample and the element response in the EDS element surface distribution map sample; then performing manual refinement in conjunction with the EDS element surface distribution map sample; and mapping the refined boundary back to the BSE image sample to correct the phase boundary, thereby obtaining a pixel-level ground truth label; wherein, the element response refers to the enrichment and combination distribution characteristics of elements in different phase regions in the EDS element surface distribution map sample.

[0014] Optionally, the construction rules for the phase labels include: if the cement-based material to be analyzed contains a phase that needs to be identified separately, a corresponding phase label category is added.

[0015] Optionally, the dual-branch multimodal fusion network includes the BSE encoding branch and the EDS encoding branch set in parallel, as well as the cross-modal adaptive fusion module and the decoding output module connected in sequence; both the BSE encoding branch and the EDS encoding branch contain multiple cascaded encoding stages, and each encoding stage is provided with an adjacent scale feature enhancement module.

[0016] Optionally, within the BSE encoding branch and the EDS encoding branch, the adjacent scale feature enhancement module performs scale alignment, channel compression, feature summation, and convolution mapping on the feature maps of adjacent encoding stages, and combines the channel attention mechanism to obtain intramodal enhanced features.

[0017] Optionally, during the training process, the network parameters of the dual-branch multimodal fusion network are optimized using the cross-entropy loss function, the Dice loss function, or a combination of both.

[0018] Optionally, based on stereochemical principles, the area fraction is converted into a three-dimensional volume fraction to obtain the microstructure phase content parameters of the cement-based material, including: The two-dimensional area fraction is obtained by statistically analyzing the ratio of the number of pixels of each phase category in the phase segmentation diagram to the total number of pixels in the effective analysis area, and then approximated as the three-dimensional volume fraction of the corresponding phase category. Based on the three-dimensional volume fraction, the porosity, hydration product volume fraction, unhydrated clinker volume fraction, and aggregate volume fraction, as well as the combined content parameters of magnesium oxide and impurities, are obtained.

[0019] This application provides a micro-phase segmentation method for cement-based materials based on BSE-EDS multimodal fusion. By deeply fusing the grayscale morphology information contained in the BSE image with the multi-element distribution information contained in the EDS elemental surface distribution map, the method allows the image under analysis to utilize both morphological and chemical composition features in the same processing flow, effectively overcoming the shortcomings of grayscale overlap in a single BSE image and boundary blurring in a single EDS image. Through dual-branch independent encoding and feature enhancement, the method fully mines microstructural information at different scales. Combined with a channel attention mechanism, it improves the model's ability to represent features of small phases (such as pores and MgO) and complex boundaries. The cross-modal adaptive fusion module can dynamically adjust the fusion weights of BSE and EDS features based on the reliability of each modality feature at different encoding stages, avoiding feature redundancy and noise interference caused by simple stitching, and significantly improving the accuracy of pixel-level segmentation. Based on the high-precision pixel-level phase segmentation map, combined with stereoscopic principles, the method accurately obtains micro-phase content parameters, providing reliable data support for interpreting material performance evolution and predicting long-term durability.

[0020] Other technical effects resulting from the additional features will be further illustrated in the corresponding embodiments. Attached Figure Description

[0021] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1This is a flowchart illustrating a method for microphase segmentation of cement-based materials based on BSE-EDS multimodal fusion, according to an exemplary embodiment. Figure 2 This is a flowchart illustrating a method for microphase segmentation of cement-based materials based on BSE-EDS multimodal fusion according to an exemplary embodiment. Figure 3 This is a schematic diagram illustrating the BSE-EDS paired image dataset and label construction process according to an exemplary embodiment; Figure 4 This is a schematic diagram of the structure of a dual-branch multimodal fusion network according to an exemplary embodiment. Detailed Implementation

[0022] The present application will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application, and these all fall within the protection scope of the present application. Parts not described in detail in the following embodiments can be implemented using existing technology.

[0023] Existing methods for segmenting the microstructures of cement-based materials suffer from problems such as overlapping gray levels in a single BSE image, blurred boundaries in a single EDS image, and insufficient segmentation accuracy due to noise interference. To address these issues, this application provides a method for segmenting the microstructures of cement-based materials based on BSE-EDS multimodal fusion.

[0024] Reference Figure 1 and Figure 2 The microphase segmentation method for cement-based materials based on BSE-EDS multimodal fusion in one embodiment of this application includes the following steps: S1. Provide the cement-based material to be analyzed; S2. Acquire BSE images and EDS elemental surface distribution maps of cement-based materials from the same field of view; S3. Construct BSE input tensors and EDS input tensors based on BSE images and EDS element surface distribution maps; S4. Provide a pre-trained dual-branch multimodal fusion network, into which the BSE input tensor and EDS input tensor are input; the dual-branch multimodal fusion network includes a BSE encoding branch, an EDS encoding branch, a cross-modal adaptive fusion module, and a decoding output module; S5. In each encoding stage of the dual-branch multimodal fusion network, feature enhancement is performed on the BSE encoding branch and the EDS encoding branch respectively, and the channel attention mechanism is combined to obtain the BSE intramodal enhancement features and the EDS intramodal enhancement features. S6. Input the BSE intramodal enhancement features and EDS intramodal enhancement features from the same encoding stage into the cross-modal adaptive fusion module of the dual-branch multimodal fusion network for adaptive fusion to obtain multi-scale fusion features; S7. The multi-scale fusion features are processed through the decoding output module of the dual-branch multimodal fusion network to obtain a pixel-level phase segmentation map; S8. Based on the phase segmentation diagram, the pixel area fraction of each phase is statistically analyzed, and based on the principle of stereochemistry, the area fraction is converted into a three-dimensional volume fraction to obtain the microscopic phase content parameters of the cement-based material.

[0025] Specifically, cement-based materials include, but are not limited to, ordinary concrete, recycled aggregate concrete, mortar, cement paste, and cement-based composite materials modified with mineral admixtures or nanomaterials. Microstructures include pores, magnesium oxide, aggregates, impurities, unhydrated clinker phases, and hydration products.

[0026] By deeply fusing the high-resolution grayscale morphology information contained in BSE images with the multi-element distribution information contained in EDS elemental surface distribution maps, the image under analysis can utilize both morphological and chemical composition features in the same processing flow, effectively overcoming the shortcomings of grayscale overlap in single BSE images and boundary blurring in single EDS images. Through bi-branch independent encoding and feature enhancement, microstructural information at different scales is fully explored. Combined with a channel attention mechanism, the model's ability to represent features of small phases (such as pores and MgO) and complex boundaries is improved. The cross-modal adaptive fusion module can dynamically adjust the fusion weights of BSE and EDS features based on the reliability of each modality feature at different encoding stages, avoiding feature redundancy and noise interference caused by simple stitching, and significantly improving the accuracy of pixel-level segmentation. Based on the high-precision pixel-level phase segmentation map, combined with stereoscopic principles, the micro-phase content parameters are accurately obtained, providing reliable data support for interpreting material performance evolution and predicting long-term durability.

[0027] In step S1, a sample of cement-based material to be analyzed is prepared, and the sample cross-section is subjected to termination hydration, drying, vacuum impregnation, grinding, polishing and cleaning to obtain a flat observation surface suitable for BSE and EDS imaging.

[0028] In order to acquire BSE images and EDS elemental surface distribution maps, in some specific embodiments of this application, in step S2, two images are acquired in a scanning electron microscope. The BSE image is a single-channel grayscale image, and the EDS elemental surface distribution map includes at least one elemental channel selected from Ca, Si, Al, Mg, Fe, O, Na, and K.

[0029] Specifically, in this step, the EDS element channels are used as a filterable and combinable multi-channel input, and the EDS element channels are filtered according to their segmentation contribution. When the object to be analyzed is ordinary concrete or recycled aggregate concrete, the above eight element channels are preferably selected as EDS inputs to balance element discrimination ability and redundancy suppression ability.

[0030] To construct the BSE input tensor and EDS input tensor, in some specific embodiments of this application, refer to Figure 3 Step S3 includes: registering, unifying the size, normalizing the grayscale and stacking the channels of the BSE image and the EDS element surface distribution map, and constructing the BSE input tensor B and the EDS input tensor E, where B is a 1×H×W single-channel tensor, E is a CE×H×W multi-channel tensor, and CE is the number of channels of the selected EDS element.

[0031] In some specific embodiments of this application, in step S4, the training process of the dual-branch multimodal fusion network includes: acquiring BSE image samples and EDS element surface distribution map samples; constructing a training dataset based on the BSE image samples, EDS element surface distribution map samples and corresponding phase labels; constructing the dual-branch multimodal fusion network, and training the dual-branch multimodal fusion network based on the training dataset.

[0032] Specifically, BSE image samples and EDS element surface distribution map samples are processed through the same operation in step S3 above to obtain BSE sample input tensors and EDS sample input tensors, which together with the corresponding phase labels constitute the training dataset.

[0033] The phase label includes at least seven types of phases: pores, magnesium oxide (MgO), aggregates, impurities, dicalcium silicate / tricalcium silicate (C2S / C3S), tricalcium aluminate / tetracalcium aluminoferrite (C3A / C4AF), and hydration products, where " / " indicates a mixture formed by the coexistence of the two.

[0034] The method for obtaining phase labels includes: firstly, obtaining initial pre-segmentation results based on the grayscale histogram of BSE image samples and the elemental responses in EDS elemental surface distribution map samples; then, manually refining the results by combining them with EDS elemental surface distribution map samples; and finally, mapping the refined boundaries back to BSE image samples to correct phase boundaries, thereby obtaining pixel-level ground truth labels. Among these, elemental response refers to the enrichment and combination distribution characteristics of elements such as Ca, Si, Al, Mg, Fe, O, Na, and K in different phase regions in the EDS elemental surface distribution map samples.

[0035] By combining pre-segmentation with manual refinement and mapping the refined boundaries back to the BSE image, the accuracy of the phase label boundaries of BSE-EDS paired images can be improved.

[0036] To adapt to the microscopic phase analysis needs of different cement-based materials, in some specific embodiments of this application, the phase label construction rules include: if the cement-based material to be analyzed contains mineral admixtures, unreacted particles, special impurities, or other phases that need to be identified separately, a corresponding phase label category is added.

[0037] In some specific embodiments of this application, such as Figure 4 As shown, the dual-branch multimodal fusion network includes parallel BSE encoding branches and EDS encoding branches, as well as a cross-modal adaptive fusion module and a decoding output module connected in sequence; both the BSE encoding branch and the EDS encoding branch contain multiple cascaded encoding stages, and each encoding stage has an adjacent scale feature enhancement module.

[0038] Specifically, the BSE encoding branch and the EDS encoding branch perform equal-weight feature encoding on the BSE image and the EDS image, respectively, to avoid any modality being used only as supplementary information. The encoding branch adopts a Transformer encoding structure based on shift window attention to capture both local details and global dependencies simultaneously; in other implementations, convolutional neural networks, U-shaped networks, hybrid convolutional-Transformer networks, or other encoding structures with multi-scale feature extraction capabilities can also be used.

[0039] In the embodiments described above, BSE images and EDS element surface distribution maps are used as paired inputs, which can simultaneously utilize grayscale morphology information and element distribution information.

[0040] Furthermore, the paired BSE input tensor, EDS input tensor, and phase labels (i.e., truth labels) are divided into training, validation, and test sets. High-resolution images are cropped into image patches of a preset size and then input into the multimodal fusion network for training. Preferably, the image patch size is 224×224 pixels, and the training, validation, and test sets are divided in a ratio of 6:1.5:2.5.

[0041] In some specific embodiments of this application, the network parameters of the bi-branch multimodal fusion network are optimized during training using the cross-entropy loss function, the Dice loss function, or a combination of both. At least one of OA, IoU, mIoU, and mF1 is used to evaluate model performance, with mIoU and mF1 being prioritized as evaluation metrics in the case of class imbalance.

[0042] For example, the model training uses the Adam optimizer with an initial learning rate of 1e-4, a batch size of 12, and 100 training epochs. These parameters can be adjusted according to the data size and hardware conditions.

[0043] In order to obtain intramodal enhancement features, in some specific embodiments of this application, in step S5, within the BSE coding branch and the EDS coding branch, the feature maps of adjacent coding stages are scale aligned, channel compressed, feature summed and convolutional mapped by the adjacent scale feature enhancement module, and intramodal enhancement features are obtained by combining the channel attention mechanism.

[0044] In the above embodiments of this application, within each modality coding branch, the feature maps of adjacent coding stages are processed by the adjacent scale feature enhancement module, and channel attention is combined to obtain intramodal enhancement features, thereby enhancing the expressive power of small-sized phases and blurred boundary regions.

[0045] To achieve deep feature complementarity between the two heterogeneous microscopic characterization information, in step S6, the intramodal enhancement features of BSE and intramodal enhancement features of EDS are input into the cross-modal adaptive fusion module. Based on spatial correlation and channel saliency, the contribution weights of the two modes at different spatial locations and different channels are calculated. The gray-scale morphology features of BSE and the element distribution features of EDS are adaptively fused to obtain multi-scale fused features.

[0046] In step S7, the multi-scale fusion features are upsampled step by step through the decoding output module and skipped with the fusion features of the corresponding level in the encoding stage to output a pixel-level phase segmentation map with the same size as the input image.

[0047] During the inference phase, the complete BSE-EDS paired images are cropped into image patches and input into the model. The image patch segmentation results output by the model are then stitched together according to the original spatial positions to restore the phase segmentation map under the original field resolution.

[0048] To obtain quantitative results of microscopic phase segmentation, in some specific embodiments of this application, step S8 includes: using the two-dimensional area fraction as an approximate estimate of the three-dimensional volume fraction according to stereoscopic principles; the two-dimensional area fraction is obtained by statistically analyzing a certain phase category N in the phase segmentation diagram. k The number of pixels and the total number of pixels in the effective analysis area N total The ratio is obtained and approximated as the three-dimensional volume fraction of the corresponding phase category; that is, the area fraction A of the k-th phase. k =N k / N total and A k V, representing the volume fraction of this type of phase k Approximate value; Based on three-dimensional volume fractions, parameters such as porosity, hydration product volume fraction, unhydrated clinker volume fraction, aggregate volume fraction, and other phase content are obtained. Porosity is obtained from pore volume fraction, hydration product volume fraction corresponds to the volume fraction of hydration product category, unhydrated clinker volume fraction is obtained by summing the approximate volume fractions of C2S / C3S and C3A / C4AF categories, aggregate volume fraction corresponds to the volume fraction of aggregate category, and MgO and impurities are used together as other phase content parameters for output.

[0049] By statistically analyzing the number of pixels of different phases, multiple types of phases suitable for microstructure analysis of cement-based materials, such as pores, MgO, aggregates, impurities, C2S / C3S, C3A / C4AF, and hydration products, are identified. The area or volume fraction of phases such as pores, unhydrated clinker, aggregates, impurities, and hydration products is obtained, forming a quantitative table of phase content, which can be used for quantitative evaluation of the microstructure of cement-based materials.

[0050] In the embodiments described above, the BSE image is used not only to provide grayscale features but also to preserve the boundary morphology information of phases such as pores, aggregates, impurities, and unhydrated clinker. The EDS elemental surface distribution map is used not only to assist in phase category determination but also to participate in cross-modal feature fusion through multi-channel input, thereby enabling the identification of complex phases, the localization of fuzzy boundaries, and the quantitative calculation of phase content. The aforementioned multimodal fusion segmentation method targets paired BSE-EDS images of cement-based materials, automatically filtering and fusing multi-element information, fully extracting BSE grayscale morphology features and EDS elemental distribution features, and obtaining pixel-level multi-category phase labels and phase content parameters. The embodiments described above solve the problems of overlapping grayscale values ​​in BSE images, blurred boundaries in EDS images, and interference from elemental signals, thereby improving the accuracy of multi-category phase identification and boundary localization in complex cement-based materials.

[0051] The embodiments described above in this application systematically integrate BSE image acquisition, EDS elemental surface distribution map acquisition, paired image preprocessing, element channel filtering, phase label construction, multimodal fusion network training, phase segmentation prediction, and quantitative calculation of phase content. The process is clear, the functional configuration is reasonable, and the degree of automation is high. Through the above-described quantitative characterization process for microscopic phases, the microscopic phases of ordinary concrete, recycled aggregate concrete, and other cement-based composite materials can be automatically identified, segmented, and quantitatively characterized.

[0052] It should be noted that in other embodiments, any appropriate adjustments can be made to the cement-based material type, the number of EDS element channels, the network backbone structure, the loss function, the training parameters, and the output phase category, as long as the same function can be achieved.

[0053] Compared to phase segmentation methods based on BSE grayscale thresholding, grayscale histograms, edge detection, or single BSE deep learning, the embodiments of this application introduce EDS elemental surface distribution maps to supplement chemical composition information, which can improve the ability to distinguish phases with similar grayscale values.

[0054] Compared to methods based on manual interpretation of EDS element ratios, element surface distribution maps, or single EDS image analysis, the above embodiments of this application utilize BSE clear boundary morphology information to constrain EDS boundary diffusion and false detection of elements in pore regions, thereby improving the stability of segmentation boundary localization.

[0055] Compared to simple channel stitching, GMM clustering, or similar statistical methods, the above embodiments of this application dynamically adjust the contributions of the two modes through spatial weights and channel weights, which can suppress the influence of redundant channels, low-confidence element responses, and local noise on the segmentation results.

[0056] In a specific application example, when using BSE and EDS multimodal inputs, the single-class F1 values ​​of the seven phases—porosity, MgO, aggregate, impurities, C2S / C3S, C3A / C4AF, and hydration products—are 75.63%, 92.69%, 92.61%, 80.79%, 89.44%, 74.49%, and 98.10%, respectively. This indicates that the method in the above embodiments of this application can effectively identify phases with different grayscale characteristics, different elemental response characteristics, and different boundary morphology characteristics. The comprehensive evaluation indices of segmentation results, mF1, mIoU, and OA, reached 86.25%, 76.81%, and 96.33%, respectively. Compared with BSE single-modal input (mF1: 56.29%, mIoU: 44.98%, OA: 90.93%) and EDS single-modal input (mF1: 78.48%, mIoU: 68.92%, OA: 95.65%), the method in the above embodiments of this application shows better segmentation performance in terms of class balance evaluation index and overall accuracy.

[0057] The method described in the above embodiments of this application can accurately obtain key phase parameters in the microstructure of cement-based materials, facilitating the analysis of microphase distribution patterns under different material compositions, mix proportions, modification methods, or service conditions. It also clarifies the influence of phases such as pores, unhydrated clinker, aggregates, impurities, and hydration products on material properties and durability. In a specific application example, the average processing time for a 1024×704 pixel image is approximately 0.55 seconds, reducing the workload of manual threshold selection, element ratio interpretation, and post-processing. Furthermore, it maintains consistency in the processing flow across different samples, fields of view, and batches of images, providing a reliable basis for quantitative evaluation of the microstructure of cement-based materials, performance analysis of recycled aggregate concrete, material mix proportion optimization, and durability mechanism research. Finally, it provides technical support for high-throughput automated characterization of the microstructure of cement-based materials.

[0058] The preferred features in the above embodiments can be used individually in any embodiment, or in any combination thereof, provided they do not conflict with each other. Furthermore, parts not described in detail in the embodiments can be implemented using existing technologies.

[0059] The foregoing has described some specific embodiments of this application. It should be understood that this application is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the substantive content of this application. The above-described preferred features can be used in any combination without conflict.

Claims

1. A method for microphase segmentation of cement-based materials based on BSE-EDS multimodal fusion, characterized in that, include: Provide the cement-based material to be analyzed; BSE images and EDS elemental surface distribution maps were acquired from the same field of view for the cement-based material. Construct the BSE input tensor and the EDS input tensor based on the BSE image and the EDS element surface distribution map; The BSE input tensor and the EDS input tensor are input into a pre-trained bi-branch multimodal fusion network; In each encoding stage of the dual-branch multimodal fusion network, feature enhancement is performed on the BSE encoding branch and the EDS encoding branch respectively, and the channel attention mechanism is combined to obtain the BSE intramodal enhancement features and the EDS intramodal enhancement features. The BSE intramodal enhancement features and the EDS intramodal enhancement features from the same encoding stage are input into the cross-modal adaptive fusion module of the dual-branch multimodal fusion network for adaptive fusion to obtain multi-scale fusion features; The multi-scale fusion features are processed by the decoding output module of the dual-branch multimodal fusion network to obtain a pixel-level phase segmentation map; Based on the phase segmentation diagram, the pixel area fraction of each phase is statistically analyzed, and the area fraction is converted into a three-dimensional volume fraction based on the principle of stereochemistry to obtain the microscopic phase content parameters of the cement-based material.

2. The method according to claim 1, characterized in that, In the steps of acquiring BSE images and EDS elemental surface distribution maps, the BSE image is a single-channel grayscale image, and the EDS elemental surface distribution map includes at least one elemental channel selected from Ca, Si, Al, Mg, Fe, O, Na, and K.

3. The method according to claim 1, characterized in that, Constructing BSE input tensors and EDS input tensors based on the BSE image and the EDS element surface distribution map includes: registering, unifying the size, normalizing the grayscale, and stacking the channels of the BSE image and the EDS element surface distribution map to construct BSE input tensor B and EDS input tensor E, where B is a 1×H×W single-channel tensor, E is a CE×H×W multi-channel tensor, and CE is the number of channels of the selected EDS elements.

4. The method according to claim 1, characterized in that, The training process of the dual-branch multimodal fusion network includes: Obtain BSE image samples and EDS element surface distribution map samples; A training dataset is constructed based on the BSE image samples, the EDS elemental surface distribution map samples, and the corresponding phase labels. The phase labels include at least porosity, magnesium oxide, aggregate, impurities, dicalcium silicate / tricalcium silicate, tricalcium aluminate / tetracalcium aluminoferrite, and hydration products. Construct the dual-branch multimodal fusion network and train it based on the training dataset.

5. The method according to claim 4, characterized in that, The method for obtaining the phase labels includes: firstly, obtaining an initial pre-segmentation result based on the grayscale histogram of the BSE image sample and the element response in the EDS element surface distribution map sample; then, manually refining the result by combining it with the EDS element surface distribution map sample; and finally, mapping the refined boundary back to the BSE image sample to correct the phase boundary, thereby obtaining pixel-level ground truth labels; wherein, the element response refers to the enrichment and combination distribution characteristics of elements in different phase regions in the EDS element surface distribution map sample.

6. The method according to claim 4, characterized in that, The rules for constructing phase labels include: if the cement-based material to be analyzed contains a phase that needs to be identified separately, a corresponding phase label category is added.

7. The method according to claim 1, characterized in that, The dual-branch multimodal fusion network includes the BSE encoding branch and the EDS encoding branch set in parallel, as well as the cross-modal adaptive fusion module and the decoding output module connected in sequence; both the BSE encoding branch and the EDS encoding branch contain multiple cascaded encoding stages, and each encoding stage is equipped with an adjacent scale feature enhancement module.

8. The method according to claim 7, characterized in that, Within the BSE encoding branch and the EDS encoding branch, the adjacent scale feature enhancement module performs scale alignment, channel compression, feature summation, and convolution mapping on the feature maps of adjacent encoding stages, and combines the channel attention mechanism to obtain intramodal enhanced features.

9. The method according to claim 4, characterized in that, During the training process, the network parameters of the dual-branch multimodal fusion network are optimized using the cross-entropy loss function, the Dice loss function, or a combination of both.

10. The method according to claim 1, characterized in that, Based on stereochemical principles, the area fraction is converted into a three-dimensional volume fraction to obtain the microstructure phase content parameters of the cement-based material, including: The two-dimensional area fraction is obtained by statistically analyzing the ratio of the number of pixels of each phase category in the phase segmentation diagram to the total number of pixels in the effective analysis area, and then approximated as the three-dimensional volume fraction of the corresponding phase category. Based on the three-dimensional volume fraction, the porosity, hydration product volume fraction, unhydrated clinker volume fraction, and aggregate volume fraction, as well as the combined content parameters of magnesium oxide and impurities, are obtained.

Images