Reconstruction method of continuous section SEM three-dimensional morphology of unhydrated cement particles

Abstract

The application provides a reconstruction method for continuous slicing SEM three-dimensional morphology of unhydrated cement particles, which has the following characteristics and comprises the following steps: step S1, constructing and training an unhydrated cement particle segmentation model based on a deep learning-based instance segmentation network according to an unhydrated cement particle sample data set, performing image segmentation on SEM image data of unhydrated cement particles to be processed by using the unhydrated cement particle segmentation model, and obtaining a predicted result image sequence; step S2, performing image alignment on the predicted result image sequence to obtain an aligned image sequence; and step S3, performing instance repair and three-dimensional morphology reconstruction on the aligned image sequence to obtain a three-dimensional morphology reconstruction image. In summary, the method can efficiently and accurately perform three-dimensional morphology reconstruction and structure quantification on unhydrated cement particles.

Description

Technical Field

[0001] This invention relates to the field of building material analysis, specifically to a method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM. Background Technology

[0002] The hydration process of cement in concrete affects the composition and microstructure of hardened cement paste, thus determining the macroscopic properties of concrete, such as mechanical properties and durability. Because the morphology and distribution characteristics of unhydrated cement particles are important indicators of the composition and microstructure of hardened cement paste, they can be directly used to quantitatively analyze the degree of cement hydration in concrete. Therefore, achieving accurate characterization of the three-dimensional morphology of unhydrated cement particles in hardened cement paste has become a key focus and challenge in current research.

[0003] Serial block-face scanning electron microscopy (SBFSEM) is currently one of the key technologies for characterizing the microscopic three-dimensional morphology of cement-based materials. This technology can directly observe the characteristics of unhydrated cement particles inside the tested cement and perform quantitative analysis. When using serial block-face scanning electron microscopy for SEM imaging, the material being tested is usually embedded in epoxy resin. This avoids the collapse of its three-dimensional structure when slicing in the vacuum environment of the electron microscope. The embedded sample block is then alternately cut by a built-in ultrathin slicer. The newly exposed surface is then imaged using a backscattered electron detector. Image processing steps, including noise reduction and thresholding, are then performed. Finally, the processed sequential slice images are stacked together to obtain the three-dimensional structure of the unhydrated cement particles.

[0004] However, continuous slicing scanning electron microscopy (SEM) has the following drawbacks: First, during the slicing process, the slicing tip inevitably shifts the sample, causing the cut surface to deviate within the imaging field of view. This results in a discrepancy between the reconstructed 3D morphology model and the actual material shape, generally requiring manual image alignment. Second, due to the blurred edges of unhydrated cement particles, traditional segmentation methods cannot segment particles based on edge features. Therefore, threshold-based binary image segmentation methods are still the only option, but this method heavily relies on the operator's experience, leading to unstable results. Third, SEM imaging is accompanied by significant noise, and different denoising methods have a substantial impact on image segmentation, resulting in inconsistent reconstructed 3D models. Therefore, traditional methods cannot efficiently and accurately reconstruct the 3D morphology and quantify the structure of unhydrated cement particles. Summary of the Invention

[0005] This invention is made to solve the above-mentioned problems, and its purpose is to provide a method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM.

[0006] This invention provides a method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles from SEM images. The method includes the following steps: Step S1, based on an unhydrated cement particle sample dataset, a segmentation model for unhydrated cement particles is constructed and trained using a deep learning instance segmentation network. The unhydrated cement particle segmentation model is then used to segment the SEM image data of the unhydrated cement particles to be processed, obtaining a sequence of predicted image results. Step S2, the predicted image sequence is aligned to obtain an aligned image sequence. Step S3, the aligned image sequence is inpainted and its three-dimensional morphology is reconstructed to obtain a reconstructed three-dimensional image. The predicted image includes multiple segmented instances, each of which includes an instance bounding box, an instance mask, and the original image.

[0007] The method for reconstructing the three-dimensional morphology of unhydrated cement particles from continuous slices using SEM provided by this invention may also have the following features: Step S1 includes the following sub-steps: Step S1-1, continuously slice and scan the unhydrated cement particles in the hardened cement paste to obtain SEM images of the unhydrated cement particles, and manually annotate the unhydrated cement particles in the SEM images to obtain an unhydrated cement particle sample dataset; Step S1-2, divide the unhydrated cement particle sample dataset into a training set, a validation set, and a test set; Step S1-3, construct a deep learning model based on a deep learning instance segmentation network; Step S1-4, ... The training set data is augmented to obtain an augmented training set. The augmentation process includes random rotation and flipping. In steps S1-5, the deep learning model is trained based on the augmented training set. The loss function value of the deep learning model is calculated for each iteration. The hyperparameters of the deep learning model are adjusted based on the loss function value using the validation set. The iteration continues until the loss function value converges. The current deep learning model is then used as the unhydrated cement particle segmentation model. In steps S1-6, the predicted result images are obtained based on the unhydrated cement particle SEM image data to be processed using the unhydrated cement particle segmentation model. The predicted result images are arranged according to the slice sequence to obtain the predicted result image sequence.

[0008] The method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles by SEM provided by the present invention may also have the following feature: wherein, in steps S1-2, the ratio of the number of data in the training set, validation set and test set is 3:1:1.

[0009] The method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles by SEM provided by the present invention may also have the following feature: wherein, in steps S1-5, the loss function is a generalized loss function TverskyLoss based on the Tversky index.

[0010] The method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM provided by this invention may also have the following features: Step S2 includes the following sub-steps: Step S2-1: Perform ORB feature detection on each prediction result image in the prediction result image sequence to obtain feature points describing the pixel features of the image. The feature points include a feature point locator and a feature point descriptor. The feature point locator uses the FAST algorithm to identify and label the feature points, and the feature point descriptor uses the BRIEF algorithm to calculate the combination relationship between the labeled feature points and surrounding points as the descriptor; Step S2-2: Perform feature point matching on the feature points in two adjacent prediction result images in the prediction result image sequence to obtain matching points, and then sort the matching points according to the feature points... The feature matching scores are sorted, and matching points with scores greater than the matching threshold are retained as the matching point set. The matching point is a pair of feature points with the same features in two predicted result images. The feature matching score is obtained by measuring the similarity of the descriptors of two feature points according to the Hamming distance. Step S2-3: The matching point set is estimated with a random sampling consensus algorithm to obtain high-precision matching points as high-precision matching points. The homography matrix between two adjacent predicted result images in the predicted result image sequence is calculated based on the high-precision matching points, so as to perform alignment correction on the two predicted result images. Step S2-4: Steps S2-1 to S2-3 are repeated to perform alignment correction on each predicted result image in the predicted result image sequence from front to back according to the slice order to obtain the aligned image sequence.

[0011] The method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles by SEM provided by the present invention may also have the following feature: wherein, in step S2-2, the matching threshold value ranges from 0.2 to 0.5.

[0012] The method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM provided by this invention may also have the following feature: step S3 includes the following sub-steps: step S3-1, in the aligned image sequence, the predicted result image Previous sequence The predicted image is used as the upper search domain. The predicted image The subsequent sequence Zhang's predicted image is used as the next search domain. Search domain Search Domain In the formula Step S3-2, process the predicted image. Search Domain and the search domain For each segmented instance in the predicted image, the overlap of the instance bounding box and the overlap of the instance mask are calculated to obtain the IoU value of the instance bounding box and the IoU value of the instance mask for each segmented instance. Then, based on the upper search domain... and the search domain Corresponding prediction result image The instance bounding box IoU value and instance mask IoU value of a certain region are used to perform or skip instance inpainting on that region in the predicted image; Step S3-3, repeat steps S3-1 to S3-2, and perform unhydrated cement particle instance inpainting on the images in the aligned image sequence from front to back according to the slice order to obtain the repaired image sequence; Step S3-4, extract each instance mask in the repaired image sequence, pixelate the instance mask to obtain a pixelated mask, and stack the pixelated mask sequentially in the normal direction of the image plane according to the slice order to form a three-dimensional topography reconstruction image. The judgment for performing or skipping instance inpainting in step S3-2 is as follows: when the predicted image There are no segmented instances in a certain region, and the upper search domain... and the search domain Each in the same area The instance bounding box IoU value and instance mask IoU value above the layer are both greater than the threshold. At that time, for the predicted image Instance repair is performed on this area when the predicted image is displayed. An instance exists in a certain region, and the search domain is above it. Prediction result image Or search domain Prediction result image The IoU values ​​of the instance bounding boxes and the instance mask in the same region are greater than the threshold. At that time, the predicted image This region will not undergo instance repair. The shape and area of ​​the instance repair will be determined based on the predicted image. Nearest neighbor prediction image and The intersection of instance masks in the corresponding regions is determined.

[0013] The method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM provided by this invention may also have the following feature: wherein, in step S3-2, The value range is 10~50, and the threshold is... The value range is 0.3 to 0.6.

[0014] The method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM provided by this invention may also have the following feature: wherein, in steps S3-4, the pixel size of the pixelated mask image is the same as that of the original image, and the thickness of the image plane in the normal direction is... 1 pixel, The value range is 10 to 20.

[0015] The method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM provided by this invention may also have the following features: the instance segmentation network algorithm is a path enhancement network algorithm, and the feature extraction network in the path enhancement network algorithm is a ResNet50 network.

[0016] The role and effect of invention

[0017] According to the method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM, the present invention first constructs and trains an unhydrated cement particle segmentation model based on a deep learning instance segmentation network using an unhydrated cement particle sample dataset. This unhydrated cement particle segmentation model then performs image segmentation on the SEM image data of the unhydrated cement particles to be processed, obtaining a sequence of predicted image results. Next, the predicted image result sequence is aligned to obtain an aligned image sequence. Finally, instance repair and three-dimensional morphology reconstruction are performed on the aligned image sequence to obtain a reconstructed three-dimensional morphology image. Therefore, the method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM can efficiently and accurately reconstruct the three-dimensional morphology and quantify the structure of unhydrated cement particles. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM in an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the SEM image data of unhydrated cement particles to be processed in an embodiment of the present invention;

[0020] Figure 3 This is a schematic diagram of the prediction result image in an embodiment of the present invention;

[0021] Figure 4 This is a comparison chart showing the effects of the traditional threshold segmentation method and the unhydrated cement particle segmentation model on segmenting unhydrated cement particles in embodiments of the present invention;

[0022] Figure 5 This is a schematic diagram of the feature point matching results of two adjacent prediction result images in an embodiment of the present invention;

[0023] Figure 6This is a schematic diagram illustrating the effect of alignment correction on two predicted result images in an embodiment of the present invention;

[0024] Figure 7 This is a schematic diagram illustrating the effect of reconstructing the three-dimensional shape of the image sequence after instance repair in an embodiment of the present invention;

[0025] Figure 8 This is a comparison diagram of the effects of traditional three-dimensional topography reconstruction method and the method of the present invention on three-dimensional topography reconstruction in embodiments of the present invention. Detailed Implementation

[0026] To make the technical means, creative features, objectives and effects of this invention easier to understand, the following embodiments, in conjunction with the accompanying drawings, specifically illustrate the method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM.

[0027] Figure 1 This is a schematic flowchart of the method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM in an embodiment of the present invention.

[0028] like Figure 1 As shown, the method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM in an embodiment of the present invention includes the following steps:

[0029] Step S1-1: Continuously slice and scan the unhydrated cement particles in the hardened cement paste to obtain SEM images of the unhydrated cement particles. The unhydrated cement particles in the SEM images of the unhydrated cement particles are labeled by manual annotation to obtain a sample dataset of unhydrated cement particles.

[0030] Step S1-2: Divide the dataset of unhydrated cement particles into a training set, a validation set, and a test set, with the ratio of data quantity in the training set, validation set, and test set being 3:1:1.

[0031] Steps S1-3: Construct a deep learning model based on a deep learning instance segmentation network. The instance segmentation network algorithm is a path enhancement network algorithm, and the feature extraction network in the path enhancement network algorithm is a ResNet50 network.

[0032] Steps S1-4 involve performing data augmentation on the training set data to obtain an augmented training set. The data augmentation process includes random rotation and flipping.

[0033] Steps S1-5: Train the deep learning model using the enhanced training set, calculate the loss function value of the deep learning model for each iteration, adjust the hyperparameters of the deep learning model based on the loss function value using the validation set, and iterate until the loss function value converges. Then, use the current deep learning model as the unhydrated cement particle segmentation model.

[0034] The loss function is the Tversky Loss, a generalized loss function based on the Tversky exponent.

[0035] Steps S1-6: Based on the SEM image data of the unhydrated cement particles to be processed, the predicted result image is obtained through the unhydrated cement particle segmentation model. The predicted result image includes multiple segmentation instances, each of which includes an instance bounding box, an instance mask, and the original image. The predicted result images are arranged according to the slice sequence to obtain the predicted result image sequence.

[0036] Figure 2 This is a schematic diagram of the SEM image data of unhydrated cement particles to be processed in an embodiment of the present invention.

[0037] Figure 3 This is a schematic diagram of the predicted result image in an embodiment of the present invention.

[0038] Figure 4 This is a comparison chart showing the effects of traditional threshold segmentation and unhydrated cement particle segmentation model on segmenting unhydrated cement particles in embodiments of the present invention.

[0039] like Figure 2 , Figure 3 As shown, Figure 2 The SEM image data of the unhydrated cement particles shown is processed using an unhydrated cement particle segmentation model to obtain the following results: Figure 3 The image showing the predicted results. Figure 3 The dark gray areas represent the segmented instances, the dark gray boxes represent the instance bounding boxes, the dark gray areas within the dark gray boxes represent the instance masks, and the non-dark gray areas within the dark gray boxes represent the original image.

[0040] like Figure 4 As shown, Figure 4 The left side shows the effect of segmenting unhydrated cement particles using the traditional threshold segmentation method. Figure 4 The image on the right shows the effect of segmenting unhydrated cement particles using a segmentation model. Figure 4 The gray area on the left represents the actual unhydrated particles, while the black area represents the unhydrated particles segmented by the traditional threshold segmentation method. Only a portion of the segmented unhydrated particles overlap with the actual unhydrated particles. Figure 4 The gray area on the right represents the actual unhydrated particles, while the black area represents the unhydrated particles segmented by the unhydrated cement particle segmentation model. Most of the segmented unhydrated particles overlap with the actual unhydrated particles. Therefore, the unhydrated cement particle segmentation model of this invention has a better effect on segmenting unhydrated cement particles compared to traditional threshold segmentation methods.

[0041] Step S2-1: Perform ORB feature detection on each predicted image in the predicted image sequence to obtain feature points that describe the pixel features of the image. The feature points include feature point locators and feature point descriptors. The feature point locators use the FAST algorithm to identify and label feature points, and the feature point descriptors use the BRIEF algorithm to calculate the combination relationship between the labeled feature points and surrounding points as descriptors.

[0042] Step S2-2: For feature points in two adjacent prediction result images in the prediction result image sequence, feature point matching is performed to obtain matching points. The matching points are sorted according to the feature matching degree, and the matching points with a matching threshold are retained as the matching point set. The matching point is a pair of feature points with the same features in the two prediction result images. The feature matching degree is obtained by measuring the similarity of the two feature point descriptors according to the Hamming distance. The matching threshold ranges from 0.2 to 0.5.

[0043] Figure 5 This is a schematic diagram of the feature point matching results of two adjacent prediction result images in an embodiment of the present invention.

[0044] like Figure 5 As shown, the two images on the left and right are two adjacent prediction result images of the sequence, and the two ends of the horizontal line are a pair of matching feature points.

[0045] Step S2-3: Use the random sampling consensus algorithm to estimate high-precision matching points for the matching point set as high-precision matching points. Calculate the homography matrix between two adjacent predicted result images in the predicted result image sequence based on the high-precision matching points, and then perform alignment correction on the two predicted result images.

[0046] Figure 6 This is a schematic diagram illustrating the effect of aligning and correcting two predicted result images in an embodiment of the present invention.

[0047] like Figure 6 As shown, after alignment correction of two adjacent prediction result images in the sequence, the black edge part is the result of cropping according to the original image size and adjusting relative to the reference image.

[0048] Step S2-4: Repeat steps S2-1 to S2-3 to align and correct each predicted image in the slice order from front to back, resulting in an aligned image sequence. During the alignment and correction process, the predicted image to be aligned and corrected is aligned and corrected with the nearest preceding predicted image that has already been aligned and corrected.

[0049] Step S3-1: In the aligned image sequence, the predicted image is... Previous sequence The predicted image is used as the upper search domain. The predicted image The subsequent sequence Zhang's predicted image is used as the next search domain. Search domain Search Domain In the formula .

[0050] Step S3-2, process the predicted image. Search Domain and the search domain For each segmented instance in the predicted image, the overlap of the instance bounding box and the overlap of the instance mask are calculated to obtain the IoU value of the instance bounding box and the IoU value of the instance mask for each segmented instance. Then, based on the upper search domain... and the search domain Corresponding prediction result image The instance bounding box IoU value and instance mask IoU value of a certain region are used to determine whether to perform or skip instance inpainting for that region in the predicted image. The determination of whether to perform or skip instance inpainting is as follows:

[0051] When the predicted image There are no segmented instances in a certain region, and the upper search domain... and the search domain Each in the same area The instance bounding box IoU value and instance mask IoU value above the layer are both greater than the threshold. At that time, for the predicted image Instance repairs will be performed in this area;

[0052] When the predicted image An instance exists in a certain region, and the search domain is above it. Prediction result image Or search domain Prediction result image The IoU values ​​of the instance bounding boxes and the instance mask in the same region are greater than the threshold. At that time, the predicted image This region will not undergo instance repair. The shape and area of ​​the instance repair will be determined based on the predicted image. Nearest neighbor prediction image and The intersection of instance masks in the corresponding regions is determined.

[0053] in, The value range is 10~50, and the threshold is... The value range is 0.3 to 0.6.

[0054] Figure 7This is a schematic diagram illustrating the effect of reconstructing the three-dimensional shape of the image sequence after instance repair in an embodiment of the present invention.

[0055] like Figure 7 As shown, the predicted image When performing instance inpainting on a region of the (i.e., the intermediate layer image), the image is based on the prediction result. (i.e., the upper layer image) and the predicted result image The intersection of instance masks in the corresponding regions (i.e., the lower layer image) was used to repair and obtain the predicted image. The instance mask for this region is used to reconstruct its 3D topography. Figure 7 The resulting image is shown.

[0056] Step S3-3: Repeat steps S3-1 to S3-2 to repair unhydrated cement particle instances in the images in the aligned image sequence from front to back according to the slice order, and obtain the repaired image sequence.

[0057] Steps S3-4: Extract each instance mask from the repaired image sequence, pixelate the instance masks to obtain pixelated masks, and stack the pixelated masks sequentially in the normal direction of the image plane according to the slice order to form a three-dimensional topography reconstruction image.

[0058] In this image, the pixel size of the pixelated mask is the same as that of the original image, and the thickness of the image plane in the normal direction is... 1 pixel, The value range is 10 to 20.

[0059] Figure 8 This is a comparison diagram of the effects of traditional three-dimensional topography reconstruction method and the method of the present invention on three-dimensional topography reconstruction in embodiments of the present invention.

[0060] like Figure 8 As shown, the left image is the 3D morphology reconstruction effect obtained using the traditional 3D morphology reconstruction method, and the right image is the 3D morphology reconstruction effect obtained using the method of the present invention. The black parts in the images represent the parts where there is an error between the 3D morphology reconstruction effect and the actual morphology of unhydrated cement particles. It can be seen that the 3D morphology reconstruction method of continuous slices of unhydrated cement particles by SEM of the present invention can efficiently and accurately reconstruct the 3D morphology and quantify the structure of unhydrated cement particles compared with the traditional 3D morphology reconstruction method.

[0061] The role and effect of the embodiments

[0062] According to the method for reconstructing the 3D morphology of continuous slices of unhydrated cement particles using SEM in this embodiment, a deep learning-based unhydrated cement particle segmentation model, image alignment operation, and instance repair operation are employed. On the one hand, the unhydrated cement particle segmentation model avoids manual annotation and improves the recognition speed and annotation accuracy of unhydrated cement particles. On the other hand, image alignment operation enables automated image alignment, and instance repair operation repairs image loss caused by sample loss during slicing, thereby reducing sample errors caused by slicing operations. In summary, this method can efficiently and accurately reconstruct the 3D morphology and quantify the structure of unhydrated cement particles.

[0063] The above embodiments are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention.

Claims

1. A method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM, characterized in that, Includes the following steps: Step S1: Based on the unhydrated cement particle sample dataset, a segmentation model for unhydrated cement particles is constructed and trained using a deep learning instance segmentation network. The unhydrated cement particle segmentation model is then used to segment the SEM image data of the unhydrated cement particles to be processed, and a sequence of predicted image results is obtained. Step S2: Align the predicted image sequence to obtain an aligned image sequence; Step S3: Perform instance repair and 3D topography reconstruction on the aligned image sequence to obtain a 3D topography reconstructed image. Step S3 includes the following sub-steps: Step S3-1, in the aligned image sequence, the predicted image is... Previous sequence The predicted image described above is used as the upper search domain. The predicted image The subsequent sequence The predicted image described above is used as the next search domain. The upper search domain The lower search domain In the formula ; Step S3-2, for the predicted result image The upper search domain and the search domain For each segmented instance in the predicted image, the overlap of the instance bounding box and the overlap of the instance mask are calculated to obtain the instance bounding box IoU value and the instance mask IoU value for each segmented instance. Then, based on the upper search domain... and the search domain Corresponding to the predicted result image The instance bounding box IoU value and the instance mask IoU value of a certain region are used to perform or skip instance inpainting on that region of the predicted image; Step S3-3: Repeat steps S3-1 to S3-2 to repair unhydrated cement particle instances in the images in the aligned image sequence from front to back according to the slice order, and obtain the repaired image sequence. Steps S3-4 involve extracting each instance mask from the repaired image sequence, pixelating the instance masks to obtain pixelated masks, and stacking the pixelated masks sequentially in the normal direction of the image plane according to the slice order to form the three-dimensional topography reconstruction image. The determination of whether to perform or skip instance repair in step S3-2 is as follows: When the predicted result image There are no segmented instances in a certain region, and the upper search domain... and the search domain Each in the same area The instance bounding box IoU value and the instance mask IoU value above the layer are both greater than the threshold. At that time, for the predicted result image Instance repair is performed in this area. When the predicted result image An instance exists in a certain region, and the upper search domain... Prediction result image Or search domain Prediction result image The instance bounding box IoU value and the instance mask IoU value in the same region are greater than a threshold. At that time, the predicted result image No instance repair will be performed in this area. The shape and area of ​​the instance repair are based on the predicted result image. Nearest neighbor prediction image and The intersection of instance masks in the corresponding regions is determined. The predicted image includes multiple segmentation instances. The segmentation instance includes the instance bounding box, the instance mask, and the original image.

2. The method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM according to claim 1. Its features are: in, Step S1 includes the following sub-steps: Step S1-1: Continuously slice and scan the unhydrated cement particles in the hardened cement paste to obtain SEM images of the unhydrated cement particles. Label the unhydrated cement particles in the SEM images of the unhydrated cement particles by manual annotation to obtain the unhydrated cement particle sample dataset. Step S1-2: Divide the dataset of unhydrated cement particles into a training set, a validation set, and a test set; Steps S1-3: Construct a deep learning model based on the instance segmentation network of the deep learning; Steps S1-4: Perform data augmentation on the training set data to obtain an augmented training set. The data augmentation includes random rotation and flipping. Steps S1-5: Train the deep learning model according to the enhanced training set, calculate the value of the loss function of the deep learning model in each iteration, adjust the hyperparameters of the deep learning model according to the value of the loss function through the validation set, and iterate until the loss function value converges. Then, use the current deep learning model as the unhydrated cement particle segmentation model. Steps S1-6: Based on the SEM image data of the unhydrated cement particles to be processed, the predicted result image is obtained through the unhydrated cement particle segmentation model. The predicted result image is then arranged according to the slice sequence to obtain the predicted result image sequence.

3. The method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM according to claim 2, characterized in that: in, In steps S1-2, the ratio of the number of data in the training set, the validation set, and the test set is 3:1:

1.

4. The method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM according to claim 2, characterized in that: in, In steps S1-5, the loss function is the generalized loss function TverskyLoss based on the Tversky exponent.

5. The method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM according to claim 1. Its features are: in, Step S2 includes the following sub-steps: Step S2-1: Perform ORB feature detection on each of the predicted result images in the predicted result image sequence to obtain feature points describing the pixel features of the image. The feature points include a feature point locator and a feature point descriptor. The feature point locator uses the FAST algorithm to identify and label the feature points. The feature point descriptor uses the BRIEF algorithm to calculate the combination relationship between the labeled feature points and surrounding points as a descriptor. Step S2-2: For feature points in two adjacent prediction result images in the prediction result image sequence, feature point matching is performed to obtain matching points. The matching points are sorted according to the feature matching degree, and the matching points with a matching threshold are retained as the matching point set. The matching points are a pair of feature points with the same features in the two prediction result images. The feature matching degree is obtained by measuring the similarity of the two feature point descriptors according to Hamming distance. Step S2-3: Use the random sampling consensus algorithm to estimate high-precision matching points for the matching point set as high-precision matching points. Calculate the homography matrix between two adjacent predicted result images in the predicted result image sequence based on the high-precision matching points, thereby performing alignment correction on the two predicted result images. Step S2-4: Repeat steps S2-1 to S2-3 to perform alignment correction on each of the predicted result images in the slice order from front to back, to obtain the aligned image sequence.

6. The method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM according to claim 5, characterized in that: in, In step S2-2, the matching threshold ranges from 0.2 to 0.

5.

7. The method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM according to claim 1, characterized in that: in, In step S3-2, the The value range is 10~50, and the threshold value is... The value range is 0.3 to 0.

6.

8. The method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM according to claim 1, characterized in that: in, In steps S3-4, the pixel size of the pixelated mask is the same as that of the original image, and the thickness of the image plane in the normal direction is... 1 pixel, the The value range is 10 to 20.

9. The method for reconstructing the three-dimensional morphology of continuous slices of unhydrated cement particles using SEM according to claim 1, characterized in that: in, The algorithm for the instance segmentation network is the path enhancement network algorithm, and the feature extraction network in the path enhancement network algorithm is the ResNet50 network.

Images