A statistical feature-based method for augmenting composite material XCT images
By using a statistical feature-based method, pseudo-XCT image data is generated using polygon fitting and VAE models, which solves the problem of data scarcity in XCT image segmentation of basalt fiber composite materials and achieves efficient and low-cost data augmentation and accurate segmentation.
Patent Information
- Application Number
- CN202610032884.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-12
- Publication Date
- 2026-06-02
AI Technical Summary
In existing technologies, XCT image segmentation methods for basalt fiber composite materials rely on manual interaction, which is inefficient and easily affected by subjective experience. Furthermore, deep learning models are difficult to train to achieve high generalization ability and high accuracy due to data scarcity and inconsistent quality.
A statistical feature-based approach is adopted, which learns the yarn centroid distribution through polygon fitting and VAE model, and combines multi-peak Gaussian fitting and Brownian bridge diffusion model to generate high-quality pseudo-XCT image data, thereby achieving data augmentation.
It breaks through the bottleneck of small sample training, reduces the cost of data acquisition and manual annotation, improves the generalization ability and recognition accuracy of the segmentation network, and generates pseudo-labels with smooth edges and clear categories, significantly improving the recognition accuracy of yarn edges and tiny gaps.
Smart Images
Figure CN122134739A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application relates to an image generation method, in particular to a composite material XCT image data enhancement method based on statistical characteristics, and belongs to the technical field of image processing. BACKGROUND
[0002] Basalt fiber composite materials, as a new type of high-performance material, have been widely concerned and applied in the engineering field due to their excellent mechanical properties and outstanding damage resistance. However, the macroscopic mechanical properties of the composite materials are largely dependent on the integrity and uniformity of the microstructure, such as fiber distribution density, fiber spatial orientation, interface bonding strength and internal crack rate, which will significantly affect the final service performance of the material. Therefore, it is of important theoretical value and practical significance to conduct in-depth research on the microstructure characteristics of basalt fiber composite materials and the influence mechanism on the performance.
[0003] X-ray computed tomography (XCT) is an advanced non-contact non-destructive testing technology, which can reconstruct the internal three-dimensional structure of materials with high resolution. Through XCT scanning, key information such as the spatial form of internal yarns, fiber arrangement and defect distribution of basalt fiber composite materials can be obtained non-destructively, providing important data support for material performance evaluation and optimization.
[0004] When analyzing the XCT images of basalt fiber composite materials, it is usually necessary to extract yarn features through image segmentation technology. However, the existing image segmentation methods have significant limitations: on the one hand, traditional segmentation methods highly depend on manual interaction, not only inefficient, but also the results are easily affected by subjective experience, making it difficult to ensure consistency; on the other hand, due to the chemical composition characteristics of basalt fibers, the attenuation coefficient of X-rays is low, resulting in low contrast, uneven gray distribution and noise interference in the imaging results, and traditional threshold-based or edge detection-based segmentation algorithms are difficult to extract single yarn-level precision from low-contrast backgrounds.
[0005] In recent years, image segmentation networks based on deep learning have become a research hotspot for solving the above problems due to their strong feature extraction capability. Such networks can theoretically automatically and efficiently segment fiber bundle targets in XCT images. However, the performance of deep learning models highly depends on the size and quality of the training data set. In practical applications, high-quality XCT images of basalt fiber composite materials are costly and time-consuming to obtain, and are limited by equipment parameters, scanning environment and material characteristics. The imaging quality of the original data is uneven. Relying solely on limited original data to create a data set makes it difficult to train a segmentation model with high generalization ability and high precision, which cannot meet the needs of actual industrial detection.
[0006] Therefore, in order to solve the above problems, it is necessary to provide an innovative statistical feature-based composite material XCT image data enhancement method to overcome the technical problems of lack of instance segmentation dataset and insufficient sample diversity in the prior art. SUMMARY
[0007] The purpose of the present application is to provide a statistical feature-based composite material XCT image data enhancement method which can accurately reproduce the internal yarn topology in three-dimensional woven composites, thereby improving the detection accuracy of composites.
[0008] To achieve the above-mentioned purpose, the technical solution adopted by the present application is as follows: a statistical feature-based composite material XCT image data enhancement method, comprising the following steps: 1) Obtain the original XCT image of the four-step three-dimensional woven composite material, and perform manual annotation to obtain the accurate yarn contour mask; then, use a polygon fitting algorithm to process the yarn contour, convert the irregular yarn edge into a geometric polygon description, and extract the key statistical features of each yarn; the key statistical features include the centroid coordinate data of the yarn contour and the approximate polygon size parameters; 2) Based on the key statistical features extracted in step 1), construct and train a VAE model, input the centroid coordinates of the yarn contour, and make the VAE model learn and capture the distribution probability density and topological arrangement of the yarn centroid in three-dimensional space; at the same time, use a multi-peak Gaussian fitting algorithm to statistically analyze the approximate polygon size parameters extracted in step 1), and fit a probability distribution model of the yarn cross-sectional size to represent the morphological variation characteristics of the yarn at different positions; 3) Use the decoder part of the trained VAE model to sample and generate on a blank background image, output the predicted centroid coordinates conforming to the true distribution law; then, use the generated centroid coordinates as positioning points, and generate the corresponding polygon geometry contour according to the size distribution constraint conditions determined by the multi-peak Gaussian fitting algorithm; map all the generated polygon contours to the blank image to construct instance segmentation pseudo-labels with clear class and position information; 4) Input the instance segmentation pseudo-labels generated in step 3) as a condition, import them into a pre-trained style conversion network based on the Brown Bridge diffusion model, use the network to learn the noise distribution, gray attenuation characteristics and texture details of the original XCT image, map the geometric pseudo-labels to pseudo-XCT images, complete the cross-domain generation from pseudo-labels to pseudo-XCT images, and realize XCT image data enhancement.
[0009] The statistical feature-based composite material XCT image data enhancement method of the present application further comprises the following steps: 1-1), the XCT device is used for tomography scanning on a four-step three-dimensional braided composite material sample; projection data obtained by scanning is reconstructed into three-dimensional body data, and a series of cross-section slice images are cut along the braiding axis by using an image labeling software Avizo; 1-2), yarn profiles in the XCT image of the three-dimensional braided composite material are manually labeled by using Avizo, points are drawn along the yarn and matrix interface region to form a closed region, and meanwhile, it is ensured that the profile interface regions of single yarns are not connected to each other; after labeling, a corresponding gray scale mask image is generated; 1-3), the above-mentioned labeling file is read, and edge profile point sets of each yarn instance are extracted by using a computer vision algorithm; then, polygon fitting dimension reduction processing is performed, and according to the yarn profile shape, the yarn profile shape can be divided into three shapes, i.e., an ellipse, a hexagon and a sector; 1-4), for each polygon yarn instance after fitting, two types of key statistical features are calculated and extracted, i.e., a centroid coordinate data, and the centroid coordinates (Cx n , Cy n ) of each yarn profile are collected by using a findContours function; and an approximate polygon size parameter, the IOU value between each yarn profile and the original yarn profile is calculated after fitting the three shapes respectively, the highest value is taken as the fitting shape of the current yarn profile, and the corresponding shape size parameter is extracted.
[0010] The composite material XCT image data enhancement method based on statistical features further comprises that in the step 1-1), every 10th slice is selected as an original data set sample.
[0011] The composite material XCT image data enhancement method based on statistical features further comprises that the step 2) specifically comprises: 2-1), the yarn centroid coordinate set of all slice images extracted in the step 1) is defined as input data X ={(Cx1, Cy1),(Cx2, Cy2), …, (Cx n , Cy n )}, wherein each represents a two-dimensional coordinate set of all yarn centroids on a slice; the number of centroids on each slice is unified, and the coordinates are normalized to the interval [-1, 1]; 2-2), after data preprocessing, the VAE model is used for training, a multilayer perceptron or a one-dimensional convolutional neural network is used as a feature extractor for the encoder, the decoder structure is symmetrical with the encoder, the input is a latent vector z, the reconstructed output predicted yarn random centroid coordinate set and the yarn profile shape type are mapped through a plurality of fully connected layers; 2-3), the parameter distribution of the ellipse, hexagon and sector extracted in step 1-3) is fitted and modeled by using a multi-peak Gaussian fitting algorithm to obtain a fitting function, so that the probability distribution model of the yarn cross-sectional size is fitted.
[0012] The statistical feature-based composite material XCT image data enhancement method of the application further comprises the following steps:
[0013] The statistical feature-based composite material XCT image data enhancement method of the application further comprises the following steps: 2-2-1), reading the slice data, and performing one-hot encoding processing on the label representing the yarn type; the original category is mapped to three binary features, namely ellipse, hexagon and sector. If a type is missing in the sample, the corresponding column is automatically filled with zero; 2-2-2), the data is perspective reorganized by taking the picture file name as the index and the yarn number as the column; the system checks whether each picture contains complete 48 yarn points, if there is data missing in a picture, it is regarded as an invalid sample and directly excluded, only the valid picture data with complete 48 points is retained; 2-2-3), the center coordinates (Cx i , Cy i ) of each yarn point are normalized; after normalization, the data is encoded, and the specific format is {(Cx i , Cy i ), type_Ellipse, type_Hexagon, type_Sector}, wherein one of the three type_ data is 1 and the others are 0, which represents the fitting shape type to which the current coordinate point belongs.
[0014] The statistical feature-based composite material XCT image data enhancement method of the application further comprises the following steps: 3-1), using the VAE model trained in step 2) to perform forward inference on the blank slice to generate random 48 yarn profile centroid coordinates (Cx' n , Cy' n ); 3-2), for each generated centroid coordinate (Cx' n , Cy' n ), a specific geometric shape is assigned to it to generate a centroid set Each centroid coordinate in the image is taken as the origin, and the corresponding contour shape is generated by combining the size constraints of the fitting function described in steps 2-3). Then, the corresponding gray value pixels are filled inside the contour with the gray value of yarn ID number + 80. The final generated image is called a pseudo label.
[0015] The composite material XCT image data enhancement method based on statistical features of the present invention further comprises: in step 4), when pre-training the Brownian bridge diffusion model, inputting the true training set, i.e., the true labels and the true original image, and outputting the training weights .pth file after learning.
[0016] The statistical feature-based composite material XCT image data enhancement method of the present invention can also be as follows: In step 4), the specific method for generating pseudo-XCT images is as follows: input the pseudo labels from step 3-2) into the trained Brownian bridge diffusion model, and the model converts the pseudo label style into pseudo-XCT images by calling the training weights.pth file.
[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. The statistical feature-based composite material XCT image data enhancement method of this invention breaks through the bottleneck of training with small sample data, and realizes low-cost and high-efficiency dataset expansion. Existing technologies for training high-precision deep learning segmentation models usually rely on massive amounts of manually labeled data, while acquiring XCT images of 3D woven composite materials is expensive and time-consuming to label. However, this invention, through statistical modeling and generative network strategies, only requires a small number (dozens) of original slices to extract feature patterns, and then automatically generates tens of thousands of significantly differentiated pseudo-data. This not only greatly reduces the cost of data acquisition and manual labeling, but also effectively solves the problem of overfitting of deep learning models caused by insufficient training samples, and significantly improves the generalization ability of the segmentation network.
[0018] 2. The statistical feature-based composite material XCT image data enhancement method of this invention employs fully automated generation of instance segmentation labels with pixel-level precision, eliminating manual annotation errors. In traditional processes, even with images, manual annotation of complex yarn contours easily introduces subjective errors and edge jitter. This invention, however, automatically generates corresponding pseudo-original images simultaneously with the pseudo-images. Because these labels are directly drawn based on mathematical coordinates, they have smooth edges, clear categories, high instance discrimination, and perfect pixel-level alignment with the generated pseudo-images. Using these zero-error perfect labels to train the downstream segmentation network significantly improves the network's accuracy in recognizing yarn edges, adhesion areas, and minute gaps. Attached Figure Description
[0019] Figure 1 This is a flowchart of the composite material XCT image data enhancement method based on statistical features according to the present invention.
[0020] Figure 2 This is a schematic diagram of the polygon fitting algorithm in step 1) of the present invention.
[0021] Figure 3 and Figure 4 This is a schematic diagram of the parameter distribution fitting function modeling in step 2) of the present invention.
[0022] Figure 5 This is the result of generating VAE coordinates, pseudo-labels, and pseudo-original images in step 3) of the present invention.
[0023] Figure 6 This is the schematic diagram of the Brown Bridge style transfer network in step 4 of this invention. Detailed Implementation
[0024] Please refer to the instruction manual appendix. Figure 1 As shown, this invention is a method for enhancing XCT image data of composite materials based on statistical features, which includes the following steps: 1) As per the instruction manual Figure 2 As shown, the original XCT image of the four-step 3D braided composite material is acquired, and manually annotated to obtain accurate yarn contour masks, ensuring that the contour interface areas of individual yarns do not stick together during the annotation process, forming independent instances. Subsequently, a polygon fitting algorithm is used to process the yarn contours, transforming irregular yarn edges into geometric polygon descriptions, and extracting key statistical features for each yarn. These key statistical features include the centroid coordinates of the yarn contour and the corresponding approximate polygon size parameters.
[0025] Specifically, step 1) is as follows: 1-1) X-ray computed tomography (XCT) was used to perform tomographic scanning on the four-step three-dimensional braided composite material sample. The scanning parameters were set to ensure that the grayscale differences between the yarn (basalt fiber) and the matrix (vinyl ester resin) could be clearly distinguished. The projected data obtained from the scan was reconstructed into three-dimensional volume data, and a series of cross-sectional slice images were taken along the braiding axis (Z-axis) using the image annotation software Avizo. To ensure the representativeness of statistical characteristics and reduce computational redundancy, one slice was selected every 10 images as a sample of the original dataset, as shown in the instruction manual. Figure 2 As shown in (a).
[0026] 1-2), as per the instruction manual Figure 2 As shown in (b), Avizo was used to manually annotate the yarn contours in the XCT image of the 3D braided composite material. Points were plotted along the interface between the yarn and the matrix to form closed regions, while ensuring that the contour interface regions of individual yarns did not stick together. This provides a basis for subsequent extraction of independent yarn instances for quantitative statistics. After annotation, the corresponding grayscale mask image was generated.
[0027] 1-3) Read the above-mentioned annotation file and use a computer vision algorithm (the findContours function in OpenCV) to extract the edge contour point set of each yarn instance. Since the original contour point set contains a large number of pixel jagged edges, which is not conducive to statistical analysis, polygon fitting and dimensionality reduction processing is required. According to the shape of the yarn contour, it can be divided into three types: elliptical, hexagonal, and fan-shaped.
[0028] 1-4), as per the instruction manual Figure 2 As shown in (c), for each fitted polygonal yarn instance, two types of key statistical features are calculated and extracted: 1. Centroid coordinate data, which are collected using the findContours function to collect the centroid coordinates (Cx) of each yarn contour. n Cy n 2. Approximate polygon size parameters: For each yarn profile, perform the aforementioned three shape fittings, and calculate the IOU value between it and the original yarn profile. Take the highest value as the fitting shape of the current yarn profile, and extract the corresponding shape size parameters.
[0029] 2) Based on the key statistical features extracted in step 1), a variational autoencoder (VAE) model is constructed and trained. The centroid coordinates of the yarn profile are used as input, enabling the VAE model to learn and capture the probability density and topological arrangement of the yarn centroid in three-dimensional space. Simultaneously, a multi-peak Gaussian fitting algorithm is used to statistically analyze the approximate polygon size parameters extracted in step 1), fitting a probability distribution model of the yarn cross-sectional dimensions (as shown in the appendix to the instruction manual). Figure 3 and attached Figure 4 As shown in the figure, this characterizes the morphological changes of the yarn at different positions.
[0030] This step is specifically as follows: 2-1), the set of yarn centroid coordinates of all slice images extracted in step 1) is defined as the input data X = {(Cx1, Cy1), (Cx2, Cy2), …, (Cx n Cy n )}, where each represents a set of two-dimensional coordinates of all yarn centroids on a slice. To adapt to network input, the number of centroids on each slice is unified (the centroids of the 48 yarn contours in each slice are unified and encoded in sequence, along with yarn contour shape type information), and the coordinates are normalized to the interval [−1,1].
[0031] 2-2) After data preprocessing, the data is fed into the VAE model for training. The encoder uses a multilayer perceptron (MLP) or a one-dimensional convolutional neural network (1D-CNN) as the feature extractor. The decoder structure is symmetrical with the encoder. The input is the latent vector z, which is mapped through multiple fully connected layers to reconstruct the set of random centroid coordinates of the yarn predicted in the output. And the yarn profile shape type, which specifically includes the following sub-steps: 2-2-1) Read the sliced data and perform one-hot encoding on the labels representing yarn types. Map the original categories to three binary feature columns: ellipse, hexagon, and sector. If a type is missing in a sample, the corresponding column is automatically padded with zeros to ensure feature alignment.
[0032] 2-2-2) Using image filenames as indexes and yarn IDs as columns, the data is reconstructed using perspective. The system verifies whether each image contains all 48 yarn points. If an image has missing data, it is considered an invalid sample and discarded directly, retaining only valid image data with all 48 detected points.
[0033] 2-2-3), for the center coordinates (Cx) of each yarn point i Cy i Normalization is performed to eliminate the influence of image resolution on the model. After normalization, the data is encoded in the format {(Cx i Cy i ), type_Ellipse, type_Hexagon, type_Sector}, where one of the three type_ data is 1 and the rest are 0, thus indicating the fitted shape type of the current coordinate point.
[0034] 2-3) Because the yarn cross-sectional shapes and sizes in the four-step three-dimensional braiding material vary, they exhibit obvious multimodal distribution characteristics, which cannot be accurately described by a single Gaussian distribution. Therefore, this invention uses a multi-peak Gaussian fitting algorithm to fit and model the elliptical, hexagonal, and sector-shaped parameter distributions extracted in steps 1-3) and obtain the fitting function, thereby fitting the probability distribution model of the yarn cross-sectional size.
[0035] 3) Please refer to the instruction manual appendix. Figure 5 As shown, the decoder part of the trained VAE model is used to sample and generate predicted centroid coordinates on a blank background image, which conform to the true distribution pattern. Subsequently, using the generated centroid coordinates as positioning points, the corresponding polygonal geometric contours are generated according to the size distribution constraints determined by the multi-peak Gaussian fitting algorithm. All generated polygonal contours are mapped onto the blank image to construct instance segmentation pseudo-labels with clear category and location information.
[0036] This step is specifically as follows: 3-1), as per the instruction manual Figure 5 As shown in (a), the VAE model trained in step 2) is used to perform forward inference on the blank slice to generate the centroid coordinates (Cx') of 48 random yarn profiles. n Cy' n Since VAE learns the spatial distribution pattern of yarn centroids, these generated points exhibit an array arrangement characteristic that conforms to the four-step weaving method on a macroscopic level.
[0037] 3-2), as per the instruction manual Figure 5 As shown in (b), for each generated centroid coordinate (Cx') n Cy' n This gives it a specific geometric shape. The generated centroid set... Each centroid coordinate in the image is taken as the origin, and the corresponding contour shape is generated by combining the size constraints of the fitting function described in steps 2-3). For easy visualization, the contour is then filled with corresponding grayscale pixels using the yarn ID number + 80. The final generated image is called a pseudo-label.
[0038] 4) Please refer to the instruction manual appendix. Figure 6 As shown, the instance segmentation pseudo-labels generated in step 3) are used as conditional inputs and imported into a pre-trained style transfer network based on the Brownian Bridge Diffusion Model (BBDM). This network learns the noise distribution, grayscale attenuation characteristics, and texture details of the original XCT image, transforming the geometric pseudo-label mapping into a pseudo-XCT image. This completes the cross-domain generation from pseudo-labels to pseudo-XCT images, achieving XCT image data augmentation.
[0039] Specifically, to address the issue of pseudo-labels generated purely by geometry lacking realistic XCT textures (such as the microscopic texture of fibers, noise in the matrix, artifacts, and grayscale gradations), this work introduces a Brownian bridge diffusion model as an image style converter to achieve a high-fidelity mapping from the "semantic domain" to the "real image domain." During pre-training of the Brownian bridge diffusion model, a true training set (i.e., true labels and true original images) is input, and the model outputs a training weights .pth file after learning.
[0040] After the model training is completed, its training weights.pth file is obtained. The pseudo-labels described in step 3-2) are then input into the trained Brownian bridge diffusion model. The model converts the pseudo-label style into a pseudo-XCT image similar to the original image style by calling the training weights.pth file, as shown in the appendix to the instruction manual. Figure 5 As shown in (c).
[0041] In summary, this invention uses pseudo-training sets composed of pseudo-labels and pseudo-original images, and combines them with the real training set composed of the aforementioned real labels and real original images to form an enhanced training set. In other words, adding the pseudo-training set to the real training set creates an enhanced dataset. When this dataset is fed into a segmentation network, the network's segmentation capabilities are enhanced during training. This method is used for training instance segmentation networks for XCT images of four-step 3D woven composite materials, solving the problems of difficult and costly dataset creation.
[0042] The above-described specific embodiments are merely preferred embodiments of this invention and are not intended to limit this invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the protection scope of this invention.
Claims
1. A method for enhancing XCT image data of composite materials based on statistical features, characterized in that: Includes the following steps: 1) Obtain the original XCT image of the four-step three-dimensional braided composite material, and manually annotate it to obtain the accurate yarn contour mask; then, use a polygon fitting algorithm to process the yarn contour, transform the irregular yarn edges into geometric polygon descriptions, and extract the key statistical features of each yarn; the key statistical features include the centroid coordinate data of the yarn contour and the corresponding approximate polygon size parameters. 2) Based on the key statistical features extracted in step 1), a VAE model is constructed and trained. The centroid coordinates of the yarn profile are used as input, so that the VAE model can learn and capture the probability density and topological arrangement of the yarn centroid in three-dimensional space. At the same time, a multi-peak Gaussian fitting algorithm is used to perform statistical analysis on the approximate polygon size parameters extracted in step 1) to fit the probability distribution model of the yarn cross-sectional size, so as to characterize the morphological change characteristics of the yarn at different positions. 3) Using the decoder part of the trained VAE model, sample and generate on the blank background image, and output the predicted centroid coordinates that conform to the real distribution law; then, using the generated centroid coordinates as the positioning point, generate the corresponding polygonal geometric contours according to the size distribution constraints determined by the multi-peak Gaussian fitting algorithm; map all the generated polygonal contours onto the blank image to construct instance segmentation pseudo-labels with clear category and location information. 4) The instance segmentation pseudo-labels generated in step 3) are used as conditional inputs and imported into a pre-trained style transfer network based on the Brownian bridge diffusion model. The network is used to learn the noise distribution, gray-level attenuation characteristics and texture details of the original XCT image, and the geometric pseudo-labels are mapped to pseudo-XCT images. This completes the cross-domain generation from pseudo-labels to pseudo-XCT images and realizes XCT image data augmentation.
2. The composite material XCT image data enhancement method based on statistical features as described in claim 1, characterized in that: Step 1) specifically refers to: 1-1) The four-step three-dimensional braided composite material sample was subjected to tomographic scanning using an XCT device; the projection data obtained from the scan was reconstructed into three-dimensional volume data, and a series of cross-sectional slice images were extracted along the braiding axis using the image annotation software Avizo. 1-2) Use Avizo to manually annotate the yarn outlines in the XCT image of the 3D braided composite material. Plot points along the interface area between the yarn and the matrix to form closed regions, while ensuring that the outline interface areas of individual yarns do not stick together. After the annotation is completed, the corresponding grayscale mask image is generated. 1-3) Read the above annotation file and use computer vision algorithms to extract the edge contour point set of each yarn instance; then perform polygon fitting and dimensionality reduction processing. According to the shape of the yarn contour, it can be divided into elliptical, hexagonal and sector-shaped. 1-4), for each fitted polygonal yarn instance, calculate and extract two types of key statistical features: centroid coordinate data, which are collected using the findContours function to collect the centroid coordinates (Cx) of each yarn profile. n Cy n Approximate polygon size parameters; perform the aforementioned three-shape fitting on each yarn profile, calculate the IOU value between it and the original yarn profile, take the highest value as the fitting shape of the current yarn profile, and extract the corresponding shape size parameters.
3. The composite material XCT image data enhancement method based on statistical features as described in claim 2, characterized in that: In step 1-1), one slice is selected every 10 slices as a sample of the original dataset.
4. The composite material XCT image data enhancement method based on statistical features as described in claim 2, characterized in that: Step 2) specifically refers to: 2-1), the set of yarn centroid coordinates of all slice images extracted in step 1) is defined as the input data X = {(Cx1, Cy1), (Cx2, Cy2), …, (Cx n Cy n )}, where each represents a set of two-dimensional coordinates of all yarn centroids on a slice; the number of centroids on each slice is unified, and the coordinates are normalized to the interval [−1,1]; 2-2) After data preprocessing, the data is fed into the VAE model for training. The encoder uses a multilayer perceptron or a one-dimensional convolutional neural network as the feature extractor. The decoder structure is symmetrical with the encoder. The input is the latent vector z, which is mapped through multiple fully connected layers to reconstruct the set of random centroid coordinates of the yarn predicted in the output. and the type of yarn outline shape; 2-3) The multi-peak Gaussian fitting algorithm is used to fit and model the elliptical, hexagonal, and sector-shaped parameter distributions extracted in step 1-3) and obtain the fitting function, thereby fitting the probability distribution model of the yarn cross-sectional size.
5. The composite material XCT image data enhancement method based on statistical features as described in claim 4, characterized in that: In step 2-1), the centroids of the 48 yarn outlines in each slice are unified and encoded sequentially, along with information on the shape type of the yarn outlines.
6. The composite material XCT image data enhancement method based on statistical features as described in claim 4, characterized in that: Step 2-2) specifically refers to: 2-2-1) Read the sliced data and perform one-hot encoding on the labels representing yarn types; map the original categories to three binary feature columns: ellipse, hexagon, and sector. If a certain type is missing in the sample, the corresponding column is automatically padded with zeros. 2-2-2), using image file names as indexes and yarn numbers as columns, perform perspective reconstruction on the data; the system verifies whether each image contains all 48 yarn points. If an image has missing data, it is considered an invalid sample and is directly removed, retaining only valid image data with all 48 points detected. 2-2-3), for the center coordinates (Cx) of each yarn point i Cy i Normalization was performed. After normalization, the data is encoded, specifically in the format {(Cx i Cy i ), type_Ellipse, type_Hexagon, type_Sector}, where one of the three type_ data is 1 and the rest are 0, thus indicating the fitted shape type of the current coordinate point.
7. The composite material XCT image data enhancement method based on statistical features as described in claim 4, characterized in that: Step 3) specifically refers to: 3-1) Using the VAE model trained in step 2), perform forward inference on the blank slice to generate the centroid coordinates (Cx') of 48 random yarn profiles. n Cy' n ); 3-2), for each generated centroid coordinate (Cx' n Cy' n ), giving it a specific geometric shape, to generate a set of centroids. Each centroid coordinate in the image is taken as the origin, and the corresponding contour shape is generated by combining the size constraints of the fitting function described in steps 2-3). Then, the corresponding gray value pixels are filled inside the contour with the gray value of yarn ID number + 80. The final generated image is called a pseudo label.
8. The composite material XCT image data enhancement method based on statistical features as described in claim 4, characterized in that: In step 4), when pre-training the Brownian bridge diffusion model, the input is the true training set, i.e., the true labels and the true original image, and the output is the training weights .pth file after learning.
9. The composite material XCT image data enhancement method based on statistical features as described in claim 8, characterized in that: In step 4), the specific method for generating pseudo-XCT images is as follows: input the pseudo-labels from step 3-2) into the trained Brownian bridge diffusion model, and the model converts the pseudo-label style into pseudo-XCT images by calling the training weights.pth file.