Image segmentation method and device for fuel particle layered micro-nano structure

Abstract

The invention discloses an image segmentation method and device for a fuel particle layered micro-nano structure. The method comprises the steps of data preprocessing, wherein 3D CT data are converted into 2D Cartesian slices, polar coordinate images are obtained, and strips are divided by means of particle radial symmetry; feature extraction: constructing a main network by U-Net, and adding a position coding module (strengthening angle position modeling) and a disturbance module (improving generalization ability); the multi-class segmentation is divided into boundary dichotomy subtasks, and boundary perception weighting loss is combined with weighted binary cross entropy training. And feature fusion: defining a structure priori matrix coding coating sequence, and training the weight by classifying cross entropy through a conditional random field fusion result (ensuring topological consistency between layers). And finally, splicing strip blocks, and converting polar coordinates into a Cartesian image. The method is applied to automatic nuclear fuel detection and radiation imaging, multi-layer accurate segmentation is achieved, and the detection efficiency and precision are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an image segmentation method and apparatus for layered micro / nano structures of fuel particles. Background Technology

[0002] In the development of nuclear technology, coated fuel particles are a core component of in-service high-temperature gas-cooled reactors (HTGRs) and a key component of accident-tolerant, all-ceramic microcapsule (FCM) fuel designed for future water-cooled reactors. Each particle typically has a diameter of 0.92 mm and a multi-layered structure, including a spherical fuel core (uranium dioxide or uranium carbide) and a series of coatings. These coatings, from the inside out, consist of a loose pyrolytic carbon buffer layer, an inner dense pyrolytic carbon layer (IPyC), a silicon carbide layer (SiC), and an outer dense pyrolytic carbon layer (OPyC). To ensure the integrity of the particles and their safety in the reactor, the geometry, thickness of each layer, interface structure, defects, voids, and surface quality of the coated fuel particles are typically accurately inspected. Common inspection techniques include optical microscopy, scanning electron microscopy (SEM), transmission electron microscopy (TEM), and X-ray computed tomography (CT). Among these, CT has become an important means of characterizing structural features such as coating thickness, defects, and voids due to its non-destructive, high-precision, and high-efficiency advantages. This technique relies on the different attenuation coefficients of the X-rays being imaged. During the imaging process, the buffer layer and IPyC layer in the fuel-coated particles are made of the same material, which results in extremely low contrast, blurred boundaries, and unclear interfaces in the CT images, thus seriously affecting the distinguishability of the coating structure and the accuracy of subsequent precise measurements.

[0003] Traditional image segmentation methods, such as thresholding, region growing, edge detection, and active contour models, have achieved significant results in many fields. However, they still face serious challenges when applied to the segmentation of CT images coated with fuel particles. Because the buffer layer and IPyC layer are made of the same material, their grayscale values ​​in CT images are highly similar, resulting in low boundary contrast, making it difficult for traditional methods to accurately identify their interlayer interfaces.

[0004] With the rapid development of deep learning, image segmentation methods have made significant progress. Classic architectures include fully convolutional networks (FCN), U-Net, SegNet, and the DeepLab series. Furthermore, multi-scale fusion networks such as Feature Pyramid Network (FPN), HRNet, and UNet++ have been proposed. The combination of attention mechanisms and Transformer-based structures, such as Attention UNet, CBAM, Swin Transformer, and Mask2Former, has further enhanced the model's representational capabilities. However, the acquisition process for fuel particle CT images is costly, and high-quality, accurately labeled data resources are scarce, severely limiting the training performance of deep learning models. Moreover, training data often comes from different batches of samples, scanning systems, or imaging parameters, which can lead to severe domain shifts, further limiting the model's generalization ability. Summary of the Invention

[0005] The purpose of this invention is to provide an image segmentation method and apparatus for layered micro / nano structures of fuel particles, in order to solve one or more of the problems of low contrast, limited samples, and domain offset in CT image segmentation of fuel particles.

[0006] To achieve the above objectives, the present invention provides an image segmentation method for layered micro / nano structures coated with fuel particles, comprising: Data preprocessing steps: Convert the raw 3D CT data of the fuel particles into 2D Cartesian coordinate slices, convert the 2D Cartesian coordinate slices into polar coordinate images, and perform striping processing on the polar coordinate images based on the radial symmetry of the particles. Feature extraction steps: Feature extraction is performed on the image data after striping using a deep learning framework; Boundary binary classification subtask segmentation steps: Decompose the multi-layer structure segmentation problem of fuel particles into multiple boundary binary classification subtasks, and obtain the segmentation results of each boundary binary classification subtask based on the feature information extracted in the feature extraction step. Feature fusion steps: Define a structural prior matrix to encode the order relationship between layers, and use a conditional random field model to fuse the segmentation results of multiple boundary binary classification sub-tasks in combination with the structural prior matrix to ensure topological consistency between layers. The reconstruction steps are as follows: the strip segmentation results after feature fusion are stitched together, and the stitched polar coordinate image is converted back to a Cartesian coordinate system image to obtain the final segmentation result.

[0007] Furthermore, in the feature extraction step, the deep learning framework is built on the U-Net architecture, and a positional encoding module is added before each convolutional block of the encoder and decoder in the U-Net architecture; The positional encoding module is used to explicitly encode radial positional relationships in polar coordinate images. It is implemented through a learnable function R(·) related to relative distance, which is fitted by a convolutional layer. The output of the positional encoding module is concatenated with the original input feature map as an additional channel.

[0008] Furthermore, the feature extraction step also includes a perturbation module; The perturbation module applies multiple affine transformations to the input image, inputs the transformed images into the deep learning framework for feature extraction, performs corresponding inverse transformations on the extraction results, and then obtains the final feature extraction result through weighted fusion.

[0009] Furthermore, during the model training phase, the boundary-aware weighted loss function is used for supervised training of the boundary binary classification subtask segmentation step. The boundary-aware weighted loss function is a weighted binary cross-entropy loss. Its weight factor is dynamically calculated based on the normalized radial distance from the pixel to a specific boundary, so that pixels closer to the boundary have a higher weight in the loss function.

[0010] Furthermore, during the model training phase, the weight matrix of the conditional random field model in the feature fusion step is trained under supervision using the classification cross-entropy loss function.

[0011] Furthermore, the structural prior matrix It is a T-row, k-column matrix, where T is the total number of categories of the coating region and k is the total number of boundary binary classification subtasks; matrix Each element in This represents the spatial relationship between the coded category t and the boundary k, and takes the value 1 or -1. 1 indicates that category t is inside the boundary k, and -1 indicates that category t is outside the boundary k.

[0012] Furthermore, the Conditional Random Field (CRF) model maps the segmentation probabilities of multiple boundary binary classification subtasks into vectors, multiplies them with the structural prior matrix, and then applies a normalization function. Normalization is used to achieve the final multi-class probability prediction: (11) in, This represents the probability that a pixel in the image after feature fusion belongs to each category. Through The generated data represents multiple binary tags. This represents multiple binary segmentation networks. Let represent the diagonal weight matrix, used to adjust the contribution of each boundary, and described by equation (10): (10) In the formula, The weights represent the segmentation results of the k-th boundary segmentation task. .

[0013] The present invention also provides an image segmentation apparatus for coating fuel particle layered micro / nano structures, comprising: The data preprocessing unit is used to convert the raw 3D CT data of the fuel particles into 2D Cartesian coordinate slices, convert the 2D Cartesian coordinate slices into polar coordinate images, and perform striping processing on the polar coordinate images based on the radial symmetry of the particles. The feature extraction unit is used to extract features from the image data after striping and blocking using a deep learning framework; The boundary binary classification subtask segmentation unit is used to decompose the multi-layer structure segmentation problem of coated fuel particles into multiple boundary binary classification subtasks, and obtain the segmentation result of each boundary binary classification subtask based on the feature information extracted by the feature extraction unit. The feature fusion unit is used to define the structural prior matrix to encode the order relationship between layers, and uses the conditional random field model to fuse the segmentation results of multiple boundary binary classification sub-tasks in combination with the structural prior matrix to ensure the topological consistency between layers. The result reconstruction unit is used to stitch together the strip segmentation results after feature fusion and convert the stitched polar coordinate image back to a Cartesian coordinate system image to obtain the final segmentation result.

[0014] Furthermore, in the feature extraction unit, the deep learning framework is built on the U-Net architecture, and a positional encoding module is added before each convolutional block of the encoder and decoder in the U-Net architecture; The positional encoding module is used to explicitly encode radial positional relationships in polar coordinate images. It is implemented through a learnable function R(·) related to relative distance, which is fitted by a convolutional layer. The output of the positional encoding module is concatenated with the original input feature map as an additional channel. The feature extraction unit also includes a perturbation module; The perturbation module applies multiple affine transformations to the input image, inputs the transformed images into the deep learning framework for feature extraction, performs corresponding inverse transformations on the extraction results, and then obtains the final feature extraction result through weighted fusion.

[0015] Furthermore, the structural prior matrix It is a T-row A matrix of columns, where, It is the total number of categories of the coated areas. It is the total number of boundary binary classification subtasks; matrix Each element in Represents the coding category With boundary The spatial relationship between them takes a value of 1 or -1, where 1 represents the category. Located at the border Within, -1 indicates the category Located at the border In addition; Conditional random field (CRF) models map the segmentation probabilities of multiple boundary binary classification subtasks into vectors, multiply them by the structural prior matrix, and then pass them through a normalization function. Normalization is used to achieve the final multi-class probability prediction: (11) in, This represents the probability that a pixel in the image after feature fusion belongs to each category. Through The generated data represents multiple binary tags. This represents multiple binary segmentation networks. Let represent the diagonal weight matrix, used to adjust the contribution of each boundary, and described by equation (10): (10) In the formula, Indicates the first The weights of the segmentation results of each boundary segmentation task. .

[0016] By adopting the above technical solutions, this invention can apply a physically interpretable and practical artificial intelligence method to automated nuclear fuel detection and radiation imaging analysis, thereby achieving precise segmentation of multi-layer structures and effectively improving detection efficiency and accuracy. Attached Figure Description

[0017] Figure 1 A flowchart of a multi-stage segmentation method for coated fuel particles based on a deep learning framework and structural priors provided in an embodiment of the present invention; Figure 2a A slice of a reconstructed image of coated fuel particles obtained using the Zeiss Xradia microXCT 400 system, provided for an embodiment of the invention; Figure 2b Labeled images of slices of reconstructed images provided for embodiments of the invention; Figure 3a Cartesian coordinate images and polar coordinate images provided for embodiments of the invention; Figure 3bA schematic diagram illustrating the striping and segmentation of a polar coordinate system image to demonstrate the radial symmetry of the polar coordinate image in an embodiment of the invention. Figure 4 A schematic diagram of a neural network structure provided for an embodiment of the invention; Figure 5 A schematic diagram illustrating the decomposition of multi-class segmentation problem labels into multiple boundary binary classification sub-task labels provided in the embodiments of the invention; Figure 6 A schematic diagram of the boundary-aware weighted loss calculation method provided in the embodiments of the invention; Figure 7 A schematic diagram comparing the segmentation results provided in the embodiments of the invention with those of conventional methods. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only possible technical implementations of the present invention and not all possible implementations. Those skilled in the art can obtain other embodiments in conjunction with the embodiments of the present invention without creative effort, and these embodiments are also within the protection scope of the present invention.

[0019] like Figure 1 As shown, the multi-stage segmentation method based on deep learning framework and structural prior provided in this embodiment of the invention includes: Data preprocessing steps: The raw 3D CT data is converted into 2D Cartesian coordinate slices, with the center of the slice image serving as the origin of the Cartesian coordinate system. The 2D Cartesian coordinate slices are then converted into polar coordinate images. The angular resolution is preset; in this embodiment, it is set to 0.25°. The radius resolution is consistent with the Cartesian coordinate system resolution. Then, the polar coordinate image is segmented into strips using the radial symmetry of the particles. In this embodiment, the strips are divided into 15° segments, but this is not a limitation.

[0020] Feature extraction steps: Feature extraction is performed on the striped block data using a deep learning framework.

[0021] The segmentation steps for the boundary binary classification subtask are as follows: The multi-class segmentation problem is decomposed into multiple boundary binary classification subtasks, and the segmentation results of each boundary binary classification subtask are obtained based on the feature information extracted in the feature extraction step.

[0022] Feature fusion steps: Define a structural prior matrix to encode the order relationship between layers, and use a conditional random field model to fuse the segmentation results of multiple boundary binary classification sub-tasks in combination with the structural prior matrix to ensure topological consistency between layers.

[0023] The reconstruction steps are as follows: the strip segmentation results after feature fusion are stitched together, and the stitched polar coordinate image is converted back to a Cartesian coordinate system image to obtain the final segmentation result.

[0024] In one embodiment, in the feature extraction step, such as Figure 4 As shown, the deep learning framework uses a U-Net-based main network for feature extraction, comprising an encoder and a decoder. The encoder contains five convolutional blocks, each containing two 3×3 convolutions with channel numbers of 64, 128, 256, 512, and 1024 respectively. The decoder also contains five convolutional blocks, each containing two 3×3 convolutions with channel numbers of 1024, 512, 256, 128, and 64 respectively.

[0025] A positional encoding module is added before each convolutional block in the encoder and decoder, corresponding to... Figure 4 The AAPE. The position encoding module is used to explicitly encode radial positional relationships in polar coordinate images, which is achieved through a learnable function related to relative distance. This function is implemented by fitting through a convolutional layer.

[0026] The output of the position encoding module is a separate channel, i.e., an additional channel connected to each convolutional block. The corresponding number of encoders becomes 65, 129, 257, 513, and 1025, and the number of decoders becomes 1025, 513, 257, 129, and 65.

[0027] Specifically, the radius positions in the polar coordinate image are explicitly encoded, enhancing structure awareness and relative positional relationships with minimal computational overhead. The definition is as follows: (1) in, The function representing the relative distance is fitted using two 9×1 convolutions, where the radial direction has 9 values ​​and the angular direction has 1 value. The convolution kernels are initialized with random values, and each kernel is activated using the ReLU function. Compared to the cosine position encoding used in Transformers, cosine position encoding is an absolute position encoding method. This encoding method reflects the relative relationship between positions, making it a relative position encoding. In classification problems, classification is related to relative position, not absolute position. Compared to the relative position encoding and rotational position encoding used in Transformers, this encoding method has lower computational cost and higher computational efficiency. The specific values ​​of the convolution kernels are obtained through training. When updating the parameters of the deep learning framework after the boundary binary classification subtask segmentation step, the parameter values ​​of the two 9×1 convolutions are updated simultaneously based on the loss function results of the boundary binary classification subtask segmentation step. Point and points The relative positions between them, the midpoint and points Located on the same straight line, cut off more than Long-distance relationships, This indicates the absolute value operation. The positional encoding is as follows: (2) in, It is a multi-channel image. express Location in multi-channel images Multi-channel image values ​​at the location, Indicates the first The first channel and the first Pixel value at 1 pixel Indicates the channel used for processing Data Functions for different channels , The fit is performed using two independent 9×1 convolutions. The equation is implemented through convolution operations. The output after adding positional encoding is as follows: (3) in, Indicates will The result is connected as a separate channel to the input multi-channel image. , Indicates multi-channel image All locations What is the result of the operation? For example... Figure 4 As shown, the final result will be Replace the original It serves as the input to the convolutional blocks of the main network encoder and decoder.

[0028] In one embodiment, in the feature extraction step, a perturbation module is added to the input of the main network. The perturbation module applies multiple affine transformations to the input image, and the transformed images are input into the deep learning framework for feature extraction. The extraction results are then subjected to corresponding inverse transformations and weighted fusion to obtain the final feature extraction result, thereby improving the generalization ability of the model.

[0029] Specifically, the calculation is performed using equation (4): (4) in, Indicates multi-channel image Increase data perturbation, Indicates the first There are affine transformations, totaling [number] . Transformation, Indicates multi-channel image Conduct the first An affine transformation, Indicates that the signal passes through the main network. The returned data is then processed. The inverse transform of an affine transformation, Indicates the first The fusion weights of an affine transformation. For example... Figure 4 As shown, the main network of this invention Using U-Net, The result is used as the feature extraction result. In a specific embodiment... The affine transformation is 3, and scaling is performed in the radial direction, specifically by scaling ratios of 0.8, 1.0, and 1.25, corresponding to the weights. The values ​​are 0.2, 0.6, and 0.2 respectively.

[0030] In one embodiment, the model includes two phases: a training phase and a prediction phase.

[0031] The same data preprocessing steps are performed in both the training and prediction phases, leading to the result reconstruction steps.

[0032] During the training phase, boundary-aware weighted loss, combined with weighted binary cross-entropy loss, is used to update the parameters of the main feature extraction network built based on U-Net in step 2. During the training phase, the weights in step 4 are updated using classification cross-entropy loss. The training phase can be terminated after the parameter updates; step 5 is not required.

[0033] In the prediction phase, once the results are obtained, there is no need to update the parameters of the feature extraction main network built on U-Net or update the weights of the feature fusion step.

[0034] The following specific embodiment illustrates the implementation process of a multi-stage segmentation method based on a deep learning framework and structural prior for coating fuel particles proposed in this invention.

[0035] This example uses reconstructed images of fuel-coated particles obtained from the Zeiss Xradia microXCT 400 system. This invention can be used not only for image segmentation of fuel-coated particles but also for segmenting images of other layered structures, such as human skeletons and carbon crystals. Scanning parameters were set to a tube voltage of 100 kV and a power of 10 W. Two datasets, dataset A and dataset B, were collected from different samples in the same batch. These datasets include reconstructed slices, as shown in Figure 2(a). Multiple sets of fuel-coated particle images were acquired using a CT scanner. The two sets of acquired fuel-coated particle data, dataset A, were used for the training phase, and dataset B was used for the prediction phase. As shown in Figure 2(b), the training phase data was labeled.

[0036] The execution hardware in this embodiment is a workstation equipped with an Intel Core i7-10700K@3.8 GHz CPU and an NVIDIA GeForce RTX 3090 24 GB GPU. This implementation is based on the PyTorch deep learning framework, which runs in Python 3.8 and features CUDA acceleration to ensure computational efficiency.

[0037] The first phase is the training stage: The model was optimized using the Adam optimizer via stochastic gradient descent. The learning rate was set to 0.001, and the first and second moment estimation coefficients were set to beta1=0.9 and beta2=0.999, respectively. The batch size was set to 8 and randomly shuffled. To prevent gradient explosion, gradient clipping was used, and the maximum threshold for the L2 norm was set to 5. The number of iterations was set to 100.

[0038] Data preprocessing step a, as shown in Figure 3, converts the image into a polar coordinate image, and then uses the radial symmetry of the particles to perform striping processing on the polar coordinate image.

[0039] Feature extraction step a: Feature extraction is performed on the striped block data using a deep learning framework.

[0040] Step a of the boundary binary classification subtask segmentation process involves decomposing the multi-class segmentation problem into multiple boundary binary classification subtasks. Then, the execution results of these subtasks are obtained using extracted feature information. Figure 4 As shown. After obtaining the results, the loss between the results and the labels is calculated using boundary-aware weighted loss combined with weighted binary cross-entropy loss, and the model parameters are updated. The weighted loss weights are as follows: Figure 5 As shown.

[0041] In feature fusion step a, a structural prior matrix is ​​defined to encode the coating order relationship, and a conditional random field is used to fuse the results. After obtaining the results, the loss between the results and labels is calculated using classification cross-entropy loss, and the weights are updated. The training phase ends here.

[0042] Then the testing phase begins: In data preprocessing step b, the image is converted into a polar coordinate image, and then the polar coordinate image is segmented into strips using the radial symmetry of the particles.

[0043] Feature extraction step b involves extracting features from the striped block data using a deep learning framework.

[0044] Step b of the boundary binary classification subtask segmentation decomposes the multi-class segmentation problem into multiple boundary binary classification subtasks, and then obtains the execution results of the boundary binary classification subtasks through the extracted feature information, such as... Figure 4 As shown.

[0045] Feature fusion step b: In the feature fusion stage, a structural prior matrix is ​​defined to encode the coating order relationship, and a conditional random field is used to fuse the results.

[0046] Result reconstruction step b: The segmented strips are stitched together, and the stitched polar coordinate image is converted into a Cartesian coordinate image. The output result is as follows. Figure 6 As shown.

[0047] from Figure 6 The figure shows the segmentation results of this invention and direct segmentation using UNet. The error plot in the figure visually illustrates the detailed differences in layer recognition between the methods. Compared with the UNet method, the method proposed in this study performs best in terms of recognition accuracy and structural consistency, and can clearly and completely separate the multi-layer encapsulation structure of TRISO fuel particles, exhibiting good boundary continuity and layer coherence. This finding verifies the effectiveness of modeling structural priors and spatial constraints under conditions of few shots.

[0048] In one embodiment, the boundary binary classification subtask segmentation step decomposes the multi-level segmentation task into multiple binary subtasks using the boundary information of the fuel particle coating. In this invention, the image multi-classification problem is decomposed into multiple boundary segmentation tasks, each of which is a binary classification task. The feature space of the binary classification task is simpler, allowing the model to focus more on distinguishing the boundary features of the image, avoiding feature confusion caused by complex relationships between categories in multi-class classification. Furthermore, since there are multiple boundaries covering the fuel particle, the importance of boundary segmentation varies. The accuracy of segmenting key boundaries can be improved by adjusting different boundary weights. In the same boundary segmentation task, images far from the boundary are easier to segment, while images near the boundary are more difficult to segment. Therefore, increasing the loss weight for images near the boundary can improve the segmentation result. The loss function is used to evaluate the matching degree of the results; a smaller loss function result indicates a better result. The loss function result is also used to calculate the gradient values ​​of the neural network parameters. Binary cross-entropy loss (BCE) is typically used for loss calculation in binary classification tasks. The definition of binary cross-entropy loss is as follows: (5) in, This represents the result of the binary cross-entropy loss calculation. It is the first pixel value, Indicates about The predicted probability, This indicates the total number of pixels in the output classified image.

[0049] In the binary classification task of this invention, a weight related to the boundary distance is added to the binary cross-entropy loss. The calculation method is as shown in formula (7). The weighted binary cross-entropy loss is called the weighted binary cross-entropy loss, which is used to measure the quality of the segmentation result and calculate the gradient of network parameter update. The weighted binary cross-entropy loss is defined as follows: (6) in, Indicates the first The results of calculating the boundary-weighted binary cross-entropy loss are as follows: Indicates the weights related to boundary distance. It is the first The pixel and the The value of each boundary, Indicates about The predicted probability, This indicates the total number of pixels in the output classified image.

[0050] The boundary-aware weighted calculation method is as follows: (7) in, Indicates from pixel To the The normalized radial distance of each boundary is calculated as the ratio between their radial pixel distance and the radial width of the image. Indicates the boundary emphasis factor. Controlling the spread, The value is 1.0. The value is 0.2.

[0051] In one embodiment, the feature fusion step uses a Conditional Random Field (CRF) to embed physical priors, including layer order and continuity, to enforce the correct topological transformation. The definition of a Conditional Random Field (CRF) is as follows: (8) in, Indicates the input image. Represents the label image The first in 1 pixel, Indicates the input image In the case of the output image, the first Each pixel is The probability, This represents the counting of potential energy functions, where the potential energy function has a total count. indivual, Represents the eigenvalue decomposition function. Indicates the first Potential energy function, Indicates the first The weights of each potential function. It is the normalization factor, defined as follows: (9) in, express The set of all possible scenarios. Specifically, applying conditional random fields to the segmentation task, the boundary segmentation task can be viewed as a potential function. This also represents the boundary number in the boundary segmentation task.

[0052] The specific steps for incorporating physical priors into conditional random fields are as follows: S41, the equations for a conditional random field with physical priors are defined as follows: (10) Among them, adding to the conditional random field equations function, Indicates the first The structural constraint functions of the potential energy functions are also known as physical prior constraints. In the application scenario of this invention, the potential energy functions represent a binary classification task, and the number of boundaries corresponds to the number of potential energy functions. It also represents the count of binary classification tasks. This indicates the process of feature extraction using a neural network. This indicates that the first binary classification task is one of the multiple binary classification tasks decomposed in the boundary binary classification subtask segmentation step. This is a binary classification task. The segmentation task involves finding the result with the highest probability among the various outcomes, as shown below: (11) in, Indicates the first The classification result with the highest probability for each pixel. This represents the input image, which is the striped image after the data preprocessing step. Represents the label image The 1 pixel, Obtained through manual annotation.

[0053] S42 maps the binary segmentation probability to [-1, 1]. A mapping probability greater than 0 indicates that the corresponding point is inside the boundary, while a mapping probability less than 0 indicates that the point is outside the boundary.

[0054] S43, the output of the previous step can be viewed as a vector, where each element represents the result of multiple binary segmentations. Given a known classification, the probability is explicitly 1 or 0, therefore the vector's value can only be 1 or -1. Define the vector with the known classification as the baseline vector. Classification can be performed by determining the angle between the unknown category and the baseline vector; the vector with the smallest angle belongs to the same category. This operation can be quickly calculated using vector multiplication. When vectors have the same magnitude, the dot product of the two vectors results in a smaller angle. Furthermore, vectors with values ​​of 1 or -1 have equal magnitudes. Multiple categories can form a matrix; matrix multiplication allows for the rapid simultaneous calculation of multiple angles. This matrix is ​​called the structural prior matrix. .

[0055] Define the structure prior matrix Substitute (10) Based on the order relationship between coding layers: (12) Each row represents a category, and the total number of categories is... Each column represents a boundary, and the total number of boundaries is . , Matrix represents categories and boundaries Spatial relationship between them It is equal to 1 or -1, where 1 indicates inside the corresponding boundary and -1 indicates outside the boundary. This matrix encodes the directionality of physical layer transitions.

[0056] In the image of coated fuel particles, from the inside to the outside, there are six categories: "core layer - buffer layer - IPyC layer - SiC layer - OPyC layer - outer edge of particle". A boundary is formed between adjacent categories, for a total of five boundaries. Figure 5 As shown. According to formula (12), the specific prior matrix form can be given, as shown in the following formula (13): (13) The matrix consists of 5 columns, each representing a boundary. From left to right, these are the boundaries: "Core Layer-Buffer Layer," "Buffer Layer-IPyC Layer," "IPyC Layer-SiC Layer," "SiC Layer-OPyC Layer," and "OPyC Layer-Particle Outer Side." The matrix also has 6 rows, representing the core layer, buffer layer, IPyC layer, SiC layer, OPyC layer, and particle outer side. Taking the third IPyC layer as an example, the IPyC layer is outside the boundaries of the "Core Layer-Buffer Layer" and "Buffer Layer-IPyC Layer," and inside the boundaries of the "IPyC Layer-SiC Layer," "SiC Layer-OPyC Layer," and "OPyC Layer-Particle Outer Side." The third row of the matrix is ​​represented as (-1 -1 1 1 1). Because convolutional neural networks (CNNs) have frequency shift invariance, they cannot effectively distinguish between similar pixels at different positions. The prior matrix combines the classification of the fuel particle image with the relative structural positions of the fuel particle image, and then... Figure 5 The method shown transforms the classification problem of fuel particle-coated images into multiple boundary segmentation problems, obtaining the relative positions of pixels with respect to the boundaries through boundary segmentation. This segmentation approach, which combines relative positions, effectively improves the ability of convolutional neural networks to recognize images at different locations.

[0057] S44, define a diagonal weight matrix. To adjust the contribution of each boundary, replacing the function of equation (10). : (14) In the formula, The weights represent the segmentation results of the k-th boundary segmentation task. The specific values ​​are obtained by training the weights using the classification cross-entropy loss. .

[0058] S45, the feature fusion result is shown below: (15) in, This represents the probability that a pixel in the image after feature fusion belongs to each category. Through The generated data represents multiple binary tags. Representing multiple binary segmentation networks, the function replacing equation (10) , This represents the normalization function. The position in the neural network is as follows: Figure 4 As shown.

[0059] This invention also provides an image segmentation device for layered micro / nano structures coated with fuel particles, comprising a data preprocessing unit, a feature extraction unit, a boundary binary classification sub-task segmentation unit, a feature fusion unit, and a result reconstruction unit, wherein: The data preprocessing unit is used to convert the raw 3D CT data of the fuel particles into 2D Cartesian coordinate slices, convert the 2D Cartesian coordinate slices into polar coordinate images, and perform striping processing on the polar coordinate images based on the radial symmetry of the particles.

[0060] The feature extraction unit is used to extract features from the image data after it has been segmented into strips using a deep learning framework.

[0061] The boundary binary classification subtask segmentation unit is used to decompose the multi-layer structure segmentation problem of coated fuel particles into multiple boundary binary classification subtasks, and obtain the segmentation result of each boundary binary classification subtask based on the feature information extracted by the feature extraction unit.

[0062] The feature fusion unit is used to define the structural prior matrix to encode the order relationship between layers, and uses a conditional random field model to fuse the segmentation results of multiple boundary binary classification subtasks in combination with the structural prior matrix to ensure topological consistency between layers.

[0063] The result reconstruction unit is used to stitch together the strip segmentation results after feature fusion, and convert the stitched polar coordinate image back to a Cartesian coordinate system image to obtain the final segmentation result.

[0064] In one embodiment, in the feature extraction unit, the deep learning framework is built based on the U-Net architecture, and a positional encoding module is added before each convolutional block of the encoder and decoder of the U-Net architecture; The positional encoding module is used to explicitly encode radial positional relationships in polar coordinate images. It is implemented through a learnable function R(·) related to relative distance, which is fitted by a convolutional layer. The output of the positional encoding module is concatenated with the original input feature map as an additional channel. The feature extraction unit also includes a perturbation module; The perturbation module applies multiple affine transformations to the input image, inputs the transformed images into the deep learning framework for feature extraction, performs corresponding inverse transformations on the extraction results, and then obtains the final feature extraction result through weighted fusion.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Those skilled in the art should understand that modifications can be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An image segmentation method for layered micro / nano structures coated with fuel particles, characterized in that, include: Data preprocessing steps: Convert the raw 3D CT data of the fuel particles into 2D Cartesian coordinate slices, convert the 2D Cartesian coordinate slices into polar coordinate images, and perform striping processing on the polar coordinate images based on the radial symmetry of the particles. Feature extraction steps: Feature extraction is performed on the image data after striping using a deep learning framework; Boundary binary classification subtask segmentation steps: Decompose the multi-layer structure segmentation problem of fuel particles into multiple boundary binary classification subtasks, and obtain the segmentation results of each boundary binary classification subtask based on the feature information extracted in the feature extraction step. Feature fusion steps: Define a structural prior matrix to encode the order relationship between layers, and use a conditional random field model to fuse the segmentation results of multiple boundary binary classification sub-tasks in combination with the structural prior matrix to ensure topological consistency between layers. The reconstruction steps are as follows: the strip segmentation results after feature fusion are stitched together, and the stitched polar coordinate image is converted back to a Cartesian coordinate system image to obtain the final segmentation result.

2. The image segmentation method for layered micro / nano structures coated with fuel particles as described in claim 1, characterized in that, In the feature extraction step, the deep learning framework is built on the U-Net architecture. Before each convolutional block of the encoder and decoder in the U-Net architecture, a positional encoding module is added. The position encoding module is used to explicitly encode radial positional relationships in polar coordinate images, which is achieved through a learnable function related to relative distance. The function is implemented by fitting a convolutional layer; the output of the position encoding module is concatenated with the original input feature map as an additional channel.

3. The image segmentation method for layered micro / nano structures coated with fuel particles as described in claim 2, characterized in that, The feature extraction step also includes a perturbation module; The perturbation module applies multiple affine transformations to the input image, inputs the transformed images into the deep learning framework for feature extraction, performs corresponding inverse transformations on the extraction results, and then obtains the final feature extraction result through weighted fusion.

4. The image segmentation method for layered micro / nano structures coated with fuel particles as described in any one of claims 1-3, characterized in that, During the model training phase, the boundary-aware weighted loss function is used for supervised training in the boundary binary classification subtask segmentation step. The boundary-aware weighted loss function is a weighted binary cross-entropy loss. Its weight factor is dynamically calculated based on the normalized radial distance from the pixel to a specific boundary, so that pixels closer to the boundary have a higher weight in the loss function.

5. The image segmentation method for layered micro / nano structures coated with fuel particles as described in any one of claims 1-3, characterized in that, During the model training phase, the weight matrix of the conditional random field model in the feature fusion step is trained under supervision using the classification cross-entropy loss function.

6. The image segmentation method for layered micro / nano structures coated with fuel particles as described in any one of claims 1-3, characterized in that, Structural Prior Matrix It is OK A matrix of columns, where, It is the total number of categories of the coating area. It is the total number of boundary binary classification subtasks; matrix Each element in Represents the coding category With boundary The spatial relationship between them takes a value of 1 or -1, where 1 represents the category. Located at the border Within, -1 indicates the category Located at the border In addition.

7. The image segmentation method for layered micro / nano structures coated with fuel particles as described in claim 6, characterized in that, Conditional random field (CRF) models map the segmentation probabilities of multiple boundary binary classification subtasks into vectors, multiply them by the structural prior matrix, and then pass them through a normalization function. Normalization is used to achieve the final multi-class probability prediction: （11） in, This represents the probability that a pixel in the image after feature fusion belongs to each category. Through The generated data represents multiple binary tags. This represents multiple binary segmentation networks. Let represent the diagonal weight matrix, used to adjust the contribution of each boundary, and described by equation (10): （10） In the formula, Indicates the first The weights of the segmentation results of each boundary segmentation task. .

8. An image segmentation device for coating fuel particle layered micro / nano structures, characterized in that, include: The data preprocessing unit is used to convert the raw 3D CT data of the fuel particles into 2D Cartesian coordinate slices, convert the 2D Cartesian coordinate slices into polar coordinate images, and perform striping processing on the polar coordinate images based on the radial symmetry of the particles. The feature extraction unit is used to extract features from the image data after striping and blocking using a deep learning framework; The boundary binary classification subtask segmentation unit is used to decompose the multi-layer structure segmentation problem of coated fuel particles into multiple boundary binary classification subtasks, and obtain the segmentation result of each boundary binary classification subtask based on the feature information extracted by the feature extraction unit. The feature fusion unit is used to define the structural prior matrix to encode the order relationship between layers, and uses the conditional random field model to fuse the segmentation results of multiple boundary binary classification sub-tasks in combination with the structural prior matrix to ensure the topological consistency between layers. The result reconstruction unit is used to stitch together the strip segmentation results after feature fusion and convert the stitched polar coordinate image back to a Cartesian coordinate system image to obtain the final segmentation result.

9. The image segmentation apparatus for coating fuel particle layered micro / nano structures as described in claim 8, characterized in that, In the feature extraction unit, the deep learning framework is built on the U-Net architecture. Before each convolutional block of the encoder and decoder in the U-Net architecture, a positional encoding module is added. The position encoding module is used to explicitly encode radial positional relationships in polar coordinate images, which is achieved through a learnable function related to relative distance. To achieve this, the function is fitted through a convolutional layer; the output of the position encoding module is concatenated with the original input feature map as an additional channel. The feature extraction unit also includes a perturbation module; The perturbation module applies multiple affine transformations to the input image, inputs the transformed images into the deep learning framework for feature extraction, performs corresponding inverse transformations on the extraction results, and then obtains the final feature extraction result through weighted fusion.

10. The image segmentation method for layered micro / nano structures coated with fuel particles as described in claim 8 or 9, characterized in that, Structural Prior Matrix It is OK A matrix of columns, where, It is the total number of categories of the coating area. It is the total number of boundary binary classification subtasks; matrix Each element in Represents the coding category With boundary The spatial relationship between them takes a value of 1 or -1, where 1 represents the category. Located at the border Within, -1 indicates the category Located at the border In addition; Conditional random field (CRF) models map the segmentation probabilities of multiple boundary binary classification subtasks into vectors, multiply them by the structural prior matrix, and then pass them through a normalization function. Normalization is used to achieve the final multi-class probability prediction: （11） in, This represents the probability that a pixel in the image after feature fusion belongs to each category. Through The generated data represents multiple binary tags. This represents multiple binary segmentation networks. Let represent the diagonal weight matrix, used to adjust the contribution of each boundary, and described by equation (10): （10） In the formula, Indicates the first The weights of the segmentation results of each boundary segmentation task. .

Images