Intelligent segmentation method and device for ceramic matrix composite CT image defects

Abstract

The invention discloses an intelligent segmentation method and device for ceramic matrix composite CT image defects, and belongs to the technical field of image detection. The method comprises the following steps: inputting preprocessed CT image data into a preset double-branch collaborative segmentation model, and respectively outputting to obtain a local texture feature map and a global relation feature map; performing three-branch material perception fusion processing on the local texture feature map and the global relation feature map according to a preset perception fusion model to obtain a fusion feature map containing local detail accuracy and global consistency; sequentially carrying out progressive decoding and refined reconstruction processing on the fused feature graph to obtain a defect probability graph; and calculating a loss function of the defect probability graph and a manually pre-labeled real defect segmentation graph, and dynamically adjusting the loss weight of the initial model according to a calculation result to obtain a CT image defect segmentation model meeting a preset convergence condition. The method can meet the detection requirements of industrial scenes on the defects of the ceramic-based composite material.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image detection, in particular to a method and device for intelligent segmentation of defects in CT images of ceramic matrix composites. BACKGROUND

[0002] The defect detection technology of ceramic matrix composites is a non-destructive testing method specially used for identifying and evaluating various types of defects inside and on the surface of ceramic matrix composites. The material is composed of a ceramic matrix and reinforcing fibers. Due to its high-temperature resistance and high specific strength, it is widely used in aerospace and other fields, but defects such as porosity, cracks, fiber breakage, and interface debonding are easily generated during manufacturing and use, which seriously affect its reliability. The detection technology aims to accurately find these micro-damages before the material is put into use or during service, to prevent problems from occurring and to ensure the safety and life of the components, providing important data support, which is a key technology for ensuring the reliability of high-end equipment.

[0003] In related technologies, traditional image processing methods fail due to low contrast in material CT imaging, and ignore material anisotropy. At the same time, deep learning methods lack consideration of material properties and cannot establish a wide range of defect associations, resulting in low computational efficiency and difficulty in meeting the detection needs of material defects in industrial scenarios.

[0004] Therefore, there is an urgent need for an intelligent segmentation method and device for defects in CT images of ceramic matrix composites to solve the above technical problems. SUMMARY

[0005] The present application provides an intelligent segmentation method and device for defects in CT images of ceramic matrix composites, which can meet the detection needs of defects in ceramic matrix composites in industrial scenarios. The technical solution is as follows: On the one hand, an intelligent segmentation method for defects in CT images of ceramic matrix composites is provided, which includes: inputting the preprocessed CT image data into a preset double-branch collaborative segmentation model to respectively output a local texture feature map for representing defect information from micro to macro and a global relationship feature map for representing global context association information of the image; wherein the double-branch collaborative segmentation model includes a CNN branch and a Transformer branch; performing three-branch material perception fusion processing on the local texture feature map and the global relationship feature map according to a preset perception fusion model to obtain a fusion feature map containing local detail accuracy and global consistency; performing progressive decoding and fine reconstruction processing on the fusion feature map in sequence to obtain a defect probability map; wherein the pixel value of the defect probability map is used to represent the probability of each pixel in the image belonging to a defect; The loss function of the defect probability map and the artificially pre-labeled real defect segmentation map is calculated, and the initial model is dynamically adjusted in loss weight according to the calculation result, to obtain a CT image defect segmentation model meeting a preset convergence condition.

[0006] In another aspect, an intelligent segmentation device for CT image defects of a ceramic matrix composite material is provided, and the device comprises: The detection module is configured to input the preprocessed CT image data into a preset double-branch collaborative segmentation model, and output a local texture feature map representing defect information from micro to macro and a global relationship feature map representing global context correlation information of the image, respectively. The fusion module is configured to perform three-branch material perception fusion processing on the local texture feature map and the global relationship feature map according to a preset perception fusion model, to obtain a fusion feature map containing local detail accuracy and global consistency. The reconstruction module is configured to sequentially perform progressive decoding and refinement reconstruction processing on the fusion feature map, to obtain a defect probability map, wherein a pixel value of the defect probability map represents a probability that each pixel in the image belongs to a defect. The update module is configured to calculate a loss function of the defect probability map and an artificially pre-labeled real defect segmentation map, and dynamically adjust the initial model in loss weight according to the calculation result, to obtain a CT image defect segmentation model meeting a preset convergence condition.

[0007] In another aspect, a computer device is provided, which comprises a memory and a processor, the memory is configured to store a computer program, and the processor is configured to execute the computer program stored in the memory to implement the steps of the intelligent segmentation method for CT image defects of a ceramic matrix composite material.

[0008] In another aspect, a computer readable storage medium is provided, which stores a computer program, and the computer program is executed by a processor to implement the steps of the intelligent segmentation method for CT image defects of a ceramic matrix composite material.

[0009] In another aspect, a computer program product is provided, which comprises a computer program, and the computer program is executed by a processor to implement the steps of the intelligent segmentation method for CT image defects of a ceramic matrix composite material.

[0010] The technical scheme provided by the application can bring at least the following beneficial effects: the local texture feature detection is used to realize accurate identification of micro features in the ceramic matrix composite material, the global relationship feature detection is used to realize accurate capture of the spatial distribution law and connectivity features of the large-scale defect network, then the complementary features of the two branches are intelligently integrated, the fusion weight is automatically adjusted according to the physical characteristics of the ceramic matrix composite material, the detection result is ensured to have both the accuracy of local details and the global consistency, finally, the initial model is trained by using the defect probability map obtained by detection and the real probability distribution map manually labeled, and an intelligent CT image defect segmentation model is obtained. The method promotes the fundamental change of the entire industry from traditional experience judgment to data-driven decision by establishing an AI detection-based quality evaluation standard for the ceramic matrix composite material. Compared with the traditional manual detection, the detection efficiency of the application is improved by more than 10 times, and the detection cost is reduced by 60%, which creates an important condition for the large-scale application of high-end composite materials. BRIEF DESCRIPTION OF DRAWINGS

[0011] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can obtain other drawings according to these drawings without creative labor.

[0012] Figure 1 is a flow chart of the intelligent segmentation method of the ceramic matrix composite material CT image defect provided by an embodiment of the present application; Figure 2 is an architecture diagram of the ceramic matrix composite material CT image processing provided by an embodiment of the present application; Figure 3 is a structure diagram of the intelligent segmentation device of the ceramic matrix composite material CT image defect provided by an embodiment of the present application; Figure 4 is a hardware architecture diagram of a computer device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0013] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0014] As described previously, the traditional image processing method fails due to low contrast of material CT imaging, and ignores material anisotropy, while the deep learning method lacks consideration of material characteristics and cannot establish a wide range of defect correlation.

[0015] Based on this, the concept of the present application is to obtain a local texture feature map and a global relationship feature map of the image by constructing a double-branch detection model containing a CNN branch and a Transformer branch, and to obtain a feature map rich in material characteristics and defect information by perceiving and fusing the two from multiple branches, so as to complete the training of the model.

[0016] The specific implementation of the above concept is described below.

[0017] Please refer to Figure 1 The ceramic matrix composite CT image defect intelligent segmentation method provided by the embodiment of the present application comprises: Step 100, input the CT image data obtained after preprocessing into a preset double-branch cooperative segmentation model, and respectively output a local texture feature map for representing defect information from micro to macro and a global relationship feature map for representing global context correlation information of the image; wherein the double-branch cooperative segmentation model comprises a CNN branch and a Transformer branch; Step 102, according to a preset perception fusion model, perform three-branch material perception fusion processing on the local texture feature map and the global relationship feature map to obtain a fusion feature map containing local detail accuracy and global consistency; Step 104, sequentially perform progressive decoding and fine reconstruction processing on the fusion feature map to obtain a defect probability map; wherein the pixel value of the defect probability map is used to represent the probability of each pixel in the image belonging to a defect; Step 106, calculate the loss function of the defect probability map and the artificially pre-labeled real defect segmentation map, and dynamically adjust the loss weight of the initial model according to the calculation result to obtain a CT image defect segmentation model meeting the preset convergence condition.

[0018] In the embodiment of the present application, the local texture feature detection is used to accurately identify the micro features in the ceramic matrix composite material, and the global relationship feature detection is used to accurately capture the spatial distribution law and connectivity features of the large-scale defect network. Then, the complementary features of the two branches are intelligently integrated, the fusion weight is automatically adjusted according to the physical characteristics of the ceramic matrix composite material, and the detection result is ensured to maintain the accuracy of local details and global consistency. Finally, the initial model is trained using the defect probability map obtained by detection and the real probability distribution map labeled by artificial marking, and an intelligent CT image defect segmentation model is obtained. The method establishes an AI detection-based quality evaluation standard for ceramic matrix composite materials, and promotes the fundamental change of the entire industry from traditional experience judgment to data-driven decision-making. Compared with the traditional manual detection, the detection efficiency of the present application is improved by more than 10 times, and the detection cost is reduced by 60%, which creates important conditions for the large-scale application of high-end composite materials.

[0019] The execution mode of each step is described below. Figure 1

[0020] Firstly, for step 100, the preprocessed CT image data is input into the preset double-branch collaborative segmentation model, and local texture feature maps and global relationship feature maps for representing defect information from micro to macro are output respectively.

[0021] As shown in Figure 2 The double-branch collaborative segmentation model includes a CNN branch and a Transformer branch, and through the collaborative work of the two, the thinking mode of a material expert in analyzing a CT image is simulated: both observing local detail features and grasping the overall defect distribution law and connectivity. The core of this design concept lies in the recognition that the essential challenge of ceramic matrix composite defect detection is not only an image processing problem, but also a complex problem of the interaction between material physical properties and imaging mechanisms.

[0022] Among them, the CNN branch adopts the ConvNeXt architecture as the backbone network, contains a deep convolution structure of 4 stages, and the feature dimension configuration is in a progressive mode of 64-128-256-512; the Transformer branch is designed based on the Swin Transformer architecture, adopts a window size of 7x7 and a multi-head attention configuration of [3, 6, 12, 24], and effectively processes long-distance spatial dependence through window attention mechanism and cyclic shift strategy (shift size is 3) ​Specifically, the CNN branch serves as a local texture expert, adopting a modern convolutional network architecture to realize accurate identification of microscopic defects through efficient convolution processing technology. The core of this technology is a deep separable convolution structure based on the ConvNeXt architecture, which significantly reduces computational complexity by decomposing standard convolution operations into two steps: depth convolution and point convolution. Depth convolution is responsible for spatial feature extraction, while point convolution is responsible for inter-channel information interaction. This decomposition strategy significantly reduces parameter quantity and computational overhead while maintaining feature extraction capability. This technology optimizes the spatial utilization efficiency of the convolution kernel and the channel information flow mechanism, specifically addressing the problem of insufficient imaging contrast in the interface region due to the close linear attenuation coefficients of SiC fibers and SiC matrix, enabling the algorithm to accurately identify key defects such as fiber-matrix interface debonding in weak signals.

[0023] This branch adopts a large-range convolution design, specifically a 7x7 kernel depth convolution operation, which can capture more texture information in a single operation compared to traditional 3x3 small kernel convolution. This technology can obtain broader contextual information without increasing network depth by expanding the spatial coverage of a single convolution operation, making it particularly suitable for capturing linear defects with directional characteristics in ceramic matrix composites. This large-range convolution design specifically addresses the challenge of detecting cracks distributed along the fiber direction, as these defects often have an elongated geometry that requires a larger receptive field to fully capture their continuity features. By enhancing sensitivity to directional textures, this technology significantly improves the ability to recognize anisotropic defect features in ceramic matrix composites.

[0024] In the embodiments of the present application, the process of obtaining a local texture feature map through the CNN branch includes: S21, performing initial downsampling on the CT image data to obtain a starting feature map; S22, inputting the starting feature map into a ConNeXt module based on large kernel depth convolution for processing, and outputting a local feature map containing defect information; S23, performing downsampling processing on the local feature map according to a convolution layer with a preset stride to halve the size of the feature map and double the number of channels; S24, re-inputting the downsampled local feature map into the ConvNeXt module and repeatedly executing steps S22-S24 until a preset number of processing times is reached, to obtain local texture feature maps of defect information at different scales.

[0025] Specifically, the CNN branch uses a four-stage 64-128-256-512 dimensional feature progressive mode through a multi-scale progressive feature extraction strategy, with different sizes of analysis range processing different morphologies of defects such as point porosity, linear crack, and planar debonding. This technology realizes progressive feature modeling from local texture to global morphology by gradually expanding the receptive field and increasing the level of feature abstraction.

[0026] For example, the first stage is dedicated to capturing 1-5 microns of point-like pores, which are mainly derived from the gas volatilization of the polymer pyrolysis process in the PIP process, and are manifested as spherical or near-spherical closed spaces.

[0027] Assuming that the size of the initial feature map F0 is [B, 64, 64, 64], it is sequentially subjected to depth convolution, layer normalization, 1x1 convolution, activation function, residual connection and regularization processing through a series of ConvNeXt blocks to obtain the local feature map F1 ([B, 64, 64, 64]), which has high feature resolution and is focused on capturing the finest point-like pores and fiber textures.

[0028] Further, the second stage processes 10-50 microns of matrix microcracks, which are usually caused by thermal stress concentration due to mismatch of thermal expansion coefficients, and exhibit linear features along the weak surface of the matrix.

[0029] Before entering the second stage, down-sampling is performed through a 2x2 convolution layer with a stride of 2. The size of the local feature map F1 is halved (64x64 becomes 32x32), and the number of channels is doubled (64 becomes 128), and then the processing process of the feature map in the first stage is repeated, and the local feature map F2 ([B, 128, 32, 32]) is output. In this stage, the receptive field is further expanded, and the micro features are integrated to form the perception of medium-sized defects (such as matrix microcracks).

[0030] Further, the third stage is optimized for fiber-matrix interface debonding, which reflects the insufficient interfacial bonding strength and is manifested as ring-shaped separation along the fiber periphery.

[0031] In this stage, the local feature map F2 is subjected to the processing process of the second stage to obtain the local feature map F3 ([B, 256, 16, 16]). Thus, larger defects and more complex morphologies can be identified.

[0032] Finally, the fourth stage processes macroscopic crack networks above 100 microns.

[0033] Similarly, the local feature map F3 is subjected to the processing process of the second stage to obtain the local feature map F4 ([B, 512, 8, 8]). This is the most abstract feature representation, and each feature point carries a large range of original image information.

[0034] In summary, the final output of the CNN branch is not a single feature map, but a multi-scale feature pyramid: {F1, F2, F3, F4}. These feature maps of different scales will be provided to the decoder and material perception fusion module through skip connection in the subsequent processing. In this way, the decoder can utilize both high-level semantic information (from F3, F4) and fine detail information (from F1, F2) when reconstructing the image, thereby achieving accurate defect boundary segmentation.

[0035] This hierarchical processing strategy solves the technical problem that traditional single-scale methods cannot simultaneously handle defects spanning three orders of magnitude, ensuring that each defect type is accurately identified at its most suitable feature scale.

[0036] The Transformer branch, as a global relationship expert, uses a window attention mechanism based on Swin Transformer to establish direct connections between any two points in the image through global feature correlation analysis. The core of this technology is the self-attention mechanism, which calculates the correlation weight between each position and all other positions in the image through the interaction of query, key, and value matrices, thereby establishing long-range spatial dependency. This mechanism uses a windowing strategy to divide the image into 7x7 non-overlapping regions, and calculates attention weights within each window. Cross-window information exchange is achieved through a cyclic shift strategy. This technical design specifically addresses the spatial correlation modeling challenge of complex defect networks in ceramic matrix composites, enabling accurate identification of defect fragments that are separate on the surface but actually belong to the same crack propagation path, which is of great significance for understanding material damage evolution and failure mechanism analysis.

[0037] In the embodiment of the application, obtaining the global relationship feature map through the Transformer branch includes: performing sequence conversion on the CT image data to obtain a starting feature sequence for processing by the Transformer branch; performing calculation and integration processing on the starting feature sequence according to a window self-attention mechanism and a displacement window self-attention mechanism to obtain a fusion feature sequence containing global context information; and performing spatial reorganization and hierarchical down-sampling processing on the fusion feature sequence to obtain global relationship feature maps at multiple different scales.

[0038] In particular, assuming that a convolutional layer with a kernel size of 4x4 and a stride of 4 is used in the process of image-to-sequence conversion, and the number of convolution kernels (output channels) is set to a base dimension, for example, 96, the input CT image can be divided into a series of non-overlapping image blocks of 4x4 pixels, for example, for an input of 256x256, there will be (256 / 4) x (256 / 4) = 64 x 64 = 4096 blocks. Each 4x4 image block (a total of 16 pixels) is flattened and projected into a 96-dimensional feature vector, and each vector represents an abstract representation of an image block. In this way, the feature map can be converted from the image format of [B, 1, 256, 256] to the initial feature sequence format of [B, 4096, 96].

[0039] Further, based on the window-based self-attention mechanism, for each sequence within a window, a query vector, a key vector, and a value vector are generated for each image block through three different linear layers. For each query vector (Query), the similarity (usually through a dot product operation) between it and all key vectors (Key) within the window is calculated. The higher the similarity, the stronger the correlation between the two pixels. Then, the similarities are normalized through a Softmax function to obtain a set of attention weights, and then the set of attention weights is used to perform weighted summation on all value vectors (Value), so that the output of each image block is fused with the context information of all other image blocks within the window.

[0040] Further, since the window self-attention is only performed within non-overlapping fixed windows, there is no information exchange between windows. This can result in a still limited "field of view" and inability to establish long-distance dependencies across the entire image. To solve this problem, the embodiment also adopts the cyclic shift window strategy of Swin Transformer.

[0041] For example, at some layers of the network (e.g., even layers), the standard non-overlapping window division method as described above is used; at the next layer (e.g., odd layers), the feature map is cyclically shifted by a plurality of pixels in the height and width directions (for a 7x7 window, shifted by 3 pixels). Then, a 7x7 window division is applied again on the new feature map after the shift.

[0042] This shift operation creates a "new" window, which is composed of sub-regions from four different corners that were originally not adjacent. This allows tokens in different windows of the previous layer to be "mixed" into the same new window in the current layer, enabling direct attention calculation and information flow across windows. This is equivalent to greatly expanding the receptive field of the model without increasing the amount of calculation.

[0043] It is worth mentioning that the processing of this branch is similar to the CNN branch, which is also a hierarchical architecture. Instead of working on one resolution, it builds a feature pyramid through downsampling to understand the global relationship at different levels of abstraction.

[0044] Specifically, after several consecutive Swin Transformer blocks (including standard windows and shift windows) are completed, a Patch Merging operation is performed. This is similar to the pooling layer in CNN. It concatenates the feature maps of adjacent 2x2 small windows, and then reduces the dimension through a linear layer, so that the spatial size (H, W) of the feature map is halved and the number of channels is doubled. Through multiple downsampling, the Transformer branch outputs a set of multi-scale feature maps corresponding to the output size of the CNN branch, for example, the channel dimension is also 64, 128, 256, 512 in turn. The output of each layer contains the long-distance context information between all regions in the image at that scale. In summary, this branch uses a regional progressive analysis method to achieve efficient computation of large-scale spatial correlation using a cyclic shift window strategy. This technology uses a standard window division at even layers and a 3-pixel shift window at odd layers to establish cross-regional information connections by dynamically adjusting the window position. This progressive analysis strategy maintains controllable computational complexity while accurately modeling complex crack networks over a range of hundreds of microns. This technology specifically addresses the integrity reconstruction problem of the main crack system of ceramic matrix composites, as these cracks often extend for a very long distance, forming a complex three-dimensional network structure. Traditional methods cannot establish this large-scale spatial correlation due to the limited receptive field, leading to the misidentification of different fragments of the same crack network as independent defects, which seriously affects the accurate assessment of material damage.

[0045] Then, for step 102, the local texture feature map and the global relationship feature map are subjected to three-branch material perception fusion processing according to a preset perception fusion model, to obtain a fusion feature map containing local detail accuracy and global consistency.

[0046] In the embodiments of the present application, a material perception fusion mechanism is innovatively designed for the physical properties of ceramic matrix composites, specifically solving the technical problem that traditional image processing methods cannot effectively identify fiber-matrix interfaces and multi-scale defects. The core idea of this mechanism is to integrate material science knowledge into intelligent algorithm design, and to realize automated defect detection by simulating the analysis ideas of material experts.

[0047] Traditional image analysis methods treat ceramic matrix composites as ordinary two-dimensional images for processing, ignoring the anisotropy and fiber arrangement of the material itself, resulting in insufficient detection accuracy. The material perception mechanism of the embodiment adopts a multi-branch parallel processing strategy, and specializes in processing the fiber directionality, defect size and interface features respectively.

[0048] In the embodiment of the application, the fusion feature map is obtained by the following method: aligning and preliminarily fusing all local texture feature maps and global relationship feature maps of the same scale to obtain basic feature maps under different scales; inputting all the basic feature maps into a fiber direction perception branch sub-model for identification to output a first feature map containing crack expansion information along different fiber directions; inputting all the basic feature maps into a defect size perception branch sub-model for identification to output a second feature map containing crack defect size information; inputting all the basic feature maps into an interface boundary enhancement branch sub-model for identification to output a third feature map containing complete texture information on both sides of the fiber-matrix interface; and performing fusion processing on the first feature map, the second feature map and the third feature map according to a preset branch contribution weight to obtain a fusion feature map with local detail accuracy and global consistency.

[0049] Specifically, since the local texture feature map and the global relationship feature map both contain different scales, they need to be fused at the corresponding scales in the preliminary fusion process. The preliminary fusion process includes splicing them in the channel dimension, and then reducing the number of channels after splicing to a unified dimension through 1x1 convolution to obtain the basic feature map. The interactive mechanism adopts a layer-by-layer feature alignment strategy to ensure that features of the same scale from different branches can be effectively fused at the same spatial resolution.

[0050] Further, fiber direction detection is performed on all scale base feature maps, and the fiber direction perception branch specifically identifies and adapts to different fiber weaving angles, and can automatically adjust the detection strategy to adapt to different fiber arrangement modes such as 0°, 30°, 45°, 90°, etc., which is crucial for accurately identifying cracks extending along the fiber direction. This branch uses a 3x3 depth separable convolution kernel, and through group convolution parameter setting of channels / / 4, the calculation efficiency is optimized. Specifically, the 3x3 convolution kernel can capture directional texture features within a local 9-pixel range, and through the group convolution mechanism, the input feature channels are divided into 4 independent processing groups, each of which specifically learns the fiber arrangement pattern in a specific direction range. When the fibers in the material are arranged along the 0° horizontal direction, the convolution filter of this branch can learn the horizontal direction response pattern; for 30° or 45° diagonal fiber weaving, different groups of group convolution automatically adapt to the corresponding inclined angle features; and 90° vertical fiber arrangement activates the vertical direction sensitive filter group. This group processing mechanism enables a single branch to adapt to multiple fiber directions, and automatically adjusts the detection sensitivity to different direction fiber arrangement through weight learning during the training process.

[0051] At the same time, defect scale detection is performed on all scale base feature maps, and the defect scale perception branch specifically processes cross-scale features from microscopic pores to macroscopic cracks, and through adaptive feature extraction strategies, ensures accurate identification of defects of different sizes. This branch uses a 5x5 medium-sized convolution kernel to establish a 25-pixel local receptive field, which provides a larger spatial coverage range than a 3x3 convolution kernel, and can effectively capture medium-scale defect feature patterns. The design of the 5x5 convolution kernel takes into account the distribution characteristics of defect sizes in ceramic matrix composites: it can detect small pores of a few pixels in size, and also identify medium crack segments spanning several dozen pixels. Through cooperation with other branches (3x3 and 7x7 convolution kernels), this branch undertakes the intermediate-scale feature learning task in the multi-scale feature extraction system. The application of group convolution enables this branch to parallelly process features within different scale ranges, and through adaptive learning of feature representations of different defect types during the training process, ensures that both microscopic residual pores produced by the polymer impregnation pyrolysis (PIP) process and macroscopic cracks formed during material use can be accurately captured.

[0052] Interface enhancement branch is designed to optimize the detection accuracy of fiber-matrix interface area by expanding the analysis range and enhancing the contrast. This branch uses a 7x7 large-size convolution kernel to establish a 49-pixel extended receptive field, which provides a significantly larger spatial context analysis range compared to small-size convolution kernels. This large receptive field design can simultaneously observe the complete texture information on both sides of the fiber-matrix interface, and compensate for the insufficient contrast in the local area by capturing more extensive material features around the interface. The combination of 7x7 convolution kernel and grouped convolution enables the branch to extract interface features from a larger spatial range, effectively reducing the detection instability caused by relying solely on local pixel differences. In addition, the output of this branch works with the CoordinateAttention coordinate attention mechanism to further enhance the attention to the interface area through spatial position encoding. The CoordinateAttention mechanism processes horizontal and vertical spatial information separately and generates position-dependent attention weights through adaptive pooling and 1x1 convolution, effectively amplifying the feature response strength of the interface area, thus achieving contrast enhancement on the whole. The combination of this branch design and attention mechanism provides an effective solution to the technical challenge of extremely low contrast between SiC fibers and SiC matrix in X-ray CT imaging.

[0053] This mechanism establishes the relationship between different areas in the image through spatial correlation analysis technology, which is particularly suitable for processing linear defects and directional features commonly found in ceramic matrix composites. The results of the three specialized processing branches are integrated through an intelligent fusion algorithm, which automatically adjusts the contribution weights of each branch to ensure that the final detection results maintain both local detail accuracy and global consistency. This design enables the algorithm to simultaneously focus on both the micro-texture features and macro-structural characteristics of the material, significantly improving the recognition ability of complex defect morphologies.

[0054] Finally, the final outputs of the three branches are weighted and summed. For example, the weights can be configured as a = 0.4 (fiber direction), b = 0.35 (defect scale), and g = 0.25 (interface boundary). These weights are pre-set hyperparameters that reflect the prior knowledge of the importance of the three factors and can be adjusted according to actual needs.

[0055] The final result is a fused feature map at each scale, which contains both the local details of CNN and the global relationships of Transformer. More importantly, it is reshaped by physical knowledge, and its feature representation is more consistent with the laws of material science.

[0056] For step 104, the fusion feature map is sequentially subjected to progressive decoding and fine reconstruction processing to obtain a defect probability map.

[0057] The embodiment adopts a progressive fine strategy to restore the complete morphology and accurate boundary of defects, which is a core technical process specially designed for the characteristics of high defect detection accuracy of ceramic matrix composites. The process is realized by a Progressive Decoder progressive decoder, which simulates the thinking process of a material expert analyzing CT images: from the whole to the part, from rough to fine, and gradually confirms the accurate information of the defect. This progressive strategy is organically combined with the aforementioned intelligent training management strategy, and through multi-stage feature reconstruction and spatial correlation analysis, the technical difficulties of missing detection or false detection of traditional one-step detection strategy in the face of complex defect morphology are solved.

[0058] In the embodiment of the application, the decoding and reconstruction process specifically comprises: S51, performing up-sampling processing on the input fusion feature map, and performing jump splicing processing on the image obtained by up-sampling and the base feature map of the same scale to obtain a spliced feature map; S52, sequentially performing convolution refinement and spatial optimization processing on the spliced feature map to obtain a decoding map; S53, repeating steps S51-S53 until a preset processing number is reached, taking the decoding map as input to obtain a final reconstruction map; S54, sequentially performing interface enhancement and boundary refinement processing on the final reconstruction map to obtain a defect probability map.

[0059] Specifically, the progressive detection first establishes the basic outline of the defect at a lower resolution, and captures the overall features of the defect area using a larger analysis range. This technology is realized through a three-stage up-sampling strategy, and each stage performs repeated operations. The feature dimension of the fusion feature map is reduced in the mode of 512→256→128→64, and each stage uses transposed convolution to recover the spatial resolution by 2 times up-sampling.

[0060] Firstly, the core of the larger analysis range is to utilize the large receptive field characteristics of deep features. At the 512-dimensional feature level, a single feature point can perceive information within a range of about 60x60 pixels. This large-range perception capability specially solves the problem of overall morphology identification of the main crack system in ceramic matrix composites. The main crack often extends hundreds of microns or even several millimeters, forming a complex three-dimensional network structure, which requires a large range of context information to accurately determine its continuity and expansion direction.

[0061] For example, using transpose convolution with kernel size of 2x2 and stride of 2 for upsampling, the spatial resolution of the fused feature map is doubled (e.g., from 8x8 → 16x16), while the number of channels is halved (512-dim → 256-dim); then the upsampled decoded fused feature map is concatenated with the original fused feature map from the fusion encoding stage, which also has 256 channels, and dimension adjustment and information integration are performed through 1x1 convolution. This technique can establish direct associations between pixels that are far apart in the image but actually belong to the same physical entity, which is of great significance for analyzing the crack propagation path in ceramic matrix composites. The propagation of cracks in ceramic matrix composites often follows the principle of minimum energy, bypassing fiber bundles or propagating along weak interfacial regions, resulting in discontinuous fragments in two-dimensional CT slices. This technique can accurately reconstruct the complete propagation path of the crack by establishing spatial relationships between these fragments, providing important evidence for understanding the damage mechanism of the material and predicting failure modes.

[0062] Further, the CoordinateAttention spatial coordinate perception technology is used to realize effective integration of cross-scale features. The core technology of spatial correlation analysis is to perform adaptive pooling on the input features along the height and width directions respectively, encode the spatial dimension information into coordinate vectors through the pooling operation, and then perform feature transformation through 1x1 convolution and SiLU activation function, finally generate attention weights in horizontal and vertical directions. This mechanism is specifically optimized for the directional features of ceramic matrix composites, and can simultaneously focus on the feature dependency relationship in horizontal and vertical directions, and is particularly suitable for processing linear defects distributed along the fiber direction. At the same time, when processing fiber-reinforced ceramic matrix composites, this technology can accurately identify linear defects that are broken in local areas due to noise interference by enhancing the connection of features along the fiber direction, ensuring the integrity reconstruction of long strip-shaped defects such as cracks.

[0063] After repeating the upsampling strategy three times, the embodiment also realizes sub-pixel level precision defect boundary reconstruction through a special interface enhancement algorithm and boundary refinement technology. The core of the interface enhancement algorithm is the interface boundary enhancement branch in the MaterialAwareAttention material perception attention mechanism. This technology uses a 7x7 convolution kernel to specifically handle the low contrast detection problem of the fiber matrix interface area. This algorithm can accurately identify key defect types such as interface debonding by expanding the convolution kernel size to enhance the ability to capture context information in the interface area, even in the case of extremely low imaging contrast. The interface area is the most prone to defects in ceramic matrix composites. Due to the similarity in X-ray attenuation coefficients between SiC fibers and SiC matrix, the interface is almost impossible to observe directly in CT imaging, and traditional methods often produce a large number of false positive or false negative results in this area. This algorithm significantly improves the accuracy of interface area defect detection through material-specific feature extraction strategies, combined with fiber arrangement rules and interface physical properties.

[0064] The boundary refinement technology realizes accurate positioning of the defect boundary through multi-level feature fusion and gradient optimization. This technology uses a combination of LayerNorm2d normalization and SiLU activation functions to ensure the numerical stability of feature propagation and the smoothness of gradient flow. The core mechanism of boundary refinement is to compare boundary features at different resolution levels, use the detailed information of high-resolution features to correct and supplement the boundaries of low-resolution features, and finally obtain a defect probability map. Each pixel value in the map directly represents "how likely is this pixel a defect". For example, a value of 0.95 indicates that the pixel has a 95% chance of being a defect. This technology is optimized specifically for the irregularity and fuzziness of ceramic matrix composite defect boundaries. Through the fusion of multi-scale boundary features, it can accurately reconstruct the true geometry of complex defects. The final output detection result is not only accurate in defect location, but also highly consistent with the actual physical characteristics of the defect in boundary shape, providing a reliable data foundation for subsequent defect quantification analysis and material performance evaluation.

[0065] For step 106, calculate the loss function of the defect probability map and the artificially pre-labeled true defect segmentation map, and dynamically adjust the loss weight of the initial model according to the calculation result to obtain a CT image defect segmentation model that meets the preset convergence condition.

[0066] The embodiment designs a training management strategy specially for the defect detection task of ceramic matrix composites, and realizes intelligent management of the training process through a DynamicModernLoss dynamic scheduling system. The strategy is based on the particularity of the material defect detection task, including serious sample imbalance, fuzzy boundary, and multi-scale coexistence, etc. A phased key adjustment method is adopted to simulate the gradual learning process from rough detection to fine segmentation. This intelligent management cooperates with the aforementioned double-branch collaborative architecture to ensure that the algorithm can learn rich ceramic matrix composite feature representations under limited labeled data through optimized loss function combination and dynamic weight scheduling.

[0067] In the embodiment of the present application, the CT image defect segmentation model is trained by the following method: according to the defect probability map and the real defect segmentation map, a multiple loss for representing the difference between the two kinds of images is calculated; wherein the multiple loss includes Dice loss, focal loss, asymmetric loss, boundary loss and Lovasz loss; the weight of each loss function is adjusted according to the training stage of the initial model, and the total loss of the model corresponding to each training stage is calculated according to the adjusted function weight; the gradient of the total loss of the model with respect to all weight parameters of the model network is calculated, and the network parameters are updated in the opposite direction of the gradient using a pre-set optimizer to reduce the total loss until the model meets the pre-set convergence condition, thereby obtaining a CT image defect segmentation model for defect segmentation.

[0068] Specifically, the training process provided by the embodiment innovatively integrates five complementary loss functions, which are carefully designed and dynamically scheduled according to the special needs of ceramic matrix composite defect detection.

[0069] The Dice loss function, as the core segmentation loss, adopts a smoothing parameter of 1e-6 to ensure numerical stability and specifically solves the accuracy problem of target area calculation.

[0070] The focal loss function effectively handles the foreground-background imbalance problem through parameter configuration of alpha=0.25 and gamma=2.0, and is particularly suitable for detection of small-sized defects.

[0071] The Tversky loss function adopts asymmetric weight settings of alpha=0.3 and beta=0.7, and optimizes the recall rate performance by adjusting the punishment intensity of false positives and false negatives, i.e., by assigning different weight coefficients to false positive predictions and false negative predictions to optimize the identification performance of the model for specific categories.

[0072] The Lovasz loss function is based on the Lovasz extension theory of submodular functions, and directly optimizes the IoU index, with per_image=True to ensure balanced processing of intra-batch samples.

[0073] The boundary loss function is designed specifically for boundary accuracy optimization, gradient calculation is performed using the Sobel operator, and the convolution kernel size is adjusted from 5x5 to 3x3 in the later training stage to improve boundary sensitivity.

[0074] After calculating the above five kinds of loss functions, instead of simply adding the five losses, a dynamic weight scheduling mechanism is adopted to train the model using a three-stage strategy to simulate the learning process from coarse to fine: in the early stage, the focus is on establishing basic segmentation ability, and the loss weight configuration is Dice 40%, Focal 30%, and Tversky 30%; in the middle stage, multi-objective balance is achieved, and boundary loss and Lovasz loss are introduced; in the later stage, the focus is on boundary refinement, and the boundary loss weight is increased to 25%, and the gamma parameter of the Focal loss is adjusted to 2.5 to enhance the learning ability of difficult samples. These loss functions achieve the identification and correction of unqualified results through multiple constraint mechanisms: fiber direction consistency check calculates the deviation angle between the detection result and the fiber arrangement direction, and applies an additional penalty weight of 0.1 times to the detection area that deviates more than 15°; morphological rationality verification analyzes the aspect ratio and shape factor of the defect, and constrains the detection result that does not meet the typical defect morphological characteristics; multi-scale consistency verification compares the detection results at different resolutions, and focuses on optimizing the areas with large scale differences to ensure the rationality and consistency of the detection results in the physical sense.

[0075] Specifically, the initial training focuses on establishing basic recognition ability of defect areas, and region overlap optimization and difficult sample learning are adopted as main targets. The region overlap optimization technology is realized through a unified loss function, and the Dice coefficient is used as an evaluation index. The intersection and union ratio of the predicted region and the real labeled region is calculated to quantify the segmentation accuracy. In specific implementation, the predicted result and the target label are flattened into one-dimensional vectors, the intersection of the two is calculated by intersection, and the union is calculated by summing the pixel values of each other. The final Dice coefficient is 2*intersection / (pred_sum + target_sum), and the smoothing parameter is set to 1e-6 to ensure numerical stability. The difficult sample learning technology is realized through a unified focal loss function. The standard cross-entropy loss is reweighted by introducing a modulation factor (1-pt)^γ, where pt represents the predicted probability of the correct class by the model, and the γ parameter is set to 1.5-2.0 to control the attention degree to difficult samples. The combination of these two technologies solves the detection problem of small defect area and weak signal in ceramic matrix composite CT imaging. Through the optimization of overlap degree calculation, it ensures that small size defects can be accurately positioned, and through the difficult sample learning mechanism, the algorithm focuses on the boundary fuzzy area which is easy to be misclassified.

[0076] In the middle of training, multi-target balanced optimization is realized, and on the basis of maintaining the accuracy of region detection, morphological constraints and boundary optimization targets are introduced. The morphological constraint technology is realized through a TverskyLoss asymmetric loss function. The technology adopts asymmetric weight configuration of α=0.3 and β=0.7, and adjusts the punishment intensity of false positive and false negative to optimize the recall rate of the model. This constraint mechanism is specially designed for the morphological characteristics of ceramic matrix composite defects, and the punishment intensity of false negative (missed detection) is increased to ensure that small fibers and micro-holes are not missed, which is of great significance to material quality evaluation. The boundary optimization target is realized through a UnifiedBoundaryLoss boundary loss function. The technology uses Sobel boundary detection operator to extract the boundary features of the predicted result and the real label, and identifies the boundary area by calculating the sum of squares of horizontal and vertical gradients. In specific implementation, the technology uses 3*3 Sobel convolution kernel to calculate the gradient in x and y directions respectively, and the boundary strength is calculated by sqrt(grad_x² + grad_y²). Finally, the consistency of the predicted boundary and the real boundary is optimized through mean square error loss. This boundary optimization design specially solves the accurate segmentation problem of low-contrast regions such as fiber matrix interface, and significantly improves the recognition accuracy of complex interface structure by strengthening the learning of boundary features.

[0077] In the later training stage, fine-grained optimization is focused on, and loss functions specially designed for ceramic matrix composites are introduced, including fiber direction consistency constraints and multi-scale detection consistency constraints. The fiber direction consistency constraint technology is realized through the MaterialAwareLoss material-aware loss function. This technology calculates the direction gradient using horizontal and vertical Sobel operators and imposes constraints by comparing the consistency of the main direction of the detection result with the known fiber arrangement direction. In the specific implementation, this technology calculates the gradient distribution of the prediction result in different directions, obtains the direction angle of each pixel through arctan(grad_y / grad_x), and matches it with the preset fiber direction (such as 0°, 30°, 45°, and 90°). An additional penalty weight is imposed on the detection result that deviates from the expected direction. The multi-scale detection consistency constraint technology calculates the loss at multiple resolution levels simultaneously by constructing a feature pyramid. A pyramid structure with [1, 2, 4] times downsampling is used to ensure that the same defect can be detected consistently at different observation scales. This technology independently calculates the Dice loss and boundary loss at each scale level, and then integrates the multi-scale information through weighted averaging, effectively solving the consistency problem of cross-scale defect detection. These specialized constraint techniques address the fundamental flaw of traditional image segmentation methods, which ignore the physical characteristics of materials. By incorporating the anisotropy and multi-scale features of ceramic matrix composites as prior knowledge into the training process, the detection results are ensured to conform to the actual physical laws of the material.

[0078] Finally, the gradient of the total loss with respect to all weight parameters of the network is calculated, and then the optimizer is used to update the network parameters in the opposite direction of the gradient to reduce the total loss. This process is repeated until the model converges, and finally a CT image defect segmentation model for defect segmentation is obtained.

[0079] For reference Figure 3 The embodiment of the present application provides a kind of intelligent segmentation device of ceramic matrix composite CT image defect, and the device comprises: The detection module 300 is used to input the CT image data obtained after preprocessing into the preset double-branch collaborative segmentation model, and output local texture feature maps for representing defect information from micro to macro and global relationship feature maps for representing image global context association information respectively;Wherein, the double-branch collaborative segmentation model includes CNN branch and Transformer branch; The fusion module 302 is used to perform three-branch material-aware fusion processing on the local texture feature map and the global relationship feature map according to the preset perception fusion model, and obtain a fusion feature map containing local detail accuracy and global consistency. The reconstruction module 304 is configured to sequentially perform progressive decoding and fine reconstruction processing on the fusion feature map to obtain a defect probability map, wherein a pixel value of the defect probability map is used to represent a probability of each pixel in the map belonging to a defect. The updating module 306 is configured to calculate a loss function of the defect probability map and an artificially pre-labeled real defect segmentation map, and dynamically adjust a loss weight of an initial model according to a calculation result to obtain a CT image defect segmentation model satisfying a preset convergence condition.

[0080] In the embodiment of the present application, the local texture feature map is obtained by the following method: S21, performing initial down-sampling on the CT image data to obtain a starting feature map; S22, inputting the starting feature map into a ConvNeXt module based on a large-core depth convolution for processing to output a local feature map containing defect information; S23, performing down-sampling processing on the local feature map according to a convolution layer with a preset stride to halve the size of the feature map and double the number of channels; S24, re-inputting the down-sampled local feature map into the ConvNeXt module, and repeatedly performing steps S22-S24 until a preset processing number is reached to obtain a local texture feature map containing defect information at different scales.

[0081] In the embodiment of the present application, the global relationship feature map is obtained by the following method: Performing sequence conversion on the CT image data to obtain a starting feature sequence for processing by a Transformer branch; Performing calculation and integration processing on the starting feature sequence according to a window self-attention mechanism and a displacement window self-attention mechanism to obtain a fusion feature sequence containing global context information; Performing spatial reorganization and hierarchical down-sampling processing on the fusion feature sequence to obtain a plurality of global relationship feature maps at different scales.

[0082] In the embodiment of the present application, when the fusion module 302 performs three-branch material perception fusion processing on the local texture feature map and the global relationship feature map according to a preset perception fusion model to obtain a fusion feature map containing material knowledge, the fusion module 302 is specifically configured to perform the following operations: Aligning and preliminarily fusing all local texture feature maps and global relationship feature maps of the same scale to obtain basic feature maps at different scales; Inputting all the basic feature maps into a fiber direction perception branch sub-model for identification to output a first feature map containing information about crack expansion along different fiber directions; input all the basic feature maps into the defect scale perception branch sub-model for identification, and output a second feature map containing crack defect size information; input all the basic feature maps into the interface boundary enhancement branch sub-model for identification, and output a third feature map containing complete texture information on both sides of the fiber-matrix interface; fuse the first feature map, the second feature map and the third feature map according to a preset branch contribution weight, to obtain a fused feature map with local detail accuracy and global consistency.

[0083] In the embodiment of the application, when the reconstruction module 304 performs progressive decoding and fine reconstruction processing on the fused feature map in sequence to obtain a defect probability map, the following operations are specifically performed: S51, up-sampling processing is performed on the input fused feature map, and the up-sampled image is jump-spliced with the basic feature map of the same scale to obtain a spliced feature map; S52, convolution refinement and spatial optimization processing are performed on the spliced feature map in sequence to obtain a decoded image; S53, the decoded image is repeatedly input to perform steps S51-S53 until a preset processing number is reached, to obtain a final reconstruction image; S54, interface enhancement and boundary refinement processing are performed on the final reconstruction image in sequence to obtain a defect probability map.

[0084] In the embodiment of the application, when the updating module 306 performs calculation of the loss function of the defect probability map and the artificially pre-labeled real defect segmentation map, and dynamically adjusts the loss weight of the initial model according to the calculation result to obtain a CT image defect segmentation model satisfying a preset convergence condition, the following operations are specifically performed: According to the defect probability map and the real defect segmentation map, a multiple loss used to represent the difference between the two images is calculated; wherein the multiple loss includes Dice loss, focal loss, asymmetric loss, boundary loss and Lovasz loss; The weight of each loss function is adjusted according to the training stage of the initial model, and the model total loss corresponding to each training stage is calculated according to the adjusted function weight; The gradient of the model total loss with respect to all weight parameters of the model network is calculated, and the network parameters are updated in the opposite direction of the gradient using a preset optimizer to reduce the total loss until the model satisfies the preset convergence condition, to obtain a CT image defect segmentation model for defect segmentation.

[0085] It should be noted that the above embodiment provides the intelligent segmentation device of the ceramic matrix composite CT image defect, and only the above each functional module is divided for example description, and in actual application, the above functions can be distributed by different functional modules to complete, that is, the internal structure of the device is divided into different functional modules to complete all or part of the functions described above. In addition, the intelligent segmentation device of the ceramic matrix composite CT image defect and the intelligent segmentation method of the ceramic matrix composite CT image defect provided in the above embodiment belong to the same concept, and the specific implementation process is described in the method embodiment, which will not be repeated here.

[0086] Embodiments of the present application also provide a computer device, which comprises a processor and a memory, the memory storing at least one instruction, at least one program, a code set or an instruction set, the at least one instruction, the at least one program, the code set or the instruction set being loaded and executed by the processor to implement the intelligent segmentation method of the ceramic matrix composite CT image defect provided in the above method embodiments. Figure 4

[0087] Embodiments of the present application also provide a computer readable storage medium, which stores at least one instruction, at least one program, a code set or an instruction set, the at least one instruction, the at least one program, the code set or the instruction set being loaded and executed by the processor to implement the intelligent segmentation method of the ceramic matrix composite CT image defect provided in the above method embodiments.

[0088] Embodiments of the present application also provide a computer program product, which comprises a computer program, the processor of the computer device reading the computer program from the computer readable storage medium, and the processor executing the computer program, so that the computer device executes the intelligent segmentation method of the ceramic matrix composite CT image defect described in any of the above embodiments.

[0089] For the convenience of description, the above system or device is described in various modules or units respectively described in functions. Of course, in the implementation of the present application, the functions of each unit can be implemented in the same or more software and / or hardware.

[0090] ​Those skilled in the art can clearly understand the application by the description of the above embodiments. The technical solutions of the application can be implemented by means of software and necessary universal hardware platforms. Based on such an understanding, the technical solutions of the application can be embodied in the form of a software product, which can be stored in a storage medium, such as a ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the application.

[0091] Finally, it should be noted that the terms such as first, second, third, and fourth, etc. are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between these entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the processes, methods, articles or devices including a series of elements not only include those elements, but also include other elements not explicitly listed or inherent to such processes, methods, articles or devices. Without more limitations, the element defined by the statement "including a" does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0092] The above description is only the preferred embodiments of the application, and it should be pointed out that those skilled in the art can make some improvements and refinements without departing from the principles of the application, and these improvements and refinements should also be regarded as the protection scope of the application.

Claims

1. An intelligent segmentation method of ceramic matrix composite (CT) image defects, characterized in that, The method comprises: inputting the preprocessed CT image data into a preset double-branch collaborative segmentation model to respectively output a local texture feature map for representing defect information from micro to macro and a global relationship feature map for representing global context correlation information of the image; wherein the double-branch collaborative segmentation model comprises a CNN branch and a Transformer branch; performing three-branch material perception fusion processing on the local texture feature map and the global relationship feature map according to a preset perception fusion model to obtain a fusion feature map containing local detail accuracy and global consistency; performing progressive decoding and fine reconstruction processing on the fusion feature map in sequence to obtain a defect probability map; wherein a pixel value of the defect probability map is used to represent a probability that each pixel in the image belongs to a defect; calculating a loss function of the defect probability map and a manually pre-labeled real defect segmentation map, and dynamically adjusting a loss weight of an initial model according to a calculation result to obtain a CT image defect segmentation model meeting a preset convergence condition.

2. The method of claim 1, wherein, The local texture feature map is obtained by the following method: S21, performing initial down-sampling on the CT image data to obtain a starting feature map; S22, inputting the starting feature map into a ConNeXt module based on a large-core depth convolution for processing to output a local feature map containing defect information; S23, performing down-sampling processing on the local feature map according to a convolution layer with a preset stride to halve the size of the feature map and double the number of channels; S24, re-inputting the down-sampled local feature map into the ConvNeXt module and repeatedly executing steps S22-S24 until a preset processing number is reached to obtain a local texture feature map of defect information at different scales.

3. The method of claim 1, wherein, The global relationship feature map is obtained by the following method: performing sequence conversion on the CT image data to obtain a starting feature sequence for Transformer branch processing; performing calculation and integration processing on the starting feature sequence according to a window self-attention mechanism and a displacement window self-attention mechanism to obtain a fusion feature sequence containing global context information; performing spatial reorganization and hierarchical down-sampling processing on the fusion feature sequence to obtain global relationship feature maps at different scales.

4. The method of claim 1, wherein, performing three-branch material perception fusion processing on the local texture feature map and the global relationship feature map according to a preset perception fusion model to obtain a fusion feature map containing material knowledge, comprising: aligning and preliminarily fusing all local texture feature maps and global relationship feature maps of the same scale to obtain basic feature maps at different scales; inputting all the basic feature maps into a fiber direction perception branch sub-model for identification to output a first feature map containing crack expansion information along different fiber directions; inputting all the basic feature maps into a defect scale perception branch sub-model for identification to output a second feature map containing crack defect size information; inputting all the basic feature maps into an interface boundary enhancement branch sub-model for identification to output a third feature map containing complete texture information on both sides of the fiber-matrix interface; The first feature map, the second feature map and the third feature map are fused according to preset branch contribution weights, to obtain a fused feature map with local detail accuracy and global consistency.

5. The method of claim 4, wherein, The progressive decoding and fine reconstruction processing are sequentially performed on the fused feature map to obtain a defect probability map, including: S51, upsampling processing is performed on the input fused feature map, and a jump splicing processing is performed on an image obtained by upsampling and the base feature map of the same scale to obtain a spliced feature map; S52, convolution refinement and spatial optimization processing are sequentially performed on the spliced feature map to obtain a decoding map; S53, the decoding map is repeatedly taken as an input to execute steps S51-S53 until a preset processing number is reached, to obtain a final reconstruction map; S54, interface enhancement and boundary refinement processing are sequentially performed on the final reconstruction map to obtain a defect probability map.

6. The method of claim 1, wherein, The loss function of the defect probability map and the artificially pre-labeled real defect segmentation map is calculated, and the initial model is dynamically adjusted in loss weight according to the calculation result, to obtain a CT image defect segmentation model meeting a preset convergence condition, including: According to the defect probability map and the real defect segmentation map, a multiple loss used to represent the difference between the two images is calculated; wherein the multiple loss includes a Dice loss, a focal loss, an asymmetric loss, a boundary loss and a Lovasz loss; The weight of each loss function is adjusted according to the training stage of the initial model, and the total loss of the model corresponding to each training stage is calculated according to the adjusted function weight; The gradient of the total loss of the model with respect to all weight parameters of the model network is calculated, and the network parameters are updated in the opposite direction of the gradient using a preset optimizer to reduce the total loss until the model meets the preset convergence condition, to obtain a CT image defect segmentation model for defect segmentation.

7. An intelligent segmentation device for defects in CT images of ceramic matrix composite materials, characterized in that, The device comprises: The detection module is configured to input the preprocessed CT image data into a preset double-branch collaborative segmentation model to respectively output a local texture feature map representing defect information from micro to macro and a global relationship feature map representing global context correlation information of the image; wherein the double-branch collaborative segmentation model comprises a CNN branch and a Transformer branch. The fusion module is configured to perform three-branch material perception fusion processing on the local texture feature map and the global relationship feature map according to a preset perception fusion model to obtain a fused feature map containing local detail accuracy and global consistency. The reconstruction module is configured to sequentially perform progressive decoding and fine reconstruction processing on the fused feature map to obtain a defect probability map; wherein the pixel value of the defect probability map is used to represent the probability of each pixel in the image belonging to a defect. The update module is configured to calculate the loss function of the defect probability map and the artificially pre-labeled real defect segmentation map, and dynamically adjust the loss weight of the initial model according to the calculation result to obtain a CT image defect segmentation model meeting a preset convergence condition.

8. A computer device, comprising: The computer device comprises a memory for storing a computer program and a processor for executing the computer program stored in the memory to implement the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, and the computer program is executed by the processor to implement the steps of the method according to any one of claims 1-6.

10. A computer program product, characterised in that, The computer program is executed by the processor to implement the steps of the method according to any one of claims 1-6.

Images