Three-phase segmentation method and system for lightweight resin-based rigid thermal insulation tile composites

Abstract

This invention relates to a three-phase segmentation method and system for lightweight resin-based rigid thermal insulation tile composite materials, belonging to the field of composite material structure analysis technology. It solves the problem that existing technologies cannot achieve accurate three-phase segmentation of lightweight resin-based rigid thermal insulation tile composite materials. The method includes: stretching the original measured XCT image to be segmented to obtain a corresponding stretched measured XCT image; segmenting the stretched measured XCT image using a trained three-phase segmentation model to obtain the three-phase segmentation result; wherein the three-phase segmentation model includes an encoder, channel attention branch, spatial attention branch, and decoder. This invention achieves accurate segmentation of the fiber, matrix, and interlayer pores of lightweight resin-based rigid thermal insulation tile composite materials.

Description

Technical Field

[0001] This invention relates to the field of composite material structure analysis technology, and in particular to a three-phase segmentation method and system for lightweight resin-based rigid heat-insulating tile composite materials. Background Technology

[0002] Lightweight resin-based rigid thermal insulation tile composite material is a special functional composite material that combines structural strength with thermal insulation and permeability. Its internal structure consists of three phases: fibers, lightweight resin matrix, and interlayer pores. As a high-performance structural and functional material, it is widely used in aerospace, high-end equipment and other fields. The accurate characterization of the material's internal structure (such as the distribution and morphology of fibers, matrix, and interlayer pores) is the core prerequisite for achieving numerical simulation of its key properties such as thermal conductivity and permeability. It directly determines the validity and reliability of the simulation results, and thus affects the design optimization and performance improvement of the material.

[0003] Currently, deep learning technology has been gradually applied to the XCT image segmentation of composite materials, becoming an important means of characterizing internal structures. However, there are significant differences in segmentation techniques for different types of composite materials, and all have limitations in adapting to lightweight resin-based rigid thermal insulation tile composites. For example, traditional threshold segmentation methods rely on the gray-level differences between phases, which completely fail due to the extremely low contrast between the matrix and interlayer pores. Segmentation models for conventional fiber-reinforced resin-based composites are suitable for scenarios with significant gray-level differences between fibers and the matrix, but are difficult to optimize for low-contrast scenarios and cannot effectively extract the boundary features of the matrix and interlayer pores, resulting in insufficient segmentation accuracy. Segmentation models for ceramic-based composites utilize the significant density difference between the ceramic matrix and air and rely on high contrast. For lightweight resin-based rigid thermal insulation tile composites with minimal gray-level differences between the matrix and interlayer pores, the mismatch in gray-level features leads to severe noise interference, large boundary segmentation errors, and a sharp decrease in segmentation accuracy, or even complete failure. Qualitative extraction of macroscopic cracks generated during the loading process can qualitatively study the damage and failure mechanisms of materials through crack morphology and distribution, but this relies on the high contrast and texture differences within the material.

[0004] However, while the fibers in lightweight resin-based rigid thermal insulation tile composites exhibit clear grayscale characteristics in XCT imaging and are easily identifiable, the extremely small density difference between the lightweight resin matrix and the interlayer pores results in very low grayscale contrast and blurred boundaries in XCT imaging. Furthermore, minor errors in the extraction of a few fibers in woven composites and in the qualitative identification of crack damage do not cause unacceptable errors in subsequent numerical simulations and damage mechanism conclusions. However, even a small error in the three-phase segmentation of lightweight resin-based rigid thermal insulation tile composites can lead to significant discrepancies in heat and mass transfer simulations, thus requiring more precise three-phase segmentation. Existing segmentation techniques cannot meet the need for accurate characterization of the three-phase structure of this material, thereby limiting its functional performance optimization and engineering applications.

[0005] Therefore, there is an urgent need for a three-phase segmentation method and system for lightweight resin-based rigid thermal insulation tile composites to solve the problems of insufficient segmentation accuracy under low contrast, serious noise interference, and inability to achieve accurate segmentation of the three-phase structure of lightweight resin-based rigid thermal insulation tile composites in existing technologies. This method would enable accurate segmentation of the three phases of fiber, matrix, and interlayer pores, providing reliable support for numerical simulation of key material properties. Summary of the Invention

[0006] Based on the above analysis, the present invention aims to provide a three-phase segmentation method and system for lightweight resin-based rigid thermal insulation tile composite materials, in order to solve the problem that the prior art cannot achieve accurate segmentation of the three-phase structure of lightweight resin-based rigid thermal insulation tile composite materials.

[0007] On one hand, embodiments of the present invention provide a three-phase segmentation method for lightweight resin-based rigid thermal insulation tile composite materials, comprising: The original measured XCT image to be segmented is stretched to obtain the corresponding stretched measured XCT image. The trained three-phase segmentation model is used to segment the measured XCT stretched image to obtain the corresponding three-phase segmentation result; wherein, the three-phase segmentation model includes an encoder, a channel attention branch, a spatial attention branch, a decoder and a classifier connected in sequence.

[0008] Furthermore, the trained three-phase segmentation model is used to segment the measured XCT stretching image to obtain the three-phase segmentation result, including: The encoder receives the measured XCT stretched image, downsamples the measured XCT stretched image, generates multiple basic feature maps, transmits one of the basic feature maps to the decoder, and transmits the other basic feature maps to the channel attention branch; The channel attention branch performs channel attention weighted fusion on the received basic feature map to generate a corresponding channel weighted feature map, and transmits it to the spatial attention branch; The spatial attention branch performs spatial attention weighted fusion on the received channel weighted feature map to generate a corresponding spatial weighted feature map, which is then transmitted to the decoder. The decoder upsamples the received basic feature map and each of the spatially weighted feature maps to generate a high-dimensional feature map, which is then transmitted to the classifier. The classifier performs three-phase classification on the high-dimensional feature map and generates the corresponding three-phase segmentation result.

[0009] Further, the trained three-phase segmentation model is obtained through the following steps: Three-phase annotation is performed on the original XCT image of each sample in the sample dataset to obtain the corresponding sample three-phase annotated image; A training dataset is constructed based on the original XCT images of each sample and the corresponding three-phase labeled images of the sample. The three-phase segmentation model is trained using the training dataset to obtain the trained three-phase segmentation model.

[0010] Furthermore, three-phase annotation is performed on each sample XCT raw image in the sample dataset to obtain the corresponding sample three-phase annotated image, including: Sliding sampling is performed on each of the original XCT images of the samples to obtain multiple sample XCT local images; Each of the sample XCT local images is contrast stretched, and all the stretched sample XCT local images are fused to obtain the sample XCT stretched image; wherein, the sample XCT stretched image includes the pixel grayscale value of each pixel; The sample XCT stretched image is annotated in three phases to obtain the corresponding sample three-phase annotated image.

[0011] Further, contrast stretching is performed on each of the sample XCT local images, including: Contrast stretching is achieved by updating the pixel grayscale value of each pixel in the sample XCT local image using the following formula: In the formula, The XCT local image of the sample represents the first... The updated pixel grayscale value corresponding to each of the aforementioned pixels. Indicates the first The pixel grayscale value before the update corresponds to each of the aforementioned pixels. , The minimum and maximum pixel gray values ​​are, in order, the local XCT images of the samples. , These represent the preset upper limit and lower limit of pixel grayscale values, respectively.

[0012] Furthermore, after constructing the training dataset based on the original XCT images of each sample and the corresponding three-phase labeled images of the sample, the dataset also includes: Perform binary transformation on any of the sample three-phase labeled images to generate the corresponding basic binary label image; Perform interlayer gap mutation operation on the basic binary label map to obtain multiple mutated binary label maps; Based on the basic binary label map and each of the variant binary label maps, the sample three-phase labeled image is labeled and replaced, and the corresponding sample XCT original image is pixel replaced to obtain the replaced sample three-phase labeled image and the sample XCT original image. All the replaced sample three-phase labeled images and the sample original XCT images are added to the training dataset.

[0013] Further, the three-phase segmentation model is trained using the training dataset to obtain the trained three-phase segmentation model, including: The original XCT images of the samples in the training dataset are input into the tri-phase segmentation model to obtain the corresponding tri-phase segmentation results of the samples. The training loss value is calculated based on the sample tri-phase segmentation results and the sample tri-phase labeled images; The model parameters of the three-phase segmentation model are updated according to the training loss value until the training stopping condition is met, thus obtaining the trained three-phase segmentation model.

[0014] Further, the training loss value is calculated based on the sample three-phase segmentation result and the sample three-phase labeled image, including: Based on the sample three-phase segmentation results and the sample three-phase labeled image, calculate the corresponding multiple single-phase Dice loss values; Calculate the full-map Dice loss value based on the single-phase Dice loss values ​​described above; Based on the sample three-phase segmentation results and the sample three-phase labeled image, calculate the corresponding multiple single-phase boundary loss values; Calculate the overall boundary loss value based on the single-phase boundary loss values ​​described above; The training loss value is calculated based on the full-graph Dice loss value and the full-graph boundary loss value.

[0015] Further, based on the sample three-phase segmentation results and the sample three-phase labeled image, multiple corresponding single-phase Dice loss values ​​are calculated, including: Based on the three-phase segmentation results of the sample, the probability value of each pixel in the three-phase segmentation results being the target phase is determined; wherein, the target phase is any one of the phases among fiber, matrix, and interlayer pores; Based on the sample three-phase labeled image, the target phase label value corresponding to each pixel of the sample three-phase labeled image is determined; wherein, each pixel of the sample three-phase segmentation result corresponds one-to-one with each pixel of the sample three-phase labeled image; The single-phase Dice loss value corresponding to the target is calculated based on the target phase probability value and the target phase label value corresponding to all the pixels.

[0016] On the other hand, embodiments of the present invention provide a three-phase segmentation system for a lightweight resin-based rigid thermal insulation tile composite material, comprising: The image stretching processing module is used to stretch the original measured XCT image to be segmented to obtain the corresponding stretched measured XCT image. The image segmentation module is used to segment the measured XCT stretched image using a trained three-phase segmentation model to obtain the corresponding three-phase segmentation result; wherein, the three-phase segmentation model includes an encoder, a channel attention branch, a spatial attention branch, a decoder, and a classifier connected in sequence.

[0017] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. First, the contrast of XCT images of lightweight resin-based rigid thermal insulation tile composite materials is increased through stretching. Then, a three-phase segmentation model consisting of an encoder, channel attention branch, spatial attention branch, decoder, and classifier is used for three-phase segmentation. The channel attention branch accurately identifies the importance of different feature channels, while the spatial attention branch focuses on local areas of the image. By cooperating with the channel attention branch and the spatial attention branch, dual optimization of channel selection and spatial localization is achieved. This breaks through the limitations of traditional segmentation methods that can only enhance features individually and cannot be specifically adapted to low-contrast blurred boundary scenes. As a result, the matrix with blurred boundaries and interlayer pores are accurately distinguished, the boundary segmentation error is significantly reduced, and the segmentation adaptability and accuracy of XCT images of lightweight resin-based rigid thermal insulation tile composite materials are improved.

[0018] 2. Downsampling is performed by the encoder to enhance the ability to capture the boundaries and texture features of the matrix and interlayer pores under low contrast. Channel information of the basic feature map is obtained through the channel attention branch, and channel attention weighted fusion is performed to focus on strengthening the channel weights corresponding to the features of the matrix and interlayer pore boundaries. Spatial attention weighted fusion is performed through the spatial attention branch to focus on the blurred boundary areas of the image, assigning high weights to boundary pixel positions and low weights to non-boundary redundant areas, further enhancing the feature response of the boundary areas. Upsampling is performed by the decoder to obtain high-dimensional feature maps, accurately distinguishing the blurred boundaries of the matrix and interlayer pores, and significantly reducing the boundary segmentation error.

[0019] 3. By performing morphological transformation on the three-phase labeled images of the samples, the interlayer pore morphology changes that may occur in the preparation and service of lightweight resin-based rigid heat insulation tile composite materials are accurately simulated. The feature of blurred boundaries is preserved, which ensures that the samples are highly matched with the actual three-phase structure of the materials and improves the sample richness of the training dataset.

[0020] 4. Based on the full-image Dice loss value and the full-image boundary loss value, the training loss value is calculated to specifically address the pain points of low contrast and blurred boundaries between the matrix and interlayer pores in XCT images of lightweight resin-based rigid thermal insulation tile composite materials. The two work together to solve the problems of sample imbalance and boundary blurring, respectively, ensuring the accuracy of model training.

[0021] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0022] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts. Figure 1 This is a schematic flowchart of a three-phase splitting method for a lightweight resin-based rigid thermal insulation tile composite material according to an embodiment of the present invention. Figure 2 This is a flowchart illustrating a three-phase splitting model in an embodiment of the present invention; Figure 3 This is a schematic diagram comparing the segmentation effects of a three-phase segmentation model in an embodiment of the present invention; Figure 4 This is a schematic diagram of the main modules of a three-phase segmentation system for a lightweight resin-based rigid thermal insulation tile composite material in an embodiment of the present invention. Detailed Implementation

[0023] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0024] A specific embodiment of the present invention discloses a three-phase segmentation method for a lightweight resin-based rigid thermal insulation tile composite material, such as... Figure 1 As shown, it includes: Step S1: Stretch the original measured XCT image to be segmented to obtain the corresponding stretched measured XCT image.

[0025] The original measured XCT image to be segmented is stretched to obtain the corresponding stretched measured XCT image. The original measured XCT image can be obtained by scanning the lightweight resin-based rigid thermal insulation tile composite material with the three-phase structure to be analyzed using X-ray computed tomography (CT), or by other existing technologies; the steps are limited here. In this embodiment, the original measured XCT image includes the original pixel grayscale value of each pixel, and the corresponding stretched measured XCT image includes the pixel grayscale value of each pixel after stretching.

[0026] In this embodiment, the measured XCT original image is stretched through steps S11-S12 to obtain the corresponding measured XCT stretched image.

[0027] Step S11: Perform sliding sampling on the measured XCT original image to obtain multiple measured XCT local images.

[0028] Multiple local XCT images are obtained by sliding sampling of the original measured XCT images, which is suitable for the uneven distribution of interlayer pores and significant local contrast differences in lightweight resin-based rigid thermal insulation tile composite materials. In this embodiment, sliding sampling is performed on the original measured XCT images according to a preset window size and a preset sliding step size to obtain multiple sample XCT local images. The preset window size and preset sliding step size can be set empirically, for example, the preset window size... The preset sliding step size is 64 pixels. In other embodiments, sliding sampling can be performed on the measured XCT original image according to the sampling coverage. For example, the sampling coverage requires that the size of the measured XCT local image does not exceed 5% of the size of the measured XCT original image, ensuring that the area of ​​local interlayer pores and matrix is ​​completely covered, while accurately capturing local low-contrast weak features. The overlap between two measured XCT local images is not less than 50%, avoiding segmentation discontinuities at the edge of the sliding window, ensuring the continuity of boundary area processing, and further guaranteeing the processing effect of boundary areas. Through sliding sampling, each pixel in the measured XCT original image is covered by at least one measured XCT local image, and except for edge areas, each pixel is covered by multiple measured XCT local images.

[0029] For example, measured raw XCT images The size is pixels, preset window size Pixels, preset sliding step size of 64 pixels, in actual XCT raw image measurement Slide 64 pixels from left to right and from top to bottom to get multiple [items / items]. Measured local XCT images of pixels Each pixel in the measured XCT raw image is covered by at least one measured XCT local image, and except for edge regions, each pixel is covered by four measured XCT local images. , The indexes are, in order, the window row index and the column index. This represents the original XCT image obtained during the first sampling in the sampling window. This indicates the measured original XCT image obtained after the sampling window is slid to the right by a preset sliding step size, and... and exist Pixel overlap area.

[0030] Furthermore, median filtering can be applied to the original measured XCT image first, followed by sliding sampling using the filtered original XCT image. This effectively suppresses Gaussian noise and isolated noise points, avoiding noise interference with feature extraction in low-contrast regions, while preserving the subtle boundary features of the matrix and interlayer pores. For example, using a 3×3 filter kernel to perform median filtering on the original measured XCT image can suppress 15%–20% of Gaussian noise in the original measured XCT image, while simultaneously increasing the average grayscale contrast of the matrix and interlayer pores by 3 times, providing high-quality input for subsequent segmentation.

[0031] Step S12: Perform contrast stretching on each measured XCT local image, and fuse all the stretched measured XCT local images to obtain a measured XCT stretched image.

[0032] Each measured XCT local image is contrast stretched, and all stretched measured XCT local images are fused to obtain a measured XCT stretched image, including steps S121-S122.

[0033] Step S121: Perform contrast stretching on each measured XCT local image.

[0034] In this embodiment, for each measured XCT local image, the formula is used. The pixel grayscale value of each pixel in the image is updated to achieve contrast stretching of each measured XCT local image, thereby effectively enhancing local low-contrast features. The measured XCT local image is shown. line, number The updated pixel grayscale value of the column. Represents the measured XCT local image of the first line, number The pixel grayscale value of the column before the pixel update. , The values ​​shown are, in order, the minimum and maximum pixel grayscale values ​​corresponding to the measured XCT local images. , These represent the preset upper limit and lower limit of pixel grayscale values, respectively, for example... , .

[0035] By independently performing contrast stretching on each measured XCT local image, local feature distortion caused by global stretching is avoided. At the same time, forced contrast stretching amplifies the grayscale difference between the matrix and interlayer pores, enhancing low-contrast features and alleviating boundary blurring. The minimum and maximum pixel grayscale values ​​of each measured XCT local image are used as stretching benchmarks to accurately capture the subtle grayscale differences between the matrix and interlayer pores in the local area, avoiding local feature distortion caused by global stretching. The stretching range is constrained by preset upper and lower limits of pixel grayscale values ​​to maximize the grayscale difference between the matrix and interlayer pores, enhance subtle features, and ensure that all measured XCT local images are stretched within a uniform range, ensuring the consistency of enhancement of each local image.

[0036] Step S122: Fuse all the measured XCT local images after stretching to obtain the measured XCT stretched image.

[0037] All stretched local XCT images are fused to obtain a stretched XCT image, which has the same size and corresponding pixel positions as the original XCT image. For each pixel position in the original XCT image, a fusion process is performed based on the pixel grayscale values ​​of at least one stretched local XCT image covering that pixel. The fused pixel grayscale value corresponding to that pixel is calculated. The fused pixel grayscale values ​​of all pixels constitute the stretched XCT image.

[0038] In this embodiment, the accumulator matrix and the counting matrix are used to perform local fusion of stretched measured XCT images.

[0039] First, initialize the accumulator matrix and the counting matrix. All elements in the matrices are initialized to 0, and the matrix size matches the measured XCT raw image. For example, the measured XCT raw image is... For pixels, the dimensions of the accumulator matrix and the counting matrix are initialized to... .

[0040] Secondly, the measured XCT local images after each stretching are superimposed onto the accumulator matrix and the counting matrix, and the values ​​of each element in the accumulator matrix and the counting matrix are updated. , ,in, This represents the row and column indices of the elements. Specifically, for each stretched measured XCT local image, at the position corresponding to the stretched measured XCT local image in the original measured XCT image, the pixel grayscale value of each pixel in the stretched measured XCT local image is superimposed on the accumulator matrix. This yields the accumulator matrix with updated element values; simultaneously, based on the position of each pixel in the stretched measured XCT local image, a counting matrix is ​​used. The element values ​​at corresponding positions are incremented by 1, assigning equal weight to each occurrence of each pixel within the overlapping region, ensuring the integrity and boundary continuity of the fused image. For example, based on sliding sampling, two stretched measured XCT local images... , The corresponding pixel grayscale values ​​are as follows: , Based on two image and accumulator matrices Counting matrix If the accumulator matrix is ​​superimposed, then the accumulator matrix is ​​superimposed. middle for , for The superimposed counting matrix middle For 1, The value is 2.

[0041] Finally, based on the values ​​of each element in the accumulator matrix and the counting matrix, the fused pixel grayscale value corresponding to each pixel is calculated. The fused pixel grayscale values ​​corresponding to all pixels constitute the measured XCT stretched image. In this embodiment, an averaging method is used to calculate the fused pixel grayscale value corresponding to each pixel, forming the measured XCT stretched image, further weakening noise interference, enhancing the grayscale difference between the matrix and interlayer pores, and clarifying blurred boundaries. Measured XCT Enhanced Image The pixel grayscale value corresponding to each pixel In the formula, This represents the measured XCT stretch image number 1. line, number The pixel grayscale value corresponding to the column pixel. The accumulator matrix that superimposes all measured XCT local images represents the first... line, number The value of the column element, This represents the first element of the counting matrix that superimposes all measured XCT local images. line, number The value of the column element.

[0042] By employing sliding sampling, local independent contrast stretching, and fusion processing, the original XCT image of the measured image is stretched to balance image enhancement and feature preservation. The gray value of each pixel is averaged after stretching based on the actual number of coverages to avoid distortion in the boundary areas. This ensures that the stretching effect of the entire image can adapt to the low contrast and blurred boundary characteristics of the material. This approach initially solves the problems of extremely low contrast between the matrix and interlayer pores, blurred boundaries, and background noise in lightweight resin-based rigid thermal insulation tile composite materials. It enhances the gray value difference between the matrix and interlayer pores, clarifies blurred boundaries, and suppresses noise.

[0043] Step S2: Use the trained three-phase segmentation model to segment the measured XCT stretched image to obtain the corresponding three-phase segmentation result; wherein, the three-phase segmentation model includes an encoder, a channel attention branch, a spatial attention branch, a decoder, and a classifier connected in sequence.

[0044] The trained three-phase segmentation model is used to segment the measured XCT stretched image to obtain the corresponding three-phase segmentation result. The three-phase segmentation model includes an encoder, a channel attention branch, a spatial attention branch, a decoder and a classifier connected in sequence. The segmentation process specifically includes steps S21-S25.

[0045] Step S21: The encoder receives the measured XCT stretched image, downsamples the measured XCT stretched image, generates multiple basic feature maps, transmits one of the basic feature maps to the decoder, and transmits the other basic feature maps to the channel attention branch.

[0046] The encoder receives a measured XCT stretched image, downsamples the image to generate multiple basic feature maps, transmits one basic feature map to the decoder, and transmits the other basic feature maps to the channel attention branch. In this embodiment, the encoder includes a first convolutional layer and a bottleneck layer. The basic feature map generated by the first convolutional layer is transmitted to the channel attention branch, and the basic feature map generated by the bottleneck layer is transmitted to the decoder.

[0047] Furthermore, the encoder includes multiple first convolutional layers and bottleneck layers. Each first convolutional layer performs convolution, activation, and max pooling on the received measured XCT stretched image or the basic feature map transmitted from the previous first convolutional layer to complete downsampling, generate a new basic feature map, and transmit it to the next first convolutional layer, until it reaches the bottleneck layer. The bottleneck layer performs downsampling on the received basic feature map transmitted from the previous first convolutional layer to generate a new basic feature map, which is then transmitted to the decoder. In this embodiment, the basic feature maps generated by each first convolutional layer are transmitted to the channel attention branch. In other embodiments, any one of the basic feature maps can be transmitted to the channel attention branch.

[0048] After multiple convolutional layers, activation, and max pooling operations complete the downsampling process, continuously reducing the spatial size of the measured XCT stretched image while gradually increasing the number of channels in the basic feature map. This enables the three-phase segmentation model to extract high-level abstract features from the measured XCT stretched image. Upon reaching the bottleneck layer, the spatial size of the basic feature map reaches its minimum, while the number of channels reaches its maximum. This allows the bottleneck layer to capture the most abstract and representative feature information from the measured XCT stretched image, highly condensing and integrating the overall features. Using the bottleneck layer as the endpoint of the encoder's feature extraction, its output feature information provides a crucial abstract feature foundation for the decoder's subsequent upsampling and feature recovery, helping the decoder better recover the image's spatial information and thus achieve accurate image segmentation.

[0049] Furthermore, the three-phase segmentation model in this embodiment is constructed based on the traditional Unet model. The multi-layer first convolutional layer of the encoder adopts the convolutional blocks of the VGG16 model. For example, the encoder includes four first convolutional layers, using the first four convolutional blocks (conv1~conv4) of the VGG16 model to replace the convolutional layers in the traditional Unet encoder. Four downsampling operations are performed through the four convolutional blocks (i.e., the four first convolutional layers) to access the bottleneck layer in the traditional Unet encoder. The number of convolutional kernels of the four convolutional blocks (conv1~conv4) are 64, 128, 256, and 512 respectively, which is suitable for downsampling and feature extraction of measured XCT stretched images with a scale of 512×512 pixels and 16-bit grayscale. The bottleneck layer performs a final compression and abstraction on the basic feature map of the lightweight resin-based rigid thermal insulation tile composite material, maximizing the preservation of the weak features of the matrix and interlayer pores. At the same time, the deep semantic features of the measured XCT stretch image compressed by the encoder are passed to the decoder for upsampling and restoration, ensuring the integrity of feature transmission.

[0050] By performing downsampling operations layer by layer through multiple first convolutional layers, the feature map size is gradually reduced and the feature abstraction is improved. This enhances the ability to capture the boundaries and texture features of the matrix and interlayer pores under low contrast, solving the problem of insufficient shallow feature extraction capability and inability to capture weak image boundary features in traditional Unet encoders. This provides reliable feature support for clear boundary segmentation and ensures that the model can accurately identify three-phase features in low contrast scenes.

[0051] Step S22: The channel attention branch performs channel attention weighted fusion on the received basic feature map to generate a corresponding channel weighted feature map, and transmits it to the spatial attention branch.

[0052] The channel attention branch performs channel attention weighted fusion on the received basic feature map to generate a corresponding channel weighted feature map, which is then transmitted to the spatial attention branch. In this embodiment, the channel attention branch includes a first global average pooling layer, a first global max pooling layer, a first connection layer, a first activation layer, and a first fusion layer.

[0053] Specifically, the first global average pooling layer and the first global max pooling layer perform pooling processing on the received basic feature map to obtain the corresponding pooling results (including channel global information corresponding to the basic feature map). These two pooling results are then input into the first connection layer (in this embodiment, a fully connected network layer, including hidden layers). After processing by the ReLU activation function of the first activation layer, a channel attention weight map is output through the sigmoid function. The first fusion layer multiplies the channel attention weight map with the received basic feature map channel by channel to obtain the corresponding channel-weighted feature map, which is then transmitted to the spatial attention branch. The dimension of the channel-weighted feature map is the same as the dimension of the received basic feature map.

[0054] For example, the dimension of the basic feature map is... Where H and W are the width and height of the base feature map, respectively, and C is the number of channels in the base feature map. The resulting channel attention weight map has the following dimensions: Each element represents the channel weight of the corresponding channel. Multiplying all elements of each channel in the basic feature map by the corresponding channel weight yields a dimension of... Channel-weighted feature map.

[0055] The channel attention branch accurately identifies the importance of different feature channels, focuses on strengthening the channel weights corresponding to the matrix and interlayer pore boundary features, suppresses the interference of meaningless background noise channels, and solves the problem of effective features being submerged by noise in low-contrast scenes.

[0056] Step S23: The spatial attention branch performs spatial attention weighted fusion on the received channel weighted feature map to generate a corresponding spatial weighted feature map, and transmits it to the decoder.

[0057] The spatial attention branch performs spatial attention weighted fusion on the received channel weighted feature maps to generate corresponding spatial weighted feature maps, which are then transmitted to the decoder. In this embodiment, the spatial attention branch includes a second global average pooling layer, a second global max pooling layer, a second connection layer, a second activation layer, and a second fusion layer.

[0058] Specifically, the second global average pooling layer performs channel-dimensional global average pooling on the received channel-weighted feature map to obtain a first single-channel feature map. The second global max pooling layer performs channel-dimensional global max pooling on the received channel-weighted feature map to obtain a second single-channel feature map. The second connection layer concatenates the first and second single-channel feature maps along their channel dimensions to obtain a dual-channel feature map. The second activation layer performs convolution on the dual-channel feature map and outputs a spatial attention weight map using the sigmoid function. The second fusion layer multiplies the spatial attention weight map element-wise with the received channel-weighted feature map to obtain the corresponding spatial weighted feature map, which is then transmitted to the decoder. The dimension of the spatial weighted feature map is the same as the dimension of the received channel-weighted feature map.

[0059] For example, the channel-weighted feature map is as follows: Then the dimensions corresponding to the first single-channel feature map and the second single-channel feature map are: The dimension corresponding to the dual-channel feature map is After convolution with a 7×7 kernel, the spatial attention weight map is output using the sigmoid function (corresponding to a dimension of ). The spatial attention weight map represents the spatial weight corresponding to each spatial point. Different channels and the same spatial point have the same spatial weight. Each element of each channel in the channel-weighted feature map is multiplied by its corresponding spatial weight, i.e., the weight of each channel in the channel-weighted feature map is calculated. Each element and the spatial attention weight map Multiplying each element one by one, we get a dimension of... Spatial weighted feature map.

[0060] The spatial attention branch focuses on local regions of the channel-weighted feature map, accurately locating the blurred boundary between the matrix and interlayer pores, further enhancing the feature response of the boundary region and weakening redundant features in non-boundary regions. By cooperating with the channel attention branch and the spatial attention branch, dual weighting and optimization of channel filtering and spatial localization are achieved, adapting to the feature distribution of low-contrast XCT images. This breaks through the limitation of traditional attention modules that can only enhance single features and cannot specifically adapt to low-contrast blurred boundary scenes, thereby accurately distinguishing the blurred boundary of the matrix and interlayer pores, significantly reducing boundary segmentation errors, and improving the segmentation adaptability and accuracy of the three-phase segmentation model for measured XCT tensile images of lightweight resin-based rigid thermal insulation tile composites.

[0061] Step S24: The decoder upsamples the received basic feature map and each of the spatially weighted feature maps to generate a high-dimensional feature map, and transmits it to the classifier.

[0062] The decoder upsamples the received base feature map and each spatially weighted feature map to generate a high-dimensional feature map, which is then transmitted to the classifier. In this embodiment, the decoder includes a second convolutional layer that upsamples the received base feature map and each spatially weighted feature map to generate a high-dimensional feature map, which is then transmitted to the classifier.

[0063] Furthermore, the decoder includes multiple layers of second convolutional layers. Each second convolutional layer performs transposed convolution on the received base feature map or the high-dimensional feature map transmitted from the next second convolutional layer, reducing the channel dimension of the base feature map or the high-dimensional feature map and increasing the image size. This ensures that the base feature map or the high-dimensional feature map after reducing the channel dimension maintains the same channel dimension as the received spatially weighted feature map, and the channel dimensions are concatenated. Then, the concatenated feature map is convolved to generate a new high-dimensional feature map, completing the upsampling operation. The generated high-dimensional feature map is then transmitted to the next second convolutional layer, and so on, until it reaches the top second convolutional layer, generating a new high-dimensional feature map, which is then transmitted to the classifier. In this embodiment, layer-by-layer upsampling through multiple second convolutional layers yields a high-dimensional feature map with high-resolution semantic features. The high-dimensional feature map generated by the top second convolutional layer is transmitted to the classifier, and the size of the high-dimensional feature map generated by the top second convolutional layer is the same as the size of the measured XCT stretched image.

[0064] For example, the decoder includes four second convolutional layers. The bottom second convolutional layer (denoted as second convolutional layer D) concatenates and convolves the received base feature map and spatially weighted feature map to perform upsampling and obtain a high-dimensional feature map, which is then transmitted to the next second convolutional layer (denoted as second convolutional layer C). Second convolutional layer C receives the spatially weighted feature map and the high-dimensional feature map generated by second convolutional layer D, performs upsampling to obtain a new high-dimensional feature map, which is then transmitted to the next second convolutional layer (denoted as second convolutional layer B). Second convolutional layer B, referring to second convolutional layer C, generates a new high-dimensional feature map and transmits it to the next second convolutional layer (denoted as second convolutional layer A). Second convolutional layer A, referring to second convolutional layer C, generates a new high-dimensional feature map and transmits it to the classifier.

[0065] Step S25: The classifier performs three-phase classification on the high-dimensional feature map and generates the corresponding three-phase segmentation result.

[0066] The classifier performs three-phase classification on the high-dimensional feature map and generates the corresponding three-phase segmentation result. In this embodiment, the classifier includes a convolutional classification layer and a third activation layer.

[0067] Specifically, the convolutional classification layer performs convolution operations on the received high-dimensional feature map, mapping it into three channels corresponding to fibers, matrix, and interlayer pores, respectively. The size of the mapped image is consistent with the size of the measured XCT stretched image. In this embodiment, the convolutional classification layer uses a 1×1 convolution kernel for convolution operations. The third activation layer is activated by the Softmax function to obtain the probability of each pixel belonging to each phase. For each pixel, the phase corresponding to the channel with the highest probability is selected as the single-phase category of that pixel. Based on the single-phase categories corresponding to all pixels, a three-phase segmentation result is constructed. For example, if the three channels sequentially represent fibers, matrix, and interlayer pores, and for a certain pixel, the corresponding three probabilities are 0.80, 0.15, and 0.05 respectively, then the single-phase category of that pixel is determined to be fiber.

[0068] Furthermore, combined Figure 2 Another implementation of the encoder, channel attention branch, spatial attention branch, decoder, and classifier is described below: Each first convolutional layer is connected to a channel attention branch, each channel attention branch is connected to a spatial attention branch, and each spatial attention branch is connected to a second convolutional layer. For example, a CBAM attention module (channel attention branch and spatial attention branch) can be built in the PyTorch framework and embedded at the jump connection points of each layer of the encoder and decoder. Through the collaboration of the two branches, the segmentation difficulty of the matrix and interlayer pores of the lightweight resin-based rigid heat insulation tile composite material XCT image is solved, which has extremely low contrast, weak boundary features, and is easily masked by noise.

[0069] Specifically, each first convolutional layer receives the measured XCT stretched image or the basic feature map generated by the previous first convolutional layer, generates a new basic feature map, and transmits it to the next convolutional layer and the channel attention branch. Each convolutional layer repeats the sampling and transmission until it reaches the bottleneck layer. In this embodiment, the encoder includes four first convolutional layers with corresponding output channel numbers of 64, 128, 256, and 512, respectively. A 3×3 convolutional kernel and ReLU function can be used for convolution and activation. The first two first convolutional layers use two convolutions, and the last two use three convolutions. Then, 2×2 max pooling is used to generate the basic feature map, which is input to the next first convolutional layer. The channel dimension of the basic feature map increases after the first convolution, but remains unchanged after subsequent convolutions. The channel dimension of the basic feature map after max pooling also remains unchanged, resulting in a 50% image size compression.

[0070] The bottleneck layer receives the base feature map generated by the previous convolutional layer, downsamples it to generate a new base feature map, and transmits it to the decoder. In this embodiment, the bottleneck layer has 1024 output channels.

[0071] Each channel attention branch performs channel attention weighted fusion on the received base feature map to generate a channel-weighted feature map, which is then transmitted to the connected spatial attention branch. It can be understood that the hidden layer dimension of the corresponding spatial attention branch differs depending on the number of channels in the first convolutional layer. For example, if the hidden layer dimension is 1 / 8 of the number of input channels, such as conv1 outputting 64 channels, then the hidden layer dimension of the corresponding spatial attention branch is 8; conv2 outputting 128 channels, then the hidden layer dimension of the corresponding spatial attention branch is 16. The dimension of the channel-weighted feature map generated by each channel attention branch is the same as the dimension of the base feature map received by that channel attention branch.

[0072] Each spatial attention branch performs spatial attention weighted fusion on the received channel-weighted feature maps to generate a corresponding spatial weighted feature map, which is then transmitted to the decoder. The dimension of the spatial weighted feature map generated by each spatial attention branch is the same as the dimension of the channel-weighted feature map received by that spatial attention branch.

[0073] Each second convolutional layer receives the spatially weighted feature map transmitted from the spatial attention branch, the basic feature map generated by the bottleneck layer, or the high-dimensional feature map generated by the next second convolutional layer. It then generates a new high-dimensional feature map, which is transmitted to the previous second convolutional layer, and so on, until the topmost second convolutional layer generates a new high-dimensional feature map, which serves as the final high-dimensional feature map output by the decoder. Specifically, each second convolutional layer first performs a transposed convolution on the received basic or high-dimensional feature map to reduce its channel dimension and increase the image size. This ensures that the reduced-channel-dimensional basic or high-dimensional feature map maintains the same channel dimension as the received spatially weighted feature map. The concatenated feature map is then convolved to generate a new high-dimensional feature map, completing the upsampling operation. In this embodiment, the decoder includes four second convolutional layers with corresponding input channel numbers of 1024, 512, 256, and 128, respectively. Each second convolutional layer employs one transposed convolution and two convolutions. The transposed convolution uses a 2×2 kernel, and the concatenated feature map is convolved using a 3×3 kernel and activated by the ReLU function. Taking the bottommost second convolutional layer as an example, it receives a 1024-channel basic feature map and a 512-channel spatially weighted feature map. First, the channel dimension of the basic feature map is reduced to 512 through transposed convolution, while the size of the basic feature map after channel dimension reduction is doubled. The basic feature map (512-channel) after channel dimension reduction is concatenated with the spatially weighted feature map (512-channel) to generate a 1024-channel concatenated feature map. The concatenated feature map is then subjected to convolutional dimensionality reduction to obtain a 512-channel high-dimensional feature map. This completes the upsampling operation and is output to the next second convolutional layer.

[0074] The classifier's convolutional classification layer performs a 1×1 convolution operation on the received high-dimensional feature map, mapping the high-dimensional feature map into 3 channels to obtain the three-phase segmentation result.

[0075] In this embodiment, the trained three-phase segmentation model is obtained through steps A-C.

[0076] Step A: Perform three-phase annotation on each sample XCT original image in the sample dataset to obtain the corresponding sample three-phase annotated image.

[0077] Three-phase annotation is performed on the original XCT image of each sample in the sample dataset to obtain the corresponding sample three-phase annotated image.

[0078] First, we construct the sample dataset.

[0079] In this embodiment, the construction of the sample dataset includes: using an X-ray micro-CT device, acquiring raw CT data of a certain type of lightweight resin-based rigid thermal insulation tile composite material (size: 3mm×3mm×5mm), a total of 1000 consecutive slice CT images were acquired, and each slice CT image was used as the sample XCT raw image to construct the sample dataset. All sample XCT raw images are 16-bit grayscale images, with pixel grayscale values ​​fixed in the range of [0, 65535]. The size of a single image is uniformly 512×512 pixels, and the pixel grayscale values ​​are distributed in the range of 0-65535, completely covering different depths and different three-phase structure regions of the thermal insulation tile, ensuring the integrity and representativeness of the raw data.

[0080] Furthermore, to balance data representativeness and processing efficiency, a fixed-interval sampling method is adopted to sample the original XCT images of the sample dataset and update the sample training set. For example, one image is extracted every three images from 1000 original XCT images. The extracted original XCT images form a new sample training set for subsequent stretching, annotation, and model training. The extracted original XCT images cover the full depth range of the 1000 original data, avoiding the concentration of samples in a local area and adapting to the three-phase structure distribution characteristics of the 3mm×3mm×5mm sample.

[0081] Secondly, three-phase annotation is performed on the original XCT image of each sample to obtain the corresponding sample three-phase annotated image.

[0082] In this embodiment, each sample's original XCT image is manually annotated with three phases: fiber, matrix, and interlayer pores, resulting in a corresponding sample three-phase annotated image. This ensures that the annotations in areas with blurred boundaries are refined, guaranteeing consistency between the sample three-phase annotated image and the actual three-phase structure of the lightweight resin-based rigid thermal insulation tile composite material. The sample three-phase annotated image includes the current annotated phase corresponding to each pixel; this can be understood as the current annotated phase being one of the phases among fiber, matrix, and interlayer pores.

[0083] Furthermore, to enhance the difference in pixel grayscale values ​​between pixels in the original XCT images of the samples, each original XCT image of the sample dataset is stretched to obtain the corresponding stretched XCT image. The stretched XCT image is then annotated in three phases to obtain the corresponding annotated three-phase image. This process includes steps A1-A3.

[0084] Step A1: Perform sliding sampling on each sample XCT original image to obtain multiple sample XCT local images.

[0085] Sliding sampling is performed on each sample XCT original image to obtain multiple sample XCT local images, which is adapted to the characteristics of uneven interlayer pore distribution and significant local contrast differences in lightweight resin-based rigid thermal insulation tile composite materials. In this embodiment, the sample XCT original images and the measured XCT original images adopt the same sliding sampling method, and the principle is the same as step S11, which will not be repeated here.

[0086] Step A2: Perform contrast stretching on each of the sample XCT local images, and fuse all the stretched sample XCT local images to obtain the sample XCT stretched image; wherein, the sample XCT stretched image includes the pixel grayscale value of each pixel.

[0087] Contrast stretching is performed on the local XCT images corresponding to each sample's original XCT image. Then, all the stretched local XCT images corresponding to the original XCT image of that sample are fused together to obtain the stretched XCT image of the sample.

[0088] First, contrast stretching is performed on the local XCT image of each sample.

[0089] In this embodiment, contrast stretching is performed on each sample XCT local image, including updating the pixel grayscale value of each pixel in the sample XCT local image using the following formula to achieve contrast stretching: In the formula, Represents the sample XCT local image of the first The updated pixel grayscale value corresponding to each pixel. Indicates the first The corresponding pixel grayscale value before the update. , The values ​​shown are, in order, the minimum and maximum pixel grayscale values ​​corresponding to the local XCT images of the samples. , These represent the preset upper limit and lower limit of pixel grayscale values, respectively. In this embodiment... , To maintain consistency between the measured XCT local images and the sample XCT local images after stretching, contrast stretching was used to increase the average grayscale difference of local low-contrast areas (original grayscale difference <50) in the 16-bit image to 200~300, significantly alleviating the boundary blurring problem and enhancing the characteristic differences between the matrix and interlayer pores.

[0090] Secondly, the stretched local XCT images of all samples corresponding to the original XCT images of each sample are fused to obtain the corresponding stretched XCT images of the samples.

[0091] The stretched XCT local images of all samples corresponding to the original XCT image of each sample are fused to obtain the stretched XCT image of the sample corresponding to the original XCT image of that sample. The principle is the same as step S122, and will not be repeated here.

[0092] By stretching the original XCT image of each sample, a corresponding stretched XCT image of the sample is obtained, covering the full depth range of the original data.

[0093] Step A3: Perform three-phase annotation on the XCT stretched image of the sample to obtain the corresponding three-phase annotated image of the sample.

[0094] Three-phase annotation is performed on the XCT tensile images of the samples to obtain the corresponding sample three-phase annotated images. In this embodiment, each sample XCT tensile image is manually annotated with three phases: fiber, matrix, and interlayer pores, to obtain the corresponding sample three-phase annotated images. This ensures that the annotations in areas with blurred boundaries are refined, and that the sample three-phase annotated images are consistent with the actual three-phase structure of the lightweight resin-based rigid thermal insulation tile composite material. The sample three-phase annotated image includes the current annotated phase corresponding to each pixel. It can be understood that the current annotated phase is one of the phases among fiber, matrix, and interlayer pores.

[0095] Step B: Construct a training dataset based on the original XCT images of each sample and the corresponding three-phase labeled images of the sample.

[0096] Based on the original XCT image of each sample and the corresponding three-phase labeled image of the sample, a sample pair is formed by each original XCT image of the sample and the corresponding three-phase labeled image of the sample to construct a training dataset for training the three-phase segmentation model.

[0097] Furthermore, the original XCT images of each sample and the corresponding three-phase labeled images of the sample can be randomly flipped (e.g., horizontally or vertically), and the flipped original XCT images of the sample and the corresponding three-phase labeled images of the sample can be added to the training dataset to increase sample diversity.

[0098] It is understandable that if step A stretches the original XCT images of each sample to obtain stretched XCT images and corresponding three-phase labeled images, then a training dataset can be constructed based on these stretched XCT images and corresponding three-phase labeled images, forming a sample pair for each sample. This dataset is then used to train the three-phase segmentation model. Similarly, the stretched XCT images and corresponding three-phase labeled images of each sample can be randomly flipped (e.g., horizontally or vertically), and the flipped stretched XCT images and corresponding three-phase labeled images can be added to the training dataset.

[0099] Furthermore, to better improve the sample diversity of the training dataset, each sample XCT original image and its corresponding sample three-phase labeled image are used to form a sample pair. After constructing the training dataset, morphological transformation of the sample three-phase labeled images is also performed, and the training dataset is updated based on the corresponding sample XCT original images.

[0100] To address the issues of sample imbalance and insufficient sample size caused by the diverse morphologies and blurred boundaries of interlayer pores in lightweight resin-based rigid thermal insulation tile composites, a data augmentation method based on morphological transformation for interlayer pore structure variation is adopted. This method simulates the morphological characteristics of different stages of interlayer pore evolution that may occur during preparation and use by performing morphological transformations on the three-phase labeled images of the samples. This supplements the training samples in low-contrast, blurred-boundary scenarios (i.e., the transformed three-phase labeled images of the samples and the corresponding transformed original XCT images of the samples), updates the training dataset, and improves sample richness. This enables the trained three-phase segmentation model to have a stronger generalization ability for the morphological changes of interlayer pores in lightweight resin-based rigid thermal insulation tile composites. Specifically, steps B1-B4 are included.

[0101] Step B1: Perform binary transformation on any of the sample three-phase labeled images to generate the corresponding basic binary label image.

[0102] Binary transformation is performed on any of the sample three-phase labeled images to generate a corresponding basic binary label image. In this embodiment, interlayer pores are taken as the phase of interest. For each sample three-phase labeled image, the pixel grayscale value corresponding to the pixel whose current labeled phase is interlayer pores is converted to 1, and the pixel grayscale value corresponding to the pixel whose current labeled phase is matrix or fiber is converted to 0, thereby generating a corresponding basic binary label image.

[0103] Step B2: Perform interlayer gap mutation operation on the basic binary label map to obtain multiple mutated binary label maps.

[0104] For each basic binary label map, interlayer porosity mutation operations are performed to obtain multiple mutated binary label maps. In this embodiment, dilation operations are used to simulate the expansion or connectivity of interlayer porosity, generating multiple mutated binary label maps with structural variations. For example, dilation operations are performed using structural elements of three sizes: 3×3, 5×5, and 7×7, generating multiple mutated binary label maps with structural variations, covering different forms such as interlayer porosity expansion and connectivity. The dilation operation can employ existing techniques; the size of the structural elements affects the scope of the dilation operation. In other embodiments, erosion operations can also be performed to simulate the contraction or disappearance of interlayer porosity.

[0105] Opening operations simulate the shrinkage or disappearance of interlayer porosity, while closing operations simulate the expansion or connectivity of interlayer porosity, generating variant binary label maps of multiple structural variations. For example, using structural elements of three sizes—3×3, 5×5, and 7×7—opening operations (simulating interlayer porosity shrinkage and disconnection) and closing operations (simulating interlayer porosity expansion and connectivity) are performed respectively, generating variant binary label maps of multiple structural variations covering different forms of interlayer porosity shrinkage, expansion, and connectivity. The opening and closing operations can employ existing techniques; the size of the structural elements affects the execution range of the opening and closing operations.

[0106] By using opening and closing operations, the interlayer pore morphology changes that may occur in the preparation and service of lightweight resin-based rigid thermal insulation tile composite materials are accurately simulated, and the boundary ambiguity features are preserved. This ensures that the subsequently obtained transformed three-phase labeled images of the samples and the corresponding transformed original XCT images of the samples are highly matched with the actual three-phase structure of the materials, thus avoiding the disconnect between new samples and actual application scenarios.

[0107] Step B3: Based on the basic binary label image and each of the variant binary label images, perform label replacement on the sample three-phase labeled image and pixel replacement on the corresponding sample XCT original image to obtain the replaced sample three-phase labeled image and the sample XCT original image.

[0108] Based on each basic binary label image and the corresponding variant binary label images, the label replacement is performed on the sample three-phase labeled image corresponding to the basic binary label image, and the pixel replacement is performed on the corresponding sample XCT original image to obtain the replaced sample three-phase labeled image and sample XCT original image.

[0109] In this embodiment, the annotation is replaced through steps B31-B32.

[0110] Step B31: For each basic binary label image, compare it pixel by pixel with the corresponding variant binary label images to determine the variant pore locations corresponding to each variant binary label image.

[0111] For each basic binary label image, a pixel-by-pixel comparison is performed with the corresponding variant binary label images to determine the variant pore locations corresponding to each variant binary label image. In this embodiment, the variant pore locations include newly added pore locations. For example, by comparing the values ​​of each pixel in the basic binary label image with the values ​​of the corresponding pixels in the variant binary label image, if the value in the basic binary label image is 0, while the value at the corresponding pixel location in the variant binary label image is 1, then that pixel location is the newly added pore location corresponding to the variant binary label image.

[0112] In other embodiments, if an erosion operation is performed, the variant porosity locations also include disappearing porosity locations. For example, if the value in the basic binary label image is 1, and the value of the corresponding pixel position in the variant binary label image is 0, then that pixel position is a disappearing porosity location.

[0113] Step B32: For the sample three-phase labeled image corresponding to the basic binary label image, based on the mutation gap point corresponding to each mutation binary label image, update the current labeled phase corresponding to each pixel in the sample three-phase labeled image, and generate multiple replaced sample three-phase labeled images.

[0114] In this embodiment, the variant void locations include newly added void locations. For the sample three-phase labeled image corresponding to the basic binary label image, based on the newly added void locations corresponding to each variant binary label image, the current labeled phase corresponding to the newly added void locations in the sample three-phase labeled image is replaced with interlayer voids, while the current labeled phases of other pixels remain unchanged, thus obtaining a replaced sample three-phase labeled image. It can be understood that a replaced sample three-phase labeled image can be obtained based on each variant binary label image.

[0115] In other embodiments, if the variable pore locations also include disappearing pore locations, then the current labeled phase corresponding to the disappearing pore location is replaced with fiber or matrix. Alternatively, the updated current labeled phase of the disappearing pore location can be determined based on the current labeled phase of the pixels adjacent to the disappearing pore location. For example, if the matrix accounts for a higher proportion of the current labeled phases corresponding to multiple pixels adjacent to the disappearing pore location, then the current labeled phase corresponding to the disappearing pore location is replaced with matrix.

[0116] In this embodiment, pixel replacement is performed through steps B33-B34.

[0117] Step B33: For each basic binary label image, compare it pixel by pixel with the corresponding variant binary label images to determine the variant pore locations corresponding to each variant binary label image.

[0118] The principle is the same as step B31, and will not be repeated here.

[0119] Step B34: For the sample XCT stretched image corresponding to the basic binary label image, based on the mutation gap point corresponding to each mutation binary label image, update the pixel gray value corresponding to each pixel in the sample XCT stretched image to generate multiple replaced sample XCT stretched images.

[0120] In this embodiment, the variant void locations include newly added void locations. For the sample XCT stretched image corresponding to the base binary label map, based on the newly added void locations corresponding to each variant binary label map, the pixel grayscale value corresponding to the newly added void location in the sample XCT stretched image is replaced with the average grayscale value of the interlayer voids, while the pixel grayscale values ​​of other pixels remain unchanged, thus obtaining a replaced sample XCT stretched image. It can be understood that a replaced sample XCT stretched image can be obtained based on each variant binary label map.

[0121] The average gray value of interlayer pores is determined based on all sample XCT stretched images and sample three-phase labeled images of the training dataset determined in step C. Based on the sample three-phase labeled images and the corresponding sample XCT stretched images, all pixels whose current labeled phase is an interlayer pore are identified, and the average gray value of these pixels is used as the average gray value of the interlayer pores. In other embodiments, it can also be determined based on the sample XCT stretched images and sample three-phase labeled images corresponding to the basic binary label image, using the average gray value of all pixels corresponding to interlayer pores in the sample XCT stretched image as the average gray value of the interlayer pores. Alternatively, the average gray value of the interlayer pores can be set empirically, for example, 25334.

[0122] By replacing the pixel grayscale values ​​corresponding to the newly added gap points with the average grayscale values ​​of the interlayer pores, the XCT stretched images of the replaced samples are guaranteed to match the material properties and maintain the core characteristics of the material, namely "extremely low contrast between the matrix and the interlayer pores and blurred boundaries". This avoids the appearance of false samples with high contrast and clear boundaries after the mutation replacement, ensuring that the training samples can accurately adapt to the model training requirements and improve the model's adaptability to low-contrast scenes.

[0123] In other embodiments, if the variable porosity points also include disappearing porosity points, then based on the updated current labeled phase corresponding to the disappearing porosity point, the pixel grayscale value corresponding to the disappearing porosity point is replaced with the fiber grayscale average value or the matrix grayscale average value. For example, if the updated current labeled phase for the disappearing porosity point is fiber, then the corresponding pixel grayscale value is replaced with the fiber grayscale average value. The determination of the fiber grayscale average value and the basic grayscale average value can refer to the method for determining the interlayer porosity grayscale average value, and will not be elaborated here.

[0124] Step B4: Add all the replaced sample three-phase labeled images and sample XCT stretched images to the training dataset.

[0125] The replaced sample three-phase labeled image and sample XCT stretched image corresponding to each mutated binary label image are added to the training dataset as a new sample pair to expand the training dataset.

[0126] By adjusting the scale and operation type (dilation operation, erosion operation) of the structural elements of morphological variation, a series of new sample pairs simulating different stages of interlayer porosity evolution can be generated, effectively expanding the sample size of the training dataset, supplementing samples in low-contrast and blurred boundary scenes, solving the problem of low boundary segmentation accuracy caused by insufficient samples and single features, and significantly improving the generalization ability and robustness of the model.

[0127] Understandably, if the original XCT images of each sample are stretched, and a training dataset is constructed based on the stretched XCT images of each sample and the corresponding sample three-phase labeled images, then a basic binary label image is generated by performing a binary transformation on the sample three-phase labeled image corresponding to any sample XCT stretched image. Interlayer gap mutation is then performed on the basic binary label image to obtain multiple mutated binary label images. Based on the basic binary label image and each mutated binary label image, the label replacement is performed on the sample three-phase labeled image corresponding to the sample XCT stretched image, and pixel replacement is performed on the sample XCT stretched image to obtain the replaced sample three-phase labeled image and sample XCT stretched image. All replaced sample three-phase labeled images and sample XCT stretched images are added to the training dataset for training the three-phase segmentation model.

[0128] Step C: Train the three-phase segmentation model using the training dataset to obtain the trained three-phase segmentation model.

[0129] The three-phase segmentation model is trained using the training dataset to obtain the trained three-phase segmentation model, specifically including steps C1-C3.

[0130] Step C1: Input the sample XCT stretched image from the training dataset into the three-phase segmentation model to obtain the corresponding sample three-phase segmentation result.

[0131] The sample XCT stretched images from the training dataset are input into the three-phase segmentation model to obtain the corresponding sample three-phase segmentation results. In this embodiment, the structure of the three-phase segmentation model remains unchanged before and after training. Inputting the sample XCT stretched images into the three-phase segmentation model to obtain the corresponding sample three-phase segmentation results is the same as the principle of step S2, and will not be repeated here.

[0132] Step C2: Calculate the training loss value based on the sample three-phase segmentation result and the sample three-phase labeled image.

[0133] The training loss value is calculated based on the sample three-phase segmentation results and the sample three-phase labeled images to optimize the model parameters of the three-phase segmentation model. In this embodiment, the training dataset is divided into a training subset and a validation subset. Each subset contains samples with different interlayer pore morphologies and different degrees of boundary blurring to meet the needs of model training and validation. The parameters are adjusted in real time through the validation subset to optimize the segmentation effect of low-contrast regions and boundaries, thus obtaining the trained three-phase segmentation model. Step C2 includes steps C21-C25.

[0134] Step C21: Based on the sample three-phase segmentation results and the sample three-phase labeled image, calculate the corresponding multiple single-phase Dice loss values.

[0135] Based on the sample three-phase segmentation results and the sample three-phase labeled images, the corresponding single-phase Dice loss value is calculated for each phase, specifically including steps C211-C213.

[0136] Step C211: Based on the three-phase segmentation result of the sample, determine the target phase probability value of each pixel in the three-phase segmentation result of the sample as the target phase; wherein, the target phase is any one of the phases of fiber, matrix, and interlayer pores.

[0137] Based on the three-phase segmentation results of the samples, the probability value of each pixel in the three-phase segmentation results belonging to the target phase is determined. For each pixel, if the probability of the pixel belonging to the target phase obtained after passing through the third activation layer of the three-phase segmentation model is the highest, then this probability is retained as the corresponding target probability value; otherwise, it is 0. For example, if three channels are set to represent fibers, matrix, and interlayer pores in sequence, the three probabilities corresponding to pixel A are 0.80, 0.15, and 0.05, respectively, and the three probabilities corresponding to pixel B are 0.10, 0.15, and 0.75, respectively. If the target phase is fiber, then the target phase probability value corresponding to pixel A is 0.80, and the target phase probability value corresponding to pixel B is 0.

[0138] Step C212: Based on the sample three-phase labeled image, determine the target phase label value corresponding to each pixel of the sample three-phase labeled image; wherein, each pixel of the sample three-phase segmentation result corresponds one-to-one with each pixel of the sample three-phase labeled image.

[0139] Based on the sample three-phase labeled images, determine the target phase label value corresponding to each pixel in the sample three-phase labeled images. Taking fiber as the target phase as an example, for any sample three-phase labeled image, if the current labeled phase corresponding to a certain pixel is fiber, then the target label value corresponding to that pixel is 1; otherwise, it is 0.

[0140] Step C213: Calculate the single-phase Dice loss value corresponding to the target based on the target phase probability value and the target phase label value corresponding to all the pixels.

[0141] Based on the target phase probability value and target phase label value corresponding to all pixels in the sample three-phase labeled image, the single-phase Dice loss value corresponding to the target is calculated. In this embodiment, the single-phase Dice loss value corresponding to the target is calculated using the following formula: In the formula, Indicates the first The single-phase Dice loss value corresponding to each target. This represents the total number of pixels in the three-phase labeled image of the sample. Indicates the first The target phase probability value corresponding to each pixel. For the first The target phase label value corresponding to each pixel. This is a minimum value used to avoid cases where the denominator is zero, such as 1e-5, to ensure stable calculation of single-phase Dice loss values.

[0142] Step C22: Calculate the full-map Dice loss value based on the single-phase Dice loss values ​​described above.

[0143] The overall image Dice loss value is calculated based on the Dice loss values ​​of each single phase. In this embodiment, the average of the single-phase Dice loss values ​​corresponding to the three targets is used as the corresponding overall image Dice loss value. In other embodiments, the weighted sum of the single-phase Dice loss values ​​of the three targets can be used as the corresponding overall image Dice loss value. The weights of each phase can be set empirically, with higher weights assigned to target phases that have a high impact on the three-phase segmentation accuracy to ensure the segmentation accuracy of the three-phase segmentation model.

[0144] Step C23: Based on the sample three-phase segmentation results and the sample three-phase labeled image, calculate the corresponding multiple single-phase boundary loss values.

[0145] Based on the sample three-phase segmentation results and sample three-phase labeled images, multiple corresponding single-phase boundary loss values ​​are calculated. This embodiment uses the symbolic distance field (SDF) to calculate the single-phase boundary loss value corresponding to each target, specifically including steps C231-C233.

[0146] Step C231: Based on the sample three-phase segmentation results and the sample three-phase labeled image, determine the pixel set corresponding to the target.

[0147] For each target phase, the corresponding pixel set is determined based on the sample three-phase segmentation results and the sample three-phase labeled image. In this embodiment, the pixel set includes the contour pixel set, the internal pixel set, and the external pixel set.

[0148] For the sample three-phase segmentation results, based on the probability of each pixel belonging to each phase, the contour pixel set, internal pixel set, and external pixel set corresponding to the target in the sample three-phase segmentation results are determined. First, for each pixel in the sample three-phase segmentation results, if the pixel is predicted to be the target phase (i.e., the probability of the pixel belonging to the target phase is the highest after the third activation layer of the three-phase segmentation model), the pixel is marked as 1; otherwise, it is marked as 0, resulting in a single-phase prediction binary map corresponding to the target. Second, for the single-phase prediction binary map, an edge detection algorithm (Canny edge detection is used in this embodiment) is used to extract the boundary between the target phase and non-target phases. All pixels on the boundary constitute the contour pixel set of the target phase. For multiple regions divided by the boundary, if the pixel labels in the region are all 1, then all pixels in the region constitute the internal pixel set of the target phase; if the pixel labels in the region are all 0, then all pixels in the region constitute the external pixel set of the target phase.

[0149] For the sample three-phase labeled image, based on the current labeled phase corresponding to each pixel, the contour pixel set, internal pixel set, and external pixel set corresponding to the target in the sample three-phase labeled image are determined. First, for each pixel in the sample three-phase labeled image, if the current labeled phase of the pixel is the target phase, the pixel is marked as 1; otherwise, it is marked as 0, resulting in a single-phase labeled binary image corresponding to the target. Second, for the single-phase labeled binary image, an edge detection algorithm (Canny edge detection is used in this embodiment) is used to extract the boundary between the target phase and non-target phases. All pixels on the boundary constitute the contour pixel set of the target phase. For multiple regions divided by the boundary, if the pixel labels within the region are all 1, then all pixels within the region constitute the internal pixel set of the target phase; if the pixel labels within the region are all 0, then all pixels within the region constitute the external pixel set of the target phase.

[0150] Step C232: Determine the symbolic distance field matrix corresponding to the target based on each pixel set.

[0151] For each target phase, the corresponding symbolic distance field matrix is ​​determined based on the set of pixels. In this embodiment, Chebyshev distance (chessboard distance) is used to determine the corresponding symbolic physical field matrix. In other embodiments, Euclidean distance or other distance calculation methods can also be used.

[0152] Taking the three-phase segmentation result of the sample as an example, Indicating the third phase of the sample three-phase segmentation results line, number The Chebyshev distance corresponding to a column pixel is determined if the pixel belongs to the contour pixel set. If the pixel belongs to the internal pixel set, then calculate the corresponding Chebyshev distance. ,in , These represent the minimum horizontal offset and minimum vertical offset of the pixel relative to each pixel in the contour pixel set, respectively; if the pixel belongs to the external pixel set, the corresponding Chebyshev distance is calculated. The symbolic distance field matrix corresponding to the three-phase segmentation result of the sample is constructed using the Chebyshev distances corresponding to all pixels in the sample three-phase segmentation result. .

[0153] Based on the same principle, determine the symbolic distance field matrix corresponding to the three-phase labeled image of the sample. .

[0154] Step C233: Calculate the single-phase boundary loss value corresponding to the target based on the distance field matrix of each symbol.

[0155] For each target phase, the corresponding single-phase boundary loss value is calculated based on the symbol range field matrix. In this embodiment, the single-phase boundary loss value is calculated using the following formula: In the formula, Indicates the first The single-phase boundary loss value corresponding to each target. This represents the number of horizontal pixels; in this embodiment, it is the width of the sample three-phase labeled image. This represents the number of vertical pixels; in this embodiment, it is the height of the sample three-phase labeled image. Indicating the third phase of the sample three-phase segmentation results line, number Chebyshev distance corresponding to column pixels Indicates the first phase of the three-phase labeled image of the sample. line, number Chebyshev distance corresponding to column pixels.

[0156] Step C24: Calculate the overall boundary loss value based on the single-phase boundary loss values.

[0157] Based on the boundary loss values ​​of each single phase, the boundary loss value of the entire image is calculated. In this embodiment, the average of the single-phase boundary loss values ​​corresponding to the three targets is used as the corresponding boundary loss value of the entire image. The three types of objects are treated equally to avoid the boundary constraints of a certain type of object (such as pores, which have a small pixel ratio) being ignored, thus adapting to scenarios where the pixel ratios of the three phases are unbalanced.

[0158] In other embodiments, the weighted sum of the single-phase boundary loss values ​​of the three phases can be used as the corresponding full-map boundary loss value. The weights of each phase can be set based on experience, with higher weights assigned to target phases that have a high impact on the three-phase segmentation accuracy, to ensure the segmentation accuracy of the three-phase segmentation model.

[0159] Step C25: Calculate the training loss value based on the full-graph Dice loss value and the full-graph boundary loss value.

[0160] The training loss is calculated based on the full-graph Dice loss and the full-graph boundary loss using the following formula: In the formula, Indicates the loss coefficient. This represents the Dice loss value for the entire image. This represents the boundary loss value of the entire map. In this embodiment... The value ranges from 0.4 to 0.6 and is used to adaptively adjust the proportion of the two losses to meet the segmentation requirements of this material with low contrast and blurred boundaries.

[0161] Step C3: Update the model parameters of the three-phase segmentation model according to the training loss value until the training stopping condition is met, and obtain the trained three-phase segmentation model.

[0162] The model parameters of the three-phase splitting model are updated based on the training loss value until the training stopping condition is met, resulting in a fully trained three-phase splitting model. In this embodiment, the Adam optimizer is used within the PyTorch framework, and the model is trained on an NVIDIA RTX 4070 Ti GPU.

[0163] For example, the initial learning rate is 1e-3; the learning rate strategy is cosine decay (period T=100, the learning rate decays from 1e-3 to 1e-5 over the period, then increases again, adapting to the actual training effect); the batch size is 8, adapted to the NVIDIA RTX4070Ti GPU memory, balancing training stability and efficiency; the maximum number of iterations is 150; the weight decay is 1e-4 to suppress model overfitting. The sample pairs of the training subset are input into the three-phase segmentation model, and training is accelerated using the NVIDIA RTX4070Ti GPU. After each training round, the training loss value and mIoU (Mean Intersection over Union) metric are calculated using the sample pairs of the validation subset, with a focus on the mIoU metric in low-contrast regions; when the validation subset mIoU shows no improvement for 10 consecutive rounds, an early stopping mechanism is triggered to save the optimal model parameters and avoid model overfitting. During training, the learning rate is dynamically adjusted based on the initial learning rate of 1e-3 using a cosine decay learning rate strategy (period T=100). For example, the learning rate decays to 2.7×10-4 in the 50th round and to 1e-5 in the 100th round, ensuring that the model converges stably on 16-bit low-contrast CT images and can accurately learn the weak boundary features of the matrix and interlayer pores.

[0164] Model parameters are adjusted using the loss function value, and an early stopping mechanism is implemented using the mIoU index to avoid model overfitting. The segmentation effect in low-contrast regions and regions with blurred boundaries is verified to ensure that the segmentation results can meet the requirements of numerical simulation of thermal conductivity and permeability of lightweight resin-based rigid heat insulation tile composite materials.

[0165] It is understandable that if a training dataset is constructed based on the XCT stretched images of each sample and the corresponding sample three-phase labeled images, then the sample XCT stretched images in the training dataset are input into the three-phase segmentation model to obtain the corresponding sample three-phase segmentation results; the training loss value is calculated based on the sample three-phase segmentation results and the sample three-phase labeled images corresponding to the sample XCT stretched images; the model parameters of the three-phase segmentation model are updated according to the training loss value until the training stopping condition is reached, and the trained three-phase segmentation model is obtained.

[0166] The trained three-phase segmentation model is used to segment lightweight resin-based rigid thermal insulation tile composites into three phases. The segmentation time for a single measured XCT stretch image is less than 0.5 seconds, meeting the real-time requirements of engineering applications. At the same time, multiple measured XCT stretch images can be input into the trained three-phase segmentation model to quickly complete the three-phase segmentation of a full set of measured XCT stretch images (e.g., 1000 images), balancing processing efficiency and segmentation accuracy. It is suitable for large-scale data processing needs and has strong engineering practicality.

[0167] The three-phase segmentation of lightweight resin-based rigid thermal insulation tile composite materials, as described in this embodiment, shows a significant improvement in segmentation performance compared to traditional threshold segmentation (Otsu's method) and the general Unet model. Figure 3 As shown, the fibers are clearly visible (white parts) in the XCT stretch image (measured XCT stretch image or sample XCT stretch image). The XCT stretch image is segmented into three phases using the three-phase segmentation model of this embodiment, the general Unet model, and the Otsu threshold segmentation method, respectively, to obtain the corresponding pore phase segmentation results, fiber phase segmentation results, and three-phase segmentation results. In the figure, purple represents pores, green represents fibers, and red represents the matrix. It can be understood that gray or white in the single-phase segmentation results represent other phases. Among the three three-phase segmentation results obtained by the three methods, it can be seen that the general Unet model and the Otsu OTSU method cannot identify the pores accurately, while the three-phase segmentation model of this embodiment can accurately identify and extract the pores and matrix, and the three-phase segmentation boundary is reasonable.

[0168] A quantitative comparison of the three methods was conducted on three indicators: mIoU, F1 score, and interlayer porosity error. The three-phase segmentation model in this embodiment showed significant superiority in all three aspects. Comparison examples are as follows:

[0169] Due to the presence of lamellar pores formed by fiber / matrix interface debonding and the matrix within the layers, even though the boundaries between these lamellar pores and the matrix can be distinguished by the naked eye, the grayscale transition between them is weak. These pores are mostly distributed along the fiber-matrix interface, with elongated shapes, tortuous boundaries, and grayscale values ​​close to those of the surrounding matrix, making it difficult for traditional thresholding methods to accurately determine pore boundaries. Otsu thresholding is limited by the inherent defects of global grayscale thresholding, generating a large number of misjudgments in the grayscale overlap region between pores and the matrix, resulting in fragmented pore phase segmentation and loss of connectivity. Its mIoU is only 0.58 and F1-score is 0.63, which is insufficient to meet the requirements for accurate characterization of microstructures. Although the general U-Net model effectively suppresses noise interference through convolutional feature extraction, it still has a significant bottleneck in restoring the connectivity of pore structures, manifested as broken pore networks and blurred edges, limiting its ability to completely reconstruct complex pore networks. The three-phase segmentation model trained in this embodiment effectively suppresses matrix background interference, accurately captures the subtle morphology of pores, and completely preserves the topological connectivity of the pore network. It achieves high-fidelity segmentation of the three phases of fiber, pores, and matrix. Its segmentation performance on 16-bit low-contrast XCT images is significantly better than the comparison methods. Specifically, the mIoU is improved by 0.25 compared to the general Unet model and by 0.24 compared to the traditional threshold segmentation (Otsu method). The interlayer porosity error is reduced to 2.9%, which meets the accuracy requirements (interlayer porosity error <5%) for the three-phase structure data in the numerical simulation of thermal conductivity and permeability of lightweight resin-based rigid thermal insulation tile composites. Meanwhile, analysis of the segmented visualization images (i.e., the three-phase segmentation results) revealed that while both the OTSU threshold segmentation and the general Unet model could accurately extract fibers, threshold segmentation completely failed to capture the layered interlayer pores with grayscale differences between deep interlayer pores and the matrix. The general Unet model improved upon threshold segmentation but still had a large number of missed detections and poor interlayer pore connectivity. In contrast, the three-phase segmentation model trained in this embodiment could accurately capture the subtle feature differences between interlayer pores and the matrix, accurately locate the fuzzy boundaries between interlayer pores and the matrix, and effectively suppress noise transmission. This provides a reliable guarantee for subsequent heat and mass transfer simulation calculations of the three phases based on the three-phase segmentation results, thereby establishing a predictive database for material structure and performance.

[0170] A specific embodiment of the present invention discloses a three-phase segmentation system for a lightweight resin-based rigid thermal insulation tile composite material, such as... Figure 4 As shown, it includes: The image stretching processing module is used to stretch the original measured XCT image to be segmented to obtain the corresponding stretched measured XCT image. The image segmentation module is used to segment the measured XCT stretched image using a trained three-phase segmentation model to obtain the corresponding three-phase segmentation result; wherein, the three-phase segmentation model includes an encoder, a channel attention branch, a spatial attention branch, a decoder, and a classifier connected in sequence.

[0171] The above-described method and system embodiments are based on the same principles, and their related aspects can be referenced from each other to achieve the same technical effects. For specific implementation processes, please refer to the foregoing embodiments, which will not be repeated here.

[0172] In summary, the three-phase segmentation method and system for a lightweight resin-based rigid thermal insulation tile composite material according to embodiments of the present invention has at least one of the following beneficial effects: 1. First, the contrast of XCT images of lightweight resin-based rigid thermal insulation tile composite materials is increased through stretching. Then, a three-phase segmentation model consisting of an encoder, channel attention branch, spatial attention branch, decoder, and classifier is used for three-phase segmentation. The channel attention branch accurately identifies the importance of different feature channels, while the spatial attention branch focuses on local areas of the image. By cooperating with the channel attention branch and the spatial attention branch, dual optimization of channel selection and spatial localization is achieved. This breaks through the limitations of traditional segmentation methods that can only enhance features individually and cannot be specifically adapted to low-contrast blurred boundary scenes. As a result, the matrix with blurred boundaries and interlayer pores are accurately distinguished, the boundary segmentation error is significantly reduced, and the segmentation adaptability and accuracy of XCT images of lightweight resin-based rigid thermal insulation tile composite materials are improved.

[0173] 2. Downsampling is performed by the encoder to enhance the ability to capture the boundaries and texture features of the matrix and interlayer pores under low contrast. Channel information of the basic feature map is obtained through the channel attention branch, and channel attention weighted fusion is performed to focus on strengthening the channel weights corresponding to the features of the matrix and interlayer pore boundaries. Spatial attention weighted fusion is performed through the spatial attention branch to focus on the blurred boundary areas of the image, assigning high weights to boundary pixel positions and low weights to non-boundary redundant areas, further enhancing the feature response of the boundary areas. Upsampling is performed by the decoder to obtain high-dimensional feature maps, accurately distinguishing the blurred boundaries of the matrix and interlayer pores, and significantly reducing the boundary segmentation error.

[0174] 3. By performing morphological transformation on the three-phase labeled images of the samples, the interlayer pore morphology changes that may occur in the preparation and service of lightweight resin-based rigid heat insulation tile composite materials are accurately simulated. The feature of blurred boundaries is preserved, which ensures that the samples are highly matched with the actual three-phase structure of the materials and improves the sample richness of the training dataset.

[0175] 4. Based on the full-image Dice loss value and the full-image boundary loss value, the training loss value is calculated to specifically address the pain points of low contrast and blurred boundaries between the matrix and interlayer pores in XCT images of lightweight resin-based rigid thermal insulation tile composite materials. The two work together to solve the problems of sample imbalance and boundary blurring, respectively, ensuring the accuracy of model training.

[0176] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0177] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A three-phase segmentation method for a lightweight resin-based rigid thermal insulation tile composite material, characterized in that, include: The original measured XCT image to be segmented is stretched to obtain the corresponding stretched measured XCT image. The trained three-phase segmentation model is used to segment the measured XCT stretched image to obtain the corresponding three-phase segmentation result; wherein, the three-phase segmentation model includes an encoder, a channel attention branch, a spatial attention branch, a decoder and a classifier connected in sequence.

2. The method according to claim 1, characterized in that, The trained three-phase segmentation model is used to segment the measured XCT stretched image to obtain the three-phase segmentation results, including: The encoder receives the measured XCT stretched image, downsamples the measured XCT stretched image, generates multiple basic feature maps, transmits one of the basic feature maps to the decoder, and transmits the other basic feature maps to the channel attention branch; The channel attention branch performs channel attention weighted fusion on the received basic feature map to generate a corresponding channel weighted feature map, and transmits it to the spatial attention branch; The spatial attention branch performs spatial attention weighted fusion on the received channel weighted feature map to generate a corresponding spatial weighted feature map, which is then transmitted to the decoder. The decoder upsamples the received basic feature map and each of the spatially weighted feature maps to generate a high-dimensional feature map, which is then transmitted to the classifier. The classifier performs three-phase classification on the high-dimensional feature map and generates the corresponding three-phase segmentation result.

3. The method according to claim 1, characterized in that, The trained three-phase segmentation model is obtained through the following steps: Three-phase annotation is performed on the original XCT image of each sample in the sample dataset to obtain the corresponding sample three-phase annotated image; A training dataset is constructed based on the original XCT images of each sample and the corresponding three-phase labeled images of the sample. The three-phase segmentation model is trained using the training dataset to obtain the trained three-phase segmentation model.

4. The method according to claim 3, characterized in that, For each sample XCT raw image in the sample dataset, perform three-phase annotation to obtain the corresponding sample three-phase annotated image, including: Sliding sampling is performed on each of the original XCT images of the samples to obtain multiple sample XCT local images; Each of the sample XCT local images is contrast stretched, and all the stretched sample XCT local images are fused to obtain the sample XCT stretched image; wherein, the sample XCT stretched image includes the pixel grayscale value of each pixel; The sample XCT stretched image is annotated in three phases to obtain the corresponding sample three-phase annotated image.

5. The method according to claim 4, characterized in that, Contrast stretching is performed on each of the sample XCT local images, including: Contrast stretching is achieved by updating the pixel grayscale value of each pixel in the sample XCT local image using the following formula: In the formula, The XCT local image of the sample represents the first... The updated pixel grayscale value corresponding to each of the aforementioned pixels. Indicates the first The pixel grayscale value before the update corresponds to each of the aforementioned pixels. , The minimum and maximum pixel gray values ​​are, in order, the local XCT images of the samples. , These represent the preset upper limit and lower limit of pixel grayscale values, respectively.

6. The method according to claim 3, characterized in that, After constructing the training dataset based on the original XCT images of each sample and the corresponding three-phase labeled images of the sample, the following steps are also included: Perform binary transformation on any of the sample three-phase labeled images to generate the corresponding basic binary label image; Perform interlayer gap mutation operation on the basic binary label map to obtain multiple mutated binary label maps; Based on the basic binary label map and each of the variant binary label maps, the sample three-phase labeled image is labeled and replaced, and the corresponding sample XCT original image is pixel replaced to obtain the replaced sample three-phase labeled image and the sample XCT original image. All the replaced sample three-phase labeled images and the sample original XCT images are added to the training dataset.

7. The method according to claim 4, characterized in that, The three-phase segmentation model is trained using the training dataset to obtain the trained three-phase segmentation model, including: The original XCT images of the samples in the training dataset are input into the tri-phase segmentation model to obtain the corresponding tri-phase segmentation results of the samples. The training loss value is calculated based on the sample tri-phase segmentation results and the sample tri-phase labeled images; The model parameters of the three-phase segmentation model are updated according to the training loss value until the training stopping condition is met, thus obtaining the trained three-phase segmentation model.

8. The method according to claim 7, characterized in that, The training loss value is calculated based on the sample three-phase segmentation result and the sample three-phase labeled image, including: Based on the sample three-phase segmentation results and the sample three-phase labeled image, calculate the corresponding multiple single-phase Dice loss values; Calculate the full-map Dice loss value based on the single-phase Dice loss values ​​described above; Based on the sample three-phase segmentation results and the sample three-phase labeled image, calculate the corresponding multiple single-phase boundary loss values; Calculate the overall boundary loss value based on the single-phase boundary loss values ​​described above; The training loss value is calculated based on the full-graph Dice loss value and the full-graph boundary loss value.

9. The method according to claim 8, characterized in that, Based on the sample three-phase segmentation results and the sample three-phase labeled image, calculate the corresponding multiple single-phase Dice loss values, including: Based on the three-phase segmentation results of the sample, the probability value of each pixel in the three-phase segmentation results being the target phase is determined; wherein, the target phase is any one of the phases among fiber, matrix, and interlayer pores; Based on the sample three-phase labeled image, the target phase label value corresponding to each pixel in the sample three-phase labeled image is determined; wherein, each pixel in the sample three-phase segmentation result corresponds one-to-one with each pixel in the sample three-phase labeled image; The single-phase Dice loss value corresponding to the target is calculated based on the target phase probability value and the target phase label value corresponding to all the pixels.

10. A three-phase segmentation method system for a lightweight resin-based rigid thermal insulation tile composite material, characterized in that, include: The image stretching processing module is used to stretch the original measured XCT image to be segmented to obtain the corresponding stretched measured XCT image. The image segmentation module is used to segment the measured XCT stretched image using a trained three-phase segmentation model to obtain the corresponding three-phase segmentation result; wherein, the three-phase segmentation model includes an encoder, a channel attention branch, a spatial attention branch, a decoder, and a classifier connected in sequence.

Images