Structural XCT fiber bundle identification and segmentation method based on improved U-Net

Abstract

The invention discloses a structural XCT fiber bundle identification and segmentation method based on improved U-Net, which is used for identifying fiber bundles in XCT slices of a woven CMC structural member, and the network is applied to the CMC field for the first time. According to the method, the down-sampling operation of the original U-Net network is greatly modified, the characteristic extraction capability of the model on the fiber bundle is improved, the parameter quantity of the model is reduced, and the network model can be trained more stably, so that the efficiency of segmenting and identifying the fiber bundle by the model is improved. Wherein an Inception structure is adopted, so that the visual field of the model is widened, and the recognition and segmentation capability of the model on fiber bundles with different sizes and boundaries thereof in CT is enhanced. And residual connection is added in an original model by adopting a ResNet structure, so that gradient propagation in training is ensured, the depth of the improved U-Net is deepened, and the accuracy of segmentation recognition is improved. And the bottleneck structure is adopted, so that the parameter quantity of the whole model is greatly reduced, the training time is shortened, and the recognition and segmentation speed is increased.

Description

Technical Field

[0001] The present invention belongs to the field of mesoscopic structure identification of ceramic matrix composite structural components, and relates to a method for identifying and segmenting XCT fiber bundles based on an improved U-Net structure. Background Art

[0002] Ceramic Matrix Composite (hereinafter simply referred to as CMC) has the outstanding advantages of low density and high temperature resistance, and is listed in the advanced projects of the development and application of new composite materials for aeroengines, which are key development areas for the major powers in the world today. The woven structure can be widely applied to various components of aeroengines.

[0003] In order to obtain the mesoscopic structure of such complex CMC structural components as fiber bundles, non-destructive testing XCT scanning technology can be used. It can accurately obtain the true mesoscopic structure inside the material without damaging the material, and effective information related to fiber bundles can be extracted from a series of XCT slices related to fiber bundles.

[0004] The identification and segmentation of XCT fiber bundles is the first and most important step. At present, XCT non-destructive testing technology has been well developed, but because there are numerous fiber bundles in the real CMC structural components, the technology for identifying and segmenting fiber bundles in XCT slices of structural components is still a major technical problem. The identification and segmentation mentioned here mainly refers to the classification problem of each pixel in each slice, that is, separating and masking the pixels belonging to the fiber bundles in the CT slices.

[0005] Most of the existing mesoscopic structure identification technologies for XCT slices are based on traditional computer vision technologies. Such as the gray level co-occurrence matrix proposed for three-dimensional orthogonal woven fabrics, the structure tensor method proposed for two-dimensional woven fabrics, and the fiber tracking algorithm proposed for multi-directional carbon fiber reinforced polymer (CFRP) specimens. The identification effects of these methods seriously depend on the scanning resolution and are only designed according to specific mesoscopic structures, and are not applicable to the identification of fiber bundles in complex CMC structural components.

[0006] Therefore, for 2.5D woven structures, some scholars have proposed a maximum threshold segmentation method to identify the fiber bundles in XCT images. Such as the patent with the Chinese patent application number "201610838554.8" and the invention name "A computer image recognition technology and three-dimensional modeling method for the mesoscopic structure of composite materials". This method first identifies the matrix part of the 2.5D woven structure XCT image, and then pairs it using the left-right symmetry of the matrix. Based on this, the boundary is fitted. The warp yarns are between the top and bottom of the matrix, and the weft yarns are between the left and right. This method is carried out based on the symmetry of the matrix, but in the actual structure, the matrix part is an asymmetric structure, and this identification algorithm fails and has poor robustness.

[0007] For three-dimensional four-directional braided structures, some scholars have established a theoretical model of the unit cell of three-dimensional braided composites based on preset geometric parameters, compared it with the XCT slice images after threshold segmentation and denoising, changed the preset values of the theoretical model, and obtained the final recognition effect diagram after optimization. For example, the patent with Chinese patent application number "201810537212.1" and invention name "Computer image recognition technology and three-dimensional modeling method for the mesoscopic structure of composite materials". However, during the recognition process using threshold segmentation in XCT images, due to shooting problems, the gray-scale change range between fiber bundles is large, and it is very difficult to find an optimal threshold for segmentation for each slice.

[0008] Neural network is an emerging segmentation and recognition method. Existing research has used neural networks to recognize fiber bundles in ceramic matrix composite materials. The multi-decoder fully convolutional neural network can perform semantic segmentation on XCT images of CMC. The multi-decoder network uses the max-pooling decoder network, channel fusion decoder network, and channel connection decoder network, which are designed in parallel. By training the network, the final recognition of the mesoscopic structure is achieved. For example, the patent with Chinese patent application number "201910165647.2". However, the multi-decoder network occupies a large amount of video memory during the training process and has a long training time. To solve this problem, some scholars have proposed a lightweight network based on YOLOv8, that is, the patent with Chinese patent application number "202410861340.7". However, it only performs recognition work on fiber bundles in CT images at the unit cell scale. The actual CMC structural parts are much larger in volume than the unit cell and have more fiber bundles. The YOLOv8 algorithm has poor recognition effect for multi-targets and low recognition accuracy, and is unable to perform fiber bundle recognition work at the structural part scale.

[0009] In summary, although existing methods can already achieve the recognition and segmentation of fiber bundles in CMC, they are not applicable to the recognition of fiber bundles in structural parts and cannot be applied in practice. Therefore, further research is needed on the semantic segmentation of XCT slices of ceramic matrix composites based on deep learning. Summary of the Invention

[0010] The technical problem to be solved by the present invention is to provide an improved U-Net-based structural XCT fiber bundle recognition and segmentation method in view of the above-mentioned existing deficiencies. Compared with existing methods, this method can quickly segment fiber bundles in XCT images of structural parts with high accuracy. The improved U-Net of the present invention is first applied to the CMC field, so it is called U-Net_CMCv1. The U-Net_CMCv1 network can be used to recognize and segment fiber bundles (warp yarns, weft yarns) in CMC structural parts of various braided types.

[0011] To achieve the above technical objectives, the technical solutions adopted by the present invention are as follows:

[0012] An improved U-Net-based structural XCT fiber bundle identification and segmentation method, characterized by comprising the following steps:

[0013] Step 1: Establish a fiber bundle XCT slice dataset of a woven CMC structural component and divide it into a training set, a test set, and a validation set;

[0014] Step 2: Perform data augmentation on the fiber bundle XCT slice dataset of the woven CMC structural component to increase the number of samples; then correct the mislabeled tags through data cleaning to form a high-quality dataset;

[0015] Step 3: Construct a network model of an improved U-Net network for fiber bundle segmentation in a woven CMC structural component and conduct debugging to ensure that the model can run normally;

[0016] Step 4: Input the fiber bundle XCT slice images of the woven structural components in the dataset and their corresponding label files into the improved U-Net network;

[0017] Step 5: According to the above network model, after constructing the improved U-Net network in the Pytorch neural network framework, adjust the training parameters: set the Adam optimizer to adjust the learning rate to train the improved U-Net network, and at the same time use 16-bit floating-point numbers for training, and adopt a linear warm-up learning strategy and a frozen training strategy to train the model, set the specified file path, and save the weight file and the corresponding evaluation parameters obtained after the training is completed;

[0018] Step 6: After the improved U-Net network is trained, select a specific folder, replace the improved convolution operation in the improved downsampling module in the network with the corresponding operation for recognition and prediction, and then use the obtained weight file to perform semantic segmentation on all images and save them in this folder.

[0019] To optimize the above technical solution, the specific measures taken also include:

[0020] The improved U-Net network is composed of an improved downsampling module, an improved upsampling module, and an improved bridging module, and specifically includes an improved convolution operation, a Transformer attention mechanism module, and a deconvolution operation.

[0021] The improved downsampling module is used to extract the features of the fiber bundle in the CT image to form a feature image, which is composed of 11 improved convolution operations. After the feature map is input, it will sequentially pass through these 11 improved convolution operations. Among them, a max pooling operation is connected after the 3rd, 6th, and 9th improved convolutions to form an improved convolution & Max pooling operation to compress the image to half of its original size, so as to extract the features of the fiber bundle subsequently.

[0022] The role of the improved upsampling module is to perform upsampling operations on the feature map after the improved downsampling module extracts the features in the graph, so as to restore the position information of the fiber bundles in the feature map. The improved upsampling module will first perform 4 upsampling operations. The upsampling operation is that the feature map passes through two improved convolution operations in sequence and then performs a transposed convolution operation. At the end of the improved upsampling module, the feature map will undergo two improved convolution operations.

[0023] The improved bridging module is used to achieve feature fusion, that is, the bridging layer merges the high-resolution feature map obtained from the improved downsampling module with the low-level detail feature map from the improved upsampling module. There are a total of 4 improved bridging modules, and each module consists of 3 fully connected layers. Each fully connected layer adopts the SENet structure. The first improved bridging module connects the feature map after the second improved convolution in the improved downsampling module with the feature map after the fourth transposed convolution module in the improved upsampling module; the second improved bridging module connects the feature map after the fourth improved convolution in the improved downsampling module with the feature map after the third transposed convolution module in the improved upsampling module; the third improved bridging module connects the feature map after the sixth improved convolution in the improved downsampling module with the feature map after the second transposed convolution module in the improved upsampling module; the fourth improved bridging module connects the feature map after the eighth improved convolution in the improved downsampling module with the feature map after the first transposed convolution module in the improved upsampling module.

[0024] The improved convolution operation is divided into the improved convolution operation used during training and the improved convolution operation used during recognition and prediction. The improved convolution operation used during training adopts the Inception structure, that is, three convolutions of sizes 3×3, 5×5, and 7×7 are added in parallel additionally, with a stride of 2 for all of them and an average pooling operation is added. The ResNet structure is also adopted in these three branches, that is, a convolution of size 1×1 with a stride of 2 is added in parallel for each of the three convolutions of different sizes as a residual connection. At the same time, the bottleneck structure is adopted, that is, a convolution of size 1×1 with a stride of 2 is added before and after each of the three convolutions. The improved convolution used during recognition and prediction is simplified compared to the improved convolution used during training, that is, the ResNet structure is cancelled.

[0025] The Transformer attention mechanism module is located between the improved downsampling module and the improved upsampling module. It is used to evaluate the fiber bundle features extracted by the improved downsampling module, screen out the features that contribute greatly to fiber bundle recognition. When the feature map passes through this module, it will first be divided into 64 small patches. After flattening and linear transformation operations, it is flattened to extract position information and embed additional learnable features. Then, after being scored by the Transformer encoder, a multi-layer perceptron is used to screen out important features. The flattened patches are input into the Transformer encoder and will successively undergo layer normalization, multi-head attention mechanism, layer normalization, and multi-layer perceptron. The ResNet structure is adopted in the first layer normalization, multi-head attention mechanism, second layer normalization, and multi-layer perceptron respectively.

[0026] The transposed convolution operation is consistent with that of U-Net, which plays a role in restoring the fiber bundle features and can expand the side length of the feature map to twice the original.

[0027] The feature map processed by the improved upsampling module will pass through a 1×1 convolution with a stride of 1 to be restored to a 3-channel RGB image, and then the size of the image is adjusted to the size of the input CT image to obtain the final fiber bundle segmentation and recognition image.

[0028] After the feature map data passes through convolution or feature map merging, it needs to go through batch normalization operation and SiLu activation function before further processing.

[0029] In step one, the fiber bundles are manually marked in the open-source software Labelme to generate label files. In the dataset, the ratio of the training set, test set, and validation set is 3:1:1.

[0030] In step two, the data augmentation methods include cropping, flipping, rotating, adding noise, brightness change, and contrast enhancement. After enhancement, the ratio of the training set, test set, and validation set in the dataset should also be controlled at 3:1:1.

[0031] In step five, the number of training times is set to 1000 times. The initial learning rate, first-order moment decay rate, second-order moment decay rate, and weight coefficient of the Adam optimizer are set to 0.0005, 0.9, 0.999, and 0.0001 respectively. For linear warm-up learning, it is stipulated that the learning rate gradually rises to the initial value in the first 20 rounds of training. After each training, the corresponding weight file is saved, and the XCT images on the validation set are segmented using the weight file to obtain three evaluation parameters: precision rate, recall rate, and mean average precision.

[0032] In step six, test the weight file generated in step five in the test set, and use three evaluation parameters, namely precision, recall, and mean average precision, to select the optimal weight file, and select the one with the best segmentation effect as the optimal weight file to segment all XCT images.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] 1. Make significant modifications to the downsampling operation of the original U-Net network, which improves the model's ability to extract features of fiber bundles while reducing the number of model parameters, and can train the network model more smoothly, thereby improving the efficiency of the model in segmenting and identifying fiber bundles. The Inception structure adopted broadens the model's vision, thus enhancing the model's ability to identify and segment fiber bundles of different sizes and their boundaries in CT. The ResNet structure adopted adds residual connections to the original model, ensuring the propagation of gradients during training, enabling the depth of the improved U-Net to be deepened, thereby enhancing the model's ability to extract features of fiber bundles and improving the accuracy of segmentation and identification. The bottleneck structure significantly reduces the overall number of model parameters, thereby reducing the computational resources consumed in training the model, reducing the training time, and accelerating the speed of identification and segmentation. When performing segmentation prediction, the residual connection is cancelled and a single-line structure is adopted, reducing the computational resources consumed in identification and segmentation and enhancing the speed of identification and segmentation.

[0035] 2. Both the Transformer attention mechanism and the SENet structure in the improved bridging module can extract several features that contribute significantly to fiber bundles, ensuring that the model will consider other extracted features when using each feature. The main features of fiber bundles screened by the Transformer attention mechanism can greatly improve the detection speed and accuracy, and at the same time enable parallel training, enhancing the efficiency of model training. The SENet structure can make the structure of the bridging module more flexible.

[0036] 3. Replace the activation function with the SiLu activation function, which improves the forward propagation speed of the model during training, solves the problem of gradient dispersion generated in the training of deep networks, ensures the update of weights, so that the model can fully learn fiber bundle-related features during training, and improves the accuracy of model recognition. A batch normalization operation is added before the activation function. On the one hand, it reduces the absolute difference between data and makes the outputs of each layer in the network more stable; on the other hand, it ensures that the data is identically distributed, which can avoid overfitting during model training and thus ensure the effect of the trained model in actual segmentation.

[0037] 4. The training adopted a strategy combining linear warm-up and frozen training, which avoided introducing excessive gradients when the model parameters were not yet adapted, resulting in unstable training. At the same time, using frozen training to randomly freeze certain layers in the improved U-Net could prevent the model from overfitting on the training data, thereby improving the generalization ability of the model. In addition, the frozen layers could reduce the computational amount because the parameters of these layers did not need to be updated, which could significantly shorten the training time. Most importantly, freezing some of the layers could reduce the uncertainty during the model training process and help the model quickly converge to a better solution. Description of the Drawings

[0038] Figure 1 It is the network structure diagram of the improved U-Net;

[0039] Figure 2 It is the improved convolution operation diagram adopted during training;

[0040] Figure 3 It is the improved convolution operation diagram adopted for recognition and prediction;

[0041] Figure 4 It is the SENet structure diagram used in the improved bridging module;

[0042] Figure 5 It is the Transformer attention mechanism module diagram;

[0043] Figure 6 It is the effect diagram of warp yarn recognition and segmentation;

[0044] Figure 7 It is the effect diagram of weft yarn recognition and segmentation. Detailed Implementation Manner

[0045] In order to make the purpose, technical solutions and advantages of this application clearer, the following describes and explains this application in combination with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not used to limit this application. Based on the embodiments provided in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of this application.

[0046] Obviously, the accompanying drawings in the following description are only some examples or embodiments of the present application. For those of ordinary skill in the art, without creative efforts, the present application can also be applied to other similar scenarios based on these drawings. In addition, it can also be understood that although the efforts made in such a development process may be complex and lengthy, for those of ordinary skill in the art related to the content disclosed in the present application, some design, manufacturing, or production changes based on the technical content disclosed in the present application are only conventional technical means and should not be understood as insufficient disclosure of the content of the present application.

[0047] When the term "embodiment" is mentioned in the present application, it means that the specific features, structures, or characteristics described in connection with the embodiment can be included in at least one embodiment of the present application. The appearance of this phrase at various positions in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment mutually exclusive with other embodiments. Those of ordinary skill in the art explicitly and implicitly understand that the embodiments described in the present application can be combined with other embodiments without conflict.

[0048] Unless otherwise defined, the technical terms or scientific terms involved in the present application should have the ordinary meaning understood by those with ordinary skills in the technical field to which the present application belongs. The terms "a", "an", "one kind", "the", and the like involved in the present application do not indicate a quantity limitation and can represent a singular or plural number. The terms "including", "comprising", "having", and any variations thereof involved in the present application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units (units) is not limited to the listed steps or units, but may further include unlisted steps or units, or may further include other steps or units inherent to these processes, methods, products, or devices. The terms "connected", "coupled", and the like involved in the present application are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The term "a plurality of" / "several" involved in the present application refers to two or more. "And / or" describes the association relationship of associated objects and indicates that three relationships may exist. For example, "A and / or B" may represent: A exists alone, A and B exist simultaneously, and B exists alone. The character " / " generally represents an "or" relationship between the associated objects before and after. The terms "first", "second", "third", etc. involved in the present application are only used to distinguish similar objects and do not represent a specific order for the objects.

[0049] The present invention proposes an improved U-Net network and a method for fiber bundle identification and segmentation of a woven CMC structural member based on this network, including the following steps:

[0050] Step 1: Construct the XCT datasets for warp and weft yarns respectively, and divide them into training set, validation set and test set.

[0051] Manually mark the fiber bundles in XCT in the software Labelme and generate the corresponding label files. Based on this, a total of 200 warp yarn datasets required for the improved U-Net network are created, and 120 of the pictures are designated as the training set, 40 pictures as the validation set, and 40 pictures as the test set. Similarly, 200 weft yarn datasets are created, and 120 of the pictures are designated as the training set, 40 pictures as the validation set, and 40 pictures as the test set.

[0052] Step 2: After performing data augmentation on the XCT data of warp and weft yarns respectively, perform data cleaning to correct the incorrect labels to obtain high-quality datasets.

[0053] The data augmentation methods include cropping, flipping, rotating, adding noise, brightness change, and contrast enhancement. After augmentation, there are 1200 pictures in the warp yarn dataset, including 720 pictures in the training set, 240 pictures in the validation set, and 240 pictures in the test set. There are 1200 pictures in the weft yarn dataset, including 720 pictures in the training set, 240 pictures in the validation set, and 240 pictures in the test set.

[0054] Step 3: Build an improved U-Net network structure for identifying and segmenting the XCT fiber bundles of the woven CMC structural parts. This network consists of an improved downsampling module, an improved upsampling module, and an improved bridging module. Specifically, it includes improved convolution operations, a Transformer attention mechanism module, and deconvolution operations.

[0055] First, build the improved downsampling module. As Figure 1 shown, the improved downsampling module is sequentially composed of 11 improved convolution operations as Figure 2 shown. The pictures input into the improved downsampling module will be unified to a size of 572×572. Among them, a max pooling operation is added after the 3rd, 6th, and 9th downsampling operations. After the max pooling operation, the size of the feature map becomes half of the original. After the 1st, 4th, 7th, and 10th downsampling operations, the number of channels of the feature map becomes twice the original. The size and number of channels of the feature images remain unchanged after the remaining improved convolution operations.

[0056] Then build the improved upsampling module. As Figure 1 shown, the improved upsampling module is sequentially composed of 4 upsampling operations and 2 improved convolution operations. Each upsampling operation is sequentially composed of 2 improved convolution operations as Figure 2 shown and 1 deconvolution operation, where the deconvolution operation is the same as that in the original U-Net network.

[0057] Next, insert 4 improved bridging modules into the improved downsampling module and the improved upsampling module. Each bridging module consists of 3 fully connected layers, and the SENet structure shown in Figure 4 is adopted in each fully connected layer. In the SENet structure, the width, height, and number of channels of the input feature map x are w, h, and c1 respectively. First, perform feature screening on the features in the feature map x through θ convolutional transformations F tr (·, θ), and at this time the number of channels of the feature map x becomes c2. Then, perform a squeezing operation F sq (·) to perform global average pooling on the feature map to obtain feature data of size 1×1×c2. Then, obtain the feature weight data through an excitation operation F ex (·, w), that is, the Figure 4 color vector of size 1×1×c2 in . Finally, according to the feature weights, obtain the feature map with redistributed feature importance through the feature map restoration operation F scale (·) The first improved bridging module is located among the feature maps after the second improved convolution in the improved downsampling module and the feature maps after the fourth transposed convolution in the improved upsampling module; the second improved bridging module is located among the feature maps after the fourth improved convolution in the improved downsampling module and the feature maps after the third transposed convolution in the improved upsampling module; the third improved bridging module is located among the feature maps after the sixth improved convolution in the improved downsampling module and the feature maps after the second transposed convolution in the improved upsampling module; the fourth improved bridging module is located among the feature maps after the eighth improved convolution in the improved downsampling module and the feature maps after the first transposed convolution in the improved upsampling module.

[0058] Then, insert the Transformer attention mechanism module shown in Figure 5 into the improved downsampling module and the improved upsampling module.

[0059] Finally, insert a 1×1 convolution after the improved upsampling module to restore the feature map to a 3-channel RGB map, and then adjust the size of the picture to the size of the input picture to perform the output.

[0060] Step 4: Import the warp dataset and the weft dataset into the improved U-Net in sequence.

[0061] The data imported into the improved U-Net network includes the original XCT maps of the warp and weft and the corresponding annotation files.

[0062] Step 5: Configure the training environment and adjust the relevant parameters used for model training.

[0063] The operating system used for training is Windows 11. The neural network framework used to build the improved U-Net is Pytorch, and its version is 2.2.2. The graphics card model used for training is GeForce RTX4090, the relevant driver version is 566.36, and the CUDA version is 11.8. The number of training epochs is set to 1000, and the batch size for each round of training is set to 16. The initial learning rate, first-moment decay rate, second-moment decay rate, and weight coefficient of the Adam optimizer are set to 0.0005, 0.9, 0.999, and 0.0001 respectively. For linear warm-up learning, the number of epochs for gradually increasing the learning rate to the initial value is set to 20.

[0064] Step 6: Conduct the training of the improved U-Net network.

[0065] Train the warp and weft yarns in sequence. At the end of each round of training, use the validation set to test the effect of the trained model. Calculate the gap between the real image and the predicted recognition segmentation image using the cross-entropy loss function to obtain the loss of the improved U-Net network, and update the weights of the YOLOv8_CMC network through the error backpropagation algorithm and the Adam optimizer. After the training is completed, obtain and save the weight file, and use the obtained weight file to perform recognition segmentation on the images in the validation set to obtain the corresponding evaluation parameters, namely precision, recall, and mean average precision.

[0066] Step 7: After the training of the improved U-Net network is completed, select a specific folder to perform semantic segmentation on all images and save them in this folder.

[0067] For the warp and weft yarns respectively, select the corresponding optimal weight file by integrating the three evaluation parameters of precision, recall, and mean average precision. After replacing the improved convolution operation used in the training of the model with the improved convolution operation used in recognition prediction as shown in Figure 3 perform recognition segmentation on the XCT images of the warp and weft yarns with the obtained optimal weight file and save them. The recognition segmentation effects of the warp and weft yarns are as shown in Figure 6 and Figure 7 shown.

[0068] In the present invention, an improved U-Net network is established to identify and segment fiber bundles in XCT images of woven CMC structural components. The adopted Inception structure enhances the receptive field of the network, thereby improving the recognition ability of fiber bundle boundaries. The residual connection and bottleneck structure are adopted to reduce the network parameters while facilitating the propagation of gradients during training and the stability of training. During prediction, the Inception structure is changed to a single-line structure, reducing the consumption of computing resources and improving the efficiency of identifying and segmenting fiber bundles. By adding the Transformer attention mechanism and the SENet structure in the bridging module, the ability of the network to utilize and extract fiber bundle features is enhanced. Finally, by adopting the SiLu activation function and batch normalization method, the smoothness of feature data transmission in the network and the network's ability to express non-linear features are improved. The improved U-Net network for identifying and segmenting fiber bundles in woven CMC structural components can provide accurate fiber bundle geometric data for the 3D modeling of related structural components.

[0069] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited by the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A structural XCT fiber bundle identification and segmentation method based on improved U-Net, characterized in that: The following steps are involved: Step 1: Establish a fiber bundle XCT slice dataset of woven CMC structural parts and divide it into a training set, a test set, and a validation set; Step 2: Perform data enhancement on the fiber bundle XCT slice dataset of the woven CMC structural parts to increase the number of samples; then correct the incorrectly marked labels through data cleaning to form a high-quality dataset; Step 3: Build a network model of the improved U-Net network for fiber bundle segmentation in woven CMC structural parts, and debug it to ensure that the model can run normally; Step 4: Input the fiber bundle XCT slice image of the woven structural part in the data set and its corresponding label file into the improved U-Net network; Step 5. According to the above network model, after building the improved U-Net network in the Pytorch neural network framework, adjust the training parameters: set the Adam optimizer to adjust the learning rate to train the improved U-Net network, use 16-bit floating-point numbers for training, and adopt the linear warm-up learning strategy and the frozen training strategy to train the model, set the specified file path, save the weight file and the corresponding evaluation parameters obtained after the training is completed; Step 6. After the improved U-Net network training is completed, select a specific folder, replace the improved convolution operation in the improved downsampling module in the network with the corresponding operation for recognition prediction, use the obtained weight file to perform semantic segmentation on all images, and save them in the folder.

2. According to the improved U-Net-based structural XCT fiber bundle identification and segmentation method of claim 1, it is characterized in that: The improved U-Net network is composed of an improved downsampling module, an improved upsampling module and an improved bridging module, which specifically includes an improved convolution operation, a Transformer attention mechanism module and a deconvolution operation. The improved downsampling module is used to extract the features of the fiber bundles in the CT image and form a feature image. It consists of 11 improved convolution operations. After the feature image is input, these 11 improved convolution operations will be performed in sequence. The third, sixth, and ninth improved convolutions are followed by a maximum pooling operation to form an improved convolution & Max pooling operation to achieve the effect of compressing the image to half of the original size, so as to subsequently extract the features of the fiber bundles. The function of the improved upsampling module is to perform an upsampling operation on the feature map after the improved downsampling module extracts the features in the map, thereby restoring the position information of the fiber bundles in the feature map. The improved upsampling module will first perform 4 upsampling operations. The upsampling operation is that the feature map is sequentially subjected to two improved convolution operations and then subjected to a deconvolution operation. At the end of the improved upsampling module, the feature map will undergo two improved convolution operations. The improved bridge module is used to realize feature fusion, that is, the bridge layer merges the high-resolution feature map obtained from the improved downsampling module with the low-level detail feature map from the improved upsampling module. There are 4 improved bridge modules in total, each of which is composed of 3 fully connected layers, where each fully connected layer adopts the SENet structure. The first improved bridge module connects the feature map after the second improved convolution in the improved downsampling module and the feature map after the fourth deconvolution module in the improved upsampling module. The second improved bridge module connects the feature map after the fourth improved convolution in the improved downsampling module and the feature map after the third deconvolution module in the improved upsampling module; the third improved bridge module connects the feature map after the sixth improved convolution in the improved downsampling module and the feature map after the second deconvolution module in the improved upsampling module; The fourth improved bridge module connects the feature map after the eighth improved convolution in the improved downsampling module and the feature map after the first deconvolution module in the improved upsampling module. The improved convolution operation is divided into the improved convolution operation used for training and the improved convolution operation used for recognition prediction. The improved convolution operation used for training adopts the Inception structure, that is, three convolutions of 3×3, 5×5, and 7×7 are added in parallel, and their step sizes are all 2, and an average pooling operation is added. The ResNet structure is used in these three branches, that is, a convolution with a step size of 2 and a size of 1×1 is added in parallel to three convolutions of different sizes as a residual connection, and a bottleneck structure is adopted, that is, a convolution with a step size of 2 and a size of 1×1 is added before and after the three convolutions. The improved convolution used for recognition prediction is simplified compared to the improved convolution used for training, that is, the ResNet structure is cancelled. The Transformer attention mechanism module is located between the improved downsampling module and the improved upsampling module, and is used to evaluate the fiber bundle features extracted by the improved downsampling module and screen out the features that contribute greatly to the fiber bundle recognition. When the feature map passes through this module, it will first be divided into 64 small patches, flattened and linearly transformed, and the position information will be extracted and the additional learnable features will be embedded. After being scored by the Transformer encoder, the important features will be screened by the multi-perceptron. After the flattened patches are input into the Transformer encoder, they will undergo inter-layer normalization, multi-head attention mechanism, inter-layer normalization and multi-perceptron in sequence. The ResNet structure is used in the first inter-layer normalization, multi-head attention mechanism, second inter-layer normalization and multi-perceptron, respectively. The deconvolution operation is consistent with the deconvolution operation of U-Net, which plays a role in restoring the fiber bundle features and can expand the side length of the feature map to twice the original length.

3. The structural XCT fiber bundle identification and segmentation method based on improved U-Net according to claim 2 is characterized in that: The feature map processed by the improved upsampling module will be restored to a 3-channel RGB image through a 1×1 convolution with a step size of 1. The image size is then adjusted to the size of the input CT image to obtain the final fiber bundle segmentation recognition image.

4. The structural XCT fiber bundle identification and segmentation method based on improved U-Net according to claim 2, characterized in that: After the feature map data is convolved or the feature maps are merged, they need to go through batch normalization and SiLu activation function before the next step of processing.

5. The structural XCT fiber bundle identification and segmentation method based on improved U-Net according to claim 1, characterized in that: In step 1, the fiber bundles were manually labeled in the open source software Labelme to generate a label file. In the dataset, the ratio of training set, test set, and validation set was 3:1:

1.

6. The structural XCT fiber bundle identification and segmentation method based on improved U-Net according to claim 1, characterized in that: In step 2, data enhancement methods include cropping, flipping, rotation, adding noise, brightness change, and contrast enhancement. After enhancement, the ratio of training set, test set, and validation set in the data set should also be controlled at 3:1:

1.

7. The structural XCT fiber bundle identification and segmentation method based on improved U-Net according to claim 2, characterized in that: In step 5, the number of training times is set to 1000 times, and the initial learning rate, first-order moment decay rate, second-order moment decay rate and weight coefficient of the Adam optimizer are set to 0.0005, 0.9, 0.999, and 0.0001 respectively. For linear warm-up learning, the learning rate is gradually increased to the initial value in the first 20 rounds of training. After each training, the corresponding weight file is saved, and the weight file is used to segment the XCT image on the validation set to obtain the three evaluation parameters of precision, recall and average precision.

8. The structural XCT fiber bundle identification and segmentation method based on improved U-Net according to claim 1, characterized in that: In step 6, the weight file generated in step 5 is tested in the test set, and the optimal weight file is selected using the three evaluation parameters of precision, recall and average precision. The file with the best segmentation effect is selected as the optimal weight file to segment all XCT images.

Images