3D printing concrete interface pore structure visual segmentation method based on point cloud

Abstract

The invention discloses a 3D printing concrete interface pore structure visual segmentation method based on point cloud. The method comprises the following steps: 1, collecting point cloud data; 2, preprocessing the collected point cloud data; 3, dividing the point cloud data set into a training set, a verification set and a test set; 4, performing point cloud data sampling on the training set of the point cloud data set in the step 3 to obtain a local relation; 5, building a feature detail extraction network in the 3D printing concrete interface pore structure visual segmentation network model PointESamba by using a multi-layer sensor, and carrying out detail extraction on sampled point cloud data features of the pore structure; step 6, optimizing the disorder of the point cloud through an adaptive sorting algorithm; and step 7, obtaining a most adaptive sequence by using the sorting algorithm in the step 6, obtaining global features by using a Mama module, further segmenting and identifying the features extracted in the step 5, and completely constructing PointESamba. According to the invention, the segmentation precision and capability of the point cloud on the pore structure are enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of visual segmentation of concrete interface pore structure, and in particular to a point cloud-based 3D printing concrete interface pore structure visual segmentation method. Background Art

[0002] In the actual concrete 3D printing process, affected by the printing materials, printing equipment, printing process and environmental conditions, the concrete structure will inevitably have defects such as interlayer deformation, local offset and surface cracks. Among them, pore defects are the most common defects. Pore defects will affect the appearance and performance of the printed structure at different levels in the macro and micro. At the macro level, it is manifested as an impact on the appearance of the printed structure, and the printed product is not airtight. At the micro level, it is manifested as an impact on mechanical anisotropy, resulting in greatly different axial bearing capacities of the printed product. It may also cause cracks and water seepage in the components during use, seriously affecting the safety and service life of the building, and even causing the failure of the entire printing process. Therefore, the identification of pore structure of 3D printed concrete is very necessary for the quality inspection of concrete 3D printing.

[0003] A method for detecting pores in a concrete 3D printing interface has been proposed (publication number: CN116863236A). This algorithm combines traditional camera images with deep learning. It is the first attempt to use deep learning methods for pore structure detection. However, traditional images are difficult to provide the necessary spatial feature information and cannot detect pores that are too small due to pixel problems. Summary of the invention

[0004] In order to overcome the above technical problems, the purpose of the present invention is to provide a point cloud-based 3D printed concrete interface pore structure visual segmentation method, which combines machine vision theory, deep learning method, point cloud method and 3D printed concrete pore structure segmentation technology, and produces a 3D printed concrete interface pore structure point cloud dataset according to engineering applications. Based on PointNet, a 3D printed concrete interface pore structure visual segmentation model PointESamba based on local feature enhancement and adaptive sorting of the Mmaba model is constructed, which enhances the segmentation accuracy and ability of the point cloud for the pore structure, so as to achieve the purpose of promoting the sustainable development of 3D printed concrete technology in the construction manufacturing industry.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A point cloud-based 3D printed concrete interface pore structure visual segmentation method comprises the following steps;

[0007] Step 1: Use a laser scanner to collect point cloud data of the pore structure of the 3D printed concrete interface;

[0008] Step 2: preprocessing the collected point cloud data;

[0009] Step 3: Generate a point cloud dataset of pore structure of 3D printed concrete interface based on the preprocessed point cloud data, and divide the point cloud dataset into a training set, a validation set, and a test set;

[0010] Step 4: Through the local feature enhancement sampling strategy, the optimization of point cloud data sampling is difficult to play a role in the small local features, and the point cloud data training set of step 3 is sampled to obtain the local relationship;

[0011] Step 5: Use a multi-layer perceptron to construct a feature detail extraction network in PointESamba, a visual segmentation network model for the pore structure of the 3D printed concrete interface, and extract details from the sampled point cloud data features of the pore structure;

[0012] Step 6: Optimize the point cloud disorder through adaptive sorting algorithm; improve the subsequent network learning ability.

[0013] Step 7: Use the sorting algorithm in step 6 to obtain the most suitable sorting, use the Mamba module to obtain global features, and then further segment and identify the features extracted in step 5 to fully construct PointESamba.

[0014] Step 8: Use the pore structure point cloud dataset training set produced in step 3 to train PointESamba;

[0015] Step 9: Use the validation set and test set in step 3 to verify and predict PointESamba, and evaluate the model performance with relevant evaluation indicators.

[0016] Optionally, step 1 specifically includes: using a laser scanner to perform point cloud scanning on a 3D printed concrete test block component printed in a 3D printed concrete laboratory, collecting interface pore structure, deleting background invalid point clouds, and completing point cloud data collection.

[0017] Optionally, step 2 specifically includes: performing data preprocessing on the collected interface pore structure point cloud data, including point cloud denoising and filtering, performing average distance statistics on the nearest 300 neighbor points through the KD tree, points above the average distance will be regarded as noise points, and the noise points will be removed to complete the point cloud data preprocessing.

[0018] Optionally, step 3 specifically includes: using CloudCompare to annotate the interface pore structure data, and perform surface normal vector estimation and point cloud color information acquisition, and dividing the annotated data into three categories: train (training set), val (validation set), and test (test set). The data in the train is used to train the model, the data in the val is used to verify the model, and the data in the test is used to test the model.

[0019] Optionally, the step 4 uses a local feature enhancement sampling strategy to optimize point cloud sampling as follows:

[0020] The farthest point sampling FPS is first performed on the train set data to obtain representative points and the K nearest neighbor algorithm KNN is used to obtain local point grouping. The features of these points in the group are concatenated with the features of the center point to generate input features for the graph neural network.

[0021] On this basis, a graph convolutional network is used to further extract and aggregate local feature relationships within the group;

[0022] Specifically, the local feature enhancement sampling module consists of three algorithm modules: the farthest point sampling algorithm FPS, the K nearest neighbor algorithm KNN, and the graph convolutional network. Its detailed construction is as follows:

[0023] The first part is FPS and KNN with intra-group feature splicing; the second part is the graph convolutional network consisting of two convolutional layer sequences, each of which is composed of two convolutional layers, a batch normalization layer and an activation function.

[0024] Optionally, the step 5 specifically includes:

[0025] The feature detail extraction network adopts the multi-layer perceptron module in PointNet. The multi-layer perceptron module consists of two fully connected layers and an activation function Gelu. It is used to build the feature detail extraction network layer in PointESamba to extract details of the sampled point cloud data features of the pore structure.

[0026] Optionally, the adaptive sorting algorithm in step 6 is specifically:

[0027] Two nonlinear projection layers are used to map the local features and global features within the group to the same feature space, and the cosine similarity between the two is calculated as the adaptability score. Accurate and effective bidirectional point cloud sorting is provided according to the adaptability score.

[0028] Optionally, the step 7 specifically includes:

[0029] The interface pore structure segmentation network PointESamba is further constructed. The feature extraction information in step 5 is used through the adaptive sorting algorithm in step 6 to obtain the adaptability score of the point cloud data. A point cloud data feature sorting module is built. Then, the Mamba network module is constructed with a normalization layer, a selective SSM state space model, deep convolution and residual connection. The global features are obtained after the adaptability is sorted by the adaptability score as the input of the Mamba network module.

[0030] Finally, a segmentation module is constructed to connect the global features and the sampled local features. The segmentation and recognition module is composed of the relu activation function, dropout layer, convolution and softmax classification function to perform pore segmentation.

[0031] Optionally, the step 8 specifically includes: training the 3D printed concrete interface pore structure visual segmentation network model PointESamba by using the training set, and obtaining an optimal weight file when the training is completed, in which important parameters for pore detection are stored.

[0032] Optionally, the relevant evaluation indicators of step 9 specifically include:

[0033] Accuracy: From the perspective of samples, it refers to the proportion of correct samples to the total number of samples. MIOU: It measures the average similarity between the predicted segmentation results and the true segmentation results.

[0034] The model is verified and tested using the validation set and test set in step 3, using the optimal weight file obtained after model training, and the above two evaluation indicators are used for test analysis.

[0035] Beneficial effects of the present invention:

[0036] (1) The present invention proposes a point cloud-based 3D printed concrete interface pore structure visual segmentation method, which aims to make up for the fact that traditional cameras cannot provide 3D printed concrete spatial feature information, solve the problem of difficulty in identifying and segmenting pore structure features in multi-pore dense scenes under complex backgrounds of 3D printed concrete interfaces, and improve the problem of insufficient model learning ability caused by point cloud disorder.

[0037] (2) The present invention produces a point cloud dataset of pore structure of 3D printed concrete interface according to practical applications, and constructs a visual segmentation model PointESamba of pore structure of 3D printed concrete interface based on PointNet, which is based on local feature enhancement and adaptive sorting of Mmaba model. The segmentation accuracy is improved to realize the intelligent, smart and automated segmentation technology of pore segmentation of 3D printed concrete interface, with the accuracy increased by 7.1% and MIOU increased by 22.5%. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a point cloud data set of pore structure of 3D printed concrete interface of the present invention.

[0039] Figure 2 It is a data processing flow chart of the present invention.

[0040] Figure 3 It is a flow chart of the local feature enhancement sampling algorithm of the present invention.

[0041] Figure 4 It is a flow chart of the adaptive sorting strategy algorithm of the present invention.

[0042] Figure 5 It is a structural diagram of the Mamba network model of the present invention.

[0043] Figure 6 The present invention obtains the 3D printed concrete interface pore structure segmentation model PointESamba. DETAILED DESCRIPTION

[0044] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0045] This embodiment discloses a method for visual segmentation of pore structure of 3D printed concrete interface based on point cloud, which specifically includes the following steps:

[0046] Step 1, using a laser scanner to collect point cloud data of the pore structure of the 3D printed concrete interface. In step 1, a laser scanner (model Freescan combo) is used to perform point cloud scanning on the 3D printed concrete test block components printed in the 3D printed concrete laboratory, with an accuracy of 0.1mm, using 26-line and 7-line mixed scanning, collecting the interface pore structure, deleting the background invalid point cloud for non-concrete point cloud data, completing point cloud data collection, and saving it as a ply file.

[0047] Step 2, preprocessing the collected data in step 1;

[0048] In step 2, data preprocessing is performed on the collected interface pore structure point cloud data, including point cloud noise reduction and filtering. The specific method of identifying noise points is to perform average distance statistics on the nearest 300 neighbor points through the KD tree, and points with a distance higher than the average will be regarded as noise points. The point cloud data preprocessing is completed by removing noise points.

[0049] Step 3, based on the preprocessed data in step 2, a point cloud dataset of pore structure of 3D printed concrete interface is prepared, which is divided into a training set, a validation set and a test set;

[0050] In step 3, CloudCompare is used to annotate the interface pore structure data. In the label Label, 0 indicates non-pore and 1 indicates pore. The surface normal vector estimation information is obtained and finally saved in the format of x, y, z, Nx, Ny, Nz, Label. The data set example is as follows: Figure 1 As shown in the figure, the labeled data is divided into three categories according to the ratio of 8:1:1: train (training set), val (validation set), and test (test set). The data in the train is used to train the model, the data in the val is used to validate the model, and the data in the test is used to test the model. The data processing flow is as follows Figure 2 shown.

[0051] Step 4: Design a local feature enhancement sampling strategy to optimize point cloud sampling, which is difficult to play a role on small local features. Sample point cloud data for the training set of the step 3 data set to obtain local relations.

[0052] The main construction of the local feature enhancement sampling module in step 4 is as follows: the first part uses the farthest point sampling FPS to obtain representative points and the K nearest neighbor algorithm KNN to obtain local point set grouping for the train training set data, and finds the k nearest neighbor points of each point, and concatenates the features of these points with the features of the center point to generate input features for the graph neural network.

[0053] Specifically, the dimension information (B, N, D) of the input data is obtained, where B represents the batch size, which is set to 32, N represents the number of points, which is set to 2048, and D is the feature dimension of each point, with an initial input of 10; the group center is selected through FPS to obtain the coordinates of the sampled center point (B, G, D), where B represents the batch size, which is set to 32, G represents the number of sampled center points, which is set to 64, and D is the feature dimension of each point, which is set to 6; KNN is used to find the nearest neighbor point of each point in the input group center, and K is set to represent the nearest neighbor point that needs to be found for each point. The number of nearest neighbors, K is set to 20, and the data dimension information (B, N, K, D) is obtained, where B represents the batch size, which is set to 32, N represents the number of points, which is set to 64, K is the number of nearest neighbors of each point, and D is the feature dimension of each point, which is set to 6. The feature difference between each point and the neighboring point is calculated and concatenated with the original feature to obtain the feature information (B, N, K, 2*D); the second part is to use the graph convolutional network to further extract and aggregate the local feature relationship within the group on this basis. The graph convolutional network consists of two convolutional layer sequences, each of which consists of two 1 D convolutional layer, batch normalization layer BN and activation function Relu. Specifically, the feature information (B, N, K, 2*D) is converted into (B*N, K, 2*D) to conform to Pytorch tensor calculation, and sent to the first convolutional layer conv1, after 1D convolutional layer-BN layer-Relu-1D convolutional layer, input channel 12, output channel 256, convolution kernel size is 1, output shape is (B*N, 256, K); perform the maximum pooling operation in the K dimension to obtain the global feature global_feature, the shape is (B*N, 256, 1 ) represents the global feature of each point. The global_feature is repeated K times and concatenated with the original feature to obtain a feature shape of (B*N, 256+256, K) and sent to the second convolution layer conv2. After 1D convolution layer-BN layer-Relu-1D convolution layer, the input channel is 512, the output channel is 384, the convolution kernel size is 1, and the output shape is (B*N, 384, K). After maximum pooling, the feature shape becomes (B, N, 384), that is, the features of each point are aggregated into a vector. The algorithm flow chart is as follows: Figure 3 shown.

[0054] Step 5, using a multi-layer perceptron to construct a feature detail extraction network in PointESamba, and extracting details from the sampled point cloud data features of the pore structure;

[0055] In step 5, a feature detail extraction network is constructed based on PointNet, using the main multi-layer perceptron module in PointNet, which is mainly composed of two fully connected layers and an activation function Gelu;

[0056] Specifically, a multi-layer perceptron is constructed to build a feature detail extraction layer, and the group center coordinates are used as input. The first part of the feature detail extraction layer is a fully connected layer with 6 input channels and 128 output channels; the second part is a Gelu activation function; the third part is a fully connected layer with 128 input channels and 384 output channels. The details of the point cloud data features after pore structure sampling are extracted to construct a feature extraction module.

[0057] Step 6: Design an adaptive sorting algorithm to optimize the point cloud disorder and improve the subsequent network learning ability.

[0058] The step 6 is an adaptive sorting algorithm, the algorithm flow chart is as follows Figure 4 As shown, two nonlinear projection layers are used to map the local features and global features within the group to the same feature space, and the cosine similarity between the two is calculated to predict the fitness score, and the bidirectional point cloud sorting is provided according to the fitness score;

[0059] Through the adaptive sorting algorithm, the point cloud data after feature extraction is sorted, and the originally disordered point cloud is given an adaptability score. It is sorted in ascending and descending order according to the adaptability score, and then the point cloud data organization method is optimized. The point cloud is no longer disordered and scattered, but the point cloud data with higher adaptability scores is processed first to ensure that the key features are highlighted, so that the sorted data can better represent the overall structure of the point cloud, taking into account not only local features but also global features, thereby enhancing the subsequent model learning ability.

[0060] Specifically, the local feature projection layer within the group and the global feature projection layer are used to respectively project the local features within the group And the global feature q of all point cloud data is projected into the same feature space to obtain and On this basis, the local and global vector dot product Divide by the product of the modulus|||| To calculate the cosine similarity Cos between points g :

[0061]

[0062] Cosine similarity can reflect the angle between two vectors. Its value range is [-1, 1]. The larger the cosine similarity, the more similar the two vectors are. The cosine similarity metric can be used to measure the correlation between the local embedding and the global features in each group.

[0063] After obtaining the cosine similarity, calculate the corresponding fitness score S of the local embedding in each group g , by minimizing the fitness score loss function Loss adaptation Learn to make Sg Close to Cos g :

[0064]

[0065] Among them, Loss adaptation Represents the overall loss, which is used to optimize the model to make the learned fitness score more accurate. N represents the number of point clouds, G represents the number of groups, and Loss contrast represents the contrastive loss, which is used to evaluate the difference between the predicted fitness score and the true similarity.

[0066] The calculated fitness score S g Then, the embedded feature data in the local group is sorted and bidirectionally sorted according to the adaptive sorting strategy:

[0067] Sort from high to low: local features within the group Sort from high to low to get the sorted local group embedding features: g=1,2,…,G

[0068] Sort from low to high: local features within the group Sort from low to high to get the sorted local group embedding features: g=1,2,…,G

[0069] Finally, the two are combined to obtain a bidirectional adaptive ranking: the local intra-group embedding features sorted from high to low and from low to high are merged to obtain: g=1,2,…,2G

[0070] Finally, the adaptability score is weighted and aggregated with the corresponding local group features through the pooling operation to obtain the global features, which enhances the expressiveness of the global features and facilitates the subsequent segmentation tasks:

[0071]

[0072] Step 7, use the sorting algorithm in step 6 to obtain the most suitable sorting, use the Mamba module to obtain global features, and then further segment and identify the features extracted in step 5 to completely construct PointESamba;

[0073] The step 7 further constructs PointESamba completely, obtains the adaptability score of the point cloud data through the adaptive sorting algorithm in step 6 by using the feature extraction information in step 5, builds a data sorting module, and uses the adaptability score to sort the features as the input of the Mamba network module. The N*Mamba network module is constructed with the normalization layer LN, the selective state space model SSM, the deep convolution DW and the residual connection. Specifically, the execution order of the Mamba network module is LN normalization layer, fully connected layer Linear, deep convolution DW, SiLU activation function, selective state space model SSM, and then residual connection. There are two ways:

[0074] Unfused path: First perform the trunk order DropPath regularization, then add the residual branch fully connected layer features, add Add, and then perform the normalized fully connected operation Linear.

[0075] Fusion path: Fusion the Add and full connection operations into one operation before outputting.

[0076] The Mamba network model structure is as follows Figure 5 shown.

[0077] Finally, the segmentation module is constructed, which mainly consists of three parts: feature processing and global feature construction, feature propagation, and convolution segmentation. Specifically, feature processing and global feature construction are processed by multiple layers of N*Mamba modules to obtain relevant information features. These features are normalized and transposed, and all features are connected together to obtain spliced ​​features (B, 1152, G). Then, maximum pooling and average pooling are performed, and global information is obtained after splicing; convolution segmentation is performed through relu activation function, Dropout layer, deep convolution and softmax classification function to perform pore segmentation, upsampling through 3 convolutions, segmentation through softmax, and the PointESamba model is completely constructed as shown in the figure. Figure 6 shown.

[0078] Specifically, the training set data of the pore structure point cloud dataset of the 3D printed concrete interface is input, and after the local feature enhancement sampling strategy (FPS-KNN-graph convolutional network), the feature information is obtained and sent to the feature detail extraction network MLP to extract the feature details, and then the unordered point cloud data is adaptively sorted in two directions to obtain the adaptability score. The ordered point cloud is then sent to the Mamba module for feature extraction, and the global information is obtained through maximum pooling and average pooling splicing, and finally segmented through three-layer convolution and softmax.

[0079] Step 8: Use the pore structure point cloud dataset training set produced in step 3 to train the 3D printed concrete interface pore structure visual segmentation network model constructed in step 7.

[0080] Step 9: Use the validation set and test set in step 3 to verify and predict the trained model, and use relevant evaluation indicators to evaluate the model performance.

[0081] The final experimental results, compared with pointnet++:

[0082] Models Acc MIou(%) PointNet++ 77.4 41.6 PointESamba 84.5 64.1

[0083] In summary, the PointESamba visual segmentation model for 3D printed concrete interface pore structure proposed in this invention realizes intelligent, precise and automated segmentation of interface pore structure by significantly improving segmentation accuracy. Experimental results show that the accuracy is improved by 7.15% and the MIOU is improved by 22.8%. The model is superior to the existing methods in terms of accuracy and average intersection-over-union ratio, providing strong support for further promoting the intelligent analysis and engineering application of 3D printed concrete structures. At the same time, the efficiency and reliability of this method have laid a solid foundation for in-depth research and practice in related fields.

Claims

1. A point cloud-based visual segmentation method for pore structure of 3D printed concrete interface, characterized in that: The steps include: Step 1: Use a laser scanner to collect point cloud data of the pore structure of the 3D printed concrete interface; Step 2: preprocessing the collected point cloud data; Step 3: Generate a point cloud dataset of pore structure of 3D printed concrete interface based on the preprocessed point cloud data, and divide the point cloud dataset into a training set, a validation set, and a test set; Step 4: Through the local feature enhancement sampling strategy, the optimization of point cloud data sampling is difficult to play a role in the small local features, and the point cloud data training set of step 3 is sampled to obtain the local relationship; Step 5: Use a multi-layer perceptron to construct a feature detail extraction network in PointESamba, a visual segmentation network model for the pore structure of the 3D printed concrete interface, and extract details from the sampled point cloud data features of the pore structure; Step 6: Optimize the point cloud disorder through adaptive sorting algorithm; improve the subsequent network learning ability. Step 7: Use the sorting algorithm in step 6 to obtain the most suitable sorting, use the Mamba module to obtain global features, and then further segment and identify the features extracted in step 5 to fully construct PointESamba.

2. According to the point cloud-based 3D printed concrete interface pore structure visual segmentation method of claim 1, it is characterized in that: The step 1 specifically includes: using a laser scanner to perform point cloud scanning on a 3D printed concrete test block component printed in a 3D printed concrete laboratory, collecting the interface pore structure, deleting background invalid point clouds, and completing point cloud data collection.

3. According to the point cloud-based 3D printed concrete interface pore structure visual segmentation method of claim 2, it is characterized in that: The step 2 specifically includes: performing data preprocessing on the collected interface pore structure point cloud data, including point cloud denoising and filtering, performing average distance statistics on the nearest neighbor points through the KD tree, and points above the average distance will be regarded as noise points, and the noise points will be removed to complete the point cloud data preprocessing.

4. According to the point cloud-based 3D printed concrete interface pore structure visual segmentation method of claim 3, it is characterized in that: The step 3 specifically includes: using CloudCompare to annotate the interface pore structure data, estimate the surface normal vector and obtain the point cloud color information, and divide the annotated data into three categories: train, val, and test. The data in the train is used to train the model, the data in the val is used to verify the model, and the data in the test is used to test the model.

5. According to the point cloud-based 3D printed concrete interface pore structure visual segmentation method of claim 4, it is characterized in that: The step 4 uses the local feature enhancement sampling strategy to optimize the point cloud sampling as follows: The farthest point sampling FPS is first performed on the train set data to obtain representative points and the K nearest neighbor algorithm KNN is used to obtain local point grouping. The features of these points in the group are concatenated with the features of the center point to generate input features for the graph neural network. On this basis, a graph convolutional network is used to further extract and aggregate local feature relationships within the group; The local feature enhancement sampling module consists of three algorithm modules: the farthest point sampling algorithm FPS, the K nearest neighbor algorithm KNN, and the graph convolutional network. Its detailed construction is as follows: The first part is FPS and KNN and intra-group feature concatenation; The second part is that the graph convolutional network consists of two convolutional layer sequences, each of which consists of two convolutional layers, a batch normalization layer and an activation function.

6. The method for visual segmentation of pore structure of 3D printed concrete interface based on point cloud according to claim 5, characterized in that: The step 5 specifically includes: The feature detail extraction network adopts the multi-layer perceptron module in PointNet. The multi-layer perceptron module consists of two fully connected layers and an activation function Gelu. It is used to build the feature detail extraction network layer in PointESamba to extract details of the sampled point cloud data features of the pore structure.

7. The method for visual segmentation of pore structure of 3D printed concrete interface based on point cloud according to claim 6, characterized in that: The adaptive sorting algorithm in step 6 is specifically: Two nonlinear projection layers are used to map the local features and global features within the group to the same feature space, and the cosine similarity between the two is calculated as the adaptability score. Accurate and effective bidirectional point cloud sorting is provided according to the adaptability score.

8. The method for visual segmentation of pore structure of 3D printed concrete interface based on point cloud according to claim 7, characterized in that: The step 7 specifically includes: Construct the interface pore structure segmentation network PointESamba, use the feature extraction information in step 5 through the adaptive sorting algorithm in step 6 to obtain the adaptability score of the point cloud data, build a point cloud data feature sorting module, and then construct the Mamba network module with normalization layer, selective SSM state space model, deep convolution and residual connection. After adaptability sorting by adaptability score, it is used as the input of the Mamba network module to obtain the global feature; Finally, a segmentation module is constructed to connect the global features and the sampled local features. The segmentation and recognition module is composed of the relu activation function, dropout layer, convolution and softmax classification function to perform pore segmentation.

Images