RAS-UNet model-based intelligent identification method for microcosmic pore structure of hollow fiber membrane

Abstract

The invention belongs to the technical field of fiber membrane material intelligent detection, and discloses an RAS-UNet model-based hollow fiber membrane microcosmic hole structure intelligent identification method, which comprises the following steps of: marking holes in a hollow fiber membrane electron microscope image to obtain a hollow fiber membrane microcosmic hole structure data set, establishing an RAS-UNet model, and identifying the hollow fiber membrane microcosmic hole structure data set according to the RAS-UNet model. Training and verifying the RAS-UNet model by using the hollow fiber membrane microscopic hole structure data set, finally inputting the hollow fiber membrane electron microscope image into the verified RAS-UNet model, and outputting a predicted segmented image; the RAS-UNet model comprises an encoder and a decoder, and the encoder and the decoder are combined through jump connection; the encoder comprises a residual convolution module and a down-sampling operation; the decoder comprises a residual error convolution module, an up-sampling operation module and an attention gate module; the jump connection includes an SE attention module. According to the invention, rapid and accurate recognition of the hollow fiber membrane holes can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent detection of fiber membrane materials, and relates to an intelligent identification method for a microscopic pore structure of a hollow fiber membrane based on a RAS-UNet model. Background Art

[0002] As one of the most effective separation methods in the field of mixed gas separation, the gas separation performance of hollow fiber membrane is closely related to the pore structure of the membrane material. However, when analyzing the electron microscope images of hollow fiber membranes, the accurate identification of the pore edges is challenging, and the difference in the depth direction of the pores causes uneven image grayscale distribution, which greatly increases the difficulty of image segmentation and recognition. In addition, the formation mechanism of the pore structure is closely related to the spinning process parameters. Although a large number of studies have focused on the relationship between the spinning process and the performance and macroscopic morphology of hollow fiber membranes, there are relatively few studies on image segmentation and recognition methods for pore microstructures.

[0003] Traditional image segmentation methods, such as threshold segmentation, are difficult to effectively process images of hollow fiber membrane holes with uneven brightness. With the rapid development of artificial intelligence technology, semantic segmentation algorithms based on deep learning have been initially applied in the identification of microscopic components of materials, but research in the field of hollow fiber membrane hole structure identification is still blank. In recent years, deep learning technology has shown significant advantages in the field of image recognition, providing a new technical path to solve the problem of accurate identification of hole structures, so as to achieve high-precision automated segmentation.

[0004] As a convolutional neural network (CNN) architecture designed specifically for biomedical image segmentation, UNet has been proven to be suitable for membrane pore segmentation tasks due to its similarity in features with hollow fiber membrane images. It is particularly noteworthy that the preparation and imaging of hollow fiber membranes is time-consuming and the image annotation process is complicated, which also leads to the general limitation of the size of available data sets. In this case, UNet stands out with its simple network structure and excellent generalization ability, and can achieve ideal segmentation effects even under limited data set conditions. However, UNet's jump connection directly fuses shallow and deep features, and the fusion effect is poor. When dealing with small targets or areas with blurred boundaries, simple downsampling will cause the loss of small target areas, which is not conducive to processing pore structures with diverse shapes.

[0005] To solve the above problems, the UNet++ model was proposed in the literature 1 (UNet++: Redesigning Skip Connections to Exploit Multiscale Features in Image Segmentation[J].IEEE Transactions on Medical Imaging, 2020, 39(6):1856-1867.DOI:10.1109 / TMI.2019.2959609.). This model replaces the original skip connection with a nested structure of multiple convolutional modules and performs supervision at multiple scales, which improves the recognition ability of small structures and complex boundaries. However, UNet++ still passively fuses features and has no ability to actively focus on areas of interest. It has no mechanism to determine which features are more important, and its structure is relatively complex.

[0006] As a key technology in the field of deep learning, the attention mechanism has been widely used in tasks such as target recognition and image classification by enabling the model to adaptively focus on important features of the input data. For example, in Reference 2 (CBAM-Unet++: easier to find the target with the attention module "CBAM"[C] / / 2021 IEEE10th Global Conference on Consumer Electronics (GCCE).0[2025-04-21].DOI:10.1109 / GCCE53005.2021.9622008.), the CBAM module is integrated into the U-Net++ architecture, and the network is guided to focus on the target area by applying attention weights in the channel and spatial dimensions. Although CBAM enhances the feature expression capability, its attention calculation is still limited to the local range when processing complex images, resulting in a decrease in training ability. Therefore, although combining the attention module with the U-Net structure can improve image segmentation performance, there is still room for optimization.

[0007] Therefore, it is of great significance to study an intelligent identification method for the microscopic pore structure of hollow fiber membranes based on the RAS-UNet model to solve the above problems. Summary of the invention

[0008] The purpose of the present invention is to solve the problems existing in the prior art and to provide a method for intelligently identifying the microscopic pore structure of a hollow fiber membrane based on a RAS-UNet model.

[0009] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0010] An intelligent identification method for the microscopic pore structure of hollow fiber membranes based on the RAS-UNet model firstly marks the pores in the electron microscope image of the hollow fiber membrane to obtain the microscopic pore structure data set of the hollow fiber membrane, then establishes the RAS-UNet model, and uses the hollow fiber membrane microscopic pore structure data set to train and verify the RAS-UNet model. Finally, the electron microscope image of the hollow fiber membrane to be identified is input into the trained and verified RAS-UNet model, and the predicted segmented image is output. According to the segmented image, the specific size and distribution of the pores on the cross section of the hollow fiber membrane can be further counted.

[0011] The RAS-UNet model includes an encoder and a decoder, which are connected via a skip connection;

[0012] The encoder consists of four residual convolution modules and four downsampling operations;

[0013] The decoder consists of four residual convolution modules, four upsampling operations, and an attention gate module. The attention gate formula is:

[0014] ;

[0015] ;

[0016] in, is the Sigmoid function, is the ReLU function, T represents transposition, Indicates the network The feature map of the layer, is the final output of the attention gate, W x and W g Both are 1×1 convolutional layers plus BatchNorm2d normalization layers, g i The gate signal g passes through W g The intermediate feature representation after that, b g W g The bias term, is a 1×1 convolutional layer, for The bias term, is the attention weight of the attention gate module; specifically, the attention gate receives two inputs: the feature map and the gate signal g, where It is passed in through the skip connection, indicating the The feature map of the layer, g comes from the upsampling branch in the decoder; W x and W g Both are 1×1 convolutional layers plus BatchNorm2d normalization layers, which are used for And g for dimensionality reduction, gi For g through W g The intermediate feature representation after that, b g W g The bias term of is a 1×1 convolutional layer used to calculate the attention weight, for The bias term of is the attention weight, The final output of the attention gate is obtained after weighted adjustment ;

[0017] The jump connection includes the SE (Squeeze-and-Excitation Module) attention module, which can enhance the feature extraction and utilization capabilities. The formula is:

[0018] ;

[0019] ;

[0020] ;

[0021] in, represents global average pooling, represents an input feature map of size H×W×C, where H, W, and C represent the height, width, and number of channels, respectively. represents the feature map after global average pooling, The size is 1×1×C, is the Sigmoid function, is the ReLU function, represents the first fully connected layer, represents the second fully connected layer, S represents the attention weight of the SE attention module, , , .

[0022] Compared with the traditional segmentation method U-Net model, the RAS-UNet model has better segmentation effect when the grayscale gradient of the hole edge changes, and can well identify the hole edge, reducing the situation of hole connectivity; the original UNet uses a relatively simple downsampling method to gradually reduce the spatial size of the feature map through a series of convolutional layers and pooling layers. Although this method can effectively reduce the resolution while retaining key feature information, the ordinary downsampling process may cause the loss of original information and reduce the segmentation performance. Using residual convolution to replace standard convolution can effectively solve the problem;

[0023] Among them, Figure 4In the structure of the residual convolution module shown in Figure e, there are two 3×3 convolution layers, each followed by batch normalization and ReLU activation function. The first convolution layer is responsible for reducing the feature dimension, and the second convolution layer is responsible for restoring the feature dimension to the original size. The residual connection directly adds the input to the output to achieve residual learning (the input refers to the feature map input entering the current residual module, and the output refers to the final result of the residual connection after two convolution operations in the residual module; the residual connection refers to directly adding the module input to the output, which can alleviate the problem of gradient disappearance or gradient explosion and improve the nonlinear expression ability of the network). This structure enables the network to maintain high accuracy and effectively learn features even when the depth increases.

[0024] As the preferred technical solution:

[0025] The above-mentioned intelligent identification method of hollow fiber membrane microscopic pore structure based on RAS-UNet model has the following specific steps:

[0026] (1) Constructing a dataset: Annotate the holes on the scanning electron microscope images of hollow fiber membranes to construct a dataset of microscopic pore structures of hollow fiber membranes. The dataset is divided into a training set, a validation set, and a test set.

[0027] (2) Establish the RAS-UNet model;

[0028] (3) Set model hyperparameters: set the learning rate, training rounds, batch size, and image scaling of the RAS-UNet model;

[0029] (4) Selecting a loss function to train and verify the model: Using the training set and the validation set to train the RAS-UNet model, and using the test set to test the model, the parameters of the trained and tested model are fixed as the intelligent recognition model for the microscopic pore structure of the hollow fiber membrane, and the output is the predicted segmentation image;

[0030] (5) Output of the recognition model: Compare the model output image with the real image of the test set to obtain the result index of hole recognition;

[0031] (6) Predicting the segmented image: Input the electron microscope image of the hollow fiber membrane to be identified into the trained and verified RAS-UNet model, and output the predicted segmented image.

[0032] In the above-mentioned intelligent identification method of hollow fiber membrane microscopic pore structure based on RAS-UNet model, constructing a hollow fiber membrane microscopic pore structure data set in step (1) means: cropping the scanning electron microscope image of the original hollow fiber membrane, filtering out images with connected pores or the difference between the grayscale mean of the pores and the grayscale mean of the background area is less than 20, and then annotating them through LabelMe; wherein, when the difference between the grayscale mean of the pore area and the grayscale mean of the background area is less than 20, it means that the contrast between the pore and the background is insufficient and it is difficult to accurately identify. At the same time, such pores are considered to be "too shallow", and images containing too shallow pores will be filtered out.

[0033] The resolution of the SEM image of the original hollow fiber membrane is 640×480, which is cropped into small images at a scale of 4×4, and the resolution of each small image is 160×120;

[0034] The hollow fiber membrane microscopic pore structure dataset includes real images (i.e., unlabeled original images) and labeled images, and the real images correspond to the labeled images one by one.

[0035] In the above-mentioned intelligent identification method of hollow fiber membrane microscopic pore structure based on RAS-UNet model, the sample ratio of training set, validation set and test set in step (1) is 6:2:2.

[0036] As described above, the intelligent identification method of hollow fiber membrane microscopic pore structure based on RAS-UNet model, the expression of the loss function in step (4) is as follows:

[0037] ;

[0038] ;

[0039] ;

[0040] Binary Cross Entropy Loss In the calculation of , N is the number of samples, is the true label of the sample, is the predicted probability of the sample; Dice loss In the calculation, A is the binary segmentation result predicted by the model, and B is the actual segmentation area. represents the intersection of the predicted area and the real area; L represents the total loss, and the total loss L is weighted summed. and .

[0041] In the above-mentioned intelligent identification method of hollow fiber membrane microscopic pore structure based on RAS-UNet model, the initial value of the learning rate in step (3) is set to 0.00001, the number of training rounds is 100, the batch size is 8, and the image scaling ratio is 1.

[0042] As described above, the intelligent identification method of hollow fiber membrane microscopic pore structure based on RAS-UNet model has certain limitations in IoU on some data sets, so it is usually combined with Dice coefficient index to provide a more comprehensive semantic segmentation performance evaluation. The calculation formula of the result index of pore identification in step (5) is as follows:

[0043] ;

[0044] ;

[0045] ;

[0046] ;

[0047] Among them, IoU (Intersection over Union) is the intersection over union ratio, a higher IoU indicates that the segmentation performance of this category is better, mIoU (mean Intersection over Union) is the average intersection over union ratio, mPA (mean PixelAccuracy) is the average pixel accuracy, Dice (Dice Coefficient) is the Dice coefficient, TP is the number of correct predictions, FP is the number of incorrect predictions, FN is the number of missed detections, K represents the number of categories, the present invention includes two categories, holes and background, the number of categories is 2, A is the binary segmentation result predicted by the model, and B is the actual segmentation area.

[0048] Principle of the invention:

[0049] The present invention's hollow fiber membrane microscopic hole structure intelligent recognition method based on RAS-UNet model improves the original UNet neural network model, uses residual convolution module to replace basic convolution structure in the network to improve the model's ability to learn spatial features, integrates attention gate mechanism in the decoder structure, and introduces SE attention module in jump connection to enhance feature extraction and utilization capabilities, and constructs a segmentation model suitable for predicting hollow fiber membrane holes. The residual convolution module can alleviate gradient disappearance or gradient explosion. For complex hollow fiber membrane hole images, the network can learn deeper features without losing information; the attention gate mechanism can help the network focus on key areas, and the holes in the image are independent key segmentation targets. Reducing the interference of background areas can improve segmentation accuracy; the SE attention module, as a channel attention mechanism, can make the model pay more attention to the edge area of ​​the hole by learning the importance of different features of each channel, and strengthen the ability to extract hole edge information. Therefore, the intelligent identification method of the microscopic pore structure of the hollow fiber membrane based on the RAS-UNet model of the present invention can effectively improve the accuracy of pore segmentation, and the edges of different holes in the predicted image can be clearly identified.

[0050] Beneficial effects:

[0051] (1) The intelligent identification method of the microscopic pore structure of hollow fiber membranes based on the RAS-UNet model of the present invention improves the original UNet neural network model and constructs a segmentation model suitable for predicting the pores of hollow fiber membranes, thereby solving the complex problem of the segmentation of the pores of hollow fiber membranes. It can more flexibly mine the features of the pores in the image and further improve the prediction accuracy of the algorithm in the segmentation of the pores of hollow fiber membranes. At the same time, through the attention mechanism and convolution module of the present invention, the feature extraction capability is enhanced and the convergence is accelerated, the segmentation performance is improved, the accurate segmentation of the pores of hollow fiber membranes is achieved, and accurate automatic identification is achieved, so as to provide more timely feedback to production and guide actual production.

[0052] (2) The present invention can realize high-precision and rapid identification of the pore structure of hollow fiber membranes, providing reliable technical support for performance optimization and quality control of membrane materials. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 The original scanning electron microscope image of the present invention and the schematic diagram after cutting; in the figure, a is the scanning electron microscope image of the original hollow fiber membrane, and b is a small image cut out according to the scale of 4×4;

[0054] Figure 2 is a schematic diagram of annotated images of the present invention; in the figure, c is Figure 1 The annotated image corresponding to the scanning electron microscope image of the original hollow fiber membrane in d is Figure 1The annotated image corresponding to the cropped small image;

[0055] Figure 3 is a schematic diagram of the attention gate module of the present invention;

[0056] Figure 4 It is a structural diagram of the residual convolution module and the SE attention module of the present invention; in the figure, e is the structural diagram of the residual convolution module, and f is the structural diagram of the SE attention module;

[0057] Figure 5 It is the RAS-UNet model diagram of the present invention;

[0058] Figure 6 The segmented images predicted by the present invention and the prior art models; in the figure, (a) is the original input scanning electron microscope image of the hollow fiber membrane, (b) is the manually annotated pore structure mask image, (c) is the segmented image predicted by the RAS-UNet model proposed in the present invention, (d) is the segmented image predicted by the original UNet model, and (e) is the segmented image predicted by the SegNet model;

[0059] Figure 7 This is a flow chart of the intelligent identification method of the microscopic pore structure of hollow fiber membranes based on the RAS-UNet model of the present invention. DETAILED DESCRIPTION

[0060] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0061] Intelligent identification method of microscopic pore structure of hollow fiber membrane based on RAS-UNet model, such as Figure 7 As shown, the steps are as follows:

[0062] (1) Dataset construction: Holes were annotated on 228 scanning electron microscope images (resolution 640 × 480) of hollow fiber membranes. The original scanning electron microscope images of hollow fiber membranes were then cropped into 16 small images with a resolution of 160 × 120 at a scale of 4 × 4. Images with connected pores or with a difference between the grayscale mean of the pores and the grayscale mean of the background area of ​​less than 20 were screened out and annotated using LabelMe to construct a dataset of microscopic pore structures of hollow fiber membranes. The images were randomly divided into a training set, a validation set, and a test set (one of the small images of a hollow fiber membrane scanning electron microscope image before and after annotation and before and after cropping, such as Figure 1 , Figure 2As shown); the hollow fiber membrane microscopic pore structure dataset includes 1060 real images (i.e., unlabeled original images) and 1060 labeled images, and the sample ratio of the training set, validation set, and test set is 6:2:2;

[0063] (2) Establish the RAS-UNet model;

[0064] like Figure 5 As shown in the figure, the encoder and decoder are combined through skip connections to form a RAS-UNet model; the input of the initial residual convolution module is 3 channels, the output is 64 channels, the number of categories is set to 1, and the features of the input image are extracted from five dimensions in total, the five dimensions refer to the initial input convolution (64 channels), the first layer downsampling (128 channels), the second layer downsampling (256 channels), the third layer downsampling (512 channels), and the fourth layer downsampling (1024 channels);

[0065] The encoder consists of four residual convolution modules and four downsampling operations. The downsampling operation is a 2×2 maximum pooling layer, which can reduce the width and height of the input feature map by half, compress the spatial dimensions layer by layer, and extract higher-level semantic information.

[0066] The decoder consists of four residual convolution modules, four upsampling operations, and an attention gate module. The upsampling operation is bilinear interpolation.

[0067] like Figure 3 As shown, the attention gate formula is:

[0068] ;

[0069] ;

[0070] in, is the Sigmoid function, is the ReLU function, T represents transposition, Indicates the network The feature map of the layer, is the final output of the attention gate, W x and W g Both are 1×1 convolutional layers plus BatchNorm2d normalization layers, g i The gate signal g passes through W g The intermediate feature representation after that, b g W g The bias term, is a 1×1 convolutional layer, for The bias term, is the attention weight of the attention gate module;

[0071] like Figure 4 As shown in f, the jump connection includes the SE attention module, and the formula is:

[0072] ;

[0073] ;

[0074] ;

[0075] in, represents global average pooling, represents an input feature map of size H×W×C, where H, W, and C represent the height, width, and number of channels, respectively (H and W are halved after each downsampling, the change process of H is 160→80→40→20, and the change process of W is 120→60→30→15), represents the feature map after global average pooling, The size is 1×1×C, represents the first fully connected layer, represents the second fully connected layer, S represents the attention weight of the SE attention module, , , .

[0076] (3) Setting model hyperparameters: setting the learning rate, training rounds, batch size, and image scaling ratio of the RAS-UNet model; the initial value of the learning rate is set to 0.00001, the training rounds are 100, the batch size is 8, the image scaling ratio is 1, and the RMSprop optimizer is used;

[0077] (4) Selecting a loss function to train and verify the model: The RAS-UNet model is trained using the training set and the validation set. After each forward propagation during training, the error between the model output and the true label is calculated using the selected loss function, and the parameters in the model are updated through the back-propagation algorithm, thereby continuously optimizing the model's feature extraction and segmentation capabilities. After the model is trained, the model is tested using the test set. The parameters of the trained and tested model are fixed and used as the intelligent recognition model for the microscopic pore structure of the hollow fiber membrane. The output is the predicted segmented image. The expression of the loss function is as follows:

[0078] ;

[0079] ;

[0080] ;

[0081] is the binary cross entropy loss, N is the number of samples, is the true label of the sample, is the predicted probability of the sample; Dice loss In the calculation, A is the binary segmentation result predicted by the model, and B is the actual segmentation area. represents the intersection of the predicted area and the true area; L represents the total loss;

[0082] (5) Output of the recognition model: Compare the image output by the model with the real image of the test set to obtain the result index of hole recognition. The calculation formula of the result index of hole recognition is:

[0083] ;

[0084] ;

[0085] ;

[0086] ;

[0087] Among them, IoU (Intersection over Union) is the intersection over union ratio, mIoU (mean Intersection over Union) is the average intersection over union ratio, mPA (mean Pixel Accuracy) is the average pixel accuracy, Dice (Dice Coefficient) is the Dice coefficient, TP is the number of correct predictions, FP is the number of wrong predictions, FN is the number of missed detections, K represents the number of categories, A is the binary segmentation result predicted by the model, and B is the actual segmentation area;

[0088] (6) Predicting the segmented image: Input the electron microscope image of the hollow fiber membrane to be identified into the trained and verified RAS-UNet model, and output the predicted segmented image.

[0089] In order to verify the above-mentioned intelligent recognition method of hollow fiber membrane microscopic pore structure based on RAS-UNet model, a comparative test of the recognition results of hollow fiber membrane was carried out using different models, and the comparison is as follows:

[0090] Identify the sample as: Figure 6 The image shown in (a);

[0091] The RAS-UNet model trained by the above-mentioned hollow fiber membrane microscopic pore structure intelligent recognition method based on the RAS-UNet model is used to identify the pores of the identified sample. The number of pore recognition is 5 times, and then the average value is taken. The final pore prediction segmentation image is as follows: Figure 6As shown in (c), the hole recognition result indicators are: mPA is 92.52%, Dice is 93.00%, IoU is 86.79%, and mIoU is 85.72%;

[0092] The FCN model described in the literature (Fully Convolutional Networks for Semantic Segmentation [J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 39 (4): 640-651. DOI: 10.1109 / CVPR.2015.7298965.) was used to identify holes in the samples. The hole identification was performed 5 times and then the average value was taken. The hole identification result indicators were: mPA was 90.49%, Dice was 91.12%, IoU was 83.50%, and mIoU was 82.25%.

[0093] The SegNet model described in the literature (SegNet: A Deep Convolutional Encoder-Decoder Architecturefor Image Segmentation[J].IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017:1-1.DOI:10.1109 / TPAMI.2016.2644615.) is used to identify holes in the sample. The hole identification is performed 5 times and then the average value is taken. The final hole prediction segmentation image is as follows: Figure 6 As shown in (e), the hole recognition result indicators are: mPA is 91.13%, Dice is 91.58%, IoU is 84.45%, and mIoU is 83.14%;

[0094] The UNet model described in the literature (U-Net: Convolutional Networks for Biomedical Image Segmentation [C] / / International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer International Publishing, 2015. DOI: 10.1007 / 978-3-319-24574-4_28.) was used to identify holes in the sample. The hole identification was performed 5 times and then the average value was taken. The final hole prediction segmentation image is as follows: Figure 6 As shown in (d), the hole recognition result indicators are: mPA is 91.72%, Dice is 92.36%, IoU is 85.46%, and mIoU is 84.29%;

[0095] Through the above tests, it can be seen that the final hole prediction segmentation image of the RAS-UNet model of the present invention is closer to Figure 6 The manually annotated mask image shown in (b) performs better, which also shows that the RAS-UNet of the present invention performs better in processing edge details and shallow hole areas, and can more accurately identify and segment difficult-to-distinguish hole areas. Compared with the existing technology, it has obvious advantages in accurately segmenting hole areas.

Claims

1. An intelligent identification method for the microscopic pore structure of hollow fiber membranes based on the RAS-UNet model, characterized by: Firstly, the holes in the electron microscope image of the hollow fiber membrane are annotated to obtain the microscopic pore structure dataset of the hollow fiber membrane. Then, the RAS-UNet model is established and trained and verified using the microscopic pore structure dataset of the hollow fiber membrane. Finally, the electron microscope image of the hollow fiber membrane to be identified is input into the trained and verified RAS-UNet model to output the predicted segmentation image. The RAS-UNet model includes an encoder and a decoder, which are connected via a skip connection; The encoder consists of four residual convolution modules and four downsampling operations; The decoder consists of four residual convolution modules, four upsampling operations, and an attention gate module. The attention gate formula is: ； ； in, is the Sigmoid function, is the ReLU function, Indicates the network The feature map of the layer, is the final output of the attention gate, W x and W g Both are 1×1 convolutional layers plus BatchNorm2d normalization layers, g i The gate signal g passes through W g The intermediate feature representation after that, b g W g The bias term, is a 1×1 convolutional layer, for The bias term, is the attention weight of the attention gate module; The skip connection includes the SE attention module, the formula is: ； ； ； in, represents global average pooling, represents an input feature map of size H×W×C, where H, W, and C represent the height, width, and number of channels, respectively. represents the feature map after global average pooling, represents the first fully connected layer, represents the second fully connected layer, S represents the attention weight of the SE attention module, , , .

2. According to claim 1, the method for intelligent identification of hollow fiber membrane microscopic pore structure based on RAS-UNet model is characterized in that: The specific steps are as follows: (1) Constructing a dataset: Annotate the holes on the scanning electron microscope images of hollow fiber membranes to construct a dataset of microscopic pore structures of hollow fiber membranes. The dataset is divided into a training set, a validation set, and a test set. (2) Establish the RAS-UNet model; (3) Set model hyperparameters: set the learning rate, training rounds, batch size, and image scaling of the RAS-UNet model; (4) Selecting a loss function to train and verify the model: Using the training set and the validation set to train the RAS-UNet model, and using the test set to test the model, the parameters of the trained and tested model are fixed as the intelligent recognition model for the microscopic pore structure of the hollow fiber membrane, and the output is the predicted segmentation image; (5) Output of the recognition model: Compare the model output image with the real image of the test set to obtain the result index of hole recognition; (6) Predicting the segmented image: Input the electron microscope image of the hollow fiber membrane to be identified into the trained and verified RAS-UNet model, and output the predicted segmented image.

3. According to claim 2, the method for intelligent identification of microscopic pore structure of hollow fiber membrane based on RAS-UNet model is characterized in that: In step (1), constructing a hollow fiber membrane microscopic pore structure dataset means: cropping the scanning electron microscope image of the original hollow fiber membrane, filtering out images with connected pores or with a difference between the grayscale mean of the pores and the grayscale mean of the background area of ​​less than 20, and then annotating them through LabelMe; The resolution of the SEM image of the original hollow fiber membrane is 640×480, which is cropped into small images at a scale of 4×4, and the resolution of each small image is 160×120; The hollow fiber membrane microscopic pore structure dataset includes real images and annotated images.

4. The method for intelligently identifying the microscopic pore structure of hollow fiber membranes based on the RAS-UNet model according to claim 2 is characterized in that: The sample ratio of the training set, validation set, and test set in step (1) is 6:2:

2.

5. The method for intelligently identifying the microscopic pore structure of hollow fiber membranes based on the RAS-UNet model according to claim 2 is characterized in that: The expression of the loss function in step (4) is as follows: ； ； ； N is the number of samples, is the true label of the sample, is the predicted probability of the sample; A is the binary segmentation result predicted by the model, and B is the actual segmentation area. represents the intersection of the predicted area and the true area; L represents the total loss.

6. The method for intelligently identifying the microscopic pore structure of hollow fiber membranes based on the RAS-UNet model according to claim 2 is characterized in that: In step (3), the initial value of the learning rate is set to 0.00001, the number of training rounds is 100, the batch size is 8, and the image scaling factor is 1.

7. The method for intelligently identifying the microscopic pore structure of hollow fiber membranes based on the RAS-UNet model according to claim 2 is characterized in that: The calculation formula for the result index of hole identification in step (5) is as follows: ； ； ； ； Among them, IoU is the intersection over union ratio, mIoU is the mean intersection over union ratio, mPA is the average pixel accuracy, Dice is the Dice coefficient, TP is the number of correct predictions, FP is the number of wrong predictions, FN is the number of missed detections, and K represents the number of categories.

Images