An essential feature-guided small sample image segmentation enhancement method

Through a priori embedded edge features and multi-branch batch standardized structure, the edge blur and detail loss problems of image segmentation model under small data sets are solved, which improves the robustness and accuracy of the model and improves the overall performance of image segmentation.

CN118262106BActive Publication Date: 2024-11-19NANJING UNIV OF AERONAUTICS & ASTRONAUTICS +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410264757.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-03-08
Publication Date
2024-11-19
Estimated Expiration
2044-03-08

AI Technical Summary

Technical Problem

Under the conditions of small datasets, deep learning models are prone to edge blur and detail loss in image semantic segmentation, and existing data augmentation methods are difficult to learn unified distributed features, resulting in loss of model accuracy.

Method used

Edge features are introduced through the prior embedded feature enhancement module, and the consistency features under different distribution domains are learned by combining multi-branch batch standardized structures to enhance the robustness and accuracy of the model for image segmentation.

Benefits of technology

Improves the accuracy and robustness of the model in small sample segmentation tasks, improves overall performance, and enhances the detail retention and edge clarity of image segmentation results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118262106B_ABST
    Figure CN118262106B_ABST
Patent Text Reader

Abstract

The present invention discloses a small sample image segmentation enhancement method guided by essential features, comprising: dividing a small sample image data set into a training set and a verification set according to an appropriate ratio; inputting the training set into a priori embedded feature enhancement module, and enhancing the input image with the contextual semantic information contained in the edge feature as a priori; inputting the enhanced input image into a multi-distribution feature learning module, performing multi-distribution offset on the enhanced input image, and using a multi-branch batch normalization structure to perform parallel training on different feature distributions, and learn consistent features under different distribution domains; extracting features and restoring the upper convolutional layer of an image segmentation network with a mainstream encoding and decoding structure for inputting feature maps, and performing loss with segmentation labels, training and repeated iterations to converge the loss function, and obtain network model parameters; and performing reasoning through a trained segmentation network model to obtain a segmentation result. The present invention effectively enhances the segmentation accuracy of a small sample data set.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of image segmentation in intelligent computer vision processing, and in particular relates to a small sample image segmentation enhancement method guided by essential features. Background Art

[0002] The purpose of image semantic segmentation is to semantically annotate each pixel of an image and segment it into several pixel regions with specific semantics, so as to interpret and analyze the image data. Therefore, building a high-precision image segmentation model has far-reaching significance, both in helping scholars understand the intrinsic nature of images and in promoting practical applications in various fields.

[0003] Recently, deep learning-based methods have achieved significant results in image processing and semantic understanding, and promoted the development of image semantic segmentation, but they often suffer from problems such as blurred edges and lost details when applied to small datasets. Moreover, in the industrial and medical fields, it is usually difficult to collect a large number of image samples, and the annotation cost is high. In this case, insufficient samples highlight the necessity of thoroughly understanding and effectively utilizing image features based on limited data. On the one hand, there are usually some key features that can provide important semantic and contextual information, such as edges, textures, and shapes. These features have a dominant role in image segmentation results. In contrast, some other features, such as noise, image brightness, and contrast, have less impact on segmentation results. Obviously, enhancing key features helps deep learning-based models achieve more accurate segmentation. Here we call such features "task-related features". On the other hand, image enhancement is indispensable in small-scale datasets. Although common image enhancement methods (such as rotation, flipping, cropping, and merging) can enrich the training set, they affect the texture distribution characteristics of the image. Existing models usually input these inconsistently distributed datasets into a single batch normalization. This operation ignores the interference information between different distributions and makes it difficult to accurately learn the "uniform distribution features". In addition, smaller datasets make it more difficult to find invariant information after augmentation. All these problems will result in a loss in model accuracy.

[0004] The prior art proposes a small sample image segmentation method based on a guided network and a fully connected conditional random field. The idea is to infer the potential features of the supporting image by optimizing the guided network; perform a preliminary segmentation of the query image without pixel annotations based on the potential features; and perform a more refined segmentation through a fully connected conditional random field based on the preliminary segmentation results. The prior art also provides a lightweight multi-scale feature enhancement network Lite-FENet, which extracts multi-scale feature representations with rich contextual information and generates descriptive features for predicting pixels. It can effectively alleviate the deviation between the supporting samples and the query samples, and can also replace the back-end feature interaction module of the small sample image segmentation method to further improve performance. The prior art also provides a small sample image segmentation method with self-compensatory learning, which optimizes image segmentation by constructing a self-compensation mechanism through a feature fusion strategy for prototype generation masks and multi-layer masks. Summary of the invention

[0005] In order to solve the above technical problems, the present invention proposes a small sample image segmentation enhancement method guided by essential features, which can effectively enhance the robustness and accuracy of the segmentation model.

[0006] To achieve the above object, the present invention provides a small sample image segmentation enhancement method guided by essential features, comprising:

[0007] Obtain a small sample image data set to be segmented, and divide the small sample image data set into a training set and a validation set according to an appropriate ratio;

[0008] Inputting the training set into a priori embedded feature enhancement module, and enhancing the input image using contextual semantic information contained in the edge features as a priori;

[0009] Inputting the enhanced input image into a multi-distribution feature learning module, performing multi-distribution shift on the enhanced input image, and then using a multi-branch batch normalization structure to perform parallel training on different feature distributions, learn consistent features under different distribution domains, and obtain a feature map;

[0010] The feature map is input into the image segmentation network of the mainstream codec structure for feature extraction and upper convolution layer restoration, and the loss function is made into a loss function with the segmentation label, and the loss function is trained and iterated repeatedly to converge, obtain the network model parameters, and obtain the trained segmentation network model;

[0011] The segmentation result is obtained by performing reasoning through the trained segmentation network model.

[0012] Optionally, the a priori embedded feature enhancement module includes:

[0013] Extract edges from input images using edge detection algorithms;

[0014] The edge extraction result is embedded into the input image with a hyperparameter q.

[0015] Optionally, the method of extracting edges from the input image using an edge detection algorithm is:

[0016]

[0017] S(x,y)=G(x,y)*f(x,y)

[0018]

[0019] Where f(x, y) is the original input image, G(x, y) is a second-order Gaussian filter, S(x, y) is the convolved image, M(x, y) is the gradient magnitude of each pixel, θ(x, y) is the direction of each pixel, G_x(x, y) and G_y(x, y) are the horizontal and vertical gradient amplitudes of the image at the pixel point (x, y), respectively.

[0020] Optionally, the method of embedding the edge extraction result into the input image with a hyperparameter q is:

[0021] H(x,y)=f(x,y)+q·D(x,y)

[0022] Among them, H(x,y) is the enhanced image, D(x,y) is the edge detection result, and q is a hyperparameter belonging to [0,1].

[0023] Optionally, the multi-distribution feature learning module includes:

[0024] With the idea of ​​seeking common ground while reserving differences, each data transformation is set as a branch. Each branch uses a convolutional layer to extract features from the data transformation results of the branch and uses an independent BN layer to normalize them to avoid interference from different branches and find consistent unified distribution features; the model parameters are adjusted to achieve the highest segmentation performance by minimizing the total loss of the joint branches.

[0025] Optionally, after performing multi-distribution shift on the enhanced input image, a multi-branch batch normalization structure is used to perform parallel training on different feature distributions to learn consistent features under different distribution domains, and a method for obtaining a feature map is as follows:

[0026] The enhanced input image is shifted into multiple distribution domains to obtain input domains with different distributions after the shift;

[0027] Perform batch normalization on the shifted input domains with different distributions to obtain batch normalization results;

[0028] The batch normalization results are summarized and unified, and a concat operation is used to obtain a unified learning input feature map.

[0029] Optionally, the loss function is obtained by weighting the cross entropy loss calculated in parallel in multiple distribution domains.

[0030] Set a weight parameter set K = [a, b, c], where is the distribution of the original clean image, is the noise branch, a is the weight parameter of the branch, is the rotation branch, b is the weight parameter of the branch, is the adversarial attack branch, c is the weight parameter of the branch;

[0031] Calculate the cross entropy loss for each branch separately;

[0032] The losses of each branch are added together to obtain the loss function.

[0033] Technical effect of the invention: The present invention discloses a method for small sample image segmentation enhancement guided by intrinsic features, emphasizes the importance of intrinsic feature guidance for small sample segmentation tasks, and proposes a promising method to solve the problem of limited robustness when the segmentation samples are insufficient. The present invention proposes a new idea of ​​embedding prior edge information and using a multi-branch BN structure to learn different distribution features, which enables the model to better utilize the inherent features in the data set, thereby improving the accuracy of the model. The present invention improves the overall performance of the model. The proposed PD block is a valuable supplement to existing methods and can be used to improve image segmentation methods in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0035] Figure 1 A schematic diagram of a model of a small sample image segmentation enhancement method guided by essential features according to an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of a multi-branch batch normalization structure according to an embodiment of the present invention;

[0037] Figure 3 Schematic diagram of different offset results of medical eyeball dataset images according to an embodiment of the present invention;

[0038] Figure 4 Schematic diagram of segmentation results of different small sample data sets according to embodiments of the present invention. DETAILED DESCRIPTION

[0039] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0040] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0041] like Figure 1 As shown, this embodiment provides a small sample image segmentation enhancement method guided by essential features, including the following steps:

[0042] Obtain a small sample image dataset to be segmented, and divide the small sample image dataset into a training set and a validation set according to an appropriate ratio;

[0043] The training set is input into the prior embedded feature enhancement module, and the contextual semantic information contained in the edge features is used as a priori to enhance the input image;

[0044] The enhanced input image is input into the multi-distribution feature learning module. After the enhanced input image is shifted by multiple distributions, a multi-branch batch normalization structure is used to perform parallel training on different feature distributions, learn consistent features under different distribution domains, and obtain feature maps.

[0045] The feature map is input into the image segmentation network of the mainstream codec structure for feature extraction and upper convolution layer restoration. The loss function is made with the segmentation label, and the loss function is trained and iterated repeatedly to converge, obtain the network model parameters, and obtain the trained segmentation network model.

[0046] The segmentation result is obtained by performing inference through the trained segmentation network model.

[0047] The specific technical solutions are as follows:

[0048] 1. Image data preprocessing

[0049] Import the small sample image dataset to be segmented and divide it into training set and validation set according to the appropriate ratio. Then perform some image preprocessing operations on the small sample image dataset, such as normalizing the input size, cropping, grayscale, etc.

[0050] 2. Prior Embedded Feature Enhancement Module (P Module)

[0051] The training set is input into the prior embedded feature enhancement module, and the contextual semantic information contained in the edge features is used as a prior to enhance the input image. The essence of image segmentation is to classify image information at the pixel level, and the segmentation result has a clear edge representation. Therefore, the edge features and the segmentation results are highly correlated and can be used as important prior information for model input. Incorporating these prior edge information can help the model learn to pay attention to the segmentation boundaries, thereby improving its performance in small sample segmentation tasks. In small-scale problems, insufficient sample size often leads to insufficient or excessive segmentation in the segmentation results. Enhancing edge features can make the model pay attention to the edge areas with higher gradients in the image, thereby making the edges clearer and helping to reduce such problems. Therefore, the present invention proposes a prior embedded feature enhancement method, that is, a priori enhancing the edge features in the original image, so that the model pays more attention to the edge information, the semantic distinction is more obvious, and the model is helped to make better segmentation decisions, effectively improving the segmentation accuracy. This module is mainly divided into two steps, one is to extract the edge of the input image, and the other is to embed the results of edge extraction into the input domain.

[0052] The present invention introduces the Canny algorithm to extract edge features, which has a series of advantages such as good robustness and high edge positioning accuracy. First, in order to reduce the noise in the original input image f(x, y), a second-order Gaussian filter is selected Because it can maintain image quality and provide effective smoothing. S(x, y) = G(x, y)*f(x, y)* represents the convolution operation; σ is the Gaussian filter coefficient. The larger the σ value, the better the smoothing effect and the stronger the noise suppression ability; S(x, y) represents the image after convolution.

[0053] After obtaining the smoothed image, it is necessary to calculate the gradient size of each pixel in the image and direction Among them, G_x(x,y) and G_y(x,y) are the horizontal and vertical gradient amplitudes of the image at the pixel point (x,y), respectively.

[0054] In image processing, it is generally believed that the extreme point of the gradient amplitude is an edge point. However, not all points with large gradient amplitudes are necessarily edge points. Non-maximum suppression is used to compare the gradient amplitude value of each pixel with its amplitude value in each gradient direction. If it is not the maximum value, the current grayscale value is represented as 0, indicating a non-edge point. If it is the maximum value, it is treated as an edge point to obtain a refined edge. In order to ensure the accuracy of the edge, a high threshold and a low threshold are generally set as hyperparameters, and each pixel in the image is judged. Pixels with a threshold greater than the high threshold are retained, and pixels with a threshold less than the low threshold are discarded. If the middle pixel is connected to a pixel above the high threshold, it is treated as an edge point, otherwise it is discarded. In this way, the edge information map D(x, y) of the input image can be obtained. Finally, the edge information map is embedded in the original input image and an appropriate weight q is assigned to achieve edge enhancement H(x, y) = f(x, y) + q·D(x, y).

[0055] Among them, H(x,y) is the enhanced image, D(x,y) is the edge detection result, and q is a hyperparameter belonging to [0,1].

[0056] 3. Multi-distribution feature learning module (D module)

[0057] The enhanced input image is input to a multi-distribution feature learning module to promote the model's learning of invariant features. Current data augmentation methods are generally regarded as data transformations of the input domain. For example, geometric transformation is a common type of perturbation. The method proposed in the present invention is based on the assumption that traditional batch normalization unifies samples of different distributions generated by different transformations, and the resulting mixed distribution may hinder the model's learning of basic features. Therefore, a feature learning module with unified distribution is proposed, which aims to integrate feature distribution information under different transformations and promote the model's learning of invariant features. Specifically, various data transformations are first performed on the current data, and then the batch statistics are explicitly decoupled in the normalization layer through a multi-branch BN structure to better absorb feature consistency.

[0058] In order to improve the robustness of the model and the richness of the data set, the present invention selects three commonly used data conversion strategies: adversarial attacks, noise addition, and geometric transformation. Then batch normalization (BN) is an important component of many advanced computer vision models, and a special design is made for this. Specifically, batch normalization normalizes the input features according to the mean and variance calculated in each small batch. An inherent assumption of using BN is that the input features should come from a single distribution or a similar distribution. However, if the distribution of small batches of data is different, this normalization operation will cause problems, resulting in inaccurate statistical estimates.

[0059] The transformed images of the input data have different basic distributions and cannot be simply input into BN for normalization. Therefore, a multi-branch BN structure is proposed, such as Figure 2 Show. Figure 2 The data input and working mechanism of our proposed multi-branch BN are compared with those of the traditional BN. The figure also shows the probability distribution curves under different types of data input (x). Figure 2 a) shows the traditional data enhancement method and its BN method. The initial clean image x T0 The input undergoes multiple data transformations to generate different transformed inputs x Tk (where k represents different data conversion methods, such as x T1 and x T2 ). Figure 2 The dotted line in a) represents the theoretical probability distribution of these transformed data. The traditional method unifies these data with different distributions into BN. The solid line represents the final probability distribution curve of the actual input model after the unified BN. The proposed multi-branch BN structure is shown in Figure 2 b). The main concept is to use parallel BN branches to process the original clean image x separately. T0 and its derivative data transformation image x Tk Unlike traditional methods, each transformation branch in this mechanism is independent and will not be interfered by other transformation methods. Then, their distribution information will be integrated and sent to the ReLU layer together. In this way, the model can learn more about the true distribution information of each branch, and thus learn more useful features from each branch.

[0060] Randomly set three boundary threshold parameters (a, b and c) to divide the interval [0,1] into four subintervals. At the beginning of each iteration, a random number s is randomly generated in the range [0,1]. The random number may belong to

[0061] The model is then trained according to the BN branch determined by the following formula:

[0062]

[0063] The proposed method aims to alleviate this shortcoming by combining multiple branches of BN, each of which undergoes a different transformation of the image. This multi-branch BN approach has two potential benefits. First, it achieves distribution-uniform feature learning, which plays a role similar to data augmentation in improving the generalization ability of the model. By generating more training samples, the model can better utilize the training data, thereby alleviating overfitting and reducing sample imbalance, and enhancing its robustness. Second, distribution-uniform feature learning in a multi-branch BN system can help the model effectively and accurately learn the invariant features of the image under different transformations. This in turn improves accuracy and can mine information from the data in a more efficient manner.

[0064] 4. Image Segmentation Network

[0065] The feature graph processed by the P module and the D module is sent to the image segmentation network of the mainstream codec structure for feature extraction and upper convolution layer restoration. The loss function is used to train and iterate with the segmentation label to make the loss function converge, thereby obtaining the network model parameters. Here, the cross entropy loss needs to be calculated for each branch separately. The total loss function of the final model is the sum of the losses of each branch, and the network model is argmin θ [L c (θ,(x c ,y))+L a (θ,(x a ,y))+L n (θ,(x n ,y))+L r (θ,(x r ,y))].

[0066] The present invention provides a small sample image segmentation enhancement method guided by essential features, and its pseudo code is shown in Table 1, which specifically includes:

[0067] Table 1

[0068]

[0069] (1) According to the task requirements, select the corresponding small sample image segmentation dataset. Here, the medical small sample public dataset CHASE DB1 is taken as an example;

[0070] (2) Normalize the image dataset. The original dataset input is 999*960. Adjust the input image size according to the ratio of (0.5, 2) and then randomly crop it to 128*128.

[0071] (3) graying the processed image dataset and inputting it into the prior embedded feature enhancement module;

[0072] (4) By using a second-order Gaussian filter Smooth the image and filter out noise;

[0073] (5) Calculate the gradient and direction of each pixel in the smoothed image;

[0074] (6) Apply non-maximum suppression to eliminate the spurious responses caused by edge detection.

[0075] (7) Apply dual threshold detection to determine real and potential edges and complete edge detection by suppressing weak edges;

[0076] (8) Setting the hyperparameter q to superimpose the edge detection result with the original input image to obtain a new image input, completing the prior feature embedding;

[0077] (9) Set the weight parameter K = [a, b, c], input the new image into the multi-distribution feature learning module, first process the image input, and perform image transformation operations such as adding noise, rotating the image, and adversarial attack on the image input generated by (8), such as Figure 3 As shown;

[0078] (10) Establish a multi-branch structure and place the original clean image and the three transformed image distributions into the corresponding four independent branch structures;

[0079] (11) Establish a multi-branch BN layer and perform batch normalization on the data distribution in each branch;

[0080] (12) superimposing the results of each group of normalization processing according to the weight parameter size to obtain a new image input;

[0081] (13) Input the image in (12) into the image segmentation network of DeepLabv3 and find the minimum value argmin of the total loss of the four branches θ [L c (θ,(x c ,y))+L a (θ,(x a ,y))+L n (θ,(x n ,y))+L r (θ,(x r ,y))] establish the model;

[0082] (14) The batch size is set to 4 and the model is trained using stochastic gradient descent. The initial learning rate is 0.01, the momentum is 0.9, and the weight decay rate is 0.0005. The multi-exponential decay method is used as the learning rate decay strategy, where the momentum is set to 0.9 and the minimum learning rate is 0.0001. The number of iterations is set to 40,000.

[0083] (15) After the model is iterated and the loss function converges, the final model is obtained. The inference can obtain a relatively ideal segmentation result of the medical small sample image.

[0084] In order to comprehensively evaluate the performance of the PD block, a series of experiments were conducted on three small sample datasets (Chase DB1, DRIVE, and STARE) using U-Net, PSPNet, and Deeplabv3 as benchmark networks. The experimental results are shown in Table 2, and the visualization results of some typical scenarios are shown in Table 2. Figure 4 shown.

[0085] Table 2 summarizes the performance of different segmentation models using PD blocks on different datasets. By introducing PD blocks, the performance of all models on CHASE DB1, DRIVE, and STARE datasets is improved. Taking U-Net as an example, the accuracy (Acc) is improved from 95.21% to 97.17%, and the dice coefficient (mDice) is improved from 89.27% ​​to 89.79%. Similar trends are observed in other models, verifying the versatility of PD blocks for different datasets.

[0086] The improvement of Rec (sensitivity) shows a consistent trend on the CHASE DB1, DRIVE, and STARE datasets. On the CHASE DB1 dataset, the models that introduce the PD block generally show a significant improvement in Rec, with an average improvement of about 3.5%. In addition, on the DRIVE and STARE datasets, adding the PD block can improve the recall performance of all baseline models, and the average recall rates of the two datasets are increased by about 2.6% and 2.3%, respectively. These results show that the PD block is universal to the model and can improve the sensitivity to true positive examples, providing support for further improving the segmentation performance of the model. Other indicators (F1, mIoU, and SP) also show similar trends to Recall. It is worth noting that the introduction of the PD block can bring performance improvements even in higher-performing baseline models (such as Genetic U-net and FANet). The results show that the PD block is compatible with advanced baseline models to further improve performance. Therefore, as a feature-based plug-in, the PD block is universal and effective in improving the performance of small sample segmentation tasks.

[0087] Table 2

[0088]

[0089]

[0090] The specific advantages of the small sample image segmentation enhancement method guided by intrinsic features provided by the present invention are as follows: a) It emphasizes the importance of intrinsic feature guidance for small sample segmentation tasks, and proposes a promising method to solve the problem of limited robustness when the segmentation samples are insufficient. b) It proposes a new idea of ​​embedding prior edge information and using multi-branch BN to learn the structure of different distribution features, which enables the model to better utilize the inherent features in the data set, thereby improving the accuracy of the model. c) Compared with the existing methods, the overall performance of the model is improved. The proposed PD block is a valuable supplement to the existing methods and can be used to improve the image segmentation methods in practical applications.

[0091] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A small sample image segmentation enhancement method guided by essential features, characterized in that: include: Obtain a small sample image data set to be segmented, and divide the small sample image data set into a training set and a validation set according to an appropriate ratio; Inputting the training set into a priori embedded feature enhancement module, and enhancing the input image using contextual semantic information contained in the edge features as a priori; Inputting the enhanced input image into a multi-distribution feature learning module, performing multi-distribution shift on the enhanced input image, and then using a multi-branch batch normalization structure to perform parallel training on different feature distributions, learn consistent features under different distribution domains, and obtain a feature map; The feature map is input into the image segmentation network of the mainstream codec structure for feature extraction and upper convolution layer restoration, the network output is compared with the segmentation label to calculate the loss function, training and repeated iterations are performed to converge the loss function, obtain network model parameters, and obtain a trained segmentation network model; Performing reasoning through the trained segmentation network model to obtain a segmentation result; The prior embedded feature enhancement module includes: Extract edges from input images using edge detection algorithms; The edge extraction result is embedded into the input image with a hyperparameter q; The method of edge extraction using edge detection algorithm for input image is: S(x,y)=G(x,y)*f(x,y) Where f(x,y) is the original input image, G(x,y) is a second-order Gaussian filter, S(x,y) is the convolved image, M(x,y) is the gradient magnitude of each pixel, θ(x,y) is the direction of each pixel, G_x(x,y) and G_y(x,y) are the horizontal and vertical gradient amplitudes of the image at the pixel point (x,y), respectively; The method of embedding the edge extraction result into the input image with a hyperparameter q is: H(x,y)=f(x,y)+q·D(x,y) Among them, H(x,y) is the enhanced image, D(x,y) is the edge detection result, and q is a hyperparameter belonging to [0,1]; The multi-distribution feature learning module includes: Each data transformation is set as a branch. Each branch uses a convolutional layer to extract features from the data transformation results of the branch and uses an independent BN layer to normalize them to avoid interference from different branches and find consistent uniform distribution features. The model parameters are adjusted to achieve the highest segmentation performance by minimizing the total loss of the joint branches. After performing multi-distribution shift on the enhanced input image, a multi-branch batch normalization structure is used to perform parallel training on different feature distributions to learn consistent features under different distribution domains. The method for obtaining the feature map is as follows: The enhanced input image is shifted to multiple distribution domains to obtain input domains with different distributions after the shift; Perform batch normalization on the shifted input domains with different distributions to obtain batch normalization results; Summarize and unify the batch normalization results, and use a concat operation to obtain a unified learning input feature map; The model will be trained according to the BN branch determined by the following formula: The loss function is obtained by weighting the cross entropy loss calculated in parallel in multiple distribution domains, and a weight parameter set K = [a, b, c] is set, where is the distribution of the original clean image, is the noise branch, a is the weight parameter of the branch, is the rotation branch, b is the weight parameter of the branch, is the adversarial attack branch, c is the weight parameter of the branch; Calculate the cross entropy loss for each branch separately; The losses of each branch are added together to obtain the loss function.