Material microscopy image prediction method based on improved sample-free incremental learning

Abstract

The invention relates to a material microscopy image prediction method based on improved sample-free incremental learning. The method comprises the following steps: constructing a material microscopic image prediction model; training a material microscopic image prediction model according to the material microscopic image set, and performing feature extraction on the first data set in a feature extractor to obtain initial features; the feature and label alignment module is used for expanding initial features, determining parameters of a linear layer and retaining a feature trunk of a feature extractor; the feature optimization and alignment module is used for optimizing the feature extractor, determining parameters of a second linear layer, performing alignment in an increment stage and storing a prototype of a new class; and performing label alignment on the extended pseudo features according to the parameters of the second linear layer in the forgetting compensation and test module, and predicting a to-be-predicted material microscopic image by using the trained material microscopic image prediction model. By adopting the method, the prediction accuracy can be improved.

Description

Technical Field

[0001] The present application relates to the technical field of image processing, and in particular to a material microscopy image prediction method based on improved sample-free incremental learning. Background Art

[0002] In the field of materials science, microscopic image analysis techniques (such as scanning electron microscopy, transmission electron microscopy, and atomic force microscopy) have become important research tools, providing a large amount of microscopic information for understanding material properties and behaviors. However, with the surge in the amount of material microscopic image data, researchers are facing increasing challenges in the image analysis process. This not only increases the complexity of the analysis, but may also lead to errors and omissions in the understanding of material properties. In order to improve the accuracy and efficiency of microscopic image analysis, researchers have begun to explore material microscopic image prediction methods based on artificial intelligence and machine learning. These methods can automatically analyze large amounts of microscopic image data and help researchers predict material categories more quickly and accurately by learning features extracted from historical samples.

[0003] However, current methods still face problems such as lack of interpretability, strong dependence on datasets, and low prediction accuracy in practical applications. Summary of the invention

[0004] Based on this, it is necessary to provide a material microscopy image prediction method based on improved sample-free incremental learning that can improve the prediction accuracy in response to the above technical problems.

[0005] A material microscopy image prediction method based on improved sample-free class incremental learning, the method comprising:

[0006] Obtain a collection of annotated microscopic images of common materials; the material microscopic image collection includes a first data set and a second data set; construct a material microscopic image prediction model; the material microscopic image prediction model includes a feature extractor, a feature and label alignment module, a feature optimization and alignment module, and a forgetting compensation and testing module;

[0007] The material microscopic image prediction model is trained according to the material microscopic image set, and the feature extractor extracts features from the first data set to obtain initial features;

[0008] In the feature and label alignment module, the initial features are expanded using the extended layer of the analysis classifier and the parameters of the linear layer are determined. The linear layer after the parameters are determined is used to align the expanded features and labels in the initial stage to retain the feature backbone and important information of the feature extractor; the important information includes the autocorrelation matrix and the prototype of the initial class;

[0009] The feature extractor is optimized in the feature optimization and alignment module, and the feature extractor and feature backbone are used to extract features from the second data set and the optimized extended features are obtained through the extension layer; the parameters of the second linear layer are determined by using the optimized extended features, and the optimized extended features and labels are aligned in the incremental stage according to the parameters of the second linear layer, and the prototype of the new class is saved;

[0010] In the forgetting compensation and testing modules, the linear layer and autocorrelation matrix in the analysis classifier are copied, the prototype features of each new class are calculated using the feature backbone, the pseudo features of the old class are generated according to the similarity between the new class and the old class, and the pseudo features are passed through the optimized feature extractor and then through the expansion layer to obtain the extended pseudo features; the labels of the extended pseudo features are aligned according to the parameters of the second linear layer to obtain the trained material microscopy image prediction model;

[0011] The trained material microscopy image prediction model is used to predict the material microscopy image to obtain the material category prediction result.

[0012] The above-mentioned material microscopy image prediction method based on improved sample-free incremental learning first constructs a material microscopy image prediction model; the material microscopy image prediction model includes a feature extractor, a feature and label alignment module, a feature optimization and alignment module, and a forgetting compensation and testing module. In view of the shortcomings of limited generalization ability of sample features and insufficient alleviation of forgetting problems in the prior art, the present application introduces self-supervised label enhancement technology and knowledge distillation technology in the feature extractor for feature optimization, so that the model can learn features with more generalization capabilities, overcoming the problem of limited feature generalization caused by freezing of feature extractors in the prior art. At the same time, the knowledge distillation technology constrains the feature optimization process to ensure that when learning new data features, it will not deviate excessively from the learned knowledge, effectively alleviating the forgetting problem. Secondly, in order to alleviate the problems of insufficient utilization of the backbone network and limited alleviation of sample imbalance problems, the feature extractor architecture of the present application consists of an invariant feature backbone and an adaptive feature projector. In the feature and label alignment module, the extended layer of the analysis classifier is used to expand the initial features and determine the parameters of the linear layer. The linear layer after determining the parameters is used to align the extended features and labels in the initial stage to retain the feature trunk and important information of the feature extractor. The unchanged feature trunk saves the prototype and generates pseudo features, providing a stable foundation for the model and making full use of the learning ability of the backbone network. The forgetting compensation and testing module further improves the model's processing ability for different categories of data by generating pseudo features and performing alignment tests, and more effectively alleviates the problem of sample imbalance. In the feature optimization and alignment module, the feature extractor is optimized to enhance the generalization and plasticity of the features, which can better fit data with complex nonlinear relationships and effectively solve the underfitting problem. At the same time, the forgetting compensation and testing module dynamically adjusts the balance between learning new knowledge and retaining old knowledge by copying the analysis classifier and performing pseudo feature alignment, more accurately balancing stability and plasticity, improving the overall performance and generalization ability of the model, and the model can effectively improve the prediction accuracy of material microscopy images. At the same time, it also ensures that the model can not only process new data, but also maintain stable performance and efficient learning ability in the emerging material microscopy images, thus having important value in material design and application practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 A schematic diagram of a process of a material microscopy image prediction method based on improved sample-free incremental learning in one embodiment;

[0014] Figure 2 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0016] In one embodiment, Figure 2 As shown, a material microscopy image prediction method based on improved sample-free class incremental learning is provided, comprising the following steps:

[0017] Step 102, obtaining a set of annotated material microscopy images of common materials; the material microscopy image set includes a first data set and a second data set; constructing a material microscopy image prediction model; the material microscopy image prediction model includes a feature extractor, a feature and label alignment module, a feature optimization and alignment module, and a forgetting compensation and testing module.

[0018] Obtaining a large number of annotated material microscopy image collections of common materials, including the first data set and the second data set, provides a solid foundation for model training. This large-scale and diverse data set can effectively reduce the risk of overfitting the model to specific samples and improve the model's ability to recognize microscopy images of different materials. The diversity of the data ensures that the model can maintain a high accuracy rate when facing new samples. The first data set and the second data set are different data sets.

[0019] Step 104 , training a material microscopy image prediction model based on the material microscopy image set, and extracting features from the first data set in a feature extractor to obtain initial features.

[0020] By performing feature extraction on the first dataset, initial features can be learned from the rich material microscopy images. These features are the basis for subsequent analysis, and good feature extraction can retain important information in the image.

[0021] Step 106, in the feature and label alignment module, the expansion layer of the analysis classifier is used to expand the initial features and determine the parameters of the linear layer, and the linear layer after the parameters are determined is used to align the expanded features and labels in the initial stage to retain the feature backbone and important information of the feature extractor; the important information includes the autocorrelation matrix and the prototype of the initial class.

[0022] The input data is passed through the feature extractor to obtain features, and then through the expansion layer to obtain extended features. Then, the parameters of the linear layer are determined by solving an optimization problem (such as minimizing a certain distance metric between the features and the labels) so that the extended features can be better aligned with the labels. After the alignment is completed, the feature backbone f in the feature extractor is frozen. FB, and save some important information, such as the autocorrelation matrix (Save Relation) and the prototype of the initial class (Save Prototype Set)

[0023] The initial features are expanded through the expansion layer and aligned with the label. This process not only retains the important information of the feature extractor, but also enhances the correlation between the features and their labels, thereby improving the model's discriminative ability. The initial class refers to the initial features corresponding to the annotated material features.

[0024] Step 108, optimize the feature extractor in the feature optimization and alignment module, extract features from the second data set based on the optimized feature extractor and feature backbone, and obtain optimized extended features through the extension layer; determine the parameters of the second linear layer using the optimized extended features, align the optimized extended features and labels in the incremental stage based on the parameters of the second linear layer, and save the prototype of the new class.

[0025] Using the invariant feature backbone (f FB ) and the trained adaptive feature projector f FP (SSLA Projector) and the current stage of the dataset Get the feature matrix Then go through the expansion layer to get the extended feature matrix The innovative proposal is based on the invariant feature backbone f FB and the adaptive feature projector f FP The optimized feature extractor architecture is composed of the following. The invariant feature backbone is used to preserve the prototype and generate pseudo features, the adaptive feature projector is used to enhance feature generalization and plasticity, and only the feature projector is optimized to alleviate forgetting. This structural design solves the problem of feature misalignment with labels and mismatch with prototypes caused by unstable feature representation in existing methods.

[0026] At the same time, the optimized feature extractor is used to extract features from the second data set, and the optimized extended features are generated through the extension layer. Using these optimized features, the model can adjust and learn for new data while saving the prototype of the new class. This incremental learning method enables the model to continuously adapt to new image data without losing existing knowledge, thus effectively alleviating the forgetting problem.

[0027] Step 110, in the forgetting compensation and testing module, the linear layer and the autocorrelation matrix in the analysis classifier are copied, the prototype features of each new class are calculated using the feature backbone, the pseudo features of the old class are generated according to the similarity between the new class and the old class, the pseudo features are passed through the optimized feature extractor and then through the expansion layer to obtain the extended pseudo features; the labels of the extended pseudo features are aligned according to the parameters of the second linear layer to obtain the trained material microscopy image prediction model.

[0028] In each incremental stage, the linear layer in the analytical classifier (Replicated AnalyticClassifier) ​​and the autocorrelation matrix are first replicated. Then, the feature backbone (f FB ) Calculate each new class C n The prototype of in W FB is the feature backbone parameter. Save the prototype to the prototype set, calculate the similarity between the new and old classes through cosine similarity, and generate pseudo features of the old class. Assuming that the old class is most similar to the new class, there is a pseudo feature

[0029] The generated pseudo features are passed through the trained feature projector (i.e., f FP ) to obtain features with better generalization ability, set Then after the expansion layer, the extended pseudo features are obtained

[0030] Align the expanded pseudo features with the copied analysis classifier and set the weights of the copied linear layer to The autocorrelation matrix is ​​R′ k , align, and then use the copied analysis classifier for testing to obtain the test results of the current stage. After the test is completed, the relevant intermediate results, such as the extended pseudo features, the copied classifier weight matrix and the autocorrelation matrix, are discarded to prepare for the next incremental stage. Model training is carried out in sequence until the preset threshold is reached to complete the model training. The new class refers to the disease of a different part from the old class. The forgetting compensation and testing module copies the second linear layer and the autocorrelation matrix of the analysis, generates pseudo features by saving the prototype to calculate the relationship between the new and old prototypes, and uses the pseudo features through the feature projector to obtain more generalizable features and align them with the copied classifier to compensate for forgetting and avoid overfitting. This unique pseudo feature generation and alignment mechanism is a key innovation to solve the forgetting problem. It can effectively improve the performance of the model in sample-free incremental learning. In specific applications, the linear layer and autocorrelation matrix in the copy analysis classifier can be balanced between new and old knowledge. By generating pseudo features of the old class and comparing them with the new class, the model can dynamically adjust the weight between learning new knowledge and retaining old knowledge. This mechanism ensures that the model can still maintain effective recognition of old data while constantly learning new data, thereby improving the overall prediction accuracy. By building an autocorrelation matrix and prototype features, the model can not only provide results when making predictions, but also provide interpretability for these results.

[0031] Step 112, using the trained material microscopic image prediction model to predict the material microscopic image to obtain a material category prediction result.

[0032] In the above-mentioned material microscopy image prediction method based on improved sample-free incremental learning, a material microscopy image prediction model is first constructed; the material microscopy image prediction model includes a feature extractor, a feature and label alignment module, a feature optimization and alignment module, and a forgetting compensation and testing module. In view of the shortcomings of limited generalization ability of sample features and insufficient alleviation of forgetting problems in the prior art, the present application introduces self-supervised label enhancement technology and knowledge distillation technology in the feature extractor for feature optimization, so that the model can learn features with more generalization capabilities, overcoming the problem of limited feature generalization caused by freezing of feature extractors in the prior art. At the same time, the knowledge distillation technology constrains the feature optimization process to ensure that when learning new data features, it will not deviate excessively from the learned knowledge, effectively alleviating the forgetting problem. Secondly, in order to alleviate the problems of insufficient utilization of the backbone network and limited alleviation of sample imbalance problems, the feature extractor architecture of the present application consists of an invariant feature backbone and an adaptive feature projector. In the feature and label alignment module, the extended layer of the analysis classifier is used to expand the initial features and determine the parameters of the linear layer. The linear layer after determining the parameters is used to align the extended features and labels in the initial stage to retain the feature trunk and important information of the feature extractor. The unchanged feature trunk saves the prototype and generates pseudo features, providing a stable foundation for the model and making full use of the learning ability of the backbone network. The forgetting compensation and testing module further improves the model's processing ability for different categories of data by generating pseudo features and performing alignment tests, and more effectively alleviates the problem of sample imbalance. In the feature optimization and alignment module, the feature extractor is optimized to enhance the generalization and plasticity of the features, which can better fit data with complex nonlinear relationships and effectively solve the underfitting problem. At the same time, the forgetting compensation and testing module dynamically adjusts the balance between learning new knowledge and retaining old knowledge by copying the analysis classifier and performing pseudo feature alignment, more accurately balancing stability and plasticity, improving the overall performance and generalization ability of the model, and the model can effectively improve the prediction accuracy of material microscopy images. At the same time, it also ensures that the model can not only process new data, but also maintain stable performance and efficient learning ability in the emerging material microscopy images, thus having important value in material design and application practice.

[0033] In one embodiment, extracting features from the first data set according to a feature extractor to obtain initial features includes:

[0034] In the feature extractor, self-supervised label enhancement technology and a fully connected classifier are used to extract features from the first data set to obtain initial features.

[0035] In a specific embodiment, the feature extractor uses the self-supervised label enhancement (SSLA) technique and the fully connected classifier (FCN) on the first dataset. The self-supervised label enhancement technique generates additional supervised information by rotating the training samples, that is, rotating the sample x by 90 degrees, 180 degrees, and 270 degrees, and expanding the label y. This can help the feature extractor learn features with better generalization capabilities.

[0036] In one embodiment, the feature and label alignment module uses the expansion layer of the analysis classifier to expand the initial features and determine the parameters of the linear layer, including:

[0037] The initial features are expanded using the extended layer of the analysis classifier, and the expanded features are obtained as follows:

[0038]

[0039] Among them, f act represents the activation function, f E represents the extended layer, f flat represents the flattening operator, represents the initial features;

[0040] The linear layer parameter optimization problem is designed as

[0041]

[0042] in, represents the parameters of the linear layer, Indicates extended features, || || F represents the Frobenius norm, γ is the regularization hyperparameter;

[0043] The linear layer parameter optimization problem is solved according to the least squares method, and the optimal solution is obtained, that is, the parameters of the linear layer are

[0044] where I is the identity matrix, (.) T is the matrix transpose operation.

[0045] In a specific embodiment, the trained feature extractor converts the input data into features, and then aligns the features and labels through an analytical classifier. The analytical classifier includes an expansion layer (Feature Projector) and a linear layer (Analytical Classifier). Let the input be X, then through the feature extractor f FE (Parameter is W FE ) to get feature X FE =f FE (X; W FE). Then an analytical classifier is used to align features and labels. This analytical classifier contains an expansion layer f E (Shape is d FE ×d E ) and a linear layer f L . Feature X FE After the expansion layer f E Get extended features Here is f act is the activation function (such as ReLU), f flat is the flattening operator. Then the linear layer f is determined by solving the optimization problem L The parameters of are estimated by the least squares method to obtain the optimal solution. After the alignment is completed, some parts of the feature extractor (such as the feature backbone f FB ), and save some important information, the autocorrelation matrix R0 and the prototype of the initial class (Save Prototype Set).

[0046] In one embodiment, the feature extractor is optimized, including:

[0047] The feature extractor is optimized by using self-supervised label enhancement and knowledge distillation techniques combined with a fully connected classifier and the second dataset of the current stage, by minimizing the current feature extractor f FE；k and the feature extractor f in the previous stage FE；k-1 The distance between the extracted new data features is used to constrain the optimization, and the formula is The overall loss is L total =L KD +L CE , where L CE is the cross entropy loss.

[0048] In a specific embodiment, a self-supervised label enhancement (SSLA) technique is used, such as rotating the training samples to expand the labels, and combining a fully connected classifier and cross entropy loss to learn generalizable features at all stages. This is one of the key innovations that distinguishes it from existing analytical learning (AL) methods. Existing methods often only focus on the initial stage when training feature extractors, and ignore the feature generalization requirements in the incremental stage. The knowledge distillation (KD) technique is used to constrain the feature optimization process, limiting the distance between the current feature extractor and the new data features extracted by the feature extractor in the previous stage, to avoid the forgetting problem caused by over-optimization. This combination of self-supervised learning and knowledge distillation provides an effective way to optimize features, which can improve the generalization ability of features while maintaining the model's memory of learned knowledge.

[0049] In one embodiment, determining the parameters of the second linear layer using the optimized extended features includes:

[0050] The parameters of the second linear layer are determined using the optimized extended features:

[0051]

[0052] Among them, R k represents the autocorrelation matrix, represents the optimized diffusion characteristics, The label matrix representing the training data at stage K, Represents the optimal estimate of the weight matrix of the k-1 linear layer in the previous stage.

[0053] In one embodiment, the prototype features of each new class are calculated using the feature backbone, including:

[0054] The feature backbone is used to calculate the prototype feature of each new class. The formula is:

[0055]

[0056]

[0057] Among them, C n represents a new class, f FB represents the feature backbone, W FB represents the characteristic backbone parameter, X Cn Represents a new class C n of the sample.

[0058] In one embodiment, generating a pseudo feature of the old class according to the similarity between the new class and the old class further includes:

[0059] The pseudo features of the old class are generated according to the similarity between the new class and the old class:

[0060]

[0061] in, Represents a new class C n The features after being processed by the invariant feature backbone FB, μ(C n ) represents the new class C n The prototype is calculated by The mean of μ(C0) is obtained, μ(C0) represents the prototype of the old class C0, and the sample features are calculated by The mean of .

[0062] In a specific embodiment, the invariant feature backbone can extract features with certain stability and representativeness. These features are used as the basis in the process of generating pseudo features. They are combined with the new class prototype to provide information related to the new class for the generation of old class pseudo features, which helps the model to establish connections between feature spaces of different categories and reduce forgetting. The new class prototype μ(C n ) and new class features Working together, the feature information of the new class is converted into pseudo features of the old class by calculating the difference with the old class prototype μ(C0), so that the model can use the information of the new class to compensate for the forgetting of the old class, and also reflects the association and conversion of features between different categories.

[0063] In one embodiment, the pseudo feature is passed through an optimized feature extractor and then through an expansion layer to obtain an extended pseudo feature, including:

[0064] The pseudo features are passed through the optimized feature extractor and then through the expansion layer to obtain the extended pseudo features:

[0065]

[0066] Among them, f act represents the activation function, f E represents the extended layer, f flat represents the flattening operator, Represents the features after the pseudo features pass through the optimized feature extractor.

[0067] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0068] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 2As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, a material microscopy image prediction method based on improved sample-free incremental learning is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad set on the computer device housing, or an external keyboard, touchpad or mouse, etc.

[0069] Those skilled in the art will understand that Figure 2 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0070] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0071] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0072] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A material microscopy image prediction method based on improved sample-free incremental learning, characterized in that: The method comprises: Acquire a set of annotated material microscopic images of common materials; the material microscopic image set includes a first data set and a second data set; construct a material microscopic image prediction model; the material microscopic image prediction model includes a feature extractor, a feature and label alignment module, a feature optimization and alignment module, and a forgetting compensation and testing module; Training the material microscopic image prediction model according to the material microscopic image set, and extracting features from the first data set in the feature extractor to obtain initial features; In the feature and label alignment module, the initial features are expanded using the expansion layer of the analysis classifier and the parameters of the linear layer are determined, and the expanded features and labels are aligned in the initial stage using the linear layer after the parameters are determined to retain the feature backbone and important information of the feature extractor; the important information includes the autocorrelation matrix and the prototype of the initial class; The feature extractor is optimized in the feature optimization and alignment module, and the feature extractor and the feature backbone are used to extract features from the second data set and obtain optimized extended features through the extension layer; the parameters of the second linear layer are determined by using the optimized extended features, and the optimized extended features and labels are aligned in the incremental stage according to the parameters of the second linear layer, and the prototype of the new class is saved; The linear layer and the autocorrelation matrix in the forgetting compensation and testing module are copied and analyzed, the prototype features of each new class are calculated using the feature backbone, the pseudo features of the old class are generated according to the similarity between the new class and the old class, the pseudo features are passed through the optimized feature extractor and then through the expansion layer to obtain the extended pseudo features; the labels of the extended pseudo features are aligned according to the parameters of the second linear layer to obtain the trained material microscopy image prediction model; The trained material microscopic image prediction model is used to predict the material microscopic image to be predicted, and a material category prediction result corresponding to the material microscopic image is obtained.

2. The method according to claim 1, characterized in that Extracting features from the first data set according to the feature extractor to obtain initial features includes: The feature extractor uses a self-supervised label enhancement technique and a fully connected classifier to extract features from the first data set to obtain initial features.

3. The method according to claim 1, characterized in that In the feature and label alignment module, the initial feature is expanded using the expansion layer of the analysis classifier and the parameters of the linear layer are determined, including: The initial features are expanded using the expansion layer of the analysis classifier to obtain the expanded features: Among them, f act represents the activation function, f E represents the extended layer, f flat represents the flattening operator, represents the initial features; The linear layer parameter optimization problem is designed as in, represents the parameters of the linear layer, Indicates extended features, || || F represents the Frobenius norm, γ is the regularization hyperparameter; The linear layer parameter optimization problem is solved according to the least squares method to obtain the optimal solution, that is, the parameters of the linear layer are where I is the identity matrix, (.) T is the matrix transpose operation.

4. The method according to claim 3, characterized in that Optimize the feature extractor, including: The feature extractor is optimized by using self-supervised label enhancement and knowledge distillation techniques combined with a fully connected classifier and the second dataset of the current stage, by minimizing the current feature extractor f FE；k and the feature extractor f in the previous stage FE；k-1 The distance between the extracted new data features is used to constrain the optimization, and the formula is The overall loss is L total =L KD +L CE , where L CE is the cross entropy loss.

5. The method according to claim 1, characterized in that The optimized extended features are used to determine the parameters of the second linear layer, including: The parameters of the second linear layer are determined using the optimized extended features: Among them, R k represents the autocorrelation matrix, represents the optimized diffusion characteristics, The label matrix representing the training data at stage K, Represents the optimal estimate of the weight matrix of the k-1 linear layers in the previous stage.

6. The method according to claim 1, characterized in that The feature backbone is used to calculate the prototype features of each new class, including: The feature backbone is used to calculate the prototype feature of each new class, and the formula is Among them, C n represents a new class, f FB represents the feature backbone, W FB represents the characteristic backbone parameter, Represents a new class C n of the sample.

7. The method according to claim 1, characterized in that Generate pseudo features of the old class based on the similarity between the new class and the old class, and also include: The pseudo features of the old class are generated according to the similarity between the new class and the old class: in, Represents a new class C n The features after being processed by the invariant feature backbone FB, μ(C n ) represents the new class C n The prototype is calculated by The mean of μ(C0) is obtained, μ(C0) represents the prototype of the old class C0, and the sample features are calculated by The mean of .

8. The method according to claim 7, characterized in that The pseudo features are passed through the optimized feature extractor and then through the expansion layer to obtain extended pseudo features, including: The pseudo features are passed through the optimized feature extractor and then through the expansion layer to obtain the extended pseudo features: Among them, f act represents the activation function, f E represents the extended layer, f flat represents the flattening operator, Represents the features after the pseudo features pass through the optimized feature extractor.

Images