Recognition method of pre-training model-based incremental learning based on neural collapse theory

Abstract

The application provides a recognition method of pre-training model class incremental learning based on neural collapse theory, introduces the theory of NC into the feature adaptation of the pre-training model, and provides a novel paradigm to solve the downstream task by using the pre-training model. Moreover, the application constructs an ETF classifier which can guide the feature representation to adapt to the optimal structure. The ETF classifier not only meets the NC2 characteristic in neural collapse, but also can be pre-fixed to the optimal structure to be stable in the training process. The FCM module is introduced to compress multiple features to the specified class prototype, reduce the intra-class variability, and meet the NC1 characteristic of neural collapse. Through the supervised learning process, the feature representation is aligned with the optimal classifier structure, so that a more discriminative feature representation is obtained. The pre-training model is trained by the application for incremental learning, and the trained model is used to execute the target recognition task, so that the recognition effect can be improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data classification and recognition, and particularly relates to a recognition method of pre-training model class incremental learning based on neural collapse theory. BACKGROUND

[0002] Class-Incremental Learning (CIL) is a learning mode of artificial intelligence system that continuously learns new tasks without forgetting previously learned tasks. These tasks have non-overlapping class sets. In CIL, the model needs to learn new classes step by step, and at the same time maintain the recognition ability of old classes. However, due to the update of model parameters, the feature representation between different incremental tasks will be shifted, resulting in performance decline on previous tasks, which is called catastrophic forgetting.

[0003] To cope with the challenges in CIL, researchers introduce pre-training models to utilize their general feature representations trained on large-scale datasets. However, the general feature representations of pre-training models may not be adaptive and discriminative for various incremental classes, as these incremental classes may exceed the distribution range of the pre-training dataset. In addition, due to the inability to learn incremental classes at once, the relationship between classes cannot be optimally constructed, resulting in a lack of discriminativeness of feature representations for downstream tasks.

[0004] Neural Collapse (NC) theory explains an elegant phenomenon between the last layer features and the classifier in deep learning classification models, which can be summarized as the following three properties:

[0005] (1) Variability Collapse (NC1): the variability of intra-class features collapses to zero. (2) Convergence to Simplex ETF (NC2): the class mean features converge to a simple shape equiangular tight frame (ETF), achieving equal length, equal diagonal and maximum distance in the feature space. (3) Self-Duality (NC3): the linear classifier converges to the class mean feature until the global scaling.

[0006] Existing class incremental learning methods based on pre-training models (PTM-CIL) can be mainly divided into three categories: prompt-based methods, representation-based methods and model mixing-based methods.

[0007] (1) Prompt-based Methods, which use prompt tuning strategies to slightly update the feature representations of PTMs, but due to the limited learnable parameters, the adaptability of the feature representations is limited. For example, Wang et al. adopted a prompt tuning strategy in their research, but this method is limited by limited parameters and cannot effectively adapt to incremental classes. (2) Representation-based Methods, which try to use the general representation of PTM to construct a classifier, but when facing downstream incremental datasets that are inconsistent with the distribution of the pre-training dataset, the performance is poor. For example, Zhou et al. used an adapter to change the general representation, but the continuous update of the adapter still faces the problem of catastrophic forgetting. (3) Model-mixture Methods, which design a set of models during the learning process and use model merging and model integration strategies to represent all tasks, but increase the memory cost of model parameters. For example, Wang et al. designed multiple models to merge different tasks, but this method brings additional computational burden.

[0008] It can be seen that the prior art has the following defects:

[0009] (1) Domain gap problem: In the prior art, there is a significant domain gap between the pre-training dataset and the downstream task dataset, which limits the performance of the pre-training model when directly applied to the downstream task. The domain gap mainly exists in the distribution, class, and features of the data, affecting the generalization ability of the model.

[0010] (2) Feature redundancy and inefficiency: Existing pre-training models often contain a large number of redundant features, which may not be completely useful in downstream tasks, or even may have a negative impact on performance. In addition, the feature representation of the model may not be compact and efficient, resulting in waste of computing resources.

[0011] (3) Poor adaptability of classifiers and feature representations: Traditional pre-training methods usually optimize the classifier and feature representation independently, lacking effective alignment and adaptability adjustment between the two. This leads to difficulties in adapting the classifier to the changing feature space in class incremental learning, affecting overall performance.

[0012] (4) Poor model scalability: Some existing incremental learning methods have poor scalability, making it difficult to adapt to a large increase in the number of classes, which limits the effectiveness of the model in handling large-scale data.

[0013] In summary, the above defects of the prior art result in poor classification performance of the model on classification tasks. SUMMARY

[0014] To solve the above problems existing in the prior art, the application provides a recognition method of pre-training model class incremental learning based on a neural collapse theory.

[0015] The recognition method of pre-training model class incremental learning based on the neural collapse theory comprises the following steps:

[0016] S100, acquiring a task data set and a constructed pre-training model, the task data set comprising a plurality of tasks and corresponding data of the tasks; each task corresponding to a plurality of samples carrying a class label; the pre-training model comprising a pre-training backbone model, a task-related adapter, an FCM module and an ETF classifier;

[0017] S200, in a training adaptation stage, training the pre-training model by using all samples corresponding to a first task in the task data set to obtain an adapted training model;

[0018] S300, in an incremental learning stage, freezing parameters of modules other than the FCM module in the adapted training model, and combining sample features of the first task with sample features of all current tasks other than the first task to perform incremental learning on the adapted training model to obtain a recognition model;

[0019] S400, using the recognition model to recognize an image corresponding to a target recognition task to obtain a class to which the target belongs.

[0020] Beneficial effects:

[0021] The recognition method of pre-training model class incremental learning based on the neural collapse theory provided by the application introduces the theory of NC into the feature adaptation of the pre-training model, provides a novel paradigm for solving downstream tasks by using the pre-training model, and constructs an ETF classifier capable of guiding the feature representation to adapt to an optimal structure. The ETF classifier not only satisfies the NC2 property in the neural collapse, but also can fix the optimal structure in advance to stabilize the training process. The FCM module is introduced to compress a plurality of features to specified class prototypes, reduce intra-class variability, satisfy the NC1 property in the neural collapse, align the feature representation with the optimal classifier structure through a supervised learning process, and thus obtain a more discriminative feature representation. The pre-training model is trained by incremental learning, and the trained model is used to perform a target recognition task, so that the recognition effect can be improved.

[0022] The application will be further described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1is a flowchart of a recognition method of pre-training model class incremental learning based on the neural collapse theory provided by the present application;

[0024] Figure 2 is a training process diagram of a pre-training model provided by the present application;

[0025] Figure 3 is a schematic diagram of the geometric principle of an equiangular tight frame (ETF) provided by the present application;

[0026] Figure 4 is a comparison diagram of the method provided by the present application and the prior art method;

[0027] Figure 5 is an effect diagram of recognizing targets in an unmanned aerial vehicle aerial image or a monitoring image provided by the present application;

[0028] Figure 6 is an effect diagram of recognizing targets in a satellite remote sensing image provided by the present application. DETAILED DESCRIPTION

[0029] The present application will be described in further detail below in conjunction with specific embodiments, but the embodiments of the present application are not limited thereto.

[0030] First, the application background of the present application is introduced.

[0031] Incremental learning (Class-Incremental Learning, CIL) has significant application value in target recognition, especially in the processing of radar images and remote sensing images. For example, in the field of military target recognition, especially in dealing with dynamic and complex battlefield environments. With the emergence of new targets (such as unmanned aerial vehicles, stealth fighters, and new missiles), traditional target recognition systems face the challenge of declining recognition accuracy. Incremental learning can quickly learn the features of new targets without forgetting existing targets, thereby improving the adaptability and real-time response capability of the recognition system. In multi-sensor data fusion, electronic warfare interference, and cross-modal target recognition, incremental learning helps the system to continuously optimize and ensure efficient recognition.

[0032] As shown in Figure 1 The present application provides a recognition method of pre-training model class incremental learning based on the neural collapse theory, which comprises:

[0033] S100, obtaining a task data set and a constructed pre-training model, the task data set comprising a plurality of tasks and corresponding data thereof; each task corresponding to a plurality of samples carrying class labels;

[0034] Reference Figure 2As shown, the pre-trained model of the present application includes: a pre-trained backbone model, a task-related adapter, an FCM module and an ETF classifier. The pre-trained backbone network is the basic part of the model, which is usually pre-trained on a large dataset and is used to extract preliminary features of the input data. The task-related adapter is used to bridge the domain gap between the pre-trained dataset and the downstream dataset, so that the model can better adapt to different downstream tasks. The feature compression module is designed to compress feature representation, reduce storage and computational costs, and improve the performance of the model. ETF, which stands for equiangular tight framework, is a special matrix structure used to build optimal classifiers.

[0035] The geometric diagram of the equiangular tight frame (ETF) is as follows Figure 3 As shown, ETFs The space consists of K vectors that satisfy a specific geometric relationship, that is, the angles between them are equal and optimal. This geometric structure enables the ETF classifier to construct an optimal classification boundary. The figure shows how four vectors (representing four categories) form a regular simplex, which is equivalent to the ETF in the neural collapse phenomenon. The ETF classifier can use the neural collapse phenomenon to guide the transfer of feature representations and construct the optimal classifier structure. In all incremental learning processes, the pre-trained backbone is frozen to maintain the generalization of feature representations.

[0036] This application involves incremental learning, which is a learning mode in which the model continuously learns new tasks on new categories. The task dataset contains T tasks, and the task dataset is represented by {D1, D2, ..., D T},in is the sample set of the tth task, containing n t samples, x i represents the i-th sample, y i Represents the category label of the i-th sample. The category sets of different tasks do not overlap. The category set is represented by C = {C i ,…,C T}, the total number of all visible categories is K t The goal of incremental learning is to learn a unified classifier that can accurately classify all visible categories. Consistent with the traditional PTM-CIL method, this application defines a pre-trained backbone model to serve as the initialization for the feature extractor f(·) and the linear classifier h(·).

[0037] Neural collapse refers to a phenomenon on balanced data, which reveals that there is a simplified geometric structure between the features of the last layer of the model and the classifier. This phenomenon is manifested as the close clustering of features of the same category, resulting in simplified classifier weights. Figure 3As shown, a simple Equiangular Tight Frame (ETF) refers to a matrix composed of K vectors in the space that satisfies the following properties: A matrix composed of K vectors in the space that satisfies the following properties:

[0038]

[0039] where, is a rotationally allowed matrix, and satisfies U T U = I K , I K is the identity matrix, and 1 K is a vector of all ones.

[0040] All column vectors in E have the same l2-norm, and the inner product between any pair of vectors is

[0041]

[0042] where, 1 when k1 = k2, and 0 otherwise.

[0043] Since neural collapse describes the optimal geometric structure between the last layer features of a model and the classifier, the present application proposes to fix a learnable classifier to the structure constructed by neural collapse to pre-allocate this optimality. Then, the present application uses an ETF classifier, constructs an ETF when initializing the classifier and fixes it during the training process. The present application hopes that the ETF can guide the general feature representation to the optimal classifier structure.

[0044] By observing neural collapse, the optimal classifier structure can be constructed to guide the migration of feature representation; the present application decides to use an ETF classifier to solve the PTM-CIL problem. Therefore, the present application first introduces a task-related adapter to bridge the domain gap between the pre-training dataset and the downstream task. Then, a feature compression module is introduced to compress the features into specified class prototypes, reduce the variability within the class, and align with a specific simple equiangular tight frame. Finally, an optimal structure alignment method is proposed to supervise the feature compression process. The overall method framework is as shown in Figure 2 .

[0045] Since there is a domain difference between the pre-training dataset and the dataset of the downstream task, the general feature representation of the pre-training model may not be discriminative for the downstream incremental task. Therefore, the present application introduces a lightweight adapter adjustment in the training adaptation stage to bridge the domain gap.

[0046] The pre-trained backbone model comprises L Transformer blocks, each of which comprises a self-attention module and a multi-layer perceptron (MLP) layer; in order to maintain the generalization ability of the pre-trained model, an adaptor module is learned for each MLP layer as a bypass branch. That is, the task-related adaptor comprises L adaptive adaptor modules; each multi-layer perceptron layer is provided with an adaptive adaptor module as a bypass branch, which is used to modulate downstream task information into general feature representation in the Transformer block; the adaptive adaptor module is a bottleneck module comprising a down-projection layer a non-linear activation function σ and an up-projection layer The adaptive adaptor module adjusts the output of the adaptive adaptor module to:

[0047] m o =σ(m i W down )W up +MLP(m i );

[0048] wherein m i and m o are the input and output of the MLP layer respectively, all L adaptive adaptor modules are collectively denoted as A, and the adaptive embedding function with the set A is denoted as f(x; A), represents the dimension of the down-projection layer.

[0049] In task-related adaptation, the pre-trained weights are frozen, and only the adaptors are optimized. In other learning steps, both the pre-trained model weights and the adaptors are frozen to ensure that the general feature representation has the ability to represent unseen classes.

[0050] Since the ETF classifier is introduced as a classifier for guiding the optimization of feature representation, the feature representation needs to be transferred to a specific class prototype to follow the characteristics of neural collapse. Therefore, a feature compression module (FCM) is introduced to compress various features. After the feature extractor, a feature compression module P(·) is added to convert the general feature representation into a feature that adapts to the pre-allocated ETF class prototype. The feature P(·) is composed of two layers of MLP and a batch normalization (BatchNorm) layer. In the incremental learning step, the feature compression module will be continuously fine-tuned to continuously adapt to the pre-allocated class prototype, and the adjustment process is represented as:

[0051]

[0052] In the formula, z i is the feature input into the FCM module, Features output by the FCM module.

[0053] To alleviate the representation overfitting of the ETF classifier, the application introduces a mixup strategy to increase feature diversity, which can be represented as:

[0054]

[0055] where λ is a random hyperparameter balancing the sample x i and x j , represents the mixed diversity feature, represents the mixed sample label.

[0056] Then the cross-entropy is used to optimize the learnable parameter P(·), denoted as:

[0057]

[0058] S200, in the training adaptation stage, using all samples corresponding to the first task in the task data set to train the pre-trained model to obtain an adapted training model;

[0059] As an optional embodiment of the application, S200 includes:

[0060] S210, inputting all samples corresponding to the first task in the task data set into the pre-trained backbone model to make the pre-trained backbone model output coarse-grained feature representation, and adjusting the domain gap of the coarse-grained features using the task-related adapter to obtain the first feature representation of all samples of the first task;

[0061] S220, inputting the first feature representation into the FCM module to convert it into a second feature representation that adapts to the pre-assigned ETF category prototype; and inputting the second feature representation into the ETF classifier to obtain the categories of all tasks corresponding to the first task;

[0062] S230, calculating the target loss function using the categories of all samples corresponding to the first task, and adjusting the parameters of the task-related adapter and the FCM module in the pre-trained model using the target loss function;

[0063] S240, replacing the task-related adapter and the FCM module in the pre-trained model with the adjusted parameters of the task-related adapter and the FCM module, and repeating the process of S210 to S220 once to obtain a trained pre-trained model.

[0064] The task-related adapter and the FCM integrated into the pre-trained model are trained in this embodiment, and their parameters are optimized using a target loss function L. The purpose of this step is to enable the pre-trained model to generate appropriate feature representations for the first task (D1).

[0065] In a specific embodiment of the present application, after S140, the recognition method based on the pre-trained model of the class incremental learning of the neural collapse theory further comprises:

[0066] a. The second feature representation output by the FCM module in the trained pre-trained model is stored in the first memory;

[0067] b. The mean and covariance of the classes of all samples corresponding to the first task are stored in the second memory.

[0068] Since the pre-trained model (Pre-Trained Model) and the adapter (Adaptor) are frozen in the incremental step, the present application needs to supervise the feature compression process to ensure that the feature transfer meets the characteristics of neural collapse (Neural Collapse). In addition, the present application also hopes to capture the class relationship between different task classes. Therefore, the present application designs two feature replay strategies to introduce the features of previous tasks in the current training step.

[0069] Full memory (Full Memory, FM): build memory M FM to store the features z of the samples.

[0070] Class memory (Class Memory, CM): build memory M CM to collect the mean μ of all sample classes.

[0071] The full memory of the present application is stored in the first memory, and the class memory is stored in the second memory.

[0072] Since the pre-trained model provides well-distributed and generalizable representations, each class tends to exhibit a unimodal distribution and can be naturally modeled as a Gaussian distribution The present application uses a reparameterization strategy to replay the features of previous tasks. By generating features from the class memory, the present application can obtain rich simulated features for each class to achieve optimal structural alignment, where the number of replay ρ is set to 256 in all experiments.

[0073] Thanks to the two feature replay strategies, the present application can reallocate the weights of the ETF classifier according to the number of seen classes to maintain the optimal classifier structure. In order to compress various features, the present application needs to combine the features with the specified corresponding class weights e iAlignment. Due to the large capacity gap between the randomly reassigned class weights and the feature representations, the optimal classifier alignment becomes challenging. The Batchnorm layer of the feature compression module can minimize collapse (Tian, Krishnan, and Isola 2019) by making fewer singular values approach zero to satisfy characteristics.

[0074] Therefore, the present application explores the use of soft maximum loss to achieve optimal structural alignment. In this way, the loss can be adjusted to compensate for poor alignment of features that may occur when the capacity gap is large. For simplicity, the present application uses a simple LogSum function, denoted as:

[0075]

[0076] where α is a smoothing factor, e k is the weight of the corresponding class, and W p is the weight of the feature compression module P. Experiments show that the soft maximum loss has better alignment performance than the mean square error (MSE) function.

[0077] Therefore, the target loss function is represented as:

[0078]

[0079] In the formula, δ is set to 10, τ is the temperature coefficient, α is the smoothing factor, e k is the weight of the corresponding class, and W p is the weight of the feature compression module.

[0080] The present embodiment constructs or updates the full-maintenance (FM) or class-maintenance (CM) according to the feature representation generated by the FCM. These matrices are used to store the feature information of the current task. Depending on the selection and integration of each base component in the pre-trained model. At the same time, the stored feature information will directly affect the adaptation effect of the subsequent task and the performance of the overall model.

[0081] S300, in the incremental learning stage, the parameters of the other modules in the adaptive training model except the FCM module are frozen, and all sample features of all current tasks except the first task are combined with the sample features of the first task to perform incremental learning on the adaptive training model to obtain a recognition model;

[0082] In a specific embodiment of the present application, S300 includes:

[0083] S310, freezing the parameters of the pre-training backbone model, the task-related adapter and the ETF classifier in the trained pre-training model unchanged, and setting the parameters of the FCM module to change with the task adaptability to obtain an incremental model;

[0084] S320, for any current task in the task data set except the first task, input all samples of the current task into the incremental model to output a first feature representation through the pre-training backbone model, extract a second feature representation from the first memory, combine the first feature representation of the current task and send them into the FCM module to obtain the second feature representation of the task, and store it into the first memory;

[0085] When processing a new task (D t , t > 1), the feature information of the previous task extracted in the previous step two (through FM / CM) is first replayed or integrated. This ensures that the model can utilize the previously learned knowledge, promoting knowledge transfer and cumulative learning.

[0086] When processing a new task (D t , t > 1), the parameters of the backbone model, the task-related adapter and the ETF classifier are frozen, and only the FCM further updated in step two is further updated to adapt to the features of the new task. Let the model maintain the memory of the old task while improving the ability to adapt to the new task quickly.

[0087] S330, input the second feature representation of the current task in S320 into the ETF classifier to obtain the class of all samples corresponding to the current task, and store the mean and covariance of the class into the second memory;

[0088] S340, calculating the target loss function using the class of all samples corresponding to the current task in S330, and adjusting the parameters of the task-related adapter and the FCM module in the pre-training model using the target loss function;

[0089] S350, replacing the task-related adapter and the FCM module in the pre-training model with the adjusted parameters of the task-related adapter and the FCM module, and repeating the process of S320 to S340 until all tasks in the task data set are completed, obtaining a recognition model, a first memory storing the second feature representation of all tasks, and a second memory storing the mean and covariance of the class of samples.

[0090] Repeat the process of S320 to S340 until all tasks (D1 to D T) are processed. In each iteration, the model learns new task-specific features while maintaining the memory of old task knowledge. This step is the core of the model's continual learning and optimization. Through continuous iterations over the entire dataset sequence, the model not only learns specific features for each task but also gradually develops a cross-task generalization ability. This ability enables the model to adapt faster and perform better on new tasks. Overall optimization is the key to improving model performance, relying on the synergistic effect of all previous steps and continuous optimization.

[0091] The present application constructs an Equiangular Tight Frame (ETF) classifier, which is an optimal classifier structure that can guide the adaptation of feature representation to the optimal structure. The ETF classifier not only satisfies the NC2 property in neural collapse (class mean features converge to an equiangular tight frame), but also avoids the instability of the classifier during training by pre-fixing the optimal structure.

[0092] A feature compression module (FCM) is introduced to compress multiple features to specified class prototypes, reducing intra-class variability and satisfying the NC1 property of neural collapse (intra-class feature variability tends to zero). The feature compression module aligns the feature representation with the optimal classifier structure through a supervised learning process, resulting in more discriminative feature representation.

[0093] Reference Figure 4 As shown, Figure 4 is a simple comparison of the method of the prior art and the method of the present application. Because different steps cannot be learned simultaneously, and the objective function may not be appropriate, the classifier structure is not optimal for all seen classes. That is, the geometric structure gradually presents a feature-tight form. However, the present method uses the neural collapse theory to assign an optimal classifier structure before model optimization, guiding the adaptation of feature representation. This allows the model to maintain a balanced distribution of features even after the introduction of new classes.

[0094] S400, using the recognition model to recognize the image corresponding to the target recognition task to obtain the class to which the target belongs.

[0095] Exemplarily, in radar images, as new types of targets (such as drones, new ships, aircraft, etc.) appear, the target recognition system needs to constantly update its recognition model without forgetting known targets. For example, an air monitoring system can use the method of the present application to incrementally learn the features of new aircraft or drones, and a maritime monitoring system needs to respond to the emergence of new ships or vessels. Incremental learning can ensure that the system maintains the ability to recognize traditional targets while recognizing new targets, thereby effectively responding to complex and dynamic environments.

[0096] In remote sensing images, target categories are also constantly changing, especially in the fields of agricultural monitoring, urban planning, and disaster response. As new crop varieties, building forms, or natural disaster types emerge, remote sensing systems need to incrementally learn these new categories. For example, the method of the present application can help remote sensing systems gradually adapt to new crop varieties in agricultural monitoring, or identify newly built buildings and roads in urban expansion monitoring. In post-disaster response, incremental learning can help the system identify disaster areas and new changes in monitoring in a timely manner, and maintain effective recognition ability for historical data.

[0097] In addition, in applications such as autonomous driving, drones, and environmental monitoring, class-incremental learning also plays an important role. Autonomous driving systems need to handle changing traffic environments, and the method of the present application helps the system adapt to new types of vehicles, pedestrians, and traffic signs while maintaining the ability to detect existing targets. When performing tasks such as agricultural monitoring and wildlife protection, drones also need to identify new species or environmental changes through incremental learning.

[0098] Reference Figure 5 The present application can be applied to target recognition in unmanned aerial vehicle aerial photography or monitoring images, reference Figure 6 The present application can also be applied to target recognition in satellite remote sensing images. From Figure 5 and Figure 6 It can be seen that the present application can locate the target in the image and accurately classify it.

[0099] Most of the existing methods, whether they are prompt-based methods, representation methods, or model mixing methods, do not explicitly use the theory of neural collapse (Neural Collapse) to guide the optimization of feature representation and classifier. However, the present application first introduces the theory of neural collapse into the feature adaptation of pre-trained models (PTM), providing a novel paradigm for using pre-trained models to solve downstream tasks.

[0100] It is to be noted that the terms "first", "second", and the like in the description do not necessarily connote an absolute sequence or order among others, but can be used to modify a common substitute to trace their introduction in the application description. By definition, the "plurality" means two or more, unless specifically defined otherwise in the description.

[0101] Although the present application has been described in connection with various embodiments thereof, it will be understood that other modifications will be apparent to those of ordinary skill in the art and can be made without departing from the spirit and scope of the application, which are indicated by the following claims. In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite articles "a" or "an" do not exclude a plurality.

[0102] The above description is further to the detailed description of the application in connection with specific preferred embodiments, and cannot be deemed to limit the specific implementation of the application to these descriptions. For those skilled in the art, some simple deductions or replacements can be made without departing from the concept of the application, and all of them shall be regarded as falling within the protection scope of the application.

Claims

1. A recognition method based on pre-training model class incremental learning of neural collapse theory, characterized in that, The method comprises the following steps: S100, obtaining a task data set and a constructed pre-training model, wherein the task data set comprises a plurality of tasks and corresponding data of the tasks; Each task corresponds to a plurality of samples carrying a category label; the pre-training model comprises a pre-training backbone model, a task-related adapter, an FCM module and an ETF classifier; the ETF classifier is an equiangular tight frame classifier, and the FCM module is a feature compression module; S200, in a training adaptation stage, training the pre-training model by using all samples corresponding to a first task in the task data set to obtain an adapted training model; S300, in an incremental learning stage, freezing parameters of modules other than the FCM module in the adapted training model, and combining sample features of the first task with sample features of all current tasks other than the first task to perform incremental learning on the adapted training model to obtain a recognition model; S400, using the recognition model to recognize an image corresponding to a target recognition task to obtain a category to which the target belongs; S200 comprises the following steps: S210, inputting all samples corresponding to the first task in the task data set into the pre-training backbone model to enable the pre-training backbone model to output coarse-grained feature representation, and adjusting a domain gap of the coarse-grained feature by using the task-related adapter to obtain first feature representation of all samples of the first task; S220, inputting the first feature representation into the FCM module to convert the first feature representation into second feature representation adapted to an ETF category prototype, and inputting the second feature representation into the ETF classifier to obtain a category of all tasks corresponding to the first task; S230, calculating a target loss function by using categories of all samples corresponding to the first task, and adjusting parameters of the task-related adapter and the FCM module in the pre-training model by using the target loss function; S240, replacing the task-related adapter and the FCM module in the pre-training model with the task-related adapter and the FCM module with adjusted parameters, and repeating the process of S210 to S220 once to obtain a trained pre-training model. 2.The identification method of pre-training model class incremental learning based on neural collapse theory according to claim 1, wherein, The task data set includes T tasks, and the task data set is represented as {D1, D2, ..., D T },in is the sample set of the tth task, containing n t samples, x i represents the i-th sample, y i Represents the category label of the i-th sample. The category sets of different tasks do not overlap. The category set is represented by C = {C i ,…,C T }, the total number of all visible categories is K t . 3.The identification method of pre-training model class incremental learning based on neural collapse theory according to claim 1, wherein, After S240, the recognition method based on the pre-training model of the neural collapse theory and incremental learning further comprises the following steps: Storing the second feature representation output by the FCM module in the trained pre-training model into a first memory; Storing a mean value and a covariance of categories of all samples corresponding to the first task into a second memory. 4.The identification method of pre-training model class incremental learning based on neural collapse theory according to claim 3, wherein, S300 comprises the following steps: S310, freezing parameters of the pre-training backbone model, the task-related adapter and the ETF classifier in the trained pre-training model, and setting parameters of the FCM module to change with task adaptability to obtain an incremental model; S320, for any current task other than the first task in the task data set, inputting all samples of the current task into the incremental model to output first feature representation by the pre-training backbone model, extracting second feature representation from the first memory, combining the first feature representation of the current task and inputting the combination into the FCM module to obtain second feature representation of the task, and storing the second feature representation into the first memory; S330, input the second feature representation of the current task in S320 into the ETF classifier to obtain the class of all samples corresponding to the current task, and store the mean and covariance of the class in the second memory; S340, calculate the target loss function using the class of all samples corresponding to the current task in S330, and adjust the parameters of the task-related adapter and the FCM module in the pre-trained model using the target loss function; S350, replace the task-related adapter and the FCM module in the pre-trained model with the task-related adapter and the FCM module with adjusted parameters, and repeat the process of S320 to S340 until all tasks in the task dataset are completed, to obtain the recognition model, the first memory storing the second feature representation of all tasks, and the second memory storing the class mean and covariance of the samples.

5. The identification method of the pre-training model based on the neural collapse theory incremental learning according to any one of claims 1-4, characterized in that, The pre-trained backbone model comprises L Transformer blocks, each of which comprises a self-attention module and a multi-layer perceptron (MLP) layer; the task-related adapter comprises L adaptive adapter modules; each multi-layer perceptron layer is provided with an adaptive adapter module as a shunt branch, which is used to modulate downstream task information into a generalization feature representation in the Transformer block; the adaptive adapter module comprises a lower projection layer a non-linear activation function σ and an upper projection layer The adaptive adapter module adjusts the output of the adaptive adapter module to: m o = σ(m i W down )W up + MLP(m i ); where m i and m o are the input and output of the MLP layer, respectively, A is the set of all L adaptive adapter modules, f(x; A) is the adaptive embedding function with set A, denotes the dimension of the lower projection layer.

6. The identification method of the pre-training model based on the neural collapse theory incremental learning according to any one of claims 1-4, characterized in that, The feature compression module includes two layers of MLP and a batch normalization layer; the feature compression module will be continuously fine-tuned to continuously adapt to the pre-allocated class prototype, and the adjustment process is represented as: z i = f(x i ); wherein z i Features input to the FCM module, Features output from the FCM module.

7. The identification method of the pre-training model based on the neural collapse theory incremental learning according to any one of claims 1-4, characterized in that, The ETF classifier introduces a hybrid strategy in the classification process to increase feature diversity, which is represented as: where λ is the balanced sample x i and x j 's random hyperparameters, denotes the mixed diversity feature, denotes the mixed sample label.

8. The identification method of incremental learning of a pre-training model based on the neural collapse theory according to any one of claims 1 or 4, characterized in that, The target loss function is represented as: where δ is set to 10, τ is the temperature coefficient, a is the smoothing factor, e k is the weight of the corresponding class, W p is the weight of the feature compression module.

Images