Precise prototype aircraft classification method based on small sample class incremental learning

Abstract

The invention belongs to the field of sample classification, and relates to a prototype accurate aircraft classification method based on small sample class incremental learning, which comprises the following steps: acquiring an aircraft image and inputting the aircraft image into a trained aircraft classification model to obtain a classification result; the training process of the aircraft classification model comprises the steps that an aircraft image data set is acquired and preprocessed, and a basic data set and an incremental data set are obtained; training each module of the aircraft classification model by using the basic data set to obtain each trained module and a basic prototype; fixing the trained feature extraction module, and performing fine adjustment on the trained prototype refining module, the feature space adaptive module and the persistent evolution classifier by using the incremental data set and the basic prototype to obtain a trained aircraft classification model; according to the method, the prototype is stored as the graph, when a new prototype is learned, graph nodes are added, the position of the previous prototype is not covered, the prototype of a new category is enhanced according to the graph, and therefore classification accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the technical fields of reinforcement learning, sample classification processing, and computer vision, and relates to an aircraft classification method with precise prototypes based on few-shot class incremental learning. Background Art

[0002] At present, aircraft have become a fast means of transportation and play an important role in the military aviation field. However, sometimes due to problems with pictures, such as scarcity, blurriness, etc., the model should be able to achieve the effect of aircraft classification without relying on too much data. However, the difficulty is obvious. The structures of each aircraft are almost similar, all having a long fuselage and a pair of wings. Therefore, it is necessary to design a model that can refine according to the representation of the important features of the aircraft and can adaptively adjust the feature space of the learned aircraft classes.

[0003] In recent years, deep neural networks have achieved excellent performance in various visual tasks, and even in single tasks such as object recognition, their performance has exceeded that of humans. However, these methods largely rely on large-scale datasets to train models that can learn a limited number of classes. In the real world, data usually appears sequentially in a stream.

[0004] Machine learning models are required to continuously learn new classes as the data changes while retaining the knowledge learned from previous data, which is called class incremental learning. The main challenge of class incremental learning is the catastrophic forgetting problem: after the model is trained on a new task, it will forget the previous knowledge because the parameters in the network are updated after the model learns the knowledge of the new class, and the knowledge of the old class is overwritten. Therefore, class incremental learning aims to embed the learned knowledge into the previously learned knowledge and gradually construct a general classifier. However, they assume that the model is continuously trained on new class data with sufficient samples.

[0005] However, this assumption does not hold in the application of aircraft classification. For example, in the field of aircraft classification data, the acquisition of samples of new classes may be limited. In this case, the model may not be able to continuously train on sufficient new class samples. In fact, in real life, new class data is usually scarce due to privacy issues or expensive collection costs, resulting in the model being trained only based on limited new data. This task of continuously learning new classes with limited data is called few-shot class incremental learning. Its basic paradigm is to initialize the model with a set of base classes with sufficient data, then learn in the form of N-way K-shot, and then continuously introduce new classes and update the model through incremental learning. In addition to the challenge of catastrophic forgetting, few-shot class incremental learning also faces the challenge of overfitting to limited training samples, making it more difficult for the model to learn new classes and making the model no longer adaptable to incremental tasks.

[0006] The research on small-sample class incremental learning based on class increment focuses on dealing with the situation of a limited number of samples in the incremental stage. This application scenario is closer to reality and more challenging. There are five mainstream methods in the current research methods, and most of the recent research work focuses on meta-learning, features and feature spaces, and dynamic network structure methods. The meta-learning-based method aims to learn a prior knowledge in the incremental stage and quickly identify new classes based on the existing knowledge in the incremental stage. A random selection strategy is used to enhance the scalability of feature representation, and a prototype refinement mechanism is introduced to update the prototype to enhance the expressive ability of new classes and retain the relationship features with old classes. The method based on features and feature spaces aims to decouple features or map them to a lower-dimensional space to obtain more useful features. Multiple virtual prototypes are pre-assigned positions in the embedding space. By optimizing these virtual prototypes, more space is reserved for future new classes. It is also possible to extract invariant information between the base session and the incremental session by optimizing multiple pseudo-tasks fabricated with the base dataset. The method based on dynamic network structure aims to dynamically adjust the network structure according to the input features during the continuous training process. The feature extractor and the classifier are decoupled. After the base session ends, the parameters of the feature extractor are frozen, and the classifier is continuously updated and expanded using the graph attention mechanism network during the increment to alleviate catastrophic forgetting and prevent overfitting.

[0007] The above methods simply aggregate samples of the same category together. When the number of prototypes gradually increases, the data distribution may be different from that in the previous stage, resulting in a change in the mapping relationship between the sample output space and the feature space. Unseen prototypes will cover the positions of previously seen prototypes, causing knowledge confusion. Therefore, the method based on simple sample centers as prototypes may have certain limitations in solving challenges such as small-sample class incremental learning. Summary of the Invention

[0008] To solve the above existing technical problems, the present invention adopts a prototype-precise aircraft classification method based on small-sample class incremental learning, including: obtaining aircraft images, inputting the aircraft images into a trained aircraft classification model to obtain an aircraft classification result; the aircraft classification model includes: a feature extraction module, a prototype refinement module, a feature space adaptation module, and a continuously evolving classifier;

[0009] The training process of the aircraft classification model includes:

[0010] S1. Obtain an image dataset of aircraft, preprocess the image dataset of aircraft to obtain a base dataset and an incremental dataset;

[0011] S2, using the basic data set to train the feature extraction module, the prototype refinement module and the feature space adaptation module to obtain the trained feature extraction module, the prototype refinement module, the feature space adaptation module and the basic prototype;

[0012] S3. Fix the trained feature extraction module, use the incremental data set and basic prototype to fine-tune the trained prototype refinement module, feature space adaptive module and continuous evolution classifier to obtain the trained aircraft classification model.

[0013] Beneficial effects:

[0014] 1. The present invention discusses the closely related inter-class separability and intra-class compact paradigm as two independent issues separately, thereby constructing a classification loss function and a prototype loss function for training respectively, so that the distribution of the feature space is more suitable for the embedding of new classes in the future; 2. In order to enable the model to have the ability of continuous learning, the present invention records and saves the learned prototypes of each category in the form of a graph. When a prototype of a new category is learned, the nodes of the graph structure will be dynamically added, so as not to cover the position of the previously seen prototype and cause knowledge confusion. The graph structure is used as the classification basis of the classifier; 3. The present invention adopts an attention mechanism to enhance the prototype of the new category according to the prototype graph structure, so that the prototype of the new category can pay attention to the information of different prototypes from different positions. For those new classes with little data, the prototype usually contains a lot of noise. The attention mechanism can reduce the noise and make the prototype more expressive, thereby improving the accuracy of classification; 4. The model refinement module of the present invention uses a self-attention mechanism to calculate the relationship matrix between samples, calculates the importance of the features of each sample to the prototype of the category to which it belongs according to the relationship, performs weighted aggregation according to the importance, and obtains the refined prototype, so that the model pays more attention to the features related to the category representation, improves the sensitivity of the model to the category features, and better captures the rich feature space of the category. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A structural diagram of a prototype accurate aircraft classification method based on incremental learning of small sample classes provided by an embodiment of the present invention;

[0016] Figure 2 A flowchart of a prototype-accurate aircraft classification method based on small sample class incremental learning provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0017] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0018] To solve the above-mentioned problems in the prior art, the present invention adopts an aircraft classification method with precise prototypes based on few-shot class incremental learning, including: obtaining aircraft images, inputting the aircraft images into a trained aircraft classification model to obtain an aircraft classification result; the aircraft classification model includes: a feature extraction module, a prototype refinement module, a feature space adaptation module, and a continuously evolving classifier.

[0019] As Figure 1 、 Figure 2 shown, the training process of the aircraft classification model includes:

[0020] S1. Obtain an image dataset of aircraft, preprocess the image dataset of aircraft to obtain a basic dataset and an incremental dataset;

[0021] Preprocessing the image dataset of aircraft includes: classifying the data samples in the image dataset of aircraft according to category labels, and selecting the category labels with the top few numbers of data samples, and randomly selecting multiple data samples from the data samples corresponding to each selected category label as the basic dataset, and randomly selecting multiple data samples from the data samples corresponding to the unselected category labels as the incremental dataset.

[0022] The categories of aircraft include Boeing737, Boeing747, Boeing777, A330, etc.

[0023] S2. Use the basic dataset to train the feature extraction module, the prototype refinement module, the feature space adaptation module, and the continuously evolving classifier to obtain a trained feature extraction module, a prototype refinement module, a feature space adaptation module, a continuously evolving classifier, and a basic prototype;

[0024] The randomly initialized parameters of the feature extraction module and the feature space adaptation module are α and β respectively.

[0025] Using the basic dataset to train the feature extraction module, the prototype refinement module, the feature space adaptation module, and the continuously evolving classifier includes:

[0026] S21. Construct the support set base of each category j in the basic dataset D and the query set

[0027] Specifically, construct the basic dataset D base The support set of class j in and the query set include:

[0028] Set the parameter m, and use the first m data samples belonging to class j in the basic dataset D base as the initial support set of class j Use the data samples after m as the initial query set of class j Randomly rotate the data samples in the support set and the query set respectively to obtain the auxiliary support set and the auxiliary query set Concatenate the auxiliary support set with the corresponding support set to obtain the final support set Concatenate the auxiliary query set and the query set to obtain the final query set Preferably, m is 15.

[0029] Samples are classified according to their distance from the prototype, and a sample belongs to the class with the closest prototype. Metric learning requires that the two being compared must belong to the same feature space. Therefore, the samples in the query set must first pass through the feature extractor module and the prototype refinement module.

[0030] S22. Input the support set and the query set of each class j into the feature extraction module to obtain the feature representations of the support set and the query set of each class j where is the feature representation of the data sample k of class j in the support set , is the feature representation of the data sample k of class j in the query set , is the feature representation of the data sample k of class j in the initial support set , is the feature representation of the data sample k of class j in the auxiliary support set is the feature representation of the data sample k of class j in the initial query set is the feature representation of the data sample k of class j in the auxiliary query set.

[0031] In one embodiment, the feature extraction module is a residual network.

[0032] S23. For each class \(j\), the support set and the query set feature representations are input into the prototype refinement module to obtain the relationship matrix for each class \(j\) and the refined prototype \(p\) for each class \(j\) j ′; where \(N\) is the number of classes in the base dataset, and the prototype set \(P′ = [p′_1, p′_2,..., p′ N ;

[0033] Specifically, the prototype refinement module processes the feature representation including:

[0034] S231. Perform different linear transformations on the feature representation to obtain the feature Perform a linear transformation on the total feature representation to obtain the feature where \(\psi_1\), \(\psi_2\), and \(\psi_3\) represent different linear transformations;

[0035] S232. Calculate the relationship matrix for class \(j\) based on the feature and The formula for the relationship matrix

[0036] The relationship matrix can be briefly described as:

[0037]

[0038] where is the dimension of the feature and is the relationship between the data sample \(k\) in the support set and all data samples of its class \(j\).

[0039] Using the above self-attention mechanism, each prototype can enhance the prototype by attending to information from different prototypes at different positions to capture such relevant information. For those new classes with little data, the prototypes usually contain a lot of noise, which can be reduced by the information aggregation of the above self-attention mechanism and make the prototypes more expressive.

[0040] S233. Calculate the class center \(p\) for class \(j\) based on the relationship matrix for class \(j\) ; j ;

[0041] According to its own and surrounding feature pairs perform correction to obtain the class center \(P = [p_0, p_1,... pN :

[0042]

[0043] Among them, is the number of samples of class j in the support set, and p represents the class center of the j-th class. j is the set of data samples in the support set that belong to class j.

[0044] S234. Calculate the weight of each data sample k in the support set j according to the class center p ;

[0045] Based on the feature and distance tensor, calculate the importance of the sample and assign a weight z to each sample kj :

[0046]

[0047] where w kj represents the distance of the data sample k in the support set from the prototype p j ; l is the index of the data sample in the support set as follows:

[0048]

[0049] where T is the temperature parameter used to reduce the overall value.

[0050] S235. Calculate the refined prototype representation p ' of class j according to the weight z kj of each data sample k in the support set j '.

[0051] After obtaining the importance of the features of each sample for the prototype of the class to which it belongs and using it as the weight for prototype refinement, the final formula is expressed as:

[0052]

[0053] where f θ is the mapping function, and N way is the total number of classes.

[0054] S24. Map the refined prototype representation p j ' of each class j to the spherical space to obtain the prototype W j of the high-dimensional feature space structure of each class j; map the prototype W j ​Sum relation matrix The input feature space adaptive module obtains the prototype loss function value, and updates the parameters of the feature extraction module, the prototype refinement module, and the feature space adaptive module according to the prototype loss function value. When the prototype loss function value is minimized, the fine-tuned feature extraction module, prototype refinement module, and feature space adaptive module are obtained;

[0055] Specifically, to achieve intra-class compactness, the feature space adaptive module of the present invention calculates the distance between the sample feature vectors of each class and their prototypes, and strives to minimize these distances to obtain the prototype loss function L intra :

[0056]

[0057] where f e is a linear transformation, and the weight parameter of f e is the parameter of the feature space adaptive module.

[0058] The loss function L intra decouples the goals of inter-class separability and intra-class compactness. It focuses on binding the features of the same class more closely to achieve intra-class compactness, and pushing the features of different classes further apart to achieve inter-class separability. In the subsequent incremental learning stage, it reserves enough space for unseen classes and enhances the adaptability of the classifier.

[0059] The core of inter-class separability lies in maximizing the gap between class means on the hypersphere, and the core of intra-class compactness lies in encouraging the features of the same class to concentrate at one point.

[0060] S25. Input the support set and the query set of each category j into the fine-tuned feature extraction module and prototype refinement module to obtain the refined prototype of each category j; map the refined prototype of each category j to the spherical space to obtain the prototype of the fine-tuned high-dimensional feature space structure of each category j where W j ′ is the prototype of category j;

[0061] S26. Input the query set Q base and the prototype W′ of the fine-tuned high-dimensional feature space structure into the persistent classifier to obtain the scores of each sample in the query set Q base and each prototype W j ′; where, N way is the number of categories in the base dataset;

[0062] Different from conventional predictions, the present invention uses a distance metric to represent the proximity to each class; specifically, for each data sample x base in the query set Q k the distance to the prototype of each class is calculated; for each sample x k the distance to each prototype W j ′ is given by the formula: d(x k ,W′ j ).

[0063]

[0064] Since distances are all positive, the smallest distance represents the most likely class, i.e., the predicted class.

[0065] For each sample x k the distance d(x k ,W′ j ) to each prototype is negated to obtain the distance vector for each sample x k . The softmax function is used to process the distance vector D k for each sample x k to obtain the scores s k of each sample x j with respect to each prototype W k,j .

[0066] S27. According to the scores s base of each sample x k in the query set Q j with respect to each prototype W k,j and the true class label y true the value of the classification loss function is calculated, and the feature extraction module, the prototype refinement module, and the continuously evolving classifier are updated according to the value of the classification loss function. When the value of the classification loss function is minimized, the trained feature extraction module, prototype refinement module, feature space adaptation module, and continuously evolving classifier are obtained; the parameters of the trained feature extraction module and feature space adaptation module are A and B.

[0067] Specifically, the formula for the classification loss function L c is:

[0068]

[0069] where y true,k,j represents the true label (0 or 1) that the data sample k belongs to class j.

[0070] The prototype representation W′ of the fine-tuned high-dimensional feature space structure obtained from the last iteration of training is used as the base prototype P base .

[0071] S3. Fix the trained feature extraction module, and use the incremental dataset and the base prototypes to finely tune the trained prototype refinement module, feature space adaptation module, and continuously evolving classifier, so as to obtain a trained aircraft classification model.

[0072] Randomly initialize the trained feature space adaptation module, and the parameter of the randomly initialized feature space adaptation module is γ.

[0073] The specific steps of using the incremental dataset and the base prototypes to train the trained prototype refinement module, feature space adaptation module, and continuously evolving classifier include:

[0074] S31. Construct the incremental dataset D n The support set of each category i in and the query set

[0075] Construct the incremental dataset D n The support set of each category i in and the query set The specific steps are similar to step S21, and we can obtain where is the final support set, is the initial support set, is the auxiliary support set, is the final query set, the initial query set, is the auxiliary query set.

[0076] S32. Input the support set and the query set of each category i into the trained feature extraction module, and obtain the feature representations and of the support set where is the feature representation of the data sample k of category i in the support set, is the feature representation of the data sample k of category i in the initial support set, is the feature representation of the data sample k of category i in the auxiliary support set, is the feature representation of each data sample k of category i in the query set, is the feature representation of the data sample k of category i in the initial query set, is the feature representation of the data sample k of category i in the auxiliary query set.

[0077] S33. Input the feature representations of the support set and query set of each category \(i\) into the prototype refinement module to obtain the relationship matrix of each category \(i\) and the refined prototype representation \(p''\); and the query set of the feature representation input into the prototype refinement module to obtain the relationship matrix of each category \(i\) and the refined prototype representation \(p''\); and the refined prototype representation \(p\) i ″;

[0078] The specific steps are similar to those in step S23, and finally the relationship matrix The center of each category \(P'' = [p''_0, p''_1,... p'' M ; where \(M\) is the number of categories in the incremental data set, is the relationship between the data sample \(k\) of category \(i\) in the support set and all data samples of category \(i\).

[0079] S34. Map the refined prototype representation \(p\) i ″ of each category \(i\) to the spherical space to obtain the prototype \(W\) of the high-dimensional feature space structure of each category \(i\) i ; Map the prototype \(W\) of the high-dimensional feature space structure of each category \(i\) i and the relationship input into the feature space adaptation module to obtain the prototype loss function value, and update the parameters of the prototype refinement module and the feature space adaptation module according to the prototype loss function value. When the prototype loss function value is the smallest, obtain the fine-tuned prototype refinement module and feature space adaptation module;

[0080] The specific steps are similar to those in step S24.

[0081] S35. Input the support set and query set of each category \(i\) into the trained feature extraction module and the fine-tuned prototype refinement module to obtain the refined prototype representation of each category \(i\); Map the refined prototype representation of each category \(i\) to the spherical space to obtain the fine-tuned prototype representation \(W'\) of the high-dimensional feature space structure of each category \(i\); and the query set input into the trained feature extraction module and the fine-tuned prototype refinement module to obtain the refined prototype representation of each category \(i\); Map the refined prototype representation of each category \(i\) to the spherical space to obtain the fine-tuned prototype representation \(W'\) of the high-dimensional feature space structure of each category \(i\); i ′;

[0082] S36. Construct a prototype graph structure based on the base prototype and the fine-tuned prototype representation \(W'\) of the high-dimensional feature space structure of each category \(i\), and input the query set \(Q\) and the prototype graph structure into the persistent classifier to obtain the scores of each data sample in the query set \(Q\); where, i ′ construct a prototype graph structure, and input the query set \(Q\) n and the prototype graph structure into the persistent classifier to obtain the scores of each data sample in the query set \(Q\); n where,

[0083] Specifically, constructing the prototype graph structure includes:

[0084] S361. According to the base prototype and the fine-tuned high-dimensional feature space structure \(W\)i 'Construct the prototype graph structure;

[0085] If it is the first iteration training in the incremental stage, then combine the basic prototype P after the end of step S2 (i.e., the initial stage) base and the W learned in this iteration i 'Construct the prototype graph structure constructed by the new prototype and the basic prototype, and the prototype graph structure is represented as P total ={P base ,P inc}; Otherwise, dynamically update the prototype graph structure P obtained from the previous iteration training according to the W learned in this iteration i ={P total ,P base ,P inc} corresponding prototype P inc ; Among them, the prototype P inc ={W1',W2',...,W' M}.

[0086] S362. Use the attention mechanism to enhance the prototype W of the fine-tuned high-dimensional feature space structure for each category i according to the prototype graph structure i ' to obtain the enhanced prototype W of each category i i ″.

[0087] Specifically, each prototype can focus on the information of different prototypes from different positions by using self-attention. In particular, for those new classes with little data, the prototypes usually contain a lot of noise. It can be reduced by information aggregation and make the prototypes more expressive. The formula is as follows:

[0088]

[0089] where d total is the dimension of the feature P total , and a i is the attention score of category i

[0090] The enhanced prototype W of the i-th category i ″ = W i '+ w(W i ')·a i ; where w() is the weight matrix

[0091] The persistent classifier processes the query set Q n and the prototype graph structure, including:

[0092] Use softmax to predict {P' inc ,Q n}, and the prediction process is the same as that in step S26, to obtain the query set Q neach sample x in k and each prototype W i ″ has a score s k,i ; where the prototype P′ inc ={W1″, W2″,..., W M ″};

[0093] S37. Calculate the classification loss function value according to the score of each sample and the true class label y true Update the prototype refinement module and the continuously evolving classifier according to the classification loss function value. When the classification loss function value is minimized, the trained prototype refinement module, the feature space adaptation module, and the continuously evolving classifier are obtained. The calculation of the classification loss function value is the same as that in step S27.

[0094] The above - mentioned embodiments further elaborate on the purpose, technical solution, and advantages of the present invention. It should be understood that the above - mentioned embodiments are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made to the present invention within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. An aircraft classification method with precise prototypes based on few-shot class incremental learning, characterized in that Including: Obtain an aircraft image, input the aircraft image into a trained aircraft classification model, and obtain an aircraft classification result; The aircraft classification model includes: a feature extraction module, a prototype refinement module, a feature space adaptation module, and a continuously evolving classifier; The training process of the aircraft classification model includes: S1. Obtain an image dataset of aircraft, preprocess the image dataset of aircraft to obtain a basic dataset and an incremental dataset; S2. Use the basic dataset to train the feature extraction module, the prototype refinement module, the feature space adaptation module, and the continuously evolving classifier to obtain a trained feature extraction module, a prototype refinement module, a feature space adaptation module, a continuously evolving classifier, and a basic prototype; S3. Fix the trained feature extraction module, and use the incremental dataset and the basic prototype to fine-tune and train the trained prototype refinement module, the feature space adaptation module, and the continuously evolving classifier to obtain a trained aircraft classification model.

2. The aircraft classification method with precise prototype based on few-shot class incremental learning according to claim 1, wherein Preprocessing the image dataset of aircraft includes: classifying the data samples in the image dataset of aircraft according to the class labels, selecting multiple class labels with the most data samples, and randomly selecting multiple data samples from the data samples corresponding to each selected class label as the basic dataset, and randomly selecting multiple data samples from the data samples corresponding to the unselected class labels as the incremental dataset.

3. A method for accurately classifying aircraft with precise prototypes based on few-shot class incremental learning according to claim 1, characterized in that, Training the feature extraction module, the prototype refinement module, the feature space adaptation module, and the continuously evolving classifier using the basic dataset includes: S21. Construct the basic dataset D base The support set of each category j in and the query set S22. Input the support set of each category j and the query set into the feature extraction module to obtain the feature representations of the support set and the query set of each category j S23. Represent the features of each category j and input them into the prototype refinement module to obtain the relationship matrix of each category j and the refined prototype p j ′; S24. Map the refined prototype p j ′ of each category j to the spherical space to obtain the prototype W of the high-dimensional feature space structure of each category j j ; Input the prototype W j of the high-dimensional feature space structure of each category j and the relationship matrix into the feature space adaptive module to obtain the prototype loss function value, and update the parameters of the feature extraction module, prototype refinement module, and feature space adaptive module according to the prototype loss function value. When the prototype loss function value is the smallest, obtain the fine-tuned feature extraction module, prototype refinement module, and feature space adaptive module; S25. Input the support set of each category j and the query set into the fine-tuned feature extraction module and the prototype refinement module to obtain the refined prototype of each category j; map the refined prototype of each category j to the spherical space to obtain the prototype W j ' of the fine-tuned high-dimensional feature space structure of each category j; S26. Input the query set of each category j and the prototype W of the fine-tuned high-dimensional feature space structure j ' into the persistent classifier, and obtain the scores of each sample in the query set of each category j with each prototype W j '; S27. Calculate the classification loss function value according to the scores, and update the feature extraction module, the prototype refinement module, and the continuously evolving classifier according to the classification loss function value. When the classification loss function value is minimized, obtain the trained feature extraction module, prototype refinement module, feature space adaptation module, and continuously evolving classifier; use the prototype of the fine-tuned high-dimensional feature space structure obtained from the last iteration training as the basic prototype P base .

4. An aircraft classification method with accurate prototypes based on small-sample class incremental learning according to claim 3, characterized in that, Construct the basic dataset D base The support set of medium category j and the query set include: Set parameter m, and use the first m data samples belonging to class j in the basic dataset D base as the initial support set of class j Use the data samples after m as the initial query set of class j For the support set and the query set Randomly rotate the data samples in them respectively to obtain the auxiliary support set and the auxiliary query set Concatenate the auxiliary support set with the corresponding support set to obtain the final support set Concatenate the auxiliary query set and the query set to obtain the final query set 5. A method for accurately classifying aircraft prototypes based on few-shot class incremental learning according to claim 3, characterized in that, The prototype refinement module processes the feature representation The processing includes: S231. Perform different linear transformations on the feature representation to obtain the feature Perform a linear transformation on the total feature representation to obtain the feature S232. Calculate the relationship matrix of category j according to the feature and ​ S233. Calculate the class center p of class j according to the relationship matrix of class j Calculate the class center p of class j j ; S234. Calculate the support set according to the category center p j Calculate the support set and the weight z of each data sample k in kj ; S235. Calculate the refined prototype representation p ' of class j according to the weight z of each data sample k in the support set kj j '.​ 6. The aircraft classification method with accurate prototypes based on few-shot class incremental learning according to claim 5, characterized in that Calculate the relationship between each data sample k and all data samples of its belonging category j Including: Among them, is the feature dimension.

7. The aircraft classification method with precise prototype based on few-shot class incremental learning according to claim 5, characterized in that, Calculation support set The weights of each data sample k in Among them, w kj represents the feature of the data sample k in the support set and the distance from the prototype p . l is the index of the data sample in the support set j , T is the temperature parameter , and is the number of data samples in the support set .

8. A method for accurately classifying aircraft with precise prototypes based on few-shot class incremental learning according to claim 1, characterized in that, Fine-tuning and training the trained prototype refinement module, the feature space adaptation module, and the continuously evolving classifier using the incremental dataset and the basic prototype includes: S31. Construct the incremental dataset D n the support set of each category i in and the query set S32. Input the support set of each category i and the query set into the trained feature extraction module to obtain the feature representations of the support set and the query set of each category i S33. Input the feature representation of each category i into the prototype refinement module to obtain the relationship matrix of each category i and the refined prototype representation p i ″; S34. Map the refined prototype representation p of each category i i ″ to the spherical space to obtain the prototype W of the high-dimensional feature space structure of each category i i ; Input the prototype W of the high-dimensional feature space structure of each category i i and the relationship into the feature space adaptive module to obtain the prototype loss function value, and update the parameters of the prototype refinement module and the feature space adaptive module according to the prototype loss function value. When the prototype loss function value is minimized, obtain the fine-tuned prototype refinement module and feature space adaptive module; S35. Input the support set and query set of each category i into the trained feature extraction module and the fine-tuned prototype refinement module to obtain the refined prototype representation of each category i; map the refined prototype representation of each category i to the spherical space to obtain the prototype representation W i ' of the fine-tuned high-dimensional feature space structure of each category i; S36. Prototype representation W of the high-dimensional feature space structure fine-tuned according to the basic prototype and each category i i ′ Construct a prototype graph structure, and input the query set of each category i and the prototype graph structure into the persistent classifier to obtain the scores of each data sample in the query set of each category i ; S37. Calculate the classification loss function value according to the score, update the prototype refinement module and the continuously evolving classifier according to the classification loss function value, and when the classification loss function value is the smallest, obtain a trained prototype refinement module, a feature space adaptation module, and a continuously evolving classifier.

9. The aircraft classification method with accurate prototypes based on few-shot class incremental learning according to claim 8, characterized in that, Constructing the prototype graph structure includes: S361. Prototype \(W\) of the fine-tuned high-dimensional feature space structure according to each category \(i\) i ' and the basic prototype \(P\) base Construct the prototype graph structure; S362. Use the attention mechanism to enhance the prototype \(W\) of the high-dimensional feature space structure fine-tuned for each category \(i\) according to the prototype diagram structure, and obtain the enhanced prototype \(W\) of each category \(i\). i ″. i ″.

10. A method for accurately classifying aircraft prototypes based on few-shot class incremental learning according to claim 9, characterized in that, The prototype W of the fine-tuned high-dimensional feature space structure for class i i ′ is enhanced to include: W i ″ = W i ′ + w(W i ′)·a i Among them, P total ={P base , P inc} is the prototype diagram structure, P inc ={W1′, W2′,..., W′ M} is the set of prototypes W i ′ of the high-dimensional feature space structure after fine-tuning for all categories i, d total is the dimension of the feature P total , a i is the attention score of the prototype W i ′, and w() is the weight matrix.

Images