Incremental small sample sar target recognition method based on angle feature separation and reservation

Abstract

The application relates to an incremental small sample SAR target recognition method based on angle feature separation and reservation. The method comprises the following steps: obtaining a trained final backbone network through basic training and incremental training, extracting features of a to-be-recognized sample and all saved sample sets according to the final backbone network, comparing the feature extraction results according to a sample angle nearest neighbor criterion, and obtaining a predicted target recognition result of the to-be-recognized sample; wherein three angle feature losses including a cosine normalization loss, an edge feature separation loss and a related feature reservation loss are introduced in the incremental training, and an angle sorting sample selection strategy is adopted to save representative samples in each incremental training process. By adopting the method, the incremental small sample recognition capability of a SAR image target can be improved under the condition that target information is small, the target inter-class similarity and intra-class difference and background interference complexity are fully considered, and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to an incremental small sample SAR target recognition method based on angle feature separation and reservation. BACKGROUND

[0002] As a long-term, fundamental and challenging problem in synthetic aperture radar (SAR) image interpretation, SAR automatic target recognition plays an indispensable role in a wide range of civilian and military applications such as military reconnaissance surveillance, border security, etc. and has been a research hotspot in recent decades. In recent years, with the rapid development of SAR imaging technology, the number of high-resolution SAR images has increased exponentially, providing more and more research opportunities for fine-grained SAR target recognition tasks. Many works including traditional and latest deep learning methods have been devoted to the research of SAR automatic target recognition problem and have made great progress. However, for the SAR automatic target recognition problem, the road is long and difficult.

[0003] In the past two years, the target recognition problem based on incremental small sample learning (FSCIL) has attracted widespread attention, which aims to solve the problem of incrementally recognizing new classes from very few labeled samples without forgetting previously learned classes. For many real-world application scenarios, the model may encounter new classes that do not exist in the initial training, and it is often very costly to collect many annotated samples for new classes, which is a very realistic but challenging problem. Especially for synthetic aperture radar automatic target recognition systems (SAR ATR) in practical applications, this problem needs to be solved. Due to the particularity of SAR images, only SAR experts can annotate them, especially in non-cooperative conditions, the cost of obtaining target annotations in SAR images is very high. Therefore, it is extremely urgent for SAR ATR systems to incrementally recognize new classes from very limited labeled samples without significantly reducing the recognition accuracy of old classes. SUMMARY

[0004] Therefore, it is necessary to provide an incremental small sample SAR target recognition method based on angle feature separation and reservation in view of the above technical problems.

[0005] An incremental small sample SAR target recognition method based on angle feature separation and reservation, the method comprising:

[0006] inputting a base class data set into a basic model comprising a backbone network and a base class classifier for processing to obtain a base class target recognition result, calculating a cosine normalization loss according to the base class target recognition result, training the basic model according to the cosine normalization loss to obtain a trained backbone network and a base class classifier;

[0007] According to the angle sorting sample selection strategy, representative samples in the base class dataset are saved to obtain a saved sample set, and the saved sample set is combined with the new class dataset to obtain an incremental dataset;

[0008] The incremental dataset is input into an incremental model including a backbone network and an incremental classifier for processing to obtain an incremental target recognition result, an angle feature loss is calculated according to the incremental target recognition result, the incremental model is trained according to the angle feature loss, and a trained backbone network and an incremental classifier are obtained; wherein the angle feature loss includes a cosine normalization loss, an edge feature separation loss, and a relevant feature preservation loss;

[0009] The incremental model is repeatedly trained until a final trained backbone network is obtained, features of a to-be-recognized sample and all saved sample sets are extracted according to the final trained backbone network, and a predicted target recognition result of the to-be-recognized sample is obtained by comparing the feature extraction results according to a sample angle nearest neighbor criterion.

[0010] In one of the embodiments, the base class dataset is input into a base model including a backbone network and a base class classifier for processing to obtain a base class target recognition result, including:

[0011] The base class dataset is input into a base model including a backbone network and a base class classifier for processing, wherein the base class dataset includes multiple pairs of base class dataset samples and corresponding labels;

[0012] Features of the base class dataset samples are extracted according to the backbone network to obtain embedded features of the base class dataset samples;

[0013] The embedded features of the base class dataset samples are classified according to the base class classifier to obtain the base class target recognition result.

[0014] In one of the embodiments, a cosine normalization loss is calculated according to the base class target recognition result, and the base model is trained according to the cosine normalization loss to obtain a trained backbone network and a base class classifier, including:

[0015] The base class target recognition result is calculated to obtain a cosine normalization loss, which is represented as

[0016]

[0017]

[0018] wherein y i represents a label corresponding to the i-th base class dataset sample x i p(c i | x) represents that the base class dataset sample x belongs to the i-th class ci a probability of the base class data set sample x, denotes the normalized embedding feature of the base class data set sample x, η(x) denotes the embedding feature of the base class data set sample x, denotes the prototype feature of the class c i corresponding to the prototype feature, denotes the j-th class c j corresponding to the prototype feature, C denotes the total number of classes, and γ is a hyperparameter.

[0019] According to the cosine normalization loss L CN The base model is trained to obtain the trained backbone network and the base class classifier.

[0020] In one of the embodiments, according to the angle ranking sample selection strategy, representative samples in the base class data set are saved to obtain a saved sample set, and the saved sample set is combined with the new class data set to obtain an incremental data set, including:

[0021] According to the angle ranking sample selection strategy, representative samples in the base class data set are saved to obtain a saved sample set, wherein the selection mode of the i-th sample e i in the saved sample set is

[0022]

[0023] wherein D 1 denotes the base class data set, k denotes the number of samples in the saved sample set, η(e i ) denotes the embedding feature of the saved sample set sample, η(x j ) denotes the embedding feature of the j-th base class data set sample x j in the base class data set, and p y denotes the prototype feature of the class y.

[0024] The saved sample set is combined with the new class data set to obtain an incremental data set.

[0025] In one of the embodiments, before the incremental data set is input into the incremental model including the backbone network and the incremental classifier for processing, it further includes:

[0026] The prototype initialization strategy is used to calculate the new class samples in the new class data set to obtain the prototypes of each class in the new class data set, and the weights of the incremental classifier are initialized according to the prototypes of each class to obtain an initialized incremental classifier.

[0027] In one of the embodiments, the incremental data set is input into the incremental model including the backbone network and the incremental classifier for processing to obtain an incremental target recognition result, including:

[0028] The incremental data set is input into an incremental model including a backbone network and an initialized incremental classifier for processing;

[0029] According to the backbone network, the incremental data set samples are subjected to feature extraction to obtain embedded features of the incremental data set samples;

[0030] According to the incremental classifier subjected to the initialization processing, the embedded features of the incremental data set samples are subjected to classification to obtain an incremental target recognition result.

[0031] In one of the embodiments, an angle feature loss is calculated according to the incremental target recognition result, the incremental model is trained according to the angle feature loss, and a trained backbone network and incremental classifier are obtained, including:

[0032] The incremental target recognition result is calculated to obtain the angle feature loss, which is represented as

[0033] L = L CN + L MFS + L RFP

[0034] wherein, L CN represents a cosine normalization loss, L MFS represents an edge feature separation loss, and L RFP represents a relevant feature retention loss.

[0035] The trained backbone network and the incremental classifier are obtained by training the backbone network according to the angle feature loss L and iteratively updating the weight of the incremental classifier.

[0036] In one of the embodiments, the edge feature separation loss is represented as:

[0037]

[0038] where o∈C (1) ∪C (2) ...,∪C (t-1) ,n∈C (t) ,i≠j

[0039] wherein, e n,i represents the embedded features of the selected samples of the new class, represents the embedded features of the samples of the same class as e n,i , and e represents the embedded features of the samples of different classes from e n,i , m is a hyperparameter for controlling the separation boundary, o represents the class to which e belongs, and C (t) represents the sample classes contained in the tthincremental stage.

[0040] In one of the embodiments, the relevant feature preservation loss is represented as:

[0041] L RFP = L RFL + L IFL

[0042] RFL = ||A t-1 (θ i,j ) - A t (θ i,j ) ||

[0043]

[0044]

[0045]

[0046] wherein L RFL represents the angle correlation loss of the embedding features of the samples of different categories, L IFL represents the angle feature loss of the embedding features of the same sample in different incremental stages, RFL represents the relevant feature loss, IFL represents the independent feature loss, A t-1 (θ i,j ) and A t (θ i,j ) represent the angle correlation losses of the samples in the t-1th and tth incremental stages respectively, θ i,j represents the angle feature value between the ith and jth samples, E (t-1) denotes the sample set saved in the tth incremental stage, η t-1 (e i ) represents the embedding feature of the sample e i in the sample set saved in the t-1th incremental stage, and η t (e i ) represents the embedding feature of the sample e i in the sample set saved in the tth incremental stage.

[0047] In one of the embodiments, the final backbone network is used to extract features of the sample to be identified and all saved sample sets, the feature extraction results are compared according to the sample angle nearest neighbor criterion, and a predicted target identification result of the sample to be identified is obtained, including:

[0048] The final backbone network is used to extract features of the sample to be identified and all saved sample sets, and normalized embedding features of the sample to be identified and embedding features of all saved sample sets are obtained respectively;

[0049] The normalized embedding feature of the sample to be identified is compared with the embedding features of all the saved sample set according to the sample angle nearest neighbor criterion, to obtain the class label y of the predicted target recognition result of the sample to be identified * , is expressed as

[0050]

[0051] wherein, The normalized embedding feature of the sample to be identified is compared with the embedding features of all the saved sample set according to the sample angle nearest neighbor criterion, to obtain the class label y of the predicted target recognition result of the sample to be identified The normalized prototype of the class y is represented by P y The sample set belonging to the class y is represented by P y p is a certain sample selected from the sample set P

[0052] The above incremental small sample SAR target recognition method based on angle feature separation and reservation can learn the generalization features and specific features of different classes by introducing a cosine normalization loss in the training of the base model, so as to obtain a trained backbone network and a base class classifier, which can improve the distinguishing ability of the model for different classes and strengthen the robustness to abnormal feature values and background interference; in the training stage of the incremental model, the model is trained by introducing an angle feature loss including a cosine normalization loss, an edge feature separation loss and a related feature reservation loss, so as to relieve the problems of catastrophic forgetting of old knowledge and inaccurate representation of new and old knowledge, and meanwhile, representative samples are saved by using an angle sorting sample selection strategy in each incremental model training process, so as to ensure the fast and accurate convergence ability of the model; in addition, in the test stage, the target recognition result of the sample to be identified is predicted by introducing a sample angle nearest neighbor criterion, which has generalization representation and insensitivity to abnormal values, so as to ensure the accuracy of the target recognition result. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 Fig. 1 is a flowchart of an embodiment of the incremental small sample SAR target recognition method based on angle feature separation and reservation;

[0054] Figure 2 Fig. 1 is a flowchart of an embodiment of the incremental small sample SAR target recognition method based on angle feature separation and reservation; DETAILED DESCRIPTION

[0055] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the 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.

[0056] In one embodiment, as shown in Figure 1 An angle feature separation and retention based incremental small sample SAR target recognition method is provided, comprising the following steps:

[0057] Step 102, input the base class data set into the base model including the backbone network and the base class classifier for processing to obtain a base class target recognition result, calculate a cosine normalization loss according to the base class target recognition result, train the base model according to the cosine normalization loss to obtain a trained backbone network and base class classifier.

[0058] It can be understood that by using the cosine normalization loss (CN) loss on the base class data set to optimize the convolutional neural network (CNN) based backbone network and the parameterized base class classifier, a trained backbone network is obtained as a well-spanned feature extractor with basic class recognition ability, and the feature extractor is retained to extract the generalization features of the classes appearing in the subsequent sessions.

[0059] Step 104, save the representative samples in the base class data set according to the angle ranking sample selection strategy to obtain a saved sample set, combine the saved sample set with the new class data set to obtain an incremental data set.

[0060] It can be understood that the angle ranking sample selection (A-RES) strategy is a strategy for preferentially selecting samples, which considers the cosine distance between the prototype and the sample instead of the L2 distance, so as to retain the specificity features of the old knowledge learned by the previous model, that is, to save the representative samples in the base class data set to obtain a saved sample set; the new class data set is a small sample containing some new classes newly input in the incremental training, including N classes and K data for each class, and the values of N and K are very small.

[0061] Step 106, input the incremental data set into the incremental model including the backbone network and the incremental classifier for processing to obtain an incremental target recognition result, calculate an angle feature loss according to the incremental target recognition result, train the incremental model according to the angle feature loss to obtain a trained backbone network and incremental classifier; wherein the angle feature loss includes a cosine normalization loss, an edge feature separation loss and a related feature retention loss.

[0062] It can be understood that in the incremental learning stage, in order to balance the stability and plasticity of the model, according to the features of the targets in the SAR image, three angle feature related losses of cosine normalization (CN), edge feature separation (MFS) and related feature retention (RFP) are designed to train the incremental model; at the same time, by designing the angle ranking sample selection (A-RES) and prototype initialization (PI) strategies, the rapid and accurate convergence of the model is ensured.

[0063] Step 108, repeat the training of the incremental model until a well-trained final backbone network is obtained, and then perform feature extraction on the to-be-identified sample and all saved sample sets according to the final backbone network, and compare the feature extraction results according to the sample angle nearest neighbor criterion to obtain the predicted target recognition result of the to-be-identified sample.

[0064] It can be understood that the above process is repeated until the i (i > 1) session, and finally a well-trained final backbone network η is obtained. i , which has good representation ability for all visible classes and all saved sample sets; the sample angle nearest neighbor (A-NME) criterion is used to assign a label to the to-be-identified sample for testing, and the advantage of this criterion is its generalization representation and insensitivity to outliers. Specifically, the cosine similarity between the embedding features of the to-be-identified sample and the prototypes of all visible classes is calculated to determine the predicted target recognition result of the to-be-identified sample.

[0065] The above incremental small sample SAR target recognition method based on angle feature separation and reservation introduces cosine normalization loss in the training of the base model to learn generalization features and specific features of different classes, and obtains a well-trained backbone network and a base class classifier, which can improve the model's ability to distinguish different classes and enhance its robustness to abnormal feature values and background interference. In the training phase of the incremental model, the model is trained by introducing angle feature loss including cosine normalization loss, edge feature separation loss and related feature reservation loss, thereby alleviating the problem of catastrophic forgetting of old knowledge and inaccurate representation of new and old knowledge. At the same time, representative samples are saved by using angle sorting sample selection strategy in each incremental model training process, thereby ensuring the model's fast and accurate convergence ability. In addition, in the testing phase, the sample angle nearest neighbor criterion is introduced to predict the target recognition result of the to-be-identified sample. Since the sample angle nearest neighbor criterion has generalization representation and insensitivity to outliers, the accuracy of the target recognition result is guaranteed.

[0066] In one embodiment, the base class data set is input into the base model including the backbone network and the base class classifier for processing to obtain the base class target recognition result, including:

[0067] The base class data set is input into the base model including the convolutional neural-based backbone network and the base class classifier for processing, wherein the base class data set is a series of labeled data sets including multiple pairs of base class data set samples and corresponding labels, providing a large number of base class training samples for the model training process.

[0068] According to the backbone network, the feature extraction is performed on the base class dataset sample to obtain embedded features of the base class dataset sample.

[0069] According to the base class classifier, the embedded features of the base class dataset sample are classified to obtain a base class target recognition result.

[0070] In one of the embodiments, a cosine normalization loss is calculated according to the base class target recognition result, and the basic model is trained according to the cosine normalization loss to obtain a trained backbone network and base class classifier, including:

[0071] The base class target recognition result is calculated to obtain a cosine normalization loss, which is represented as

[0072]

[0073]

[0074] wherein y i represents a label corresponding to the i-th base class dataset sample x i , p(c i |x) represents a probability that the base class dataset sample x belongs to the i-th class c i , and is calculated by a Softmax function of a cosine correlation between γ and and , γ is a hyperparameter, which restricts the amplitude of sample similarity, represents a normalized embedded feature of the base class dataset sample x, and η(x) represents an embedded feature of the base class dataset sample x, represents a prototype feature corresponding to the class c i , represents a prototype feature corresponding to the j-th class c j , and C represents a total number of classes.

[0075] According to the cosine normalization loss L CN , the basic model is trained to obtain a trained backbone network η b and a base class classifier φ b .

[0076] It can be understood that according to the cosine normalization loss, the generalization features and specific features of different classes can be learned, and by maximizing the cosine similarity between the sample embedded features and the classifier inter-class weights, the distinguishing ability of the model for different classes can be improved, and the influence of abnormal target features and background interference can be reduced; and after the model is trained by the cosine normalization loss, the model can distinguish old classes and new classes.

[0077] In one of the embodiments, according to the angle ranking sample selection strategy, representative samples in the base class dataset are saved to obtain a saved sample set, the saved sample set is combined with the new class dataset to obtain an incremental dataset, including:

[0078] According to the angle ranking sample selection strategy, representative samples in the base class dataset are saved to obtain a saved sample set E1, wherein the i-th saved sample set sample e i in the saved sample set is selected in the following manner

[0079]

[0080] wherein D 1 represents the base class dataset, k represents the number of saved sample set samples, η(e i ) represents the embedding feature of the saved sample set sample, η(x j ) represents the embedding feature of the j-th base class dataset sample x 1 in the base class dataset D j , and p y represents the prototype feature of the class y.

[0081] The saved sample set E1 is combined with the new class dataset D 2 to obtain an incremental dataset

[0082] It can be understood that, compared with randomly selecting samples in each incremental session, the angle ranking sample selection (A-REM) strategy can retain the most representative features in the old class, and by saving representative samples, the original topological structure of the old class features learned by the previous model can be accurately retained.

[0083] In one of the embodiments, before the incremental dataset is input into the incremental model including the backbone network and the incremental classifier for processing, the method further includes:

[0084] The prototype initialization strategy is used to calculate the new class dataset to obtain the mean value of each class embedding in the new class dataset, and the weight of the incremental classifier is initialized according to the mean value of each class embedding to obtain an initialized incremental classifier.

[0085] It can be understood that, before the incremental learning, the prototype initialization strategy (PI) is used to initialize the weight of the incremental classifier with the mean value of the specific class embedding, wherein the mean value of each class embedding can represent the common features of each class, and initializing the weight of the incremental classifier with the mean value of each class embedding can reduce the difficulty of learning new classes in the subsequent incremental model training process.

[0086] In one of the embodiments, the incremental data set is input into an incremental model including a backbone network and an incremental classifier for processing to obtain an incremental target recognition result, including:

[0087] The incremental data set is input into an incremental model including a backbone network and an initialized incremental classifier for processing;

[0088] The incremental data set samples are feature-extracted according to the backbone network to obtain embedded features of the incremental data set samples;

[0089] The embedded features of the incremental data set samples are classified according to the initialized incremental classifier to obtain the incremental target recognition result.

[0090] In one of the embodiments, an angle feature loss is calculated according to the incremental target recognition result, and the incremental model is trained according to the angle feature loss to obtain a trained backbone network and an incremental classifier, including:

[0091] The incremental target recognition result is calculated to obtain an angle feature loss, which is represented as

[0092] L = L CN + L MFS + L RFP

[0093] L CN represents a cosine normalization loss, L MFS represents an edge feature separation loss, and L RFP represents a related feature retention loss.

[0094] The trained backbone network η 2 and the incremental classifier φ b + φ n are obtained by training the backbone network and iteratively updating the weight of the incremental classifier according to the angle feature loss L, respectively.

[0095] It can be understood that the edge feature separation loss aims to build a clear feature boundary between old and new classes through marginal loss ordering, and considering that new and old targets may have common components, structures and appearances, it is urgent to aggregate the feature space spanned by a small number of new samples with the same label and expand the difference between classes; the related feature retention loss is designed to alleviate the catastrophic forgetting of old knowledge by the model, wherein the mutual feature relationship and individual feature relationship spanned by the previous model and the existing model are considered to fully retain the relationship, absolute position and structure topology of the learned knowledge.

[0096] In one of the embodiments, the edge feature separation loss is represented as:

[0097]

[0098] where o∈C (1) ∪C (2) ...,∪C (t-1) ,n∈C (t) i≠j

[0099] Indicates with e n,i Embedding features of samples of the same category For e n,i Embedding features of samples from different categories, where m is a hyperparameter controlling the separation boundary, and o represents... The category to which it belongs, C (t) This represents the sample categories included in the t-th incremental stage. Specifically, it represents the new sample e that saves the sample set. i Using the anchor points, the top k positive sample pairs arranged in ascending order of angular correlation and the top k negative sample pairs arranged in descending order of angular correlation are selected to calculate the edge feature separation loss, so that the model mainly focuses on samples located at the class boundary and avoids overfitting.

[0100] It is understandable that edge feature separation loss takes into account the ambiguity of features of the same category and the extreme similarity of features of different categories. In the FSCIL environment, maintaining a clear separation between learned categories is crucial for the accurate identification of targets in SAR images. Specifically, edge feature separation loss maintains a clear boundary by fully exploring the angular feature relationship between new and old categories, expanding the difference between angular features rather than amplitude features of different categories, making the generalized features of the same category more compact, while the features of different categories are more discriminative, thus ensuring the generalization of the learning model to a certain extent.

[0101] In one embodiment, the relevant feature preservation loss is expressed as:

[0102] L RFP =L RFL +L IFL

[0103] RFL = ||A t-1 (θ i,j )-A t (θ i,j )||

[0104]

[0105]

[0106]

[0107] Among them, L RFLthe angle correlation loss of different category sample embedding features, L IFL the angle feature loss of embedding features of the same sample in different incremental stages, RFL represents the relevant feature loss, and IFL represents the independent feature loss, A t-1 (θ i,j ) and A t (θ i,j ) respectively represent the angle correlation loss of the sample in the t-1th and tth incremental stages, θ i,j represents the angle feature value between the ith and jth samples, E (t-1) represents the sample set saved in the t-1th incremental stage, η t-1 (e i ) represents the embedding feature of the sample e i in the sample set saved in the t-1th incremental stage, η t (e i ) represents the embedding feature of the sample e i in the sample set saved in the tth incremental stage.

[0108] It can be understood that the relevant feature preservation loss enhances the pair-wise and mutual angle features, which is crucial for obtaining the inter-class and intra-class multi-dimensional relationships of the target in the SAR image in the incremental learning stage, and alleviates the catastrophic forgetting of the old knowledge of the model by preserving the semantic topology.

[0109] In one of the embodiments, the feature extraction is performed on the to-be-identified sample and all saved samples according to the final backbone network, and the prediction target recognition result of the to-be-identified sample is obtained by comparing the feature extraction results according to the sample angle nearest neighbor criterion, including

[0110] The feature extraction is performed on the to-be-identified sample x' and all saved sample sets E1, E2,..., E i according to the final backbone network η i-1 , to obtain the normalized embedding feature of the to-be-identified sample and the embedding features of all saved sample sets, respectively;

[0111] The normalized embedding feature of the to-be-identified SAR image target sample and the embedding features of all saved sample sets are compared according to the sample angle nearest neighbor criterion, to obtain the class label y * of the prediction target recognition result of the to-be-identified sample, which is represented as

[0112]

[0113] wherein, represents the normalized embedding feature of the to-be-identified sample x', η(p) represents the embedding feature of the sample belonging to the category y in the saved sample set, represents the normalized prototype of the category y, P yrepresent a sample set belonging to the category y, p is a sample selected from the sample set P y

[0114] It can be understood that the sample angle nearest neighbor criterion adopts a classification strategy based on angle features, which can better retain the fault tolerance to abnormal values in the SAR image and the ability to distinguish different categories of targets, while suppressing severe interference in the interference stage.

[0115] In summary, according to the incremental small sample SAR target recognition method based on angle feature separation and retention proposed in the present application, the overall framework established is as follows Figure 2 As described above, the method comprises three stages of basic training, incremental training and session testing:

[0116] The basic training stage is used for inputting the base class data set into the basic model comprising the backbone network and the base classifier for processing to obtain the base class target recognition result, calculating the cosine normalization loss according to the base class target recognition result, training the basic model according to the cosine normalization loss, and obtaining the trained backbone network and the base classifier.

[0117] The incremental training stage is used for saving the representative samples in the base class data set according to the angle sorting sample selection strategy to obtain the saved sample set, combining the saved sample set with the new category data set to obtain the incremental data set; inputting the incremental data set into the incremental model comprising the backbone network and the incremental classifier for processing to obtain the incremental target recognition result, calculating the angle feature loss according to the incremental target recognition result, training the incremental model according to the angle feature loss, and obtaining the trained backbone network and the incremental classifier; wherein the angle feature loss comprises the cosine normalization loss, the edge feature separation loss and the correlation feature retention loss.

[0118] The session testing stage is used for repeatedly training the incremental model until the final trained backbone network is obtained, extracting features from the to-be-recognized sample and all saved sample sets according to the final backbone network, comparing the feature extraction results according to the sample angle nearest neighbor criterion, and obtaining the predicted target recognition result of the to-be-recognized sample.

[0119] It should be understood that, although Figure 1 the steps in the flowchart of the present application are displayed in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, Figure 1 ​At least one of the steps in the above embodiments can include a plurality of sub-steps or a plurality of stages, which are not necessarily executed at the same time, but can be executed at different times, and the order of execution of the sub-steps or stages is not necessarily sequential, but can be executed in rotation or alternation with other steps or at least a part of the sub-steps or stages of other steps.

[0120] The technical features of the above embodiments can be combined in any manner. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described, but it is understood that any combination of the technical features is within the scope of the present disclosure as long as there is no contradiction.

[0121] The above embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the patent scope of the present application. It should be noted that, for those skilled in the art, several modifications and improvements can be made without departing from the concept of the present application, and these are within the protection scope of the present application. Therefore, the patent protection scope of the present application should be subject to the appended claims.

Claims

1. An incremental small sample SAR target recognition method based on angle feature separation and reservation, characterized in that, The method comprises: inputting the base class data set into a base model comprising a backbone network and a base classifier for processing to obtain a base class target recognition result, calculating a cosine normalization loss based on the base class target recognition result, training the base model based on the cosine normalization loss to obtain a trained backbone network and base classifier; saving representative samples in the base class data set according to an angle sorting sample selection strategy to obtain a saved sample set, combining the saved sample set with a new class data set to obtain an incremental data set; inputting the incremental data set into an incremental model comprising a backbone network and an incremental classifier for processing to obtain an incremental target recognition result, calculating an angle feature loss based on the incremental target recognition result, training the incremental model based on the angle feature loss to obtain a trained backbone network and incremental classifier; wherein the angle feature loss comprises a cosine normalization loss, an edge feature separation loss and a relevant feature retention loss; repeating the training of the incremental model until a final trained backbone network is obtained, extracting features of a to-be-identified sample and all the saved sample set based on the final trained backbone network, comparing the feature extraction results based on a sample angle nearest neighbor criterion to obtain a predicted target recognition result of the to-be-identified sample; wherein saving representative samples in the base class data set according to an angle sorting sample selection strategy to obtain a saved sample set, combining the saved sample set with a new class data set to obtain an incremental data set comprises: Representative samples from the base class dataset are saved according to the angle-sorted sample selection strategy to obtain a saved sample set, wherein the first sample in the saved sample set is... i Sample The selection method is wherein, represents the base class dataset, k represents the number of samples in the base class dataset, represents the embedding feature of a sample in the base class dataset, represents the embedding feature of the jth sample in the base class dataset, represents the embedding feature of the jth sample in the base class dataset, represents the prototype feature of class y; combining the saved sample set with the new class data set to obtain the incremental data set.

2. The method of claim 1, wherein, inputting the base class data set into a base model comprising a backbone network and a base classifier for processing to obtain a base class target recognition result comprises: inputting the base class data set into a base model comprising a backbone network and a base classifier for processing, wherein the base class data set comprises multiple pairs of base class data set samples and corresponding labels; extracting features of the base class data set samples based on the backbone network to obtain embedded features of the base class data set samples; classifying the embedded features of the base class data set samples based on the base classifier to obtain a base class target recognition result.

3. The method of claim 1, wherein, calculating a cosine normalization loss based on the base class target recognition result, training the base model based on the cosine normalization loss to obtain a trained backbone network and base classifier comprises: calculating the base class target recognition result to obtain the cosine normalization loss, which is represented as in, Indicates the first i Sample data from each base class dataset The corresponding tags Represents the base class dataset samples Belongs to the i Categories The probability, Represents the base class dataset samples Normalized embedding features, Represents the base class dataset samples Embedding features, Indicate category Corresponding prototype features Indicates the first j Categories Corresponding prototype features C Indicates the total number of categories. For hyperparameters; According to the cosine normalization loss The base model is trained to obtain a trained backbone network and a base class classifier.

4. The method according to any one of claims 1 to 3, characterized in that, before inputting the incremental data set into an incremental model comprising a backbone network and an incremental classifier for processing, further comprising: calculating new class samples in the new class data set through a prototype initialization strategy to obtain prototypes of each class in the new class data set, initializing weights of the incremental classifier based on the prototypes of each class to obtain an initialized incremental classifier.

5. The method of claim 1, wherein, inputting the incremental data set into an incremental model comprising a backbone network and an incremental classifier for processing to obtain an incremental target recognition result comprises: inputting the incremental data set into an incremental model including a backbone network and an initialized incremental classifier for processing; performing feature extraction on the incremental data set samples according to the backbone network to obtain embedding features of the incremental data set samples; performing classification on the embedding features of the incremental data set samples according to the initialized incremental classifier to obtain an incremental target recognition result.

6. The method of claim 1, wherein, calculating an angle feature loss according to the incremental target recognition result, training the incremental model according to the angle feature loss to obtain a trained backbone network and an incremental classifier, including: calculating the angle feature loss according to the incremental target recognition result, which is represented as wherein, denotes the cosine normalization loss, denotes the edge feature separation loss, denotes the relevant feature preservation loss; According to the angle feature loss Respectively training the backbone network and iteratively updating the weight of the incremental classifier, to obtain the trained backbone network and the incremental classifier.

7. The method of claim 6, wherein, the edge feature separation loss is represented as: wherein, represent embedding features of selected samples of a new class, represent embedding features of samples of the same class as represent embedding features of samples of the same class as represent embedding features of samples of the same class as represent embedding features of samples of a different class than are hyperparameters that control the separation boundary, represent the class to which belongs, represent the class of samples contained in the t th incremental stage.

8. The method of claim 6, wherein, the correlation feature retention loss is represented as: wherein, represents the angle correlation loss of different categories of sample embedding features, represents the angle feature loss of embedding features of the same sample at different incremental stages, RFL represents the relevant feature loss, and IFL represents the independent feature loss, and respectively represent the angle correlation loss of the sample at the first and the second incremental stage, represents the angle feature value between the first i and the second j sample, represents the sample set saved at the first incremental stage, represents the embedding feature of the sample in the sample set saved at the first incremental stage, represents the embedding feature of the sample in the sample set saved at the second incremental stage.

9. The method of claim 1, wherein, performing feature extraction on the to-be-recognized sample and all the saved sample sets according to the final backbone network, comparing the feature extraction results according to a sample angle nearest neighbor criterion to obtain a predicted target recognition result of the to-be-recognized sample, including: performing feature extraction on the to-be-recognized sample and all the saved sample sets according to the final backbone network to obtain normalized embedding features of the to-be-recognized sample and embedding features of all the saved sample sets, respectively; comparing the normalized embedding feature of the sample to be identified and the embedding features of all the saved sample set according to a sample angle nearest neighbor criterion to obtain a class label of a predicted target recognition result of the sample to be identified is represented as wherein, denotes a normalized embedding feature of a sample to be identified, denotes embedding features of samples belonging to class in the set of saved samples, denotes a normalized prototype of class denotes a set of samples belonging to class p is a certain sample selected from the set of samples .​​​

Images