Radar target recognition method, device and equipment based on scale invariant feature transform

Abstract

The application relates to a radar target recognition method, device and equipment based on a scale-invariant feature transformation. The method comprises the following steps: constructing a radar target recognition model; the radar target recognition model comprises a scale-invariant feature transformation feature extraction module, an encoder, a projection head and a classifier; after feature extraction on the augmented sample pair by the scale-invariant feature transformation feature extraction module and the encoder, the first mixed features are spliced, the first loss function is set after the first mixed features are mapped to the embedding space by the projection head; the encoder is trained by using the first loss function, and the trained encoder is obtained; the input radar image is subjected to feature extraction and splicing according to the trained encoder and the scale-invariant feature transformation feature extraction module, the second mixed features are input into the classifier, the second loss function is designed, the classifier is trained, and radar target recognition is performed according to the trained radar target recognition model. The method can improve the radar target recognition accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar signal processing, in particular to a radar target recognition method, device and equipment based on scale-invariant feature transformation. BACKGROUND

[0002] Radar auto target recognition (RATR) plays an important role in the field of aviation safety, and how to quickly, accurately and stably identify space targets is the main task of RATR. In the RATR task, it is of great significance to identify the synthetic aperture radar (SAR) target which presents a long-tailed distribution. In the SAR image, some targets appear very low in frequency (tail targets) due to the difficulty of acquisition, while some common targets appear frequently (head targets), resulting in a serious imbalance between target categories, which is the long-tailed distribution. The target with low frequency of occurrence is likely to be a target with important significance. The data of the head target is relatively easy to obtain, and the corresponding sample data of the tail data is also relatively scarce, which brings challenges to the training of the model.

[0003] The traditional machine learning method cannot fully utilize a small amount of samples for effective long-tail target recognition, resulting in a low recognition accuracy of the tail targets. Therefore, how to simultaneously achieve a good recognition accuracy and recognition robustness of the head targets and the tail targets is a problem to be solved in the RATR field at present. The traditional long-tail recognition method improves the recognition accuracy of the tail data by downsampling the head targets, resampling the tail targets and weighting the loss, but the recognition rate obtained by the traditional method is low. In recent years, the performance of the existing SAR target recognition method based on deep learning is highly dependent on the quality and quantity of the training data set. However, the SAR target classification scene usually neither has enough training samples nor has a balanced distribution of categories, the robustness of the deep learning method is low, and the scattering characteristics of the SAR image are not considered. Therefore, the existing method cannot well implement the long-tail SAR target recognition. SUMMARY

[0004] Therefore, it is necessary to provide a radar target recognition method, device and equipment based on scale-invariant feature transformation which can improve the radar target recognition accuracy in view of the above technical problems.

[0005] A radar target recognition method based on scale-invariant feature transformation, the method comprising:

[0006] obtaining long-tail SAR target data; constructing a radar target recognition model; the radar target recognition model comprises a scale-invariant feature transformation feature extraction module, an encoder, a projection head and a classifier;

[0007] Data augmentation is performed on the long-tail SAR target data to obtain an augmented sample pair; a scale-invariant feature transformation feature extraction module and an encoder are used to extract features from the augmented sample pair to obtain scattering features and texture features of the radar image; the scattering features and the texture features are spliced to obtain first mixed features;

[0008] The first mixed features are mapped to an embedding space by using a projection head, and a first loss function is set according to all embeddings in a batch; the encoder is trained by using the first loss function to obtain a trained encoder;

[0009] The input radar image is subjected to feature extraction and splicing by using the trained encoder and the scale-invariant feature transformation feature extraction module to obtain second mixed features; the second mixed features are input into a classifier to obtain a prediction result vector; a second loss function is designed according to the prediction result vector and a true label of the radar image;

[0010] The classifier is trained according to the second loss function to obtain a trained classifier; the trained encoder and the classifier are substituted into a radar target recognition model to obtain a trained radar target recognition model; and radar target recognition is performed according to the trained radar target recognition model.

[0011] In one of the embodiments, the scale-invariant feature transformation feature extraction module and the encoder are used to extract features from the augmented sample pair to obtain scattering features and texture features of the radar image, including:

[0012] The scale-invariant feature transformation feature extraction module is used to extract features from the augmented sample pair, to extract K scattering point information of the augmented sample, to calculate scattering information of the kth scattering point, and to contain coordinates of a scattering intensity maximum point a maximum scattering intensity direction and a module length corresponding to the direction

[0013] Definition P k represents the kth scattering point feature, P = [p1, p2, …, p K ] represents scattering features extracted from the augmented sample i, and T represents a transpose operation.

[0014] In one of the embodiments, the coordinates of the scattering intensity maximum point are

[0015]

[0016] wherein S represents a difference Gaussian space, C k represents an extreme point with a maximum scattering intensity in the Gaussian difference space.

[0017] In one embodiment, the maximum scattering intensity direction is

[0018]

[0019] wherein, represents a scale space, N(x k ,y k ,σ k ) represents a Gaussian distribution with a scale factor σ k , x k , y k represents the coordinates of the point, L1=L(x k ,y k +1,σ k ), L2=L(x k ,y k -1,σ k ), L3=L(x k +1,y k ,σ k ), L4=L(x k -1,y k ,σ k ).

[0020] In one embodiment, the length of the direction corresponding to the modulus is

[0021]

[0022] wherein, wherein, represents a scale space, N(x k ,y k ,σ k ) represents a Gaussian distribution with a scale factor σ k , x k , y k represents the coordinates of the scattering point, L1=L(x k ,y k +1,σ k ), L2=L(x k ,y k -1,σ k ), L3=L(x k +1,y k ,σ k ), L4=L(x k -1,y k ,σ k ).

[0023] In one embodiment, the texture feature of the radar image is obtained by performing feature extraction on the augmented sample pair according to the encoder

[0024]

[0025] where f E (·) denotes the encoder feature extraction operation, denotes the augmented sample.

[0026] In one embodiment, the first loss function is set according to all embeddings in a batch, including:

[0027] The first loss function is set according to all embeddings in a batch

[0028]

[0029] where z a denotes the augmented sample The embedding of z a is also the anchor point in this batch, P(i) denotes the set of all positive samples of z h in this batch, |P(i)| denotes the number of P(i), A(i) denotes the set of all samples other than the anchor point in this batch, and τ is a temperature hyperparameter.

[0030] In one embodiment, the second loss function is designed according to the prediction result vector and the true label of the radar image, including:

[0031] The second loss function is designed according to the prediction result vector and the true label of the radar image

[0032]

[0033] where q = [q1, q2, …, q C ] denotes the true label vector containing C elements, and v denotes the prediction result vector.

[0034] A radar target recognition device based on scale-invariant feature transformation, the device comprising:

[0035] A model construction module is configured to obtain long-tail SAR target data, and construct a radar target recognition model, wherein the radar target recognition model comprises a scale-invariant feature transformation feature extraction module, an encoder, a projection head, and a classifier.

[0036] The feature extraction module is configured to perform data augmentation on the long-tail SAR target data to obtain an augmented sample pair, and perform feature extraction on the augmented sample pair according to the scale-invariant feature transformation feature extraction module and the encoder to obtain scattering features and texture features of the radar image.

[0037] an encoder training module configured to map the first mixed feature to an embedding space by using the projection head, set a first loss function according to all embeddings in a batch, and train the encoder by using the first loss function to obtain a trained encoder;

[0038] a classifier training module configured to perform feature extraction and splicing on an input radar image according to the trained encoder and the scale-invariant feature transformation feature extraction module to obtain a second mixed feature, input the second mixed feature into a classifier to obtain a prediction result vector, design a second loss function according to the prediction result vector and a real label of the radar image, and train the classifier according to the second loss function to obtain a trained classifier;

[0039] a radar target recognition module configured to substitute the trained encoder and the classifier into a radar target recognition model to obtain a trained radar target recognition model, and perform radar target recognition according to the trained radar target recognition model.

[0040] A computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the following steps when executing the computer program:

[0041] obtain long-tail SAR target data, construct a radar target recognition model, and the radar target recognition model includes a scale-invariant feature transformation feature extraction module, an encoder, a projection head and a classifier;

[0042] perform data augmentation on the long-tail SAR target data to obtain an augmented sample pair, perform feature extraction on the augmented sample pair according to the scale-invariant feature transformation feature extraction module and the encoder to obtain scattering features and texture features of a radar image, and splice the scattering features and the texture features to obtain a first mixed feature;

[0043] map the first mixed feature to an embedding space by using the projection head, set a first loss function according to all embeddings in a batch, and train the encoder by using the first loss function to obtain a trained encoder;

[0044] perform feature extraction and splicing on an input radar image according to the trained encoder and the scale-invariant feature transformation feature extraction module to obtain a second mixed feature, input the second mixed feature into a classifier to obtain a prediction result vector, and design a second loss function according to the prediction result vector and a real label of the radar image;

[0045] train the classifier according to the second loss function to obtain a trained classifier, substitute the trained encoder and the classifier into a radar target recognition model to obtain a trained radar target recognition model, and perform radar target recognition according to the trained radar target recognition model.

[0046] A computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the following steps:

[0047] The radar target recognition method, device and equipment based on the scale invariant feature transformation, the application provides a double-channel decoupling learning representation method based on a scale invariant feature transformation feature extraction module, the SAR image features provided by the scale invariant feature transformation feature extraction module are used in the feature extraction learning stage and the classifier fine-tuning stage, and the SAR image features have good separability, are used for long-tail SAR image classification, the SAR physical scattering properties are fully utilized to improve the feature extraction capability of the SAR image under unbalanced sample distribution, a first loss function is designed in the feature extraction learning stage according to the extracted features and the fusion features, the augmented samples corresponding to the anchor points and the samples with the same labels as the anchor points are all used as positive samples, and other samples are used as negative samples for supervised contrast learning, so that the learning features have better balance on all categories, and a second loss function is set in the classifier fine-tuning stage by using the fusion features, the difference between the predicted result vector and the real label of the radar image is directly measured to adjust the classifier output, and the loss function designed according to the fusion features with better representation capability can improve the accuracy of model training, and then improve the accuracy of subsequent radar target recognition. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 A flowchart of a radar target recognition method based on scale invariant feature transformation in an embodiment is shown in the figure;

[0049] Figure 2 A framework diagram of a radar target recognition model in an embodiment is shown in the figure;

[0050] Figure 3 A framework diagram of a scale invariant feature transformation feature extraction module in an embodiment is shown in the figure;

[0051] Figure 4 A diagram of the detailed network structure of an encoder and a projection head in another embodiment is shown in the figure;

[0052] Figure 5 A structural block diagram of a radar target recognition device based on scale invariant feature transformation in an embodiment is shown in the figure;

[0053] Figure 6 An internal structure diagram of a computer device in an embodiment is shown in the figure. DETAILED DESCRIPTION

[0054] 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 examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application.

[0055] In one embodiment, as shown in Figure 1 a scale-invariant feature transform-based radar target recognition method is provided, comprising the following steps:

[0056] Step 102, obtaining long-tail SAR target data; constructing a radar target recognition model; the radar target recognition model comprises a scale-invariant feature transform feature extraction module, an encoder, a projection head and a classifier.

[0057] The present application mainly studies the SAR image classification task with an unbalanced distribution of target data sets, and constructs a radar target recognition model as shown in Figure 2 The training of the radar target recognition model includes a feature extraction learning stage and a classifier fine-tuning stage. The scale-invariant feature transform feature extraction module (SIFT feature extraction module) provides SAR image features with good separability, which is used for long-tail SAR image classification. The SAR physical scattering properties are fully utilized to improve the feature extraction capability of SAR images under unbalanced sample distribution. The loss function designed based on the more representative fused features can improve the accuracy of model training, and thus improve the accuracy of subsequent radar target recognition.

[0058] Step 104, data augmentation is performed on the long-tail SAR target data to obtain augmented sample pairs; the augmented sample pairs are subjected to feature extraction by the scale-invariant feature transform feature extraction module and the encoder to obtain scattering features and texture features of the radar image; and the scattering features and the texture features are spliced to obtain first mixed features.

[0059] In the feature extraction learning stage of supervised contrastive learning, supervised contrastive learning can simultaneously achieve good representation effect for head targets and tail targets. In this stage, a feature fusion model based on a supervised contrastive learning module and a SIFT feature extraction module is proposed to train the encoder and the projection head for long-tail SAR target data. In this stage, for a batch of input data x, the data is first subjected to two data augmentations to obtain a pair of sample pairs and That is:

[0060]

[0061] wherein Aug(·) represents a data augmentation operation, and the augmentation operation includes grayscale transformation, random cropping and horizontal flipping, and the purpose is to obtain and with similar but not identical information, and improve the robustness of the model. The subsequent operations will be performed simultaneously and .

[0062] The augmented data and will go through a feature extraction operation to get high-dimensional fusion features, including texture features obtained after the encoder and scattering features obtained after the SIFT algorithm. First, (and ) will go through an encoder to get texture features r a,e (and r b,e ), that is:

[0063]

[0064] where f E (·) represents the encoder feature extraction operation, and the extracted r a,e ∈R w . At the same time, (and ) will also go through a SIFT feature extraction module to obtain the scattering features of the SAR image, that is:

[0065]

[0066] where SIFT(·) represents the SIFT feature extraction operation, and the extracted r a,s ∈R u has rich scattering information of the SAR image, including the coordinates of the scattering intensity maximum point, the maximum scattering intensity direction, and the modulus corresponding to the direction. The structure of the scale-invariant feature transformation feature extraction module (SIFT feature extraction module) is shown in Figure 3 , which realizes scattering feature extraction by calculating the coordinates of the scattering intensity maximum point, the maximum scattering intensity direction, and the modulus corresponding to the direction of the kth scattering point of the augmented sample, and uses the SAR physical scattering properties to improve the feature extraction ability of the SAR image under unbalanced sample distribution.

[0067] Then r a,e and r a,s are spliced together to get the final mixed features r a , that is:

[0068] r a =Concat(r a,e ,r a,s ),

[0069] where Concat(·) represents the feature splicing operation, r a∈ R w+u by r a,e and r a,s are spliced.

[0070] Step 106, using the projection head to map the first mixed feature to the embedding space, setting the first loss function according to all embeddings in a batch; using the first loss function to train the encoder to obtain the trained encoder.

[0071] Input r a will pass through a projection head to map the fused feature to the embedding space, which can be described as:

[0072] z a = f P (r a ),

[0073] where z a represents the embedding obtained by mapping r a , f P (·) represents the mapping operation, and the projection head can usually be composed of multiple fully connected layers or a multi-layer perception. The structure diagram of the above encoder and projection head is shown in Figure 4 .

[0074] The first loss function is set according to all embeddings in a batch. The loss function of the present application is different from the ordinary loss function in that the contrast loss measures the similarity between a batch of samples, and all similar samples (positive samples) are aggregated together, and all different samples (negative samples) in the embedding space are separated. In self-supervised contrast learning, the self-supervised contrast loss only regards the augmented embedding corresponding to the anchor point as a positive sample, and regards all other samples as negative samples, which is an instance-based method inconsistent with the classification task. The loss function of the present application regards the augmented samples corresponding to the anchor point and the samples with the same label as the anchor point as positive samples, and other samples as negative samples. Such proxy task design is consistent with the classification task. For samples belonging to the same category, although the number of samples is limited, there will be a center in the embedding space. It is a kind of long-tail recognition classification task suitable for the task, and can have good performance on the task.

[0075] Step 108, according to the trained encoder and the scale-invariant feature transformation feature extraction module, the input radar image is feature-extracted and spliced to obtain the second mixed feature; the second mixed feature is input into the classifier to obtain a prediction result vector; a second loss function is designed according to the prediction result vector and the real label of the radar image; the classifier is trained according to the second loss function to obtain the trained classifier.

[0076] For a single input image g, it will first pass through the encoder and SIFT feature extraction module trained in the last stage, that is:

[0077] m e =f E*

[0078] where f E* (·) represents the encoder feature extraction operation obtained in the last stage, m e ∈R w is the texture feature obtained in this stage. At the same time, the input image g will also pass through the SIFT feature extraction operation to obtain m s ∈R u , that is:

[0079] m s =SIFT(g),

[0080] Similarly, SIFT(·) also represents the SIFT feature extraction operation. m s and m e will also be spliced to obtain the mixed feature of this stage:

[0081] m=Concat(m s ,m e ),

[0082] where m represents the mixed feature of this stage, and then m will be input into a linear layer-based classifier to obtain a prediction vector, that is:

[0083] v=f C (m),

[0084] where f C (·) represents the forward operation of the classifier mapping the fused feature to the prediction vector.

[0085] The second mixed feature is input into the classifier to obtain the prediction vector; a second loss function is designed according to the prediction vector and the true label of the radar image, and the second loss function is set in the fine-tuning stage of the fusion feature to adjust the classifier output by directly measuring the gap between the prediction vector and the true label of the radar image, which can improve the output accuracy of the classifier.

[0086] Step 110, the trained encoder and classifier are substituted into the radar target recognition model to obtain a trained radar target recognition model; and radar target recognition is performed according to the trained radar target recognition model.

[0087] ​In a specific downstream identification task experiment, as shown in Table 1, compared with the existing method, the application has a greater improvement in the accuracy of radar target identification. It is proved that the application can obtain long-tail SAR data targets with good representation.

[0088] Table 1

[0089]

[0090] In the above radar target identification method based on scale invariant feature transformation, the application proposes a dual-channel decoupling learning representation method based on a scale invariant feature transformation feature extraction module. In the feature extraction learning stage and the classifier fine-tuning stage, the scale invariant feature transformation feature extraction module provides SAR image features with good separability, which is used for long-tail SAR image classification. The SAR physical scattering properties are fully utilized to improve the feature extraction capability of SAR images under unbalanced sample distribution. According to the extracted features, a first loss function is designed in the feature extraction learning stage using the fusion features. The augmented samples corresponding to the anchor points and the samples with the same label as the anchor points are all used as positive samples, and other samples are used as negative samples for supervised contrast learning, so that the learning features have better balance in all categories. At the same time, a second loss function is set in the classifier fine-tuning stage using the fusion features. The difference between the predicted result vector and the true label of the radar image is directly measured to adjust the classifier output. The loss function designed according to the fusion features with better representation capability can improve the accuracy of model training, and thus improve the accuracy of subsequent radar target identification.

[0091] In one embodiment, the scale invariant feature transformation feature extraction module and the encoder are used to extract features from the augmented sample pair to obtain the scattering features and the texture features of the radar image, including:

[0092] According to the scale invariant feature transformation feature extraction module, the K scattering point information of the augmented sample is extracted, and the scattering information of the kth scattering point is calculated, including the coordinates of the scattering intensity maximum point the maximum scattering intensity direction and the module length corresponding to the direction

[0093] Definition represents the kth scattering point feature, P = [p1, p2, …, p K represents the scattering features extracted from the augmented sample i, and T represents the transpose operation.

[0094] In one embodiment, the coordinates of the scattering intensity maximum point are

[0095]

[0096] where S denotes the difference of Gaussian space, C k denotes the extreme point with the maximum scattering intensity in the difference of Gaussian space.

[0097] In one embodiment, the direction of the maximum scattering intensity is

[0098]

[0099] wherein, denotes the scale space, N(x k ,y k ,σ k ) denotes the Gaussian distribution with scale factor σ k , x k ,y k denotes the coordinates of the point, L1=L(x k ,y k +1,σ k ), L2=L(x k ,y k -1,σ k ), L3=L(x k +1,y k ,σ k ), L4=L(x k -1,y k ,σ k ).

[0100] In one embodiment, the length of the direction is

[0101]

[0102] wherein, wherein, denotes the scale space, N(x k ,y k ,σ k ) denotes the Gaussian distribution with scale factor σ k , x k ,y k denotes the coordinates of the scattering point, L1=L(x k ,y k +1,σ k ), L2=L(x k ,y k -1,σ k ), L3=L(x k +1,y k ,σ k ), L4=L(x k -1,y k ,σ k ).

[0103] In one embodiment, the texture feature of the radar image is obtained according to feature extraction of the encoder on the augmented sample pair.

[0104]

[0105] wherein f E represents an encoder feature extraction operation, represents an augmented sample.

[0106] In one embodiment, the first loss function is set according to all embeddings in a batch, including:

[0107] The first loss function is set according to all embeddings in a batch as

[0108]

[0109] wherein z a represents an augmented sample embedding is also an anchor point in this batch, P(i) represents a set of all positive samples of z a in this batch, |P(i)| represents the number of P(i), A(i) represents a set of all samples other than the anchor point in this batch, τ is a temperature hyperparameter, and z h represents a sample other than the anchor point.

[0110] In one embodiment, the second loss function is designed according to the prediction result vector and the real label of the radar image, including:

[0111] The second loss function is designed according to the prediction result vector and the real label of the radar image as

[0112]

[0113] wherein q = [q1, q2, …, q C ] represents a real label vector containing C elements, and v represents a prediction result vector.

[0114] It should be understood that although each step in the flowchart of Figure 1 is displayed in sequence according to the direction of the arrow, these steps are not necessarily executed in sequence according to the direction of the arrow. 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 1At least one of the steps in the method can comprise a plurality of sub-steps or a plurality of stages, which sub-steps or stages are not necessarily performed at the same time, but can be performed at different times, and the order of the sub-steps or stages is not necessarily sequential, but can be performed alternately or in rotation with other steps or sub-steps or stages of other steps.

[0115] In one embodiment, as shown in FIG. 5, a scale-invariant feature transform-based radar target recognition device is provided, comprising a model construction module 502, a feature extraction module 504, an encoder training module 506, a classifier training module 508, and a radar target recognition module 510, wherein: Figure 5 The model construction module 502 is configured to obtain long-tail SAR target data, construct a radar target recognition model, and the radar target recognition model comprises a scale-invariant feature transform feature extraction module, an encoder, a projection head, and a classifier.

[0116] The feature extraction module 504 is configured to perform data augmentation on the long-tail SAR target data to obtain an augmented sample pair, perform feature extraction on the augmented sample pair according to the scale-invariant feature transform feature extraction module and the encoder to obtain scattering features and texture features of a radar image, and splice the scattering features and the texture features to obtain a first mixed feature.

[0117] The encoder training module 506 is configured to map the first mixed feature to an embedding space using the projection head, set a first loss function according to all embeddings in a batch, train the encoder using the first loss function, and obtain a trained encoder.

[0118] The classifier training module 508 is configured to perform feature extraction and splicing on an input radar image according to the trained encoder and the scale-invariant feature transform feature extraction module to obtain a second mixed feature, input the second mixed feature into the classifier to obtain a prediction result vector, design a second loss function according to the prediction result vector and a true label of the radar image, train the classifier according to the second loss function, and obtain a trained classifier.

[0119] The radar target recognition module 510 is configured to substitute the trained encoder and the classifier into the radar target recognition model to obtain a trained radar target recognition model, and perform radar target recognition according to the trained radar target recognition model.

[0120]

[0121] ​The specific limitation of the radar target recognition device based on the scale invariant feature transformation can refer to the limitation of the radar target recognition method based on the scale invariant feature transformation, which will not be repeated here. Each module in the above radar target recognition device based on the scale invariant feature transformation can be realized by software, hardware and their combination. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to call and execute the operation corresponding to each module by the processor.

[0122] In one embodiment, a computer device, which can be a terminal, is provided, and its internal structure diagram can be as shown in Figure 6 The computer device includes a processor, a memory, a network interface, a display screen and an input device connected through a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with external terminals through network connection. The computer program is executed by the processor to implement a radar target recognition method based on the scale invariant feature transformation. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer overlaid on the display screen, or a key, trackball or touchpad arranged on the shell of the computer device, or an external keyboard, touchpad or mouse, etc.

[0123] Those skilled in the art can understand that Figure 6 The structure shown in the above

[0124] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, storage, databases, or other media in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0125] The technical features of the above embodiments can be combined in any way. In order to make the description simple, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist, they should be considered as the scope of the present application.

[0126] The above embodiments only express several implementation manners of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the present application, some modifications and improvements can be made, which are all within the scope of the present application. Therefore, the scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for radar target recognition based on scale invariant feature transform, characterized in that, The method comprises: obtaining long-tail SAR target data; constructing a radar target recognition model; the radar target recognition model comprises a scale-invariant feature transformation feature extraction module, an encoder, a projection head and a classifier; performing data augmentation on the long-tail SAR target data to obtain an augmented sample pair; performing feature extraction on the augmented sample pair according to the scale-invariant feature transformation feature extraction module to obtain scattering features of a radar image; performing feature extraction on the augmented sample pair according to the encoder to obtain texture features of the radar image; and splicing the scattering features and the texture features to obtain first mixed features; mapping the first mixed features to an embedding space by using the projection head, setting a first loss function according to all embeddings in a batch, training the encoder by using the first loss function, and obtaining a trained encoder; performing feature extraction and splicing on an input radar image according to the trained encoder and the scale-invariant feature transformation feature extraction module to obtain second mixed features; inputting the second mixed features into the classifier to obtain a prediction result vector; and designing a second loss function according to the prediction result vector and a real label of the radar image; training the classifier according to the second loss function to obtain a trained classifier; substituting the trained encoder and the classifier into the radar target recognition model to obtain a trained radar target recognition model; and performing radar target recognition according to the trained radar target recognition model.

2. The method of claim 1, wherein, performing feature extraction on the augmented sample pair according to the scale-invariant feature transformation feature extraction module to obtain scattering features of a radar image, comprising: According to the scale-invariant feature transformation feature extraction module, the augmented sample pair is subjected to feature extraction, and the augmented sample is subjected to feature extraction The kth scattering point information is calculated, and the scattering information of the kth scattering point contains the coordinates of the scattering intensity maximum point , the maximum scattering intensity direction And the direction corresponding to the modulus ; Definitions represents the i-th scattering point feature, represents the augmented sample extracted scattering features, represents the transpose operation.​​​ 3. The method of claim 2, wherein, the coordinates of the scattering intensity maximum point are wherein, denotes a difference Gaussian space, denotes an extreme point having the maximum scattering intensity in the difference Gaussian space.

4. The method of claim 2, wherein, the maximum scattering intensity direction is wherein, denotes a scale space, denotes a Gaussian distribution with a scale factor denotes a coordinate of a point, denotes a coordinate of a point, , , , .

5. The method of claim 2, wherein, the module length corresponding to the direction is wherein, denotes a scale space, denotes a Gaussian distribution with a scale factor , denotes the coordinates of the scattering points, , , , .

6. The method of claim 2, wherein, The method further comprises: performing feature extraction on the augmented sample pair according to the encoder to obtain texture features of the radar image wherein, (·) denotes an encoder feature extraction operation, denotes an augmented sample.

7. The method of claim 1, wherein, setting the first loss function according to all embeddings in a batch, comprising: setting the first loss function according to all embeddings in a batch is wherein, represents the embedding of the augmented sample is also the anchor point in this batch, represents the set of all positive samples of in this batch, represents the number of in this batch, represents the set of all samples in this batch except the anchor point, is the temperature hyperparameter, represents all samples except the anchor point.

8. The method of claim 1, wherein, designing the second loss function according to the prediction result vector and the real label of the radar image, comprising: designing the second loss function according to the prediction result vector and the real label of the radar image is wherein, represents a true label vector containing elements, represents a predicted result vector.

9. A scale-invariant feature transform-based radar target recognition device, characterized by comprising: The device comprises: a model construction module configured to obtain long-tail SAR target data; and construct a radar target recognition model; the radar target recognition model comprises a scale-invariant feature transformation feature extraction module, an encoder, a projection head and a classifier; a feature extraction module configured to perform data augmentation on the long-tail SAR target data to obtain an augmented sample pair; perform feature extraction on the augmented sample pair according to the scale-invariant feature transformation feature extraction module to obtain scattering features of a radar image; perform feature extraction on the augmented sample pair according to the encoder to obtain texture features of the radar image; and splice the scattering features and the texture features to obtain first mixed features; An encoder training module is configured to map the first mixed feature to an embedding space by using the projection head, set a first loss function according to all embeddings in a batch, and train the encoder by using the first loss function to obtain a trained encoder. A classifier training module is configured to perform feature extraction and splicing on an input radar image according to the trained encoder and the SIFT feature extraction module to obtain a second mixed feature, input the second mixed feature into a classifier to obtain a prediction result vector, design a second loss function according to the prediction result vector and a real label of the radar image, and train the classifier according to the second loss function to obtain a trained classifier. A radar target recognition module is configured to substitute the trained encoder and the classifier into a radar target recognition model to obtain a trained radar target recognition model, and perform radar target recognition according to the trained radar target recognition model. 10.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-9. The processor executes the computer program to implement the steps of the method in any one of claims 1 to 8.

Images