Zero-shot spatial target recognition method combining attribute vectors

Abstract

This invention provides a zero-shot method for identifying space targets using attribute vectors. The method involves acquiring ISAR images of the target from radar equipment; generating a category attribute vector for space targets belonging to the same class; using a trained attribute model network to identify the target in the ISAR image, obtaining a predicted attribute vector; and comparing the predicted attribute vector with the category attribute vector to determine the category of the space target in the ISAR image. Because this invention utilizes attribute knowledge from training data and sample data information, it can predict the attributes of the target in the ISAR image without requiring an ISAR image sequence, thus having wide applicability. Furthermore, this invention leverages the attribute knowledge of space target categories to identify space target categories that lack training samples, overcoming the problem of existing network-based space target identification methods requiring a large number of training samples for each target class.

Description

Technical Field

[0001] This invention belongs to the field of radar target recognition technology, specifically relating to a zero-shot spatial target recognition method that combines attribute vectors. Background Technology

[0002] Space targets such as communication and navigation satellites provide convenience for people's lives. Space target identification technology is a crucial component of space target surveillance systems and is of great significance for maintaining the normal operation of space targets in orbit. Using radar echo signals to obtain ISAR images allows for a more intuitive understanding of target characteristics, providing a reliable guarantee for accurate space target identification. By extracting features from the acquired target ISAR images and training a classifier, space target ISAR image identification can be achieved.

[0003] In his master's thesis, "Research on Spatial Target Physical Feature Extraction and Recognition Method Based on 3D Reconstruction," Zhang Xiaole of Xi'an University of Electronic Science and Technology proposed a spatial target recognition method based on membership functions. This method first extracts the dimensions of target components from ISAR image sequences using component segmentation and size extraction methods. Then, it uses the extracted features to train a membership function as a classifier. The extracted features of the sample to be identified are input into the trained membership function classifier to obtain the judgment result for the sample. However, this method requires a relatively long ISAR image sequence with significant viewpoint changes to estimate accurate size results and obtain correct recognition results. In practical applications, it is difficult to obtain long-term, large-angle ISAR image sequences of targets, and this method is difficult to achieve real-time recognition.

[0004] The Huzhou Institute of Zhejiang University proposed a space satellite target recognition method based on compact feature learning in its patent application, "A Space Satellite Target Recognition Method Based on Compact Feature Learning" (Publication No. CN114373119A, Application No. CN202210006296.2). This method first proposes a convolutional autoencoder with compact constraints for learning features from space satellite target images with small sample sizes. Then, it trains the convolutional autoencoder using ISAR images. Finally, it uses the trained convolutional autoencoder to classify the ISAR images of the target to be identified. By applying compact constraints to the loss function of the traditional convolutional autoencoder, this method can simultaneously minimize reconstruction error and intra-class sample error, reducing the distance between intra-class samples in the learned feature space while increasing the distance between inter-class samples, thereby improving feature discriminative power and enabling small-sample recognition of space targets. However, this method requires training samples for each class of space targets to train the classifier, and cannot recognize spatial categories with missing training samples.

[0005] First, existing space target recognition methods require extracting target size features from ISAR image sequences. These methods demand long ISAR image sequences with significant viewpoint variations to obtain accurate size estimates and correct recognition results. However, this approach is difficult to implement in practice: on the one hand, it requires long ISAR image sequences to achieve accurate target estimation, making real-time recognition of space targets challenging; on the other hand, the stable flight attitude of space targets makes it difficult to obtain ISAR image sequences with significant viewpoint variations.

[0006] Second, existing neural network-based spatial target recognition methods require a large number of samples of various spatial targets as training data to train classifiers. However, there are many types of spatial targets, making it difficult to collect a large number of samples for each type. Spatial targets without training samples cannot be accurately identified. Summary of the Invention

[0007] To address the aforementioned problems in the existing technology, this invention provides a zero-shot spatial target recognition method that combines attribute vectors. The technical problem to be solved by this invention is achieved through the following technical solution:

[0008] This invention provides a method for zero-shot spatial target recognition combining attribute vectors, comprising:

[0009] S100: Acquires ISAR images of space targets from radar equipment and uses these ISAR images as ISAR images to be identified;

[0010] S200, based on prior information and the type of space target, generates a category attribute vector for space targets belonging to the same category;

[0011] S300, the ISAR image to be identified is identified by the trained attribute model network to obtain the predicted attribute vector of the ISAR image to be identified;

[0012] S400 compares the predicted attribute vector with the category attribute vector to obtain the category of the space target in the ISAR image to be identified.

[0013] Beneficial effects:

[0014] First, this invention proposes a spatial target attribute model network. Utilizing attribute knowledge from training data and sample data information, it can predict the attributes of ISAR images of spatial targets without requiring ISAR image sequences, and then use attribute vectors of various categories to measure and identify the target samples. This invention overcomes the problem of existing size-based spatial target recognition methods requiring long-duration ISAR image sequences with significant perspective changes. Compared to existing methods that can complete the recognition task with a single ISAR image, this invention enables real-time recognition and has a wide range of applications.

[0015] Second, this invention proposes a zero-shot identification approach. By utilizing the attribute knowledge of spatial target categories, it can identify spatial target categories that lack training samples. This invention overcomes the problem in existing network-based spatial target identification methods that require a large number of training samples for each target category. Compared to existing methods, it can identify newly entered targets or targets without established databases, which lack training samples.

[0016] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating a zero-sample spatial target recognition method combining attribute vectors provided by the present invention. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0019] like Figure 1 As shown, this invention provides a zero-shot spatial target recognition method combining attribute vectors, comprising:

[0020] S100: Acquire ISAR images of space targets from radar equipment and use these ISAR images as the ISAR images to be identified;

[0021] S200, based on prior information and the type of space target, generates a category attribute vector for space targets belonging to the same category;

[0022] S300, the ISAR image to be identified is identified by the trained attribute model network to obtain the predicted attribute vector of the ISAR image to be identified;

[0023] The attribute model network includes a feature extractor and a multilayer perceptron.

[0024] S400 compares the predicted attribute vector with the category attribute vector to obtain the category of the space target in the ISAR image to be identified.

[0025] This invention proposes a zero-shot target recognition scheme for space targets. For new target categories without training samples, there is no need to train a network model; only the attribute vectors of all target categories need to be obtained. This allows for accurate identification of target category samples without training samples. In practical applications, it can achieve rapid and accurate identification of newly orbiting satellite targets, with a wide range of application scenarios. The network is trained using generated point model simulation data and space echo data. The point model simulation data is inexpensive to obtain and can simulate and generate a large amount of echo data of different types and attitude angles to improve the learning of the attribute model network and enhance the network's zero-shot recognition performance.

[0026] In a specific embodiment of the present invention, the training process of the trained attribute model network includes:

[0027] S510, acquire ISAR images of spatial targets with first attribute vectors, and randomly select multiple ISAR images to generate a training sample set with attribute vectors; each ISAR image corresponds to a first attribute vector.

[0028] In practice, the embodiments of the present invention select a type of space target sample from Beidou as the first training sample. For this type of space target, 30 training samples with space target attribute vectors are randomly selected to form a training sample set. The attribute vector is 5-dimensional and contains five attributes: the number of solar panels, the position of the solar panel relative to the main body, whether the solar panel is symmetrically distributed about the main body, whether there is a nozzle component, and the aspect ratio of the main body.

[0029] S520: Randomly generate multiple second attribute vectors, generate an ISAR image for each second attribute vector, and use it as a second training sample;

[0030] In practice, this invention selects five types of space targets: Apollo, Beidou, Keyhole, Land, and Tiangong. Among them, Beidou has training samples, while Apollo, Keyhole, Land, and Tiangong do not have training samples. The attribute vectors generated for each type are 5-dimensional. The attributes included in the attribute vectors are the number of solar panels, the position of the solar panels relative to the main body, whether the solar panels are symmetrically distributed about the main body, whether there are nozzle components, and the aspect ratio of the main body.

[0031] S530, build attribute model network;

[0032] S540, take the first training sample and its corresponding first attribute vector as a pair, and the second training sample and its corresponding second attribute vector as a pair, and train the attribute model network to obtain the trained attribute model network.

[0033] In a specific embodiment of the present invention, S510 includes:

[0034] S511, acquire an ISAR image with spatial target attribute vectors;

[0035] S512, randomly select P1 ISAR images from the ISAR images of each space target to form a training sample set;

[0036] Where P1≥20, the i-th first training sample in the training sample set is x. i Let i = 1, 2, ..., I, where I is the number of training samples in the training sample set, and the label of the i-th first training sample is y. i The first attribute vector of the i-th first training sample is z. i =[z i1 ,z i2 ,...,z ik ,...,z iK ], k = 1, 2, ..., K, z ik Let K represent the k-th dimension attribute of the i-th first training sample, where K represents the attribute dimension.

[0037] In a specific embodiment of the present invention, S520 includes:

[0038] S521, randomly generate P2 second attribute vectors, P2≥200, and the j-th second attribute vector generated is z. j =[z j1 ,z j2 ,...,z jk ,...,z jK ], j = 1, 2, ..., P2, z ik This represents the k-th dimension attribute of the j-th training sample;

[0039] S522, using the ISAR imaging method, generate scattering point model training samples for each second attribute vector, and use each scattering point model training sample as a second training sample; where the scattering point model training sample generated by the j-th second attribute vector is x. I+j .

[0040] In a specific embodiment of the present invention, S522 includes generating scattering point model training samples for each second attribute vector using a range-Doppler imaging method. The generation process includes:

[0041] S522a, Generate a spatial scattering point model set L = {l1, l2, ..., l p ,...,l lnum};

[0042] Among them, l p=[l px ,l py ,l pz [] represents the coordinates of the p-th scattering point in the point model, where p = 1, 2, ..., P, and P is the total number of scattering points in the scattering point model, P ≥ 200. px , l py , l pz These are the three-dimensional coordinates of the p-th scattering point, and lnum represents the number of scattering points in the point model.

[0043] S522b uses the following formula to calculate the distance from each scattering point in the point model to the radar:

[0044] r p (t)=R0+l px sinw e tf s -l py cosw e tf s ;

[0045] Where, r p (t) represents the distance from the p-th scattering point to the radar via the t-th imaging pulse, where t = 1, 2, ..., F. s F s f is the total number of imaging moments. s R0 is the imaging time interval, and w is the distance from the spatial target to the radar. e The rotational angular velocity of the target in space;

[0046] S522c, the point model echo is calculated using the following formula:

[0047]

[0048] Among them, s R (τ,t) represents the echo signal of the space target at the τth sampling point of the tth imaging pulse, where τ = 1, 2, ..., F n F n f is the total number of sampling points within the pulse. n f is the sampling time interval. c denoted as the signal carrier frequency, μ as the signal modulation frequency, and c as the speed of light;

[0049] S522d, the ISAR image is obtained by performing a two-dimensional inverse Fourier transform on the model echo using the following formula:

[0050]

[0051] Among them, s I (r,r a ) represents the r-th row and r-th position of the ISAR image. aColumn pixel values, r = 1, 2, ..., F n r a =1,2,...,F s ;

[0052] S522e uses the ISAR image obtained from the two-dimensional inverse Fourier transform in S522d to determine the second training sample.

[0053] In a specific embodiment of the present invention, S540 includes:

[0054] S541, the weights of the attribute model network are randomly initialized using normally distributed random points, and the bias of the network is initialized to 0, thus obtaining the initialized attribute model network.

[0055] S542, the first training sample and its corresponding first attribute vector are used as a pair of input samples, and the second training sample and its corresponding second attribute are used as a pair of input samples, both of which are input into the attribute model network to obtain the output predicted attribute vector.

[0056] S543, calculates the distance metric between the predicted attribute vector and the attribute vector corresponding to the input training sample using the attribute metric cost function;

[0057] S543, the parameters of the attribute model network are iteratively updated through the backpropagation algorithm until the attribute metric cost function of the attribute model network converges, thus obtaining the trained attribute model network.

[0058] The attribute measurement cost function in S543 is expressed as follows:

[0059]

[0060] Where Loss represents the cross-entropy loss function, ∑ represents the summation operation, ||·||2 represents taking the l2 norm, P1 is the number of the first attribute vectors, and P2 is the number of the second attribute vectors. Let f be the predicted attribute vector of the i-th input sample. i Let be the attribute vector corresponding to the i-th input sample.

[0061] In one specific embodiment, S542 includes:

[0062] S542a uses the VGG16 network as a feature extractor;

[0063] This step can use the VGG16 network as a feature extractor. Other feature extraction networks, such as ResNet and GoogleNet, can also be used to achieve feature extraction.

[0064] S542b extracts feature vectors from each pair of input samples using a feature extractor, and uses a multilayer perceptron as a nonlinear mapping network to input the feature vectors into the nonlinear mapping network to obtain attribute prediction vectors.

[0065] In one specific embodiment, S400 compares the predicted attribute vector with the category attribute vector using the following formula to obtain the category to which the spatial target in the ISAR image to be identified belongs:

[0066]

[0067] in, This indicates the label corresponding to the minimum result. Indicates the category to which a space target in the ISAR image to be identified belongs. a test The attribute vector, attr, represents the predicted attribute vector obtained from the ISAR image to be identified through the attribute model network. b Indicates that the space target belongs to The category attribute vector.

[0068] The effectiveness of the present invention will be further explained below with reference to actual test data:

[0069] 1. Experimental conditions and experimental content

[0070] The software platform used in this experiment is: Windows 10 operating system, Matlab R2017a, Python 3.7, PyTorch, and FEKO 14.0.

[0071] The hardware platform used in this experiment is: Dell T7910 workstation, CPU: Intel Core™ i7-4770, GPU: NVIDIA GTX 2080Ti.

[0072] The radar parameters set for this experiment are: carrier frequency 16GHz, frequency modulation bandwidth 2GHz, repetition frequency 51.2Hz, number of pulses during dwell time 128, and number of sampling points per pulse 128.

[0073] The space target data used in this experiment are echo data obtained through simulation using FEKO software. The point model simulation data used in this experiment are obtained through simulation using Matlab software using the method given in this invention. The training set includes 100 ISAR image training samples of BeiDou space targets and 100 training samples of each of the 64 generated point model targets. The test set includes 50 test samples of each of the four types of space targets: Apollo, Keyhole, Land, and Tiangong.

[0074] 2. Experimental Results and Analysis

[0075] The invention aims to train an attribute model to achieve zero-shot recognition of spatial object categories for which no training samples are available. Experimental results are shown in Table 1.

[0076] Table 1. Experimental results under three wearing conditions

[0077]

[0078] As shown in Table 1, the method proposed in this invention can accurately identify samples of these categories even without training samples, achieving a comprehensive recognition rate of 91.85%. Existing methods cannot complete the task of identifying spatial targets without training samples. This invention is the first to propose a zero-shot recognition method, enabling the classification of categories without training samples, and has high practical application value.

[0079] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0080] Although this application has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, the disclosure, and the appended claims in carrying out the claimed application. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality.

[0081] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for zero-shot spatial target recognition combining attribute vectors, characterized in that, include: S100: Acquires ISAR images of space targets from radar equipment and uses these ISAR images as ISAR images to be identified; S200, based on prior information and the type of space target, generates a category attribute vector for space targets belonging to the same category; S300, the ISAR image to be identified is identified by the trained attribute model network to obtain the predicted attribute vector of the ISAR image to be identified; S400 compares the predicted attribute vector with the category attribute vector to obtain the category of the space target in the ISAR image to be identified; The training process of the trained attribute model network includes: S510, acquire ISAR images of spatial targets with first attribute vectors, and randomly select multiple ISAR images as first training samples with attribute vectors; each ISAR image corresponds to a first attribute vector; S520: Randomly generate multiple second attribute vectors, generate an ISAR image for each second attribute vector, and use it as a second training sample; S530, build attribute model network; S540, take the first training sample and its corresponding first attribute vector as a pair, and the second training sample and its corresponding second attribute vector as a pair, and train the attribute model network to obtain the trained attribute model network. The S520 includes: S521, randomly generated A second attribute vector, The generated first The second attribute vector is , , Indicates the first The training sample of the th training sample Dimensional attributes; S522, using the ISAR imaging method, generate scattering point model training samples for each second attribute vector, and use each scattering point model training sample as a second training sample; wherein, the generated first... The training samples for the scattering point model generated by the second attribute vector are: ; S522 includes generating scattering point model training samples for each second attribute vector using a range-Doppler imaging method. The generation process includes: S522a, Generate a set of spatial scattering point models ; in, Representing the point model Coordinates of the scattering points , The total number of scattering points in the scattering point model. , , , The first The three-dimensional coordinates of each scattering point This indicates the number of scattering points in the point model; S522b uses the following formula to calculate the distance from each scattering point in the point model to the radar: ； in, Indicates the first The scattering point of the th scattering point The distance from each imaging pulse to the radar. , The total number of imaging moments. For the imaging time interval, The distance from the space target to the radar. The rotational angular velocity of the target in space; S522c, the point model echo is calculated using the following formula: ； in, Indicates that the space target is in the first place. The imaging pulse of the first Echo signal at each sampling point , This represents the total number of sampling points within the pulse. The sampling time interval, For signal carrier frequency, To adjust the frequency of the signal, The speed of light; S522d, the ISAR image is obtained by performing a two-dimensional inverse Fourier transform on the model echo using the following formula: ； in, Represents the first ISAR image Line 1 Column pixel values, , ; S522e uses the ISAR image obtained from the two-dimensional inverse Fourier transform in S522d to determine the second training sample.

2. The zero-shot spatial target recognition method combining attribute vectors according to claim 1, characterized in that, The S510 includes: S511, acquire an ISAR image with spatial target attribute vectors; S512, randomly select from the ISAR images of each type of space target The training sample set consists of ISAR images; in, The training sample set The first training sample is , , Let be the number of training samples in the training sample set, and be the th . The label of the first training sample is , No. The first attribute vector of the first training sample is , Indicates the first The first training sample Dimensional attributes, Indicates the attribute dimension.

3. The zero-shot spatial target recognition method combining attribute vectors according to claim 1, characterized in that, The S540 includes: S541, the weights of the attribute model network are randomly initialized using normally distributed random points, and the bias of the network is initialized to 0, thus obtaining the initialized attribute model network. S542, the first training sample and its corresponding first attribute vector are used as a pair of input samples, and the second training sample and its corresponding second attribute are used as a pair of input samples, both of which are input into the attribute model network to obtain the output predicted attribute vector. S543, calculates the distance metric between the predicted attribute vector and the attribute vector corresponding to the input training sample using the attribute metric cost function; S543, the parameters of the attribute model network are iteratively updated through the backpropagation algorithm until the attribute metric cost function of the attribute model network converges, thus obtaining the trained attribute model network.

4. The method for zero-shot spatial target recognition combining attribute vectors according to any one of claims 1 to 3, characterized in that, Attribute model networks include feature extractors and multilayer perceptrons.

5. The zero-shot spatial target recognition method combining attribute vectors according to claim 4, characterized in that, S542 includes: S542a uses the VGG16 network as a feature extractor; S542b extracts feature vectors from each pair of input samples using a feature extractor, and uses a multilayer perceptron as a nonlinear mapping network to input the feature vectors into the nonlinear mapping network to obtain attribute prediction vectors. .

6. The zero-shot spatial target recognition method combining attribute vectors according to claim 5, characterized in that, The attribute measurement cost function in S543 is expressed as follows: ； in, Represents the cross-entropy loss function. This represents the summation operation. Indicates taking Norm, The number of elements in the first attribute vector. The number of elements in the second attribute vector. For the first The predicted attribute vector of each input sample. For the first The attribute vector corresponding to each input sample.

7. The method for zero-shot spatial target recognition combining attribute vectors according to claim 1, characterized in that, In S400, the predicted attribute vector is compared with the category attribute vector using the following formula to obtain the category of the spatial target in the ISAR image to be identified: ； in, This indicates the label corresponding to the minimum result. Indicates the category to which a space target in the ISAR image to be identified belongs. , This represents the predicted attribute vector obtained from the ISAR image to be identified through the attribute model network. Indicates that the space target belongs to The category attribute vector.

Images