Space-based space target detection and identification method based on open set domain adaptation

Abstract

The invention relates to the technical field of spaceflight artificial intelligence, and provides a space-based space target detection and identification method based on open set domain adaptation, which comprises the following steps: S1, acquiring an ISAR image and an RGB image of a space target; s2, marking the acquired ISAR image of the space target, generating a space target type label, and dividing an ISAR image data set into an ISAR image training set and an ISAR image test set; s3, dividing the RGB image data set into an RGB image training set and an RGB image test set; step S4, an open set domain adaptation model oriented to ISAR-RGB data is established; the open set domain adaptive model structure comprises a backbone network module for shared feature extraction, an RGB image recognition branch and an ISAR image recognition branch; and S5, using an RGB image training set and an ISAR image training set to carry out iterative training on the open set domain adaptive model until the model converges, and obtaining a trained open set domain adaptive model. According to the method, the multi-dimensional sensing capability and the detection robustness of the space target in a complex space environment can be improved.

Description

Technical Field

[0001] The present invention relates to the field of aerospace artificial intelligence technology, and in particular to a space-based space target detection and recognition method based on open set domain adaptation. Background Art

[0002] As human beings continue to explore and use space, the types and numbers of space targets are also increasing rapidly. These targets include not only traditional communication and navigation satellites, but also scientific research satellites, earth observation satellites, and spacecraft used for deep space exploration. However, the resulting problem of space debris is also becoming increasingly serious. Rocket debris, failed satellites and other debris are moving at high speeds in orbit, which may pose a threat to existing space assets such as spacecraft and satellites. Countries and international organizations around the world are studying and implementing various methods to track, monitor and reduce space debris to ensure the sustainability of the space environment.

[0003] Compared with traditional ground-based radar, optical and other monitoring methods, space-based monitoring has unique advantages. Usually, space-based monitoring systems are deployed in low-Earth orbits where space debris is densely distributed. Therefore, it can avoid interference from the Earth's atmosphere, provide higher-precision and higher-frequency observation data, and have the ability to work all day and all weather, which makes space-based monitoring gradually become an important development direction for space situational awareness. At present, the on-orbit tracking and detection of space targets relies on radar and optical high-quality continuous high-resolution imaging perception.

[0004] Inverse Synthetic Aperture Radar (ISAR) is an advanced radar technology that can be used for space-based imaging. It can perform high-resolution imaging of moving targets in the sub-terahertz frequency band, break through the limitations of optical images, achieve optimal observation according to needs, and thus realize all-weather, multi-angle tracking and monitoring of targets. In addition, the space target recognition method based on deep learning models can automatically extract target features from ISAR image data and perform intelligent classification and recognition. However, due to factors such as Doppler frequency shift caused by target motion, sidelobe interference of radar signals, strong environmental clutter, and potential military confrontation, the imaging quality of the obtained ISAR images is poor, and the data collection process is complex and difficult, which greatly limits the acquisition of high-quality ISAR images. Due to these challenges, when the training samples are insufficient, the performance of the deep learning model based on ISAR images is seriously affected. In contrast, the optical image data acquisition of space targets is relatively simple and can provide more high-quality samples for model training.

[0005] Therefore, it is necessary to find an effective method to transfer the knowledge learned in the field of optical images to the field of ISAR images to alleviate the problem of insufficient training samples and thus improve the recognition performance of the model, which is of great strategic significance for enhancing space situational awareness capabilities.

[0006] The patent with application publication number CN117437549A discloses "A method for ISAR image target recognition based on transfer learning", which involves using a transfer network unit to transfer knowledge learned from the optical image field to the ISAR image recognition field. It is completed in two parts. First, the ISAR and optical image domain features are extracted respectively by a deep convolutional neural network, and then the knowledge transfer is completed by using a transfer network unit. The shortcomings of this method include two aspects: first, during the feature extraction process, the same deep convolutional neural network framework may have different feature expression capabilities for different image domains, which may lead to information loss or distortion during the transfer process, thereby affecting the overall performance of the model; second, only one layer of transfer network units is used for feature space mapping, which may not fully capture the complex relationship between domains, thereby limiting the depth and effectiveness of knowledge transfer; third, this method assumes that the category distribution of the two is completely consistent, ignoring possible unknown categories or new categories. In this case, the generalization ability of the model in the target domain may be limited. Summary of the invention

[0007] The present invention mainly solves the technical problems that the ISAR image data set of the existing space target recognition method is limited in size and the categories are not rich enough, which leads to the limited performance of the detection and recognition model based on deep learning in practical applications, especially when dealing with complex scenes or unknown targets, and the domain adaptation model of the existing technology cannot effectively deal with the inconsistency between the source domain and the target domain categories. A space-based space target detection and recognition method based on open set domain adaptation is proposed. This method is based on open set recognition technology and uses feature alignment mechanism to enhance the feature consistency of common categories in the source domain and the target domain. By introducing an adaptive strategy, the recognition boundary is dynamically adjusted, thereby improving the detection accuracy and robustness of unknown categories.

[0008] The present invention provides a method for detecting and identifying space-based objects based on open set domain adaptation, comprising:

[0009] Step S1, acquiring an ISAR image and an RGB image of a space target to form a space target data set;

[0010] Step S2, annotating the acquired ISAR images of the space targets, generating space target type labels, and dividing the ISAR image data set into an ISAR image training set and an ISAR image test set;

[0011] Step S3, dividing the RGB image data set into an RGB image training set and an RGB image test set;

[0012] Step S4, building an open set domain adaptation model for ISAR-RGB data; the open set domain adaptation model structure includes a backbone network module for shared feature extraction, an RGB image recognition branch, and an ISAR image recognition branch;

[0013] Step S5: using the RGB image training set and the ISAR image training set, iteratively train the open set domain adaptation model until the model converges, thereby obtaining a trained open set domain adaptation model.

[0014] Furthermore, the step S1 specifically includes the following steps S11 to S13:

[0015] Step S11, obtaining RGB images through public data sets to obtain RGB image data sets;

[0016] Step S12, the ISAR image is constructed by building a 3D model, calculating the echo data of the sub-THz band, and obtaining a high-resolution ISAR image from the echo data;

[0017] Step S13: Combine the RGB image dataset and the ISAR image dataset to form a cross-modal spatial target dataset, and determine the types of known source domains and unknown source domains.

[0018] Furthermore, the step S12 specifically includes the following steps S121-S124:

[0019] Step S121: Use three-dimensional modeling software to perform 3D modeling on the space target to construct a 3D model of the space target;

[0020] Step S122: construct an electromagnetic simulation environment, perform electromagnetic simulation calculations, and obtain response data of space targets under various electromagnetic environments;

[0021] Step S123, applying a range Doppler algorithm to form an ISAR image of the space target;

[0022] Step S124, repeating the above steps S121 to S124, obtaining ISAR images of various types of space targets at different viewing angles, and obtaining an ISAR image data set.

[0023] Further, the step S2 includes the following steps S21-S23:

[0024] S21, screening available ISAR images and adjusting the size of the original images;

[0025] S22, marking the space target of each space target type to generate a space target type label;

[0026] S23, dividing the ISAR image data set processed by step S21-step S22 into an ISAR image training set and an ISAR image test set.

[0027] Further, the step S4 includes the following steps S41-S44:

[0028] Step S41, using ResNet50 as the backbone network module for shared feature extraction;

[0029] Step S42: construct an RGB image recognition branch, wherein the RGB image recognition branch takes the RGB image data as input, uses the backbone network module to perform feature extraction, and then outputs a classification result;

[0030] Step S43, building an ISAR image recognition branch, wherein the ISAR image recognition branch takes the ISAR image data as input, uses the backbone network module to perform feature extraction, introduces an orthogonal projection operator to perform feature decomposition, and combines a dynamic threshold discrimination strategy based on probability density estimation, so as to effectively handle open set classification problems and cross-domain adaptation tasks;

[0031] Step S44: define the loss function L of the entire model.

[0032] Furthermore, the overall loss function includes: cross entropy loss L for pseudo label learning ce , category prototype alignment loss L proto , eigendecomposition regularization loss L reg And the feature consistency regularization loss L con ;

[0033] The cross entropy loss L ce Defined as:

[0034]

[0035] Here, L ce represents the cross entropy loss for pseudo-label learning, It is a sample The pseudo-label of Represents the model prediction sample after SoftMax normalization The probability of belonging to category c;

[0036] Category prototype alignment loss L proto By minimizing the distance between prototypes of the same category in the source and target domains, the semantic features are made consistent between the two domains:

[0037]

[0038] Among them, L ptoto represents the category prototype alignment loss, is the prototype of category c in the source domain, is the prototype of category c in the target domain;

[0039] For each known type c∈C, calculate the mean semantic feature of the category in the source domain:

[0040]

[0041] In the formula, is the number of samples belonging to category c in the source domain, is the semantic feature of the i-th source domain sample;

[0042] For the target domain, it is necessary to estimate the mean of the semantic features of each category; since the target domain does not contain labels, consider using the pseudo-label method:

[0043]

[0044] In the formula, is the number of samples assigned to category c in the target domain, is the semantic feature of the i-th target domain sample;

[0045] Eigendecomposition regularization loss L reg It is used to ensure the effective separation of semantic information and domain information, avoid information redundancy and feature confusion, and thus achieve domain invariance of semantic features and domain specificity of domain features. It is defined as follows:

[0046]

[0047] p i represents the probability that sample i belongs to the semantic feature category; L reg represents the eigendecomposition regularization loss;

[0048] The feature consistency regularization loss L con The definition is as follows:

[0049]

[0050] in, Representation sample The set of nearest neighbor samples in feature space; Represents the target sample Neighborhood The average consistency score of all samples for category c; L con represents the feature consistency regularization loss;

[0051] The total loss function L of the model can be expressed as:

[0052] L=λ1·L ce +λ2·L proto +L reg +L con

[0053] Among them, λ1 and λ2 are weight hyperparameters.

[0054] Furthermore, step S5 specifically includes the following steps S51-S52:

[0055] Step S51: Use the RGB image training set to perform pre-training, and update the model parameters by the gradient descent method so that the model can effectively identify the spatial targets in the RGB image;

[0056] Step S52: Use the ISAR image training set to perform training to obtain a trained open set domain adaptation model.

[0057] Furthermore, after step S5, the method further includes:

[0058] Step S6: Based on the trained open set domain adaptation model, the ISAR image test set is used for testing to calculate the model recognition accuracy.

[0059] Furthermore, step S6 specifically includes the following steps S61-S62:

[0060] Step S61, using the ISAR image test set to test the trained ISAR image recognition branch; for each input sample, calculate its probability density value in the feature space, and perform sample-level adaptive discrimination according to the discrimination threshold ω obtained in the training phase;

[0061]

[0062] sample The prediction results depend on when When the sample Identified as a type unknown to the source domain; when When the sample Identify as a type known to the source domain;

[0063] Step S62: For the open-set domain adaptation task, H-score is used as the performance evaluation indicator.

[0064] The invention provides a method for detecting and identifying space-based objects based on open set domain adaptation, which has the following beneficial effects:

[0065] 1. The open set domain adaptation-based space-based target detection and recognition method of the present invention adopts sub-THz electromagnetic simulation technology to obtain target echo data, and further inverts the shape and characteristics of the target by simulating the interaction between electromagnetic waves and targets, in order to solve the problems of ISAR images being difficult to obtain and having low resolution. This method effectively overcomes the difficulty of obtaining ISAR images and the resolution limitations in practical applications, and provides high-precision echo data support for target identification and analysis. At the same time, the use of sub-THz frequency bands can provide more detailed spatial resolution, helping to improve the accuracy of target shape reconstruction.

[0066] 2. The open set domain adaptation-based space-based space target detection and recognition method of the present invention innovatively applies the open set domain adaptation technology to the space target ISAR image recognition task under few-sample conditions to address the problems that the amount of ISAR image data is insufficient to support deep learning needs, cannot provide effective support for space target recognition, and the target categories are insufficient to cover the diversity in actual scenes. This method overcomes the challenges of data scarcity and incomplete categories through cross-domain transfer learning, enables the model to be effectively trained on limited ISAR image data, and improves its adaptability and accuracy in practical applications.

[0067] 3. The open set domain adaptation-based space-based target detection and recognition method of the present invention aims to solve the problem that traditional methods usually rely on manually setting thresholds or adopting time-consuming iterative clustering strategies when identifying known data in the source domain with covariate shift in the target domain and distinguishing unknown data in the source domain at the same time, resulting in low efficiency and unstable performance. The present invention estimates the probability density of data distribution and adaptively adjusts the threshold to automatically distinguish known categories from unknown categories, thereby avoiding the problems of human intervention and computational overhead in traditional methods and improving the efficiency and accuracy of target recognition tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is a flow chart for implementing the method for detecting and identifying space-based space targets based on open set domain adaptation provided by the present invention.

[0069] Figure 2a is a schematic diagram of a 3D model of the U.S. Aura satellite in an embodiment of the present invention;

[0070] Figure 2b is a schematic diagram of a 3D model of the U.S. Calipso satellite in an embodiment of the present invention;

[0071] Figure 3a is a schematic diagram of an RGB image of a Sentinel-6 satellite in an embodiment of the present invention;

[0072] Figure 3bis a schematic diagram of an ISAR image of a Sentinel-6 satellite in an embodiment of the present invention;

[0073] Figure 3c is a schematic diagram of an RGB image of the Jason-1 satellite in an embodiment of the present invention;

[0074] Figure 3d is a schematic diagram of an ISAR image of the Jason-1 satellite in an embodiment of the present invention;

[0075] Figure 4 Schematic diagram of the overall network structure of the open set domain adaptation model in an embodiment of the present invention. DETAILED DESCRIPTION

[0076] In order to make the technical problems solved by the present invention, the technical solutions adopted and the technical effects achieved clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It should also be noted that, for the convenience of description, only the parts related to the present invention are shown in the accompanying drawings, rather than all the contents.

[0077] like Figure 1 As shown, an embodiment of the present invention provides a method for detecting and identifying space-based space targets based on open set domain adaptation, comprising the following process:

[0078] Step S1, obtain the ISAR image (target domain) and RGB image (source domain) of the space target to form a cross-modal space target dataset. Step S1 specifically includes the following steps S11-S13:

[0079] Step S11: Acquire RGB images through public channels to obtain an RGB image dataset.

[0080] The RGB images were obtained through the SPARK2021 Challenge released by the Interdisciplinary Center for Security, Reliability and Trust of the University of Luxembourg. The data can provide up to 10 target categories for research, namely AcrimSAT, Aquarius, Aura, Calipso, CloudSat, CubeSat, Jason-1, Sentinel-6, Terra, TRMM, and include data reflecting real scenes such as different orbit scenes, background noise, low signal-to-noise ratio and high image contrast.

[0081] Step S12: The ISAR image is constructed by building a 3D model, calculating the echo data of the sub-THz band, and obtaining a high-resolution ISAR image from the echo data. Step S12 specifically includes the following steps S121-S124:

[0082] Step S121: Use 3D modeling software to perform 3D modeling on the space target and construct a 3D model of the space target. Some examples can be seen. Figure 2a -b Schematic diagram.

[0083] In this embodiment, the 3D model resource of the space target comes from the NASA 3D library. The library provides space target 3D model data in various formats, including STL (StereoLithography, stereolithography file format), GLB (GLTransmission Format Binary File, GL transmission format binary file) and other file types, which are convenient for various applications in space target modeling and simulation. However, STL data cannot be directly used for electromagnetic simulation calculations. To solve this problem, this embodiment uses Geomagic Wrap reverse modeling software to check and repair the STL data grid to ensure the integrity and applicability of the model. Subsequently, SolidWorks software is used for copying modeling to further generate a high-precision 3D model suitable for electromagnetic simulation calculations.

[0084] Step S122: construct an electromagnetic simulation environment, perform electromagnetic simulation calculations, and obtain response data of space targets under various electromagnetic environments.

[0085] The electromagnetic simulation environment is built through model import, material setting, solver configuration, excitation source definition and surface element segmentation. Electromagnetic simulation is used for efficient and fast calculation to obtain the response data of space targets in various electromagnetic environments.

[0086] As a specific example, this embodiment uses FEKO electromagnetic simulation software to perform triangulated surface element segmentation on the three-dimensional model of the space target, adopts the sub-THz frequency band to provide higher spatial resolution, and uses the large surface element physical optics method to calculate the echo data.

[0087] Step S123: Apply the range Doppler algorithm to form an ISAR image of the space target.

[0088] Assuming that there is no cross-range unit movement, the range Doppler (RD) algorithm is applied to recover the ISAR image of the space target from the one-dimensional echo data;

[0089] Step S124: Repeat the above steps S121 to S124 to obtain ISAR images of various types of space targets at different viewing angles, and obtain an ISAR image dataset. Table 1 shows the specific parameters of the ISAR imaging process in this example. Figure 3a -d shows a schematic diagram of the corresponding RGB image data and ISAR image data.

[0090] Table 1 ISAR imaging parameter settings

[0091]

[0092]

[0093] Step S13: Combine the RGB image dataset and the ISAR image dataset to form a cross-modal spatial target dataset, and determine the types of known source domains and unknown source domains.

[0094] Step S2: annotate the acquired ISAR images of the space targets, generate space target type labels, and divide the ISAR image data set into an ISAR image training set and an ISAR image test set. Step S2 includes the following steps S21-S23:

[0095] S21. Depending on the imaging angle, some of the calculated ISAR images may not be used for subsequent processing. Therefore, it is necessary to screen the available ISAR images and adjust the size of the original images.

[0096] The original image is resized to, for example, 101×101.

[0097] S22. Mark the space targets of each space target type to generate space target type labels.

[0098] The space target type label is specifically the name of the space target (specific to a certain model of satellite).

[0099] S23, dividing the ISAR image data set processed by step S21-step S22 into an ISAR image training set and an ISAR image test set.

[0100] When dividing ISAR data, the test set must contain more types than the training set to adapt to real scenarios.

[0101] Step S3: divide the RGB image data set into an RGB image training set and an RGB image test set.

[0102] RGB image data only participates in the backbone network pre-training. Before segmentation, the original RGB image is also resized to the same size as the ISAR image.

[0103] As a specific example, Table 2 shows the division of the data set in the embodiment of the present invention.

[0104]

[0105] Step S4, building an open set domain adaptation model for space-based radar and optical (ISAR-RGB) data; the open set domain adaptation model structure includes a backbone network module for shared feature extraction, an RGB image recognition branch, and an ISAR image recognition branch. Figure 4 As shown, step S4 includes the following steps S41-S44:

[0106] Step S41, using ResNet50 as the backbone network module for shared feature extraction.

[0107] Step S42: construct an RGB image recognition branch, wherein the RGB image recognition branch takes the RGB image data as input, uses the backbone network module to perform feature extraction, and then outputs a classification result;

[0108] Assume that the backbone network model is expressed as in represents the feature extraction part, is the classification part, where h θ Some parameters are frozen, g θ Participation in training;

[0109] Step S43, building an ISAR image recognition branch, which takes ISAR image data as input, uses a backbone network module for feature extraction, introduces an orthogonal projection operator for feature decomposition, and combines a dynamic threshold discrimination strategy based on probability density estimation, which can effectively handle open set classification problems and cross-domain adaptation tasks.

[0110] The backbone network module realizes cross-domain feature representation alignment through a parameter sharing mechanism. As a specific example, in this embodiment, cross-domain feature extraction is realized by sharing the network architecture of the RGB image branch, and the ISAR image data is encoded to be mapped to the potential feature space, thereby obtaining the feature vector

[0111] As a specific example, in this embodiment, an orthogonal projection operator (such as SVD) is used to decompose the features extracted from the ISAR image into semantic features and domain features, and ensure that these two types of features are as independent as possible in the feature space. Assume that there is a "semantic" feature subspace S and a "domain" feature subspace P, satisfying S⊥P. In this way, the semantic feature is expressed as Domain features are represented as

[0112] The classification threshold is adjusted dynamically by using the probability density estimation based on the distribution of semantic features in the target domain, so as to adapt to the changes and distribution characteristics of different categories in the target domain. This process is mainly achieved through Gaussian kernel density estimation (KDE). For each type of known target, the historical data of the source domain (i.e., labeled samples) is used to calculate its semantic feature distribution, and the classification threshold in the target domain is adjusted based on this information. Specifically, for each type of known target, Gaussian kernel density estimation (KDE) is used to estimate the probability distribution of the source domain features, which is calculated as follows:

[0113]

[0114] in, represents the probability distribution of source domain features, is the semantic feature of the sample to be estimated, is the sample feature vector, N is the number of source domain samples, and σ is the bandwidth parameter of the kernel function.

[0115] For each category in the target domain, the Gaussian kernel density estimate (KDE) calculated in the source domain is used to evaluate the category probability of the target domain sample. The classification decision threshold is set by the classification confidence (i.e., category probability), and the threshold is dynamically updated as the category changes. If the probability value of the target domain sample is greater than the threshold, it is classified as the corresponding category, otherwise it is judged as an unknown category.

[0116] Step S44: define the loss function L of the entire model.

[0117] The overall loss function includes: cross entropy loss L for pseudo label learning ce , category prototype alignment loss L proto , eigendecomposition regularization loss L reg And the feature consistency regularization loss L con .

[0118] As a specific example, the cross entropy loss L of pseudo-label learning in this embodiment is ce Used to guide the model to learn discriminative features, defined as:

[0119]

[0120] Here, L ce represents the cross entropy loss for pseudo-label learning, It is a sample The pseudo-label of Represents the model prediction sample after SoftMax normalization The probability of belonging to category c.

[0121] Category prototype alignment loss L protoBy minimizing the distance between prototypes of the same category in the source and target domains, the semantic features are made consistent between the two domains:

[0122]

[0123] Among them, L proto represents the category prototype alignment loss, is the prototype of category c in the source domain, is the prototype of category c in the target domain.

[0124] For each known type c∈C, calculate the mean semantic feature of the category in the source domain (i.e., category prototype):

[0125]

[0126] In the formula, is the number of samples belonging to category c in the source domain, is the semantic feature of the i-th source domain sample.

[0127] Similarly, for the target domain, it is necessary to estimate the mean of the semantic features of each category. Since the target domain does not contain labels, consider using a pseudo-labeling method:

[0128]

[0129] In the formula, is the number of samples in the target domain that are assigned to category c (based on pseudo labels), is the semantic feature of the i-th target domain sample.

[0130] Eigendecomposition regularization loss L reg It is used to ensure the effective separation of semantic information and domain information, avoid information redundancy and feature confusion, and thus achieve domain invariance of semantic features and domain specificity of domain features. Its definition is as follows:

[0131]

[0132] p i represents the probability that sample i belongs to the semantic feature category; L reg represents the eigendecomposition regularization loss.

[0133] The feature consistency regularization loss L con By maintaining the consistency of semantic features between the nearest neighbor samples in the feature space, the generalization ability of the model and its robustness to unknown samples are improved. Its definition is as follows:

[0134]

[0135] in, Representation sample The set of nearest neighbor samples in feature space; Represents the target sample Neighborhood The average consistency score of all samples for category c; L con represents the feature consistency regularization loss.

[0136] Combining the above loss functions, the total loss function L of the model can be expressed as:

[0137] L=λ1·L ce +λ2·L proto +L reg +L con

[0138] Among them, λ1 and λ2 are weight hyperparameters, which are used to balance the impact of each loss on the overall optimization.

[0139] Step S5: using the RGB image training set and the ISAR image training set, iteratively train the open set domain adaptation model until the model converges, thereby obtaining a trained open set domain adaptation model.

[0140] The training of the open set domain adaptation model includes two parts: pre-training of the RGB image recognition branch and training of the ISAR image recognition branch. The pre-training is performed on the RGB image data, and the model parameters are adjusted by gradient updating. The joint training feature extraction module initially uses the pre-trained parameters, combined with the orthogonal projection operator and pseudo-label learning strategy, to continuously adjust the initial parameters of the model. Step S5 specifically includes the following steps S51-S52:

[0141] Step S51: Use the RGB image training set to perform pre-training, and update the model parameters by the gradient descent method so that the model can effectively identify the spatial targets in the RGB image;

[0142] Step S52: Use the ISAR image training set to train and obtain a trained open set domain adaptation model; wherein the feature extraction module initially uses pre-trained parameters and freezes the parameters of the classifier part. This process optimizes cross-domain feature learning through continuous iterative training, thereby improving the generalization ability and accuracy of the model in open set domain adaptation.

[0143] Step S6: Based on the trained open set domain adaptation model, the ISAR image test set is used for testing to calculate the model recognition accuracy.

[0144] Step S6 specifically includes the following steps S61-S62:

[0145] Step S61, using the ISAR image test set to test the trained ISAR image recognition branch. For each input sample, calculate its probability density value in the feature space, and perform sample-level adaptive discrimination according to the discrimination threshold ω obtained in the training phase;

[0146]

[0147] sample The prediction results depend on when When the sample Recognized as a type unknown to the source domain. When the sample Identified as a type known to the source domain.

[0148] Step S62: For the open-set domain adaptation task, H-score is used as the performance evaluation indicator.

[0149] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that modifying the technical solutions described in the aforementioned embodiments, or replacing some or all of the technical features therein by equivalents, does not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A space-based space target detection and recognition method based on open set domain adaptation, characterized in that: include: Step S1, acquiring an ISAR image and an RGB image of a space target to form a space target data set; Step S2, annotating the acquired ISAR images of the space targets, generating space target type labels, and dividing the ISAR image data set into an ISAR image training set and an ISAR image test set; Step S3, dividing the RGB image data set into an RGB image training set and an RGB image test set; Step S4, building an open set domain adaptation model for ISAR-RGB data; the open set domain adaptation model structure includes a backbone network module for shared feature extraction, an RGB image recognition branch, and an ISAR image recognition branch; Step S5: using the RGB image training set and the ISAR image training set, iteratively train the open set domain adaptation model until the model converges, thereby obtaining a trained open set domain adaptation model.

2. The method for space-based space target detection and recognition based on open set domain adaptation according to claim 1, characterized in that: The step S1 specifically includes the following steps S11 to S13: Step S11, obtaining RGB images through public data sets to obtain RGB image data sets; Step S12, the ISAR image is constructed by building a 3D model, calculating the echo data of the sub-THz band, and obtaining a high-resolution ISAR image from the echo data; Step S13: Combine the RGB image dataset and the ISAR image dataset to form a cross-modal spatial target dataset, and determine the types of known source domains and unknown source domains.

3. The method for space-based space target detection and recognition based on open set domain adaptation according to claim 2, characterized in that: The step S12 specifically includes the following steps S121-S124: Step S121: Use three-dimensional modeling software to perform 3D modeling on the space target and construct a 3D model of the space target; Step S122: construct an electromagnetic simulation environment, perform electromagnetic simulation calculations, and obtain response data of space targets under various electromagnetic environments; Step S123, applying a range Doppler algorithm to form an ISAR image of the space target; Step S124, repeating the above steps S121 to S124, obtaining ISAR images of various types of space targets at different viewing angles, and obtaining an ISAR image data set.

4. The method for space-based space target detection and recognition based on open set domain adaptation according to claim 1, characterized in that: The step S2 includes the following steps S21-S23: S21, screening available ISAR images and adjusting the size of the original images; S22, marking the space target of each space target type to generate a space target type label; S23, dividing the ISAR image data set processed by step S21-step S22 into an ISAR image training set and an ISAR image test set.

5. The method for space-based space target detection and recognition based on open set domain adaptation according to claim 4, characterized in that: The step S4 includes the following steps S41-S44: Step S41, using ResNet50 as the backbone network module for shared feature extraction; Step S42: construct an RGB image recognition branch, wherein the RGB image recognition branch takes the RGB image data as input, uses the backbone network module to perform feature extraction, and then outputs a classification result; Step S43, building an ISAR image recognition branch, wherein the ISAR image recognition branch takes the ISAR image data as input, uses the backbone network module to perform feature extraction, introduces an orthogonal projection operator to perform feature decomposition, and combines a dynamic threshold discrimination strategy based on probability density estimation, so as to effectively handle open set classification problems and cross-domain adaptation tasks; Step S44: define the loss function L of the entire model.

6. The method for space-based space target detection and recognition based on open set domain adaptation according to claim 5, characterized in that: The overall loss function includes: cross entropy loss L for pseudo label learning ce , category prototype alignment loss L proto , eigendecomposition regularization loss L reg And the feature consistency regularization loss L con ; The cross entropy loss L ce Defined as: Here, L ce represents the cross entropy loss for pseudo-label learning, It is a sample The pseudo-label of Represents the model prediction sample after SoftMax normalization The probability of belonging to category c; Category prototype alignment loss L proto By minimizing the distance between prototypes of the same category in the source and target domains, the semantic features are made consistent between the two domains: Among them, L proto represents the category prototype alignment loss, is the prototype of category c in the source domain, is the prototype of category c in the target domain; For each known type c∈C, calculate the mean semantic feature of the category in the source domain: In the formula, is the number of samples belonging to category c in the source domain, is the semantic feature of the i-th source domain sample; For the target domain, it is necessary to estimate the mean of the semantic features of each category; since the target domain does not contain labels, consider using the pseudo-label method: In the formula, is the number of samples assigned to category c in the target domain, is the semantic feature of the i-th target domain sample; Eigendecomposition regularization loss L reg It is used to ensure the effective separation of semantic information and domain information, avoid information redundancy and feature confusion, and thus achieve domain invariance of semantic features and domain specificity of domain features. It is defined as follows: p i represents the probability that sample i belongs to the semantic feature category; L reg represents the eigendecomposition regularization loss; The feature consistency regularization loss L con The definition is as follows: in, Representation sample The set of nearest neighbor samples in feature space; Represents the target sample Neighborhood The average consistency score of all samples for category c; L con represents the feature consistency regularization loss; The total loss function L of the model can be expressed as: L=λ1·L ce +λ2·L proto +L reg +L con Among them, λ1 and λ2 are weight hyperparameters.

7. The method for space-based space target detection and recognition based on open set domain adaptation according to claim 6, characterized in that: Step S5 specifically includes the following steps S51-S52: Step S51: Use the RGB image training set to perform pre-training, and update the model parameters by the gradient descent method so that the model can effectively identify the spatial targets in the RGB image; Step S52: Use the ISAR image training set to perform training to obtain a trained open set domain adaptation model.

8. The method for space-based space target detection and recognition based on open set domain adaptation according to claim 7, characterized in that: After step S5, the method further includes: Step S6: Based on the trained open set domain adaptation model, the ISAR image test set is used for testing to calculate the model recognition accuracy.

9. The method for space-based space target detection and recognition based on open set domain adaptation according to claim 8, characterized in that: Step S6 specifically includes the following steps S61-S62: Step S61, using the ISAR image test set to test the trained ISAR image recognition branch; for each input sample, calculate its probability density value in the feature space, and perform sample-level adaptive discrimination according to the discrimination threshold ω obtained in the training phase; sample The prediction results depend on when When the sample Identified as a type unknown to the source domain; when When the sample Identify as a type known to the source domain; Step S62: For the open-set domain adaptation task, H-score is used as the performance evaluation indicator.

Images