Adaptive training satellite target recognition method and system based on scattering feature extraction and fusion

Abstract

The present invention discloses a satellite target identification method and system based on adaptive training for scattering feature extraction and fusion. Specifically, the method comprises the following steps: based on the satellite target echoes received by an inverse synthetic aperture radar, a two-dimensional inverse Fourier transform is performed to obtain an inverse synthetic aperture radar image dataset of the original attitude; an attribute scattering center model dataset of the original attitude is obtained through an inverse parameterized modeling algorithm; the scalability of the attribute scattering center model is utilized to obtain the attribute scattering center model dataset and the inverse synthetic aperture radar image dataset after attitude transformation; a satellite target identification network model is established and trained using an adaptive training method; and an open set recognition algorithm is used to output a probability distribution vector based on the trained model to distinguish between the types of satellite targets of interest and those of non-interest. The present invention improves satellite target identification accuracy and can distinguish non-interested satellite targets, enabling target identification in more complex environments.

Description

Technical Field

[0001] The present invention belongs to the technical field of electromagnetic target recognition, and in particular to an adaptive training satellite target recognition method and system based on scattering feature extraction and fusion. Background Art

[0002] With the continuous development of aerospace technology, the number of space targets, such as satellites, is rapidly increasing. Electromagnetic target identification (EM target identification) faces challenges such as complex electromagnetic environments and the difficulty of target identification. Accurately identifying satellite targets of interest and discriminating against non-targets is a pressing issue. Traditional satellite target identification methods achieve target recognition by matching inverse synthetic aperture radar (ISAR) images extracted from ISAR measurements with target feature datasets. However, these methods have weak anti-interference capabilities and are unable to distinguish between satellite target types not included in the target feature dataset or non-targets.

[0003] In recent years, deep learning methods have been widely used in the field of target recognition due to their efficient feature extraction capabilities. IEEE Transactions on Geoscience and Remote Sensing , vol. 60, pp.1-12, 2022, Art no. 5211212.) proposed a synthetic aperture radar target automatic recognition method based on a multi-scale convolutional neural network. This method extracts typical structural components of the target from synthetic aperture radar images, thereby achieving automatic target recognition. However, this algorithm cannot extract the structural components of non-interested satellite targets and cannot achieve the discrimination of non-interested satellite targets. Reference 2 (J. -H. Choi, M. -J. Lee, N. -H. Jeong, G. Lee, and K. -T. Kim, “Fusion of target and shadow regions for improved SAR ATR,” IEEE Transactions on Geoscience and Remote Sensing, vol. 60, pp. 1-17, 2022, Art no. 5226217.) proposed a target recognition method that fuses target and shadow features. This method extracts target and shadow features from synthetic aperture radar images and adaptively fuses these two features through a convolutional neural network to achieve automatic target recognition. However, this method requires a large number of data samples to train the network model, and the recognition accuracy of the network model is easily affected by the number of training samples. Patent CN118887504A discloses a target recognition method and device based on multimodal orthogonal fusion. This method uses a deep learning method to identify target types using visible light and infrared images. However, this method can only identify target types that exist in the dataset and cannot distinguish non-targets of interest. Summary of the Invention

[0004] The purpose of the present invention is to provide a satellite target recognition method and system based on adaptive training of scattering feature extraction and fusion, which improves the accuracy of satellite target recognition by adaptively training neural networks in multiple dimensions, and can distinguish non-interested satellite targets, thus realizing target recognition in more complex environmental scenes.

[0005] The technical solution to achieve the purpose of the present invention is: an adaptive training satellite target recognition method based on scattering feature extraction and fusion, comprising the following steps:

[0006] Step 1: Obtain an inverse synthetic aperture radar image dataset of the original attitude of the satellite target from the satellite target echo received by the inverse synthetic aperture radar through two-dimensional inverse Fourier transform;

[0007] Step 2: Using an inverse parameterized modeling algorithm, obtain an attribute scattering center model dataset of the original attitude of the satellite target from the satellite target echo received by the inverse synthetic aperture radar;

[0008] Step 3: Using the extensibility of the attribute scattering center model, obtain the attribute scattering center model dataset after the satellite target attitude transformation, and determine the inverse synthetic aperture radar image dataset after the satellite target attitude transformation;

[0009] Step 4: Establish a satellite target recognition network model based on the convolutional network and the graph convolutional network. Then, use the adaptive training method to train the satellite target recognition network model based on the inverse synthetic aperture radar image dataset and the attribute scattering center model dataset of the original satellite target posture, as well as the inverse synthetic aperture radar image dataset and the attribute scattering center model dataset after the satellite target posture transformation.

[0010] Step 5: Using the open set recognition algorithm, the trained satellite target recognition network model outputs a probability distribution vector, and the probability distribution vector is used to identify the types of satellite targets of interest and the types of non-satellite targets of interest.

[0011] An adaptive training satellite target recognition system based on scattering feature extraction and fusion is provided. The system is used to implement the adaptive training satellite target recognition method based on scattering feature extraction and fusion. The system includes first to fifth units, and the functions of each unit are as follows:

[0012] The first unit uses two-dimensional inverse Fourier transform to obtain the inverse synthetic aperture radar image dataset of the original attitude of the satellite target from the satellite target echo received by the inverse synthetic aperture radar;

[0013] The second unit uses the inverse parameterized modeling algorithm to obtain the attribute scattering center model dataset of the satellite target's original attitude from the satellite target echo received by the inverse synthetic aperture radar;

[0014] Unit 3: Using the extensibility of the attribute scattering center model, we obtain the attribute scattering center model dataset after the satellite target attitude transformation, and determine the inverse synthetic aperture radar image dataset after the satellite target attitude transformation;

[0015] Unit 4: Building a satellite target recognition network model based on convolutional networks and graph convolutional networks. Then, using an adaptive training method to train the satellite target recognition network model, the model is trained based on an inverse synthetic aperture radar image dataset and an attribute scattering center model dataset of the original satellite target pose, as well as an inverse synthetic aperture radar image dataset and an attribute scattering center model dataset after the satellite target pose is transformed.

[0016] In the fifth unit, the open set recognition algorithm is used to output the probability distribution vector through the trained satellite target recognition network model, and the probability distribution vector is used to realize the identification of the types of satellite targets of interest and the discrimination of the types of satellite targets of non-interest.

[0017] Compared with the existing technology, the present invention has the following significant advantages: (1) the attribute scattering center model of the original attitude is extracted by using the inverse parameterized modeling algorithm, and the attribute scattering center model and the inverse synthetic aperture radar image after attitude transformation are obtained through the extensibility of the attribute scattering center model, thereby expanding the number of samples in the data set; (2) the deep learning features of the inverse synthetic aperture radar image and the scattering features of the attribute scattering center model are fused through deep learning, and the satellite targets are identified through the fused features, thereby improving the satellite target identification accuracy; (3) the satellite identification network model is trained based on the adaptive training method, and the network model has a higher ability to distinguish similar targets and a higher identification accuracy; (4) the satellite target identification results are output through the open set recognition algorithm, and non-focused satellite targets are distinguished, thereby realizing target identification in more complex environmental scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is the inverse synthetic aperture radar image in its original pose.

[0019] Figure 2 It is a visualization of the attribute scattering center model of the extracted original pose.

[0020] Figure 3 is the rotation angle Visualization of the attribute scattering center model for other postures.

[0021] Figure 4 is the reconstructed rotation angle Inverse synthetic aperture radar images of other postures.

[0022] Figure 5 This is a schematic diagram of the satellite target recognition network model structure.

[0023] Figure 6 It is a schematic diagram of the convolution module structure.

[0024] Figure 7 It is an inverse synthetic aperture radar image of 9 types of satellite targets of interest.

[0025] Figure 8 This is a visualization of the extracted attribute scattering center model for 9 types of satellite targets of interest.

[0026] Figure 9 The inverse synthetic aperture radar images of two types of non-interest satellite targets

[0027] Figure 10 This is a visualization of the extracted attribute scattering center model for two types of non-satellite targets.

[0028] Figure 11 This is the test result diagram of the satellite target recognition network trained using the original posture dataset.

[0029] Figure 12 This is a test result diagram of the satellite target recognition network of the present invention. DETAILED DESCRIPTION

[0030] The present invention is based on an adaptive training satellite target identification method and system for scattering feature extraction and fusion. Starting from the received echo of the inverse synthetic aperture radar, the two-dimensional inverse Fourier transform and inverse parameterized modeling algorithm are used to quickly construct the inverse synthetic aperture radar image data set and the attribute scattering center model data set. Satellite targets are identified based on the satellite target identification network model and the open set recognition algorithm. Compared with the traditional neural network-based identification method, the method and system can train the neural network in a multi-dimensional and adaptive manner, and the satellite target identification accuracy is higher.

[0031] The present invention provides a method for adaptively training satellite target recognition based on scattering feature extraction and fusion, comprising the following steps:

[0032] Step 1: Obtain an inverse synthetic aperture radar image dataset of the original attitude of the satellite target from the satellite target echo received by the inverse synthetic aperture radar through two-dimensional inverse Fourier transform;

[0033] Step 2: Using an inverse parameterized modeling algorithm, obtain an attribute scattering center model dataset of the original attitude of the satellite target from the satellite target echo received by the inverse synthetic aperture radar;

[0034] Step 3: Using the extensibility of the attribute scattering center model, obtain the attribute scattering center model dataset after the satellite target attitude transformation, and determine the inverse synthetic aperture radar image dataset after the satellite target attitude transformation;

[0035] Step 4: Establish a satellite target recognition network model based on the convolutional network and the graph convolutional network. Then, use the adaptive training method to train the satellite target recognition network model based on the inverse synthetic aperture radar image dataset and the attribute scattering center model dataset of the original satellite target posture, as well as the inverse synthetic aperture radar image dataset and the attribute scattering center model dataset after the satellite target posture transformation.

[0036] Step 5: Using the open set recognition algorithm, the trained satellite target recognition network model outputs a probability distribution vector, and the probability distribution vector is used to identify the types of satellite targets of interest and the types of non-satellite targets of interest.

[0037] As a specific example, in step 1, the inverse synthetic aperture radar image of the original attitude of the satellite target is obtained from the satellite target echo received by the inverse synthetic aperture radar through two-dimensional inverse Fourier transform. The inverse synthetic aperture radar image of the original attitude is as follows: Figure 1 As shown, the size of each inverse synthetic aperture radar image is , the inverse synthetic aperture radar images of the original attitudes of various satellite targets constitute the inverse synthetic aperture radar image dataset of the original attitudes of satellite targets.

[0038] As a specific example, in step 2, the attribute scattering center model obtained from the satellite target echo received by the inverse synthetic aperture radar is a two-dimensional attribute scattering center model, and the attribute scattering center model expression is:

[0039] (1)

[0040] in is the total scattered field, is the radar frequency, is the radar azimuth; The radar frequency is , the radar azimuth is The total scattered field when is the number of scattering centers, It is The scattered field of a scattering center, The radar frequency is , the radar azimuth is Time The scattered field of a scattering center; is the radar carrier frequency, is an imaginary number, is the electromagnetic wave propagation velocity, It is The amplitude of the scattering center, It is The frequency dependence factor of the scattering center, It is The length of the scattering center, It is The angular dependence factor of the scattering center, It is The distance dimension position of the scattering center, It is The azimuthal position of the scattering center, and Composition The two-dimensional position of the scattering center .

[0041] As a specific example, the parameters of the attribute scattering center model in step 2 include the amplitude of the scattering center, the frequency dependence factor of the scattering center, the length of the scattering center, the angle dependence factor of the scattering center, and the two-dimensional position of the scattering center. The extracted two-dimensional attribute scattering center model is as follows: Figure 2 shown.

[0042] The reverse parameterized modeling algorithm described in step 2 is as follows:

[0043] Step 2.1, obtaining a one-dimensional range image and a one-dimensional azimuth image of the satellite target from the satellite target echo through a one-dimensional inverse Fourier transform;

[0044] Step 2.2: Using a peak detection method, extract the distance dimension distribution and azimuth dimension distribution of the attribute scattering center from the one-dimensional range image and the one-dimensional azimuth image of the satellite target, respectively. Based on the distance dimension distribution and azimuth dimension distribution of the attribute scattering center, obtain the distance dimension position and azimuth dimension position of the scattering center, respectively, thereby determining the two-dimensional position of the scattering center;

[0045] Step 2.3: Based on the two-dimensional position of the extracted scattering center, locate the position of the attribute scattering center in the inverse synthetic aperture radar image, and obtain the length of the scattering center through the azimuth dimension distribution length of the attribute scattering center in the inverse synthetic aperture radar image;

[0046] Step 2.4: Obtain the angle dependence factor of the scattering center based on the length of the extracted scattering center: If the length of the scattering center is 0, then the scattering center is a local scattering center, and the angle dependence factor of the scattering center is If the length of the scattering center is greater than 0, then the scattering center is a distributed scattering center and the angular dependence factor of the scattering center is 0.

[0047] Step 2.5: Obtain the amplitude of the scattering center and the frequency dependence factor of the scattering center by the least squares method. The expression of the optimization objective function of the least squares method is:

[0048] (2)

[0049] in is the optimization objective function, The amplitude of the scattering center , No. Frequency dependence factor of the scattering center is the optimized parameter; Indicates the radar frequency is , the radar azimuth is When the amplitude is , the frequency dependence factor is No. The scattered field of a scattering center; It represents optimization to the minimum value. represents the two-norm.

[0050] The parameters of the attribute scattering center model of the satellite target are extracted through the inverse parameterized modeling algorithm to construct the attribute scattering center model of the original posture of the satellite target. The attribute scattering center models of the original postures of multiple satellite targets constitute the attribute scattering center model dataset of the original posture of the satellite target.

[0051] As a specific example, in step 3, the extensibility of the attribute scattering center model is utilized to obtain the attribute scattering center model dataset after the satellite target attitude transformation, and to determine the inverse synthetic aperture radar image dataset after the satellite target attitude transformation, as follows:

[0052] Step 3.1: As the satellite target's attitude changes, the two-dimensional position of the satellite target's attribute scattering center also changes. The attitude change of the satellite target is described by the rotation matrix, and the expression of the rotation matrix is:

[0053] (3)

[0054] in is the rotation matrix, is the rotation angle;

[0055] After the satellite target attitude changes The two-dimensional position of the scattering center is ,but , is the first The distance dimension position of the scattering center, is the first The azimuthal position of the scattering center;

[0056] The satellite target attitude changes after The length of a scattering center is ,but ; If the satellite target attitude changes after the The length of the scattering center , then the first angular dependence factor of the scattering center ; If the first The length of the scattering center , then the first angular dependence factor of the scattering center ;

[0057] Before and after the satellite target attitude changes, the amplitude of the scattering center and the frequency dependence factor of the scattering center do not change. The attribute scattering center model after attitude change is as follows: Figure 3 shown.

[0058] By processing the parameters of the attribute scattering center model in the attribute scattering center model dataset of the original attitude of the satellite target, the attribute scattering center model dataset after the attitude transformation of the satellite target is obtained;

[0059] Step 3.2: Determine the inverse synthetic aperture radar image after the satellite target attitude transformation through the attribute scattering center model after the satellite target attitude transformation. The expression is:

[0060] (4)

[0061] in is the inverse synthetic aperture radar image, is the distance dimension, is the azimuth dimension, is the radar bandwidth, is the wavelength of the radar carrier frequency, is the electromagnetic wave propagation velocity, is the radar sweep angle width;

[0062] is a bandpass filter function, if ,but ;like or ,but ;

[0063] Based on the above expression, the attribute scattering center model dataset after the satellite target attitude transformation is used to construct the inverse synthetic aperture radar image dataset after the satellite target attitude transformation. The constructed inverse synthetic aperture radar image after the attitude transformation is as follows: Figure 4 shown.

[0064] As a specific example, step 4 builds a satellite target recognition network model based on convolutional network and graph convolutional network, such as Figure 5 As shown in the figure, the satellite target recognition network model includes a convolution module, a graph convolution module, and an identification module, as follows:

[0065] The convolution module is as follows Figure 6 The figure shows a convolutional network consisting of a convolutional layer (Conv), a maximum pooling layer (MaxPool), an average pooling layer (AvgPool), and a residual structure (Residual). The convolution kernel size of the convolutional layer includes 、 and , the kernel size of the maximum pooling layer is , the kernel size of the average pooling layer is The residual structure consists of 4 convolutional layers, 4 batch normalization layers, and ReLU activation function. The convolution module inputs the inverse synthetic aperture radar image and outputs Deep learning features of size.

[0066] The graph convolution module is as follows Figure 5 As shown in the figure, it consists of four graph convolution layers. The graph convolution module takes each scattering center as a node in the graph convolution layer, and takes the amplitude, frequency dependence factor, length, angle dependence factor and two-dimensional position of the scattering center as node features. The node feature dimension is 6. The correlation between nodes is related to the node features. The correlation between nodes is the correlation between scattering centers, which is expressed as:

[0067] (5)

[0068] in For the The scattering center and The correlation degree of the scattering centers, 、 The values ​​of ; and They are The amplitude of the scattering center and the The amplitude of the scattering center, and They are The two-dimensional position of the scattering center and the The two-dimensional position of the scattering center, is the standard deviation of the Gaussian distribution, is the distance threshold; the standard deviation of the Gaussian distribution in the present invention , distance threshold ; The graph convolution module inputs the attribute scattering center model and outputs Scattering characteristics of size;

[0069] The identification module is as follows Figure 5 As shown in the figure, it consists of one fully connected layer (FC) and one dropout layer. The recognition module inputs and connects the deep learning features output by the fusion convolution module and the scattering features output by the graph convolution module, and outputs the activation vector for identifying satellite targets. Finally, through the Softmax function, the probability distribution vector of the satellite target of interest is output. , , For the The probability of satellite targets of interest, is the number of types of satellite targets of interest.

[0070] As a specific example, in step 4, based on the inverse synthetic aperture radar image dataset and attribute scattering center model dataset of the original satellite target posture, and the inverse synthetic aperture radar image dataset and attribute scattering center model dataset after the satellite target posture is transformed, an adaptive training method is used to train the satellite target recognition network model, as follows:

[0071] Step 4.1, adding the inverse synthetic aperture radar image dataset of the original attitude of the satellite target and the attribute scattering center model dataset of the original attitude of the satellite target to the training set, and pre-training the satellite target recognition network model through the training set;

[0072] Step 4.2: Calculate the recognition error for each satellite target type of interest in the training set, determine the satellite target type with the largest recognition error, and supplement the training set with the attitude-transformed inverse synthetic aperture radar image and attitude-transformed attribute scattering center model of that satellite target type. Continue training the satellite target recognition network model using this supplemented training set.

[0073] Step 4.3: If the number of network training rounds does not reach the threshold, repeat step 4.2; otherwise, end the training of the satellite target recognition network model;

[0074] The cross entropy loss function is used to train the satellite target recognition network model in steps 4.1 and 4.2, and the expression is:

[0075] (6)

[0076] in is the loss function, is the number of samples in the training set, For the The labels of samples, For the The predicted probability of a sample.

[0077] As a specific example, in step 5, the open set recognition algorithm is used to output a probability distribution vector through the trained satellite target recognition network model, and the probability distribution vector is used to realize the identification of the type of satellite target of interest and the discrimination of the type of satellite target of non-interest, as follows:

[0078] The satellite target types of interest are those that exist in the inverse synthetic aperture radar image dataset and the attribute scattering center model dataset. In the actual satellite target identification scenario, there are non-concerned satellite target types, and the samples of these non-concerned satellite target types do not exist in the established inverse synthetic aperture radar image dataset and the attribute scattering center model dataset. Assume that the satellite target types of interest are Class, then the dimension of the probability distribution vector of the satellite target of interest output by the satellite target recognition network model is .

[0079] The present invention utilizes an open set recognition algorithm to output the identification results of satellite target types of interest and the discrimination results of satellite target types of non-interest through the trained satellite target recognition network model, as follows:

[0080] Step 5.1: Obtain the activation vector of each sample in the training set through the trained satellite target recognition network model, and calculate the average activation vector of each type of satellite target in the training set;

[0081] Step 5.2: Calculate the Euler distance between the activation vector of each sample in the training set and the average activation vector of the satellite target category corresponding to each sample, and obtain the maximum Euler distance for each category of satellite targets of interest;

[0082] Step 5.3: Based on the average probability distribution vector and the maximum Euler distance of each type of satellite target in the training set, the Weibull probability distribution of each type of satellite target is calculated by fitting;

[0083] Step 5.4: Obtain the activation vector of the test sample through the trained satellite target recognition network model, and calculate the Euler distance between the activation vector of the test sample and the average activation vector of each type of satellite target. Obtain the Weibull probability distribution vector of the test sample through the Weibull probability distribution of each type of satellite target. , The test sample belongs to Weibull probability of satellite targets of interest;

[0084] Step 5.5: Use Weibull probability distribution vector to process probability distribution vector , obtain the updated probability distribution vector of the test sample , After the update The probability of satellite targets of interest, The probability of being a satellite target of no concern; when hour, ; ;right , , , and Normalize and get the updated probability distribution vector of the test sample ;

[0085] Step 5.6: If the updated probability distribution vector of the test sample is The maximum value in ,and , then the test sample belongs to The satellite target of the class; if the updated probability distribution vector of the test sample The maximum value in , then the test sample belongs to a non-concerned satellite target; thus, the recognition result of the concerned satellite target type and the discrimination result of the non-concerned satellite target type are obtained, and the recognition and discrimination results of the test sample are output.

[0086] The present invention also provides an adaptive training satellite target recognition system based on scattering feature extraction and fusion, which is used to implement the adaptive training satellite target recognition method based on scattering feature extraction and fusion. The system includes first to fifth units, and the functions of each unit are as follows:

[0087] The first unit uses two-dimensional inverse Fourier transform to obtain the inverse synthetic aperture radar image dataset of the original attitude of the satellite target from the satellite target echo received by the inverse synthetic aperture radar;

[0088] The second unit uses the inverse parameterized modeling algorithm to obtain the attribute scattering center model dataset of the satellite target's original attitude from the satellite target echo received by the inverse synthetic aperture radar;

[0089] Unit 3: Using the extensibility of the attribute scattering center model, we obtain the attribute scattering center model dataset after the satellite target attitude transformation, and determine the inverse synthetic aperture radar image dataset after the satellite target attitude transformation;

[0090] Unit 4: Building a satellite target recognition network model based on convolutional networks and graph convolutional networks. Then, using an adaptive training method to train the satellite target recognition network model, the model is trained based on an inverse synthetic aperture radar image dataset and an attribute scattering center model dataset of the original satellite target pose, as well as an inverse synthetic aperture radar image dataset and an attribute scattering center model dataset after the satellite target pose is transformed.

[0091] In the fifth unit, the open set recognition algorithm is used to output the probability distribution vector through the trained satellite target recognition network model, and the probability distribution vector is used to realize the identification of the types of satellite targets of interest and the discrimination of the types of satellite targets of non-interest.

[0092] The present invention will be described in further detail below with reference to the accompanying drawings and specific embodiments.

[0093] Example

[0094] The field of target recognition faces problems such as complex electromagnetic environment, difficulty in target recognition, and low recognition efficiency. Traditional target recognition methods have many limitations, such as low target recognition accuracy, inability to identify non-targets of interest, a large number of required training samples, and lack of wide applicability.

[0095] In this embodiment, the radar center frequency ,bandwidth , the number of frequency points is 401, the radar azimuth angle range is -3° to 3°, the number of azimuth angle samples is 512, and the polarization mode is VV polarization. There are 8 types of satellite targets of interest. The inverse synthetic aperture radar images of the 8 types of satellite targets of interest are as follows Figure 7 As shown in the figure, the attribute scattering center model extracted from the 8 types of satellite targets is as follows: Figure 8 There are two types of satellite targets of no concern. The inverse synthetic aperture radar images of the two types of satellite targets of no concern are as follows: Figure 9 As shown, the extracted attribute scattering center model of the two types of non-focused satellite targets is as follows Figure 10 There are 250 samples for each type of satellite target, of which 150 samples from each type of satellite targets of interest are used as training sets, and 100 samples from each type of satellite targets of interest and non-satellite targets of interest are used as test sets. It is worth noting that samples from non-satellite targets of interest do not exist in the training sets.

[0096] The present invention uses the satellite target recognition accuracy as the evaluation index of the invention effect. The recognition accuracy formula is:

[0097]

[0098] in For identification accuracy, is the total number of samples, For the The number of correct samples for satellite target identification of the class of interest, is the number of satellite target categories of interest, Classification is a satellite target of no concern. In this embodiment The test results of the satellite target recognition network trained with only the original pose dataset are as follows: Figure 11 As shown in Figure 2, the recognition accuracy is 80.8%. It can be found that the model has a low ability to distinguish similar satellite targets, including category 6 and category 8. The test results of the satellite target recognition network proposed by the present invention are shown in Figure 2. Figure 12 As shown in the figure, the recognition accuracy is 86.2%. It can be found that the satellite target recognition network model proposed in the present invention improves the ability to distinguish similar satellite targets. Figure 10 and Figure 11 In the table, “U” represents the satellite target type that is not of interest.

[0099] From the analysis of the results of the above examples, the adaptive training satellite target identification method based on scattering feature extraction and fusion of the present invention can extract the attribute scattering center model from the inverse synthetic aperture radar image of the satellite target, and expand the sample after the attitude transformation of the data sample through the attribute scattering center model, thereby improving the satellite target identification network model's ability to distinguish similar satellite targets, thereby improving the accuracy of satellite target identification. On the other hand, based on the open set recognition algorithm, the identification of the target of interest and the discrimination of the target of non-interest are simultaneously realized. The method proposed by the present invention is more suitable for the identification of satellite targets in complex space environments. Compared with the traditional deep learning target identification method, the method proposed by the present invention has improved both accuracy and stability.

Claims

1. A satellite target recognition method based on adaptive training of scattering feature extraction and fusion, characterized in that: The following steps are involved: Step 1: Obtain an inverse synthetic aperture radar image dataset of the original attitude of the satellite target from the satellite target echo received by the inverse synthetic aperture radar through two-dimensional inverse Fourier transform; Step 2: Using an inverse parameterized modeling algorithm, obtain an attribute scattering center model dataset of the original attitude of the satellite target from the satellite target echo received by the inverse synthetic aperture radar; Step 3: Using the extensibility of the attribute scattering center model, obtain the attribute scattering center model dataset after the satellite target attitude transformation, and determine the inverse synthetic aperture radar image dataset after the satellite target attitude transformation; Step 4: Establish a satellite target recognition network model based on the convolutional network and the graph convolutional network. Then, use the adaptive training method to train the satellite target recognition network model based on the inverse synthetic aperture radar image dataset and the attribute scattering center model dataset of the original satellite target posture, as well as the inverse synthetic aperture radar image dataset and the attribute scattering center model dataset after the satellite target posture transformation. Step 5: Using the open set recognition algorithm, the trained satellite target recognition network model outputs a probability distribution vector, and the probability distribution vector is used to identify the types of satellite targets of interest and those of non-satellite targets of interest. The reverse parameterized modeling algorithm described in step 2 is as follows: Step 2.1, obtaining a one-dimensional range image and a one-dimensional azimuth image of the satellite target from the satellite target echo through a one-dimensional inverse Fourier transform; Step 2.2: Using a peak detection method, extract the distance dimension distribution and azimuth dimension distribution of the attribute scattering center from the one-dimensional range image and the one-dimensional azimuth image of the satellite target, respectively. Based on the distance dimension distribution and azimuth dimension distribution of the attribute scattering center, obtain the distance dimension position and azimuth dimension position of the scattering center, respectively, thereby determining the two-dimensional position of the scattering center; Step 2.3: Based on the two-dimensional position of the extracted scattering center, locate the position of the attribute scattering center in the inverse synthetic aperture radar image, and obtain the length of the scattering center through the azimuth dimension distribution length of the attribute scattering center in the inverse synthetic aperture radar image; Step 2.4: Obtain the angle dependence factor of the scattering center based on the length of the extracted scattering center: If the length of the scattering center is 0, then the scattering center is a local scattering center, and the angle dependence factor of the scattering center is If the length of the scattering center is greater than 0, then the scattering center is a distributed scattering center and the angular dependence factor of the scattering center is 0. Step 2.5: Obtain the amplitude of the scattering center and the frequency dependence factor of the scattering center by the least squares method. The expression of the optimization objective function of the least squares method is: (2) in is the optimization objective function, The amplitude of the scattering center , No. Frequency dependence factor of the scattering center is the optimized parameter; The radar frequency is , the radar azimuth is The total scattered field when Indicates the radar frequency is , the radar azimuth is When the amplitude is , the frequency dependence factor is No. The scattered field of a scattering center; It represents optimization to the minimum value. represents the two-norm; The parameters of the satellite target's attribute scattering center model are extracted through the inverse parameterized modeling algorithm to construct the attribute scattering center model of the satellite target's original attitude. The attribute scattering center models of multiple satellite target's original attitudes constitute the attribute scattering center model data set of the satellite target's original attitude. In step 3, the extensibility of the attribute scattering center model is utilized to obtain the attribute scattering center model dataset after the satellite target attitude transformation, and to determine the inverse synthetic aperture radar image dataset after the satellite target attitude transformation, as follows: Step 3.1: As the satellite target's attitude changes, the two-dimensional position of the satellite target's attribute scattering center also changes. The attitude change of the satellite target is described by the rotation matrix, and the expression of the rotation matrix is: (3) in is the rotation matrix, is the rotation angle; After the satellite target attitude changes The two-dimensional position of the scattering center is ,but , is the first The distance dimension position of the scattering center, is the first The azimuthal position of the scattering center; The satellite target attitude changes after The length of a scattering center is ,but ; If the satellite target attitude changes after The length of the scattering center , then the first angular dependence factor of the scattering center ; If the first The length of the scattering center , then the first angular dependence factor of the scattering center ; The amplitude of the scattering center and the frequency dependence factor of the scattering center do not change before and after the satellite target attitude changes; By processing the parameters of the attribute scattering center model in the attribute scattering center model dataset of the original attitude of the satellite target, the attribute scattering center model dataset after the attitude transformation of the satellite target is obtained; Step 3.2: Determine the inverse synthetic aperture radar image after the satellite target attitude transformation through the attribute scattering center model after the satellite target attitude transformation. The expression is: (4) in is the inverse synthetic aperture radar image, is the distance dimension, is the azimuth dimension, is the radar bandwidth, is the wavelength of the radar carrier frequency, is the electromagnetic wave propagation velocity, is the radar sweep angle width; is a bandpass filter function, if ,but ;like or ,but ; Based on the attribute scattering center model dataset after satellite target attitude transformation, an inverse synthetic aperture radar image dataset after satellite target attitude transformation is constructed.

2. The method for adaptively training satellite target recognition based on scattering feature extraction and fusion according to claim 1 is characterized in that: The inverse synthetic aperture radar image dataset described in step 1, wherein each inverse synthetic aperture radar image has a size of 64×64, and the inverse synthetic aperture radar images of the original postures of multiple satellite targets constitute the inverse synthetic aperture radar image dataset of the original postures of satellite targets.

3. The method for adaptively training satellite target recognition based on scattering feature extraction and fusion according to claim 2 is characterized in that: In step 2, the attribute scattering center model obtained from the satellite target echo received by the inverse synthetic aperture radar is a two-dimensional attribute scattering center model, which is expressed as: (1) in is the total scattered field, is the radar frequency, is the radar azimuth; is the number of scattering centers, It is The scattered field of a scattering center, The radar frequency is , the radar azimuth is Time The scattered field of a scattering center; is the radar carrier frequency, is an imaginary number, is the electromagnetic wave propagation velocity, It is The amplitude of the scattering center, It is The frequency dependence factor of the scattering center, It is The length of the scattering center, It is The angular dependence factor of the scattering center, It is The distance dimension position of the scattering center, It is The azimuthal position of the scattering center, and Composition The two-dimensional position of the scattering center .

4. The method for adaptively training satellite target recognition based on scattering feature extraction and fusion according to claim 3 is characterized in that: The parameters of the attribute scattering center model in step 2 include the amplitude of the scattering center, the frequency dependence factor of the scattering center, the length of the scattering center, the angle dependence factor of the scattering center, and the two-dimensional position of the scattering center.

5. The method for adaptively training satellite target recognition based on scattering feature extraction and fusion according to claim 4 is characterized in that: In step 4, a satellite target recognition network model is established based on the convolutional network and the graph convolutional network. The satellite target recognition network model includes a convolution module, a graph convolution module, and an recognition module, as follows: The convolution module is composed of a convolutional network, including a convolution layer, a maximum pooling layer, an average pooling layer, and a residual structure, wherein the convolution kernel size of the convolution layer includes 、 and , the kernel size of the maximum pooling layer is , the kernel size of the average pooling layer is The residual structure consists of 4 convolutional layers, 4 batch normalization layers, and activation function ReLU; the convolution module inputs the inverse synthetic aperture radar image and outputs Deep learning features of size; The graph convolution module consists of four graph convolution layers. The graph convolution module takes each scattering center as a node of the graph convolution layer, and uses the amplitude, frequency dependence factor, length, angle dependence factor, and two-dimensional position of the scattering center as node features. The node feature dimension is 6. The correlation between nodes is related to the node features. The correlation between nodes is the correlation between scattering centers, which is expressed as: (5) in For the The scattering center and The correlation degree of the scattering centers, 、 The values ​​of ; and They are The amplitude of the scattering center and the The amplitude of the scattering center, and They are The two-dimensional position of the scattering center and the The two-dimensional position of the scattering center, is the standard deviation of the Gaussian distribution, is the distance threshold; the graph convolution module inputs the attribute scattering center model and outputs Scattering characteristics of size; The recognition module consists of one fully connected layer and one Dropout layer. The recognition module inputs and connects the deep learning features output by the fusion convolution module and the scattering features output by the graph convolution module, and outputs the activation vector for identifying satellite targets. Finally, through the Softmax function, the probability distribution vector of the satellite target of interest is output. , , For the The probability of satellite targets of interest, is the number of types of satellite targets of interest.

6. The method for adaptively training satellite target recognition based on scattering feature extraction and fusion according to claim 5 is characterized in that: In step 4, based on the inverse synthetic aperture radar image dataset and attribute scattering center model dataset of the original satellite target posture, as well as the inverse synthetic aperture radar image dataset and attribute scattering center model dataset after the satellite target posture is transformed, an adaptive training method is used to train the satellite target recognition network model, as follows: Step 4.1, adding the inverse synthetic aperture radar image dataset of the original attitude of the satellite target and the attribute scattering center model dataset of the original attitude of the satellite target to the training set, and pre-training the satellite target recognition network model through the training set; Step 4.2: Calculate the recognition error for each satellite target type of interest in the training set, determine the satellite target type with the largest recognition error, and supplement the training set with the attitude-transformed inverse synthetic aperture radar image and attitude-transformed attribute scattering center model of that satellite target type. Continue training the satellite target recognition network model using this supplemented training set. Step 4.3: If the number of network training rounds does not reach the threshold, repeat step 4.2; otherwise, end the training of the satellite target recognition network model; The cross entropy loss function is used to train the satellite target recognition network model in steps 4.1 and 4.2, and the expression is: (6) in is the loss function, is the number of samples in the training set, For the The labels of samples, For the The predicted probability of a sample.

7. The method for adaptively training satellite target recognition based on scattering feature extraction and fusion according to claim 6, characterized in that: In step 5, the open set recognition algorithm is used to output a probability distribution vector through the trained satellite target recognition network model, and the probability distribution vector is used to realize the identification of the type of satellite target of interest and the discrimination of the type of satellite target of non-interest, as follows: The satellite target types of interest are the satellite target types that exist in the inverse synthetic aperture radar image dataset and the attribute scattering center model dataset. In the actual satellite target identification scenario, there are non-concerned satellite target types, and the samples of the non-concerned satellite target types do not exist in the established inverse synthetic aperture radar image dataset and the attribute scattering center model dataset. The satellite target types of interest are Class, then the dimension of the probability distribution vector of the satellite target of interest output by the satellite target recognition network model is ; Using the open set recognition algorithm, the trained satellite target recognition network model outputs the recognition results of the satellite target types of interest and the discrimination results of the satellite target types of non-interest, as follows: Step 5.1: Obtain the activation vector of each sample in the training set through the trained satellite target recognition network model, and calculate the average activation vector of each type of satellite target in the training set; Step 5.2: Calculate the Euler distance between the activation vector of each sample in the training set and the average activation vector of the satellite target category corresponding to each sample, and obtain the maximum Euler distance for each category of satellite targets of interest; Step 5.3: Based on the average probability distribution vector and the maximum Euler distance of each type of satellite target in the training set, the Weibull probability distribution of each type of satellite target is calculated by fitting; Step 5.4: Obtain the activation vector of the test sample through the trained satellite target recognition network model, and calculate the Euler distance between the activation vector of the test sample and the average activation vector of each type of satellite target. Obtain the Weibull probability distribution vector of the test sample through the Weibull probability distribution of each type of satellite target. , The test sample belongs to Weibull probability of satellite targets of interest; Step 5.5: Use Weibull probability distribution vector to process probability distribution vector , obtain the updated probability distribution vector of the test sample , After the update The probability of satellite targets of interest, The probability of being a satellite target of no concern; when hour, ; ;right , , , and Normalize and get the updated probability distribution vector of the test sample ; Step 5.6: If the updated probability distribution vector of the test sample is The maximum value in ,and , then the test sample belongs to The satellite target of the class; if the updated probability distribution vector of the test sample The maximum value in , then the test sample belongs to a non-concerned satellite target; thus, the recognition result of the concerned satellite target type and the discrimination result of the non-concerned satellite target type are obtained, and the recognition and discrimination results of the test sample are output.

8. An adaptive training satellite target recognition system based on scattering feature extraction and fusion, characterized in that: The system is used to implement the adaptive training satellite target recognition method based on scattering feature extraction and fusion as described in any one of claims 1 to 7. The system includes the first to fifth units, and the functions of each unit are as follows: The first unit uses two-dimensional inverse Fourier transform to obtain the inverse synthetic aperture radar image dataset of the original attitude of the satellite target from the satellite target echo received by the inverse synthetic aperture radar; The second unit uses the inverse parameterized modeling algorithm to obtain the attribute scattering center model dataset of the satellite target's original attitude from the satellite target echo received by the inverse synthetic aperture radar; Unit 3: Using the extensibility of the attribute scattering center model, we obtain the attribute scattering center model dataset after the satellite target attitude transformation, and determine the inverse synthetic aperture radar image dataset after the satellite target attitude transformation; Unit 4: Building a satellite target recognition network model based on convolutional networks and graph convolutional networks. Then, using an adaptive training method to train the satellite target recognition network model, the model is trained based on an inverse synthetic aperture radar image dataset and an attribute scattering center model dataset of the original satellite target pose, as well as an inverse synthetic aperture radar image dataset and an attribute scattering center model dataset after the satellite target pose is transformed. In the fifth unit, the open set recognition algorithm is used to output the probability distribution vector through the trained satellite target recognition network model, and the probability distribution vector is used to realize the identification of the types of satellite targets of interest and the discrimination of the types of satellite targets of non-interest.

Images