Semi-supervised recognition method based on frequency agility scattering characteristics and complex-valued autoencoder

Abstract

The invention discloses a semi-supervised recognition method based on frequency agile scattering characteristics and a complex-valued autoencoder, relates to the technical field of radar signal processing, and solves the problem in the prior art that it relies on a large number of labeled sample training and has insufficient performance under low signal-to-noise ratio in actual application scenarios; the method comprises: receiving scattered echoes, and obtaining polarization scattering matrices at various frequency points of the frequency agile waveform according to the scattered echoes; performing polarization rotation processing on the polarization scattering matrices at various frequency points of the frequency agile waveform to obtain frequency agile polarization rotated scattering coefficient tensors; and splicing the frequency agile polarization rotated scattering coefficient tensors to obtain splicing vectors; inputting the splicing vectors into a trained complex-valued autoencoder network to obtain recognition results and denoised splicing vectors corresponding to the scattered echoes; the method realizes intelligent, rapid and accurate recognition of passive interference and targets under low signal-to-noise ratio and small sample scenarios.

Description

Technical Field

[0001] The present invention relates to the field of radar signal processing technology, and in particular to a semi-supervised recognition method based on frequency agility scattering characteristics and a complex-valued autoencoder. Background Art

[0002] Corner reflectors, chaff strips, towed decoys and other false targets are common high-fidelity, low-cost passive interference in radar target detection and recognition scenarios. They have similar appearance, structure and electromagnetic scattering characteristics to real targets such as aircraft and ships, posing a severe challenge to radar recognition. Existing radar target recognition methods rely on information sources such as high-resolution range profiles (HRRPs), synthetic aperture radar (SAR) images, micro-Doppler characteristics, polarization characteristics, and radar cross-section (RCS). However, the first three methods based on one-dimensional and two-dimensional image features and motion features require large system bandwidth and computing resources, and are difficult to apply to scenarios with limited resources or strong real-time requirements. In contrast, RCS and polarization scattering coefficient characteristics reflect the detailed essential characteristics of target structure, material, etc., and do not require high transmission power, resolution (bandwidth) or computing resources. Therefore, researchers have developed target and passive interference recognition methods based on radar scattering data feature extraction and classifier design, and have made certain application progress. In recent years, deep learning with superior performance has been widely used in the field of radar target recognition. In the field of target / passive interference identification, most existing methods use long-term, multi-angle RCS sequences as raw information input into deep learning models to achieve target and passive interference identification.

[0003] However, this type of method has two main defects: it uses long-term, multi-angle RCS sequences as model input, and suffers from problems of insufficient real-time performance and insufficient generalization in actual application scenarios. Summary of the invention

[0004] The present invention solves the problems of insufficient real-time performance and insufficient generalization in practical application scenarios in the prior art by providing a semi-supervised recognition method based on frequency-agile scattering characteristics and a complex-valued autoencoder, and realizes intelligent, rapid and accurate recognition of passive interference and targets in low signal-to-noise ratio and small sample scenarios.

[0005] The present invention provides a semi-supervised recognition method based on frequency agility scattering characteristics and a complex-valued autoencoder, the method comprising:

[0006] receiving scattered echoes, and obtaining a polarization scattering matrix at each frequency point of the frequency agile waveform according to the scattered echoes;

[0007] The polarization scattering matrix at each frequency point of the frequency agile waveform is subjected to polarization rotation processing to obtain a frequency agile polarization rotation scattering coefficient tensor; and the frequency agile polarization rotation scattering coefficient tensor is spliced ​​to obtain a splicing vector; the splicing vector is input into the trained complex-valued autoencoder network to obtain the recognition result and denoised splicing vector corresponding to the scattered echo; wherein the complex-valued autoencoder network includes: an encoder for hierarchically extracting features from the splicing vector to obtain a first three-dimensional feature vector, a second three-dimensional feature vector and a third three-dimensional feature vector respectively; a first efficient hybrid attention module performs global average pooling and adaptive convolution kernel processing of different dimensions on the first three-dimensional feature vector to obtain a first weight feature tensor set; the first weight feature tensor set is multiplied by the first three-dimensional feature vector to obtain a first feature tensor; a second efficient hybrid attention module The second three-dimensional feature vector is subjected to global average pooling and adaptive convolution kernel processing of different dimensions to obtain a second weight feature tensor set; the second weight feature tensor set is multiplied by the second three-dimensional feature vector to obtain a second feature tensor; a first complex fully connected module is used to perform dimension transformation on the third three-dimensional feature vector to obtain a first-dimensional feature vector; a second complex fully connected module is used to restore the dimension of the first-dimensional feature vector to obtain a second-dimensional feature vector; a classifier is used to perform multiple dimensionality reduction according to the first-dimensional feature vector to obtain a three-dimensional feature vector, and determine the recognition result corresponding to the scattered echo according to the three-dimensional feature vector; a decoder is used to perform calculations according to the second-dimensional feature vector, the first feature tensor and the second feature tensor to obtain a denoised spliced ​​vector corresponding to the agile frequency polarization rotation scattering coefficient tensor.

[0008] One or more technical solutions provided in the present invention have at least the following technical effects or advantages:

[0009] The semi-supervised recognition method based on agile frequency scattering characteristics and a complex-valued autoencoder proposed in the present invention has a smaller amount of computation in the encoder than in the traditional convolutional layer; the denoising matrix is ​​restored in the decoder; and the efficient hybrid attention module can quickly generate attention weights for different channels; unsupervised target scattering matrix denoising in the complex domain and semi-supervised recognition of real targets and passive interference are realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 A flowchart of the steps of a semi-supervised recognition method based on frequency agility scattering characteristics and a complex-valued autoencoder provided in an embodiment of the present invention;

[0011] Figure 2 A schematic diagram of the structure of a complex-valued autoencoding network provided by an embodiment of the present invention;

[0012] Figure 3A schematic diagram of a method for implementing an efficient hybrid attention mechanism provided by an embodiment of the present invention;

[0013] Figure 4 A schematic diagram of the overall structure of a complex-valued autoencoding network in a specific use embodiment provided by the present invention;

[0014] Figure 5 A schematic diagram of comparison results of agile frequency polarization rotational scattering coefficient tensors (AFPRSCs) before and after denoising provided by an embodiment of the present invention;

[0015] Figure 6 Schematic diagram of the results of three comparison algorithms when SNR=-25~10dB provided in an embodiment of the present invention;

[0016] Figure 7 Schematic diagram of the results of three comparison algorithms when the proportion of the supervised classification training set provided for the embodiment of the present invention is 70%. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] The present invention provides a semi-supervised recognition method based on frequency agility scattering characteristics and complex-valued autoencoders, such as Figure 1 As shown, the method includes the following steps S101 to S102.

[0019] S101, receiving scattered echoes, and obtaining a polarization scattering matrix at each frequency point of the frequency agile waveform according to the scattered echoes; specifically, the polarization scattering matrix is ​​expressed as:

[0020] ;

[0021] in, Indicates The scattering coefficient of the HH polarization component corresponding to the frequency point; Indicates The scattering coefficient of the HV polarization component corresponding to each frequency point; Indicates The scattering coefficient of the VH polarization component corresponding to the frequency point; Indicates The scattering coefficient of the VV polarization component corresponding to the frequency point; Indicates The polarization scattering matrix corresponding to the frequency point.

[0022] Exemplarily, taking frequency point 1 as an example, the polarization scattering matrix of frequency point 1 is expressed as: ,in, Indicates the scattering coefficient of the HH polarization component corresponding to frequency 1; Indicates the scattering coefficient of the HV polarization component corresponding to frequency 1; Indicates the scattering coefficient of the VH polarization component corresponding to frequency 1; Indicates the scattering coefficient of the VV polarization component corresponding to frequency 1; Represents the polarization scattering matrix corresponding to frequency 1.

[0023] S102, performing polarization rotation processing on the polarization scattering matrix at each frequency point of the frequency agile waveform to obtain a frequency agile polarization rotation scattering coefficient tensor; and splicing the frequency agile polarization rotation scattering coefficient tensor to obtain a splicing vector; inputting the splicing vector into the trained complex-valued autoencoder network to obtain the recognition result corresponding to the scattered echo and the denoised splicing vector.

[0024] Here, the frequency-agile polarization rotation scattering coefficient tensor is expressed as:

[0025] ;

[0026] in, The rotation angle is No. Polarization rotation scattering coefficient of the HH polarization component corresponding to the frequency point; The rotation angle is No. Polarization rotation scattering coefficient of HV polarization component corresponding to each frequency point; The rotation angle is No. Polarization rotation scattering coefficient of the VH polarization component corresponding to the frequency point; The rotation angle is No. Polarization rotation scattering coefficient of the VV polarization component corresponding to the frequency point; The rotation angle is No. The frequency-agile polarization rotation scattering coefficient tensor corresponding to each frequency point; The rotation angle is The rotation matrix of The rotation angle is The transpose of the rotation matrix; Indicates The polarization scattering matrix corresponding to the frequency point.

[0027] The polarization scattering matrix of frequency point 1 is subjected to polarization rotation processing to obtain the frequency agile polarization rotation scattering coefficient matrix of frequency point 1, which is specifically expressed as:

[0028] ;

[0029] in, The rotation angle is The polarization rotation scattering coefficient of the HH polarization component corresponding to the frequency point 1; The rotation angle is The polarization rotation scattering coefficient of the HV polarization component corresponding to the frequency point 1; The rotation angle is The polarization rotation scattering coefficient of the VH polarization component corresponding to the frequency point 1; The rotation angle is The polarization rotation scattering coefficient of the VV polarization component corresponding to the frequency point 1; The rotation angle is The frequency-agile polarization rotation scattering coefficient tensor corresponding to the frequency point 1; The rotation angle is The rotation matrix of The rotation angle is The transpose of the rotation matrix; represents the polarization scattering matrix corresponding to frequency point 1. Here, .

[0030] The frequency-agile polarization rotation scattering coefficient tensor is: , each polarization scattering matrix component of the frequency-agile polarization rotation scattering coefficient tensor is expressed in the same way. Among them: HH polarization scattering matrix component It is expressed as:

[0031] ;

[0032] in, N Indicates the number of frequency points.

[0033] Complex-valued autoencoder networks include: Figure 2 As shown, an encoder, a first efficient hybrid attention module, EHA (Efficient hybrid attention) 1, a second efficient hybrid attention module EHA2, a first complex fully connected module, a second complex fully connected module, a classifier and a decoder.

[0034] An encoder, used for performing feature extraction on the concatenated vector in layers to obtain a first three-dimensional feature vector, a second three-dimensional feature vector and a third three-dimensional feature vector respectively;

[0035] Specifically, the encoder includes a first complex-valued depthwise separable convolutional layer, a second complex-valued depthwise separable convolutional layer, and a third complex-valued depthwise separable convolutional layer connected in sequence;

[0036] The first complex-valued depthwise separable convolutional layer comprises a first depthwise convolutional unit and a first pointwise convolutional unit connected in sequence;

[0037] Here, the convolution kernel sizes of the first depth convolution unit and the first point convolution unit are 3 and 5 respectively. The size of the feature vector output by the first depth convolution unit is , the size of the feature vector output by the first point convolution unit is .

[0038] The input of the first complex-valued depthwise separable convolutional layer is: the concatenated vector; the output is: the first three-dimensional feature vector;

[0039] The second complex-valued depthwise separable convolutional layer comprises a second depthwise convolutional unit and a second pointwise convolutional unit connected in sequence;

[0040] Here, the convolution kernel sizes of the second depth convolution unit and the second point convolution unit are 3 and 5 respectively. The size of the feature vector output by the first depth convolution unit is , the size of the feature vector output by the first point convolution unit is .

[0041] The input of the second complex-valued depthwise separable convolutional layer is: the first complex tensor, and the output is: the second three-dimensional feature vector; wherein the first complex tensor is calculated by the complex-valued convolution calculation formula according to the first three-dimensional feature vector and the number of convolution kernels in the first complex-valued depthwise separable convolutional layer;

[0042] The third complex-valued depthwise separable convolutional layer includes a third depthwise convolutional unit and a third pointwise convolutional unit connected in sequence;

[0043] Here, the convolution kernel sizes of the third depth convolution unit and the second point convolution unit are 3 and 5 respectively. The size of the feature vector output by the third point convolution unit is , the size of the feature vector output by the first point convolution unit is .

[0044] The input of the third complex-valued depthwise separable convolutional layer is: the second complex tensor, and the output is: a third three-dimensional feature vector; wherein the second complex tensor is calculated by the complex-valued convolution calculation formula according to the second three-dimensional feature vector and the number of convolution kernels in the second complex-valued depthwise separable convolutional layer.

[0045] Here, the complex-valued convolution calculation formula is expressed as:

[0046] ;

[0047] in, represents the convolution kernel complex-valued weight matrix, ; Represents the real part of the convolution kernel complex-valued weight matrix; Represents the imaginary part of the convolution kernel complex-valued weight matrix; Represents the input complex tensor; Represents the imaginary part of the input complex tensor; Represents the real part of the input complex tensor; Represents an imaginary unit.

[0048] Exemplarily, the frequency-agile polarization rotation scattering coefficient tensor The polarization scattering matrix components in are used as the complex-valued convolution calculation formula (input complex tensor), calculate according to the complex-valued convolution calculation formula, obtain the first calculated value, the second calculated value, the third calculated value and the fourth calculated value corresponding to each polarization scattering matrix component, splice the first calculated value, the second calculated value, the third calculated value and the fourth calculated value to obtain a spliced ​​vector; input the spliced ​​vector into the first complex-valued depthwise separable convolution layer for depthwise separable convolution and maximum pooling to obtain a first three-dimensional feature vector;

[0049] Preferably, in a specific embodiment provided by the present invention, an activation function layer is added between the first complex-valued depth-separable convolution layer, the second complex-valued depth-separable convolution layer and the third complex-valued depth-separable convolution layer, and the activation function used in the activation function layer is: CReLU function.

[0050] Frequency-agile polarization rotation scattering coefficient tensor The concatenated vector is calculated in the first complex-valued depthwise separable convolutional layer. , the complex number of convolution kernels of the first complex-valued depthwise separable convolutional layer is expressed as: , As the actual input value of the first complex-valued depthwise separable convolutional layer, the output value of the first complex-valued depthwise separable convolutional layer is: the first three-dimensional feature vector ; The first three-dimensional feature vector According to the CReLU function, the real part activation and imaginary part activation are performed according to the activation function, and the result of complex domain activation is obtained. ;

[0051] The complex number expression of the number of convolution kernels of the second complex-valued depthwise separable convolutional layer is: , the first complex tensor is represented as: , the output is the second three-dimensional feature vector ; The first three-dimensional feature vector According to the CReLU function, the real part activation and imaginary part activation are performed according to the activation function, and the result of complex domain activation is obtained. ;

[0052] The complex number of convolution kernels of the third complex-valued depthwise separable convolutional layer is expressed as: , the second complex tensor is expressed as: , the output is the third three-dimensional feature vector ; The third three-dimensional feature vector According to the CReLU function, the real part activation and imaginary part activation are performed according to the activation function, and the result of complex domain activation is obtained. .

[0053] The activation function of the activation function layer is expressed as:

[0054] ;

[0055] in, Represents the ReLU function; A complex tensor representing the input; represents the real part of the input complex tensor; Represents the imaginary part of the input complex tensor; Represents an imaginary unit.

[0056] The first efficient hybrid attention module EHA1 performs global average pooling and adaptive convolution kernel processing of different dimensions on the first three-dimensional feature vector to obtain a first weight feature tensor set; multiplies the first weight feature tensor set and the first three-dimensional feature vector to obtain a first feature tensor;

[0057] Specifically, the first efficient hybrid attention module performs global average pooling and adaptive convolution kernel processing of different dimensions on the first three-dimensional feature vector to obtain a first weighted feature tensor set, including:

[0058] (1) performing global average pooling on the first three-dimensional feature vector along the channel dimension, height dimension, and width dimension respectively, to obtain a channel dimension global average pooling result, a height dimension global average pooling result, and a width dimension global average pooling result;

[0059] (2) Using a one-dimensional convolution kernel with adaptive convolution kernel size, convolution processing is performed on the global average pooling results of the channel dimension, the global average pooling results of the height dimension, and the global average pooling results of the width dimension, respectively, to obtain a first set of weighted feature tensors; wherein the first set of weighted feature tensors includes: a channel dimension weighted feature tensor, a height dimension weighted feature tensor, and a width dimension feature vector.

[0060] Specifically, the first weight feature tensor set is multiplied by the first three-dimensional feature vector to obtain a first feature tensor, including:

[0061] (1) multiplying each weight feature tensor in the first weight feature tensor set respectively to obtain a first weight value; wherein the first weight feature tensor set includes three groups of weight feature tensors;

[0062] (2) Multiply the first weight value by the first three-dimensional feature vector to obtain a first feature tensor.

[0063] The second efficient hybrid attention module EHA2 performs global average pooling and adaptive convolution kernel processing of different dimensions on the second three-dimensional feature vector to obtain a second weight feature tensor set; multiplying the second weight feature tensor set and the second three-dimensional feature vector to obtain a second feature tensor;

[0064] Specifically, the second weight feature tensor set is multiplied by the second three-dimensional feature vector to obtain a second feature tensor, including:

[0065] Multiplying each weight feature tensor in the second weight feature tensor set respectively to obtain a second weight value; wherein the second weight feature tensor set includes three groups of weight feature tensors;

[0066] The second weight value is multiplied by the second three-dimensional eigenvector to obtain a second eigentensor.

[0067] An example of an efficient hybrid attention mechanism implementation is Figure 3 As shown: Combined with the first efficient hybrid attention module EHA1 Figure 3 to explain.

[0068] First, the first three-dimensional feature vector is transposed along the channel dimension, height dimension and width dimension respectively, and global average pooling is performed on the transposed first three-dimensional feature vector to obtain the channel dimension global average pooling result, the height dimension global average pooling result and the width dimension global average pooling result;

[0069] Here, the channel dimension global average pooling result is expressed as: ; The global average pooling result in the height dimension is expressed as: ; The global average pooling result in the width dimension is expressed as: ;in, Represents the feature tensor The spatial coordinates in the channel dimension are ; Represents the feature tensor The spatial coordinates in the height dimension are ; Represents the feature tensor The spatial coordinates in the width dimension are ; Indicates the number of channel dimensions; Represents the height dimension; represents the width dimension; Represents the feature tensor The global average pooling result in the channel dimension; Represents the feature tensor The global average pooling result in the height dimension; Represents the feature tensor Global average pooling result in the width dimension.

[0070] Then, a one-dimensional convolution kernel with adaptive convolution kernel size is used to perform convolution processing on the global average pooling results of the channel dimension, the global average pooling results of the height dimension, and the global average pooling results of the width dimension, respectively, to obtain the channel dimension weight feature tensor, the height dimension weight feature tensor, and the width dimension feature vector;

[0071] Here, the channel dimension weight feature tensor, height dimension weight feature tensor, and width dimension feature vector are expressed as:

[0072] ;

[0073] in, Represents the convolution operation; Represents the channel dimension weight feature tensor; Represents the height-dimensional weight feature tensor; represents the width dimension feature vector; Represents the Sigmoid activation function.

[0074] Finally, the channel dimension weight feature tensor, height dimension weight feature tensor and width dimension feature vector are dimensionally permuted respectively, and the channel dimension weight feature tensor, height dimension weight feature tensor and width dimension feature vector are multiplied to obtain the first weight value, and the first weight value is multiplied with the first three-dimensional feature vector to obtain the first feature tensor. The first eigentensor It is expressed as: ,in, represents the first three-dimensional eigenvector.

[0075] A first complex fully connected module, used for performing dimension transformation on the third three-dimensional feature vector to obtain a first-dimensional feature vector;

[0076] Here, the first complex fully connected module includes three fully connected layers connected in sequence, and the sizes of the output features of the three fully connected layers are respectively , and .

[0077] A second complex fully connected module is used to restore the dimension of the first dimensional feature vector to obtain a second dimensional feature vector;

[0078] Here, the second complex fully connected module includes three fully connected layers connected in sequence, and the sizes of the output features of the three fully connected layers are respectively , and .

[0079] A classifier is used to perform multiple dimensionality reduction according to the first dimension feature vector to obtain a three-dimensional feature vector, and determine the recognition result corresponding to the scattered echo according to the three-dimensional feature vector;

[0080] Here, the classifier includes three fully connected layers connected in sequence, and the sizes of the output features of the three fully connected layers are , and .

[0081] The decoder is used to calculate according to the second dimensional feature vector, the first feature tensor and the second feature tensor to obtain a denoised concatenated vector corresponding to the frequency-agile polarization rotation scattering coefficient tensor. The decoder includes: a first decoding unit, a second decoding unit and a third decoding unit; each decoding unit is composed of a deconvolution layer.

[0082] Here, the convolution kernel of the first decoding unit is , the size of the output feature ;

[0083] The convolution kernel of the second decoding unit is , the size of the output feature ;

[0084] The convolution kernel of the third decoding unit is , the size of the output feature .

[0085] The loss function of the complex-valued autoencoder network is expressed as:

[0086] ;

[0087] in, represents the concatenation vector; represents the real part of the concatenated vector; represents the imaginary part of the concatenated vector; represents the denoised concatenated vector; represents the real part of the denoised concatenated vector; Represents the imaginary part of the denoised concatenated vector.

[0088] In a specific use embodiment provided by the present invention, Figure 4As shown, a complex ReLU activation function is included after each convolution layer; a batch normalization layer (BN, Batch Normalization) is added after the second complex-valued depthwise separable convolution layer; a flattening layer (Flatten) is included after the third complex-valued depthwise separable convolution layer, and the flattening layer flattens the third three-dimensional feature vector into a one-dimensional feature vector; before the decoder, a recovery layer (unFlatten) is provided, and the recovery layer restores the one-dimensional feature vector to a three-dimensional feature vector; and a batch normalization layer is also provided after the first decoding unit.

[0089] In a simulation embodiment provided by the present invention, a 1:1 scale 3D model of an Arleigh Burke class destroyer, an icosahedral corner reflector and its array was constructed using SolidWorks to verify the method for identifying targets and passive interference. After adding the sea surface model, these models were imported into the CST electromagnetic simulation software to obtain the scattering matrix.

[0090] In the simulation, the frequency range is set to 12~13 GHz, the step size is 20 MHz, the elevation angle is 0°~90°, and the azimuth angle is 0°~180°, both with a step size of 1°. Therefore, for each target, a scattering matrix of 16,290 angles (90×181) at 51 frequency points is obtained.

[0091] by , , , For the research object, each target has 16290 × 4 = 65160 complex sample sequences. In order to verify the superiority of the proposed denoising and separation methods, the support vector machine (SVM) method based on polarization invariant, OS-CNN-FF based on frequency agile RCS, and ResNet18 using frequency agile polarization rotation matrix are used as comparison algorithms. Based on the settings of the above comparison algorithms, label-level experimental results, experiments under various SNRs and training set sizes are given. The datasets for training, validation, and testing are split into 0.4:0.1:0.5. All models are built using Python 3.9 and Torch 2.2. In Table 1, the first column represents the target; the second column represents the recognition result of the comparison algorithm SVM; the third column represents the recognition result of the OS-CNN-FF algorithm; the fourth column represents the recognition result of the ResNet18 algorithm; the fifth column RVNN-AE-A represents the recognition result of the model which has exactly the same structure as the method of the present invention but only uses real number network operations; the sixth column CVNN-AE represents the recognition result of the complex-valued neural network based on the autoencoder architecture without the attention mechanism; the eighth column CVNN-AE-A represents the model designed by the method of the present invention.

[0092] 1) Multi-algorithm comparison and ablation experiment (SNR=0 dB):

[0093] Table 1 Comparison of recognition results

[0094]

[0095] It can be seen that compared with the first three comparison algorithms with SNR=0dB, CVNN-AE-A achieves improvements of 26.88%~21.90%, 13.95%~18.92%, 17.58%~24.19% and 17.68%~22.70% in overall accuracy, average precision, average recall and F1, respectively.

[0096] In the last three columns of the ablation experiment, the results show that CVNN-AE-A outperforms the real number processing network RVNN-AE-A with the same structure, with improvements of about 8.37% to 9.09% in overall accuracy, average precision, average recall, and F1. This proves the effectiveness and performance contribution of introducing complex network processing. Compared with CVNN-AE, the floating-point operation volume FLOPs of the designed EHA of CVNN-AE-A has almost no change, while the performance indicators such as overall accuracy, average precision, average recall, and F1 have increased by 0.77% to 1.14%.

[0097] Figure 5 The results of the three denoising algorithms for AFPRSCs (SNR = 0 dB) are intuitively displayed, indicating that the denoising effects of RVNN-AE-A, CVNN-AE, and CVNN-AE-A are enhanced in turn, which is also the fundamental reason for the difference in classification performance in Table 1.

[0098] Performance at different signal-to-noise ratios:

[0099] 2) This method uses data with SNR=0dB to train three denoising models, and uses data with SNR ranging from -25 dB to 10 dB to train the other three algorithms. Finally, the performance analysis was performed using the SNR=-25~10dB test data. The results are as follows: Figure 6 As shown in Figure 3, the performance of the three existing methods without denoising degrades severely when the SNR decreases. In contrast, the accuracy of the three denoising algorithms decreases slightly when the SNR decreases from 10 dB to -10 dB. It is worth noting that the proposed CVNN-AE-A still achieves an overall accuracy of 88.79% at SNR = -10 dB, which is 34.55% to 52.53% higher than the three existing methods. Figure 6 In the figure, SVM represents support vector machine; OS-CNN-FF represents the model of OS-CNN-FF based on frequency agility RCS; ResNet18 represents the ResNet18 algorithm.

[0100] 3) Performance of different training set sizes: e.g. Figure 7 As shown in the figure, when the ratio of supervised classification training set increases from 10% to 70%, the performance of the three compared algorithms improves significantly. On the other hand, the three denoising algorithms mainly use unlabeled data for unsupervised denoising learning and only require a small number of labeled samples for classification learning. Therefore, the change in the training set ratio has the least impact on their performance, and they show excellent accuracy in the case of small samples.

[0101] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. All or part of the present invention can be used in many general or special computer system environments or configurations. For example: personal computers, server computers, handheld devices or portable devices, tablet devices, mobile communication terminals, multi-processor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, distributed computing environments including any of the above systems or devices, etc.

[0102] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, a person skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some or all of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the present invention.

Claims

1. A semi-supervised recognition method based on frequency agility scattering characteristics and complex-valued autoencoder, characterized in that: include: receiving scattered echoes, and obtaining a polarization scattering matrix at each frequency point of the frequency agile waveform according to the scattered echoes; The polarization scattering matrix at each frequency point of the frequency agile waveform is subjected to polarization rotation processing to obtain a frequency agile polarization rotation scattering coefficient tensor; and the frequency agile polarization rotation scattering coefficient tensor is spliced ​​to obtain a splicing vector; the splicing vector is input into the trained complex-valued autoencoder network to obtain the recognition result and denoised splicing vector corresponding to the scattered echo; wherein the complex-valued autoencoder network comprises: an encoder for performing feature extraction on the splicing vector in layers to obtain a first three-dimensional feature vector, a second three-dimensional feature vector and a third three-dimensional feature vector respectively; a first efficient hybrid attention module for performing global average pooling and adaptive convolution kernel processing of different dimensions on the first three-dimensional feature vector to obtain a first weight feature tensor set, and multiplying the first weight feature tensor set and the first three-dimensional feature vector to obtain a first feature tensor; a second efficient hybrid attention module A block is used to perform global average pooling and adaptive convolution kernel processing of different dimensions on the second three-dimensional feature vector to obtain a second weight feature tensor set, and multiply the second weight feature tensor set and the second three-dimensional feature vector to obtain a second feature tensor; a first complex fully connected module is used to perform dimensionality transformation on the third three-dimensional feature vector to obtain a first-dimensional feature vector; a second complex fully connected module is used to restore the dimension of the first-dimensional feature vector to obtain a second-dimensional feature vector; a classifier is used to perform multiple dimensionality reduction on the first-dimensional feature vector to obtain a three-dimensional feature vector, and determine the recognition result corresponding to the scattered echo according to the three-dimensional feature vector; a decoder is used to calculate according to the second-dimensional feature vector, the first feature tensor and the second feature tensor to obtain a denoised spliced ​​vector corresponding to the agile frequency polarization rotation scattering coefficient tensor.

2. The semi-supervised recognition method based on frequency agility scattering characteristics and complex-valued autoencoder according to claim 1 is characterized in that: The polarization scattering matrix is ​​expressed as: ； in, Indicates The scattering coefficient of the HH polarization component corresponding to the frequency point; Indicates The scattering coefficient of the HV polarization component corresponding to each frequency point; Indicates The scattering coefficient of the VH polarization component corresponding to the frequency point; Indicates The scattering coefficient of the VV polarization component corresponding to the frequency point; Indicates The polarization scattering matrix corresponding to the frequency point is: The value is 1~ N , N Indicates the number of frequency points.

3. The semi-supervised recognition method based on frequency agility scattering characteristics and complex-valued autoencoder according to claim 1 is characterized in that: The frequency-agile polarization rotation scattering coefficient tensor The frequency-agile polarization rotation scattering coefficient matrix corresponding to each frequency point is expressed as: ； in, The rotation angle is No. Polarization rotation scattering coefficient of the HH polarization component corresponding to the frequency point; The rotation angle is No. Polarization rotation scattering coefficient of HV polarization component corresponding to each frequency point; The rotation angle is No. Polarization rotation scattering coefficient of the VH polarization component corresponding to the frequency point; The rotation angle is No. Polarization rotation scattering coefficient of the VV polarization component corresponding to the frequency point; The rotation angle is No. The frequency-agile polarization rotation scattering coefficient tensor corresponding to each frequency point; The rotation angle is The rotation matrix of Represents the rotation matrix The transpose of The rotation angle is No. The polarization scattering matrix corresponding to the frequency point is: The value is 1~ N , N Indicates the number of frequency points.

4. The semi-supervised recognition method based on frequency agility scattering characteristics and complex-valued autoencoder according to claim 1, characterized in that: The encoder comprises a first complex-valued depthwise separable convolutional layer, a second complex-valued depthwise separable convolutional layer and a third complex-valued depthwise separable convolutional layer connected in sequence; The first complex-valued depthwise separable convolutional layer comprises a first depthwise convolutional unit and a first pointwise convolutional unit connected in sequence; The input of the first complex-valued depthwise separable convolutional layer is: the concatenated vector; the output is: the first three-dimensional feature vector; The second complex-valued depthwise separable convolutional layer comprises a second depthwise convolutional unit and a second pointwise convolutional unit connected in sequence; The input of the second complex-valued depthwise separable convolutional layer is: the first complex tensor, and the output is: the second three-dimensional feature vector; wherein the first complex tensor is calculated by a complex-valued convolution calculation formula according to the first three-dimensional feature vector and the number of convolution kernels in the first complex-valued depthwise separable convolutional layer; The third complex-valued depthwise separable convolutional layer comprises a third depthwise convolutional unit and a third pointwise convolutional unit connected in sequence; The input of the third complex-valued depthwise separable convolutional layer is: the second complex tensor, and the output is: the third three-dimensional feature vector; wherein, the second complex tensor is calculated by the complex-valued convolution calculation formula according to the second three-dimensional feature vector and the number of convolution kernels in the second complex-valued depthwise separable convolutional layer.

5. The semi-supervised recognition method based on frequency agility scattering characteristics and complex-valued autoencoder according to claim 4 is characterized in that: An activation function layer is further provided between each of the first complex-valued depthwise separable convolution layer, the second complex-valued depthwise separable convolution layer, and the third complex-valued depthwise separable convolution layer; wherein the activation function of the activation function layer is expressed as: ； in, A complex tensor representing the input; Represents the imaginary part of the input complex tensor; represents the real part of the input complex tensor; represents an imaginary unit; Represents the activation function.

6. The semi-supervised recognition method based on frequency agility scattering characteristics and complex-valued autoencoder according to claim 4, characterized in that: The complex-valued convolution calculation formula is expressed as: ； in, Represents the convolution kernel complex-valued weight matrix; Represents the real part of the convolution kernel complex-valued weight matrix; Represents the imaginary part of the convolution kernel complex-valued weight matrix; A complex tensor representing the input; Represents the imaginary part of the input complex tensor; represents the real part of the input complex tensor; Represents an imaginary unit.

7. The semi-supervised recognition method based on frequency agility scattering characteristics and complex-valued autoencoder according to claim 1, characterized in that: In the first efficient hybrid attention module, global average pooling and adaptive convolution kernel processing of different dimensions are performed on the first three-dimensional feature vector to obtain a first weighted feature tensor set, including: Performing global average pooling on the first three-dimensional feature vector along the channel dimension, height dimension, and width dimension respectively to obtain a channel dimension global average pooling result, a height dimension global average pooling result, and a width dimension global average pooling result; A one-dimensional convolution kernel with adaptive convolution kernel size is used to perform convolution processing on the channel dimension global average pooling result, the height dimension global average pooling result and the width dimension global average pooling result, respectively, to obtain a first weight feature tensor set; wherein the first weight feature tensor set includes: a channel dimension weight feature tensor, a height dimension weight feature tensor and a width dimension feature vector.

8. The semi-supervised recognition method based on frequency agility scattering characteristics and complex-valued autoencoder according to claim 1, characterized in that: The step of multiplying the first weight feature tensor set and the first three-dimensional feature vector to obtain a first feature tensor includes: Multiplying each weight feature tensor in the first weight feature tensor set respectively to obtain a first weight value; wherein the first weight feature tensor set includes three groups of weight feature tensors; The first weight value is multiplied by the first three-dimensional feature vector to obtain a first feature tensor.

9. The semi-supervised recognition method based on frequency agility scattering characteristics and complex-valued autoencoder according to claim 1, characterized in that: The loss function used by the complex-valued autoencoder network during training is expressed as: ； in, represents the concatenation vector; represents the real part of the concatenated vector; represents the imaginary part of the concatenated vector; represents the denoised concatenated vector; represents the real part of the denoised concatenated vector; represents the imaginary part of the denoised concatenated vector, represents the imaginary unit, Represents the sum of squares of modulo values.

Images