A multi-frequency multi-polarization extremely narrow pulse echo target fusion recognition method based on feature constraint

Abstract

The application provides a multi-domain extremely narrow pulse radar echo target fusion recognition method based on scattering center matching degree constraint and structural feature auxiliary constraint, constructs an end-to-end classification model for multi-frequency and multi-polarization fusion recognition of an extremely narrow pulse radar echo, can perform feature extraction on input multi-frequency and multi-polarization extremely narrow pulse radar echoes and fusion of frequency dimension and polarization dimension information, and then directly recognizes the fusion features, realizes end-to-end fusion recognition integration, adopts deep learning technology based on feature level fusion, can more fully utilize feature information contained in samples, makes the model make more accurate and stable decisions in the classification task. The method fully mines the correlation between multi-frequency and multi-polarization information, fuses the frequency dimension and polarization dimension information mined, utilizes rich scattering information of multi-frequency and multi-polarization extremely narrow pulse radar echoes, and achieves a relatively ideal target recognition effect.

Description

Technical Field

[0001] This invention relates to the field of radar target recognition technology, specifically to a multi-frequency, multi-polarization, ultra-narrow pulse radar echo target fusion recognition method based on scattering center matching degree constraints and structural feature constraints. Background Technology

[0002] Automatic radar target identification (ADI) is a method based on electromagnetic scattering theory that uses radar echo signals from the target and its surrounding environment to extract stable target features and automatically determine the target type and attributes. When the width of a single echo pulse is much smaller than the target size, it is equivalent to the radar range resolution being much smaller than the target size. In this case, the target occupies multiple range cells, which is equivalent to an object composed of multiple discrete scattering centers. The vector sum of the projections of the target scattering points onto the radar line of sight is called the extremely narrow pulse radar echo, also known as the high-resolution range profile (HRRP).

[0003] Extremely narrow pulse radar echoes reflect the distribution of the radar cross section (RCS) of a target scattering point along the radar line of sight, revealing the relative position and intensity of the scattering point. Compared to (inverse) synthetic aperture radar (SAR / ISAR) images, target recognition is easier to acquire and process; therefore, automatic radar target recognition based on extremely narrow pulse radar echoes has significant research value and practical application.

[0004] Traditional radar target recognition methods based on extremely narrow pulse radar echoes typically rely on single-band, single-polarization, or multi-polarization data, which lacks sufficient effective target information, leading to unsatisfactory recognition results. With the development of modern radar technology, multi-domain extremely narrow pulse radars, incorporating multi-band and multi-polarization information, have increased the dimensions of information acquisition, enabling the extraction of richer target scattering mechanism information and gradually becoming an important development direction. Furthermore, deep learning has also demonstrated excellent performance in radar target recognition. By designing a deep learning model based on radar scattering feature constraints, combining frequency band and polarization extremely narrow pulse radar echoes, richer target information can be utilized to improve target recognition performance.

[0005] Among existing multi-domain fusion methods, multi-polarization fusion technology has made some progress, but a mature and standardized approach has not yet emerged, and research on multi-band fusion technology is relatively limited. Existing multi-polarization fusion methods mainly focus on data-level fusion, including kernel mapping, principal component analysis, and wavelet transform, but these methods have low fusion levels and require high data quality. Existing multi-band fusion methods primarily perform fusion at the decision-level, including traditional fusion algorithms such as Bayesian inference and the Dempster-Shafer method. However, rich discriminative information exists between multi-frequency, multi-polarization HRRPs, and existing methods have not fully integrated the complementary information across multiple frequencies and polarizations, limiting the fusion and recognition performance. Summary of the Invention

[0006] This invention provides a multi-domain ultra-narrow pulse radar echo target fusion and recognition method based on scattering center matching degree constraints and structural feature auxiliary constraints. The method firstly uses feature-level fusion, leveraging scattering center constraints and feature auxiliary constraints to mine and fuse frequency dimension information of the ultra-narrow pulse radar echoes. Then, through channel and spatial attention modules, adaptive weighting is applied to the multi-polarization ultra-narrow pulse radar echoes in the channel and range dimensions, and the fusion of polarization ultra-narrow pulse radar echoes is achieved through weighted summation. Finally, the fused ultra-narrow pulse radar echoes are input into a classifier to achieve target recognition.

[0007] A multi-frequency, multi-polarization, ultra-narrow pulse echo target fusion and identification method includes the following steps:

[0008] Step 1: Obtain measured ultra-narrow pulse radar echo data of radar targets to obtain multi-band fully polarized ultra-narrow pulse radar echoes of different targets of type M.

[0009] Step 2: Construct a multi-domain fusion network, which sequentially includes an input layer, a salient feature extraction layer, a frequency band fusion layer, a polarization fusion layer, and a fully connected classification layer.

[0010] The salient feature extraction layer contains K feature extraction paths, each path receiving multi-polarization ultra-narrow pulse radar echoes in one frequency band, in order to... This represents the fully polarized input of the k-th frequency band. This represents the four polarization echo data of the k-th frequency band; each route consists of two cascaded convolutional layers, used for target saliency scattering feature extraction, and the output feature sequence. :

[0011] The frequency band fusion layer includes a channel and a spatial attention module. With two convolutional layers and ;

[0012] In the channel and spatial attention module In the process, the input K feature sequences Compositional characteristics One-dimensional convolution through channel attention module The obtained weights and features Multiplication yields intermediate feature maps Subsequently, a channel-based two-dimensional convolution is performed using a spatial attention module. The obtained weights are compared with the intermediate feature maps. Multiplication yields the output result. Then, the feature sequences of different polarization channels are weighted, and subsequently, the feature sequences of different frequency bands of the same polarization channel are added together to obtain preliminary frequency band fusion results. ;

[0013] Through convolutional layers right Further deep feature extraction is performed to obtain the frequency band fusion feature sequence. ; through convolutional layers right Achieving information from the strong scattering centers of the fused sequence Extraction:

[0014]

[0015] Finally, the information of the strong scattering center will be... Superimposed on frequency band fusion feature sequences The feature sequence of frequency band fusion is obtained above. ;

[0016] The polarization fusion layer includes a convolutional layer. With a channel and spatial attention module Features are further extracted through convolutional layers, and then through channel and spatial attention modules. Adaptive weighting of the extremely narrow pulse radar echo and summation of J polarization channels yield a polarization fusion feature sequence. ;

[0017]

[0018] The fully connected classification layer targets the input polarization fusion feature sequence. Output the target classification result;

[0019] Step 3: Train the multi-domain fusion network based on the loss function; input the radar echo data to be identified into the trained multi-domain fusion network to obtain the target classification result.

[0020] Furthermore, when training the multi-domain fusion network, a loss function is also set for the frequency band fusion layer, which is obtained by weighted summation of similarity loss, scattering center loss and structural feature loss;

[0021] The similarity loss is achieved by fusing extremely narrow pulse radar echoes. Information loss compared to the original ultra-narrow pulse radar echo; scattering center loss is the difference between the scattering information of the multi-band ultra-narrow pulse radar echo and the scattering information of the fused ultra-narrow pulse radar echo; structural feature loss is the difference in structural features between the frequency band fused ultra-narrow pulse radar echo and the original ultra-narrow pulse radar echo.

[0022] Preferably, the scattering center loss includes the loss in the number of scattering centers and the loss in the location of scattering centers.

[0023] Preferably, for the loss of the number of scattering centers, the absolute value of the absolute value of the union of the number of scattering centers of the ultra-narrow pulse radar echo and the number of scattering centers of the multi-frequency ultra-narrow pulse radar echo is used for calculation.

[0024] Preferably, for the loss of the scattering center position, the L2 norm is calculated by subtracting the union of the scattering center positions of the extremely narrow pulse radar echo and the scattering center positions of the multi-frequency extremely narrow pulse radar echo.

[0025] Preferably, the structural feature loss includes symmetry features, dispersion features, and descaled structural features of the amplitude waveform.

[0026] Preferably, the loss function of the entire multi-domain fusion network is obtained by weighted summation of the classification loss function and the loss function of the frequency band fusion layer.

[0027] Ideally, the backpropagation algorithm is used to update the weight parameters and train the constructed multi-domain fusion network.

[0028] Ideally, during the training of the multi-domain fusion network, the weights are updated along the direction of the decrease in the loss function.

[0029] Preferably, during the training of the multi-domain fusion network, updating the weights along the direction of loss function descent includes using a stochastic gradient descent algorithm.

[0030] The present invention has the following beneficial effects:

[0031] This invention proposes a multi-domain ultra-narrow pulse radar echo target fusion recognition method based on scattering center matching degree constraints and structural feature-assisted constraints. It constructs an end-to-end classification model for multi-frequency, multi-polarization fusion recognition of ultra-narrow pulse radar echoes. This model can extract features from input multi-frequency, multi-polarization ultra-narrow pulse radar echoes and fuse frequency and polarization dimensions, then directly recognize the fused features, achieving end-to-end integrated fusion recognition. This method employs deep learning technology based on feature-level fusion, which can fully utilize the feature information contained in the samples, enabling the model to make more accurate and stable decisions in classification tasks. By fully exploring the correlation between multi-band and multi-polarization information and fusing the mined frequency and polarization dimensions, it utilizes the rich scattering information of multi-frequency, multi-polarization ultra-narrow pulse radar echoes to achieve relatively ideal target recognition results.

[0032] By utilizing scattering center matching constraints, L2 norm operations are performed on the scattering information of multi-band ultra-narrow pulse radar echoes and the fused ultra-narrow pulse radar echoes to fuse the differential scattering information of multiple bands, making the multi-band scattering information complementary and reducing information redundancy. Structural feature-assisted constraints are further used to constrain the feature sequence of the frequency-fused ultra-narrow pulse radar echoes, making the physical meaning of the frequency-fused ultra-narrow pulse radar echoes clearer and further defining the physical boundary of the frequency-fused range profile, thus enhancing the model's generalization ability.

[0033] By using convolution operations of channel and spatial attention modules, the polarization and range dimensions of polarimetric ultra-narrow pulse radar echoes are adaptively weighted. The fusion of polarimetric ultra-narrow pulse radar echoes is achieved through weighted summation, making full use of the multi-polarization information of ultra-narrow pulse radar echoes. Attached Figure Description

[0034] Figure 1 Block diagram of a multi-domain ultra-narrow pulse radar echo fusion and identification model;

[0035] Figure 2 A detailed structural diagram of the model;

[0036] Figure 3 This is a structural diagram of the channel and spatial attention modules. Detailed Implementation

[0037] The present invention will now be described in detail with reference to the accompanying drawings and examples.

[0038] For multi-frequency, multi-polarized target ultra-narrow pulse radar echoes, the dynamic range of the scattering center intensity varies with frequency, with different frequency bands containing complementary information from scattering centers of varying intensities. In terms of polarization, the scattering mechanisms of different scattering centers differ, with different polarization modes containing complementary information from scattering centers with different scattering mechanisms. Based on these characteristics of multi-frequency, multi-polarized ultra-narrow pulse radar echoes, this invention utilizes a deep learning model to achieve adaptive feature extraction. Simultaneously, it employs scattering center-assisted constraints and structural feature-assisted constraints to fully leverage the scattering and structural feature information of multi-polarized ultra-narrow pulse radar echoes across different frequency bands, thereby improving the physical interpretability and target recognition performance of the data-driven model.

[0039] Therefore, this invention provides a multi-domain ultra-narrow pulse radar echo target fusion identification method based on scattering center matching degree constraints and structural feature-assisted constraints. The framework of this method is as follows: Figure 1 As shown. The multi-domain fusion model mainly includes an input layer, a salient feature extraction layer, a frequency band fusion layer, a polarization fusion layer, and a fully connected classification layer. Specifically, as shown... Figure 2 As shown:

[0040] The technical approach to implementing this invention is as follows: First, acquire multi-polarization ultra-narrow pulse radar echo data in different frequency bands, perform preprocessing operations such as normalization, and divide the data into training, validation, and test sets. Then, construct a system as follows: Figure 2 The multi-domain fusion network structure shown enables the fusion of multi-frequency, multi-polarization, ultra-narrow pulse radar echo data. During frequency band fusion, scattering center matching constraints and structural feature-assisted constraints are used to constrain the fusion process, enhancing the physical meaning of the fused ultra-narrow pulse radar echoes. During polarization fusion, channel and spatial attention mechanisms are employed to adaptively weight each polarization channel and the sequence within each channel of the frequency-band fused ultra-narrow pulse radar echo, fully utilizing the feature information of each polarization channel to maximize the effective information of the polarization-fused ultra-narrow pulse radar echoes. Then, multi-frequency, multi-polarization, ultra-narrow pulse radar echo samples and category labels from the training set are input into the network for training, resulting in a parameter-optimized multi-domain fusion recognition network. Finally, multi-frequency, multi-polarization, ultra-narrow pulse radar echoes from the test set are input into the multi-domain fusion recognition network to achieve recognition and classification.

[0041] Specifically, the steps include the following:

[0042] Step 1: Acquire the ultra-narrow pulse radar echo, perform preprocessing such as data alignment, amplitude normalization, and extraction of scattering information, and divide the dataset into training set, validation set, and test set.

[0043] 101. Obtain measured extremely narrow pulse radar echo data of radar targets to obtain multi-frequency, multi-polarization extremely narrow pulse radar echoes of M different types of targets, where M is at least 1; use Different frequency bands are represented by K, and the number of polarizations is J. The commonly used value of J is 2 or 4, which represent dual-polarization and fully polarization data modes, respectively.

[0044] 102. Align the extremely narrow pulse radar echo data using the first frequency band. Based on the data, the centroid alignment method is used to align the echoes of multi-band ultra-narrow pulse radar.

[0045] 103. Amplitude normalization processing is performed on the raw data of ultra-narrow pulse radar echoes in different frequency bands. Usually, the amplitude of cross-polarized ultra-narrow pulse radar echoes is lower than that of co-polarized ultra-narrow pulse radar echoes. In order to ensure that this characteristic is maintained after normalization, the amplitude normalization processing of fully polarized ultra-narrow pulse radar echoes is performed based on the maximum value of ultra-narrow pulse radar echoes in vertical polarization (HH polarization).

[0046] The maximum value of the extremely narrow pulse radar echo under HH polarization is used to normalize the extremely narrow pulse echoes of different polarizations. Let... This represents the original extremely narrow pulse echo in the k-th frequency band, where P and Q represent horizontal polarization H or vertical polarization V, for example... The amplitude-normalized echo of a very narrow pulse transmitted vertically and received horizontally is expressed as:

[0047]

[0048] in, This represents the HH-polarized extremely narrow pulse radar echo in the k-th frequency band. .

[0049] 104. Extract the scattering center of the ultra-narrow pulse radar echo from multi-band input and set a threshold. If the amplitude of the ultra-narrow pulse radar echo exceeds a threshold and is a peak point, this location is designated as a strong scattering center. The scattering center is then encoded into a one-hot vector with a length equal to that of the ultra-narrow pulse radar echo. :

[0050]

[0051] in, Represents the first one-hot vector One point, The function representing the scattering center extraction function, This represents the modulo value. is the length of the one-hot vector.

[0052] It is the scattering center vector obtained by extracting the input extremely narrow pulse radar echo of the PQ polarization channel in the k-th frequency band, which serves as the label information in addition to the category label.

[0053] 105. Divide the dataset into training and test sets.

[0054] Step 2: Construct a multi-domain fusion network. This network structure consists of an input layer, a salient feature extraction layer, a frequency band fusion layer, a polarization fusion layer, and a fully connected classification layer. The specific structure of the multi-domain fusion network is as follows: Figure 2 As shown.

[0055] Step 201: Construct a salient feature extraction layer, which contains K feature extraction paths. Each path receives multi-polarization ultra-narrow pulse radar echoes in one frequency band, in order to... express The frequency band route takes a fully polarized input, with each route consisting of two cascaded convolutional layers, used for target saliency scattering feature extraction, and outputs a feature sequence. :

[0056]

[0057] For the first A cascaded convolutional layer for each feature extraction path, Specifically This module is capable of mining and extracting deep features of ultra-narrow pulse radar echoes.

[0058] Step 202: Construct a frequency band fusion layer, which consists of a channel and a spatial attention module. With two convolutional layers and constitute.

[0059] Channel and spatial attention modules, such as Figure 3 As shown. It consists of a channel attention module and a spatial attention module, with input features... One-dimensional convolution through channel attention module The obtained weights With features Multiplication yields intermediate feature maps Subsequently, a channel-based two-dimensional convolution is performed using a spatial attention module. The obtained weights With intermediate feature map Multiplication yields the output result. This can be expressed in functional form as follows:

[0060]

[0061] Through channel and spatial attention modules The feature sequences of different channels are weighted, and then the feature sequences of the corresponding polarization channels of different frequency bands are added together to obtain the preliminary frequency band fusion result. :

[0062]

[0063] Through convolutional layers Further deep feature extraction is performed; through convolutional layers. Achieving information from the strong scattering centers of the fused sequence Extraction:

[0064]

[0065] Finally, the information of the strong scattering center will be... Superimposed on the initial frequency band fusion feature sequence of channel J The multi-band, multi-channel feature sequences from the feature extraction module are fused to obtain the frequency band fused feature sequence. J represents the number of polarization channels. The length of the feature sequence is:

[0066]

[0067]

[0068] in, This is the i-th unit of the frequency band fusion sequence.

[0069] Step 203: Construct a polarization fusion layer, which includes convolutional layers. With a channel and spatial attention module Features are further extracted through convolutional layers, and then through channel and spatial attention modules. Adaptive weighting of the extremely narrow pulse radar echo and summation of J polarization channels yield a polarization fusion feature sequence. .Right now:

[0070]

[0071] Step 204: Construct a fully connected classification layer, with the last layer being a softmax layer containing M neurons, where M is the number of target categories.

[0072] Output an M-dimensional vector, representing the probability that the training target belongs to each of the M classes. The class with the highest probability is taken as the target recognition result. ,Right now:

[0073]

[0074] Step 3: Constructing the loss function. In the frequency band fusion layer, a similarity loss function, a scattering center loss constraint, and a structural feature auxiliary constraint are added for frequency band fusion. The similarity loss is used to initially constrain the information loss between the fused and original ultra-narrow pulse radar echoes, ensuring that the fused ultra-narrow pulse radar echo sequence fully covers the information of the original ultra-narrow pulse radar echoes from each channel. The scattering center loss constraint is used to complement the scattering information of the multi-band ultra-narrow pulse radar echoes with the fused ultra-narrow pulse radar echoes, reducing information redundancy and achieving a fusion effect for multi-polarization ultra-narrow pulse radar echoes from different frequency bands. The structural auxiliary constraint is used to further constrain the fusion effect of the frequency band ultra-narrow pulse radar echoes, increasing the physical interpretability of the fused ultra-narrow pulse radar echoes. Finally, the cross-entropy loss is used to fit the sample labels, enabling the multi-domain fusion network to learn the ultra-narrow pulse radar echo data.

[0075] Step 301: Construct a similarity loss. The information loss of the fused ultra-narrow pulse radar echo sequence should be lower than that of the original input. Using amplitude similarity loss as a constraint for network optimization, by maximizing the similarity of the ultra-narrow pulse radar echoes before and after frequency band fusion, the network can initially possess the ability to reduce information loss. For the input multi-frequency fully polarized ultra-narrow pulse radar echo sequence, let the... The input of each channel is The fusion output is Then the similarity loss It can be represented as:

[0076]

[0077] For the first The m-th point of the input sequence of a channel This represents the m-th point in the fused output sequence.

[0078] Step 302: Construct the scattering center information loss. For the same target, the ultra-narrow pulse radar echoes of different frequency bands have strong correlation, and the target scattering center information contained in the ultra-narrow pulse radar echoes has small differences. The ultra-narrow pulse radar echo after frequency band fusion should be as close as possible to the scattering center information of the multi-frequency band ultra-narrow pulse radar echoes. By minimizing the difference between the fused ultra-narrow pulse radar echo scattering information and the input ultra-narrow pulse radar echo scattering information, the correlation of feature information of different frequency bands can be strengthened.

[0079] Using step two, the information sequence of strong scattering center of ultra-narrow pulse radar echo is extracted and fused through the network. ,Will Sequence and input scattering information Perform L2 norm operations.

[0080] The formula for the scattering center loss function is as follows:

[0081]

[0082]

[0083]

[0084] (Take the union of the number of scattering centers for all frequency bands)

[0085] (Take the union of the scattering center locations for all frequency bands)

[0086]

[0087] in, For the loss of scattering information, For the loss of the number of scattering centers, This represents the loss due to the location of the scattering center.

[0088] Based on The extracted information on the number of scattering centers of the fused ultra-narrow pulse radar echo. For the first Information on the number of echo scattering centers of ultra-narrow pulse radar in each frequency band. Based on The extracted information on the location of the fused ultra-narrow pulse radar echo scattering center. For the first Information on the location of the echo scattering center of ultra-narrow pulse radar in each frequency band. It contains two polarization channels, HH and HV. It contains four polarization channels: HH, HV, VH, and VV.

[0089] The scattering center regularization term consists of two parts: the loss in the number of scattering centers and the loss in the location of scattering centers. The above losses between the fused ultra-narrow pulse radar echo and the multi-frequency ultra-narrow pulse radar echo should be minimized.

[0090] To account for the loss in the number of scattering centers, the absolute value of the absolute value is taken by subtracting the union of the number of scattering centers of the fusion-based ultra-narrow pulse radar echo and the number of scattering centers of the multi-frequency ultra-narrow pulse radar echo.

[0091] For the loss of the scattering center position, the method of subtracting the union of the scattering center position of the ultra-narrow pulse radar echo and the scattering center position of the multi-frequency ultra-narrow pulse radar echo and taking the L2 norm is used for calculation.

[0092] The L2 norm formula is as follows:

[0093]

[0094] in, Represents the strong scattering center of the fused extremely narrow pulse radar echo. This represents the union of the center vectors of the echo scattering from a multi-band input extremely narrow pulse radar.

[0095] Step 303: The existence of extremely narrow pulse radar echoes can reflect the structural characteristics of their internal structure. By using structural characteristics as auxiliary constraints, the physical characteristics of the frequency-band fused extremely narrow pulse radar echoes can be made more apparent. Three structural features with good stability were selected for auxiliary constraints: symmetry feature, dispersion feature, and descaled structural feature of amplitude waveform.

[0096] The dispersion characteristic reflects the degree of dispersion of the target's scattering cross-section, as shown in the following formula:

[0097]

[0098]

[0099] in, This is the PQ-polarized extremely narrow pulse radar echo sequence in the k-th frequency band. The dispersion characteristics of PQ-polarized extremely narrow pulse radar echoes in the k-th frequency band; represents the dispersion characteristic of the frequency band fusion sequence; N is the length of the extremely narrow pulse radar echo sequence.

[0100] The symmetry characteristic reflects the degree of symmetry in the distribution of the target's scattering cross-section, as shown in the following formula:

[0101]

[0102]

[0103] in, The symmetry characteristics of the PQ-polarized extremely narrow pulse radar echo in the k-th frequency band; This represents the degree of dispersion of the frequency band fusion sequence.

[0104] The descaled structural features of the amplitude waveform reflect the internal structural features of the range image, independent of its scale variation, as shown in the following formula:

[0105]

[0106]

[0107] in, The symmetry characteristics of the PQ-polarized extremely narrow pulse radar echo in the k-th frequency band; This represents the degree of dispersion of the frequency band fusion sequence.

[0108] Concatenate the three features:

[0109]

[0110] First, the single-value characteristics of the multi-band data are averaged, i.e.

[0111]

[0112] The average of the input structural feature vectors, For the first Single-valued feature vector of a frequency band.

[0113] The result is an eigenvector composed of single-valued structural features. This eigenvector is then combined with the eigenvector of the fused narrow-pulse radar echo and subjected to the L2 norm to obtain the structural feature-assisted constraint loss function.

[0114]

[0115] Step 304: In order to achieve the effect of integrated recognition, after the ultra-narrow pulse radar echo is fused, it is directly input into the classification network. The classification loss is used to constrain the target classification effect to achieve end-to-end recognition.

[0116] The loss function used is the cross-entropy loss function:

[0117]

[0118] in For known category labels This is the actual output of the softmax layer. For input The number of elements. By minimizing the classification backpropagation loss function, the feature representation capability of the network can be enhanced.

[0119] The error sensitivity of the output layer is propagated forward using the backpropagation algorithm to update the weights of each layer. Through repeated iterations, the cost function converges, and finally, the network training is complete, resulting in a trained network model.

[0120] Step 4: Input the training and validation sets of multi-frequency, multi-polarization, ultra-narrow pulse radar echoes into the constructed multi-domain fusion network and train it using the loss function proposed in Step 3.

[0121] Step 401: Input the multi-frequency fully polarized data of the training and validation sets into the network as required, extract the hierarchical features of the ultra-narrow pulse radar echo through the designed network, realize the fusion of frequency band dimension through the fusion layer, realize the fusion of polarization dimension through the channel and spatial attention modules, and realize the classification through the classifier.

[0122] Step 402: Adjust network parameters through backpropagation using four loss functions. The total loss function for frequency band fusion is:

[0123]

[0124] These are the weighting coefficients of each loss function in the frequency band fusion loss.

[0125] The total loss function for network training is:

[0126]

[0127] These are the weighting coefficients for the frequency band fusion loss and the cross-entropy loss, respectively.

[0128] Step 5: After preprocessing the multi-frequency, multi-polarization, ultra-narrow pulse radar echoes of the test set in the manner described in Step 1, input them into the multi-domain fusion network for classification testing.

[0129] Step 501: Input the multi-frequency multi-polarization channel data from the test sample according to the corresponding channel in the training phase.

[0130] Step 502: Input the feature map that has passed through the feature extraction module, frequency band fusion layer and polarization fusion layer into the fully connected layer. The last softmax layer of the fully connected layer performs classification. The category corresponding to the maximum value output by the softmax layer is the target category.

[0131] Example:

[0132] The invention will be further described below using measured dual-frequency fully polarized ultra-narrow pulse radar echo data from three types of targets.

[0133] This embodiment takes dual-band full polarization as an example, using Ku and W dual-band full polarization radar echoes of three typical targets at different azimuth angles with an ultra-narrow pulse radar echo bandwidth of 0.75 GHz. The simulation data includes 60 azimuth angles, starting at 1° and spaced 6° apart, with 30 frames of ultra-narrow pulse radar echoes at each azimuth angle. A comparative experiment was set up between a single-frequency single-polarization CNN network and a fusion network. For the single-frequency single-polarization CNN network, two model testing methods were set up, as shown in the table below:

[0134]

[0135] For the fused network, a dual-frequency, full-polarization training and testing method using Ku and W frequencies was adopted. Ten ultra-narrow pulse radar echoes were randomly selected from each azimuth angle as the training set, and 20 ultra-narrow pulse radar echoes were selected as the test set.

[0136] The block diagram for this example is as follows: Figure 1 As shown, Ku and W dual-frequency fully polarized ultra-narrow pulse radar echoes are input in parallel into a multi-domain fusion network. Convolutional layers are used to extract deep-level target features. Frequency-dimensional fusion is achieved using channel and spatial attention modules, scattering center loss constraints, and structural feature-assisted constraints. Polarization-dimensional fusion is realized through channel and spatial attention mechanisms, thereby deeply mining target information, extracting target features, and ultimately achieving target classification and recognition. The results of single-frequency single-polarization and multi-domain fusion ultra-narrow pulse radar echo recognition are shown in Table 1.

[0137] Table 1 Comparison of recognition performance of different methods

[0138]

[0139] As shown in Table 1, the features extracted by the multi-domain fusion network are more separable than those extracted by the single-frequency single-polarization network. The recognition performance of the fusion model is improved by more than 22% compared with the single-frequency single-polarization basic CNN model. The multi-domain fusion network with added physical constraints has better recognition effect, with a recognition rate improved by more than 25% compared with the single-frequency single-polarization basic CNN model.

[0140] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-frequency, multi-polarization, ultra-narrow pulse echo target fusion and recognition method, characterized in that, Includes the following steps: Step 1: Obtain measured ultra-narrow pulse radar echo data of radar targets to obtain multi-band fully polarized ultra-narrow pulse radar echoes of different targets of type M. Step 2: Construct a multi-domain fusion network, which sequentially includes an input layer, a salient feature extraction layer, a frequency band fusion layer, a polarization fusion layer, and a fully connected classification layer. The salient feature extraction layer contains K feature extraction routes, each route receiving multi-polarization ultra-narrow pulse radar echoes in one frequency band, in order to... This represents the fully polarized input of the k-th frequency band. This represents the four polarization echo data of the k-th frequency band; each route consists of two cascaded convolutional layers used for target saliency scattering feature extraction, and the output feature sequence... : The frequency band fusion layer includes a channel and a spatial attention module. With two convolutional layers and ; In the channel and spatial attention module In the process, the input K feature sequences Compositional characteristics One-dimensional convolution through channel attention module The obtained weights and features Multiplication yields intermediate feature maps Subsequently, a channel-based two-dimensional convolution is performed using a spatial attention module. The obtained weights are compared with the intermediate feature maps. Multiplication yields the output result. Then, the feature sequences of different polarization channels are weighted, and subsequently, the feature sequences of different frequency bands of the same polarization channel are added together to obtain preliminary frequency band fusion results. ; Through convolutional layers right Further deep feature extraction is performed to obtain the frequency band fusion feature sequence. ; through convolutional layers right Achieving information from the strong scattering centers of the fused sequence Extraction: Finally, the information of the strong scattering center will be... Superimposed on frequency band fusion feature sequences The feature sequence of frequency band fusion is obtained above. ; The polarization fusion layer includes a convolutional layer. With a channel and spatial attention module Features are further extracted through convolutional layers, and then through channel and spatial attention modules. Adaptive weighting of the extremely narrow pulse radar echo and summation of J polarization channels yield a polarization fusion feature sequence. ; The fully connected classification layer targets the input polarization fusion feature sequence. Output the target classification result; Step 3: Train the multi-domain fusion network based on the loss function; The radar echo data to be identified is input into a trained multi-domain fusion network to obtain the target classification result.

2. The multi-frequency, multi-polarization, ultra-narrow pulse echo target fusion and recognition method as described in claim 1, characterized in that, When training the multi-domain fusion network, a loss function is also set for the frequency band fusion layer, which is obtained by weighted summation of similarity loss, scattering center loss and structural feature loss. The similarity loss is achieved by fusing extremely narrow pulse radar echoes. Information loss compared to the original ultra-narrow pulse radar echo; scattering center loss is the difference between the scattering information of the multi-band ultra-narrow pulse radar echo and the scattering information of the fused ultra-narrow pulse radar echo; structural feature loss is the difference in structural features between the frequency band fused ultra-narrow pulse radar echo and the original ultra-narrow pulse radar echo.

3. The multi-frequency, multi-polarization, ultra-narrow pulse echo target fusion and recognition method as described in claim 2, characterized in that, The loss of scattering centers includes the loss of the number of scattering centers and the loss of the location of scattering centers.

4. The multi-frequency, multi-polarization, ultra-narrow pulse echo target fusion and recognition method as described in claim 3, characterized in that, For the loss of the number of scattering centers, the absolute value of the absolute value is taken by subtracting the union of the number of scattering centers of the ultra-narrow pulse radar echo and the number of scattering centers of the multi-frequency ultra-narrow pulse radar echo.

5. The multi-frequency, multi-polarization, ultra-narrow pulse echo target fusion and recognition method as described in claim 3, characterized in that, For the loss of the scattering center position, the method of subtracting the union of the scattering center position of the ultra-narrow pulse radar echo and the scattering center position of the multi-frequency ultra-narrow pulse radar echo and taking the L2 norm is used for calculation.

6. The multi-frequency, multi-polarization, ultra-narrow pulse echo target fusion and recognition method as described in claim 2, characterized in that, The structural feature loss includes symmetry features, dispersion features, and descaled structural features of amplitude waveforms.

7. A multi-frequency, multi-polarization, ultra-narrow pulse echo target fusion and recognition method as described in any one of claims 2 to 6, characterized in that, The loss function of the entire multi-domain fusion network is obtained by weighted summation of the classification loss function and the loss function of the frequency band fusion layer.

8. The multi-frequency, multi-polarization, ultra-narrow pulse echo target fusion and recognition method as described in claim 7, characterized in that, The weight parameters are updated using the backpropagation algorithm, and the constructed multi-domain fusion network is trained.

9. The multi-frequency, multi-polarization, ultra-narrow pulse echo target fusion and recognition method as described in claim 8, characterized in that, During the training of the multi-domain fusion network, the weights are updated along the direction of the decrease in the loss function.

10. The multi-frequency, multi-polarization, ultra-narrow pulse echo target fusion and recognition method as described in claim 9, characterized in that, During the training of a multi-domain fusion network, updating the weights along the direction of loss function descent includes using the stochastic gradient descent algorithm.

Images