Narrowband radar aerial target classification method based on deep feature fusion network

Abstract

The present invention proposes a narrowband radar aerial target classification method based on a deep feature fusion network, which is implemented by the following steps: obtaining a training sample set and a test sample set; constructing a deep feature fusion network; defining a maximum margin orthogonal loss function; initializing parameters: training the deep feature fusion network; updating the parameters of the deep feature fusion network; and obtaining aerial target classification results. The present invention trains the deep feature fusion network and obtains aerial target classification results, wherein the convolution module adaptively fuses the shallow features of the narrowband radar and extracts deep global features. Then, the channel and spatial attention modules are combined to obtain local highly separable features based on the deep global features, avoiding the defect of the existing technology that uses shallow features for fusion, resulting in low classification accuracy. Target reconstruction error is used to screen out targets outside the library, thereby improving the reliability of classification. Finally, the real-time performance of the classification is effectively improved through the end-to-end network structure.

Description

Technical Field

[0001] The present invention belongs to the field of narrowband radar target classification and relates to an aerial target classification method, in particular to a narrowband radar aerial target classification method based on a deep feature fusion network. Background Art

[0002] Radar air target classification involves extracting features from radar air target echo data and automatically determining the target's class. Low-resolution narrowband radars generally have a range resolution greater than the size of targets detected by conventional radars. Targets are approximated as "points," making it impossible to extract fine structures. However, they can extract features at many unique attribute levels, which are crucial for narrowband radar target classification.

[0003] Existing narrowband radar aerial target classification methods often use statistical analysis to construct a probabilistic model. These methods then directly fuse shallow target features, such as those in the time, frequency, and Doppler domains, based on the probabilistic model. Finally, target classification is achieved based on these fused features. For different aerial categories, the shallow features acquired by narrowband radar contribute differently to target classification. Some features are not separable and must be suppressed. Existing narrowband radar target classification methods typically directly combine shallow features, lacking the adaptive ability to highlight highly separable features and suppress ambiguous features. Furthermore, compared to deep neural networks, simple statistical analysis methods fail to extract deep features that are more representative of the target. The separability of shallow features is relatively limited, making it difficult to fully reflect the target's intrinsic information, resulting in limited narrowband radar target classification accuracy. Furthermore, existing narrowband radar target classification methods are unable to reject or filter out irrelevant targets outside the target's classification, making the models susceptible to interference from these targets. For example, existing techniques have proposed a method that extracts shallow target features in the time, Doppler, and time-frequency domains, fuses these features, and then utilizes pattern recognition and classification algorithms for target classification. For example, patent application publication number CN 116167012A, entitled "A Narrowband Radar Aerial Target Classification Method Based on Multi-Domain Fusion Features," discloses a narrowband radar target classification method based on multi-domain feature fusion. This method extracts features in the time, Doppler, and time-frequency domains and performs normalization processing to obtain a multi-domain feature matrix. A random forest is then used to evaluate the multi-domain features, obtain weights for each feature, and fuse the multi-domain features to obtain a fused feature matrix. The fused feature matrix is ​​used to train an SVM for classification. This method extracts features from multiple domains, which not only comprehensively reflects the characteristics of the target but also utilizes multi-domain feature fusion to achieve complementary advantages among features in different signal domains, thereby achieving better classification and recognition results and improving the accuracy of aerial target classification. However, this method has the following shortcomings: 1. It uses the random forest Gini index to directly combine shallow features and then uses support vector machines for classification. This method cannot extract deeper features that are more representative of the target, resulting in poor generalization performance and low classification accuracy of the trained model. 2. This method does not have the ability to reject decoys and other out-of-database targets, resulting in unreliable classification results. 3. This method is not an end-to-end target classification method and has poor real-time performance. Summary of the Invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and propose a narrowband radar aerial target classification method based on a deep feature fusion network to solve the technical problems existing in the prior art, such as low classification accuracy, lack of the ability to reject out-of-repository targets and poor real-time performance.

[0005] To achieve the above object, the technical solution adopted by the present invention includes the following steps:

[0006] (1) Obtain training sample set and test sample set:

[0007] The narrowband radar is used to obtain D groups of target feature data of N categories, which are composed of C decoy targets outside the library and NC targets inside the library. The feature sequences obtained after preprocessing each group of target feature data are labeled, and then the S groups of the M groups of library feature sequences obtained by preprocessing and their labels are composed of the training sample set X. tr , the pre-processed feature sequences outside the library and the remaining feature sequences in the MS group library are combined into the test sample set X te , where 1≤C≤N-2, N≥3, D≥10,

[0008] (2) Constructing a deep feature fusion network O:

[0009] Construct a deep feature fusion network O including an encoding module, a decoding module and a fully connected module cascaded and arranged in parallel therewith, and a Softmax function layer cascaded at the output end of the fully connected module, wherein the encoding module is used to extract features of input data and includes multiple cascaded composite modules, each composite module cascaded includes one or more convolution modules, a channel attention module and a spatial attention module; the decoding module includes multiple cascaded deconvolution modules for reconstructing input data based on the features extracted by the encoding module;

[0010] (3) Define the maximum edge orthogonal loss function L MOP :

[0011]

[0012] Among them, x s is the sth sample, P(x s ) is the sample x s The classification probability, For category y s The center of , ||·||2 is the L2 norm, θ s,j For category y s With category y j The angle between the eigenvectors of , λ and γ are equalization factors;

[0013] (4) Initialization parameters:

[0014] The number of initial training iterations is t, the maximum number of iterations is T, T ≥ 1000, and the parameters of the encoding module, decoding module and fully connected module in the deep feature fusion network of the tth iteration are φ t , α t and μ t , and let t = 0;

[0015] (5) Train the deep feature fusion network O:

[0016] The training sample set X tr As the input of the deep feature fusion network O, forward propagation is performed to obtain S samples that are consistent with the input training sample set X tr The corresponding classification probability P and reconstruction output

[0017] (6) Update the parameters of the deep feature fusion network O:

[0018] The S classification probabilities P obtained by step (5) and the reconstructed output Use P to adjust the parameters φ of the deep feature fusion network coding module and the fully connected module t and α t Update, and use The parameter μ of the deep feature fusion network decoding module t Update to get the deep feature fusion network O of this iteration t , judge whether t≥T is established, if so, get the deep feature fusion network O of the training number * Otherwise, let t = t + 1, O t =0 and execute step (5);

[0019] (7) Obtaining aerial target classification results:

[0020] The test sample set X te As the input of the trained deep feature fusion network O*, forward propagation is performed to obtain the reconstructed output corresponding to Q test samples And classification probability P Q , and calculated by the MSE loss function With X te The reconstruction error L r , then judge L r Does it meet the pre-set threshold τ? r ≥τ, if so, the target corresponding to this test sample is an out-of-library target, and it is screened out; otherwise, the target corresponding to this test sample is an in-library target, and it is classified to obtain the classification result, Q=MS.

[0021] Compared with the prior art, the present invention has the following advantages:

[0022] 1. In the process of training the deep feature fusion network and obtaining the aerial target classification results, the convolution module in the present invention extracts global deep features, and then combines the channel attention module and the spatial attention module to obtain local features with high separability based on the preliminary deep features, so that the network can obtain deep high-separability features, avoiding the defect of the existing technology that only shallow features with limited separability can be obtained, and effectively improving the classification accuracy.

[0023] 2. The present invention reconstructs the deep and highly separable features obtained by the encoding module through the decoding module, and uses the target reconstruction error calculated by the MSE loss function to determine whether the target is an out-of-library target, and then screens it out, thereby rejecting out-of-library targets and improving the reliability of classification.

[0024] 3. The deep feature fusion network constructed by the present invention is an end-to-end network from narrowband radar target echo data to classification results. Feature fusion and classification do not need to be completed separately, which improves the real-time performance of classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is an implementation flow chart of the present invention;

[0026] Figure 2 It is a schematic diagram of the structure of the deep feature fusion network of the present invention;

[0027] Figure 3 It is a schematic diagram of the composite module structure of the present invention;

[0028] Figure 4 Schematic diagram of the deconvolution module structure of the present invention;

[0029] Figure 5 Schematic diagram of the channel attention module and spatial attention module structure of the present invention;

[0030] Figure 6 It is a box plot of the reconstruction error of the present invention. DETAILED DESCRIPTION

[0031] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0032] Reference Figure 1 , the implementation steps of the present invention are as follows:

[0033] (1) Obtain training sample set and test sample set:

[0034] Obtain N categories of D groups of target feature data from narrowband radar target echo data, which consist of C decoy targets outside the library and NC targets within the library. Label the feature sequence obtained after preprocessing each group of target feature data. Then, S groups of the M groups of library feature sequences obtained by preprocessing and their labels form the training sample set X. tr The preprocessed M′ group library feature sequences and the remaining MS group library feature sequences form the test sample set X te , where M and M′ are obtained by preprocessing D groups of target feature data of N categories. In this embodiment, N=6, C=1, D=10, M=540, M′=108, and S=360;

[0035] The implementation steps for preprocessing each set of target feature data are:

[0036] (1a) Using a sliding window with a window length of l and a step size of step, each group of target feature data with a length of L is trimmed to obtain M groups of in-library feature sequences and M′ groups of out-library feature sequences. In this embodiment, l = 100, step = 25, L = 1000, M = 540, and M′ = 108;

[0037] (1b) Each active RCS sequence and semi-active RCS sequence in each group of clipped feature sequences is normalized by L2 norm. At the same time, each radial velocity, radial distance, heading angle, and pitch angle in each group of feature sequences is normalized by Z-score to obtain the preprocessed M groups of in-library feature sequences and M′ groups of out-library feature sequences. The L2 norm normalization method is:

[0038]

[0039] Among them, S is the sample size of the training sample set, here is, For the active RCS sequence and semi-active RCS sequence in the training sample set after normalization, the Z-score normalization method is:

[0040]

[0041] Where σ is the standard deviation, The radial velocity, radial distance, heading angle, and pitch angle sequence in the training sample set after normalization, in this embodiment, S=360.

[0042] (2) Construct a deep feature fusion network O, whose structure is as follows Figure 2 shown.

[0043] Construct a deep feature fusion network O including an encoding module, a decoding module and a fully connected module cascaded and arranged in parallel with the encoding module, and a Softmax function layer cascaded at the output end of the fully connected module, where:

[0044] The encoding module is used to extract the features of the input data, including two cascaded composite modules. The structure of the composite module is as follows: Figure 3 As shown, each composite module cascade includes two convolution modules, a channel attention module and a spatial attention module. The structure of the convolution module is as follows Figure 4 As shown, the channel attention module and spatial attention module structures are as follows Figure 5 As shown; the decoding module includes four cascaded deconvolution modules for reconstructing input data based on the features extracted by the encoding module;

[0045] The convolution module consists of a cascade of convolutional layers with a kernel size of 3×1, a ReLU activation function layer, a normalization layer, and a max pooling layer;

[0046] Deconvolution module, including cascaded deconvolution layer, activation function ReLU layer, normalization layer and upsampling layer;

[0047] The channel attention module includes an adaptive average pooling layer and an adaptive maximum pooling layer arranged in parallel, as well as a convolutional layer, a summation layer, and a sigmoid activation layer cascaded with the output of the adaptive average pooling layer and the adaptive maximum pooling layer;

[0048] The spatial attention module includes an adaptive average pooling layer and an adaptive maximum pooling layer arranged in parallel, as well as a convolutional layer and a sigmoid activation layer cascaded with the output ends of the adaptive average pooling layer and the adaptive maximum pooling layer.

[0049] (3) Define the maximum edge orthogonal loss function L MOP :

[0050]

[0051] Among them, x s is the sth sample, P(x s ) is the sample x s The classification probability, For category y s The center of , ||·||2 is the L2 norm, θ s,j For category y s With category y j The angle between the eigenvectors of , λ and γ are equalization factors;

[0052] Category y j Center The calculation formula is:

[0053]

[0054] in, For category yj The hth sample of For category y j Features, belongs to category y j The sample size of category y s With category y j The angle cosθ between the eigenvectors s,j The calculation method is:

[0055]

[0056] Among them, x s For category y s The eigenvector of x j For category y j The eigenvector of .

[0057] (4) Initialization parameters:

[0058] The number of initial training iterations is t, the maximum number of iterations is T, T ≥ 1000, and the parameters of the encoding module, decoding module and fully connected module in the deep feature fusion network of the tth iteration are φ t , α t and μ t , and let t = 0, where T = 1000;

[0059] (5) Train the deep feature fusion network O:

[0060] The training sample set X tr As the input of the deep feature fusion network O, forward propagation is performed to obtain S samples that are consistent with the input training sample set X tr The corresponding classification probability P and reconstructed feature sequence In this embodiment, S=360;

[0061] (5a) One or more convolutional modules in the first composite module in the encoding module are used to train the training sample set X tr Perform convolution operation on the S feature sequences to obtain the feature map F1;

[0062] The channel attention module in the first composite module in the encoding module uses the adaptive average pooling layer and the adaptive maximum pooling layer to calculate the average matrix F for the feature map F1. avg and the maximum matrix F max , and then the mean value matrix F avg and the maximum matrix F maxAs the input of the convolution layer with a convolution kernel size of k×1, the feature maps F2 and F3 are obtained by convolution operation respectively, and the feature map F4 is obtained by summing F2 and F3. Then, F4 is normalized by the Sigmoid function to obtain the output feature map F of the channel attention module. c , the calculation method of the channel attention module can be expressed as:

[0063]

[0064] in, is the matrix element-by-element multiplication, Sm represents the Sigmoid function, Conv 3×3 represents a k×1 convolution operation, Avg(·) represents finding the average, and max(·) represents finding the maximum. In this embodiment, k=3;

[0065] The spatial attention module in the first composite module in the encoding module pays attention to the feature map F c Use the adaptive average pooling layer and the adaptive maximum pooling layer to obtain the average matrix F c_avg and the maximum matrix F c_max , the mean value matrix F c_avg and the maximum matrix F c_max Splicing to get feature map F con , and then the feature map F con The feature map F is obtained as the input of the convolution layer with a convolution kernel size of k×1. c_con , and then use the Sigmoid function to adjust F c_con Normalize and get the output feature map F of the channel attention module s , the calculation method of the spatial attention module can be expressed as:

[0066]

[0067] Except for the first composite module, the other composite modules of the coding layer perform the above processing on the feature map output by the previous composite module in sequence, and finally obtain the output feature map F of the coding layer. e , in this embodiment, k=3;

[0068] (5b) The deconvolution module and the fully connected module in the decoding module respectively perform the feature map F obtained by the encoding layer e Perform upsampling and classification to obtain the training sample set X tr The reconstructed output corresponding to the S feature sequences And the classification probability P, in this embodiment S=360.

[0069] (6) Update the parameters of the deep feature fusion network O:

[0070] The S classification probabilities P obtained by step (5) and the reconstructed output Use P to adjust the parameters φ of the deep feature fusion network coding module and the fully connected module t and α t Update, and use The parameter μ of the deep feature fusion network decoding module t Update to get the deep feature fusion network O of this iteration t , judge whether t≥T is established, if so, get the deep feature fusion network O of the training number * Otherwise, let t = t + 1, O t =0 and execute step (5), in this embodiment S=360;

[0071] (6a) Using the maximum margin orthogonal loss function L MOP , and through the training sample set X tr The predicted probability P and training sample set X tr The corresponding label y obtains the loss value loss MOP , using the MSE loss function, and passing O t The resulting reconstructed output And the training sample set X tr Obtain reconstruction error loss r , using loss value loss MOP Obtaining the gradient and Utilize loss r Finding the gradient

[0072] loss MOP =L MOP (P,y;f(X s )=P)

[0073]

[0074]

[0075]

[0076]

[0077] in, and They are respectively the use of loss value loss MOP For parameter α t and φ t Find the gradient, To use the loss value loss r For the parameter μ t Find the gradient, and They are loss values ​​respectively MOP For parameter φ t and α t The gradient is obtained;

[0078] (6b) For parameter φ t and α t To update:

[0079]

[0080]

[0081]

[0082] Among them, β is the learning rate, φ t+1 , α t+1 , μ t+1 They are the updated parameters respectively. In this embodiment, the initial learning rate β is 0.001.

[0083] (7) Obtaining aerial target classification results:

[0084] The test sample set X te As the input of the trained deep feature fusion network O*, forward propagation is performed to obtain Q groups of reconstructed feature sequences corresponding to Q test samples and Q classification probabilities P Q , and calculate each group through the MSE loss function With X te The loss value is obtained, and Q reconstruction errors L are obtained. r , then judge each L r Does it meet the pre-set threshold τ? r ≥τ, if so, then it is determined that L r The target corresponding to the test sample with ≥τ is an out-of-library target and is screened out. Otherwise, it is determined to meet L r The target corresponding to the test sample with τ < τ is the target in the library, which is classified to obtain the classification result. In this embodiment, Q = 180 and τ = 10.

[0085] The following simulation experiments are used to verify the technical effects of the present invention:

[0086] 1. Simulation conditions and contents:

[0087] The simulation used an NVIDIA GeForce RTX 3090 GPU with 24GB of video memory and an Intel(R) Core(TM) i9-10920X CPU with a 3.50GHz clock speed and 64GB of memory. The simulation code was compiled using PyCharmCommunity 2020.1 and Matlab R2015a, using Python 3.7.1. The frameworks used were PyTorch 1.1.0 and Cuda 9.0, and the main libraries used were NumPy 1.19.1, Scikit-learn 0.23.2, Scipy 1.5.2, and Matplotlib 3.3.2.

[0088] Data conditions: Both active and semi-active radars are in narrowband mode. The electromagnetic simulation software Feko is used to simulate five typical aerial targets as in-database targets: early warning aircraft, tankers, small jets, helicopters, and civil aircraft. Two categories of out-of-database targets are simulated: drag interference and out-of-database data generated based on the assumption of uniform distribution across the entire space. The simulated maneuvers include straight ahead, left turn, right turn, left detour, and right detour. For each maneuver, electromagnetic simulations are performed at an initial heading angle of 45°, an initial pitch angle of 4°, an initial heading angle of 47°, and initial pitch angles of 5°, 7°, and 10°. The data sampling period of feature sequences of different dimensions is 20ms. The sequence length is subsequently trimmed by sliding windows to construct feature sequence samples.

[0089] The classification performance of the present invention is compared with that of an existing narrowband radar air target classification method based on multi-domain fusion features. The results are as follows: Figure 6 , as shown in Table 1 and Table 2.

[0090] 3) Analysis of simulation results:

[0091] a) Identification of off-site targets:

[0092] This embodiment first uses the trained deep feature fusion network to identify the out-of-library targets contained in the test set. The reconstruction error box plot obtained by identification is as follows: Figure 6As shown, from left to right, the reconstruction errors for samples generated based on uniform distribution across the entire space, towed interference, and five typical aerial targets within the reservoir are shown. As can be seen from the figure, the reconstruction errors for artificially generated samples outside the reservoir are generally larger and have a wider distribution range, with reconstruction errors ranging from 2000 to 3000. The reconstruction errors for towed decoys range from 500 to 1000, while the reconstruction errors for targets within the reservoir are all less than 10 and are smaller than those for targets outside the reservoir. Therefore, using reconstruction errors to identify targets outside the reservoir is valid. Table 1 shows the accuracy of the trained deep feature fusion network for identifying targets outside the reservoir included in the test set. The comparison method, RF-SVM, lacks the ability to identify targets outside the reservoir, resulting in an identification accuracy of 0. The identification accuracy achieved by the present invention is 95.23%. This demonstrates that the present invention has the ability to effectively identify targets outside the reservoir, such as towed decoys.

[0093] Table 1 Comparison of identification accuracy

[0094]

[0095] b) Classification of targets within the library:

[0096] After the rejection process for false targets such as towed decoys is completed, the model performs target classification. Under the condition that the sequence length is 100 sample points, the classification accuracy of the present invention and the comparative method is shown in Table 2. The RF-SVM classification accuracy is 80.65%, which is relatively low classification performance and takes a long time, 3.73 seconds. This is because the existing RF-SVM is not an end-to-end classification method. This method first combines features using the random forest Gini index, then trains the SVM model, and uses the SVM model to obtain the classification results. Compared with RF-SVM, the classification accuracy of the present invention is improved by 13.53%, while the test time is relatively reduced by 34.32%. This shows that the present invention can effectively improve the classification performance of the model. At the same time, it can directly obtain end-to-end classification results for narrowband radar feature data through a deep feature fusion network, effectively improving the real-time performance of the method.

[0097] Table 2 Comparison of classification accuracy

[0098]

Claims

1. A narrowband radar aerial target classification method based on deep feature fusion network, characterized by The steps include: (1) Obtain training sample set and test sample set: Obtain N categories of D groups of target feature data from narrowband radar target echo data, which consist of C decoy targets outside the library and NC targets within the library. Label the feature sequence obtained after preprocessing each group of target feature data. Then, S groups of the M groups of library feature sequences obtained by preprocessing and their labels form the training sample set X. tr The preprocessed M′ group library feature sequences and the remaining MS group library feature sequences form the test sample set X te , where N-2≥C≥1, N≥3, D≥10, (2) Constructing a deep feature fusion network O: Construct a deep feature fusion network O including an encoding module, a decoding module and a fully connected module cascaded and arranged in parallel therewith, and a Softmax function layer cascaded at the output end of the fully connected module, wherein the encoding module includes multiple cascaded composite modules, each composite module cascaded includes one or more convolution modules, a channel attention module and a spatial attention module for extracting features of input data; the decoding module includes multiple cascaded deconvolution modules for reconstructing the features extracted by the encoding module; (3) Define the maximum edge orthogonal loss function L MOP : Among them, P(x s ) is the sth sample x s The classification probability, for No. s indivual Category y s The center of , ||·||2 is the L2 norm, θ s,j for No. s indivual Category y s and No. j indivual Category y j The angle between the eigenvectors of , λ and γ are equalization factors; (4) Initialization parameters: The number of initial training iterations is t, the maximum number of iterations is T, T ≥ 1000, and the parameters of the encoding module, decoding module and fully connected module in the deep feature fusion network of the tth iteration are φ t , α t and μ t , and let t = 0; (5) Train the deep feature fusion network O: The training sample set X tr As the input of the deep feature fusion network O, forward propagation is performed to obtain S samples that are consistent with the input training sample set X tr The corresponding classification probability P and reconstructed feature sequence (6) Update the parameters of the deep feature fusion network O to obtain the trained deep feature fusion network: The S classification probabilities P obtained through step (5) and the reconstructed feature sequence The parameters φ of the encoding module are respectively t and the parameter α of the fully connected module t , parameters μ of the decoding module t Update to get the deep feature fusion network O of this iteration t , judge whether t≥T is true, if so, get the trained deep feature fusion network O * Otherwise, let t = t + 1, O t =0, and execute step (5); (7) Obtaining aerial target classification results: The test sample set X te As the input of the trained deep feature fusion network O*, forward propagation is performed to obtain Q groups of reconstructed feature sequences corresponding to Q test samples and Q classification probabilities P Q , and calculate each group through the MSE loss function With X te The loss value is obtained, and Q reconstruction errors L are obtained. r , then judge each L r Does it meet the pre-set threshold τ? r ≥τ, if so, then it is determined that L r The target corresponding to the test sample with ≥τ is an out-of-library target and is screened out. Otherwise, it is determined to meet L r The target corresponding to the test sample with <τ is the target in the library. It is classified to obtain the classification result, Q=MS.

2. The narrowband radar aerial target classification method according to claim 1, characterized in that: The target characteristic data acquired by the narrowband radar in step (1) includes the target's active RCS, semi-active RCS, radial velocity, radial range, heading angle, and pitch angle.

3. The narrowband radar aerial target classification method according to claim 2, characterized in that: The preprocessing of each set of target feature data in step (1) is implemented as follows: (1a) Each set of target feature data of length L is trimmed by a sliding window with a window length of l and a step size of step to obtain M sets of in-library feature sequences and M′ sets of out-library feature sequences, where 10≤step≤100; (1b) Each active RCS sequence and semi-active RCS sequence in each group of clipped feature sequences is normalized by L2 norm, and each radial velocity, radial distance, heading angle, and pitch angle in each group of feature sequences is normalized by Z-score to obtain the preprocessed M groups of in-library feature sequences and M′ groups of out-library feature sequences.

4. The narrowband radar aerial target classification method according to claim 1, characterized in that: The deep feature fusion network O described in step (2), wherein: Convolution module, including a cascade of convolution layers with a kernel size of k×1, a ReLU activation function layer, a normalization layer, and a maximum pooling layer; Deconvolution module, including cascaded deconvolution layer, activation function ReLU layer, normalization layer and upsampling layer; The channel attention module includes an adaptive average pooling layer and an adaptive maximum pooling layer arranged in parallel, as well as a convolutional layer, a summation layer, and a sigmoid activation layer cascaded with the output of the adaptive average pooling layer and the adaptive maximum pooling layer; The spatial attention module includes an adaptive average pooling layer and an adaptive maximum pooling layer arranged in parallel, as well as a convolutional layer and a sigmoid activation layer cascaded with the output ends of the adaptive average pooling layer and the adaptive maximum pooling layer.

5. The narrowband radar aerial target classification method according to claim 4, characterized in that: The training of the deep feature fusion network O described in step (5) is implemented as follows: (5a) One or more convolutional modules in the first composite module in the encoding module are used to train the training sample set X tr Perform convolution operation on the S feature sequences to obtain S feature maps F1; The channel attention module in the first composite module in the encoding module uses the adaptive average pooling layer and the adaptive maximum pooling layer to calculate the average value matrix F for each feature map F1 avg and the maximum matrix F max , and then the mean value matrix F avg and the maximum matrix F max As the input of the convolution layer with a convolution kernel size of k×1, the feature maps F2 and F3 are obtained by convolution operation respectively, and the feature map F4 is obtained by summing F2 and F3. Then, F4 is normalized by the Sigmoid function to obtain the output feature map F of the channel attention module. c , perform the above operation on each feature map F1 to obtain S feature maps F c ; The spatial attention module in the first composite module in the encoding module pays attention to the feature map F c Use the adaptive average pooling layer and the adaptive maximum pooling layer to obtain the average matrix F c_avg and the maximum matrix F c_max , the mean value matrix F c_avg and the maximum matrix F c_max Splicing to get feature map F con , and then the feature map F con The feature map F is obtained as the input of the convolution layer with a convolution kernel size of k×1. c_con , and then use the Sigmoid function to adjust F c_con Normalize and get the output feature map F of the channel attention module s , for each feature map F c Perform the above operation to obtain S feature maps F s ; Except for the first composite module, the other composite modules in the coding layer perform the above processing on the feature map output by the previous composite module in sequence, and finally obtain the output feature map F of S coding modules e ; (5b) The deconvolution module and the fully connected module in the decoding module respectively perform the feature map F obtained by the encoding layer e Reconstruct and classify to obtain the training sample set X tr The reconstructed feature sequence corresponding to the S feature sequences And classification probability P.

6. The narrowband radar aerial target classification method according to claim 1, characterized in that: Category y described in step (3) j Center The calculation formula is: in, is the jth category y j The hth sample of y j Features, For y j The sample size of .

7. The narrowband radar aerial target classification method according to claim 1, characterized in that: The parameter φ of the encoding module described in step (6) t and the parameter α of the fully connected module t , parameters μ of the decoding module t To update, the steps are: (6a) Using the maximum margin orthogonal loss function L MOP , and through the training sample set X tr The predicted probability P and X tr The corresponding label y calculates X tr The loss value loss MOP , using the MSE loss function and reconstructing the feature sequence and X tr Calculate F e Reconstruction error loss r , using loss value loss MOP Obtaining the gradient and Utilize loss r Finding the gradient loss MOP =L MOP (P,y;f(X s )=P) in, and They are respectively the use of loss value loss MOP For parameter α t and φ t Find the gradient, To use the loss value loss r For the parameter μ t Find the gradient, and They are loss values ​​respectively MOP For parameter φ t and α t The gradient is obtained; (6b) For parameter φ t , α t and μ t To update: Among them, β is the learning rate, φ t+1 , α t+1 , μ t+1 are the parameters after updating.

Images