Robust target recognition method for radar bandwidth variation based on selective kernel attention

Abstract

The present invention discloses a method for robust target recognition with radar bandwidth variation based on selective kernel attention. The method comprises: obtaining radar high-resolution range image data and bandwidth information of the target to be identified as the test data; inputting the data into a pre-trained selective kernel attention network to obtain the target recognition result; the network comprises a sequentially connected parallel branch structure, a selective kernel attention layer, a spatial pyramid pooling layer, and a fully connected layer structure; the parallel branch structure comprises a parallel convolutional neural network structure and a bandwidth information branch, respectively for receiving the radar high-resolution range image data and the bandwidth information; the selective kernel attention network is trained using a training dataset generated from radar high-resolution range image data and corresponding bandwidth information of multiple target categories with different bandwidths. By extracting multi-scale features from the different bandwidth data and utilizing bandwidth information to assist in multi-scale feature fusion, the method can reduce database construction costs and effectively be used for target recognition tasks with radar bandwidth variation.

Description

Technical Field

[0001] The present invention belongs to the field of radar target recognition, and in particular relates to a radar bandwidth variation robust target recognition method based on selective kernel attention. Background Art

[0002] A high-resolution radar range image is acquired using a broadband radar signal. It is defined as the vector sum of the sub-echoes from the target's scattering points projected along the radar's line of sight. This image reflects the distribution of the target's scattering points along the radar's line of sight and contains rich target information, useful for target identification. The radar's range resolution is determined by its bandwidth. The larger the bandwidth, the richer the target information contained in the high-resolution radar range image. In actual radar operation, the radar typically needs to adjust its bandwidth to identify targets of varying sizes. This results in changes in the range resolution and dimensionality of the echo signal. Using a database with a single bandwidth can lead to a mismatch between the test samples and the training set, resulting in a decrease in or even loss of the network's recognition capability.

[0003] Therefore, when the radar bandwidth changes, existing technologies typically require databases to be established at multiple bandwidths in order for the network to recognize data from the new bandwidth. For example, the method for recognizing radar high-resolution range profiles using a convolutional neural network, presented in "Radar High-Resolution Range Profile Target Recognition Method Based on One-Dimensional Convolutional Neural Networks," requires establishing databases at multiple bandwidths and retraining the network. However, the cost of establishing such a large database and retraining the network is extremely high.

[0004] Furthermore, the paper "Robust Target Recognition with Repetition-Frequency Variation Based on Spatial Pyramid Pooling Network" proposes a network structure for radar signal dimensionality variations caused by variations in the repetition-frequency operating parameters of narrowband radars. While this method can address the issue of varying network input signal dimensions, its multi-scale feature extraction capabilities are limited and it cannot fully extract multi-scale features from data of varying bandwidths, making it unsuitable for radar bandwidth variation recognition.

[0005] It can be seen that there is currently a lack of effective methods for target recognition in scenarios where radar bandwidth changes. Summary of the Invention

[0006] In order to solve the above problems existing in the prior art, the present invention provides a radar bandwidth variation robust target recognition method based on selective kernel attention. The technical problem to be solved by the present invention is achieved through the following technical solutions:

[0007] Obtaining radar high-resolution range image data and corresponding bandwidth information of the target to be identified as the data to be measured;

[0008] The test data is input into a pre-trained selective kernel attention network to obtain a target recognition result; wherein the selective kernel attention network includes a parallel branch structure, a selective kernel attention layer, a spatial pyramid pooling layer and a fully connected layer structure connected in sequence, and the parallel branch structure includes a parallel convolutional neural network structure and a bandwidth information branch, which are respectively used to receive radar high-resolution range image data and bandwidth information; the selective kernel attention network is trained using a training data set generated by radar high-resolution range image data of multiple categories of targets with different bandwidths and corresponding bandwidth information.

[0009] In one embodiment of the present invention,

[0010] The convolutional neural network structure includes a first convolution layer, a first pooling layer, a ReLU activation layer, a second convolution layer, a second pooling layer, a ReLU activation layer, a third convolution layer, a third pooling layer and a ReLU activation layer connected in sequence; wherein the number of convolution kernels of the first convolution layer, the second convolution layer and the third convolution layer is set to 8, 16 and 32 respectively; the convolution kernel size is set to 1×5; the convolution kernel moving step is set to 1; the first pooling layer, the second pooling layer and the third pooling layer all adopt the maximum pooling strategy; the pooling kernel size is set to 1×2; the pooling kernel moving step is set to 2;

[0011] The bandwidth information branch includes three fully connected layers, each of which is followed by a ReLU activation function, and the weight dimensions of each fully connected layer are set to 1×8, 8×16, and 16×32 respectively;

[0012] The selective kernel attention layer includes four convolution branches, the scaling factor is set to 4, the number of convolution kernels of the four convolution branches is set to 32, the convolution kernel sizes are set to 1×3, 1×5, 1×7 and 1×9 respectively, and the convolution kernel moving step size is set to 1;

[0013] The spatial pyramid pooling layer includes pooling at three scales, and each scale pooling adopts a maximum pooling strategy. The number of spatial bins of the three scale poolings is 1×1, 1×4, and 1×16, respectively.

[0014] The fully connected layer structure includes a first fully connected layer, a ReLU activation layer, a second fully connected layer and a softmax layer; wherein the output dimension of the first fully connected layer is 512, and the output dimension of the second fully connected layer is 10; the softmax layer adopts a softmax function.

[0015] In one embodiment of the present invention, the spatial pyramid pooling layer includes pooling at three scales, each scale of pooling adopts an average pooling strategy, and the numbers of spatial bins of the three scales of pooling are 1×1, 1×4, and 1×16, respectively.

[0016] In one embodiment of the present invention, the processing of the selective kernel attention layer includes:

[0017] Using the four convolution branches of the selective kernel attention layer, feature extraction is performed on the input of the selective kernel attention layer to obtain features at multiple scales;

[0018] Add the features of the multiple scales and use global average pooling to obtain a one-dimensional vector S;

[0019] Obtaining a vector having the same length as the one-dimensional vector S as generated by processing the bandwidth information by the bandwidth information branch, and adding the vector to the one-dimensional vector S to obtain a bandwidth information fusion vector S′;

[0020] The bandwidth information fusion vector S′ is subjected to linear transformation to generate a one-dimensional vector Z;

[0021] The one-dimensional vector Z is subjected to different linear transformations and then softmax processing to obtain the attention weights of different channel features at different scales;

[0022] The output result of the selective kernel attention layer is obtained by weighted summing of each attention weight and the original multi-scale features.

[0023] In one embodiment of the present invention, the selective kernel attention network training process includes:

[0024] Perform Gaussian initialization on the parameters of each layer of the constructed selective kernel attention network;

[0025] Obtaining a training data set and segmenting the data sets corresponding to different bandwidths in the training data set to obtain batch data for each bandwidth;

[0026] Each batch of data after the segmentation process is sequentially input into the initialized network, and preliminary features are extracted through the convolutional neural network structure; wherein the feature dimension of the preliminary features is 32×a, where a is determined by the input data dimension of the network;

[0027] Inputting bandwidth information corresponding to the input batch data into the bandwidth information branch to obtain a bandwidth information branch output vector;

[0028] Inputting the preliminary features and the bandwidth information branch output vector into the selective kernel attention layer to obtain a fine feature after multi-scale feature fusion, wherein the feature dimension of the fine feature is 32×a;

[0029] Input the fine features into the spatial pyramid pooling layer, perform pooling at three different scales, and obtain features with dimensions of 32×1, 32×4, and 32×16, respectively. The three obtained features are concatenated to obtain an output result of the spatial pyramid pooling layer with a feature dimension of 32×21;

[0030] Flatten the output of the spatial pyramid pooling layer into a one-dimensional feature with a dimension of 1×672, input it into the fully connected layer structure, and obtain the predicted value of the training sample in the input batch data;

[0031] After batch data of all bandwidths are input into the network, the back-propagation algorithm is used to optimize the total loss function of all bandwidths in a hybrid training, and the parameters of each layer of the network are iteratively updated until the network converges or reaches the set maximum number of training rounds, obtaining a trained selective kernel attention network.

[0032] In one embodiment of the present invention, the process of generating the training data set includes:

[0033] For each bandwidth, extract radar high-resolution range image data of M types of targets to form a data set corresponding to the bandwidth, and record the bandwidth information of the data set; the number of radar high-resolution range image data of each type of target in each bandwidth is at least 800;

[0034] All data sets obtained from N bandwidths are combined into an echo data set, all bandwidth information corresponding to the echo data set is combined into a bandwidth information data set, and the echo data set and the bandwidth information data set are combined into a merged data set;

[0035] Class labels are set for the radar high-resolution range image data in the merged data set to obtain a training data set; wherein M≥5 and N≥6.

[0036] In one embodiment of the present invention, the total loss function of all bandwidths in the hybrid training is:

[0037]

[0038] Among them, N represents the total number of bandwidths contained in the training set, B represents the number of samples in each bandwidth batch after segmentation, M represents the number of target categories in the training set, and H represents the cross entropy loss function. represents the predicted value of the i-th training sample in the n-th bandwidth batch, represents the label value of the i-th training sample in the n-th bandwidth batch, and ∑· represents the summation operation.

[0039] Beneficial effects of the present invention:

[0040] 1. A hybrid training strategy increases the amount of data in disguise, eliminating the need to collect high-resolution radar range profile data at all bandwidths. Only one network needs to be trained to robustly recognize data at different bandwidths.

[0041] 2. Combining spatial pyramid pooling with the selective kernel attention mechanism solves the problem of robust recognition in scenarios with varying bandwidth. This allows the network to not only input echo signals of different dimensions but also extract features from echo signals of different resolutions at multiple scales.

[0042] 3. Integrating bandwidth information into the attention calculation process is a new way to utilize echo information, allowing the network to learn the importance of features at different scales for the current bandwidth and ultimately extract the most separable features under the current bandwidth. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 A flowchart of a radar bandwidth variation robust target recognition method based on selective kernel attention provided by an embodiment of the present invention;

[0044] Figure 2 Schematic diagram of the structure of the selective kernel attention network in an embodiment of the present invention;

[0045] Figure 3 Schematic diagram of the selective kernel attention layer processing process in an embodiment of the present invention;

[0046] Figure 4 This is a diagram of the experimental results of the present invention. DETAILED DESCRIPTION

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only 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 making creative efforts are within the scope of protection of the present invention.

[0048] In order to reduce costs and fully extract multi-scale features of data with different bandwidths to be applicable to target recognition tasks with radar bandwidth variations, an embodiment of the present invention provides a radar bandwidth variation robust target recognition method based on selective kernel attention.

[0049] like Figure 1 As shown, a radar bandwidth variation robust target recognition method based on selective kernel attention provided by an embodiment of the present invention may include the following steps:

[0050] S1, obtaining radar high-resolution range image data and corresponding bandwidth information of the target to be identified as the test data;

[0051] S2, inputting the test data into a pre-trained selective kernel attention network to obtain a target recognition result;

[0052] The selective kernel attention network includes a sequentially connected parallel branch structure, a selective kernel attention layer, a spatial pyramid pooling layer, and a fully connected layer structure. The parallel branch structure includes a parallel convolutional neural network structure and a bandwidth information branch, which are respectively used to receive radar high-resolution range image data and bandwidth information. The selective kernel attention network is trained using a training data set generated by radar high-resolution range image data of multiple categories of targets with different bandwidths and corresponding bandwidth information.

[0053] In order to facilitate understanding of the embodiments of the present invention, the training data set, the structure of the selective kernel attention network, and the network training process are first described.

[0054] (1) Training dataset

[0055] In an optional implementation manner, the process of generating the training data set includes the following steps:

[0056] 1) For each bandwidth, extract the radar high-resolution range image data of M types of targets to form a data set corresponding to the bandwidth, and record the bandwidth information of the data set;

[0057] In the embodiments of the present invention, targets can be selected as needed, such as aircraft and vehicles, and are not specifically limited here. High-resolution radar range profile data for each target is used as echo data, with M ≥ 5. It will be appreciated that when acquiring high-resolution radar range profile data, corresponding bandwidth information can also be acquired and recorded. The number of high-resolution radar range profile data for each type of target within each bandwidth is at least 800.

[0058] 2) All data sets obtained from N bandwidths are combined into an echo data set, all bandwidth information corresponding to the echo data set is combined into a bandwidth information data set, and the echo data set and the bandwidth information data set are combined into a merged data set;

[0059] Among them, N≥6.

[0060] 3) setting category labels for the radar high-resolution range image data in the combined data set to obtain a training data set;

[0061] The method for setting the category labels may adopt an existing method, and different category labels may be marked with numerical values, so that the label values of the M categories of targets are: 0, 1, ..., M-1.

[0062] (2) Structure of the Selective Kernel Attention Network

[0063] This embodiment of the present invention is specifically designed for scenarios with a small number of training samples at multiple bandwidths. It builds a selective kernel attention network for identifying high-resolution radar range profiles. This aims to enable the network to recognize high-resolution radar range profiles of various bandwidths in mixed training scenarios, while significantly reducing the cost of database construction.

[0064] For network structure, see Figure 2 ; wherein, the selective kernel attention network includes a parallel branch structure, a selective kernel attention layer, a spatial pyramid pooling layer and a fully connected layer structure connected in sequence, and the parallel branch structure includes a parallel convolutional neural network structure and a bandwidth information branch, which are respectively used to receive radar high-resolution range image data and bandwidth information; wherein, the radar high-resolution range image data is simply referred to as echo data.

[0065] The following describes in detail the structure and parameter settings of each part, in one optional implementation:

[0066] ① The convolutional neural network structure

[0067] The convolutional neural network structure is used to extract features from the input echo data to obtain preliminary features; the convolutional neural network structure includes a first convolution layer, a first pooling layer, a ReLU activation layer, a second convolution layer, a second pooling layer, a ReLU activation layer, a third convolution layer, a third pooling layer and a ReLU activation layer connected in sequence; wherein, the number of convolution kernels of the first convolution layer, the second convolution layer and the third convolution layer are set to 8, 16 and 32 respectively; the convolution kernel size is set to 1×5; the convolution kernel moving step is set to 1; the first pooling layer, the second pooling layer and the third pooling layer all adopt the maximum pooling strategy; the pooling kernel size is set to 1×2; the pooling kernel moving step is set to 2.

[0068] ② The bandwidth information branch

[0069] The bandwidth information branch is used to extract the input bandwidth information and obtain a one-dimensional vector of a specific dimension as the bandwidth information branch output vector. The bandwidth information branch includes three fully connected layers, each followed by a ReLU activation function, and the weight dimensions of each fully connected layer are set to 1×8, 8×16, and 16×32, respectively.

[0070] ③The selective kernel attention layer

[0071] The selective kernel attention layer is used to extract multi-scale features from data of different bandwidths. It considers the importance of features of different scales based on bandwidth information, thereby obtaining more refined features than the preliminary feature information. The selective kernel attention layer includes four convolution branches with a scaling factor of 4. The number of convolution kernels in each of the four convolution branches is set to 32, and the convolution kernel sizes are set to 1×3, 1×5, 1×7, and 1×9, respectively. The convolution kernel movement step size is set to 1.

[0072] ④The spatial pyramid pooling layer

[0073] The spatial pyramid pooling layer is used to pool the fine features output by the selective kernel attention layer at different scales and concatenate the pooling results to obtain a fixed-dimensional spatial pyramid pooling layer output result. Since the output result dimension is fixed, the spatial pyramid pooling layer allows the network to input echo data of variable dimension. The spatial pyramid pooling layer includes three scales of pooling, each of which adopts a maximum pooling strategy. The number of spatial bins of the three scales of pooling is 1×1, 1×4, and 1×16 respectively.

[0074] In an optional embodiment, the spatial pyramid pooling layer includes pooling of three scales, each scale of pooling adopts an average pooling strategy, and the numbers of spatial boxes of the three scales of pooling are 1×1, 1×4, and 1×16, respectively.

[0075] ⑤The fully connected layer structure

[0076] The fully connected layer structure is used for final classification. It includes a first fully connected layer, a ReLU activation layer, a second fully connected layer, and a softmax layer. The output dimension of the first fully connected layer is 512, and the output dimension of the second fully connected layer is 10. The softmax layer uses the softmax function. It is understood that the softmax layer calculates the probability of the input sample being identified as each target class and selects the class corresponding to the highest probability as the recognition result.

[0077] It can be seen that the embodiment of the present invention combines spatial pyramid pooling and the selective kernel attention mechanism to solve the problem of robust recognition in bandwidth-varying scenarios, so that the trained selective kernel attention network can serve as a target recognition network and not only input echo signals of different dimensions, but also perform feature extraction on echo signals of different resolutions at multiple scales.

[0078] (3) Training process of the selective kernel attention network

[0079] The selective kernel attention network is trained using a training data set generated by radar high-resolution range image data of multiple types of targets with different bandwidths and corresponding bandwidth information.

[0080] In an optional embodiment, the training process of the selective kernel attention network includes:

[0081] A1, perform Gaussian initialization on the parameters of each layer of the constructed selective kernel attention network;

[0082] The purpose of initializing the parameters of each network layer is to obtain initial parameters to facilitate network training. The process of Gaussian initialization of the parameters of each network layer can be found in related technologies and will not be described in detail here.

[0083] A2, obtaining a training data set and segmenting the data sets corresponding to different bandwidths in the training data set to obtain batch data for each bandwidth;

[0084] The generation process of the training dataset is described above.

[0085] Since data of different bandwidths have different dimensions, it is necessary to split the data of different bandwidths to make the network training more efficient. The splitting operation is to package the data of the same bandwidth into small batches, so that the batch data of each bandwidth can be fed into the network in sequence.

[0086] A3, inputting each batch of segmented data into the initialized network in sequence, and extracting preliminary features through the convolutional neural network structure;

[0087] The feature dimension of the preliminary features is 32×a, where a is determined by the input data dimension of the network.

[0088] A4, inputting the bandwidth information corresponding to the input batch data into the bandwidth information branch to obtain a bandwidth information branch output vector;

[0089] The bandwidth information branch output vector is a one-dimensional vector with a specific dimension, which is set according to the processing process of the selective kernel attention layer. Please understand it in conjunction with the description in A5 below.

[0090] A5, inputting the preliminary features and the bandwidth information branch output vector into the selective kernel attention layer to obtain a fine feature after multi-scale feature fusion, wherein the feature dimension of the fine feature is 32×a;

[0091] Regarding the selective kernel attention layer, the traditional selective kernel attention layer works as follows: Assuming the input feature channel dimension of the selective kernel attention layer is c, the number of branches is k, and the scaling factor is r. The selective kernel attention layer first extracts features from the initially extracted features using convolution kernels of different sizes, obtaining k features of different scales, each with c channels. The k features of different scales are directly added together and then subjected to global average pooling to obtain a one-dimensional vector S of dimension 1×c that represents the information of each channel. This is then passed through a fully connected layer to obtain a dimensionally scaled one-dimensional vector Z of dimension 1×(c / r). This scaled vector Z is then mapped to k different one-dimensional vectors of dimension 1×c through k different fully connected layers. The k 1×c vectors are then normalized along the channel dimension using the softmax function to obtain attention weights for different channel features of different scales, with dimension k×c. Finally, the attention weights are weighted summed with the original k features of different scales to obtain a more refined feature after feature fusion.

[0092] Based on the above-mentioned traditional selective kernel attention layer, the embodiment of the present invention integrates bandwidth information into the attention calculation process. This is a new way to utilize echo information, enabling the network to learn the importance of different scale features for the current bandwidth data, and ultimately extract the most separable features under the current bandwidth.

[0093] For details, see Figure 3 , Figure 3 Schematic diagram of the selective kernel attention layer processing process in an embodiment of the present invention. Figure 3 , the processing process of the selective kernel attention layer includes:

[0094] B1, using the four convolution branches of the selective kernel attention layer, extracting features from the input of the selective kernel attention layer to obtain features at multiple scales;

[0095] The input of the selective kernel attention layer is the preliminary features. Since the convolution kernel sizes of the four convolution branches of the selective kernel attention layer are 1×3, 1×5, 1×7 and 1×9 respectively, after feature extraction of the preliminary features with a dimension of 32×a, four scale features can be obtained, each with a dimension of 32×a ( Figure 3 It is referred to as multi-scale features in the figure, and for simplicity, only two scale features are listed. It can be understood that the above process achieves the purpose of multi-scale feature extraction, which is conducive to extracting features that are robust to bandwidth changes.

[0096] B2, add the features of the multiple scales and use global average pooling to obtain a one-dimensional vector S with a dimension of 1×32;

[0097] Since the embodiment of the present invention needs to consider the importance of features of different scales, it is necessary to first add the features of different scales. Since the added features are multi-channel, global average pooling can consider the information of the entire channel.

[0098] B3, obtaining a vector having the same length as the one-dimensional vector S as generated by the bandwidth information branch after the bandwidth information is processed, and adding the vector to the one-dimensional vector S to obtain a bandwidth information fusion vector S′;

[0099] The dimension of the bandwidth information branch output vector generated after the bandwidth information is processed by the bandwidth information branch is the same as the dimension of the one-dimensional vector S, so as to ensure that the two can be added together.

[0100] The bandwidth information branch output vector and the one-dimensional vector S are added together to utilize the bandwidth information to assist the calculation process of the attention weight, so that the final selective feature fusion is more targeted and the final extracted features are more separable.

[0101] B4, performing a linear transformation on the bandwidth information fusion vector S′ to generate a one-dimensional vector Z;

[0102] Among them, the linear transformation is implemented using a fully connected layer. The activation function after the fully connected layer is the ReLU activation function. Since the scaling factor is 4, the weight dimension of the fully connected layer is 32×8, and the dimension of the generated one-dimensional vector Z is 1×8.

[0103] B5, the one-dimensional vector Z is subjected to different linear transformations and then softmax processing to obtain the attention weights of different channel features at different scales;

[0104] Specifically, since the selective kernel attention layer has four convolutional branches, four different linear transformations are required. The weight dimension of each linear transformation is 8×32, and four vectors with dimensions of 1×32 are obtained. After softmax processing, the attention weights of different channel features at different scales can be obtained.

[0105] B6, performing weighted summation on each attention weight and the original multi-scale features to obtain the output result of the selective kernel attention layer.

[0106] The purpose of weighted summation is to make the final extracted features take into account the importance of features at different scales and to extract fine features with richer information.

[0107] A6: Input the fine features into the spatial pyramid pooling layer and perform pooling at three different scales to obtain features with dimensions of 32×1, 32×4, and 32×16, respectively. The three obtained features are concatenated to obtain an output result of the spatial pyramid pooling layer with a feature dimension of 32×21.

[0108] Among them, since the number of spatial boxes of the three pooling scales is 1×1, 1×4, and 1×16 respectively, the features with dimensions of 32×1, 32×4, and 32×16 are obtained respectively.

[0109] The three obtained features are concatenated to obtain a spatial pyramid pooling layer output result with a feature dimension of 32×21. This dimension is fixed and independent of the input signal dimension. Therefore, the spatial pyramid pooling layer of the embodiment of the present invention allows the network to input echo data with variable dimensions.

[0110] A7, flattens the output of the spatial pyramid pooling layer into a one-dimensional feature with a dimension of 1×672, inputs the one-dimensional feature into the fully connected layer structure, and obtains the predicted value of the training sample in the input batch data;

[0111] A8, after batch data of all bandwidths are input into the network, the backpropagation algorithm is used to optimize the total loss function of all bandwidths in a hybrid training. The parameters of each layer of the network are iteratively updated until the network converges or reaches the set maximum number of training rounds, resulting in a trained selective kernel attention network.

[0112] In the embodiment of the present invention, a mixed training refers to inputting batch data of different bandwidths into the network in sequence to obtain respective training sample prediction values, and then using the back propagation algorithm to update the parameters of each layer of the network based on the prediction results and label values of the batch data of all input bandwidths.

[0113] Among them, the total loss function of all bandwidths in a mixed training is:

[0114]

[0115] Among them, N represents the total number of bandwidths contained in the training set, B represents the number of samples in each bandwidth batch after segmentation, M represents the number of target categories in the training set, and H represents the cross entropy loss function. represents the predicted value of the i-th training sample in the n-th bandwidth batch, represents the label value of the i-th training sample in the n-th bandwidth batch, and ∑· represents the summation operation.

[0116] It can be seen that the embodiment of the present invention adopts a hybrid training strategy, which can increase the data volume in disguise. It is not necessary to collect radar high-resolution range image data at each bandwidth, and data with different bandwidths can be robustly identified.

[0117] The backpropagation algorithm, loss function, parameter update, etc. in the specific training process can be combined with the relevant concepts of the network training process in the existing technology to assist in understanding.

[0118] Based on the above, it can be understood that for S2, the data to be tested is input into the pre-trained selective kernel attention network. Specifically, the echo data is input into the input end of the convolutional neural network structure, and the bandwidth information is input into the input end of the bandwidth information branch. It is processed using the selective kernel attention layer, the spatial pyramid pooling layer and the fully connected layer structure. Finally, the probability of the target to be identified as each category is calculated through the softmax layer in the fully connected layer structure, and the category corresponding to the highest probability is selected as the recognition result.

[0119] In the method for robust target recognition under radar bandwidth variations based on selective kernel attention, provided in an embodiment of the present invention, a pre-built selective kernel attention network comprises a sequentially connected parallel branch structure, a selective kernel attention layer, a spatial pyramid pooling layer, and a fully connected layer structure. The parallel branch structure comprises a parallel convolutional neural network structure and a bandwidth information branch, which are respectively used to receive radar high-resolution range image data and bandwidth information. This embodiment of the present invention leverages the advantages of the network's multi-scale feature extraction and selective feature fusion, enabling the network to recognize data of varying bandwidths. Furthermore, by utilizing the bandwidth information of the echo, the network can selectively fuse features specific to the current bandwidth, further improving network performance and effectively adapting it to robust target recognition tasks under radar bandwidth variations.

[0120] In order to verify the effectiveness of the method of the embodiment of the present invention, the following is an illustration with reference to simulation experiments.

[0121] Experimental Preparation

[0122] The hardware platform of the simulation experiment of the present invention is: a processor Intel(R) Core(TM) i7-8700 CPU with a main frequency of 3.2GHz and a memory of 16GB.

[0123] The simulation experiment software platform of the present invention is: Windows 10 operating system and Python 3.6.

[0124] Experimental results

[0125] The data used in the experiments of this invention are high-resolution radar range profiles of 10 types of aircraft, simulated using CST electromagnetic simulation software. The attitude angle coverage range is 0°-180°. To test the performance of the recognition method, the pitch angles corresponding to the training and test sets for each type of aircraft are 3° and 5°, respectively. The data contains data from 11 bandwidths: 300MHz, 400MHz, 500MHz, 600MHz, 700MHz, 800MHz, 900MHz, 1000MHz, 1100MHz, 1200MHz, and 1300MHz. At each bandwidth, the training and test sets for each type of aircraft contain 900 and 1000 records, respectively. For all 10 types of aircraft, the training and test sets contain 9000 and 10,000 records, respectively. The total training and test sets for all bandwidths are 99,000 and 110,000 records, respectively.

[0126] Figure 4 It is the experimental result diagram of the present invention. Figure 4 In the paper, the "non-selective kernel" method refers to the robust recognition method for narrowband radar repetition frequency changes proposed in the "Robust Target Recognition Method for Repetition Frequency Changes Based on Spatial Pyramid Pooling Network" (Application Publication No. CN112666533A).

[0127] The recognition accuracy results of different bandwidth data of the method of the present invention and the "non-selective kernel" method are plotted as follows: Figure 4 shown. Figure 4 The horizontal axis represents the bandwidth, which is 300MHz, 400MHz, 500MHz, 600MHz, 700MHz, 800MHz, 900MHz, 1000MHz, 1100MHz, 1200MHz, and 1300MHz, corresponding to 11 groups of data with different bandwidths in the test dataset. The vertical axis represents the recognition accuracy of the test dataset. The star-shaped dotted line represents the "no selective kernel" method, and the dotted solid line represents the method of the present invention, called "selective kernel - with bandwidth information." In addition, to verify the role of bandwidth information, a simplified version of the method of the present invention is also drawn in the figure, represented by a dotted line, called "selective kernel - without bandwidth information."

[0128] Compared with the method of the present invention, the network structure of the "no selective kernel" method does not have the bandwidth information branch and the selective kernel attention layer, while the other structures are exactly the same. Compared with the method of the present invention, the network structure of the "selective kernel-no bandwidth information" method does not have the bandwidth information branch, while the other structures are exactly the same.

[0129] Depend on Figure 4It can be seen that the recognition accuracy of the method of the present invention is better than that of the "non-selective kernel" method at any test bandwidth. When the bandwidth is 300M, the recognition accuracy of the method of the present invention is improved by 47 percentage points compared with the "non-selective kernel" method, and the average improvement is about 20 percentage points per bandwidth. Compared with the "non-selective kernel" method, the recognition accuracy of the method of the present invention fluctuates very little with the change of bandwidth. When the bandwidth exceeds 800M, the recognition accuracy of the method of the present invention remains basically stable, while the recognition accuracy of the "non-selective kernel" method decreases significantly. In addition, compared with the "selective kernel-no bandwidth information" method, the recognition performance of the method of the present invention is improved to a certain extent. When the bandwidth is 300M, the recognition rate is improved by 6.2 percentage points, and the average improvement is about 2 percentage points per bandwidth, which proves that the bandwidth information in the method of the present invention plays a role in the network.

[0130] The above description is only a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention are included in the scope of protection of the present invention.

Claims

1. A radar bandwidth-variable robust target recognition method based on selective kernel attention, characterized in that: include: Obtaining radar high-resolution range image data and corresponding bandwidth information of the target to be identified as the data to be measured; The test data is input into a pre-trained selective kernel attention network to obtain a target recognition result; wherein the selective kernel attention network includes a parallel branch structure, a selective kernel attention layer, a spatial pyramid pooling layer, and a fully connected layer structure connected in sequence, and the parallel branch structure includes a parallel convolutional neural network structure and a bandwidth information branch, which are respectively used to receive radar high-resolution range image data and bandwidth information; the selective kernel attention network is trained using a training data set generated from radar high-resolution range image data of multiple types of targets with different bandwidths and corresponding bandwidth information; The bandwidth information branch includes three fully connected layers, each of which is followed by a ReLU activation function, and the weight dimensions of each fully connected layer are set to 、 and ; The selective kernel attention layer includes four convolution branches, the scaling factor is set to 4, the number of convolution kernels of the four convolution branches is set to 32, and the convolution kernel sizes are set to 、 、 and , the convolution kernel moving step size is set to 1; The processing of the selective kernel attention layer includes: Using the four convolution branches of the selective kernel attention layer, feature extraction is performed on the input of the selective kernel attention layer to obtain features at multiple scales; Add the features of the multiple scales and use global average pooling to obtain a one-dimensional vector S; Obtaining a vector having the same length as the one-dimensional vector S as generated by processing the bandwidth information by the bandwidth information branch, and adding the vector to the one-dimensional vector S to obtain a bandwidth information fusion vector S′; The bandwidth information fusion vector S′ is subjected to linear transformation to generate a one-dimensional vector Z; The one-dimensional vector Z is subjected to different linear transformations and then softmax processing to obtain the attention weights of different channel features at different scales; The output result of the selective kernel attention layer is obtained by weighted summing of each attention weight and the original multi-scale features.

2. The radar bandwidth variation robust target recognition method based on selective kernel attention according to claim 1 is characterized in that: The convolutional neural network structure includes a first convolutional layer, a first pooling layer, a ReLU activation layer, a second convolutional layer, a second pooling layer, a ReLU activation layer, a third convolutional layer, a third pooling layer and a ReLU activation layer connected in sequence; wherein the number of convolution kernels of the first convolutional layer, the second convolutional layer and the third convolutional layer is set to 8, 16 and 32 respectively; the size of the convolution kernel is set to ; The convolution kernel movement step size is set to 1; The first pooling layer, the second pooling layer and the third pooling layer all adopt the maximum pooling strategy; The pooling kernel size is set to ;The pooling kernel moving step size is set to 2; The spatial pyramid pooling layer includes three scales of pooling, each scale of pooling adopts the maximum pooling strategy, and the number of spatial boxes of the three scales of pooling is respectively 、 and ; The fully connected layer structure includes a first fully connected layer, a ReLU activation layer, a second fully connected layer and a softmax layer; wherein the output dimension of the first fully connected layer is 512, and the output dimension of the second fully connected layer is 10; the softmax layer adopts a softmax function.

3. The radar bandwidth variation robust target recognition method based on selective kernel attention according to claim 2 is characterized in that: The spatial pyramid pooling layer includes three scales of pooling, each scale of pooling adopts the average pooling strategy, and the number of spatial boxes of the three scales of pooling is respectively 、 and .

4. The radar bandwidth variation robust target recognition method based on selective kernel attention according to claim 1 is characterized in that: The selective kernel attention network training process includes: Perform Gaussian initialization on the parameters of each layer of the constructed selective kernel attention network; Obtaining a training data set and segmenting the data sets corresponding to different bandwidths in the training data set to obtain batch data for each bandwidth; Each batch of data after segmentation is sequentially input into the initialized network, and the preliminary features are extracted through the convolutional neural network structure; wherein the feature dimension of the preliminary features is , Determined by the dimension of the network’s input data; Inputting bandwidth information corresponding to the input batch data into the bandwidth information branch to obtain a bandwidth information branch output vector; The preliminary features and the bandwidth information branch output vector are input into the selective kernel attention layer to obtain the fine features after multi-scale feature fusion. The feature dimension of the fine features is ; The fine features are input into the spatial pyramid pooling layer and pooled at three different scales to obtain dimensions of 、 、 The features of , and the three features are spliced together to obtain the feature dimension The output of the spatial pyramid pooling layer; The output of the spatial pyramid pooling layer is flattened into a dimension of The one-dimensional feature of is input into the fully connected layer structure to obtain the predicted value of the training sample in the input batch data; After batch data of all bandwidths are input into the network, the back-propagation algorithm is used to optimize the total loss function of all bandwidths in a hybrid training, and the parameters of each layer of the network are iteratively updated until the network converges or reaches the set maximum number of training rounds, obtaining a trained selective kernel attention network.

5. The radar bandwidth variation robust target recognition method based on selective kernel attention according to claim 1 or 4, characterized in that: The process of generating the training data set includes: For each bandwidth, extract The radar high-resolution range image data of each target type constitutes a data set corresponding to the bandwidth, and the bandwidth information of the data set is recorded; wherein, the number of radar high-resolution range image data of each type of target in each bandwidth is at least 800; Will All data sets obtained by the bandwidth are combined into an echo data set, all bandwidth information corresponding to the echo data set are combined into a bandwidth information data set, and the echo data set and the bandwidth information data set are combined into a merged data set; Setting category labels for the radar high-resolution range image data in the merged data set to obtain a training data set; wherein, .

6. The radar bandwidth variation robust target recognition method based on selective kernel attention according to claim 5 is characterized in that: The total loss function of all bandwidths in the hybrid training is: in, Indicates the total bandwidth contained in the training set, Indicates the number of samples in each bandwidth batch after segmentation, represents the number of categories of targets in the training set, represents the cross entropy loss function, Indicates the Bandwidth batch The predicted value of the training samples, Indicates the Bandwidth batch The label value of the training sample, Represents a sum operation.

Images