A multi-view ISAR vehicle target classification method based on space-time feature enhancement network

Abstract

The application discloses a multi-view ISAR vehicle target classification method based on a space-time feature enhancement network and belongs to the field of inverse synthetic aperture radar target classification. Aiming at the problems existing in the ISAR vehicle target classification, the application provides a multi-view ISAR vehicle target classification method based on a space-time feature enhancement network. Firstly, the method provides a multi-view feature extraction framework which fuses a dynamic convolution operation and an adaptive view angle weighting strategy; the dynamic convolution can adaptively adjust a convolution kernel according to input; and then, based on the adaptive view angle weighting strategy, the comprehensive view angle features of the vehicle target are extracted in combination with multiple view angles; then, a double-flow space-time feature refining module is provided; the multi-view space-time features are optimized from the global and local angles respectively; finally, the long short-term memory network is used to effectively aggregate the multi-view space-time features, so that the vehicle target classification is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of inverse synthetic aperture radar (ISAR) target classification, and particularly relates to a multi-view ISAR vehicle target classification method based on a space-time feature enhancement network. BACKGROUND

[0002] Inverse synthetic aperture radar can capture rich information such as the size, trajectory and number of wheels of a vehicle target, and compared with optical and infrared sensors, in addition to the advantages such as all-weather and all-day work inherent in traditional radar, ISAR working in the millimeter wave band has high bandwidth and large antenna array, and can provide excellent range, Doppler and azimuth resolution, and thus becomes an important detection system in intelligent transportation. Compared with synthetic aperture radar (SAR) applied in the field of static target monitoring, ISAR has a significant advantage in processing moving targets. The complex motion process of an ISAR target can be regarded as a combination of translational motion and rotational motion, and by accurately estimating and compensating the translational motion of the target, the target can be moved to the turntable coordinate to directly perform turntable target imaging. Vehicle target classification is one of the important downstream tasks of car radar echoes, and it has extremely important academic value and wide application prospect to study an accurate and efficient vehicle target classification method based on ISAR images.

[0003] In recent years, the deep learning technology which is continuously developed, as one of the important means for ISAR image interpretation, greatly promotes the progress of ISAR vehicle target classification. However, ISAR imaging has a non-cooperative characteristic, and the changes of the motion direction of the target, radar bandwidth and distance, azimuth angle sampling rate and the like will affect the effect of ISAR imaging, resulting in unknown deformation problems such as stretching, compression and rotation of the image. Therefore, how to extract the robust features of the vehicle target from the ISAR images with unknown deformation and realize high-precision vehicle target classification is still a topic to be further studied. In order to solve the above problems, the application proposes a multi-view space-time feature refinement network to realize the vehicle target classification task for the time-series ISAR images with unknown deformation. SUMMARY

[0004] The present application is directed to the problems existing in the above ISAR vehicle target classification, and proposes a multi-view ISAR vehicle target classification method based on a space-time feature enhancement network. First, the method proposes a multi-view feature extraction framework that fuses dynamic convolution operation and adaptive view weighting strategy. The dynamic convolution can adaptively adjust the convolution kernel according to the input, and then based on the adaptive view weighting strategy of the multi-view feature, the comprehensive view features of the vehicle target are extracted from multiple views; then, a double-flow space-time feature refinement module is proposed, which optimizes the multi-view space-time features from the global and local angles respectively; finally, the multi-view space-time features are effectively aggregated by using the long short term memory (LSTM) network, so as to realize the classification of the vehicle target.

[0005] The technical scheme of the present application is a multi-view ISAR vehicle target classification method based on a space-time feature enhancement network, which comprises the following steps: step 1: given a vehicle ISAR sequence image I={I0,...,I i ,...,I M}∈R M×W×H , wherein M represents the length of the sequence image, i.e. the number of frames, I i represents the i-th frame image, WxH represents the size of a single frame image, and then the ISAR image size is normalized;

[0006] Step 2: establish a multi-view feature extraction framework that fuses dynamic convolution operation and adaptive view weighting strategy, which comprises two parts of multi-view feature extraction and multi-view feature weighted fusion;

[0007] Step 3: use a double-flow space-time feature refinement module to extract and refine the spatial features;

[0008] Step 4: the time refinement is divided into global and local branches;

[0009] Step 5: use LSTM to aggregate the view, spatial and temporal features; based on the memory unit of the LSTM, the hidden state of the cell state of the current time step input is used to calculate the cell state and the hidden state of the next time step respectively;

[0010]

[0011]

[0012] wherein fg(·), ig(·), og(·) respectively represent the forget gate, input gate and output gate of the LSTM, w c and b crespectively represent the weight and bias parameters of the memory cell, and t(·) represents a Tanh activation function;

[0013] Step 6: According to the formula in step 5, the feature of the aggregated multi-view information at the last time step is obtained by reasoning step by step, and finally the vehicle target is classified by using a Softmax classifier, and the process is as follows:

[0014]

[0015] Label(I)=Max(p1,...,p K )

[0016] where F o is the aggregated feature output by the LSTM at the last time step, w i ∈R C×K and b i are the weight and bias parameters of the Softmax classifier respectively, K is the number of vehicle target categories, Max(·) is a maximum value function, and p i represents the classification value of the i-th vehicle.

[0017] Further, the specific method of step 2 is:

[0018] Step 2.1: Obtain the dynamic convolution kernel parameters by input feature learning, and then perform dynamic convolution operation on the input feature, and the dynamic convolution operation is:

[0019] F i =σ(W T (I i )I i +b(I i ))

[0020]

[0021]

[0022] where each frame in the feature sequence represents a view, F i represents the i-th frame mixed feature, i.e., the i-th view, ω c represents the dynamic weight of the c-th convolution kernel, C is the number of convolution kernels, ω c satisfies 0≤ω c ≤1, σ(·) represents a Sigmoid function, T represents a transpose operation, W c and b c represent the weight and bias of the c-th convolution kernel, respectively.

[0023] Step 2.2: learn the importance weight of each view in the feature sequence F = {F0,..., F i ,...,F M} using a fully connected layer and a Sigmoid function, and the learning process of the importance weight γ i is as follows:

[0024]

[0025] wherein, represents a fully connected layer with parameter w;

[0026] Step 2.3: weight and aggregate all view features according to the importance weight obtained in step 2.2 to obtain the global view feature F g :

[0027]

[0028] wherein, F g is consistent in size with the single view feature F i ;

[0029] Step 2.4: connect the single view feature and the global view feature to obtain the multi-view joint feature sequence F f :

[0030] F f = {concat(F1,F g ),...,concat(F M ,F g )}

[0031] wherein, concat(·,·) represents a feature concatenation operation.

[0032] Further, the specific method of step 3 is as follows:

[0033] Step 3.1: spatial feature refinement uses a spatial feature refinement method based on a self-attention mechanism. First, a three-layer convolutional layer with a 1×1 convolutional kernel is used to learn F i f , and three features K∈R N×C , Q∈R C×N , and V∈R C×N are obtained after each convolution, wherein N = H × W; then K∈R N×C and Q∈R C×N are convoluted to obtain the global attention weight map M G :

[0034] M G = δ(Conv(K)·Conv(Q))

[0035] wherein, Conv(·) denotes a convolution operation with a 1x1 kernel;

[0036] Step 3.2: According to the global attention weight map M calculated above G , V e R C×N Element-wise product is performed to obtain a global spatial refinement feature containing global spatial information

[0037]

[0038] wherein, is the Hadamard product;

[0039] Step 3.3: In order to ensure that the local spatial information is not lost, a skip connection operation is introduced to concatenate the local spatial features in the channel dimension to obtain the globally and locally refined spatial features As shown below:

[0040]

[0041] Further, the specific method of step 4 is:

[0042] Step 4.1: The global branch of the time refinement learns adaptive convolution kernel parameters with a global view through two fully connected layers, dynamically learns the long-term correlation between ISAR images in the continuous imaging interval, and the convolution kernel Θ of the cth channel c According to the following formula:

[0043]

[0044]

[0045] wherein, denotes a global average pooling operation, denotes a RELU activation function, and w2 and w1 are learnable parameters of the two fully connected layers;

[0046] Step 4.2: The local branch of the time feature refinement module focuses on the contribution of each time step feature to the identification task, and learns the importance weight β of each time step through two one-dimensional convolution layers i , the calculation formula is as follows:

[0047]

[0048] wherein, Conv1d(·) denotes a one-dimensional convolution operation;

[0049] Step 4.3: The importance weight β of step 4.2 is used to i the feature of 4.1 Adaptive time convolution is performed to obtain time refinement features As follows:

[0050]

[0051] In summary, the application can obtain the following beneficial effects:

[0052] The application establishes a multi-view ISAR vehicle target classification model based on a space-time feature enhancement network. First, a multi-view joint feature extraction framework based on dynamic convolution and a multi-view adaptive weighting strategy is adopted, multi-view features are extracted using dynamic convolution operation, and multi-view invariant features of the vehicle target are extracted by combining the multi-view adaptive weighting strategy, thereby effectively alleviating the influence of multiple different views on the classification effect. Through a double-flow space-time feature refinement network, the time and space features of the sequence ISAR image can be efficiently extracted, and the space-time features are optimized from the global and local levels respectively, and the features with strong discrimination are focused on. Finally, the LSTM module is used to aggregate the view, time and space features, and all information is comprehensively utilized to stably and effectively realize the time-series ISAR vehicle target classification task. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 The figure is the overall framework of the application;

[0054] Figure 2 The figure is the tau-SNE visualization result of the application and other two methods in two experimental scenes. DETAILED DESCRIPTION

[0055] In order to better show the technical key points of the application, the invention content will be described in detail below in combination with the algorithm framework of the application. The application proposes a multi-view ISAR vehicle target classification method based on a space-time feature enhancement network, and the algorithm framework of the method is as shown in the figure. Figure 1 The specific implementation process of the application is as follows:

[0056] Step 1: Given the vehicle ISAR sequence image I={I0,…,I M}∈R M×W×H , wherein M represents the length of the sequence image, i.e. the number of frames, and WxH represents the size of a single frame image. All ISAR images are cropped to 388x388 pixels in size;

[0057] Step 2: Establish a multi-view feature extraction framework that integrates dynamic convolution operation and adaptive view weighting strategy, which is divided into two parts: multi-view feature extraction and multi-view feature weighting fusion;

[0058] Step 2.1: learn the dynamic convolution kernel parameters through the input features, and then perform dynamic convolution operation on the input features, and the dynamic convolution operation is defined as:

[0059] F i = σ(W T (I i )I i +b(I i ))

[0060]

[0061]

[0062] where ω c c represents the dynamic weight of the c-th convolution kernel, C is the number of convolution kernels, ω c c satisfies 0≤ω c ≤1, σ(·) represents a Sigmoid function, T represents a transpose operation, W c and b c represent the weight and bias of the c-th convolution kernel, respectively;

[0063] Step 2.2: Use the fully connected layer and the Sigmoid function to adaptively learn the importance weight of each view in the feature sequence F = {F0,..., F M}, where each frame in the feature sequence represents a view. The importance weight learning process is represented as follows:

[0064]

[0065] where σ(·) represents a Sigmoid function, represents a fully connected layer with parameter w, F i represents the i-th frame of the mixed feature, i.e., the i-th view;

[0066] Step 2.3: According to the importance weight obtained in step 2.2, weight and aggregate all view features to obtain the global view feature F g :

[0067]

[0068] where F g is consistent in size with the single view feature F i ;

[0069] Step 2.4: Connect the single view feature and the global view feature to obtain the multi-view joint feature sequence F f :

[0070] F f = {concat(F1,F g ),...,concat(F M ,F g )}

[0071] where concat(·,·) denotes the feature concatenation operation;

[0072] Step 3: The extraction and optimization of spatial features and temporal features are performed by using a double-flow spatio-temporal feature refinement module, which is divided into two parts: spatial feature refinement and temporal feature refinement;

[0073] Step 3.1: The spatial feature refinement module uses a spatial feature refinement method based on a self-attention mechanism. First, a three-layer convolutional layer with a kernel size of 1×1 is used to process F i f learning, obtaining K∈R N×C ,Q∈R C×N ,V∈R C×N three features, where N=H×W. Then K∈R N×C and Q∈R C×N perform convolution operation to obtain the global attention weight map M G , as follows:

[0074] M G =δ(Conv(K)·Conv(Q))

[0075] where Conv(·) denotes the convolution operation with a kernel size of 1×1, and δ(·) denotes the Softmax function;

[0076] Step 3.2: According to the global attention weight map M G calculated above, element-wise multiplication is performed on V∈R C×N to obtain the global spatial refinement feature containing global spatial information

[0077]

[0078] where is the Hadamard product;

[0079] Step 3.3: To ensure that local spatial information is not lost, a skip connection operation is introduced to concatenate the local spatial features in the channel dimension to obtain the globally and locally refined spatial features , as follows:

[0080]

[0081] where concat(·,·) denotes the feature concatenation operation;

[0082] Step 4: The temporal refinement module has a similar architecture to the spatial refinement module, which is divided into two parts: global and local branches;

[0083] Step 4.1: The global branch of the time refinement module learns adaptive convolution kernel parameters with a global view through two fully connected layers, dynamically learns the long-term correlation between ISAR images in the continuous imaging interval, and the convolution kernel of the c-th channel is calculated according to the following formula:

[0084]

[0085]

[0086] wherein, denotes a global average pooling operation, denotes a RELU activation function, δ(·) denotes a Softmax function, and w2 and w1 are learnable parameters of the two fully connected layers;

[0087] Step 4.2: The local branch of the time feature refinement module focuses on the contribution of each time step feature to the identification task, and learns the importance weight of each time step through two one-dimensional convolution layers. The calculation formula is as follows:

[0088]

[0089] wherein, Conv1d(·) denotes a one-dimensional convolution operation;

[0090] Step 4.3: Using the importance weight β i of step 4.2 to perform adaptive time convolution on the features of 4.1 to obtain time refinement features as follows:

[0091]

[0092] Step 5: Using LSTM to aggregate perspective, spatial and temporal features. Based on the memory unit of LSTM, the hidden state of the previous time step and the input of the current time step are used to respectively calculate the cell state and the hidden state of the next time step.

[0093]

[0094]

[0095] wherein, fg(·), ig(·), og(·) respectively denote the forget gate, input gate and output gate of LSTM, w c and b c respectively denote the weight and bias parameters of the memory unit, and t(·) denotes a Tanh activation function.

[0096] Step 6: According to the formula in step 5, the features of the aggregated multi-view information at the last time step are obtained. Finally, the vehicle target is classified by using a Softmax classifier, and the process is as follows:

[0097]

[0098] Label(I)=Max(p1,…,p K )

[0099] where F o is the aggregated feature output by the LSTM at the last time step, w i ∈R C×K and b i are the weight and bias parameters of the Softmax classifier, respectively, K is the number of vehicle target categories, and Max(·) is the maximum value function.

[0100] The application will be described below in conjunction with examples:

[0101] I. Experimental conditions

[0102] The simulation experiment uses a short-range automobile radar simulation ISAR image database to evaluate the performance of the algorithm. The ISAR image dataset consists of five vehicle targets, each including four turning modes and east, west, south, and north starting directions, simulating four signal-to-noise ratios and four clutter conditions. The application is based on time-series ISAR images for vehicle target classification. In the experiment, the sliding window method is used to obtain ISAR sequence image samples, and the sequence length is set to 2, 3, and 4, respectively, and the sliding window step is set to 1. In the simulation experiment, the 0db signal-to-noise ratio images and 2.5m / s clutter wind speed images in the mixed ISAR image dataset are used for model training and verification, and are tested under the conditions of signal-to-noise ratio of 0dB, +5dB, +10dB, and clutter wind speed of 2.5m / s, 5m / s, 7.5m / s, and 10m / s. In the first experimental scenario, left-turn data is used for training and verification, and straight, right-turn, and U-turn data are used for testing. In the second experimental scenario, straight data is used for training and verification, and left-turn, right-turn, and U-turn data are used for testing.

[0103] The simulation experiment adopts a stochastic gradient descent (SGD) optimization algorithm to train the model proposed in the application, the batch size is set to 16, the initial learning rate is 0.001, the momentum parameter is 0.9 to accelerate the SGD weight update speed, the L2 regularization parameter is set to 0.00001 to prevent network overfitting, and the iteration number is 50. The simulation environment and platform are a Win10 64-bit computer system, the GPU model is NVIDIA GeForce RTX 3090Ti, and the vehicle target classification algorithm proposed in the application is implemented using the Python programming language based on the PyTorch framework.

[0104] II. Experimental content and results

[0105] Experiment 1: Evaluate the vehicle target classification performance of the application.

[0106] The simulation experiment evaluates the classification effectiveness of the algorithm proposed in the application based on the ISAR image database simulated by the short-range automotive radar. The experiment compares three commonly used target classification algorithms for processing sequence images, namely the Convolution LSTM network (ConvLSTM), the Double layers bidirectional LSTM (DLBi-LSTM), and the Time convolutional network (TCN), and uses a 4-layer convolutional layer as a baseline method for comparison. Tables 1 to 3 are the comparison experiment results under the first experimental scenario with sequence lengths of 2, 3, and 4, respectively, and Tables 4 to 6 are the comparison experiment results under the second experimental scenario with sequence lengths of 2, 3, and 4, respectively.

[0107] The experimental results of Table 1 and Table 2 show that the classification accuracy of the algorithm of the application is higher than that of the other four methods when the sequence length is 2 and 3. As can be observed from Table 3, the algorithm of the application also has a high classification accuracy when the sequence length is 4, showing the best vehicle target recognition performance. By comparing the comparative experimental results of the two experimental scenes, it is not difficult to observe that the overall classification accuracy of various methods has decreased in the second experimental scene. The main reason is that the model is trained and verified only using straight-line data in the second experimental scene, and cannot learn the dynamic changes of the vehicle target in the turning condition. According to the experimental results of Table 4 and Table 5, the classification accuracy of the algorithm of the application is stably higher than that of the other comparative methods when the sequence length is 2 and 3 in the second experimental scene. When the sequence length is 4, the classification accuracy of the algorithm of the application is higher than that of the other five methods when the signal-to-noise ratio is +5dB, +10dB, and the clutter wind speed is 2.5m / s, 7.5m / s and 10m / s. Therefore, the comparative experimental results verify that the algorithm of the application has significant superiority and robustness compared with the existing commonly used sequence image classification algorithms.

[0108] Experiment two: verify the effectiveness of each sub-module in the algorithm of the application.

[0109] In order to verify the effectiveness of each sub-module in the algorithm of the application, the ablation experiment is performed on the short-range automobile radar simulation ISAR image database. In order to ensure the uniformity of the ablation experiment setting, the ablation experiment is performed in the first experimental scene, and the sequence image sample length is set to 3. Here, the multi-view feature joint extraction framework and the double-flow spatio-temporal feature refinement module in the algorithm of the application are marked as module 1 and module 2 respectively, and the ablation experiment results are shown in Table 7. According to the results in Table 7, module 1 and module 2 can improve the time sequence ISAR vehicle target classification performance to different degrees.

[0110] Experiment three: separable feature embedding space visualization.

[0111] In order to further illustrate the classification performance of the algorithm of the application, the simulation experiment uses the feature visualization method τ-Distributed stochastic neighbor embedding (t-SNE) in Python to perform dimension reduction and visualization on the feature space learned by the algorithm, so as to more intuitively show the classification effect of the algorithm on various vehicle targets. Figure 2 The t-SNE visualization results of the application, DLBLSTM and TCN in the two experimental scenes are shown in Figures 6 and 7, and different colors represent different vehicle targets. As can be observed from Figures 6 and 7, the algorithm of the application can effectively distinguish different vehicle targets in the two experimental scenes. Figure 2It can be known that in the comparative classification method, there is a mixed situation between different types of vehicles, and by observing the feature space distribution of the application, it is found that the internal target of each type of vehicle is compact, the different types of vehicles are far apart and the boundary is clear. Therefore, the algorithm proposed in the application has the best vehicle target classification performance compared with other methods.

[0112] Table 1 Comparative test results of sequence length 2 under the first experimental scene

[0113]

[0114] Table 2 Comparative test results of sequence length 3 under the first experimental scene

[0115]

[0116] Table 3 Comparative test results of sequence length 4 under the first experimental scene

[0117]

[0118] Table 4 Comparative test results of sequence length 2 under the second experimental scene

[0119]

[0120] Table 5 Comparative test results of sequence length 3 under the second experimental scene

[0121]

[0122] Table 6 Comparative test results of sequence length 4 under the second experimental scene

[0123]

[0124] Table 7 Ablation experiment results

[0125]

Claims

1. A multi-view ISAR vehicle target classification method based on a spatio-temporal feature enhancement network, the method comprising: Step 1: Given the ISAR sequence images of a vehicle I = {I0,..., I i ,...,I M} ∈ R M×W×H , where M represents the length of the sequence images, i.e., the number of frames, I i represents the i-th frame image, W x H represents the size of a single frame image, and then the ISAR image size is normalized; Step 2: establishing a multi-view feature extraction framework that fuses dynamic convolution operation and adaptive view weighting strategy, comprising two parts of multi-view feature extraction and multi-view feature weighted fusion; Step 3: using a double-flow spatio-temporal feature refinement module to extract and refine spatial features; Step 4: using a double-flow spatio-temporal feature refinement module to perform time refinement, which is divided into global and local branches; Step 5: using LSTM to aggregate view, spatial and temporal features; LSTM-based memory cell that utilizes the hidden state from the previous time step cell state current time step input computes the cell state for the next time step and hidden state wherein fg(·), ig(·), og(·) represent the forget gate, input gate and output gate of the LSTM respectively, w c and b c represent the weight and bias parameters of the memory cell respectively, t(·) represents the Tanh activation function; Step 6: according to the formula of step 5, reasoning at each time step to obtain the feature of the aggregated multi-view information at the last time step; finally, using a Softmax classifier to classify the vehicle target, the process being as follows: Label(I) = Max(p1,..., p K ) where F o is the aggregated feature of the last time step output of LSTM, w i ∈ R C×K and b i are the weight and bias parameters of the Softmax classifier, K is the number of vehicle target classes, Max(·) is the max function, and p i represents the classification value of the i-th vehicle class. 2.The multi-view ISAR vehicle target classification method based on the space-time feature enhanced network of claim 1, wherein, The specific method of step 2 is: Step 2.1: learning dynamic convolution kernel parameters through input feature learning, and then performing dynamic convolution operation on the input feature, the dynamic convolution operation being: F i = σ(W T (I i )I i +b(I i )) wherein each frame in the feature sequence represents a view, F i denotes the i-th frame mixed feature, i.e., the i-th view, ω c denotes the dynamic weight of the c-th convolution kernel, C is the number of convolution kernels, ω c satisfies 0≤ω c ≤1, σ(·) denotes a Sigmoid function, T denotes a transpose operation, W c and b c respectively denote the weight and bias of the c-th convolution kernel; Step 2.2: Learning the importance weight of each view in the feature sequence F = {F0,..., F i ,...,F M} using a fully connected layer and a Sigmoid function. The learning process of the importance weight γ i is as follows: wherein, denotes a fully connected layer with parameter w; Step 2.3: Weighted aggregation of all view features according to the importance weights obtained in Step 2.2 to obtain global view features F g : wherein F g with a single view feature F i are of uniform size; Step 2.4: Concatenate the single-view features and the global-view features to obtain the multi-view joint feature sequence F f : F f = {concat(F1, F g ),...,concat(F M , F g )} Wherein, concat(·,·) represents a feature concatenation operation. 3.The multi-view ISAR vehicle target classification method based on the space-time feature enhanced network of claim 1, wherein, The specific method of step 3 is: Step 3.1: Spatial feature refinement uses a spatial feature refinement method based on a self-attention mechanism. First, a three-layer convolution kernel with a size of 1x1 is used to obtain K∈R i f learning, and then K∈R N×C ,Q∈R C×N ,V∈R C×N three features, where N = HxW; then K∈R N×C and Q∈R C×N convolution operation to obtain a global attention weight map M G : M G = δ (Conv (K) Conv (Q)) Wherein, Conv(·) represents a convolution operation with a 1×1 convolution kernel; Step 3.2: Compute the global attention weight map M according to the above calculation G , V e R C×N Perform element-wise multiplication to obtain the global spatial refinement feature containing global spatial information wherein is the Hadamard product; Step 3.3: To ensure no local spatial information is lost, a skip connection operation is introduced to concatenate the local spatial features in the channel dimension to obtain the global and local refined spatial features As shown below:

4. The multi-view ISAR vehicle target classification method based on the space-time feature enhancement network according to claim 1, wherein, The specific method of step 4 is: Step 4.1: Time-refined global branch learns adaptive convolution kernel parameters with global view through two fully connected layers, dynamically learns the long-term correlation between ISAR images in the continuous imaging interval, and the convolution kernel Θ c The following formula is used for calculation: wherein, represents a global average pooling operation, represents a RELU activation function, w2 and w1 are learnable parameters of two fully connected layers; Step 4.2: The local branch of the time feature refinement module focuses on the contribution of each time step feature to the identification task, and learns the importance weight β of each time step through two one-dimensional convolution layers i The calculation formula is as follows: Wherein, Conv1d(·) represents a one-dimensional convolution operation; Step 4.3: Utilize the importance weights β of Step 4.2 i Features of Step 4.1 Perform adaptive temporal convolution to obtain temporally refined features As follows:

Images