A radar low-altitude target recognition method based on a multi-dimensional feature nonlinear coupling network

Abstract

The application discloses a radar low-altitude target identification method based on a multi-dimensional feature nonlinear coupling network, and relates to the technical field of intelligent radar target identification. The method comprises the following steps: inputting a range-Doppler (RD) image into a trained nonlinear coupling network to obtain a prediction probability of at least one candidate category corresponding to a target; adjusting the prediction probability of the candidate category according to a candidate category obtained through a radar cross section (RCS) value of the target, and determining that a target category of the target is a candidate category with the maximum prediction probability. The nonlinear coupling network comprises a feature extraction enhancement module, a feature decoupling fusion module, a feature modeling module and a prediction module. In this way, the identification precision, the anti-interference capability and the robustness of the radar low-altitude target can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of radar target intelligent identification, and particularly relates to a radar low-altitude target identification method based on a multi-dimensional feature nonlinear coupling network. BACKGROUND

[0002] With the diversified development of low-altitude flying targets such as unmanned aerial vehicles and air balls, the complexity of the environment and the difficulty of identification faced by radar low-altitude target identification are significantly improved. The core of radar target identification is to extract effective features in target echoes to accurately distinguish different types of targets, and the identification accuracy directly affects the effectiveness of applications such as low-altitude defense and control and airspace management.

[0003] The existing radar target identification method has the following disadvantages: first, the traditional feature extraction method relies on manual design and is difficult to capture complex high-dimensional nonlinear features in radar echoes, and the identification accuracy of low-altitude small targets is limited; second, the existing network model is insufficient in channel and spatial correlation mining of features, and cannot effectively highlight the key information in the echoes, and the anti-interference ability is weak; third, the single decision result is easily affected by noise and environmental interference, and the robustness is insufficient, and the identification stability is not improved by fully utilizing multi-source information and historical trajectory consistency. SUMMARY

[0004] The radar low-altitude target identification method based on the multi-dimensional feature nonlinear coupling network provided in the embodiments of the present application is used to solve the technical problems of how to improve the identification accuracy, anti-interference ability and robustness of radar low-altitude targets.

[0005] In a first aspect, the embodiments of the present application provide a radar low-altitude target identification method based on a multi-dimensional feature nonlinear coupling network, which comprises: inputting a range-Doppler (RD) image into a trained nonlinear coupling network to obtain a prediction probability of at least one candidate class corresponding to a target output by the nonlinear coupling network, wherein the target is an object to be identified in the RD image; adjusting the prediction probability of the candidate class according to a candidate class obtained through a radar cross section (RCS) value of the target; determining a target class of the target as a candidate class with the maximum prediction probability; wherein the nonlinear coupling network comprises: a feature extraction enhancement module, a feature decoupling fusion module, a feature modeling module and a prediction module, the feature extraction enhancement module is used to perform feature mapping and feature enhancement on the RD image through a multi-level cascaded network structure to obtain a high-level feature map; the feature decoupling fusion module is used to perform channel splitting and parallel feature extraction on the high-level feature map to obtain a fusion feature with static and dynamic information; the feature modeling module is used to capture long-distance spatiotemporal dependence of the fusion feature to obtain a global correlation feature; and the prediction module is used to perform classification prediction on the global correlation feature to obtain the prediction probability of at least one candidate class corresponding to the target.

[0006] In a second aspect, the embodiments of the present application further provide a radar low-altitude target recognition device based on a multi-dimensional feature nonlinear coupling network, the device comprising: at least one processor; and a memory in communication connection with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the radar low-altitude target recognition method based on the multi-dimensional feature nonlinear coupling network according to the first aspect.

[0007] In a third aspect, the embodiments of the present application further provide a computer storage medium storing computer executable instructions, and the computer executable instructions are executed to implement the radar low-altitude target recognition method based on the multi-dimensional feature nonlinear coupling network according to any one of the above aspects.

[0008] The radar low-altitude target recognition method based on the multi-dimensional feature nonlinear coupling network provided by the embodiments of the present application has the following beneficial effects:

[0009] In the embodiments of the present application, the range-Doppler (RD) image can be input into the trained nonlinear coupling network to obtain the prediction probability of at least one candidate class corresponding to the target to be recognized, and then the prediction probability of the candidate class is adjusted according to the candidate class obtained through the RCS value of the target, so that the target class of the target is finally determined as the candidate class with the maximum prediction probability. In this way, the RD image is recognized by using the nonlinear coupling network, which can realize accurate and strong anti-interference radar recognition, and the candidate class obtained based on the RCS value of the target is used for auxiliary decision-making, which can improve the robustness of the recognition. BRIEF DESCRIPTION OF DRAWINGS

[0010] The accompanying drawings, which are included to provide a further understanding of the present application, constitute a part of the present application and illustrate the illustrative embodiments of the present application and their description serve to explain the present application, and do not constitute improper limitations on the present application. In the drawings:

[0011] Figure 1 A radar low-altitude target recognition method based on a multi-dimensional feature nonlinear coupling network provided by the embodiments of the present application is shown in a flowchart;

[0012] Figure 2 A general structure diagram of a nonlinear coupling network provided by the embodiments of the present application is shown;

[0013] Figure 3 An RCS frequency histogram obtained by calculating the original echo data of each classified target provided by the embodiments of the present application is shown;

[0014] Figure 4 A decision-level fusion structure diagram provided by the embodiments of the present application is shown;

[0015] Figure 5 A network output classification precision column chart on four types of targets provided for an embodiment of the present application;

[0016] Figure 6 A network output normalized confusion matrix diagram provided for an embodiment of the present application;

[0017] Figure 7 A scatter plot of output features (visualized in a t-SNE manner) provided for an embodiment of the present application;

[0018] Figure 8 A decision-level fusion output classification precision column chart on four types of targets provided for an embodiment of the present application;

[0019] Figure 9 A decision-level fusion output normalized confusion matrix diagram provided for an embodiment of the present application;

[0020] Figure 10 A scatter plot of decision-level fusion output features provided for an embodiment of the present application;

[0021] Figure 11 An internal structure diagram of a radar low-altitude target recognition device based on a multi-dimensional feature nonlinear coupling network provided for an embodiment of the present application. DETAILED DESCRIPTION

[0022] In order to make the objectives, technical solutions and advantages of the present application clearer, the technical solutions of the present application will be described below in detail with reference to the embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0023] The embodiments of the present application provide a radar low-altitude target recognition scheme based on a multi-dimensional feature nonlinear coupling network, and the technical solutions proposed by the embodiments of the present application will be described in detail below with reference to the drawings.

[0024] Figure 1 A radar low-altitude target recognition method based on a multi-dimensional feature nonlinear coupling network provided for an embodiment of the present application. As shown in Figure 1 The radar low-altitude target recognition method based on a multi-dimensional feature nonlinear coupling network provided by the embodiments of the present application specifically includes the following steps:

[0025] Step 101, input a range-doppler (RD) image into a trained nonlinear coupling network to obtain a prediction probability of at least one candidate class corresponding to the target output by the nonlinear coupling network.

[0026] The target is an object to be identified in the RD image.

[0027] In the embodiments of the present application, the range-doppler (RD) image obtained by radar can be input into a trained nonlinear coupling network, and then the prediction probability of at least one candidate class corresponding to the object to be identified in the RD image output by the nonlinear coupling network is obtained. For example, for a certain RD image, the candidate classes of the target are light unmanned aerial vehicle, small unmanned aerial vehicle, bird and balloon, and the prediction probabilities obtained by inputting the RD image into the nonlinear coupling network are 0.8, 0.1, 0.05 and 0.05 respectively. In this way, the prediction of the class of the target to be identified in the RD image can be simple and convenient.

[0028] Step 102, adjusting the prediction probability of the candidate class according to the candidate class obtained by the RCS value of the target.

[0029] In the embodiments of the present application, the prediction probability of the candidate class of the target can be adjusted by the candidate class obtained by the RCS value of the target. That is, the candidate class of the target obtained by the RCS value can be obtained by the RCS value of the target in the RD image, and then the prediction probability of at least one candidate class corresponding to the target obtained by the nonlinear coupling network is adjusted by the candidate class. For example, the candidate class of the target obtained by the RCS value is light unmanned aerial vehicle. Then the prediction probabilities of light unmanned aerial vehicle, small unmanned aerial vehicle, bird and balloon corresponding to the target can be adjusted to 0.85, 0.08, 0.035 and 0.035, and the specific adjustment method is not limited. In this way, the prediction probability of the candidate class can be adjusted to increase the prediction probability of some candidate classes.

[0030] Step 103, determining that the target class of the target is the candidate class with the maximum prediction probability.

[0031] In the embodiments of the present application, the candidate class with the maximum prediction probability can be used as the target class of the target. In this way, the prediction probability of the candidate class is obtained by combining the nonlinear coupling network and the radar cross section (RCS) value of the target, and the candidate class with the maximum prediction probability is used as the target class, which can effectively handle the relationship between multiple classes, avoid hesitation between multiple classes, and make the classification result more clear and consistent. In this way, the target class of the target is determined by multiple source information, which can improve the problem of insufficient robustness of single-frame decision.

[0032] In the embodiment of the present application, the nonlinear coupling network includes a feature extraction enhancement module, a feature decoupling fusion module, a feature modeling module, and a prediction module. The feature extraction enhancement module is configured to perform feature mapping and feature enhancement on the RD image through a multi-stage cascaded network structure to obtain a high-level feature map. The feature decoupling fusion module is configured to perform channel splitting and parallel feature extraction on the high-level feature map to obtain a fusion feature with static and dynamic information. The feature modeling module is configured to capture long-distance space-time dependence of the fusion feature to obtain a global correlation feature. The prediction module is configured to perform dimension reduction and classification prediction on the global correlation feature to obtain a prediction probability of at least one candidate category corresponding to the target.

[0033] In the embodiment of the present application, the multi-stage cascaded network structure is introduced into the nonlinear coupling network, so that the extracted radar echo features have more rich structured information and nonlinear expression capability, and the recognition rate of low-altitude small targets is significantly improved. Through the feature decoupling fusion module, two types of features are decoupled and extracted, and the radar self-attention module is combined to capture long-distance correlation, thereby enhancing the recognition robustness of the model under conditions such as target micro-motion and complex scenes. Therefore, the RD image is recognized by the above nonlinear coupling network, which has higher accuracy and stronger anti-interference capability.

[0034] In the embodiment of the present application, the range-Doppler (RD) image can be input into the trained nonlinear coupling network to obtain a prediction probability of at least one candidate category corresponding to the target to be recognized. Then, the prediction probability of the candidate category is adjusted according to the candidate category obtained by the RCS value of the target, and finally the target category of the target is determined as the candidate category with the maximum prediction probability. In this way, the RD image is recognized by the nonlinear coupling network, which can realize accurate and strong anti-interference radar recognition. At the same time, the candidate category obtained based on the RCS value of the target is used for auxiliary decision-making, which can improve the robustness of recognition.

[0035] In one possible implementation, the feature extraction enhancement module is configured to perform feature mapping and feature enhancement on the RD image through a multi-stage cascaded network structure to obtain a high-level feature map, including:

[0036] An initial feature map is obtained by using a stem convolutional layer to perform initial feature extraction on the RD image.

[0037] A high-level feature map is obtained by using a multi-stage cascaded improved ConvNeXt module and a nonlinear coupling attention module to perform feature extraction and feature enhancement on the initial feature map.

[0038] In the first layer structure, the improved ConvNeXt module performs feature extraction on the initial feature map to obtain a middle layer feature; and the nonlinear coupled attention module performs feature enhancement on the middle layer feature to obtain a deep layer feature.

[0039] In actual application, the input RD image can be subjected to initial feature extraction through a stem convolution layer, as shown in Figure 2 , wherein the stem convolution layer applies a two-dimensional convolution operation with a stride of 2 and padding of 3, and the convolution kernel size is 7. K stem Thus, the RD image is mapped into an initial feature map with 32 channels. X RD F stem Then, the initial feature map can be input into a multi-stage cascaded improved ConvNeXt module and a nonlinear coupled attention module, for example, a four-stage cascaded improved ConvNeXt module. F stem In the first layer structure, the improved ConvNeXt module can be used to perform feature extraction on the initial feature map to obtain a middle layer feature F convnext , so as to enhance the feature extraction capability while keeping the computational overhead controllable. F convnext At this time, the nonlinear coupled attention module can be used to perform feature enhancement on the middle layer feature F high , so as to realize multi-dimensional nonlinear coupled attention calibration to highlight the key signals in the radar echo image, and enhance the interaction between the channel and spatial features through nonlinear transformation.

[0040] ConvNeXt is a pure convolutional neural network architecture, which can be directly translated into Chinese as “Convolution Next Generation Network” or “Convolution New Generation Network”. The improved ConvNeXt module is improved on the basis of the existing ConvNeXt.

[0041] In one possible implementation, the improved ConvNeXt module performs feature extraction on the initial feature map to obtain a middle layer feature, including:

[0042] applying a depth separable convolution to the initial feature map to obtain an output feature map, wherein the depth separable convolution has a convolution kernel size of 1x7;

[0043] performing layer normalization on the output feature map and using two consecutive 1x1 convolutions for nonlinear transformation;

[0044] ​The initial feature map is added element-wise to the transformed output feature map through residual connections to obtain the intermediate-level features.

[0045] In the above embodiments, the core component in each feature extraction stage is the improved ConvNeXt module. This module first processes the input initial feature map. We apply depthwise convolution, where a 1×7 spatial kernel size is used. Since the 1×7 convolution spans only 7 pixels in the Doppler direction (while maintaining the same size in the distance direction), it has a large lateral receptive field and low computational cost. Formally, the 1×7 depthwise convolution can be represented as:

[0046]

[0047] in, X dw The feature map is output by depthwise convolution. W dw ( c , k ) is a channel c Displacement of the high-depth convolution kernel in the Doppler direction k The weights are determined by the time. This convolution operation preserves the number of channels. c This remains unchanged, corresponding to performing one-dimensional convolution independently for each channel.

[0048] After depthwise convolution, the output feature map is normalized along the channel direction using LayerNorm, and then two consecutive 1×1 convolutions (equivalent to a channel-wise fully connected layer) are applied with the GELU activation function for non-linear transformation. This step first increases the dimensionality of the feature channels and then reduces it, thereby enhancing the non-linear expressive power of the network.

[0049] Finally, residual connections can be used to connect the input. X The transformed feature map is added element-wise to form the final output of ConvNeXt. This residual structure keeps the information flow of the network smooth, effectively alleviates the gradient vanishing problem in deep networks, and enhances the feature extraction capability in a wide field of view.

[0050] In one possible implementation, the nonlinear coupled attention module performs feature enhancement on the mid-level features to obtain deep features, including:

[0051] Channel attention is calculated on the mid-layer features to obtain channel attention weights;

[0052] Spatial attention is calculated on the mid-layer features to obtain spatial attention weights;

[0053] applying the channel attention weight and the spatial attention weight to the middle layer feature;

[0054] using a multi-layer perceptron to perform nonlinear calibration on the weighted middle layer feature to obtain a deep layer feature.

[0055] In the above embodiment, first, the input middle layer feature is subjected to channel attention calculation: compressed into a vector of F ×1×1 through global average pooling, and then sequentially passed through two 1×1 convolution layers (the first layer followed by ReLU activation, and the second layer followed by Sigmoid activation) to obtain a channel attention weight vector C . This channel attention vector can automatically learn to adjust the response intensity of each channel.

[0056] At the same time, the position attention of the feature map is obtained through the spatial attention branch: first, respectively perform maximum pooling and average pooling on F to obtain two feature maps of 1×H×W; after splicing the two maps, pass through a 7×7 convolution layer (followed by Sigmoid normalization) to obtain a spatial attention map . This spatial attention highlights the spatial positions in the feature map that are more important for the classification task.

[0057] Finally, the channel attention CA ( F ) and the spatial attention SA ( F ) are simultaneously applied to the original feature F : first, multiply the original feature with CA ( F ) channel by channel, and then multiply with SA ( F ) position by position; then, in order to further enhance the fusion of the features, a nonlinear coupling branch can be introduced, and a multi-layer perceptron (MLP) is used to perform nonlinear calibration on the features weighted by the channel and spatial attention to obtain the final output feature map F out . Overall, it can be represented as:

[0058]

[0059] wherein, represents element-wise multiplication. NCA [ ] is a nonlinear coupling branch, which further enhances the interaction of channel and spatial attention through a multi-layer perceptron to obtain more complex coupled features.

[0060] ​In a possible implementation, the feature decoupling fusion module is configured to perform channel splitting and parallel feature extraction on the high-level feature map to obtain fusion features with static and dynamic information, including:

[0061] The high-level feature map is equally divided into two groups of channels in the channel dimension, one group is sent to the shape branch, and the other group is sent to the motion branch;

[0062] In the shape branch, a 3x3 size convolution kernel is used to extract the shape features of the target;

[0063] In the motion branch, a 1x7 size convolution kernel is used to extract the dynamic features of the target along the Doppler direction;

[0064] The shape features and the dynamic features are spliced along the channel dimension to obtain fusion features.

[0065] In the above embodiment, in order to enhance the sensitivity of the network to the motion features of the target, the feature decoupling fusion module is used to decouple and process the shape features and the motion features. In this module, the input feature map is equally divided into two groups of channels in the channel dimension: one group is sent to the "shape branch", and the other group is sent to the "motion branch". The shape branch uses a standard 3x3 convolution kernel to perform convolution operation on the feature map to extract the static contour and texture and other shape information of the target; the motion branch uses a 1x7 convolution kernel to perform convolution on the feature map along the Doppler direction to extract the motion (Doppler) features of the target. Two branches each generate a feature map of C / 2)× H × W The output is restored to C × H × W This parallel branch structure realizes the shape-motion decoupling in the feature dimension, avoids the mutual interference between them, and thus improves the modeling and recognition ability of the network to the micro-motion spectrum information of the flying target.

[0066] In a possible implementation, the feature modeling module is configured to capture long-distance space-time dependence on the fusion features to obtain global correlation features, including:

[0067] The fusion feature map is flattened into a two-dimensional sequence, and the spatial dimension features are converted into a sequence form that can be calculated and associated;

[0068] Layer normalization is performed on the two-dimensional sequence, and three independent linear transformations are used to generate query, key, and value matrices, respectively;

[0069] The query, key, value matrix is split by the number of attention heads, each head independently calculates the correlation weight between features, and the scaling dot product attention is used to avoid large weight values, and then the calculation results of all heads are spliced, and the multi-head attention features are output through linear mapping.

[0070] The global correlation features are obtained by enhancing the nonlinearity through the feedforward network, combining residual connection and layer normalization.

[0071] In the above embodiment, a feature modeling module is introduced in the nonlinear coupling network to supplement the global spatio-temporal context information. This module is based on the multi-head self-attention mechanism, which can capture long-distance dependencies and correlation features between targets. Specifically, let the input fusion feature map be First, it is reconstructed into a sequence with a shape of N × C , where N = H × W (i.e., the spatial position is flattened). Then, the sequence is subjected to layer normalization (LayerNorm), and query matrix , key matrix and value matrix are generated through three linear transformations, where d is the intra-head dimension. Next, the scaling dot product attention is used to calculate the attention weight:

[0072]

[0073] In practical applications, in order to improve the expression ability, the multi-head mechanism is used to divide the input embedding into h heads through different weight matrices, and each head calculates the attention and outputs head i , which is mathematically represented as:

[0074]

[0075] where W i Q is the query projection matrix of the i th head, W i K is the key projection matrix of the i th head, W i K is the value projection matrix of the i th head.

[0076] Then, the outputs of all heads are spliced and multiplied by the weight matrix W OThe multi-head output is completed. It can be mathematically expressed as:

[0077] Finally, the representation of each position is nonlinearly mapped through a feed-forward network (FFN), and a residual connection and a normalization layer are adopted to obtain the output of the Transformer module.

[0078] Feed-forward network FFN The mathematical expression of (X) is:

[0079]

[0080] The output of the feature modeling module is X 3 can be expressed as:

[0081]

[0082] In this way, the feature modeling module enables the nonlinear coupling network to pay attention to the relationship between distant targets in space, and incorporates the relevant information of the distant targets into the judgment, thereby improving the discrimination ability for complex radar echoes.

[0083] In one possible implementation, the prediction module includes the following steps of dimension reduction and classification prediction on the global correlation feature: performing global average pooling dimension reduction on the global correlation feature, and converting the two-dimensional sequence feature into a one-dimensional feature vector. The one-dimensional feature vector is X vec Two fully connected layers are input, the first layer introduces nonlinearity through a ReLU activation function, and the second layer directly outputs the prediction probability of the candidate class of the target.

[0084] In one possible implementation, in the training process of the above non-network, a cross-entropy loss function based on class recall rate weighting is used to optimize the network parameters, the recall rate of each class of target in the training set is calculated first, and then the prediction score of the corresponding class is normalized to the weight, and then the cross-entropy loss is calculated. All learnable parameters of the network are updated through backpropagation iteration.

[0085] In one possible implementation, the prediction probability of the candidate class is adjusted according to the RCS value of the target, including:

[0086] Based on the radar echo data, the RCS value of the target is calculated;

[0087] The RCS value of each target is fitted with a Rayleigh distribution, and a naive Bayes classifier is used to obtain an RCS auxiliary classification result and a recall rate corresponding to the auxiliary classification result;

[0088] According to the recall rate, the prediction probability of the candidate class is adjusted.

[0089] In the above embodiment, for each set of radar echoes of a track point, in addition to the deep network output, the present application can also assist in decision-making by means of radar cross section (Radar Cross Section, RCS). RCS is the effective reflection area of an object to radar waves, reflecting the radar echo intensity of the target. The formula for calculating the radar cross section is:

[0090]

[0091] wherein, P r is the radar receiving power, P t is the radar transmitting power, R is the target distance, λ is the wavelength of the transmitted electromagnetic wave, G is the antenna gain, L is the total loss.

[0092] Since the parameters such as transmitting power, antenna gain and total loss are the same in each frame of data, in actual application, these parameters can be assumed to be constants in the calculation process, and the radar echo energy calculation formula is corrected by the target distance and the wavelength, to obtain the simplified RCS calculation formula:

[0093]

[0094] In actual application, the square of the distance-Doppler unit amplitude of the target in the RD diagram is used as the radar receiving echo energy for RCS calculation. The specific steps are as follows:

[0095] Step 1: Read the global distance unit index of the target from the original echo file, and intercept 31 distance unit echo data of the target and the 15 distance units before and after the target.

[0096] Step 2: Window the intercepted echo data to suppress spectral leakage.

[0097] Step 3: Perform Fourier transform on the slow time dimension to obtain the RD diagram.

[0098] Step 4: Set the Doppler center ±3 channels to zero to eliminate fixed clutter.

[0099] Step 5: Find the index of the maximum value of the RD diagram amplitude within ±5 distance units of the target approximate position to determine the accurate position of the target.

[0100] Step 6: To avoid the influence of the window function on the RCS value calculation, use the RD diagram without windowing to calculate the target echo energy, and divide the square of the amplitude at the accurate position of the target by the number of Doppler dimension points to obtain the echo energy.

[0101] Multiply the target echo energy by 4 to the 3rd power of the target range and the 4th power of the target range, divided by the square of the wavelength, to get the estimate of the target RCS, and combine the prediction probability obtained from the network P k , calculate the fusion weight after decision r k .

[0102] In the above embodiment, in order to make full use of the radar cross section (RCS) information in the radar echo signal for target recognition, first, the RCS values of each type of target are counted, and the RCS value distribution of each category is modeled by fitting the Rayleigh distribution. The distribution of RCS values usually presents Rayleigh distribution, especially when the target reflects radar waves under the condition of random phase and amplitude. According to the probability density function of Rayleigh distribution, the classification probability density function when the RCS value r is mathematically expressed as:

[0103]

[0104] wherein, r is the RCS value, σ is the scale parameter of Rayleigh distribution. By counting the RCS values of each type of target, the probability density function obtained by fitting can describe the RCS distribution characteristics of each target category.

[0105] Using the probability density function obtained by fitting the Rayleigh distribution, a naive Bayes classifier based on the probability density function is further used to classify the RCS value of each frame of radar image. Specifically, given the RCS value r of a target in the radar image, the goal of the naive Bayes classifier is to calculate the posterior probability that the target belongs to a certain category k . The posterior probability P ( k | r ) can be expressed according to Bayes theorem as:

[0106]

[0107] wherein, P ( k ) is the prior probability of the category k , indicating the frequency of the category k in the data set, p k ( r ) is the probability density function of the category k , given the RCS value r . By maximizing the posterior probability, the naive Bayes classifier can select the category with the maximum posterior probability as the classification result. The classification decision can be formalized as:

[0108]

[0109] To further improve the accuracy of RCS classification, this application can divide the distribution of RCS values, dividing the range of RCS values ​​into multiple discrete segments. This division method helps to refine the feature space of RCS values, enabling more accurate classification and recall analysis within different RCS intervals. Let the range of RCS values ​​be divided into... N Each section, section i The corresponding RCS value range is [ r i , r i+1 ), then within each segment, the category k recall rate R k ( i It can be calculated using the following formula:

[0110]

[0111] in, TP k ( i ) indicates category k In the i The number of true cases within each segment (i.e., correctly classified into categories) k (number of samples) FN k ( i ) indicates category k The number of false negatives (i.e., classes misclassified as other classes) within this segment. k (Sample size). This formula calculates the category. k Recall rate in a specific RCS segment.

[0112] By applying the aforementioned RCS estimation algorithm, Naive Bayes classifier, and RCS distribution algorithm, and inputting the raw echo data, the RCS integration result (classification result) of each radar image frame is obtained. R C Recall rate corresponding to RCS distribution R k The RCS frequency histogram is as follows: Figure 3 As shown. When performing decision-level fusion with the classification results obtained from the network, the RCS confidence of the network prediction results is adjusted:

[0113] like R k<0.6 (low confidence), the network prediction probability of the RCS indicating category is penalized, for example, reduced to 0.8 of the original probability, and if the category does not appear in the current trajectory history, the penalty is repeated once to reduce the interference of low-confidence RCS on decision-making.

[0114] If R k ≥ 0.9 (high confidence), the network prediction probability of the category is multiplied by 1.2 to increase its weight.

[0115] After the above operation, the adjusted probability vector is exponentially normalized (softmax) to obtain the adjusted probability distribution of each frame of RD image q i,k .

[0116] In one possible implementation, after the prediction probability of the candidate category is adjusted according to the recall rate, the method further comprises:

[0117] Based on the prediction probability corresponding to at least one frame of historical RD image, an RCS fusion weight is obtained, wherein the historical RD image is an image before the RD image;

[0118] A sliding window strategy is used to count historical decision information, a window size is set, and a historical frequency of the candidate category is calculated;

[0119] The historical frequency and the RCS fusion weight are fused to obtain the final prediction probability of the candidate category.

[0120] In actual application, the RD image may be the first few frames collected, especially the first frame image. In this case, only the nonlinear coupling network and the RCS auxiliary prediction can be used for identification, and the accuracy of identification can be ensured. However, for subsequent images, in order to improve the robustness of identification, multi-source information can be combined for identification.

[0121] In the above embodiment, when calculating the RCS fusion weight, all images can be weighted and aggregated: first, the probability of each frame is accumulated to obtain the weight sum of each category, then each highest probability category of each frame is counted once, that is, the category of arg max( q i ) is added by 1. Finally, the RCS fusion weight r k :

[0122] Wherein, 1( ) is an indicator function, which is 1 when the condition is true, and 0 otherwise.

[0123] Considering the consistency of the target trajectory, the historical decisions are fused by a sliding window strategy. Let the size of the history window be N , and the current processing is the t th trajectory point. The final category sequence n (maximum N ) in the past window is maintained, and the frequency of each category in the history is counted f k . The historical weight decay coefficient is defined as:

[0124]

[0125] where K is the decay rate constant, which is set to 0.15 according to the number of trajectory points in the training set, and the denominator is used for normalization so that t is 1 when α is very large. The meaning of the formula is: when just over the window size, α is small, mainly relying on the current point information; with the growth of the trajectory, α exponentially increases, and more relies on the historical trend.

[0126] In practical applications, historical and current information can be fused. If the current trajectory point index t<N( is still less than the window size), directly use the RCS fusion weight as the final weight. When t≥N , according to the fusion formula, the historical frequency and the current weight are considered at the same time: for each category k , according to experience, the historical frequency is amplified by 10 times, and the current contribution r k According to the proportion, the comprehensive weight of each category is calculated:

[0127]

[0128] where f k is the historical frequency, α ∈[0,1] is the historical contribution weight ratio, r k is the current frame RCS fusion weight.

[0129] Let W k Again, do softmax normalization to get the final decision probability distribution:

[0130]

[0131] The overall structure of the above multi-source information fusion optimization of single-frame recognition result is shown in Figure 4 , which can improve the robustness.

[0132] In summary, the algorithm will add the final category to the history queue (first-in first-out keeps the maximum length N ) after each track point processing is completed. This algorithm effectively combines "voting" and "probability" to improve the robustness of decision-making using track consistency and category information.

[0133] To make the above technical solutions clearer, the above method will be described in detail below in combination with specific detection cases. The measured data is the radar raw echo, and the radar range-Doppler (RD) two-dimensional image is obtained after time-frequency processing, and the size is 31x360x1 (where 31 is the number of distance sampling points, and 360 is the number of Doppler sampling points).

[0134] The four-class recall rates obtained by the nonlinear coupling network are 0.9021, 0.9293, 0.8677, and 0.8887, respectively, corresponding to the four types of targets "light unmanned aerial vehicle, small unmanned aerial vehicle, bird, and balloon". The specific results output by the network are shown in Figure 5 、 Figure 6 、 Figure 7 .

[0135] Among them, the full name of t-SNE in English is t-Distributed Stochastic Neighbor Embedding, which is interpreted as t-distribution random neighborhood embedding in Chinese.

[0136] The four-class recall rates obtained by decision-level fusion are 0.9068, 0.9224, 0.8796, and 0.9225, respectively, corresponding to the four types of targets "light unmanned aerial vehicle, small unmanned aerial vehicle, bird, and balloon". The specific results output by the decision-level fusion are shown in Figure 8 、 Figure 9 、 Figure 10 .

[0137] In summary, the present application starts from two levels of network structure design and decision fusion algorithm, and realizes efficient radar recognition of multiple small targets through detailed feature extraction and multi-source information integration.

[0138] The above is a method embodiment of the present application. Based on the same inventive concept, the present application embodiment also provides a radar low-altitude target recognition device based on a multi-dimensional feature nonlinear coupling network, and the structure is shown in Figure 11 .

[0139] Figure 11 It is a device internal structure schematic diagram provided by the present application embodiment. As shown in Figure 11 , the device comprises:

[0140] at least one processor 1101;

[0141] and a memory 1102 connected with the at least one processor in communication;

[0142] The memory 1102 stores instructions executable by the at least one processor 1101, and the instructions are executed by the at least one processor 1101 to enable the at least one processor 1101 to perform the radar low-altitude target identification method based on the multi-dimensional feature nonlinear coupling network described above.

[0143] Some embodiments of the present application provide a non-volatile computer storage medium corresponding to Figure 1 The computer executable instructions are arranged to perform the radar low-altitude target identification method based on the multi-dimensional feature nonlinear coupling network described above.

[0144] Each of the embodiments in the present application is described in a progressive manner, and the same and similar parts of each of the embodiments can be referred to each other. Each of the embodiments mainly describes the differences from other embodiments. In particular, the IoT device and medium embodiments are basically similar to the method embodiments, and thus are described simply. The relevant parts can be referred to the description of the method embodiments.

[0145] The system and medium provided by the embodiments of the present application are one-to-one corresponding to the method, and thus the system and medium have similar beneficial technical effects as the method. Since the beneficial technical effects of the method have been described in detail above, the beneficial technical effects of the system and medium will not be described here.

[0146] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0147] The present application is described with reference to flowcharts and / or block diagrams of the method, device (system), and computer program product according to the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of the flows and / or blocks can be implemented by computer program instructions. These computer program instructions can be provided to a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to produce a machine, so that the instructions executed by the computer or other programmable data processing devices produce a device that implements the functions described in the flowcharts and / or block diagrams. Figure 1 Each flow or multiple flows and / or blocks Figure 1an apparatus to perform each block or blocks of the flow or flows and / or steps of the function specified in the block or blocks.

[0148] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a Figure 1 an apparatus to perform each block or blocks of the flow or flows and / or steps of the function specified in the block or blocks. Figure 1 an apparatus to perform each block or blocks of the flow or flows and / or steps of the function specified in the block or blocks.

[0149] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions that execute on the computer or other programmable apparatus provide steps for implementing the flow or flows and / or functions specified in the block or blocks. Figure 1 an apparatus to perform each block or blocks of the flow or flows and / or steps of the function specified in the block or blocks. Figure 1 an apparatus to perform each block or blocks of the flow or flows and / or steps of the function specified in the block or blocks.

[0150] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0151] The memory can include non-persistent memory and / or volatile memory, such as random access memory (RAM) about which the computer stores the information. The memory is an example of computer readable media.

[0152] Computer readable media includes permanent and non-permanent, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital versatile discs (DVD) or other optical storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible to a computing device. According to the definition herein, computer readable media does not include transitory media, such as modulated data signals and carrier waves.

[0153] It should also be noted that the terms "comprising," "including," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can also include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises a... " does not, without more constraints, exclude the existence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0154] The above description is merely illustrative of the application, and not restrictive. Various modifications and changes can become apparent to those skilled in the art. Incorporating any modification, equivalent substitution, improvement, etc. within the spirit and principle of the application, shall be included in the scope of the claims of the application.

Claims

1. A radar low-altitude target identification method based on a multi-dimensional feature nonlinear coupling network, characterized in that, include: The distance-Doppler RD image is input into a trained nonlinear coupling network to obtain the predicted probability of at least one candidate category corresponding to the target output by the nonlinear coupling network, wherein the target is the object to be identified in the RD image; The predicted probability of the candidate category is adjusted based on the candidate category obtained by the radar cross-section (RCS) value of the target. The target category of the target is determined to be the candidate category with the highest predicted probability; The nonlinear coupling network includes a feature extraction and enhancement module, a feature decoupling and fusion module, a feature modeling module, and a prediction module. The feature extraction and enhancement module is used to perform feature mapping and feature enhancement on the RD image through a multi-level cascaded network structure to obtain a high-level feature map. The feature decoupling and fusion module is used to perform channel splitting and parallel feature extraction on the high-level feature map to obtain fused features that combine static and dynamic information. The feature modeling module is used to capture the long-distance spatiotemporal dependencies of the fused features to obtain globally related features; The prediction module is used to classify and predict the global association features to obtain the predicted probability of at least one candidate category corresponding to the target. The step of adjusting the predicted probability of the candidate category based on the candidate category obtained through the RCS value of the target includes: The RCS value of the target is calculated based on radar echo data; For each target class, a Rayleigh distribution is fitted to the RCS value, and a Naive Bayes classifier is used to obtain the RCS auxiliary classification result and the recall corresponding to the auxiliary classification result; The predicted probability of the candidate category is adjusted based on the recall rate; Based on the prediction probability corresponding to at least one historical RD image, the RCS fusion weight is obtained, wherein the historical RD image is an image preceding the RD image; A sliding window strategy is used to collect historical decision information, the window size is set, and the historical frequency of the candidate category is calculated. By fusing the historical frequency with the RCS fusion weight, the final predicted probability of the candidate category is obtained.

2. The method according to claim 1, characterized in that, The feature extraction and enhancement module is used to perform feature mapping and feature enhancement on the RD image through a multi-level cascaded network structure to obtain a high-level feature map, including: The initial feature map is obtained by using a Stem convolutional layer to extract initial features from the RD image. The initial feature map is used to extract and enhance features using a multi-level cascaded improved ConvNeXt module and a non-linear coupled attention module to obtain a high-level feature map. In the first layer structure, the improved ConvNeXt module is used to extract features from the initial feature map to obtain intermediate features; the nonlinear coupled attention module is used to enhance the intermediate features to obtain deep features.

3. The method according to claim 2, characterized in that, The improved ConvNeXt module is used to extract features from the initial feature map to obtain mid-level features, including: A depthwise separable convolution is applied to the initial feature map to obtain an output feature map, wherein the kernel size of the depthwise separable convolution is 1×7; The output feature map is layer normalized and then subjected to a nonlinear transformation using two consecutive 1×1 convolutions. The initial feature map is added element-wise to the transformed output feature map through residual connections to obtain the intermediate-level features.

4. The method according to claim 2, characterized in that, The nonlinear coupled attention module is used to enhance the mid-level features to obtain deep features, including: Channel attention is calculated on the mid-layer features to obtain channel attention weights; Spatial attention is calculated on the mid-layer features to obtain spatial attention weights; The channel attention weights and spatial attention weights are applied to the middle-layer features; The weighted mid-layer features are nonlinearly calibrated using a multilayer perceptron to obtain deep-layer features.

5. The method according to claim 1, characterized in that, The feature decoupling and fusion module is used to perform channel splitting and parallel feature extraction on the high-level feature map to obtain fused features that combine static and dynamic information, including: The high-level feature map is divided into two groups of channels in the channel dimension. One group is sent to the morphology branch and the other group is sent to the motion branch. In the morphological branch, a 3×3 convolution kernel is used to extract the morphological features of the target; In the motion branch, a 1×7 convolution kernel is used to extract the dynamic features of the target along the Doppler direction; The morphological features and the dynamic features are spliced ​​along the channel dimension to obtain the fused features.

6. The method according to claim 1, characterized in that, The feature modeling module is used to capture long-distance spatiotemporal dependencies of the fused features to obtain globally related features, including: The fused feature map is flattened into a two-dimensional sequence, transforming spatial features into a computable, correlated sequence. The two-dimensional sequence is layer normalized, and query, key, and value matrices are generated through three independent linear transformations, respectively. The query, key, and value matrix is ​​split according to the number of attention heads. The association weights between features are calculated independently for each head. The dot product attention is scaled to avoid excessively large weight values. The calculation results of all heads are then concatenated and output as multi-head attention features through linear mapping.

7. A radar low-altitude target identification device based on a multi-dimensional feature nonlinear coupling network, characterized in that, The device includes: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform a radar low-altitude target identification method based on a multidimensional feature nonlinear coupling network as described in any one of claims 1-6.

8. A computer storage medium storing computer-executable instructions, characterized in that, When the computer-executable instructions are executed, they implement a radar low-altitude target identification method based on a multi-dimensional feature nonlinear coupling network as described in any one of claims 1-6.

Images