Radar low-altitude target identification method based on multi-dimensional feature nonlinear coupling network

Abstract

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

Description

Technical Field

[0001] This application relates to the field of radar target intelligent recognition technology, and in particular to a radar low-altitude target recognition method based on a multi-dimensional feature nonlinear coupling network. Background Technology

[0002] With the diversification of low-altitude flying targets such as drones and holsters, the environmental complexity and identification difficulty faced by radar low-altitude target identification have significantly increased. The core of radar target identification is to accurately distinguish different types of targets by extracting effective features from the target echo. Its identification accuracy directly affects the effectiveness of applications such as low-altitude air defense and airspace management.

[0003] Existing radar target recognition methods have the following shortcomings: First, traditional feature extraction methods rely on manual design, making it difficult to capture complex high-dimensional nonlinear features in radar echoes, resulting in limited recognition accuracy for low-altitude small targets; second, existing network models do not adequately mine the channel and spatial correlation of features, failing to effectively highlight key information in the echoes and exhibiting weak anti-interference capabilities; third, single decision results are easily affected by noise and environmental interference, lacking robustness and failing to fully utilize multi-source information and historical trajectory consistency to improve recognition stability. Summary of the Invention

[0004] This application provides a radar low-altitude target identification method based on a multi-dimensional feature nonlinear coupling network to solve the following technical problem: how to improve the identification accuracy, anti-interference capability, and robustness of radar low-altitude targets.

[0005] In a first aspect, embodiments of this application provide a radar low-altitude target recognition method based on a multi-dimensional feature nonlinear coupling network. The method includes: inputting a range-Doppler (RD) image into a trained nonlinear coupling network to obtain the predicted probability of at least one candidate category 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 predicted probability of the candidate category based on the candidate category obtained through the radar cross-section (RCS) value of the target; and determining the target category as 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 performs 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 performs channel splitting and parallel feature extraction on the high-level feature map to obtain a fused feature that combines static and dynamic information. The feature modeling module performs long-range spatiotemporal dependency capture on the fused feature to obtain globally correlated features. The prediction module performs classification prediction on the globally correlated features to obtain the predicted probability of at least one candidate category corresponding to the target.

[0006] Secondly, embodiments of this application also provide a radar low-altitude target identification device based on a multidimensional feature nonlinear coupling network. The device includes: at least one processor; and a memory communicatively connected to 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 a radar low-altitude target identification method based on a multidimensional feature nonlinear coupling network as described in the first aspect above.

[0007] Thirdly, embodiments of this application also provide a computer storage medium storing computer-executable instructions, which, when executed, implement a radar low-altitude target identification method based on a multi-dimensional feature nonlinear coupling network as described above.

[0008] The radar low-altitude target identification method based on a multi-dimensional feature nonlinear coupling network provided in this application has the following beneficial effects: In this embodiment, the range-Doppler RD image can be input into a trained nonlinear coupled network to obtain the predicted probability of at least one candidate category corresponding to the target to be identified. Then, based on the candidate category obtained through the target's RCS value, the predicted probability of the candidate category is adjusted, and finally, the target category is determined as the candidate category with the highest predicted probability. Thus, using a nonlinear coupled network to identify RD images can achieve accurate radar identification with strong anti-interference capabilities. Furthermore, using the candidate category obtained based on the target's RCS value for auxiliary decision-making can improve the robustness of the identification. Attached Figure Description

[0009] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart of a radar low-altitude target identification method based on a multi-dimensional feature nonlinear coupling network provided in this application embodiment; Figure 2 This is a schematic diagram of the overall structure of a nonlinear coupled network provided in an embodiment of this application; Figure 3 The RCS frequency histogram is calculated from the raw echo data of various targets provided in this application embodiment. Figure 4 This is a schematic diagram of a decision-level fusion structure provided in an embodiment of this application; Figure 5 This application provides a network output classification accuracy histogram for four types of targets; Figure 6 A schematic diagram of a network output normalized confusion matrix provided in an embodiment of this application; Figure 7 A scatter plot of output features provided for an embodiment of this application (visualized in t-SNE manner); Figure 8 This application provides a decision-level fusion output classification accuracy histogram for four types of targets; Figure 9 A schematic diagram of a decision-level fusion output normalized confusion matrix provided in an embodiment of this application; Figure 10 A scatter plot of decision-level fusion output features provided in an embodiment of this application; Figure 11 This is a schematic diagram of the internal structure of a radar low-altitude target identification device based on a multi-dimensional feature nonlinear coupling network, provided as an embodiment of this application. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0011] This application provides a scheme for radar low-altitude target identification based on a multi-dimensional feature nonlinear coupling network. The technical solution proposed in this application will be described in detail below with reference to the accompanying drawings.

[0012] Figure 1 This is a flowchart illustrating a radar low-altitude target identification method based on a multi-dimensional feature nonlinear coupling network, provided as an embodiment of this application. Figure 1 As shown in the figure, the radar low-altitude target identification method based on a multi-dimensional feature nonlinear coupling network provided in this application embodiment specifically includes the following steps: Step 101: Input the distance-Doppler RD image into the 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.

[0013] The target is the object to be identified in the RD image.

[0014] In this embodiment, a range-Doppler (RD) image acquired by radar can be input into a trained nonlinear coupled network to obtain the predicted probability of at least one candidate category corresponding to the object to be identified in the RD image output by the nonlinear coupled network. For example, for a certain RD image, if the candidate categories of the target are light drone, small drone, bird, and balloon, inputting the RD image into the nonlinear coupled network yields prediction probabilities of 0.8, 0.1, 0.05, and 0.05, respectively. This allows for a simple and convenient method to predict the category of the target to be identified in the RD image.

[0015] Step 102: Adjust the predicted probability of the candidate category based on the candidate category obtained through the RCS value of the target.

[0016] In this embodiment, the predicted probability of the target's candidate categories can be adjusted using the candidate categories obtained from the target's RCS value. Specifically, the candidate categories of the target obtained from the RCS value in the RD image can be acquired. Then, the predicted probability of at least one candidate category corresponding to the target obtained through the nonlinear coupling network can be adjusted using these candidate categories. For example, if the candidate category of the target obtained from the RCS value is a light drone, the predicted probabilities of light drones, small drones, birds, and balloons corresponding to this target can be adjusted to 0.85, 0.08, 0.035, and 0.035, respectively. The specific adjustment method is not limited. This allows for adjustment of the predicted probabilities of candidate categories, increasing the predicted probabilities of some candidate categories.

[0017] Step 103: Determine the target category of the target as the candidate category with the highest predicted probability.

[0018] In this embodiment, the candidate category with the highest predicted probability can be used as the target category. Thus, by combining the nonlinear coupled network and the target's radar cross-section (RCS) value, the predicted probabilities of candidate categories are obtained, and the candidate category with the highest predicted probability is used as the target category. This effectively handles the relationships between multiple categories, avoids indecisiveness among multiple categories, and makes the classification results more explicit and consistent. In this way, determining the target category through multi-source information can improve the problem of insufficient robustness in single-frame decision-making.

[0019] In this embodiment, 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 performs 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 performs channel splitting and parallel feature extraction on the high-level feature map to obtain a fused feature that combines static and dynamic information. The feature modeling module performs long-distance spatiotemporal dependency capture on the fused feature to obtain globally correlated features. The prediction module performs dimensionality reduction and classification prediction on the globally correlated features to obtain the predicted probability of at least one candidate category corresponding to the target.

[0020] In this embodiment, a multi-level cascaded network structure is introduced into the nonlinear coupling network, which gives the extracted radar echo features richer structured information and nonlinear representation capabilities, significantly improving the recognition rate of low-altitude small targets. A feature decoupling and fusion module achieves decoupled extraction of two types of features, and a radar self-attention module captures long-range correlations, enhancing the model's robustness under conditions such as target micro-movement and complex scenes. Therefore, the recognition of RD images using the aforementioned nonlinear coupling network achieves higher accuracy and stronger anti-interference capabilities.

[0021] In this embodiment, the range-Doppler RD image can be input into a trained nonlinear coupled network to obtain the predicted probability of at least one candidate category corresponding to the target to be identified. Then, based on the candidate category obtained through the target's RCS value, the predicted probability of the candidate category is adjusted, and finally, the target category is determined as the candidate category with the highest predicted probability. Thus, using a nonlinear coupled network to identify RD images can achieve accurate radar identification with strong anti-interference capabilities. Furthermore, using the candidate category obtained based on the target's RCS value for auxiliary decision-making can improve the robustness of the identification.

[0022] In one possible implementation, 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 extracts features from the initial feature map to obtain mid-layer features; the nonlinear coupled attention module enhances the mid-layer features to obtain deep features. In practical applications, the input RD image can be processed through a STEM convolutional layer for initial feature extraction, such as... Figure 2 As shown, the stem convolutional layer applies a two-dimensional convolution operation with stride 2 and padding 3, and the convolution kernel... K stem The size is 7, so the RD image X RD Mapped to an initial feature map of 32 channels F stem Then, the initial feature map can be... F stem The input is fed into a multi-level cascaded improved ConvNeXt module and a non-linear coupled attention module, such as a four-level cascaded improved ConvNeXt module. In the first layer structure, the improved ConvNeXt module can be used to extract features from the initial feature map to obtain the mid-layer features. F convnext This allows for enhanced feature extraction capabilities while maintaining manageable computational overhead. In this case, a non-linearly coupled attention module can be used to extract the aforementioned mid-level features. F convnext Perform feature enhancement to obtain deep features F high This allows for multi-dimensional nonlinear coupled attention calibration to highlight key signals in radar echo images and enhance the interaction between channel and spatial features through nonlinear transformation.

[0023] ConvNeXt is a pure convolutional neural network architecture, which can be directly translated into Chinese as "next-generation convolutional network" or "new generation convolutional network". The improved ConvNeXt module mentioned above is an improvement on the existing ConvNeXt.

[0024] In one possible implementation, the improved ConvNeXt module performs feature extraction on 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.

[0025] 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:

[0026] 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.

[0027] 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.

[0028] 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.

[0029] In one possible implementation, the nonlinear coupled attention module performs feature enhancement on 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 features.

[0030] In the above embodiments, the input mid-level features are first processed. Perform channel attention calculation: This is done using global average pooling. F Compress to CThe 1×1 vector is then passed through two 1×1 convolutions (the first layer is followed by ReLU activation, and the second layer is followed by Sigmoid activation) to obtain the channel attention weight vector. This channel attention vector can automatically learn and adjust the response intensity of each channel.

[0031] Simultaneously, positional attention of the feature map is obtained through a spatial attention branch: first, on the channel dimension... F Max pooling and average pooling are performed separately to obtain two 1×H×W feature maps; these two maps are concatenated and passed through a 7×7 convolutional layer (followed by Sigmoid normalization) to obtain the spatial attention map. This spatial attention focuses on highlighting spatial locations in the feature map that are more relevant to the classification task.

[0032] Finally, channel attention CA ( F Spatial attention SA ( F Simultaneously applied to the original feature F Above: First, combine the original features with... CA ( F Multiply by each channel, then multiply by the product. SA ( F The features are multiplied position by position. Next, to further enhance feature fusion, a nonlinear coupling branch can be introduced. A multilayer perceptron (MLP) is used to nonlinearly calibrate the channel- and spatial attention-weighted features to obtain the final output feature map. F out Overall, it can be represented as:

[0033] in, This indicates element-wise multiplication. NCA [ [] is a nonlinear coupling branch that further enhances the interaction between channel and spatial attention through a multilayer perceptron, resulting in more complex coupling features.

[0034] In one possible implementation, 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.

[0035] In the above embodiments, to enhance the network's sensitivity to target motion features, this application employs a feature decoupling and fusion module to decouple morphological features from motion features. In this module, the input feature map is equally divided into two groups of channels: one group is fed into the "morphological branch," and the other into the "motion branch." The morphological branch uses a standard 3×3 convolution kernel to convolve the feature map to extract static contours and textures of the target; the motion branch uses a 1×7 convolution kernel to convolve the feature map along the Doppler direction to specifically extract the target's motion (Doppler) features. Each branch generates (… C / 2)× H × W The feature maps are finally concatenated along the channel dimension to restore the original data. C × H × W The output of this parallel branch structure achieves morphological-motion decoupling in the feature dimension, avoiding mutual interference between the two, thereby improving the network's ability to model and identify the micro-motion spectrum information of flying targets.

[0036] In one possible implementation, 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.

[0037] By enhancing nonlinearity through feedforward networks and combining residual connections with layer normalization, global correlation features are obtained.

[0038] In the above embodiments, a feature modeling module is introduced into the nonlinear coupled network to supplement global spatiotemporal context information. This module is based on a multi-head self-attention mechanism and can capture long-range dependencies and correlation features between targets. Specifically, let the input fused feature map be... First, reconstruct it into a shape. N × C The sequence, where N = H × W(This involves flattening the spatial location). Then, layer normalization (LayerNorm) is performed on the sequence, and a query matrix is ​​generated through three linear transformations. Key matrix Sum matrix ,in d This is the in-head dimension. Next, we use scaled dot product attention to calculate the attention weights:

[0039] In practical applications, to improve expressive power, a multi-head mechanism is used, which divides the input embedding into different weight matrices. h Each head calculates attention and outputs its own output. head i Mathematically, this is represented as:

[0040] in, W i Q For the first i The query projection matrix of the size W i K For the first i The key projection matrix of the head, W i K For the first i The projection matrix of the values ​​of each head.

[0041] Then, concatenate the outputs of all heads and multiply by the weight matrix. W O This enables multi-head output. Mathematically, it can be represented as: Finally, a nonlinear mapping is performed on the representation of each position through a feedforward network (FFN), and the output of the Transformer module is obtained by using residual connections and normalization layers.

[0042] Feedforward network FFN The mathematical expression for (X) is:

[0043] The output of the feature modeling module X 3 can be represented as:

[0044] In this way, the feature modeling module enables the nonlinear coupled network to focus on the relationship between distant targets in space, incorporate relevant information of distant targets into the judgment, and improve the ability to distinguish complex radar echoes.

[0045] In one possible implementation, the prediction module performs dimensionality reduction and classification prediction on globally correlated features, including: performing global average pooling dimensionality reduction on the globally correlated features, transforming the two-dimensional sequence features into a one-dimensional feature vector. The one-dimensional feature vector is then... X vec The input consists of two fully connected layers. The first layer introduces non-linearity through the ReLU activation function, while the second layer directly outputs the predicted probability of the candidate class of the target.

[0046] In one possible implementation, during the training process of the non-network mentioned above, the network parameters are optimized using a cross-entropy loss function weighted by class recall. First, the recall rate of each class of target in the training set is calculated and normalized into a weight assigned to the prediction score of the corresponding class. Then, the cross-entropy loss is calculated, and all learnable parameters of the network are iteratively updated through backpropagation.

[0047] In one possible implementation, 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.

[0048] In the above embodiments, for each trajectory point's radar echo set, in addition to the depth network output, this application can also use the radar cross section (RCS) for auxiliary decision-making. RCS is the effective reflective area of ​​an object against radar waves, reflecting the radar echo intensity of the target. The formula for calculating the radar cross section is:

[0049] in, P r For radar receiving power, P t For radar transmission power, R For the target distance, λ The wavelength of the emitted electromagnetic wave G For antenna gain, L This represents the total loss.

[0050] Since parameters such as transmit power, antenna gain, and total loss are the same in each frame of data, in practical applications, these parameters can be assumed to be constants during the calculation process. By correcting the radar echo energy calculation formula based on target distance and wavelength, a simplified RCS calculation formula can be obtained:

[0051] In practical applications, the target's distance minus the square of the Doppler cell amplitude in the RD diagram is used as the radar received echo energy for RCS calculation. The specific steps are as follows: Step 1: Read the target global range cell index from the original echo file, and extract the echo data of 31 range cells, 15 range cells before and after the target.

[0052] Step 2: Window the extracted echo data to suppress spectral leakage.

[0053] Step 3: Perform a Fourier transform on the slow time dimension to obtain the RD diagram.

[0054] Step 4: Set the Doppler center ±3 channels to zero to eliminate fixed clutter.

[0055] Step 5: Locate the index of the maximum value of the RD map within ±5 distance cells of the approximate target location to determine the precise location of the target.

[0056] Step 6: To avoid the influence of the window function on the RCS value calculation, the target echo energy is calculated using the RD map without windowing, and the echo energy is obtained by dividing the square of the amplitude at the precise location of the target by the number of Doppler dimension points.

[0057] Multiply the target echo energy by 4 Dividing the cube of the distance to the target and the fourth power of the distance to the target by the square of the wavelength yields an estimate of the target's RCS, which is then combined with the predicted probability obtained from the network. P k Calculate the fusion weights after the decision r k .

[0058] In the above embodiments, to fully utilize the radar cross-section (RCS) information in the radar echo signal for target identification, the RCS values ​​of each type of target are first statistically analyzed, and the RCS value distribution of each category is modeled by fitting a Rayleigh distribution. The distribution of RCS values ​​typically follows a Rayleigh distribution, especially when the target reflects radar waves under random phase and amplitude conditions. Based on the probability density function of the Rayleigh distribution, the RCS values ​​are known. r The classification probability density function at time t can be mathematically expressed as:

[0059] in, r RCS value σ represents the scale parameter of the Rayleigh distribution. By statistically analyzing the RCS values ​​of each target class, the fitted probability density function can describe the RCS distribution characteristics of each target category.

[0060] Using the probability density function fitted by the Rayleigh distribution, a Naive Bayes classifier based on the probability density function is further employed to classify the RCS value of each radar image frame. Specifically, given the RCS value of a target in a radar image... r The goal of the Naive Bayes classifier is to calculate which category the object belongs to. k The posterior probability. P ( k | r This can be expressed using Bayes' theorem as:

[0061] in, P ( k ) is a category k The prior probability, representing the class k Frequency in the dataset p k ( r ) is a category k Given an RCS value r The probability density function is given by the given time. By maximizing the posterior probability, the Naive Bayes classifier can select the class with the highest posterior probability as the classification result. The classification decision can be formalized as:

[0062] 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:

[0063] 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) FNk ( 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.

[0064] 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: like R k If the confidence level is less than 0.6, a penalty is imposed on the network prediction probability of the RCS-indicating category, for example, reducing it to 0.8 of the original probability. 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.

[0065] like R k ≥ If the confidence level is 0.9 (high confidence level), the network prediction probability for that category is multiplied by 1.2 to increase its weight.

[0066] After the above operations, the adjusted probability vector is exponentially normalized (softmax) to obtain the adjusted probability distribution of each frame of the RD image. q i,k .

[0067] In one possible implementation, after adjusting the predicted probability of the candidate category based on the recall rate, the method further includes: 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.

[0068] In practical applications, the RD image may be the first few frames acquired, especially the first frame. In this case, recognition can be performed using only a nonlinear coupled network and RCS-assisted prediction, ensuring accuracy. However, for subsequent images, to improve the robustness of recognition, multi-source information can be combined.

[0069] 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 sum of the weights of each category; then, the highest probability category in each frame is recorded as a vote, i.e., for each category... arg max( q i Add 1 to the category of ) to obtain the final RCS fusion weight. r k : Among them, 1( ) is an indicator function, which is 1 when the condition is true and 0 otherwise.

[0070] Considering the consistency of information along the target trajectory, a sliding window strategy is adopted to fuse historical decisions. Let the size of the historical window be... N Currently processing the first t A number of trajectory points. Maintain the final category sequence within the past window. n (most N (Number of categories), counting the frequency of each category in history. f k The historical weight decay coefficient is defined as:

[0071] in, K Let be the decay rate constant, set to 0.15 based on the number of track points in the training set. The denominator is used for normalization to ensure that when... t When very big α =1. The formula means: when just past the window size, α The trajectory is relatively small, mainly relying on information from the current point; as the trajectory grows, α The index is increasing, relying more on historical trends.

[0072] In practical applications, historical and current information can be merged, such as the current trajectory point index. t <N( (If the window size is insufficient), the RCS fusion weights are used directly as the final weights. t≥N At that time, the fusion formula simultaneously considers historical frequency and current weight: for each class k Based on experience, historical frequencies are magnified tenfold, and the current contribution... r k The weights are then calculated by merging the categories proportionally:

[0073] in, f k For historical frequencies, α ∈[0,1] represents the historical contribution weight ratio. r k This is the RCS fusion weight for the current frame.

[0074] Will W k Perform softmax normalization again to obtain the final decision probability distribution:

[0075] The overall structure of the above-mentioned optimization of single-frame recognition results through multi-source information fusion is as follows: Figure 4 As shown, this can improve robustness.

[0076] In summary, this algorithm adds the final category to the historical queue (first-in, first-out, maintaining maximum length) after processing each trajectory point. N This algorithm effectively combines "voting" with "probability," leveraging trajectory consistency and category information to improve decision robustness.

[0077] To make the above technical solution clearer, the method is explained in detail below with specific testing cases. The measured data is the original radar echo, which is processed by time and frequency to obtain a radar range-Doppler (RD) two-dimensional image with a size of 31×360×1 (where 31 is the number of range sampling points and 360 is the number of Doppler sampling points).

[0078] The four-class recall rates obtained using the nonlinear coupled network were 0.9021, 0.9293, 0.8677, and 0.8887, corresponding to the four target classes: "light drones, small drones, birds, and balloons," respectively. The specific network output results are as follows... Figure 5 , Figure 6 , Figure 7 As shown.

[0079] Among them, t-SNE stands for t-Distributed Stochastic Neighbor Embedding, which is explained in Chinese as t-distributed random neighborhood embedding.

[0080] The four-category recall rates obtained after decision-level fusion were 0.9068, 0.9224, 0.8796, and 0.9225, corresponding to the four target categories of "light drones, small drones, birds, and balloons," respectively. The specific output results of the decision-level fusion are as follows... Figure 8 , Figure 9 , Figure 10 As shown.

[0081] In summary, this application, starting from two levels—network structure design and decision fusion algorithm—achieves efficient radar identification of various small targets through detailed feature extraction and multi-source information integration.

[0082] The above are embodiments of the method proposed in this application. Based on the same inventive concept, embodiments of this application also provide a radar low-altitude target identification device based on a multi-dimensional feature nonlinear coupling network, the structure of which is as follows: Figure 11 As shown.

[0083] Figure 11 This is a schematic diagram of the internal structure of a device provided in an embodiment of this application. Figure 11 As shown, the device includes: At least one processor 1101; And a memory 1102 that is communicatively connected to at least one processor; The memory 1102 stores instructions that can be executed by at least one processor. The instructions are executed by at least one processor 1101 so that at least one processor 1101 can: execute the radar low-altitude target identification method based on a multi-dimensional feature nonlinear coupling network.

[0084] Some embodiments of this application provide corresponding to Figure 1 A non-volatile computer storage medium stores computer-executable instructions, which are configured to execute the aforementioned radar low-altitude target identification method based on a multi-dimensional feature nonlinear coupling network.

[0085] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for IoT devices and media are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0086] The systems, media, and methods provided in this application are one-to-one correspondences. Therefore, the systems and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the systems and media will not be repeated here.

[0087] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied 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.

[0088] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0089] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0090] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0091] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0092] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0093] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, 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 technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0094] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0095] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this 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.

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. The method according to claim 1, characterized in that, The step of adjusting the predicted probability of the candidate categories based on the candidate categories 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.

8. The method according to claim 7, characterized in that, After adjusting the predicted probabilities of the candidate categories, the method further includes: 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.

9. 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 1-8.

10. 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 1-8.

Images