Target comprehensive identification method, device and equipment based on high-dimensional feature spectrum
By using a high-dimensional feature map-based approach, and combining autoencoders and target neural networks with generative adversarial networks, the real-time performance and accuracy issues of traditional radar target recognition algorithms when data scale is insufficient are solved, achieving fast and high-precision target recognition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING AEROSPACE INST OF THE LONG MARCH VEHICLE
- Filing Date
- 2023-03-22
- Publication Date
- 2026-06-23
AI Technical Summary
Traditional radar target recognition algorithms struggle to meet real-time requirements when data volume is insufficient, and feature extraction relies on manual design, which can easily lead to incompleteness. Existing deep learning models are time-consuming to train, making it difficult to meet the rapid recognition needs of practical applications.
We employ a high-dimensional feature map-based approach, which collects observation data through radar sensors, uses an autoencoder for dimensionality reduction and deep feature extraction, combines a target neural network for classification, and introduces a generative adversarial network to generate approximate real feature maps under small sample conditions, thereby improving the sufficiency of training data.
It significantly improves target recognition accuracy, enabling fast and accurate target recognition under small sample conditions, reducing feature extraction loss, and improving the real-time performance and accuracy of recognition.
Smart Images
Figure CN116311067B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of target recognition technology, and in particular to a target comprehensive recognition method, apparatus and device based on high-dimensional feature maps. Background Technology
[0002] With the continuous improvement of ballistic missile target penetration methods, anti-missile target identification technology has become one of the most critical technologies in ballistic missile defense, playing a decisive role in the success or failure of missile defense. The accuracy of ballistic missile target identification affects the performance of each stage, including target early warning, precision tracking, target interception, and kill assessment. Without correct target identification information, accurate threat assessment and warnings cannot be provided, precision tracking is impossible, and without tracking information of the missile target, impact point prediction is impossible, thus hindering effective target interception and kill assessment.
[0003] Traditional radar target recognition can be divided into the following stages: information acquisition, data preprocessing, feature learning, classifier design, and discrimination output.
[0004] The data acquired during the information acquisition phase are divided into two categories: 1) one-dimensional data such as radar cross section (RCS) and high resolution range profile (HRRP); 2) two-dimensional image data such as synthetic aperture radar (SAR) images and inverse synthetic aperture radar (ISAR) images.
[0005] The data preprocessing stage mainly involves noise suppression, clutter suppression, and sensitivity processing of the input data.
[0006] The feature learning stage primarily analyzes and extracts RCS features, broadband features, micro-motion features, and ballistic features from target radar data to obtain information reflecting the target's characteristics, such as shape and size, motion parameters, polarization information, and statistical features. This stage is crucial in the entire identification process, hence many researchers dedicate significant time and effort to research on target feature extraction.
[0007] The classifier design phase primarily determines the classifier parameters based on the form of the features extracted during the feature learning phase. Commonly used classification methods include Bayesian classification, Support Vector Machine (SVM), and BackPropagation (BP) neural networks. Bayesian classification is a posterior probability-based classification technique that requires a large amount of data accumulation in practical applications, making it difficult to implement. SVM performs well in handling small samples and nonlinear high-dimensional problems, but it is sensitive to parameter settings and its performance is not as strong as neural network methods when handling multi-class problems. The BP algorithm classifies new samples by training the network; its drawback is that the model uses gradient descent during training, requiring multiple iterations to update network parameters. Furthermore, this method is prone to getting trapped in local optima and is sensitive to parameters such as the learning rate.
[0008] The effectiveness of traditional radar target recognition depends on the learning of target features and the choice of classifier. While traditional feature extraction methods work well in some situations, these algorithms are shallow and cannot effectively represent targets. Furthermore, these features are manually designed and rely on the researcher's expertise; without sufficient prior knowledge in the relevant field, the extracted features are often incomplete.
[0009] Deep learning algorithms construct a neural network with multiple hidden layers, mapping input data to each hidden layer to obtain feature representations of the input data at each layer. Due to its powerful feature representation capabilities, deep learning has received widespread attention and successful applications in machine learning, pattern recognition, and other fields in recent years. Currently, commonly used deep learning models include Convolutional Neural Networks (CNNs), Deep Belief Networks (DBNs), and Stacked Auto-Encoders (SAEs). These methods have been successfully applied to image processing, face recognition, and speech recognition. However, for radar target recognition, the sheer scale of the data makes it difficult to meet the model's data requirements, and traditional deep learning models have long training times, making it difficult to meet the real-time requirements of practical applications. Therefore, it is necessary to research new algorithms that can effectively extract target-level features while also possessing rapid learning capabilities. Summary of the Invention
[0010] To address the aforementioned technical problems, embodiments of the present invention provide a target comprehensive recognition method based on high-dimensional feature maps, comprising:
[0011] Obtain observational data about the target group based on radar sensor data;
[0012] The observation data is processed, and features are extracted from the processing results to obtain multidimensional feature data about the target group;
[0013] The multidimensional feature data is processed to form a target multidimensional feature vector;
[0014] The target multidimensional feature vector is input into an autoencoder to achieve data dimensionality reduction and deep feature extraction based on the autoencoder, thereby obtaining a target multidimensional feature map;
[0015] The target multidimensional feature map is input into the target neural network to classify the target multidimensional feature map based on the target neural network, thereby achieving target recognition.
[0016] As an optional embodiment, obtaining observation data about the target group based on radar sensor acquisition includes:
[0017] Obtain observational data about a target group based on radar sensors, including flight data, target precession angle, and precession frequency of the target group.
[0018] As an optional embodiment, the processing of the observation data includes:
[0019] Calculate the target precession attitude angle of the target group at each observation time based on the observation data;
[0020] Based on the target precession attitude angle and observation data, the RCS data, HRRP data and narrowband polarization data of the target group during actual flight are calculated and generated.
[0021] As an optional embodiment, the step of extracting features from the processing results to obtain multidimensional feature data about the target group includes:
[0022] Determine the sliding window and sliding step size in the feature extraction module to extract measurement sequences for feature extraction of different types of features in the processing results;
[0023] Extraction algorithms for different types of features are determined, and feature extraction is performed by combining each extraction algorithm with the corresponding measurement sequence to obtain multidimensional feature data about the target group.
[0024] As an optional embodiment, the target group has multiple types of features, and different types of features have corresponding multidimensional feature data; the process of processing the multidimensional feature data to form a target multidimensional feature vector includes:
[0025] The multidimensional feature vector of the target group is formed by merging the multidimensional feature data corresponding to different types of features of the target group.
[0026] As an optional embodiment, the step of inputting the target multidimensional feature vector into an autoencoder to achieve data dimensionality reduction and deep feature extraction of the target multidimensional feature vector based on the autoencoder, thereby obtaining a target multidimensional feature map, includes:
[0027] The target multidimensional feature vector is input into an autoencoder, and the data dimensionality reduction and deep feature extraction of the target multidimensional feature vector are realized based on the multiple intermediate layer structures of the autoencoder to obtain the target multidimensional feature map.
[0028] As an optional embodiment, it also includes:
[0029] The target neural network architecture is trained using the multidimensional feature map of the target group as training data. The maximum number of iterations is set to 50, and the number of iterations is 30. This results in a target neural network used to classify the multidimensional feature map and achieve target recognition. The target neural network convolves the input multidimensional feature map to extract features. After standardization, the extracted feature data is activated by an activation function, then transformed into a one-dimensional vector through a fully connected layer for classification. The probability of each input data point is then obtained through an objective function, and the corresponding target classification result is obtained based on the probability analysis. The objective function includes a softmax function.
[0030] As an optional embodiment, it also includes:
[0031] Construct a generative adversarial network, wherein the generative adversarial network includes a generator and a discriminator;
[0032] The generative adversarial network is trained using a cross-training method to obtain samples that approximate the distribution of high-dimensional feature map samples of the real target group;
[0033] When performing small-sample target recognition based on the target neural network, the target neural network is trained by generating multiple new multidimensional feature maps that approximate the real feature maps based on the generative adversarial network, so that the trained target neural network can accurately perform small-sample target recognition.
[0034] Another embodiment of the present invention also provides a target comprehensive recognition device based on high-dimensional feature maps, comprising:
[0035] The acquisition module is used to acquire observational data about the target group based on radar sensor data.
[0036] The first processing module is used to process the observation data and extract features from the processing results to obtain multidimensional feature data about the target group.
[0037] The second processing module is used to process the multidimensional feature data to form a target multidimensional feature vector.
[0038] The first input module is used to input the target multidimensional feature vector into the autoencoder, so as to realize the data dimensionality reduction and deep feature extraction of the target multidimensional feature vector based on the autoencoder, and obtain the target multidimensional feature map;
[0039] The second input module is used to input the target multidimensional feature map into the target neural network, so as to classify the target multidimensional feature map based on the target neural network and realize target recognition.
[0040] Another embodiment of the present invention also provides a target comprehensive recognition device based on high-dimensional feature map, comprising:
[0041] At least one processor; and,
[0042] A memory communicatively connected to the at least one processor; wherein,
[0043] The memory stores instructions that can be executed by the at least one processor to implement the target comprehensive recognition method based on high-dimensional feature maps as described in any of the embodiments above.
[0044] Based on the disclosure of the above embodiments, it can be understood that the beneficial effects of the embodiments of the present invention include the ability to effectively reduce the loss that occurs during feature extraction, to significantly improve the target recognition accuracy by using a target neural network to identify the target multidimensional feature map, and to introduce a generative adversarial network under small sample conditions to generate a new multidimensional feature map that approximates the real multidimensional feature map based on the generative adversarial network, so that the target neural network still has sufficient training data for training under small sample conditions, ensuring that the trained target neural network can accurately identify the target.
[0045] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings.
[0046] The technical solution of this application will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0047] The accompanying drawings are provided to further illustrate the present application and form part of the specification. They are used together with the embodiments of the present application to explain the application and do not constitute a limitation thereof. In the drawings:
[0048] Figure 1 This is a flowchart of a target comprehensive recognition method based on high-dimensional feature maps in an embodiment of the present invention.
[0049] Figure 2 This is the basic process of the feature extraction module in this embodiment of the invention.
[0050] Figure 3 This is a structural diagram of the self-encoder in an embodiment of the present invention.
[0051] Figure 4 This is a multidimensional feature map of the target group in an embodiment of the present invention.
[0052] Figure 5 This is the confusion matrix of the network recognition results in this embodiment of the invention.
[0053] Figure 6 This is a graph showing the network recognition accuracy in an embodiment of the present invention.
[0054] Figure 7 This is a novel multidimensional feature map in an embodiment of the present invention.
[0055] Figure 8 This is the confusion matrix for the corresponding small sample target recognition in another embodiment of the present invention.
[0056] Figure 9 This is a structural block diagram of the target comprehensive recognition device based on high-dimensional feature map in an embodiment of the present invention. Detailed Implementation
[0057] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings, but these are not intended to limit the scope of the invention.
[0058] It should be understood that various modifications can be made to the embodiments disclosed herein. Therefore, the following description should not be considered as limiting, but merely as an example of embodiments. Other modifications within the scope and spirit of this disclosure will be apparent to those skilled in the art.
[0059] The accompanying drawings, which are included in and form part of this specification, illustrate embodiments of the present disclosure and, together with the general description of the disclosure given above and the detailed description of the embodiments given below, serve to explain the principles of the disclosure.
[0060] These and other features of the invention will become apparent from the following description of preferred forms of embodiments given as non-limiting examples, with reference to the accompanying drawings.
[0061] It should also be understood that although the invention has been described with reference to some specific examples, those skilled in the art can certainly implement many other equivalent forms of the invention, which have the features described in the claims and are therefore all within the scope of protection defined herein.
[0062] The above and other aspects, features and advantages of this disclosure will become more apparent when taken in conjunction with the accompanying drawings and in view of the following detailed description.
[0063] Specific embodiments of the present disclosure are described thereafter with reference to the accompanying drawings; however, it should be understood that the disclosed embodiments are merely examples of the present disclosure, which may be implemented in various ways. Well-known and / or repeated functions and structures are not described in detail to avoid unnecessary or redundant details that could obscure the present disclosure. Therefore, the specific structural and functional details disclosed herein are not intended to be limiting, but merely to serve as the basis and representative basis for the claims to teach those skilled in the art to use the present disclosure in a variety of substantially any suitable detailed structures.
[0064] This specification may use the phrases “in one embodiment,” “in another embodiment,” “in yet another embodiment,” or “in still another embodiment,” all of which may refer to one or more of the same or different embodiments according to this disclosure.
[0065] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0066] like Figure 1 As shown, this embodiment of the invention provides a target comprehensive recognition method based on high-dimensional feature maps, including:
[0067] S101: Obtain observational data about the target group based on radar sensor data;
[0068] S102: Process the observation data and extract features from the processing results to obtain multidimensional feature data about the target group;
[0069] S103: Process multidimensional feature data to form a target multidimensional feature vector;
[0070] S104: Input the target multidimensional feature vector into the autoencoder to achieve data dimensionality reduction and deep feature extraction of the target multidimensional feature vector based on the autoencoder, and obtain the target multidimensional feature map;
[0071] S105: Input the multidimensional feature map of the target into the target neural network to classify the multidimensional feature map of the target based on the target neural network, thereby achieving target recognition.
[0072] The beneficial effects of this embodiment include at least the ability to effectively reduce the loss that occurs during feature extraction, and to significantly improve the target recognition accuracy by using a target neural network to identify the multidimensional feature map of the target.
[0073] Specifically, obtaining observational data about the target group based on radar sensor data includes:
[0074] S106: Obtain observational data about the target group based on radar sensors. The observational data includes the target group's flight data, target precession angle, and precession frequency.
[0075] The observation data is processed, including:
[0076] S107: Calculate the target precession attitude angle of the target group at each observation time based on the observation data;
[0077] S108: Calculate and generate RCS data, HRRP data and narrowband polarization data of the target group during actual flight based on the target precession attitude angle and observation data.
[0078] For example, the target group in this embodiment consists of four types of targets: Indian fire, cone, seamless cone-shaped sphere, and slotted cone-shaped sphere. Target flight parameters: launch point coordinates (6378137,0,0); landing point coordinates (-926908.13,6310425.732,0.001); flying from west to east, with a total flight time of 2000s and an observation interval of 0.1s.
[0079] Static measurement data for four types of targets were obtained, with azimuth angles varying in 0.2° increments between 0° and 180°, and frequencies varying in 20MHz increments between 8.75GHz and 10.75GHz. Each frequency and azimuth angle corresponded to nine data points: frequency value, RCS(HH), phase(HH), RCS(HV), phase(HV), RCS(VH), phase(VH), RCS(VV), and phase(VV), resulting in 901×909 target static measurement data. Next, based on the target flight data, target precession angle, and precession frequency, the target precession attitude angle at each observation moment was calculated, thus obtaining RCS data, HRRP data, and narrowband polarization data for the actual flight process of the target group based on the target static measurement data. Subsequently, observation data from 950s to 1050s was used for simulation analysis, yielding RCS data, HRRP data, and narrowband polarization data with sizes of 1×1000, 101×1000, and 4×1000, respectively.
[0080] Specifically, RCS (Radar Cross Section) is a physical quantity that reflects the scattering ability of a radar target to reflect electromagnetic waves. It directly reflects the electromagnetic scattering characteristics of a target and is mainly affected by the following factors: the physical size and surface material of the target (such as an object or group of targets), the target's mass, relevant parameters of the radar equipment, and the attitude angle between the target and the radar wave vector. For example, the target's velocity, acceleration, and micro-motion parameters can be reflected in the RCS sequence. By extracting the mean, variance, skewness coefficient, and kurtosis coefficient from the RCS time series, radar target identification can be performed. The target's dynamic RCS sequence is a random variable, constituting a random process. Its statistical characteristics can be obtained, including positional characteristic parameters: mean, variance, median, and mode; distribution characteristic parameters: skewness coefficient, amplitude coefficient, range, and coefficient of variation (the ratio of standard deviation to mean); and transform domain characteristic parameters, which can specifically be the mean and variance of the Mellin transform.
[0081] HRRP features are called High-Resolution Range Profiles (HRRP). A 1D range profile is a one-dimensional amplitude signal obtained by matched filtering of broadband radar echoes operating in the radio frequency optical region. Physically, "distance" refers to the projection of each scattering point of the target onto the radar's line-of-sight, while amplitude information comes from the amplitude of the echoes from each scattering point, characterizing the scattering intensity. When the radar bandwidth is large, the radar range resolution is improved, and the 1D range profile can be divided into a series of short-range units, called high-resolution 1D range profiles. HRRP features not only reflect the electromagnetic topography of the target surface but also have strong sensitivity to the target's attitude, thus reflecting the target's attitude. They contain rich target information, have a simple imaging principle, and clear physical meaning, making them important for target identification. HRRP is also affected by various factors, including radar frequency and bandwidth, the electromagnetic scattering properties of the target material, and motion attitude. Because the 1D profile is sensitive to changes in the target's precession angle, feature extraction requires a large number of 1D range profiles. One-dimensional range images are very similar within a certain range of attitude angles, and the changes in one-dimensional range images within a larger range of attitude angles also have their own inherent relationships. Therefore, as long as enough one-dimensional range images are collected and their changing patterns are grasped, targets can be identified using one-dimensional range images.
[0082] For narrowband polarization characteristics, the target's variable polarization effect, that is, the modulation of the polarization of the incident electromagnetic wave by the target, is reflected through the polarization scattering matrix. The polarization scattering matrix is related to factors such as the target's shape, size, material, structure, and attitude, as well as the radar frequency, thus containing rich information. Furthermore, although the polarization scattering matrix characterizes the target's scattering characteristics at a given orientation, it changes with the target's precession angle and is related to the selected polarization base of the transmitting and receiving antennas, making it inconvenient to use. For a radar observer, changes in the radar polarization base or the target's rotation around the line of sight do not add any new information. Therefore, there exists a set of polarization invariants that are independent of the target's rotation around the line of sight and the radar polarization base, which can eliminate one-dimensional changes in the target's three-dimensional attitude; these polarization invariants are called narrowband polarization characteristics.
[0083] For example, suppose the scattering matrix of the radar target is: If the target satisfies the reciprocity principle, then the scattering matrix is symmetric, i.e., S = S T S HV =S VH ,therefore
[0084] The trace of the power matrix is: P = Tr[G] = |S hh | 2 +|S vv | 2 +2|S hv | 2
[0085] The determinant of the scattering matrix is: Δ = det(S) = S hh S vv -S hv S vh
[0086] The depolarization coefficient is: D = 1 - (abs(S)) hh +S vv )) 2 / Tr(G)
[0087] Specifically, the trace P of the power matrix roughly reflects the size of the target, the determinant Δ of the scattering matrix roughly reflects the thickness of the target, and the depolarization coefficient D indicates the number of target scattering centers.
[0088] Furthermore, in this embodiment, when extracting features from the processing results to obtain multidimensional feature data about the target group, the following steps are included:
[0089] S108: Determine the sliding window and sliding step size in the feature extraction module to extract the measurement sequence used for feature extraction of different types of features in the processing results;
[0090] S107: Determine the extraction algorithms corresponding to different types of features, combine each extraction algorithm with the corresponding measurement sequence to extract features, and obtain multidimensional feature data about the target group.
[0091] In this embodiment, the target group has multiple types of features, and different types of features have corresponding multi-dimensional feature data. The process of processing multi-dimensional feature data to form a target multi-dimensional feature vector includes:
[0092] S110: Merge the multidimensional feature data corresponding to different types of features of the target group to form a target multidimensional feature vector.
[0093] For example, in this embodiment, feature extraction is performed based on a feature extraction module, such as... Figure 2 The diagram illustrates the basic flow of the feature extraction module. For radar, the main features that can be extracted include target structure and motion characteristics. Each type of feature contains multiple feature quantities, and even for the same feature quantity, there can be multiple extraction algorithms. Therefore, to achieve more accurate feature extraction, the feature extraction module requires the system or user to first determine the sliding window and sliding step size, and then extract the specific measurement sequence used for feature extraction. Afterwards, the system or user specifies which feature extraction algorithm to use and calls the relevant algorithm module to calculate the corresponding feature quantities.
[0094] Specifically, in practical applications, for example, a 10s observation window and a 1s sliding step can be used to obtain 91 window data within the observation time of 950s-1050s. The data sizes of RCS data, HRRP data and narrowband polarization data in each window are 1×100, 100×101 and 4×100, respectively.
[0095] For RCS sequences, 16 features can be extracted, including center distance, range, mean, and extreme values.
[0096] For HRRP sequences, 10 features can be extracted, such as symmetry, dispersion characteristics, and number of scattering centers.
[0097] For narrowband polarization sequences, three features can be extracted: power matrix, scattering matrix, and depolarization coefficient.
[0098] During the observation period of 950s-1050s, feature extraction and normalization were performed on 91 observation windows of RCS data, HRRP data and narrowband polarization data, respectively, to obtain 91 sets of 1×16, 100×10 and 100×3 feature vectors. The obtained feature vectors were then merged to obtain a 91×1316 feature vector. The merged feature vector was saved to prepare for the subsequent deep feature extraction by the autoencoder.
[0099] Furthermore, an autoencoder (AE) is a neural network based on unsupervised learning, which aims to reconstruct dimensionally compressed input samples by continuously adjusting parameters. Figure 3 This is a typical autoencoder with a three-layer structure. The mapping from the input layer to the intermediate layer is called encoding, and the mapping from the intermediate layer to the output layer is called decoding. First, the compressed vector is obtained through encoding, and then the data is reconstructed through decoding.
[0100] The connection weights and biases between the input layer and the intermediate layer are denoted as W and b, respectively. The input data x is encoded to obtain the compressed vector y, as shown in the following equation: y = f(Wx + b).
[0101] The connection weights and biases between the intermediate layer and the reconstruction layer are denoted as follows: and Reconstructed value (decoding result) As shown in the following formula:
[0102] Where f(·) represents the encoder activation function, This represents the activation function of the decoder. (Reconstruction layer) It can be represented as:
[0103] Training an autoencoder involves determining the parameters W of the encoder and decoder. b、 The process involves first calculating the reconstructed value of the input sample x. Then, the backpropagation algorithm is used to adjust the parameter values, and the above process is iterated until the error function converges to a minimum value.
[0104] The target multidimensional feature vector is input into an autoencoder to achieve dimensionality reduction and deep feature extraction based on the autoencoder, resulting in a target multidimensional feature map, including:
[0105] The target multidimensional feature vector is input into the autoencoder. Based on the multiple intermediate layer structure of the autoencoder, the data dimensionality reduction and deep feature extraction of the target multidimensional feature vector are realized to obtain the target multidimensional feature map.
[0106] In other words, by using the multidimensional measurement data of the target as input to the autoencoder, the encoding obtained from the intermediate layer of the autoencoder can be used as the feature vector of the target. Although the extracted features are abstract, the resulting feature vector can more comprehensively represent the original measurement data and can more meticulously characterize the target.
[0107] For example, an autoencoder is designed and trained 10,000 times. The autoencoder has a 5-layer intermediate structure, with the 3rd intermediate layer used for feature vector output. In this embodiment, there are four types of targets. After the feature vector data of each type of target is extracted by the autoencoder during the observation time of 950s-1050s, 91 20×20 multidimensional feature maps of the targets can be obtained. For details, please refer to [reference needed]. Figure 4 As shown.
[0108] Optionally, the method in this embodiment further includes:
[0109] S111: The multidimensional feature map of the target group is used as the training data for the target neural network architecture. The maximum number of iterations is set to 50, and the number of iterations is 30 to train the target neural network architecture, resulting in a target neural network used to classify the multidimensional feature map of the target and achieve target recognition. The target neural network convolves the input multidimensional feature map of the target group to extract features. After standardization, the extracted features are activated using the ReLU activation function, then transformed into a one-dimensional vector through a fully connected layer for classification. The probability of each input data point is then obtained through the softmax function, and finally, the corresponding output, i.e., the target classification result, is obtained through classoutput. Specifically, the target neural network architecture in this embodiment is a convolutional neural network (convolutional neural network). A convolutional neural network is a network model inspired by biological visual systems and is a variant of the multilayer perceptron. The two main characteristics of a convolutional neural network are local connectivity and weight sharing.
[0110] The basic structure of the convolutional neural network in this embodiment is as follows:
[0111] Convolutional layers are the core modules of convolutional neural networks, primarily responsible for extracting image features. A convolutional layer consists of learnable convolutional kernels (also called filters). Each kernel has a relatively small two-dimensional spatial dimension, but its depth is the same as the input data. During the convolution operation, each kernel slides across the input data, and then the inner product of the entire kernel and its corresponding position in the input data is calculated.
[0112] Nonlinear activation functions are used because there is a highly nonlinear mapping between the input image and the final output label. A nonlinear function is added after each hidden layer to enhance the network's expressive power. This embodiment uses the ReLU activation function.
[0113] Fully connected layers act as classifiers in a high-dimensional feature space. They perform feature weighting operations, mapping high-dimensional features to the sample label space. Within a fully connected layer, lower-level neurons are fully connected to all neurons in the upper layer.
[0114] A classifier, specifically a convolutional neural network classifier, mainly consists of two parts: a scoring function and a loss function. The scoring function maps input image data to class scores.
[0115] Another crucial component of a classifier is the loss function (also known as the cost function or objective function), which measures the consistency between the model's predicted label score and the target true label. In other words, the loss function quantifies the difference between the predicted and true labels using a specific numerical value. The greater the difference between the predicted and true labels output by the scoring function, the larger the loss function value, and vice versa. Optimizing the parameters of the loss function can be transformed into an optimization problem, finding the minimum point of the loss function by iteratively updating its parameters.
[0116] Optionally, a Softmax classifier is preferably used in this embodiment. The Softmax classifier outputs normalized classification probabilities, and the output score is more intuitive and can be interpreted probabilistically. In the Softmax classifier, the output value of the system function mapping is s = f(x). i The input values (W, b) remain unchanged, but the final score is obtained by processing the output value through a Softmax function. The Softmax function compresses the input values and outputs a vector where each element is between 0 and 1, and the sum of all original elements is 1.
[0117] Once the target neural network architecture is constructed, the target multidimensional feature map obtained after deep feature extraction via an autoencoder or the historical target multidimensional feature map can be used as input to train the model. In this embodiment, the target group includes four categories of targets. During the observation period of 950s-1050s, each category has 91 feature maps. Therefore, 80% (72 maps) can be used as network training data, and 20% (19 maps) as network test data. Specifically, the maximum number of iterations and the number of iterations can be set to 50 and 30 respectively.
[0118] During training, if the training data is limited (i.e., small sample training), although the number of network iterations and training epochs can be adjusted, it is still difficult to change the training outcome. Figure 5 and Figure 6 As shown, although the network recognition accuracy improves slightly with the increase of network iterations and training rounds, the learning effect of the convolutional neural network is still poor due to the small number of training data samples, resulting in an extremely low recognition accuracy of only 55.26% for the target group. Therefore, it is crucial to improve the recognition accuracy in small samples.
[0119] Therefore, to address target recognition in small sample sizes, the method in this embodiment further includes:
[0120] S112: Construct a generative adversarial network, which includes a generator and a discriminator;
[0121] S113: The generative adversarial network is trained by cross-training to obtain samples that approximate the distribution of high-dimensional feature map samples of the real target group;
[0122] S114: When performing small-sample target recognition based on the target neural network, a new multi-dimensional feature map that approximates the real feature map is generated based on the generative adversarial network to train the target neural network, so that the trained target neural network can accurately perform small-sample target recognition.
[0123] Specifically, addressing the small sample size problem, this embodiment uses data augmentation to expand the training sample set. This ensures the model's ability to fit target features while enriching the sample's feature space, preventing overfitting and providing a new direction for research on small sample radar target recognition technology. In the field of image processing, sample expansion is typically achieved by performing geometric transformations such as rotation, flipping, scaling, translation, and cropping on images. However, due to the pose, translation, and amplitude sensitivity of HRRP data, this data augmentation method is prone to causing the destruction and loss of HRRP features.
[0124] Generative Adversarial Networks (GANs), as generative models, can generate high-quality new samples that closely approximate the distribution of real sample data by fitting the distribution of input sample data through training. This can be used for data augmentation in small-sample problems. The GAN in this embodiment consists of a generator and a discriminator. The GAN model structure fully embodies the idea of a two-player zero-sum game. The discriminator's function is to distinguish the probability that the input is a real sample, while the generator's goal is to generate samples that closely approximate the real sample distribution to deceive the discriminator. During the adversarial process, the generator and discriminator continuously improve their performance through training, with the ultimate goal of reaching a Nash equilibrium. In this state, the discriminator is deceived by the generated samples and can no longer distinguish them from real samples; the generator's training is complete.
[0125] During the training of the GAN model, no specific criteria were designed to evaluate the quality of the generated samples. Instead, the discriminator evaluated the samples, which not only avoided the need for complex evaluation criteria but also provided a certain degree of diversity to the generated samples.
[0126] In practical applications, such as Figure 7 As shown, in this embodiment, a new multidimensional feature map of four types of targets in the target group is generated after training the adversarial network 2000 times. Figure 4The real feature maps shown are very similar. Then, the generative adversarial network (GAN) will be trained multiple times based on the target's real feature maps to generate new multidimensional feature maps, which will then be used as input sample data for the convolutional neural network to perform target recognition on the target group.
[0127] In this embodiment, the generative adversarial network is trained 2000 times. 50% of the generated feature maps are used to obtain 1000 feature maps for each of the four target classes in the target group. Then, 80% (800 maps) are used as network training data, and 20% (200 maps) are used as network test data. For example... Figure 8 As shown, it is the confusion matrix of the network recognition results, compared with... Figure 5 It can be seen that using generative adversarial networks can obtain more sufficient sample data, which can further improve the recognition accuracy of convolutional neural networks. At this time, the network recognition accuracy reaches 90.75%, an improvement of 35.49%.
[0128] like Figure 9 As shown, another embodiment of the present invention also provides a target comprehensive recognition device 100 based on high-dimensional feature maps, comprising:
[0129] The acquisition module is used to acquire observational data about the target group based on radar sensor data.
[0130] The first processing module is used to process the observation data and extract features from the processing results to obtain multidimensional feature data about the target group.
[0131] The second processing module is used to process the multidimensional feature data to form a target multidimensional feature vector.
[0132] The first input module is used to input the target multidimensional feature vector into the autoencoder, so as to realize the data dimensionality reduction and deep feature extraction of the target multidimensional feature vector based on the autoencoder, and obtain the target multidimensional feature map;
[0133] The second input module is used to input the target multidimensional feature map into the target neural network, so as to classify the target multidimensional feature map based on the target neural network and realize target recognition.
[0134] As an optional embodiment, obtaining observation data about the target group based on radar sensor acquisition includes:
[0135] Obtain observational data about a target group based on radar sensors, including flight data, target precession angle, and precession frequency of the target group.
[0136] As an optional embodiment, the processing of the observation data includes:
[0137] Calculate the target precession attitude angle of the target group at each observation time based on the observation data;
[0138] Based on the target precession attitude angle and observation data, the RCS data, HRRP data and narrowband polarization data of the target group during actual flight are calculated and generated.
[0139] As an optional embodiment, the step of extracting features from the processing results to obtain multidimensional feature data about the target group includes:
[0140] Determine the sliding window and sliding step size in the feature extraction module to extract measurement sequences for feature extraction of different types of features in the processing results;
[0141] Extraction algorithms for different types of features are determined, and feature extraction is performed by combining each extraction algorithm with the corresponding measurement sequence to obtain multidimensional feature data about the target group.
[0142] As an optional embodiment, the target group has multiple types of features, and different types of features have corresponding multidimensional feature data; the process of processing the multidimensional feature data to form a target multidimensional feature vector includes:
[0143] The multidimensional feature vector of the target group is formed by merging the multidimensional feature data corresponding to different types of features of the target group.
[0144] As an optional embodiment, the step of inputting the target multidimensional feature vector into an autoencoder to achieve data dimensionality reduction and deep feature extraction of the target multidimensional feature vector based on the autoencoder, thereby obtaining a target multidimensional feature map, includes:
[0145] The target multidimensional feature vector is input into an autoencoder, and the data dimensionality reduction and deep feature extraction of the target multidimensional feature vector are realized based on the multiple intermediate layer structures of the autoencoder to obtain the target multidimensional feature map.
[0146] As an optional embodiment, it also includes:
[0147] The target neural network architecture is trained using the multidimensional feature map of the target group as training data. The maximum number of iterations is set to 50, and the number of iterations is 30. This results in a target neural network used to classify the multidimensional feature map and achieve target recognition. The target neural network convolves the input multidimensional feature map to extract features. After standardization, the extracted feature data is activated by an activation function. A fully connected layer then transforms the feature data into a one-dimensional vector for classification. Finally, an objective function is used to obtain the probability of each input data point, and the corresponding target classification result is obtained based on the probability analysis. The objective function includes a softmax function.
[0148] As an optional embodiment, it also includes:
[0149] Construct a generative adversarial network, wherein the generative adversarial network includes a generator and a discriminator;
[0150] The generative adversarial network is trained using a cross-training method to obtain samples that approximate the distribution of high-dimensional feature map samples of the real target group;
[0151] When performing small-sample target recognition based on the target neural network, the target neural network is trained by generating multiple new multidimensional feature maps that approximate the real feature maps based on the generative adversarial network, so that the trained target neural network can accurately perform small-sample target recognition.
[0152] Another embodiment of the present invention also provides a target comprehensive recognition device based on high-dimensional feature map, comprising:
[0153] At least one processor; and,
[0154] A memory communicatively connected to the at least one processor; wherein,
[0155] The memory stores instructions that can be executed by the at least one processor to implement the target comprehensive recognition method based on high-dimensional feature maps as described in any of the embodiments above.
[0156] Furthermore, one embodiment of the present invention also provides a storage medium storing a computer program, which, when executed by a processor, implements the target comprehensive recognition method based on high-dimensional feature maps as described above. It should be understood that the various solutions in this embodiment have the corresponding technical effects in the above-described method embodiments, and will not be repeated here.
[0157] Furthermore, embodiments of the present invention also provide a computer program product, which is tangibly stored on a computer-readable medium and includes computer-readable instructions that, when executed, cause at least one processor to perform a target comprehensive recognition method based on high-dimensional feature maps, such as the embodiments described above.
[0158] It should be noted that the computer storage medium in this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. Computer-readable media can be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access storage media (RAM), read-only storage media (ROM), erasable programmable read-only storage media (EPROM or flash memory), optical fibers, portable compact disk read-only storage media (CD-ROM), optical storage media, magnetic storage media, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. In this application, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program configured for use by or in connection with an instruction execution system, system, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, antenna, optical fiber, RF, etc., or any suitable combination thereof.
[0159] Furthermore, 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. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage and optical storage) containing computer-usable program code.
[0160] 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 system that specifies functions in one or more boxes.
[0161] 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 an instruction set implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0162] 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.
[0163] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.
[0164] The above embodiments are merely exemplary embodiments of the present invention and are not intended to limit the present invention. The scope of protection of the present invention is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to the present invention within its spirit and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of the present invention.
Claims
1. A target recognition method based on high-dimensional feature maps, characterized in that, include: Obtain observational data about a target group based on radar sensors, the observational data including the flight data of the target group, target precession angle, and precession frequency; The observation data is processed, and features are extracted from the processing results to obtain multidimensional feature data about the target group; The multidimensional feature data is processed to form a target multidimensional feature vector; The target multidimensional feature vector is input into an autoencoder to achieve data dimensionality reduction and deep feature extraction based on the autoencoder, thereby obtaining a target multidimensional feature map; The target multidimensional feature map is input into the target neural network to classify the target multidimensional feature map based on the target neural network, thereby achieving target recognition; The processing of the observation data includes: Calculate the target precession attitude angle of the target group at each observation time based on the observation data; Based on the target precession attitude angle and observation data, the RCS data, HRRP data and narrowband polarization data of the target group during actual flight are calculated and generated.
2. The target comprehensive recognition method based on high-dimensional feature map as described in claim 1, characterized in that, The process of extracting features from the processing results to obtain multidimensional feature data about the target group includes: Determine the sliding window and sliding step size in the feature extraction module to extract measurement sequences for feature extraction of different types of features in the processing results; Extraction algorithms for different types of features are determined, and feature extraction is performed by combining each extraction algorithm with the corresponding measurement sequence to obtain multidimensional feature data about the target group.
3. The target comprehensive recognition method based on high-dimensional feature map as described in claim 1, characterized in that, The target group has multiple types of features, and different types of features have corresponding multidimensional feature data; The process of processing the multidimensional feature data to form a target multidimensional feature vector includes: The multidimensional feature vector of the target group is formed by merging the multidimensional feature data corresponding to different types of features of the target group.
4. The target comprehensive recognition method based on high-dimensional feature map as described in claim 1, characterized in that, The step of inputting the target multidimensional feature vector into an autoencoder to achieve dimensionality reduction and deep feature extraction of the target multidimensional feature vector based on the autoencoder, thereby obtaining a target multidimensional feature map, includes: The target multidimensional feature vector is input into an autoencoder, and the data dimensionality reduction and deep feature extraction of the target multidimensional feature vector are realized based on the multiple intermediate layer structures of the autoencoder to obtain the target multidimensional feature map.
5. The target comprehensive recognition method based on high-dimensional feature map as described in claim 1, characterized in that, Also includes: The target neural network architecture is trained using the multidimensional feature map of the target group as training data. The maximum number of iterations is set to 50, and the number of iterations is 30. This results in a target neural network used to classify the multidimensional feature map and achieve target recognition. The target neural network convolves the input multidimensional feature map to extract features. After standardization, the extracted feature data is activated by an activation function. A fully connected layer then transforms the feature data into a one-dimensional vector for classification. Finally, an objective function is used to obtain the probability of each input data point, and the corresponding target classification result is obtained based on the probability analysis. The objective function includes a softmax function.
6. The target comprehensive recognition method based on high-dimensional feature map as described in claim 1, characterized in that, Also includes: Construct a generative adversarial network, wherein the generative adversarial network includes a generator and a discriminator; The generative adversarial network is trained using a cross-training method to obtain samples that approximate the distribution of high-dimensional feature map samples of the real target group; When performing small-sample target recognition based on the target neural network, the target neural network is trained by generating multiple new multidimensional feature maps that approximate the real feature maps based on the generative adversarial network, so that the trained target neural network can accurately perform small-sample target recognition.
7. A target comprehensive recognition device based on high-dimensional feature maps, characterized in that, include: The acquisition module is used to acquire observational data about the target group based on radar sensor data. The first processing module is used to process the observation data and extract features from the processing results to obtain multidimensional feature data about the target group. The observation data includes the flight data of the target group, the target precession angle, and the precession frequency. The second processing module is used to process the multidimensional feature data to form a target multidimensional feature vector. The first input module is used to input the target multidimensional feature vector into the autoencoder, so as to realize the data dimensionality reduction and deep feature extraction of the target multidimensional feature vector based on the autoencoder, and obtain the target multidimensional feature map; The second input module is used to input the target multidimensional feature map into the target neural network, so as to classify the target multidimensional feature map based on the target neural network and realize target recognition; The processing of the observation data includes: Calculate the target precession attitude angle of the target group at each observation time based on the observation data; Based on the target precession attitude angle and observation data, the RCS data, HRRP data and narrowband polarization data of the target group during actual flight are calculated and generated.
8. A target comprehensive recognition device based on high-dimensional feature maps, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to implement the target comprehensive recognition method based on high-dimensional feature maps as described in any one of claims 1-6.