One-dimensional range profile target prediction method based on deep learning

Abstract

The invention belongs to the technical field of signal processing, and particularly discloses a one-dimensional range profile target prediction method based on deep learning, and the method comprises the steps: converting the sparse feature vectors of the dimension of a strong scattering center, the feature size of a target, and the distribution entropy of the scattering center into corresponding dense feature vectors; an FM model and an MLP model which are parallel are established, and the FM model is used for learning feature interaction of a strong scattering center dimension dense feature vector, a target feature size dense feature vector and a scattering center distribution entropy dense feature vector; the MLP model is used for learning a nonlinear relationship between the dense feature vector of the dimension of the strong scattering center, the dense feature vector of the feature size of the target and the dense feature vector of the distribution entropy of the scattering center and the target; and under the condition that respective model parameters of the FM model and the MLP model are determined through training, fusing the output of the FM model and the output of the MLP model to obtain a target prediction result. According to the scheme, the technical problem that an existing one-dimensional range profile target prediction method in a data sparse scene is low in precision is solved.

Description

Technical Field

[0001] This invention belongs to the technical field of signal processing, specifically relating to a one-dimensional distance image target prediction method based on deep learning. Background Technology

[0002] In the optical region, when the radar transmit signal spectrum is wide enough, the radial distance of the target occupies multiple radar range resolution cells. The backscattering of high-frequency electromagnetic radiation from the target exhibits continuous fluctuation characteristics in the time domain, forming a target amplitude image along the radar line of sight, which is a one-dimensional range image.

[0003] Currently, the common practice for target prediction is to correlate and align one-dimensional range images, and then predict the target by calculating the matching degree between the target's one-dimensional range image and the corresponding attitude angular range image in the database. However, the correlation and alignment of range images involves a large computational load; at the same time, storing the entire range image in the database also requires a significant amount of storage. In the past few years, scholars have proposed several methods to improve the accuracy of target prediction, and these methods have been widely used. Although they have achieved great success, these methods still have the following problems:

[0004] (1) Existing target prediction methods lack qualitative analysis of the relationship between target data and the strong scattering center dimension, target feature size, and scattering center distribution entropy. Studies have demonstrated the importance of the strong scattering center dimension, target feature size, and scattering center distribution entropy in target prediction. However, the relationship between the target and the strong scattering center dimension, target feature size, and scattering center distribution entropy has not been clearly established. The strong scattering center dimension, target feature size, and scattering center distribution entropy only indirectly participate in the prediction process and are not involved in subsequent predictions. This further limits the model's ability to perceive the strong scattering center dimension, target feature size, and scattering center distribution entropy.

[0005] (2) Collaborative filtering is a commonly used method for data prediction. The core idea of ​​collaborative filtering is to use historical data of similar users for prediction. Although existing collaborative filtering methods improve the accuracy of target prediction by improving similarity, increasing the reliability of neighbors, and mitigating the influence of target data range, these methods only use historical target data of some similar neighbors for prediction and cannot utilize all data for more accurate prediction. The accuracy of collaborative filtering methods is closely related to the available data. When the available target data is very scarce, the accuracy of collaborative filtering methods will be greatly reduced. In addition, in the case of sparse data, collaborative filtering methods may fail to work due to the lack of common targets among one-dimensional distance image features.

[0006] (3) Existing prediction methods neglect the influence of nonlinear interactions. Matrix factorization methods can learn one-dimensional distance profile features and latent interaction feature vectors of the target from the original target interaction data, and use a simple linear inner product operation between the one-dimensional distance profile features and the latent feature vectors of the target to achieve target prediction. Factorization machine methods represent each feature as a latent vector, and use the inner product of pairwise feature latent vectors to represent their interaction. Although matrix factorization and factorization machine methods can improve prediction accuracy to some extent by capturing the latent features of features and the target, this simple linear model is difficult to capture the nonlinear relationship between one-dimensional distance profile features and target interactions, which limits the improvement of target prediction accuracy. Summary of the Invention

[0007] The purpose of this invention is to provide a one-dimensional distance image target prediction method based on deep learning, so as to solve the technical problem of low accuracy of existing one-dimensional distance image target prediction methods in data-sparse scenarios.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A deep learning-based one-dimensional range profile target prediction method includes transforming sparse feature vectors of strong scattering center dimension, target feature size, and scattering center distribution entropy into corresponding dense feature vectors; establishing parallel FM and MLP models, wherein the FM model is used to learn the feature interactions of the dense feature vectors of strong scattering center dimension, target feature size, and scattering center distribution entropy, and the MLP model is used to learn the nonlinear relationship between the dense feature vectors of strong scattering center dimension, target feature size, and scattering center distribution entropy and the target; and, after training and determining the model parameters of the FM and MLP models respectively, fusing the outputs of the FM and MLP models to obtain the target prediction result.

[0010] Preferably, the sparse feature vectors of the strong scattering center dimension, target feature size, and scattering center distribution entropy are transformed into corresponding dense feature vectors, including: using feature embedding to transform the binary sparse feature vectors of the strong scattering center dimension, target feature size, and scattering center distribution entropy into corresponding dense feature vectors.

[0011] Preferably, feature embedding is used to transform the sparse feature vectors of the binary strong scattering center dimension, target feature size, and scattering center distribution entropy into corresponding dense feature vectors, including: mapping the sparse feature vectors of each binary strong scattering center dimension, target feature size, and scattering center distribution entropy through the corresponding embedding matrix to obtain the corresponding dense feature vector.

[0012] Preferably, establishing an FM model includes: establishing a first-order module for each single feature to be correlated with the prediction result; establishing a higher-order module for the interaction of pairs of features to be correlated with the prediction result; and associating the first-order module and the higher-order module to obtain the FM model.

[0013] Preferably, establishing an MLP model includes: establishing a neural network to learn the nonlinear relationship between the dense feature vector of the strong scattering center dimension, the dense feature vector of the target feature size, the dense feature vector of the scattering center distribution entropy, and the target; and determining the activation function of the above MLP model.

[0014] Preferably, the training determines the model parameters of the FM model and the MLP model, including: defining the corresponding objective function based on the model parameters to be determined; and using a stochastic gradient descent optimizer to train in an end-to-end manner to determine the local optimum of the objective function.

[0015] Preferably, the objective function corresponding to the training is defined based on the model parameters to be determined, including: determining the squared loss between the target predicted value and the target actual value as the objective function of the model parameters.

[0016] Preferably, the stochastic gradient descent optimizer is used for end-to-end training, including updating the model parameters according to the learning rate and gradient in each iteration of training, until the iteration ends.

[0017] Preferably, determining the local optimum of the objective function includes: defining a loss function for the objective function; and obtaining the current solution of the objective function when the loss function converges or the number of iterations reaches a preset threshold.

[0018] A one-dimensional distance-image target prediction system based on deep learning includes: a transformation module for transforming sparse feature vectors of strong scattering center dimension, target feature size, and scattering center distribution entropy into corresponding dense feature vectors; a model building module for building parallel FM and MLP models, wherein the FM model is used to learn the feature interactions of dense feature vectors of strong scattering center dimension, target feature size, and scattering center distribution entropy, and the MLP model is used to learn the nonlinear relationship between the dense feature vectors of strong scattering center dimension, target feature size, and scattering center distribution entropy and the target; and a prediction module for fusing the outputs of the FM model and the MLP model to obtain the target prediction result, given that the model parameters of the FM model and the MLP model have been determined during training.

[0019] A computer-readable storage medium storing a computer program configured to execute the aforementioned deep learning-based one-dimensional distance image target prediction method at runtime.

[0020] An electronic device includes a memory and a processor, wherein the memory stores a computer program and the processor is configured to execute the aforementioned deep learning-based one-dimensional distance image target prediction method through the computer program.

[0021] In this invention, sparse feature vectors representing the strong scattering center dimension, target feature size, and scattering center entropy distribution are transformed into corresponding dense feature vectors. Parallel FM and MLP models are established. The FM model learns the feature interactions between the dense feature vectors representing the strong scattering center dimension, target feature size, and scattering center entropy distribution, while the MLP model learns the nonlinear relationships between these features and the target. With the model parameters of both the FM and MLP models determined during training, the outputs of the FM and MLP models are fused to obtain the target prediction. As a result, the deep learning model learns the feature interactions of the dense feature vectors of the strong scattering center dimension, the dense feature vector of the target feature size, and the dense feature vector of the scattering center distribution entropy, as well as the nonlinear relationship between these features and the target. This effectively solves the data coefficient problems of the dense feature vectors of the strong scattering center dimension, the dense feature vector of the target feature size, and the scattering center distribution entropy, achieving accurate target prediction in sparse data scenarios. It also solves the technical problem of low accuracy in existing one-dimensional range image target prediction methods in sparse data scenarios, thus improving the prediction accuracy of one-dimensional range image targets. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a one-dimensional distance image target prediction method based on deep learning in an embodiment of the present invention;

[0023] Figure 2 This is a schematic diagram of the structure of a deep learning model for a one-dimensional distance image target prediction method based on deep learning in an embodiment of the present invention;

[0024] Figure 3 This is a schematic diagram of the structure of a one-dimensional distance image target prediction system based on deep learning in an embodiment of the present invention. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The advantages and features of the present invention will become clearer from the following description and claims. It should be noted that the drawings are all in a very simplified form and are not to a precise scale, and are only used to facilitate and clarify the illustration of the embodiments of the present invention.

[0026] It should be noted that, in order to clearly illustrate the content of this invention, several embodiments are provided to further explain different implementations of the invention. These embodiments are enumerated rather than exhaustive. Furthermore, for the sake of brevity, content mentioned in the preceding embodiments is often omitted in the following embodiments. Therefore, content not mentioned in the later embodiments can be referred to in the preceding embodiments.

[0027] Example 1

[0028] A deep learning-based method for predicting one-dimensional distance-image targets, such as Figure 1 As shown, the method includes:

[0029] S102 transforms the sparse feature vectors of strong scattering center dimension, target feature size, and scattering center distribution entropy into corresponding dense feature vectors.

[0030] S104. Establish parallel FM and MLP models. The FM model is used to learn the feature interactions of the dense feature vectors of the strong scattering center dimension, the dense feature vectors of the target feature size, and the dense feature vectors of the scattering center distribution entropy. The MLP model is used to learn the nonlinear relationship between the dense feature vectors of the strong scattering center dimension, the dense feature vectors of the target feature size, the dense feature vectors of the scattering center distribution entropy, and the target.

[0031] S106. With the model parameters of the FM model and MLP model determined during training, the outputs of the FM model and the MLP model are fused to obtain the target prediction result.

[0032] This is not limited to feature analysis of one-dimensional range images. Let the radar echo be S. M*N M is the echo number, N is the number of sampling points, and the one-dimensional range profile is obtained through FFT operation. Assume the magnitude matrix of the one-dimensional range profile is P. M*N Each row of the matrix represents the amplitude value of a single observation of the one-dimensional range image. In a single observation of the one-dimensional range image, the range image envelope fluctuates with the scattering point, and the resolution is determined by the bandwidth of the radar transmitted signal. Due to changes in the target attitude angle and the influence of external disturbances, the one-dimensional range image observed M times exhibits instability. Studies using a large amount of measurement data show that the mean amplitude of each observation of the one-dimensional range image follows a certain random distribution.

[0033] Let the mean of any one-dimensional range image be E. k (1≤k≤M), then the dimension Z(k) of the target strong scattering center is expressed as:

[0034]

[0035] In equation (1), U is the unit step function.

[0036] The dimension Z(k) of the target's strong scattering centers represents the number of range cells greater than the mean in the one-dimensional range profile observed in the kth time. It is closely related to the number of scattering centers of the target, but not equal to the number of scattering centers.

[0037] Optionally, let L k (i)={(k,i)|(P(k,i)-E k )≥0,k=0,1...M-1,i=0,1...Z k -1},L k (i) indicates that the one-dimensional range profile of the target observed in the kth observation is greater than or equal to the mean E. k The position sequence.

[0038] The target feature size C(k) is expressed as:

[0039] C(k)=L k (Z k -1)-L k (0) (2)

[0040] C(k) reflects the radial dimension of the target. Since the one-dimensional range image of the target is closely related to its attitude, C(k) is a sequence that fluctuates with the attitude. However, from the overall trend analysis, different targets with significantly different dimensions have large differences in their corresponding C(k), thus enabling target identification.

[0041] Entropy reflects the randomness of a variable; a uniformly distributed variable has the highest entropy, while a relatively concentrated distribution has the lowest entropy. For any one-dimensional distance image modulus sequence of an observation, let m... k (n) = P(k,n), n = 0, 1...N-1, for m k (n) Normalization, m k (n) divided by the total modulus is represented as: m' k (n)=mk(n) / ∑ |m k(n)|.

[0042] m k The entropy H(k) of the scattering center distribution of (n) is expressed as:

[0043]

[0044] H(k) reflects the distribution of scattering centers in the one-dimensional range image of the k-th observation, and also the degree of dispersion of the scattering centers in the radial scale. A higher entropy indicates a greater likelihood that the scattering centers are evenly distributed in the radial scale; conversely, a lower entropy indicates a more concentrated distribution of scattering centers. Therefore, H(k) reflects the general structure of the target.

[0045] To determine the correlation between the strong scattering center dimension, target feature size, scattering center distribution entropy, and the target, the Pearson similarity coefficient (PCC) is used to measure similarity based on these three features, and the following definition and calculation method are given:

[0046]

[0047] Among them, U x It is a feature set of strong scattering center dimension, target feature size, and scattering center distribution entropy, PCCU x This represents targets with a similarity greater than a specified threshold, where α represents the threshold of PCC(u,v). |·| represents the size of the set.

[0048] As an optional implementation, the sparse feature vectors of strong scattering center dimension, target feature size, and scattering center distribution entropy are transformed into corresponding dense feature vectors, including: using feature embedding to transform the binary sparse feature vectors of strong scattering center dimension, target feature size, and scattering center distribution entropy into corresponding dense feature vectors.

[0049] As an optional implementation, feature embedding is used to transform the sparse feature vectors of the binary strong scattering center dimension, target feature size, and scattering center distribution entropy into corresponding dense feature vectors. This includes mapping each binary strong scattering center dimension, target feature size, and scattering center distribution entropy sparse feature vector through the corresponding embedding matrix to obtain the corresponding dense feature vector.

[0050] As an optional implementation method, the FM model is established by: establishing a first-order module for each single feature to be associated with the prediction result; establishing a higher-order module for the interaction of pairs of features to be associated with the prediction result; and associating the first-order module and the higher-order module to obtain the FM model.

[0051] As an optional implementation, an MLP model is established, including: establishing a neural network to learn the nonlinear relationship between the dense feature vector of the strong scattering center dimension, the dense feature vector of the target feature size, the dense feature vector of the scattering center distribution entropy, and the target; and determining the activation function of the MLP model.

[0052] As an optional implementation, training determines the model parameters of the FM model and the MLP model, including: defining the corresponding objective function based on the model parameters to be determined; and using a stochastic gradient descent optimizer to train in an end-to-end manner to determine the local optimum of the objective function.

[0053] As an optional implementation, the objective function corresponding to the training is defined based on the model parameters to be determined, including: determining the squared loss between the target predicted value and the target actual value as the objective function of the model parameters.

[0054] As an alternative implementation, the stochastic gradient descent optimizer is used for end-to-end training, which includes updating the model parameters according to the learning rate and gradient in each iteration of training until the iteration ends.

[0055] As an optional implementation method, determining the local optimum of the objective function includes: defining a loss function for the objective function; and obtaining the current solution of the objective function when the loss function converges or the number of iterations reaches a preset threshold.

[0056] In this embodiment, the sparse feature vectors of the strong scattering center dimension, target feature size, and scattering center distribution entropy are transformed into corresponding dense feature vectors. Parallel FM and MLP models are established. The FM model learns the feature interactions of the dense feature vectors of the strong scattering center dimension, target feature size, and scattering center distribution entropy. The MLP model learns the nonlinear relationships between these features and the target. With the model parameters of the FM and MLP models determined during training, the outputs of the FM and MLP models are fused to obtain the target. The prediction results, through a deep learning model, learn the feature interactions of the dense feature vectors of the strong scattering center dimension, the target feature size, and the dense feature vectors of the scattering center distribution entropy, as well as the nonlinear relationships between these features and the target. This effectively solves the data coefficient problems of the dense feature vectors of the strong scattering center dimension, the target feature size, and the scattering center distribution entropy, achieving accurate target prediction in sparse data scenarios. It also solves the technical problem of low accuracy in existing one-dimensional range image target prediction methods in sparse data scenarios, thus improving the prediction accuracy of one-dimensional range image targets.

[0057] Specifically, the implementation of one-dimensional distance image target prediction using deep learning models includes three parts, not limited to: Figure 2 The diagram illustrates input preprocessing, model building, and prediction / learning. In input preprocessing, the sparse input feature vectors are transformed into dense feature vectors and fed into the model for learning. The model consists of two parallel components: a Functional Factor (FM) model and a Multi-Level Processing (MLP) model. FM is used to learn higher-order feature interactions, while MLP is used to learn the complex nonlinear relationships between the strong scattering center dimension, target feature size, scattering center distribution entropy, and the target. In prediction learning, the results of FM and MLP are fused to generate the final predicted value.

[0058] Input preprocessing: Using feature embedding techniques, the sparse feature vectors representing the binary sparse strong scattering center dimension, target feature size, and scattering center distribution entropy are transformed into corresponding dense feature vectors. Dense feature vector e i Mapping is not limited to the following methods:

[0059] e i =W i,embed ·x i (5)

[0060] Where, x i It is the sparse vector of the i-th feature. It is the corresponding embedding matrix.

[0061] The embedding vector of the strong scattering center dimension, target feature size, scattering center distribution entropy, and target input features is not limited to being expressed by the following formula:

[0062] [e1,e2,...,e m ] = W embed ·[x1,x2,...,x m (6)

[0063] Where m represents the number of input features, W embed This represents the embedding matrix of all input features.

[0064] Model Establishment: The relationship between the strong scattering center dimension, target feature size, scattering center distribution entropy, and the target is a complex interactive process, and high-order and nonlinear interactive learning are both important. To simultaneously satisfy high-order and nonlinear interactions, the model adopts a parallel approach of FM and MLP. The FM part combines features to form new combined features, fully considering the interactions between features. The specific implementation of FM is not limited to the following:

[0065]

[0066] Where w0 is the bias coefficient, w i Representing a single feature x i Its function.

[0067] The sum of the first two terms in formula (7) represents the influence of the first-order feature on the prediction, w i,j Representative feature x i and x j The interaction between them represents the influence of feature interactions.

[0068] MLPs learn nonlinear interactions of features through neural networks. The activation function in MLPs is not limited to ReLU, avoiding gradient explosion and vanishing problems. When the model performs regression predictions, after multiple propagation layers, the final output layer is not limited to the following representation:

[0069]

[0070] Among them, e l W is the dense feature vector representation of the output layer of the MLP. o and b o These are the weight matrix and bias of the output layer, respectively.

[0071] Prediction and Learning: Considering the advantages of FM in high-order feature interactions and MLP in nonlinear interactions, the model ultimately fuses the results of FM and MLP to improve the accuracy of service quality prediction. The final prediction result, combining the results of both FM and MLP, is shown in the following equation:

[0072]

[0073] The problem of solving for the model parameters is transformed into an optimization problem, and overfitting is considered during the optimization process. Therefore, the objective function is minimized, and the loss function of the objective function is shown in the following equation:

[0074]

[0075] Where R+ represents the training instance set, R represents the target predicted value. i,j Represents the actual target value. Θ = {E, w} i w l b l} represents all the parameters in the model that need to be trained, and the regularization coefficient λ controls the strength of L2 regularization.

[0076] In regression problems, the service quality prediction is numerical. It is not limited to choosing the squared loss of the target prediction and the actual value as the objective function, and adding a regularization term to the parameters can reduce overfitting and improve the model's generalization ability.

[0077] The model is not limited to using stochastic gradient descent optimizers trained end-to-end to approach local optima of the objective function. Therefore, Θ = {E, w} i w l b l Each parameter ρ in} can be updated according to the following formula:

[0078]

[0079] Where i represents the number of iterations, and γ is the learning rate. It is the gradient of parameter ρ.

[0080] Specifically, the workflow algorithms for deep learning are not limited to those shown below:

[0081] Input: Training set R+

[0082] Number of iterations I

[0083] Output: Model parameters {E, w} i w l b l}

[0084] Begin

[0085]

[0086]

[0087] End

[0088] The computational overhead of model learning is concentrated in lines 5 through 12 of the algorithm. Line 7 embeds the input features, with a computational complexity of O(d*f), where d is the number of input features and f is the feature dimension. Line 8 fuses the results of FM and MLP, with a time complexity of O(1). Lines 9 and 10 calculate the loss and update the parameters. Since lines 7 through 10 are executed sequentially, the time complexity of learning the model is O(|I|×|R|×(d*f+1+2f). Because |I|, d, and f are finite values ​​in the model, the model complexity is linearly related to the training set |R|. Therefore, the deep learning model is efficient and applicable to large-scale target prediction.

[0089] In this embodiment, the one-dimensional range profile of the target is approximated as a quasi-random sequence that varies with attitude. Statistical methods are used to study the one-dimensional range profile, proposing three features: the target's strong scattering center dimension, target feature size, and target scattering center distribution entropy. These three features are used together to better characterize the target. Qualitative analysis is performed to determine the correlation between the strong scattering center dimension, target feature size, scattering center distribution entropy, and the target, enabling accurate discovery of information beneficial to target prediction based on these features. Qualitative analysis of the data determines the correlation between target data and the strong scattering center dimension, target feature size, and scattering center distribution entropy, improving the accuracy of target prediction. A deep learning-based model effectively addresses the problem of data sparsity for the strong scattering center dimension, target feature size, and scattering center distribution entropy. The deep learning model actively perceives the strong scattering center dimension, target feature size, and scattering center distribution entropy, achieving accurate target prediction in sparse data scenarios. Qualitative data analysis of target information determines the correlation between the strong scattering center dimension, target feature size, scattering center distribution entropy, and target data, achieving target prediction and recognition.

[0090] Example 2

[0091] A one-dimensional distance image target prediction system based on deep learning, such as Figure 3 As shown, the system includes:

[0092] The conversion module 302 is used to convert the sparse feature vectors of strong scattering center dimension, target feature size, and scattering center distribution entropy into corresponding dense feature vectors.

[0093] The model building module 304 is used to build parallel FM and MLP models. The FM model is used to learn the feature interactions of the dense feature vectors of the strong scattering center dimension, the dense feature vectors of the target feature size, and the dense feature vectors of the scattering center distribution entropy. The MLP model is used to learn the nonlinear relationship between the dense feature vectors of the strong scattering center dimension, the dense feature vectors of the target feature size, the dense feature vectors of the scattering center distribution entropy and the target.

[0094] The prediction module 306 is used to fuse the outputs of the FM model and the MLP model to obtain the target prediction result, given that the model parameters of the FM model and the MLP model are determined during training.

[0095] Optionally, the above-mentioned conversion module 302 converts the sparse feature vectors of strong scattering center dimension, target feature size, and scattering center distribution entropy into corresponding dense feature vectors, including: using feature embedding to convert the binary sparse feature vectors of strong scattering center dimension, target feature size, and scattering center distribution entropy into corresponding dense feature vectors.

[0096] Optionally, the above-mentioned conversion module 302 utilizes feature embedding to convert the sparse feature vectors of the binary strong scattering center dimension, target feature size, and scattering center distribution entropy into corresponding dense feature vectors, including: mapping each binary strong scattering center dimension, target feature size, and scattering center distribution entropy sparse feature vector through the corresponding embedding matrix to obtain the corresponding dense feature vector.

[0097] Optionally, the model building module 304 above builds an FM model, including: building a first-order module for each single feature to be related to the prediction result; building a higher-order module for the interaction of pairs of features to be related to the prediction result; and associating the first-order module and the higher-order module to obtain the FM model.

[0098] Optionally, the model building module 304 above builds an MLP model, including: building a neural network to learn the nonlinear relationship between the dense feature vector of the strong scattering center dimension, the dense feature vector of the target feature size, the dense feature vector of the scattering center distribution entropy, and the target; and determining the activation function of the MLP model.

[0099] Optionally, the prediction module 306 above trains and determines the model parameters of the FM model and the MLP model respectively, including: defining and training the corresponding objective function based on the model parameters to be determined; and using a stochastic gradient descent optimizer to train in an end-to-end manner to determine the local optimum of the objective function.

[0100] Optionally, the prediction module 306 defines and trains the corresponding objective function based on the model parameters to be determined, including: determining the squared loss between the target predicted value and the target actual value as the objective function of the model parameters.

[0101] Optionally, the prediction module 306 is trained end-to-end using a stochastic gradient descent optimizer, including updating the model parameters according to the learning rate and gradient in each iteration until the iteration ends.

[0102] Optionally, the prediction module 306 determines the local optimum of the objective function by: defining a loss function for the objective function; and obtaining the current solution of the objective function when the loss function converges or the number of iterations reaches a preset threshold.

[0103] In this embodiment, the sparse feature vectors of the strong scattering center dimension, target feature size, and scattering center distribution entropy are transformed into corresponding dense feature vectors. Parallel FM and MLP models are established. The FM model learns the feature interactions of the dense feature vectors of the strong scattering center dimension, target feature size, and scattering center distribution entropy. The MLP model learns the nonlinear relationships between these features and the target. With the model parameters of the FM and MLP models determined during training, the outputs of the FM and MLP models are fused to obtain the target. The prediction results, through a deep learning model, learn the feature interactions of the dense feature vectors of the strong scattering center dimension, the target feature size, and the dense feature vectors of the scattering center distribution entropy, as well as the nonlinear relationships between these features and the target. This effectively solves the data coefficient problems of the dense feature vectors of the strong scattering center dimension, the target feature size, and the scattering center distribution entropy, achieving accurate target prediction in sparse data scenarios. It also solves the technical problem of low accuracy in existing one-dimensional range image target prediction methods in sparse data scenarios, thus improving the prediction accuracy of one-dimensional range image targets.

[0104] Example 3

[0105] In another aspect, the present invention provides an electronic device for implementing the above-described deep learning-based one-dimensional range image target prediction method. This electronic device is not limited to a terminal device or server in a radar system. The electronic device includes, but is not limited to, a memory and a processor. The memory stores a computer program, and the processor is configured to execute the steps of any of the above method embodiments via the computer program.

[0106] Example 4

[0107] In another aspect, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in various alternative embodiments of the deep learning-based one-dimensional distance image target prediction method described above. The computer program is configured to execute the steps in any of the above method embodiments at runtime.

[0108] The exemplary embodiments described herein may be embodied in different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that the invention will be thorough and complete, and that those skilled in the art will fully understand the scope of the invention.

[0109] Where there is no conflict, the embodiments of the present invention and the features thereof can be combined with each other.

[0110] As used herein, the term “and / or” includes any and all combinations of one or more related enumerated entries.

[0111] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a” and “the” used herein are also intended to include the plural forms unless the context clearly indicates otherwise. It will also be understood that when the terms “comprising” and / or “made of” are used in this specification, the presence of the stated features, integrals, steps, operations, elements, and / or components is specified, but the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof is not excluded.

[0112] Unless otherwise specified, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art. It will also be understood that terms such as those defined in commonly used dictionaries should be understood to have the meaning consistent with their meaning in the context of the relevant art and the invention, and will not be understood to have an idealized or overly formal meaning unless expressly so defined herein.

Claims

1. A method for predicting one-dimensional distance profile targets based on deep learning, characterized in that, include: Transform the sparse feature vectors of strong scattering center dimension, target feature size, and scattering center distribution entropy into corresponding dense feature vectors; A parallel FM model and MLP model are established. The FM model is used to learn the feature interactions of the dense feature vectors of the strong scattering center dimension, the dense feature vectors of the target feature size, and the dense feature vectors of the scattering center distribution entropy. The MLP model is used to learn the nonlinear relationship between the dense feature vectors of the strong scattering center dimension, the dense feature vectors of the target feature size, the dense feature vectors of the scattering center distribution entropy, and the target. Given the model parameters of the FM model and the MLP model determined during training, the outputs of the FM model and the MLP model are fused to obtain the target prediction result.

2. The one-dimensional distance image target prediction method based on deep learning as described in claim 1, characterized in that, The sparse feature vectors representing the strong scattering center dimension, target feature size, and scattering center distribution entropy are transformed into corresponding dense feature vectors, including: By using feature embedding, the sparse feature vectors of binary strong scattering center dimension, target feature size, and scattering center distribution entropy are transformed into corresponding dense feature vectors.

3. The one-dimensional distance image target prediction method based on deep learning as described in claim 2, characterized in that, By utilizing feature embedding, the sparse feature vectors of binary strong scattering center dimension, target feature size, and scattering center distribution entropy are transformed into corresponding dense feature vectors, including: The sparse feature vectors of each binary strong scattering center dimension, target feature size, and scattering center distribution entropy are mapped to the corresponding dense feature vectors through the corresponding embedding matrix.

4. The one-dimensional distance image target prediction method based on deep learning as described in claim 1, characterized in that, Establishing an FM model includes: Establish a first-order module that relates each individual feature to the prediction result; Establish a high-order module that correlates the prediction results with pairwise feature interactions; By associating the first-order module and the higher-order module, the FM model is obtained.

5. The one-dimensional distance image target prediction method based on deep learning as described in claim 1, characterized in that, Building an MLP model includes: Establish a neural network to learn the nonlinear relationship between the dense feature vector of the strong scattering center dimension, the dense feature vector of the target feature size, the dense feature vector of the scattering center distribution entropy, and the target; Determine the activation function of the MLP model.

6. The one-dimensional distance image target prediction method based on deep learning as described in claim 1, characterized in that, Training determines the model parameters of the FM model and the MLP model, including: Define the target function for training based on the model parameters to be determined; The stochastic gradient descent optimizer is trained end-to-end to determine the local optimum of the objective function.

7. The one-dimensional distance image target prediction method based on deep learning as described in claim 6, characterized in that, Define the target function for training based on the model parameters to be determined, including: The squared loss between the target predicted value and the target actual value is determined as the objective function of the model parameters.

8. The one-dimensional distance image target prediction method based on deep learning as described in claim 6, characterized in that, The optimizer is trained end-to-end using stochastic gradient descent, including: In each iteration of training, the model parameters are updated based on the learning rate and gradient until the iteration ends.

9. The one-dimensional distance image target prediction method based on deep learning as described in claim 6 or 8, characterized in that, Determining the local optimum of the objective function includes: Define the loss function for the objective function; When the loss function converges or the number of iterations reaches a preset threshold, the current solution of the objective function is obtained.

10. A one-dimensional distance image target prediction system based on deep learning, characterized in that, include: The transformation module is used to transform sparse feature vectors of strong scattering center dimension, target feature size, and scattering center distribution entropy into corresponding dense feature vectors. The model building module is used to build parallel FM and MLP models. The FM model is used to learn the feature interactions of the strong scattering center dimension dense feature vector, the target feature size dense feature vector, and the scattering center distribution entropy dense feature vector. The MLP model is used to learn the nonlinear relationship between the strong scattering center dimension dense feature vector, the target feature size dense feature vector, the scattering center distribution entropy dense feature vector and the target. The prediction module is used to fuse the outputs of the FM model and the MLP model to obtain the target prediction result, given that the model parameters of the FM model and the MLP model have been determined during training.

Images