CN109784311B - Target identification method based on linear frequency modulation wavelet atomic network - Google Patents

Target identification method based on linear frequency modulation wavelet atomic network Download PDF

Info

Publication number
CN109784311B
CN109784311B CN201910108831.3A CN201910108831A CN109784311B CN 109784311 B CN109784311 B CN 109784311B CN 201910108831 A CN201910108831 A CN 201910108831A CN 109784311 B CN109784311 B CN 109784311B
Authority
CN
China
Prior art keywords
substep
wavelet
atomic
linear frequency
frequency modulation
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN201910108831.3A
Other languages
Chinese (zh)
Other versions
CN109784311A (en
Inventor
郭尊华
李怡霏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shandong University
Original Assignee
Shandong University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shandong University filed Critical Shandong University
Priority to CN201910108831.3A priority Critical patent/CN109784311B/en
Publication of CN109784311A publication Critical patent/CN109784311A/en
Application granted granted Critical
Publication of CN109784311B publication Critical patent/CN109784311B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Landscapes

  • Radar Systems Or Details Thereof (AREA)

Abstract

This creation provides a target identification method based on linear frequency modulation wavelet atomic network, including: step S1: training a linear frequency modulation wavelet atomic network off line; and, step S2: and (5) classifying the targets by using the linear frequency modulation wavelet atomic network obtained in the step (S1) and outputting an identification result. The device is based on a three-layer feedforward neural network structure, linear frequency modulation wavelet atoms are used as a characteristic extraction basis function of an input layer, the input layer realizes characteristic extraction through linear frequency modulation wavelet atom transformation, and a neural network classifier is formed by a hidden layer and an output layer. This creation has the following advantages: 1. richer characteristic information of the target can be obtained through linear frequency modulation wavelet atomic transformation, better effective data support is provided for a classifier, and the accuracy of target identification is improved; 2. the combined feature extraction and classification can adjust parameters in real time according to different target features and recognition environments, and improves the recognition performance and the anti-noise performance of the recognition system.

Description

Target identification method based on linear frequency modulation wavelet atomic network
Technical Field
The creation relates to the field of signal processing and pattern recognition, in particular to a target recognition method based on a linear frequency modulation wavelet atomic network.
Background
With the rapid development of modern signal processing technology and the urgent need of practical application, the automatic target recognition technology plays an increasingly important role in modern high-technology war. At present, an identification technology based on an electromagnetic scattering image is widely applied to identification of various targets in the sea, land and air, and is one of research hotspots in the field of target identification.
Automatic object recognition systems typically include two separate stages of feature extraction and classification decision. For the feature extraction method, an article "radar high-range resolution one-dimensional image target identification" on pages 35 (1): 53-60 of "system engineering and electronic technology" in Guo Zunhua, li Da, zhang Bayan 2013 is mentioned: modern signal processing techniques are usually adopted to extract effective and reliable electromagnetic scattering image features, such as frequency domain features like Fourier transform amplitude, power spectrum, bispectrum and the like, and time frequency features like wavelet transform and Gabor transform. For the classification method, pattern recognition methods are often used as classifiers, such as neural networks, support vector machines, deep learning, etc., and these pattern recognition methods are described in the following papers: DUIN R, P and KALSKA E, 2007, studies in Computational Intelligence, pages 63, INVENTION, pages 221-259, INVENTION, guo Zunhua 2018, TELECOMMUNICATIONS, pages 58 (10) 1121-1126, ONE-DIMENSIONAL convolutional neural Networks for Radar high resolution distance image Recognition, FENG B, CHEN B, LIU H W, 393, pages 61 379-393, RADIRP HRRP targeting with pages, et al, IDS, KALSKA E, page 58.
However, in the prior art, the biggest defect is that the feature extraction and the target classification stages are performed independently, the identification method can only perform the feature extraction and classification operations on the targets respectively, the feature extraction and classification parts of the identification system are difficult to achieve coordination work aiming at the complexity of the target features and the uncertainty of the identification process, and because the parameters are not synergistically optimal, the identification performance cannot be optimal when the classifier classifies according to the feature information, and the identification rate is influenced. This is the first drawback of the prior art.
In addition, in the prior art, there is a second drawback as follows: for targets with complex scattering characteristics, a certain damage can be caused to a stable scattering model by a scattering point shielding phenomenon and a moving scattering point on the target, and the traditional method is difficult to extract effective characteristics of the targets; for target identification under a complex environment background, how to ensure the identification accuracy as much as possible under a noise condition needs to be considered, that is, how to acquire characteristic information which contains richer targets and reflects the essence of the targets, so that the anti-noise performance of the identification system is improved.
In the prior art, due to the existence of the first defect and the second defect: once target misrecognition occurs, serious accidents can be caused, huge property loss is caused, and in military wars, misjudgment can cause serious casualties.
Therefore, how to quickly and accurately identify the complex environmental targets is one of the great challenges facing the research field, and is a major problem to be researched and solved urgently.
For the problem, if the recognition system can simultaneously acquire feature information of the target electromagnetic scattering center, the geometric shape, the size, the time shift, the frequency shift and the like, and select appropriate classifier parameters according to different targets and environmental conditions, a better target recognition effect is expected to be realized.
However, in a common method for extracting features by time-frequency transformation, the basic fourier transform can only obtain frequency domain features of a target, the Gabor transform can simultaneously obtain time-frequency features, and the wavelet transform contains time domain, frequency domain and scale information of the target, but cannot well describe complex scattering points.
The applicant is aware of The fact that The expanding method of wavelet Transform method proposed by MANN S and HAYKIN S1995 in IEEE transformations on Signal Processing 43 (11): 2745-2761 page [ The chirp Transform: physical configurations-chirp wavelet atomic Transform ] adds chirp parameters on The basis of wavelet Transform, can obtain a series of Transform coefficients of different time-frequency plane slopes, and thus can obtain more complete characteristic information of The target.
However, this will again present new challenges: the richer the extracted feature information is, the more challenge the recognition system can bring to the classification of the features, and how to select a proper classifier becomes the key to success or failure of recognition.
Disclosure of Invention
One purpose of the present authoring is to implement feature extraction and object recognition for objects that are relatively complex with respect to scattering characteristics.
In order to achieve the purpose, the method adopts the linear frequency modulation wavelet atomic transformation as a characteristic extraction function of the neural network to extract the target time-frequency-scale-linear frequency modulation four-dimensional space characteristic.
The applicant considers that the time-frequency joint domain description can more accurately reflect the target characteristics, and the key of the time-frequency characteristic extraction lies in how to determine a group of optimal basis functions for determining signal classification. Because the linear frequency modulation wavelet transformation is popularized from a three-dimensional time-frequency-scale space to a four-dimensional time-frequency-scale-linear frequency modulation space, linear frequency modulation wavelet atoms contain richer characteristic information than wavelets and Gabor atoms, transformation parameters of the linear frequency modulation wavelet atoms are related to target physical characteristics, time-frequency local performance is more excellent, and extraction of complex target characteristics is more advantageous. Therefore, the creation adopts the linear frequency modulation wavelet atomic transformation as a feature extraction function of the neural network to extract the time-frequency-scale-linear frequency modulation four-dimensional spatial feature of the target.
Another objective of the present creation is how to ensure the recognition accuracy as much as possible under noisy conditions, i.e. to improve the noise immunity and recognition performance of the recognition system.
In order to achieve the purpose, the method collaboratively adjusts the atomic parameters of the linear frequency modulation wavelet and the parameters of the classifier, and jointly realizes feature extraction and target classification.
The applicant considers that, when identifying the features, the parameters of the classification algorithm also influence the identification performance of the target to a certain extent. Therefore, the creation optimizes and adjusts the parameters of the features while training the classifier, realizes synchronous training, can coordinate the parameters according to different target data, and can obtain better recognition effect and noise resistance than a recognition system with mutually independent feature extraction and target classification. Therefore, the creation organically combines the chirp wavelet atoms with a multilayer feedforward neural network with the advantages of strong self-adaptive learning capacity, distributed storage, parallel cooperative processing and the like, and provides a chirp wavelet atom network method for cooperatively adjusting chirp wavelet atom parameters and classifier parameters to realize combined feature extraction and classification of targets.
In order to achieve the above purpose, the present creation provides a target identification method based on a chirp wavelet atomic network, which adopts a chirp wavelet atomic network to perform target identification. The chirp wavelet atomic network is based on a three-layer feedforward neural network structure shown in fig. 2, chirp wavelet atoms are used as a feature extraction basis function of an input layer, the input layer realizes feature extraction through chirp wavelet atomic transformation, and a neural network classifier is formed by a hidden layer and an output layer.
The object identification method based on the linear frequency modulation wavelet atomic network comprises the following steps:
step S1: training a linear frequency modulation wavelet atomic network off line; and the number of the first and second groups,
step S2: classifying the targets by using the linear frequency modulation wavelet atomic network obtained in the step S1 and outputting an identification result;
wherein, step S1 comprises substep S11 of inputting a training sample vector x n And randomly initializing a chirp wavelet atomic parameter set beta of a chirp wavelet atomic network k ={u kk ,s k ,c k And classifier weights
Figure BDA0001967316190000031
And &>
Figure BDA0001967316190000032
Wherein N =1,2.. N is the number of samples, K =1,2.. K is the number of input layer atoms, H =1,2.. H is the number of hidden layer nodes, M =1,2.. M is the number of output layer nodes; beta is a beta k ={u kk ,s k ,c k Includes a time shift parameter u k Frequency shift parameter xi k Size parameter s k And linear frequency c k
Figure BDA0001967316190000033
The connection weight between the input layer and the hidden layer is used; />
Figure BDA0001967316190000034
The weight of the connection between the hidden layer and the output layer.
Substep S12, using the data sequence t according to equation 1 and equation 2 l And chirp wavelet atomic parameter set beta k To calculate the atoms of the chirp wavelet,
Figure BDA0001967316190000035
and forming linear frequency-modulated wavelet atomic vector g based on the linear frequency-modulated wavelet atomic vector k
Wherein, the linear frequency modulation wavelet atom parameter set beta k Can be obtained from substep S11;
wherein, t l Is a data sequence, where L =1,2, and L is the sample length; g (t) is the basic window function of the chirp wavelet transform:
Figure BDA0001967316190000041
wherein, the linear frequency modulation wavelet atomic vector
Figure BDA0001967316190000042
Superscript T is the matrix transpose, and e is the natural constant.
Substep S13, using the chirp wavelet atomic vector g obtained in substep S12, according to equation 3 k And the training sample vector x input in sub-step S11 n Calculating the characteristic value phi of all samples at each atomic node nk
Figure BDA0001967316190000043
A substep S14 of obtaining a characteristic value phi of the sample from the substep S13 nk The classifier of the input neural network calculates the output result o of the hidden layer according to the formula 4-7 by the activation function nh And output result y of the output layer nm
Figure BDA0001967316190000044
Figure BDA0001967316190000045
Wherein N =1,2.. N is the number of samples; h =1,2,.. H is the number of hidden nodes, m =1,2, M is the number of output layer nodes; intermediate variables
Figure BDA0001967316190000046
And &>
Figure BDA0001967316190000047
The calculation formula of (a) is expressed as:
Figure BDA0001967316190000048
Figure BDA0001967316190000049
/>
substep S15, comprising substeps S15A-S15B;
in which, in sub-step S15A, the desired output d of the sample is made nm Is 1 or 0; wherein, d nm 1 belongs to a preset target; d nm 0 does not belong to the preset target;
substep S15B of outputting d according to the desire nm And the actual output y nm Using equation 8, the mean square error E is calculated:
Figure BDA00019673161900000410
and a substep S16 of judging whether the expected effect is achieved, and finishing the following steps if the expected effect is achieved: comparing the mean square error E with a preset threshold; if the mean square error E is less than the preset threshold value, the learning process is ended, and the substep S18 is switched to; if the mean square error E is more than or equal to a preset threshold value, continuing the following substeps;
substep S17, weight adjustment, and adjustment of chirp wavelet atomic parameters, and then iterative loop:
substep S18: preserving chirp wavelet atomic parameters beta k ={u kk ,s k ,c k And the connection weight between the input layer and the hidden layer
Figure BDA0001967316190000051
And the connection weight between hidden layer and output layer->
Figure BDA0001967316190000052
And finishing the training.
The substep S17 comprises substeps S17A-S17C;
substep S17A of reversely adjusting the connection weight between the input layer and the hidden layer by using a Levenberg-Marquardt algorithm
Figure BDA0001967316190000053
And the connection weight between hidden layer and output layer->
Figure BDA0001967316190000054
Substep S17B, adjusting the set of atomic node parameters β according to the gradient descent algorithm using equation 9 k ={u kk ,s k ,c k }:
Figure BDA0001967316190000055
Wherein n is the iteration number, eta is the learning rate,
Figure BDA0001967316190000056
is a partial derivative operator, E is a mean square error;
substep S17C, iterative loop: returning to the substep S12, and performing the next training period;
wherein, in the next training period, the set of chirp wavelet atomic parameters β for sub-step S12 k Is obtained from substep S17B.
The target identification method based on the linear frequency modulation wavelet atomic network comprises the following steps of S2, classifying targets by using the linear frequency modulation wavelet atomic network obtained in the S1 and outputting identification results; wherein, step S2 includes the following substeps:
substep S21: inputting data and preprocessing the data;
the substep S21 comprises substeps S21A-S21C; the execution sequence of the substeps S21A-S21C is arbitrary;
substep S21A:
input training sample vector x n And data preprocessing is carried out; wherein N =1,2., N is the number of samples,
substep S21B:
inputting the linear frequency modulation wavelet atom parameter set beta of the linear frequency modulation wavelet atom network obtained according to the step S1 k ={u kk ,s k ,c k Where K =1,2, K is the number of input layer atoms, β k ={u kk ,s k ,c k Includes a time shift parameter u k Frequency shift parameter xi k Scale parameter s k And linear frequency c k
Substep S21C:
inputting the classifier weight of the linear frequency modulation wavelet atomic network obtained according to the step S1
Figure BDA0001967316190000057
And &>
Figure BDA0001967316190000058
Wherein the classifier weights of a chirp wavelet atomic network->
Figure BDA0001967316190000059
And &>
Figure BDA00019673161900000510
Comprising connection weights between an input layer and a hidden layer>
Figure BDA00019673161900000511
And the connection weight between hidden layer and output layer->
Figure BDA00019673161900000512
/>
Substep S22, according toEquation 1 and equation 2, using a data sequence t l And the set of chirp wavelet atomic parameters beta input at substep S21 k To calculate the atoms of the chirp wavelet,
Figure BDA0001967316190000061
and forming a chirp wavelet atomic vector g therefrom k
Wherein, t l Is a data sequence, where L =1,2, and L is the sample length; wherein g (t) is the basic window function of the chirp wavelet transform:
Figure BDA0001967316190000062
wherein, the linear frequency modulation wavelet atomic vector
Figure BDA0001967316190000063
Superscript T is the matrix transpose, and e is the natural constant.
Substep S23: using the chirp wavelet atomic vector g obtained in sub-step S22 according to equation 3 k And the training sample vector x input in sub-step S21 n Calculating the characteristic value phi of all samples at each atomic node nk
Figure BDA0001967316190000064
Substep S24: the characteristic value phi of the sample obtained from the substep S23 nk The classifier of the input neural network calculates the output result o of the hidden layer according to the formula 4-7 by the activation function nh And output result y of the output layer nm : according to the output result y of the output layer nm To determine whether the sample belongs to the p-th class target:
Figure BDA0001967316190000065
Figure BDA0001967316190000066
wherein the intermediate variable
Figure BDA0001967316190000067
And &>
Figure BDA0001967316190000068
The calculation formula of (2) is expressed as:
Figure BDA0001967316190000069
Figure BDA00019673161900000610
substep S25: and outputting a classification result to finish the identification.
The beneficial effect of this creation is:
compared with the prior art, the creation has the following remarkable advantages:
1. richer characteristic information of the target can be obtained through linear frequency modulation wavelet atomic transformation, better effective data support is provided for a classifier, and certain advantages are achieved for improving the accuracy of target identification of a system;
2. the combined feature extraction and classification can adjust parameters in real time according to different target features and recognition environments, and the recognition performance and the anti-noise performance of the recognition system are improved.
Drawings
FIG. 1 is a diagram of an implementation of a target recognition technique that combines feature extraction and classification;
FIG. 2 is a diagram of a chirp wavelet atomic network structure;
FIG. 3A shows the recognition rate of Gabor atomic network based on different azimuth data of radar under different SNR conditions;
FIG. 3B is a graph showing the recognition rate of a chirp wavelet atomic network based on different azimuth data of a radar under different signal-to-noise ratios;
description of reference numerals: 1-input vector (sample data set), 2-vector product, 3-input layer, 4-quantity product, 5-hidden layer, 6-quantity product, 7-output layer, 8-output (classification result), 9-Sigmoid activation function, 10-connection weight between hidden layer and output layer, 11-Sigmoid activation function, 12-connection weight between input layer and hidden layer, 13-characteristic value and 14-chirp wavelet atomic vector set.
Detailed Description
In order to make the technical means, creation features, achievement purposes and effects of the creation easier to understand, the creation is further described below with reference to the accompanying drawings and the detailed description.
The applicant considers that the neural network has the advantages of strong self-adaptive learning capability, distributed storage, parallel cooperative processing and the like through multi-party screening and multiple simulation tests, and has certain advantages for the learning of complex data, so that the neural network can be used as a classifier to realize the classification of complex target characteristics.
For how to improve the anti-noise performance of the identification system, on one hand, the applicant considers that the target data needs to be preprocessed, and noise is reduced as much as possible;
on the other hand, the recognition system should be able to adapt adaptively to different environmental conditions. According to the Wavelet atomic Network articles of Wavelet Neural Networks of pages a Practical guide, 2013 and SHI Y, ZHANG X D, IEEE Transactions on Signal Processing 2001 at 49 (12) of pages 2994-3004, proposed by ALEXANDRIDIS A K and ZAPRANIS A D.2013 at 42.
By combining the factors, the applicant finally considers that a method of combining linear frequency modulation wavelet atomic transformation and a neural network is adopted to extract four-dimensional time-frequency space effective information of the target and synchronously realize the classification of the target.
Aiming at the technical problems of how to extract effective target features and designing a reliable classification algorithm in the two target identification fields, the creation provides a method capable of realizing combined feature extraction and target classification: the linear frequency modulation wavelet atomic network is used for improving the recognition rate and the anti-noise performance of a target recognition system and reducing the great loss caused by misidentification when recognition technology is adopted in real life.
The method adopts the linear frequency modulation wavelet atomic transformation to extract the time-frequency-scale-linear frequency modulation rate four-dimensional spatial characteristics of the target signal, and adjusts the linear frequency modulation wavelet atomic parameters and the classifier parameters in a coordinated manner, so that the method has higher recognition rate and better anti-noise performance.
The creation provides a target identification method based on a linear frequency modulation wavelet atomic network, and the linear frequency modulation wavelet atomic network is adopted for target identification. The chirp wavelet atomic network is based on a three-layer feed-forward neural network as shown in figure 2. FIG. 2 is a diagram of a chirp wavelet atomic network structure; the method comprises the following steps of 1-inputting a vector (sample data set), 2-vector product, 3-inputting layer, 4-quantity product, 5-hiding layer, 6-quantity product, 7-outputting layer, 8-outputting (classification result), 9-Sigmoid activation function, 10-connection weight between the hiding layer and the outputting layer, 11-Sigmoid activation function, 12-connection weight between the inputting layer and the hiding layer, 13-characteristic value and 14-linear frequency modulation wavelet atom vector set.
As shown in fig. 1, the present creation proposes a target identification method based on a chirp wavelet atomic network, including:
step S1, off-line training:
training a linear frequency modulation wavelet atomic network off line; and the number of the first and second groups,
step S2, online identification:
and classifying the targets by using the linear frequency modulation wavelet atomic network obtained in the step S1 and outputting an identification result.
For the steps, P linear frequency modulation wavelet atomic networks can be simultaneously realized in parallel to obtain characteristic parameters of the P-type targets.
P =1,2,.., P, step S1 of off-line training of a chirp wavelet atomic network of a class P target, includes the following sub-steps:
substep S11, network initialization:
as shown in the input vector (sample data set) 1 of FIG. 2, a training sample vector x is input n And randomly initializing a chirp wavelet atomic parameter set beta of a chirp wavelet atomic network k ={u kk ,s k ,c k } and classifier weights
Figure BDA0001967316190000081
And &>
Figure BDA0001967316190000091
Wherein N =1,2., N is the number of samples, K =1,2., K is the number of input layer atoms, H =1,2., H is the number of hidden layer nodes, M =1,2., M is the number of output layer nodes;
wherein, the linear frequency modulation wavelet atom parameter set beta of the linear frequency modulation wavelet atom network k ={u kk ,s k ,c k Includes a time shift parameter u k Frequency shift parameter xi k Size parameter s k And linear frequency c k (ii) a Wherein the time shift parameter u k Representing the time centre of the signal, the frequency shift parameter xi k Representing the center of frequency of the signal, the scale parameter s k Representing the stretching of the atoms of chirped wavelets, chirp c k Representing the slope of the signal on a time-frequency plane;
wherein the classifier weights of the chirp wavelet atomic network
Figure BDA0001967316190000092
And &>
Figure BDA0001967316190000093
Is the connection weight of each layer of the network, including an input layer and a hidden layerIs connected with the weight->
Figure BDA0001967316190000094
And the connection weight between hidden layer and output layer->
Figure BDA0001967316190000095
Substep S12, calculating chirp wavelet atoms:
using the data sequence t according to equation 1 and equation 2 l And a linear frequency modulation wavelet atomic parameter set beta k To calculate the atoms of the chirp wavelet,
Figure BDA0001967316190000096
and forming a chirp wavelet atomic vector g therefrom k
Wherein, the linear frequency modulation wavelet atom parameter set beta k Can be derived from substep S11-when substep S17C of the iterative loop is not started, the set of chirp wavelet atomic parameters β k Is derived from substep S11; after the start of iteration loop substep S17C, a set of chirp wavelet atomic parameters β k Is obtained from substep S17B.
Wherein, formula 1 is a linear frequency modulation wavelet atomic expression; t is t l Is a data sequence, where L =1,2, and L is the sample data length;
wherein g (t) is a basic window function of the chirp wavelet transform, a gaussian window function is generally adopted as the basic window function of the chirp wavelet atomic transform, and the expression is as follows:
Figure BDA0001967316190000097
/>
wherein, the linear frequency modulation wavelet atomic vector
Figure BDA0001967316190000098
Wherein, the linear frequency modulation wavelet atom is taken as the characteristic extraction basis function of the input layer 3 in figure 2, the input layer 3 realizes the characteristic extraction through the linear frequency modulation wavelet atom transformation,
wherein, the superscript T is the matrix transposition, e is the natural constant;
substep S13, calculating a feature value of the input layer 3 as shown in fig. 2, to realize feature extraction:
using the chirp wavelet atomic vector g obtained in sub-step S12 according to equation 3 k And the training sample vector x input in sub-step S11 n Calculating the characteristic value phi of all samples at each atomic node nk
Figure BDA0001967316190000101
I.e. the characteristic value phi of the sample nk Is a linear frequency-modulated wavelet atomic vector
Figure BDA0001967316190000102
And training sample vector x n =[x n (t 1 ),x n (t 2 ),...,x n (t l )] T Absolute value of the vector inner product of (1); this is by using a chirp wavelet atomic vector g k For sample vector x n Carrying out linear frequency modulation wavelet atomic transformation;
wherein, the superscript T is the matrix transposition, e is the natural constant;
substep S14, calculating the output o of hidden layer 5 of FIG. 2 nh And output result y of output layer 7 of FIG. 2 nm
The characteristic value phi of the sample obtained from substep S13 nk The classifier of the input neural network calculates the output result o of the hidden layer according to the formula 4-7 through the activation functions of the node 11 of the hidden layer 5 and the node 9 of the output layer 7 nh And output result y of the output layer nm
Figure BDA0001967316190000103
Figure BDA0001967316190000104
Wherein, a neural network classifier is formed by a hidden layer 5 and an output layer 7; in the creation, the characteristic value parameters of the input layer 3 and the neural network classifier parameters of the hidden layer 5 and the output layer 7 are coordinated and optimized; in the present creation, feature extraction (see substep S13) and object classification (S14-S16) of a chirp wavelet atomic network are jointly implemented.
Wherein N =1,2.. N is the number of samples; h =1,2,.. H is the number of hidden layer nodes, M =1,2,. Wherein M is the number of output layer nodes; the activation function adopts a Sigmoid function;
intermediate variables
Figure BDA0001967316190000105
And &>
Figure BDA0001967316190000106
The calculation formula of (2) is expressed as:
Figure BDA0001967316190000107
Figure BDA0001967316190000108
substep S15, calculating the mean square error E:
substep S15, comprising substeps S15A-S15B;
substep S15A of letting the desired output d of the sample be nm Is 1 or 0; wherein d is nm 1 belongs to a preset p-th class target; d nm 0 does not belong to the preset p-th class target;
substep S15B, outputting d according to the desire nm And the actual output y nm Using equation 8, the mean square error E is calculated:
Figure BDA0001967316190000111
and a substep S16 of judging whether the expected effect is achieved, and finishing the following steps if the expected effect is achieved:
comparing the mean square error E with a preset threshold;
if the mean square error E is smaller than the preset threshold, ending the learning process, and going to substep S18;
if the mean square error E is more than or equal to a preset threshold value, continuing the following substeps;
substep S17, weight adjustment, adjustment of linear frequency modulation wavelet atomic parameters, and iteration loop:
substep S17, comprising substeps S17A-S17C;
in the creation, the automatic adjustment of classifier weight (substep S17A) and the automatic adjustment of chirp wavelet atomic time-frequency parameters (substep S17B) are performed jointly. Reverse layer-by-layer adjustment: output layer > hidden layer > input layer.
Substep S17A, weight adjustment:
method for reversely adjusting connection weight of each layer of network by utilizing Levenberg-Marquardt algorithm
Figure BDA0001967316190000112
And &>
Figure BDA0001967316190000113
Substep S17B, adjusting the chirp wavelet atomic parameters:
adjusting the atomic node parameter set beta according to a gradient descent algorithm using equation 9 k ={u kk ,s k ,c k }:
Figure BDA0001967316190000114
Wherein n is the iteration number, eta is the learning rate,
Figure BDA0001967316190000115
is a partial derivative operator, E is a mean square error; the gradient descent algorithm is based on a delta learning rule;
substep S17C, iterative loop:
returning to the substep S12, performing the next training period;
substep S17C, iterative loop: returning to the substep S12, and performing the next training period;
wherein, when the iterative loop is not started, the linear frequency modulation wavelet atomic parameter set beta k Is derived from substep S11; after starting the iterative loop, in the next training period, the set of chirp wavelet atomic parameters β for sub-step S12 k Is obtained from sub-step S17B.
Substep S18: preserving chirp wavelet atomic parameters beta k ={u kk ,s k ,c k Connection weights with layers of the network
Figure BDA0001967316190000116
And &>
Figure BDA0001967316190000117
And finishing the training.
After the offline recognition step S1 is completed, an online recognition step S2 may be performed.
P =1,2,.. P, identification of a chirp wavelet atomic network for a class P object is as follows:
step S2, online identification:
and classifying the targets by using the linear frequency modulation wavelet atomic network obtained in the step S1 and outputting an identification result.
Accordingly, in step S2, the identification of the p-th class object includes the following sub-steps:
substep S21: inputting data and preprocessing the data;
a substep S21 comprising substeps S21A-S21C; the execution sequence of the substeps S21A-S21C is arbitrary;
substep S21A:
as shown in the input vector (sample data set) 1 of FIG. 2, a training sample vector x is input n And data preprocessing is carried out;
wherein N =1,2., N is the number of samples,
substep S21B:
inputting the linear frequency modulation wavelet atom parameter set beta of the linear frequency modulation wavelet atom network obtained according to the step S1 k ={u kk ,s k ,c k }
Wherein K =1,2.. K is the input layer atom number, the chirp wavelet atom parameter set β of the chirp wavelet atom network k ={u kk ,s k ,c k Includes a time shift parameter u k Frequency shift parameter xi k Size parameter s k And linear frequency c k (ii) a Wherein the time shift parameter u k Representing the time centre of the signal, the frequency shift parameter xi k Representing the center of frequency of the signal, the scale parameter s k Representing the extensional variation of the chirp wavelet atom, chirp frequency c k Representing the slope of the signal on a time-frequency plane;
substep S21C:
inputting the classifier weight of the linear frequency modulation wavelet atomic network obtained according to the step S1
Figure BDA0001967316190000121
And &>
Figure BDA0001967316190000122
Wherein the classifier weights of the chirp wavelet atomic network
Figure BDA0001967316190000123
And &>
Figure BDA0001967316190000124
Is the connection weight of each layer of the network, including the connection weight between the input layer and the hidden layer>
Figure BDA0001967316190000125
And the connection weight between hidden layer and output layer->
Figure BDA0001967316190000126
Substep S22, calculating chirp wavelet atoms:
using the data sequence t according to equation 1 and equation 2 l And the set of chirp wavelet atomic parameters beta input at substep S21 k To calculate the atoms of the chirp wavelet,
Figure BDA0001967316190000127
and forming a chirp wavelet atomic vector g therefrom k
Wherein, formula 1 is a linear frequency modulation wavelet atomic expression; t is t l Is a data sequence, where L =1,2, and L is the sample data length;
wherein g (t) is a basic window function of the chirp wavelet transform, a gaussian window function is generally adopted as the basic window function of the chirp wavelet atomic transform, and the expression is as follows:
Figure BDA0001967316190000131
wherein, the linear frequency modulation wavelet atomic vector
Figure BDA0001967316190000132
Wherein, the linear frequency modulation wavelet atom is taken as the characteristic extraction basis function of the input layer 3 in figure 2, the input layer 3 realizes the characteristic extraction through the linear frequency modulation wavelet atom transformation,
wherein, the superscript T is the matrix transposition, e is the natural constant;
substep S23: the eigenvalues φ of all samples at each atomic node of the input layer 3 of FIG. 2 are calculated according to equations 1-3 nk
Using the value obtained in substep S22 according to equation 3To linear frequency-modulated wavelet atomic vector g k And the training sample vector x input in sub-step S21 n Calculating the characteristic value phi of all samples at each atomic node nk
Figure BDA0001967316190000133
I.e. the characteristic value phi of the sample nk Is a linear frequency-modulated wavelet atomic vector
Figure BDA0001967316190000134
And training sample vector x n =[x n (t 1 ),x n (t 2 ),...,x n (t l )] T Absolute value of the vector inner product of (1); this is by using a chirp wavelet atomic vector g k For sample vector x n Carrying out linear frequency modulation wavelet atomic transformation;
substep S24, calculating the output o of hidden layer 5 of FIG. 2 nh And output result y of output layer 7 nm Judging whether the sample belongs to the p-th class target:
the characteristic value phi of the sample obtained from the substep S23 nk The classifier of the input neural network calculates the output result o of the hidden layer according to the formula 4-7 through the activation functions of the node 11 of the hidden layer 5 and the node 9 of the output layer 7 nh And output result y of the output layer nm : according to the output result y of the output layer nm To determine whether the sample belongs to the p-th class target:
Figure BDA0001967316190000135
Figure BDA0001967316190000136
wherein the activation function adopts a Sigmoid function;
intermediate variables
Figure BDA0001967316190000141
And &>
Figure BDA0001967316190000142
The calculation formula of (2) is expressed as:
Figure BDA0001967316190000143
Figure BDA0001967316190000144
so far, the task of classifying the target according to each sample output value is completed, and the following steps are carried out:
substep S25: outputting a classification result:
and outputting a classification result to finish the identification.
To illustrate the beneficial effects of the present creation, the following four examples are listed. Wherein examples 1-3 are: the identification test is carried out by a control variable method (azimuth angle and signal-to-noise ratio). Example 4 is a comprehensive test, and the test results are shown in fig. 3.
Example 1
The object recognition method provided by the invention is used for training and testing a linear frequency modulation wavelet atomic network, four types of electromagnetic scattering images with azimuth angles of 0-45 degrees and signal-to-noise ratio of 30dB are used as sample data, and compared with the existing method (a back propagation neural network, a wavelet neural network and a Gabor atomic network), the recognition rate result shown in table 1 is obtained. This example shows that the present creation has better recognition performance for radar narrow-angle scattering images.
TABLE 1 azimuthal angle 0-45 deg. range data discrimination (%)
Figure BDA0001967316190000145
Example 2
The four networks described in example 1 were trained and tested using electromagnetic scattering images of four types of targets with azimuth angles of 0-180 ° and signal-to-noise ratios of 30dB as sample data to obtain the recognition rate results shown in table 2. This embodiment shows that the present creation basically has a relatively high recognition rate for the radar wide-angle scattering image compared to other prior art target recognition methods.
TABLE 2 azimuthal angle range data discrimination (%) -180
Figure BDA0001967316190000146
Figure BDA0001967316190000151
Example 3
The anti-noise performance test is performed on the four networks obtained by training in example 2 by using electromagnetic scattering images of four types of targets with azimuth angles of 0-180 degrees and signal-to-noise ratios of 5dB as test sample data, and the recognition rate results shown in Table 3 are obtained. This example shows that the present creation has good noise immunity for radar scattering image identification.
TABLE 3 SNR 5dB data discrimination (%)
Figure BDA0001967316190000152
Example 4
In the prior art, electromagnetic scattering images of four types of targets with azimuth angles of 0-45 degrees, 0-90 degrees, 0-180 degrees and signal-to-noise ratios of 30 dB-5 dB are used as sample data, and comprehensive training and testing are performed on a Gabor atomic network as shown in figure 3A, so that average recognition rate results of the four types of targets are obtained.
In example 4, electromagnetic scattering images of four types of targets with azimuth angles of 0 to 45 °, 0 to 90 °, 0 to 180 °, and signal-to-noise ratios of 30dB to 5dB are used as sample data, and as shown in fig. 3B, comprehensive training and testing are performed on a linear frequency modulation wavelet atomic network to obtain an average recognition rate result of the four types of targets.
This embodiment 4 shows that, with the increase of image noise, the radar scattering image recognition performance of the present creation is reduced to a smaller extent, and the recognition rate is overall higher than that of the existing Gabor atomic network recognition method.
In conclusion, the beneficial effects of the creation are:
compared with the prior art, the creation has the following remarkable advantages:
1. richer characteristic information of the target can be obtained through linear frequency modulation wavelet atomic transformation, better effective data support is provided for a classifier, and certain advantages are achieved for improving the accuracy of target identification of a system;
2. the combined feature extraction and classification can adjust parameters in real time according to different target features and recognition environments, and improves the recognition performance and the anti-noise performance of the recognition system.
The above description is intended to be illustrative, and not restrictive, and it will be understood by those skilled in the art that many modifications, variations, or equivalents may be made without departing from the spirit and scope of the present disclosure as defined in the following claims.

Claims (10)

1. A target identification method based on a linear frequency modulation wavelet atomic network is characterized by comprising the following steps:
step S1: training a linear frequency modulation wavelet atomic network off line; and the number of the first and second groups,
step S2: classifying the targets by using the linear frequency modulation wavelet atomic network obtained in the step S1 and outputting an identification result;
wherein, step S1 comprises substep S11 of inputting a training sample vector x n And randomly initializing a chirp wavelet atomic parameter set beta of a chirp wavelet atomic network k ={u kk ,s k ,c k And classifier weights
Figure FDA0001967316180000011
And &>
Figure FDA0001967316180000012
Wherein N =1,2, …, N is the number of samples, K =1,2.. K is the number of input layer atoms, H =1,2.. H is the number of hidden layer nodes, M =1,2.. M is the number of output layer nodes; beta is a beta k ={u kk ,s k ,c k Includes a time shift parameter u k Frequency shift parameter xi k Scale parameter s k And linear frequency c k
Figure FDA0001967316180000013
The connection weight between the input layer and the hidden layer is obtained; />
Figure FDA0001967316180000014
The weight of the connection between the hidden layer and the output layer.
2. The chirp wavelet atomic network-based object identification method of claim 1, wherein after substep S11, step S1 further comprises:
substep S12, using the data sequence t according to equation 1 and equation 2 l And chirp wavelet atomic parameter set beta k To calculate the atoms of the chirp wavelet,
Figure FDA0001967316180000015
and forming a chirp wavelet atomic vector g therefrom k
Wherein, the linear frequency modulation wavelet atom parameter set beta k Can be obtained from substep S11;
wherein, t l Is a data sequence, where L =1,2, and L is the sample length; g (t) is the basic window function of the chirp wavelet transform:
Figure FDA0001967316180000016
wherein, the linear frequency modulation wavelet atomic vector
Figure FDA0001967316180000017
The superscript T is the matrix transpose and e is the natural constant.
3. The chirp wavelet atomic network-based object identification method of claim 2, wherein after substep S12, step S1 further comprises:
substep S13, using the chirp wavelet atomic vector g obtained in substep S12, according to equation 3 k And the training sample vector x input in sub-step S11 n Calculating the characteristic value phi of all samples at each atomic node nk
Figure FDA0001967316180000021
4. A method for identifying an object based on a chirp wavelet atomic network as claimed in claim 3, wherein after substep S13, step S1 further comprises:
a substep S14 of obtaining a characteristic value phi of the sample from the substep S13 nk The classifier of the input neural network calculates the output result o of the hidden layer according to the formula 4-7 by the activation function nh And output result y of the output layer nm
Figure FDA0001967316180000022
Figure FDA0001967316180000023
/>
Wherein N =1,2.. N is the number of samples; h =1,2,.. H is the number of hidden nodes, M =1,2,. Wherein M is the output layer nodeCounting; intermediate variables
Figure FDA0001967316180000024
And &>
Figure FDA0001967316180000025
The calculation formula of (a) is expressed as:
Figure FDA0001967316180000026
Figure FDA0001967316180000027
5. the chirp wavelet atomic network-based object identification method of claim 4, wherein after substep S14, step S1 further comprises substep S15:
substep S15, comprising substeps S15A-S15B;
wherein, in the substep S15A, the expected output d of the sample is made nm Is 1 or 0; wherein, d nm 1 belongs to a preset target; d nm 0 does not belong to the preset target;
substep S15B, outputting d according to the desire nm And the actual output y nm Using equation 8, the mean square error E is calculated:
Figure FDA0001967316180000028
6. the chirp wavelet atomic network-based object identification method as claimed in claim 5, wherein after substep S15, step S1 further comprises substeps S16-S18,
and a substep S16 of judging whether the expected effect is achieved, and finishing the following steps if the expected effect is achieved: comparing the mean square error E with a preset threshold; if the mean square error E is smaller than the preset threshold, ending the learning process, and going to substep S18; if the mean square error E is more than or equal to a preset threshold value, continuing the following substeps;
substep S17, weight adjustment, adjustment of linear frequency modulation wavelet atomic parameters, and iteration loop:
substep S18: preserving chirp wavelet atomic parameters beta k ={u kk ,s k ,c k And the connection weights between the input and hidden layers
Figure FDA0001967316180000031
And connection weights between hidden and output layers>
Figure FDA0001967316180000032
And finishing the training.
7. The method for identifying an object based on a chirp wavelet atomic network as claimed in claim 6, wherein the substep S17 comprises substeps S17A-S17C;
substep S17A: reversely adjusting connection weight between input layer and hidden layer by utilizing Levenberg-Marquardt algorithm
Figure FDA0001967316180000033
And the connection weight between hidden layer and output layer->
Figure FDA0001967316180000034
Substep S17B: adjusting the atomic node parameter set beta according to a gradient descent algorithm using equation 9 k ={u kk ,s k ,c k }:
Figure FDA0001967316180000035
Wherein n is the iteration number, eta is the learning rate,
Figure FDA0001967316180000036
is a partial derivative operator, E is a mean square error;
substep S17C: returning to the substep S12, and performing the next training period;
wherein, in the next training period, the set of chirp wavelet atomic parameters β for sub-step S12 k Is obtained from substep S17B.
8. The method for identifying objects based on a chirp wavelet atomic network as claimed in claim 7, wherein the step S2 comprises the sub-steps of:
substep S21: inputting data and preprocessing the data;
the substep S21 comprises substeps S21A-S21C; the execution sequence of the substeps S21A-S21C is arbitrary;
substep S21A:
input training sample vector x n And carrying out data preprocessing; wherein N =1,2, …, N is the number of samples,
substep S21B:
inputting the linear frequency modulation wavelet atom parameter set beta of the linear frequency modulation wavelet atom network obtained according to the step S1 k ={u kk ,s k ,c k Where K =1,2, K is the number of input layer atoms, β k ={u kk ,s k ,c k Includes a time shift parameter u k Frequency shift parameter xi k Size parameter s k And linear frequency c k
Substep S21C:
inputting the classifier weight of the linear frequency modulation wavelet atomic network obtained according to the step S1
Figure FDA0001967316180000037
And &>
Figure FDA0001967316180000038
In which the division of the chirp wavelet atomic networkWeight of analog device>
Figure FDA0001967316180000039
And &>
Figure FDA00019673161800000310
Comprising a connection weight between an input layer and a hidden layer>
Figure FDA00019673161800000311
And the connection weight between hidden layer and output layer->
Figure FDA00019673161800000312
9. The chirp wavelet atomic network-based object identification method of claim 8, wherein after the sub-step S21, the step S2 further comprises the sub-steps of:
substep S22, using the data sequence t according to equation 1 and equation 2 l And the set of chirp wavelet atomic parameters beta input at substep S21 k To calculate the atoms of the chirp wavelet,
Figure FDA0001967316180000041
and forming a chirp wavelet atomic vector g therefrom k
Wherein, t l Is a data sequence, wherein L =1,2, …, and L is a sample length; wherein g (t) is the basic window function of the linear frequency modulation wavelet transform:
Figure FDA0001967316180000042
wherein, the linear frequency modulation wavelet atomic vector
Figure FDA0001967316180000043
The superscript T is the matrix transpose and e is the natural constant.
10. The chirp wavelet atomic network-based object identification method of claim 9, wherein after substep S22, step S2 further comprises the substeps of:
substep S23: using the chirp wavelet atomic vector g obtained in sub-step S22 according to equation 3 k And the training sample vector x input in sub-step S21 n Calculating the characteristic value phi of all samples at each atomic node nk
Figure FDA0001967316180000044
Substep S24: the characteristic value phi of the sample obtained from the substep S23 nk The classifier of the input neural network calculates the output result o of the hidden layer according to the formula 4-7 by the activation function nh And output result y of the output layer nm : according to the output result y of the output layer nm To determine whether the sample belongs to the p-th class target:
Figure FDA0001967316180000045
Figure FDA0001967316180000046
wherein the intermediate variable
Figure FDA0001967316180000047
And &>
Figure FDA0001967316180000048
The calculation formula of (2) is expressed as: />
Figure FDA0001967316180000049
Figure FDA00019673161800000410
Substep S25: and outputting a classification result to finish the identification.
CN201910108831.3A 2019-02-03 2019-02-03 Target identification method based on linear frequency modulation wavelet atomic network Active CN109784311B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910108831.3A CN109784311B (en) 2019-02-03 2019-02-03 Target identification method based on linear frequency modulation wavelet atomic network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910108831.3A CN109784311B (en) 2019-02-03 2019-02-03 Target identification method based on linear frequency modulation wavelet atomic network

Publications (2)

Publication Number Publication Date
CN109784311A CN109784311A (en) 2019-05-21
CN109784311B true CN109784311B (en) 2023-04-18

Family

ID=66504235

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910108831.3A Active CN109784311B (en) 2019-02-03 2019-02-03 Target identification method based on linear frequency modulation wavelet atomic network

Country Status (1)

Country Link
CN (1) CN109784311B (en)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110297238B (en) * 2019-06-24 2023-03-31 山东大学 Joint feature extraction and classification method based on self-adaptive chirp wavelet filtering
CN113238200B (en) * 2021-04-20 2024-09-17 上海志良电子科技有限公司 Classification method of radar linear frequency modulation signals based on validity verification

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104408481A (en) * 2014-12-05 2015-03-11 西安电子科技大学 Deep wavelet neural network-based polarimetric SAR (synthetic aperture radar) image classification method
CN104792522A (en) * 2015-04-10 2015-07-22 北京工业大学 Intelligent gear defect analysis method based on fractional wavelet transform and BP neutral network
CN107194433A (en) * 2017-06-14 2017-09-22 电子科技大学 A kind of Radar range profile's target identification method based on depth autoencoder network
CN109188414A (en) * 2018-09-12 2019-01-11 北京工业大学 A kind of gesture motion detection method based on millimetre-wave radar

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104408481A (en) * 2014-12-05 2015-03-11 西安电子科技大学 Deep wavelet neural network-based polarimetric SAR (synthetic aperture radar) image classification method
CN104792522A (en) * 2015-04-10 2015-07-22 北京工业大学 Intelligent gear defect analysis method based on fractional wavelet transform and BP neutral network
CN107194433A (en) * 2017-06-14 2017-09-22 电子科技大学 A kind of Radar range profile's target identification method based on depth autoencoder network
CN109188414A (en) * 2018-09-12 2019-01-11 北京工业大学 A kind of gesture motion detection method based on millimetre-wave radar

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
Hervé Glotin.FAST CHIRPLET TRANSFORM INJECTS PRIORS IN DEEP LEARNING OF ANIMAL CALLS AND SPEECH.ICLR 2017.2017,全文. *
谢将剑 等.基于Chirplet语图特征和深度学习的鸟类物种识别方法.北京林业大学学报.2018,第第40卷卷(第第40卷期),全文. *

Also Published As

Publication number Publication date
CN109784311A (en) 2019-05-21

Similar Documents

Publication Publication Date Title
CN109389058B (en) Sea clutter and noise signal classification method and system
CN108256436B (en) Radar HRRP target identification method based on joint classification
CN102608589B (en) Radar target identification method on basis of biomimetic pattern identification theory
CN111835444B (en) Wireless channel scene identification method and system
CN109784311B (en) Target identification method based on linear frequency modulation wavelet atomic network
CN109977724A (en) A kind of Underwater Target Classification method
CN112014801A (en) Composite interference identification method based on SPWVD and improved AlexNet
CN114117912A (en) Sea clutter modeling and inhibiting method under data model dual drive
CN110045336B (en) Convolutional neural network-based radar interference identification method and device
Ramezanpour et al. Two-stage beamforming for rejecting interferences using deep neural networks
CN115372925A (en) Array robust adaptive beam forming method based on deep learning
Zheng et al. Time-frequency feature-based underwater target detection with deep neural network in shallow sea
Turhan‐Sayan et al. Electromagnetic target classification using time–frequency analysis and neural networks
Zhu et al. Gabor filter approach to joint feature extraction and target recognition
Li et al. Aerospace target identification-Comparison between the matching score approach and the neural network approach
Ali et al. An improved gain vector to enhance convergence characteristics of recursive least squares algorithm
CN109784318A (en) The recognition methods of Link16 data-link signal neural network based
CN110297238B (en) Joint feature extraction and classification method based on self-adaptive chirp wavelet filtering
CN112346056B (en) Resolution characteristic fusion extraction method and identification method of multi-pulse radar signals
Lee et al. ISAR autofocus by minimizing entropy of eigenimages
Ye et al. Research on machine learning algorithm based on contour matching modal matrix
Liu et al. A direction of arrival estimation method based on deep learning
CN114994657B (en) Radar repetition frequency variation steady target identification method based on repetition frequency self-adaptive network
Li et al. Chirplet-atoms network approach to high-resolution range profiles automatic target recognition
CN116679278B (en) Target radar detection method under strong ground clutter interference

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant