CN113109780B - High-resolution range profile target identification method based on complex number dense connection neural network - Google Patents
High-resolution range profile target identification method based on complex number dense connection neural network Download PDFInfo
- Publication number
- CN113109780B CN113109780B CN202110230325.9A CN202110230325A CN113109780B CN 113109780 B CN113109780 B CN 113109780B CN 202110230325 A CN202110230325 A CN 202110230325A CN 113109780 B CN113109780 B CN 113109780B
- Authority
- CN
- China
- Prior art keywords
- data set
- frequency spectrum
- complex
- time
- range profile
- 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
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
- G01S7/00—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00
- G01S7/02—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00
- G01S7/41—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section
- G01S7/417—Details of systems according to groups G01S13/00, G01S15/00, G01S17/00 of systems according to group G01S13/00 using analysis of echo signal for target characterisation; Target signature; Target cross-section involving the use of neural networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Abstract
The invention discloses a high-resolution range profile target identification method based on a plurality of densely connected neural networks, which comprises the following steps: acquiring a radar range profile data set; carrying out short-time Fourier transform on the radar range profile data set to obtain a complex time-frequency spectrum data set; dividing the complex time-frequency spectrum data set into a complex time-frequency spectrum training data set and a complex time-frequency spectrum verification data set; constructing a plurality of densely connected neural networks; training a plurality of densely-connected neural networks by using a plurality of time spectrum training data sets, and verifying the trained plurality of densely-connected neural networks by using a plurality of time spectrum verification data sets to obtain a trained plurality of densely-connected neural networks; and identifying the radar range profile test data set by using the trained complex dense connection neural network to obtain a target identification result. The invention utilizes the built complex dense connection neural network to train and recognize the complex high-resolution range profile and fully utilizes the characteristic structure in the signal, thereby improving the accuracy of the recognition network.
Description
Technical Field
The invention belongs to the technical field of radars, and particularly relates to a high-resolution range profile target identification method based on a plurality of densely connected neural networks.
Background
The radar target identification is to judge the type of a target by using a radar echo signal of the target. Broadband radars typically operate in the optical region, where the target can be viewed as being made up of a large number of scattering points of varying intensity. The High Resolution Range Profile (HRRP) is a vector sum of echoes of scattering points on a target acquired by using a broadband radar signal. The method reflects the distribution situation of scattering points on a target body along the radar sight line, contains important structural features of the target, and is widely applied to the field of radar target identification.
The method is an important link in a radar target identification system by extracting identification features from the high-resolution range profile. The original high-resolution range image data is complex, wherein the amplitude and the phase contain abundant target detail information. Most of the traditional identification methods directly perform a modulus operation on an original complex high-resolution range profile to obtain real range profile data for identification. In recent years, expanding the traditional real number network to the complex number domain is an emerging research direction. Before the invention of the present application, a complex convolutional neural network and a complex residual error network applied to a high-resolution range profile have been proposed in the "radar high-resolution range profile target method research based on a complex network" (lewanese), and the recognition rate is improved compared with a real network.
However, the convolutional neural network and the residual neural network are limited by the basic network structure, and when the identification features are complex and the number of network layers is increased, the problem of network degradation is easy to occur, so that the target identification performance is limited.
Disclosure of Invention
In order to solve the above problems in the prior art, the present invention provides a high resolution range profile target identification method based on a plurality of densely connected neural networks.
One embodiment of the invention provides a high-resolution range profile target identification method based on a complex number dense connection neural network, which comprises the following steps:
acquiring a radar range profile data set;
carrying out short-time Fourier transform on the radar range profile data set to obtain a complex time-frequency spectrum data set;
dividing the complex time-frequency spectrum data set into a complex time-frequency spectrum training data set and a complex time-frequency spectrum verification data set;
constructing a plurality of densely connected neural networks;
training a plurality of densely-connected neural networks by using the plurality of time-frequency spectrum training data sets, and verifying the trained plurality of densely-connected neural networks by using the plurality of time-frequency spectrum verification data sets to obtain a trained plurality of densely-connected neural networks;
and identifying the radar range profile test data set by using the trained complex dense connection neural network to obtain a target identification result.
In an embodiment of the present invention, the dividing the complex time-frequency spectrum data set into a complex time-frequency spectrum training data set and a complex time-frequency spectrum verification data set includes:
separating a real part and an imaginary part of each complex time spectrum data in the complex time spectrum data set to obtain a real part time spectrum data set and an imaginary part time spectrum data set;
dividing the real part time frequency spectrum data set into a real part time frequency spectrum training data set and a real part time frequency spectrum verification data set;
correspondingly dividing the imaginary part time frequency spectrum data set into an imaginary part time frequency spectrum training data set and an imaginary part time frequency spectrum verification data set;
and the real part time frequency spectrum training data set and the imaginary part time frequency spectrum training data set correspondingly form the complex time frequency spectrum training data set, and the real part time frequency spectrum verification data set and the imaginary part time frequency spectrum verification data set correspondingly form the complex time frequency spectrum verification data set.
In one embodiment of the invention, the constructed plural densely-connected neural networks comprise an input layer, a first dense block, a first vector splicing layer, a second vector splicing layer, a data output preprocessing block and an output layer which are connected in sequence, wherein the first dense block is further connected with the first vector splicing layer through the second dense block, and the first vector splicing layer is further connected with the second vector splicing layer through the third dense block.
In one embodiment of the present invention, the first dense block includes a plurality of batch normalization layers, a plurality of activation function layers, a plurality of convolution layers, a plurality of batch normalization layers, a plurality of activation function layers, and a plurality of convolution layers, which are connected in sequence.
In an embodiment of the present invention, the second dense block includes a plurality of batch normalization layers, a plurality of activation function layers, a plurality of convolution layers, a plurality of batch normalization layers, a plurality of activation function layers, and a plurality of convolution layers, which are connected in sequence.
In an embodiment of the present invention, the third dense block includes a plurality of batch normalization layers, a plurality of activation function layers, a plurality of convolution layers, a plurality of batch normalization layers, a plurality of activation function layers, and a plurality of convolution layers, which are connected in sequence.
In an embodiment of the invention, the data output preprocessing block includes a plurality of batch normalization layers, a plurality of activation function layers, a plurality of convolution layers, a flat one-dimensional layer, a full connection layer, a plurality of batch normalization layers, a full connection layer, and an activation function layer, which are connected in sequence.
In one embodiment of the present invention, training a complex densely-connected neural network using the complex time-frequency spectrum training data set comprises:
and training a plurality of densely connected neural networks by using the real part time-frequency spectrum training data set and the imaginary part time-frequency spectrum training data set.
In one embodiment of the invention, verifying the trained complex densely-connected neural network using the complex time-frequency spectrum verification dataset comprises:
and verifying the trained complex dense connection neural network by using the real part time-frequency spectrum verification data set and the imaginary part time-frequency spectrum verification data set.
In an embodiment of the present invention, the identifying the radar range profile test data set by using the trained plural densely connected neural networks to obtain the target identification result includes:
carrying out short-time Fourier transform on the radar range profile test data set to obtain a complex time-frequency spectrum test data set;
separating a real part and an imaginary part of each complex time spectrum test data in the complex time spectrum test data set to obtain a real part time spectrum test data set and an imaginary part time spectrum test data set;
and inputting the real part time frequency spectrum test data set and the imaginary part time frequency spectrum test data set into the trained complex dense connection neural network for identification to obtain the target identification result.
Compared with the prior art, the invention has the beneficial effects that:
according to the high-resolution range profile target identification method based on the complex dense connection neural network, the constructed complex dense connection neural network can be used for training and identifying the complex high-resolution range profile, and the characteristic structure in the signal is fully utilized, so that the identification precision of the network is improved.
The present invention will be described in further detail with reference to the accompanying drawings and examples.
Drawings
Fig. 1 is a schematic flowchart of a high-resolution range profile target identification method based on a plurality of densely connected neural networks according to an embodiment of the present invention;
fig. 2 is a schematic structural diagram of a plurality of densely connected neural networks in a high-resolution range profile target identification method based on the plurality of densely connected neural networks according to an embodiment of the present invention;
fig. 3 is a schematic structural diagram of another complex dense-connected neural network in a high-resolution range profile target identification method based on the complex dense-connected neural network according to an embodiment of the present invention;
FIG. 4 is a diagram illustrating comparison results of correct recognition rates of several target recognition methods provided by the embodiment of the present invention;
FIG. 5 is a schematic diagram of a recognition rate variation curve in all target recognition processes provided by an embodiment of the present invention;
fig. 6 is a schematic diagram of a loss value variation curve in all target identification processes provided by the embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to specific examples, but the embodiments of the present invention are not limited thereto.
Example one
Referring to fig. 1, fig. 1 is a schematic flow chart of a high-resolution range profile target identification method based on a plurality of densely connected neural networks according to an embodiment of the present invention. The embodiment provides a high-resolution range profile target identification method based on a plurality of densely connected neural networks, which comprises the following steps:
step 1, a radar range profile data set is obtained.
Specifically, the radar high-resolution range profile HRRP is a vector sum of echoes of scattering points on a target body acquired by using a broadband radar signal, reflects the distribution of the scattering points on the target body along the radar sight line, and includes important structural features of the target. The original high-resolution range profile data is complex, so that not only can a great deal of target structure information be embodied, but also the advantages of quick processing and easy acquisition are achieved. In the embodiment, a radar range profile data set is obtained by acquiring a plurality of radar range profile data.
And 2, carrying out short-time Fourier transform on the radar range profile data set to obtain a complex time-frequency spectrum data set.
Specifically, in the embodiment, the short-time fourier transform is performed on the radar high-resolution range profile to obtain complex time-frequency spectrum data of the radar high-resolution range profile, and the short-time fourier transform converts a time-domain signal into a frequency domain to obtain more signal details. Before short-time Fourier transform is carried out on the radar range profile data set, each radar range profile data in the radar range profile data set can be aligned to a zero phase by adopting an absolute alignment method, and the main operation is to move the frequency point of original data to the center of a frequency spectrum to overcome translational sensitivity. After alignment processing, in order to overcome intensity sensitivity, an intensity normalization method is adopted, and amplitude normalization of the aligned radar range profile data in a dimension k is expressed as:
wherein f is k Representing the magnitude of the radar range profile data in dimension k,
representing the normalized amplitude of the radar range profile data in dimension k. In this embodiment, the data obtained by the absolute alignment and normalization is subjected to short-time fourier transform to obtain complex time-frequency spectrum data h, where h is an mxn complex matrix, and each absolute alignment and normalization data is subjected to short-time fourier transform to obtain a complex time-frequency spectrum data set. The short-time Fourier transform is to add a sliding time window to a signal and perform Fourier transform on the signal in the window to obtain time-frequency spectrum data of the signal, so that the time resolution and the frequency resolution of the short-time Fourier transform are constrained by the Heisenberg inaccuracy detection principle, once a window function is selected, the time-frequency resolution is determined, and each complex time-frequency spectrum data obtained by the short-time Fourier transform is expressed as follows:
where τ represents time, ω represents frequency, f (-) represents radar range profile data after absolute alignment and normalization that requires short-time fourier transform, and L (-) represents a hamming window function used in the short-time fourier transform. For example, a window function is set as a hamming window, an overlap interval is 0, a window function length is 32, a fourier point number is 16, and finally, complex time-frequency spectrum data with a size of 32 × 15 is obtained.
And 3, dividing the complex time-frequency spectrum data set into a complex time-frequency spectrum training data set and a complex time-frequency spectrum verification data set.
Specifically, in this embodiment, each complex time-frequency spectrum data in the complex time-frequency spectrum data set obtained in step 2 may be represented as h ═ x + yi, where x is an mxn real matrix, and y is an mxn real matrix, and the dividing the complex time-frequency spectrum data set into a complex time-frequency spectrum training data set and a complex time-frequency spectrum verification data set specifically includes:
separating a real part and an imaginary part of each complex time spectrum data in the complex time spectrum data set to obtain a real part time spectrum data set and an imaginary part time spectrum data set; dividing the real part time-frequency spectrum data set into a real part time-frequency spectrum training data set and a real part time-frequency spectrum verification data set; correspondingly dividing the imaginary part time-frequency spectrum data set into an imaginary part time-frequency spectrum training data set and an imaginary part time-frequency spectrum verification data set; and the real part time frequency spectrum training data set and the imaginary part time frequency spectrum training data set correspondingly form a complex time frequency spectrum training data set, and the real part time frequency spectrum verification data set and the imaginary part time frequency spectrum verification data set correspondingly form a complex time frequency spectrum verification data set. For example, the embodiment processes three types of airplane data, and the three types of airplane data are sampled and divided into a complex time-frequency spectrum training data set and a complex time-frequency spectrum verification data set according to a ratio of 30:1, where the ratio of the real part time-frequency spectrum training data set to the real part time-frequency spectrum verification data set is 30:1, the ratio of the corresponding imaginary part time-frequency spectrum training data set to the imaginary part time-frequency spectrum verification data set is 30:1, the three types of airplane data are divided into batches with fixed sizes, each batch contains 100 signals, and the batches are respectively used for training and verifying the subsequent complex dense connection neural network.
And 4, constructing a plurality of densely connected neural networks.
Specifically, referring to fig. 2, fig. 2 is a schematic structural diagram of a plurality of dense connection neural networks in a high-resolution range profile target identification method based on the plurality of dense connection neural networks according to an embodiment of the present invention, where the plurality of dense connection neural networks constructed in this embodiment includes an input layer, a first dense block, a first vector splicing layer, a second vector splicing layer, a data output preprocessing block, and an output layer, which are sequentially connected, the first dense block is further connected to the first vector splicing layer through the second dense block, and the first vector splicing layer is connected to the first vector splicing layer through the second dense blockThe splice layer is also connected to the second vector splice layer through a third dense block. The embodiment builds a complex dense connection neural network with two channels of a real part and an imaginary part, and the input of the complex dense connection neural network is a real part matrix x and an imaginary part matrix y. The built double-channel complex dense connection neural network directly connects each dense block through one line, so that a direct-connected short path exists from an early layer to a later layer, information flow between network layers can be maximized, and the degradation problem of the complex dense connection neural network is prevented. The method comprises the steps that a receiving signal h of a complex dense connection neural network is a real part matrix x and an imaginary part matrix y of x + yi, an initialized complex convolution kernel w in the network is also divided into a real part matrix a and an imaginary part matrix b, convolution is respectively carried out on the complex convolution kernel w and the real part matrix a and the imaginary part matrix b of the receiving signal of the network, and finally the convolution result is w h (a x-b y) + i (b x + a y). The complex parameter initialized in the complex dense-connected neural network can be expressed as a product w ═ w | e of a modulus value and a phase iθ The variance of w can be determined by
The equation is calculated, Var (r) represents the variance calculation, E (r) represents the expectation calculation, and it can be seen that the variance of w depends only on the modulus and the phase, so in the complex parameter initialization, the embodiment initializes the modulus of the complex parameter w using the rayleigh distribution with a suitable parameter σ, and then initializes the phase using the uniform distribution on (- π, π), and in order to keep the variance of the input and output consistent with the gradient in the network training, the initialization is performed by setting
Wherein N is in 、N out Respectively representing the number of samples input and output in the network.
Further, referring to fig. 3, fig. 3 is a schematic structural diagram of another complex dense connection neural network in a high-resolution range profile target identification method based on a complex dense connection neural network according to an embodiment of the present invention, where a first dense block in this embodiment includes a plurality of batch normalization layers, a plurality of activation function layers, a plurality of convolution layers, a plurality of batch normalization layers, a plurality of activation function layers, and a plurality of convolution layers, which are connected in sequence, specifically, for example, activation functions of the plurality of activation function layers are all CReLU activation functions, the plurality of convolution layers all have 32 convolution kernels, each convolution kernel has a size of 1 × 3, a convolution step size of 2, and the plurality of batch normalization layers are normalized to a BN normalization function. In this embodiment, the first dense block performs the above-mentioned sequential layer connection processing on the real part time-frequency spectrum training data set and the imaginary part time-frequency spectrum training data set input in step 3. For the plural batch normalization layer, in order to normalize each complex time-frequency spectrum training data to a normal distribution compliant with the standard, this embodiment first normalizes and expresses the complex time-frequency spectrum training data h as:
wherein V represents a 2 x 2 covariance matrix,
cov (-) denotes the covariance, E (-) denotes the expectation, R (-) denotes the real part of the complex data, and S (-) denotes the imaginary part of the complex data. After normalization, similar to batch normalization of real data, we also introduce two variables, namely a scale coefficient and an offset, and the final complex batch normalization formula is
Wherein the content of the first and second substances,
representing normalized complex time-frequency spectrum training data; γ represents a scale factor, which is a semi-positive definite matrix with a size of 2 × 2, and β represents a bias parameter, which is a complex parameter with two degrees of learning (imaginary part and real part). In initialization, because of complex number data
Is 1, so let
γ ri ＝0，β＝0。
For the complex activation function layer, a ReLU function is used in the real part and the imaginary part respectively, and when the real part and the imaginary part of the complex data are both greater than 0 or less than 0, the activation function satisfies the cauchy-riemann equation, and the specific activation function is expressed as:
CReLU(z)＝ReLU(R(z))+iReLU(S(z))；
where R (-) represents the real part of the complex data, and S (-) represents the imaginary part of the complex data.
For the complex convolution layer, the real part and the imaginary part are regarded as two real parts with different logic meanings, the real parts and the imaginary parts are respectively input into the two-dimensional convolution layer through different channels, and complex data are abstractly expressed as follows:
h＝x+yi；
in this embodiment, the complex time-frequency spectrum data set is input to the complex dense connection neural network, x represents a real part of the complex data, y represents an imaginary part of the complex data, and i represents an imaginary unit. In order to make the convolution operation achieve the same effect in the complex domain as the real domain, the adopted convolution kernel is also expressed as a complex filter matrix:
w＝a+bi；
where a represents the real matrix, b represents the imaginary matrix, and i represents the imaginary unit. The specific operation of complex convolution is that the real part and the imaginary part of the complex and the real part and the imaginary part of the convolution kernel are respectively subjected to pairwise cross convolution, the real part takes the subtraction result, the imaginary part takes the addition result, and the calculation formula is represented as follows:
w*h＝(a*x-b*y)+(b*x+a*y)i；
rewriting the above formula into a matrix form as:
where R (-) represents the real part of the complex data, and S (-) represents the imaginary part of the complex data.
Further, referring to fig. 3, the second dense block of the present embodiment includes a plurality of batch normalization layers, a plurality of activation function layers, a plurality of convolution layers, a plurality of batch normalization layers, a plurality of activation function layers, and a plurality of convolution layers, which are sequentially connected. In this embodiment, the second dense block performs the above-mentioned processing of the successive connection layer on the plurality of data sets output by the processing of the first dense block, and the implementation of the specific connection layer is similar to the processing of the first dense block, which is not described herein again.
Further, referring to fig. 3, the third dense block of the present embodiment includes a plurality of batch normalization layers, a plurality of activation function layers, a plurality of convolution layers, a plurality of batch normalization layers, a plurality of activation function layers, and a plurality of convolution layers, which are sequentially connected. In this embodiment, after the complex data set output by processing the second dense block and the complex data set output by processing the first dense block are spliced by the first vector splicing layer, the third dense block performs the above processing of the successive connection layers, and the implementation of the specific connection layer is similar to the processing of the first dense block and the second dense block, which is not described herein again.
Further, referring to fig. 3 again, the data output preprocessing block of the present embodiment includes a plurality of batch normalization layers, a plurality of activation function layers, a plurality of convolution layers, a flat one-dimensional layer, a full connection layer, a plurality of batch normalization layers, a full connection layer, and an activation function layer, which are connected in sequence. In this embodiment, after the complex data set output by the third dense block processing and the complex data set output by the second dense block processing are spliced by the second vector splicing layer, the processing of the successive connection layers is carried out, and the realization of a plurality of batch normalization layers, a plurality of activation function layers and a plurality of convolution layers in the connection layers is similar to the processing of a first dense block, a second dense block and a third dense block, and the flat unidimensional layer, the full connection layer is conventional, the full connection layer connected with the flat unidimensional layer reduces the dimension to 300, the dimension is reduced to 3 by a full connection layer connected with a plurality of batches of normalization layers, the activation function of the activation function layer used as the output in the data output preprocessing block is a softmax activation function, and judging which kind of vector with the dimension of 3 is, and outputting the classified vector, wherein the detailed implementation of each layer is not repeated herein.
And 5, training a plurality of dense connection neural networks by using a plurality of time spectrum training data sets, and verifying the trained plurality of dense connection neural networks by using a plurality of time spectrum verification data sets to obtain the trained plurality of dense connection neural networks.
Specifically, in this embodiment, the complex dense-connected neural network is trained by using the complex time-frequency spectrum training data set, and since the complex time-frequency spectrum training data set includes the real-part time-frequency spectrum training data set and the imaginary-part time-frequency spectrum training data set, the complex dense-connected neural network is trained by using the real-part time-frequency spectrum training data set and the imaginary-part time-frequency spectrum training data set. In the embodiment, the real part time frequency spectrum training data set and the imaginary part time frequency spectrum training data set are used for carrying out precision training on the complex dense connection neural network, so that the complex dense connection neural network is gradually fitted, and different structural features in signals can be accurately extracted.
Further, in this embodiment, the complex densely-connected neural network trained is verified by using the complex time-frequency spectrum verification data set, and similarly, the real time-frequency spectrum verification data set and the imaginary time-frequency spectrum verification data set are used to verify the trained complex densely-connected neural network. And (3) training all real part time-frequency spectrum training data sets and imaginary part time-frequency spectrum training data sets each time, correspondingly using the real part time-frequency spectrum verification data sets and the imaginary part time-frequency spectrum verification data sets for verification, and if the verification result is reduced, indicating that the network training is over-fitted, indicating that the network training parameters need to be adjusted. And finally, determining the optimal network parameters of the complex dense connection neural network of the embodiment through multiple times of adjustment, and finally obtaining the trained complex dense connection neural network.
In order to utilize more information in the original complex radar range profile data, but because the complex radar range profile data is directly used, the data calculation amount in the network is huge, and the learning speed of the network is influenced.
And 6, recognizing the radar range profile test data set by using the trained complex dense connection neural network to obtain a target recognition result.
Specifically, the target recognition result is obtained by recognizing the radar range profile test data set by using the trained complex dense connection neural network. For any radar range profile test data, similar processing needs to be performed like a radar range profile training data set and a radar range profile verification data set, specifically: carrying out short-time Fourier transform on the radar range profile test data set to obtain a complex time-frequency spectrum test data set; separating a real part and an imaginary part of each complex time spectrum test data in the complex time spectrum test data set to obtain a real part time spectrum test data set and an imaginary part time spectrum test data set; and inputting the real part time spectrum test data set and the imaginary part time spectrum test data set into a trained complex dense connection neural network for identification to obtain a target identification result. The short-time Fourier transform is carried out on the complex-time spectrum test data set, and the specific process of separating the real part from the imaginary part is carried out on the complex-time spectrum test data in the complex-time spectrum test data set, and the process is similar to the process of the radar range profile training data set and the radar range profile verification data set, and is not repeated here.
In order to verify the effectiveness of the method for identifying a target of a high-resolution range profile based on a plurality of densely connected neural networks, the following experiment on measured data is further described.
1. Content of the experiment
The experiment adopts the high-resolution range profiles of three types of airplane targets to train the recognition system. The parameters of the three types of airplane targets and the radar parameters for recording the high-resolution range images of the three types of airplane targets are shown in the table 1.
TABLE 1 parameters and Radar parameters of three types of aircraft targets
In Table 1, the "Jack-42" aircraft target contains seven segments of high resolution range profile data, the "An-26" aircraft target contains seven segments of high resolution range profile data, and the "medal" aircraft target contains five segments of high resolution range profile data. In the experiment, second and fifth sections of high-resolution range profile data of a Jack-42 airplane target, sixth and seventh sections of high-resolution range profile data of a prize-like airplane target and fifth and sixth sections of high-resolution range profile data of an 'an-26' airplane target are selected as training samples of a training and recognition system, and the rest data are used as test samples for testing the performance of the recognition system. All high resolution range profile data is 256 range bins.
Meanwhile, in the embodiment, under the three types of airplane target test data, experimental test comparison is performed on the complex number dense connection neural network based on the time spectrum and the traditional classical real number convolution neural network, classical real number dense connection neural network and complex number convolution neural network based on the high-resolution range profile.
2. Analysis of Experimental results
Referring to fig. 4, fig. 4 is a schematic diagram illustrating a comparison result of correct recognition rates of several target recognition methods according to an embodiment of the present invention, and it can be seen from fig. 4 that the recognition rate of the complex dense-connected neural network based on the time-frequency spectrum is about 98.21%, which is significantly higher than 94.6% of the classical real-convolution neural network based on the high-resolution range profile and 94.19% of the classical real-dense-connected neural network based on the high-resolution range profile, and is also significantly improved compared with 97.6% of the recognition rate of the complex convolution neural network based on the high-resolution range profile.
Referring to fig. 5 and 6, fig. 5 is a schematic diagram of an identification rate change curve in all target identification processes provided by an embodiment of the present invention, and fig. 6 is a schematic diagram of a loss value change curve in all target identification processes provided by an embodiment of the present invention, and as can be seen from fig. 5 and 6, in this embodiment, a real signal is obtained by directly taking a modulus from an original signal, but a real part and an imaginary part of the original signal are respectively utilized, and a direct-connected line is built in each layer of network structure, so that more target structure information in the original signal is retained, and improvement of identification accuracy of a network is facilitated.
In summary, according to the high-resolution range profile target identification method based on the complex dense connection neural network provided by this embodiment, the complex dense connection neural network constructed by the method can be used for training and identifying the complex high-resolution range profile, the degradation problem can be better dealt with by the complex dense connection neural network due to the direct connection structure, and the identification network in this embodiment is based on the time spectrum, so that more feature details can be identified compared with the identification network based on the high-resolution range profile, and the feature structure in the signal is fully utilized, thereby improving the accuracy of the identification network.
The foregoing is a more detailed description of the invention in connection with specific preferred embodiments and it is not intended that the invention be limited to these specific details. For those skilled in the art to which the invention pertains, several simple deductions or substitutions can be made without departing from the spirit of the invention, and all shall be considered as belonging to the protection scope of the invention.
Claims (9)
1. A high-resolution range profile target identification method based on a plurality of densely connected neural networks is characterized by comprising the following steps:
acquiring a radar range profile data set;
carrying out short-time Fourier transform on the radar range profile data set to obtain a complex time-frequency spectrum data set;
dividing the complex time-frequency spectrum data set into a complex time-frequency spectrum training data set and a complex time-frequency spectrum verification data set;
constructing a plurality of densely connected neural networks;
training a plurality of densely-connected neural networks by using the plurality of time-frequency spectrum training data sets, and verifying the trained plurality of densely-connected neural networks by using the plurality of time-frequency spectrum verification data sets to obtain a trained plurality of densely-connected neural networks;
recognizing a radar range profile test data set by using the trained plural dense connection neural networks to obtain a target recognition result;
the constructed dense connection neural network comprises an input layer, a first dense block, a first vector splicing layer, a second vector splicing layer, a data output preprocessing block and an output layer which are sequentially connected, wherein the first dense block is also connected with the first vector splicing layer through the second dense block, and the first vector splicing layer is also connected with the second vector splicing layer through the third dense block.
2. The method of claim 1, wherein the dividing the complex time-frequency spectrum data set into a complex time-frequency spectrum training data set and a complex time-frequency spectrum verification data set comprises:
separating a real part and an imaginary part of each complex time spectrum data in the complex time spectrum data set to obtain a real part time spectrum data set and an imaginary part time spectrum data set;
dividing the real part time frequency spectrum data set into a real part time frequency spectrum training data set and a real part time frequency spectrum verification data set;
correspondingly dividing the imaginary part time frequency spectrum data set into an imaginary part time frequency spectrum training data set and an imaginary part time frequency spectrum verification data set;
and the real part time frequency spectrum training data set and the imaginary part time frequency spectrum training data set correspondingly form the complex time frequency spectrum training data set, and the real part time frequency spectrum verification data set and the imaginary part time frequency spectrum verification data set correspondingly form the complex time frequency spectrum verification data set.
3. The method for identifying the high-resolution range profile target based on the plurality of the densely-connected neural networks as claimed in claim 1, wherein the first dense block comprises a plurality of batch normalization layers, a plurality of activation function layers, a plurality of convolution layers, a plurality of batch normalization layers, a plurality of activation function layers and a plurality of convolution layers which are connected in sequence.
4. The method for identifying the high-resolution range profile target based on the plurality of the densely-connected neural networks as claimed in claim 3, wherein the second dense block comprises a plurality of batch normalization layers, a plurality of activation function layers, a plurality of convolution layers, a plurality of batch normalization layers, a plurality of activation function layers and a plurality of convolution layers which are connected in sequence.
5. The method for identifying the high-resolution range profile target based on the plurality of the densely-connected neural networks as claimed in claim 3, wherein the third dense block comprises a plurality of batch normalization layers, a plurality of activation function layers, a plurality of convolution layers, a plurality of batch normalization layers, a plurality of activation function layers and a plurality of convolution layers which are connected in sequence.
6. The method for identifying the high-resolution range profile target based on the plurality of the densely-connected neural networks as claimed in claim 3, wherein the data output preprocessing block comprises a plurality of batch normalization layers, a plurality of activation function layers, a plurality of convolution layers, a flat one-dimensional layer, a full connection layer, a plurality of batch normalization layers, a full connection layer and an activation function layer which are connected in sequence.
7. The method of claim 2, wherein training the plurality of densely connected neural networks with the complex time-frequency spectrum training data set comprises:
and training a plurality of densely connected neural networks by using the real part time-frequency spectrum training data set and the imaginary part time-frequency spectrum training data set.
8. The method of claim 2, wherein validating the trained complex dense-connected neural network using the complex temporal spectral validation dataset comprises:
and verifying the trained complex dense connection neural network by using the real part time-frequency spectrum verification data set and the imaginary part time-frequency spectrum verification data set.
9. The method for recognizing the high-resolution range profile target based on the plurality of densely connected neural networks as claimed in claim 1, wherein recognizing the radar range profile test data set by using the trained plurality of densely connected neural networks to obtain a target recognition result comprises:
carrying out short-time Fourier transform on the radar range profile test data set to obtain a complex time-frequency spectrum test data set;
separating a real part and an imaginary part of each complex time spectrum test data in the complex time spectrum test data set to obtain a real part time spectrum test data set and an imaginary part time spectrum test data set;
and inputting the real part time frequency spectrum test data set and the imaginary part time frequency spectrum test data set into the trained complex dense connection neural network for identification to obtain the target identification result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110230325.9A CN113109780B (en) | 2021-03-02 | 2021-03-02 | High-resolution range profile target identification method based on complex number dense connection neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110230325.9A CN113109780B (en) | 2021-03-02 | 2021-03-02 | High-resolution range profile target identification method based on complex number dense connection neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113109780A CN113109780A (en) | 2021-07-13 |
| CN113109780B true CN113109780B (en) | 2022-08-05 |
Family
ID=76709675
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110230325.9A Active CN113109780B (en) | 2021-03-02 | 2021-03-02 | High-resolution range profile target identification method based on complex number dense connection neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113109780B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114626418A (en) * | 2022-03-18 | 2022-06-14 | 中国人民解放军32802部队 | Radiation source identification method and device based on multi-center complex residual error network |
| CN116559794B (en) * | 2023-07-12 | 2023-09-29 | 西安电子科技大学 | Radar anti-interference intelligent decision method for double-multi-domain complex neural network |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106934419A (en) * | 2017-03-09 | 2017-07-07 | 西安电子科技大学 | Classification of Polarimetric SAR Image method based on plural profile ripple convolutional neural networks |
| CN107728142A (en) * | 2017-09-18 | 2018-02-23 | 西安电子科技大学 | Radar High Range Resolution target identification method based on two-dimensional convolution network |
| CN109086700A (en) * | 2018-07-20 | 2018-12-25 | 杭州电子科技大学 | Radar range profile's target identification method based on depth convolutional neural networks |
| CN109376574A (en) * | 2018-08-14 | 2019-02-22 | 西安电子科技大学 | Refuse to sentence radar HRRP target identification method based on CNN |
| CN110245608A (en) * | 2019-06-14 | 2019-09-17 | 西北工业大学 | A kind of Underwater targets recognition based on semi-tensor product neural network |
| CN110334741A (en) * | 2019-06-06 | 2019-10-15 | 西安电子科技大学 | Radar range profile's recognition methods based on Recognition with Recurrent Neural Network |
| CN111126570A (en) * | 2019-12-24 | 2020-05-08 | 江西理工大学 | SAR target classification method for pre-training complex number full convolution neural network |
| CN112329538A (en) * | 2020-10-10 | 2021-02-05 | 杭州电子科技大学 | Target classification method based on microwave vision |
Family Cites Families (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11585896B2 (en) * | 2019-02-04 | 2023-02-21 | Metawave Corporation | Motion-based object detection in a vehicle radar using convolutional neural network systems |
| US11852750B2 (en) * | 2019-06-28 | 2023-12-26 | Smart Radar System, Inc. | Method and apparatus for radar signal processing using recurrent neural network |
| CN110969121A (en) * | 2019-11-29 | 2020-04-07 | 长沙理工大学 | High-resolution radar target recognition algorithm based on deep learning |
| CN111768340B (en) * | 2020-06-30 | 2023-12-01 | 苏州大学 | Super-resolution image reconstruction method and system based on dense multipath network |
-
2021
- 2021-03-02 CN CN202110230325.9A patent/CN113109780B/en active Active
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106934419A (en) * | 2017-03-09 | 2017-07-07 | 西安电子科技大学 | Classification of Polarimetric SAR Image method based on plural profile ripple convolutional neural networks |
| CN107728142A (en) * | 2017-09-18 | 2018-02-23 | 西安电子科技大学 | Radar High Range Resolution target identification method based on two-dimensional convolution network |
| CN109086700A (en) * | 2018-07-20 | 2018-12-25 | 杭州电子科技大学 | Radar range profile's target identification method based on depth convolutional neural networks |
| CN109376574A (en) * | 2018-08-14 | 2019-02-22 | 西安电子科技大学 | Refuse to sentence radar HRRP target identification method based on CNN |
| CN110334741A (en) * | 2019-06-06 | 2019-10-15 | 西安电子科技大学 | Radar range profile's recognition methods based on Recognition with Recurrent Neural Network |
| CN110245608A (en) * | 2019-06-14 | 2019-09-17 | 西北工业大学 | A kind of Underwater targets recognition based on semi-tensor product neural network |
| CN111126570A (en) * | 2019-12-24 | 2020-05-08 | 江西理工大学 | SAR target classification method for pre-training complex number full convolution neural network |
| CN112329538A (en) * | 2020-10-10 | 2021-02-05 | 杭州电子科技大学 | Target classification method based on microwave vision |
Non-Patent Citations (5)
| Title |
|---|
| A novel feature vector using complex HRRP for radar target recognition;Du, Lan等;《ADVANCES IN NEURAL NETWORKS - ISNN 2007, PT 1, PROCEEDINGS》;20071031;全文 * |
| Radar HRRP recognition based on CNN;Jia Song等;《The Journal of Engineering》;20191231;全文 * |
| 基于复数域深度学习的SAR舰船目标识别方法研究;化青龙;《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》;20210115;全文 * |
| 基于复数网路的雷达高分辨距离向目标识别方法研究;李婉萍;《中国优秀硕士学位论文全文数据库信息科技辑》;20200215;全文 * |
| 基于统计建模的雷达高分辨距离向目标识别方法研究;王鹏辉;《中国优秀博硕士学位论文全文数据库(博士)信息科技辑》;20130415;全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113109780A (en) | 2021-07-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108133232B (en) | Radar high-resolution range profile target identification method based on statistical dictionary learning | |
| Xu et al. | Target-aware recurrent attentional network for radar HRRP target recognition | |
| CN109376574B (en) | CNN-based (probabilistic neural network-based) HRRP (high-resolution Radar) target identification method for radar capable of refusing judgment | |
| CN107728142B (en) | Radar high-resolution range profile target identification method based on two-dimensional convolutional network | |
| CN109903259B (en) | Magnetic resonance spectrum reconstruction method based on deep learning | |
| CN113109780B (en) | High-resolution range profile target identification method based on complex number dense connection neural network | |
| Zhang et al. | Noise reduction in the spectral domain of hyperspectral images using denoising autoencoder methods | |
| Liu et al. | Deep learning and recognition of radar jamming based on CNN | |
| CN109343046B (en) | Radar gait recognition method based on multi-frequency multi-domain deep learning | |
| CN108447057B (en) | SAR image change detection method based on significance and depth convolution network | |
| CN102645649A (en) | Radar target recognition method based on radar target range profile time-frequency feature extraction | |
| CN112990334A (en) | Small sample SAR image target identification method based on improved prototype network | |
| CN110736707B (en) | Spectrum detection optimization method for transferring spectrum model from master instrument to slave instrument | |
| CN112394354B (en) | Method for identifying HRRP fusion target small samples based on meta-learning in different polarization modes | |
| CN114237046B (en) | Partial discharge pattern recognition method based on SIFT data feature extraction algorithm and BP neural network model | |
| CN106951822B (en) | One-dimensional range profile fusion identification method based on multi-scale sparse preserving projection | |
| CN113486917B (en) | Radar HRRP small sample target recognition method based on metric learning | |
| CN112991483A (en) | Non-local low-rank constraint self-calibration parallel magnetic resonance imaging reconstruction method | |
| CN108507607B (en) | Weak signal detection method based on kernel function | |
| CN110458071A (en) | A kind of fiber-optic vibration signal characteristic abstraction and classification method based on DWT-DFPA-GBDT | |
| Fursov et al. | Thematic classification with support subspaces in hyperspectral images | |
| CN113240081B (en) | High-resolution range profile target robust identification method aiming at radar carrier frequency transformation | |
| CN116205863A (en) | Method for detecting hyperspectral image abnormal target | |
| CN115456029A (en) | Conveyor belt roller fault identification method and system based on 1DCNN-TSA-GRU | |
| CN114818845A (en) | Noise-stable high-resolution range profile feature selection method |
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 |