CN117192548A - Sparse ISAR high-resolution imaging method based on depth expansion - Google Patents
Sparse ISAR high-resolution imaging method based on depth expansion Download PDFInfo
- Publication number
- CN117192548A CN117192548A CN202311093320.1A CN202311093320A CN117192548A CN 117192548 A CN117192548 A CN 117192548A CN 202311093320 A CN202311093320 A CN 202311093320A CN 117192548 A CN117192548 A CN 117192548A
- Authority
- CN
- China
- Prior art keywords
- network
- layer
- sparse
- isar
- controllable
- 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.)
- Pending
Links
- 238000003384 imaging method Methods 0.000 title claims abstract description 99
- 230000007547 defect Effects 0.000 claims abstract description 46
- 238000012549 training Methods 0.000 claims abstract description 44
- 238000005457 optimization Methods 0.000 claims abstract description 9
- 238000013507 mapping Methods 0.000 claims description 57
- 238000000034 method Methods 0.000 claims description 29
- 230000006870 function Effects 0.000 claims description 19
- 238000012545 processing Methods 0.000 claims description 7
- 239000011159 matrix material Substances 0.000 claims description 6
- 230000004913 activation Effects 0.000 claims description 4
- 230000010354 integration Effects 0.000 claims description 4
- 238000005070 sampling Methods 0.000 claims description 4
- 230000009191 jumping Effects 0.000 claims description 2
- 238000006467 substitution reaction Methods 0.000 claims description 2
- 230000009466 transformation Effects 0.000 claims description 2
- 238000005259 measurement Methods 0.000 claims 1
- 238000004088 simulation Methods 0.000 description 7
- 238000011156 evaluation Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 230000008602 contraction Effects 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 238000012935 Averaging Methods 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 238000000691 measurement method Methods 0.000 description 1
- 230000008569 process Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000002123 temporal effect Effects 0.000 description 1
Landscapes
- Radar Systems Or Details Thereof (AREA)
Abstract
The invention discloses a sparse ISAR high-resolution imaging network based on depth expansion, which mainly solves the problems of instability of echo signal-to-noise ratio and defect rate, high space complexity and low efficiency in the prior art. The implementation scheme comprises the following steps: establishing a sparse observation model under the condition of echo defect; constructing an objective function conforming to an L1 norm optimization criterion according to a sparse observation model Y under the echo defect condition; constructing a sparse ISAR high-resolution imaging network based on a 2D-ISTA iterative algorithm; generating a training set by utilizing the scattering points and training the sparse ISAR high-resolution imaging network by using the training set; inputting the measured data into a trained imaging network, and solving an objective function through forward propagation of the network to obtain a final ISAR imaging result. The invention obviously improves the robustness of the network to the echo signal-to-noise ratio and the defect rate, reduces the time and space complexity, and can be used for extracting important information of carrier-based and airborne ISAR systems.
Description
Technical Field
The invention belongs to the technical field of radar remote sensing, and further relates to an ISAR high-resolution imaging method which can be used for an ISAR system on a carrier or an onboard carrier to extract important information of the shape, structure and gesture of a target.
Background
Inverse synthetic aperture radar ISAR can realize long-distance high-resolution imaging of non-cooperative targets, and therefore plays an important role in space situation awareness. Under complex observation conditions such as low signal-to-noise ratio, echo defect and the like, although focusing imaging can be realized through a sparse signal reconstruction method, the correlation method comprises multiple iterations, and the steps such as matrix inversion and the like are high in complexity. In addition, it is often necessary to manually set a plurality of super parameters depending on the target and echo characteristics, and in practice, it is difficult to find an optimal solution and time-consuming.
The model-driven deep expansion network high-resolution ISAR imaging is performed by intercepting iterative steps of a sparse signal reconstruction method and expanding the iterative steps into a neural network, and parameters which need to be manually optimized in an algorithm are converted into network learnable parameters, so that optimal imaging performance is obtained through network training. Because the values of the parameters in different network layers are different, the flexibility is higher, and the convergence speed is faster. Various high-resolution ISAR imaging methods of a depth expansion network are proposed at present, but the problems of instability of echo defect rate and signal to noise ratio and the like still exist, so that network training is required to be carried out respectively for different echo defect rates and signal to noise ratios, and time and space complexity are greatly increased.
Patent document with application number of CN202010501764.4 discloses a 'low-rank and sparse combined constraint-based SA-ISAR imaging method with a micro-motion component target', which is completed in three parts, and a target one-dimensional range profile sequence with the micro-motion component target translational compensation is modeled; modeling a target sparse aperture ISAR imaging problem with a micro-motion component; and finally, solving the target sparse aperture ISAR imaging problem with the micro-motion component by adopting a linear ADMM. Although the method can obtain the ISAR image with good focusing under a complex observation environment, the method comprises multiple iterations, and the steps such as matrix inversion and the like have higher complexity. In addition, it is often necessary to manually set a plurality of super parameters depending on the target and echo characteristics, and in practice, it is difficult to find an optimal solution and time-consuming.
Patent document with application number of CN202111464674.3 discloses a structured sparse aperture ISAR imaging method based on C-ADMN, which is divided into three parts, firstly, modeling is carried out on sparse aperture ISAR echo signals, then a C-ADMN forward propagation model is constructed, the model comprises a reconstruction layer, a noise reduction layer and a multiplier updating layer, and finally, the C-ADMN is utilized to solve the structured sparse aperture ISAR imaging problem and obtain a final imaging result. Although the method improves ISAR imaging performance and shortens imaging time, the method has the following problems: 1) The robustness is lacking for different signal to noise ratios, and the imaging network is required to be trained according to the different signal to noise ratios, so that the time and space complexity is higher; 2) Lack of robustness for different echo defect rates, require training the imaging network separately according to the different echo defect rates, further increasing the temporal and spatial complexity.
Disclosure of Invention
The invention aims to overcome the defects in the prior art, and provides an ISAR high-resolution imaging method based on depth expansion, so as to improve the robustness of a network to echo signal-to-noise ratio and defect rate and reduce the time and space complexity.
The invention adopts the technical ideas that a traditional iterative algorithm is unfolded to form an imaging network, a controllable near-end mapping module is designed to construct a sparse ISAR high-resolution imaging network based on depth unfolding, and the implementation steps comprise the following steps:
(1) Establishing a sparse observation model under the condition of echo defect:
Y＝Φ 1 XΦ 2 +E
wherein,for echo signals +.>For ISAR image, ++>Is a complex domain distance dictionary which is used for the data processing,is complex domain Doppler dictionary->M and N are respectively the number of sampling points and the total number of pulses, U and V are respectively the length and the width of the image, and when sparse observation conditions exist in the frequency band or azimuth, M is less than U, and N is less than V;
(2) Constructing an objective function conforming to an L1 norm optimization criterion according to a sparse observation model Y under the echo defect condition:
where y=vect (Y), x=vect (X), vect (·) represents the vectorization operation,representing the observation matrix of the image of the object, T representing a transpose operation->Represents Kronecker product;
(3) Constructing a sparse ISAR high-resolution imaging network based on depth expansion:
3a) Expanding the iterative algorithm into an imaging network comprising a gradient descent layer and a proximal mapping layer;
3b) Constructing a controllable near-end mapping module comprising a near-end mapping unit and a controllable unit;
3c) Embedding a controllable near-end mapping module into a near-end mapping layer in an imaging network and connecting withHigh frequency feature extractorResidual reconstructor->Sequentially connecting, and jumping and connecting the output of the gradient descent layer with the output of the residual error reconstructor to form a controllable near-end mapping layer;
3d) Sequentially connecting the gradient descent layer and the controllable near-end mapping layer to obtain an nth layer of sub-network, and cascading a plurality of sub-networks to obtain a sparse ISAR high-resolution imaging network;
(4) Generating a training set:
4a) Scattering points with random positions and amplitude conforming to Gaussian distribution are used as training labels, different echo defect rates and signal to noise ratios are set to obtain different ISAR scenes, and then original images in different scenes are generated;
4b) Forming image pairs by the original image and the label image, and generating n image pairs as training sets, wherein n is more than or equal to 600;
(5) Training a sparse ISAR high-resolution imaging network:
5a) Using normalized mean square error NMSE as a loss function of network training, selecting batch-sized data from a training set each time, inputting the data into a sparse ISAR high-resolution imaging network to calculate a loss value l in each batch, updating network parameters through a random gradient descent algorithm until all data in the training set are selected, and summing and averaging the loss values obtained by all the batches to obtain loss after one-time network training;
5b) Repeating the step 5 a) until the network converges to obtain a trained sparse ISAR high-resolution imaging network;
(6) The actually measured Yak-42 data is input into a trained imaging network, and a final ISAR imaging result is obtained through forward propagation of the network.
Compared with the prior art, the invention has the following advantages:
firstly, the invention builds the near-end mapping unit based on the residual error network, extracts noise characteristics by using the convolutional neural network, and further suppresses noise components in the image by jump connection, thereby improving the robustness of the sparse ISAR high-resolution imaging network to echo signal-to-noise ratio and reducing the time and space complexity.
Secondly, the controllable unit is constructed and added into the near-end mapping module to obtain the controllable near-end mapping module, the controllable vector is generated by using the condition vector containing echo defect information as input, and the output of the last convolution layer in the residual error network is modulated, so that the network can dynamically adjust network parameters according to different echo defect rates in the imaging process, the robustness of the sparse ISAR high-resolution imaging network to the echo defect rate is improved, and the time and space complexity is further reduced.
Drawings
FIG. 1 is a flow chart of an implementation of the present invention;
FIG. 2 is a block diagram of a depth expansion based sparse ISAR high-resolution imaging network constructed in the present invention;
FIG. 3 is a graph comparing ISAR imaging effects of the present invention with those of the prior art.
Detailed Description
The following embodiments and effects of the present invention will be described in further detail with reference to the accompanying drawings, which are only for illustrating the implementation of the present invention and not to constitute any limitation of the present invention, and it will be apparent to those skilled in the art that various modifications and changes in form and detail are possible without departing from the principle and structure of the present invention, but these modifications and changes based on the idea of the present invention remain within the scope of the present invention after understanding the contents and principle of the present invention.
Referring to fig. 1, the implementation steps of the present embodiment include the following:
step one, a sparse observation model under the condition of echo defect is established.
1.1 Assuming that the radar emits a chirp signal, the target is represented by Q scattering points, and an echo signal of the p-th target scattering point in a range fast time-azimuth slow time domain is obtained
Wherein f c Is the carrier frequency, T p Pulse width, gamma frequency modulation, t n =nt represents slow time, T is pulse repetition interval, n=1, 2,..n represents pulse number;representing a fast time; rect (·) represents a unit rectangular functionR p (t n ) Represents the distance between the p-th scattering point and the radar, A p A represents the back reflection coefficient of the p-th scattering point, a a (t n ) Is a function of azimuth window and->T a For the observation time, exp (·) represents an exponential operation with the base of the natural constant e, j represents an imaginary unit symbol;
1.2 For echo signals)Performing line-separating tone processing to obtain an echo s of the p-th scattering point in a distance frequency-slow time domain p (f,t n )：
Wherein,distance frequency, B is bandwidth; r is R ps For the instantaneous skew between the scattering point p and the reference point s, according to the turntable model, the instantaneous skew is expressed as: r is R ps (θ)≈x p sinθ+y p cosθ≈x p θ+y p ，(x p ,y p ) Representing the coordinates of the p-th scattering point of the target, θ=ωt n Omega represents the ISAR target rotational angular velocity;
since the ISAR imaging integration time is short, the rotation angle ω of the target relative to the radar is small in the imaging integration time, so sin (ωt) n )≈ωt n ＝θ、cos(ωt n ) 1, R is ps Substitution of (θ) into s p (f,t n ) Can be obtained by:
1.3 Given the frequency interval Δf and the pulse repetition interval T, discretizing the distance and azimuth directions into f=mΔf and T, respectively n =nt, s p (f,t n ) Can be further converted into a discrete form s p (m,n)：
Where m=1, 2, M, n=1, 2, N, M and N represent the number of sampling points and the total number of pulses, respectively;
1.4 Let ISAR image size be U×V, and element in image be i p (u,v)，u∈[1,U]，v∈[1,V]Echo signal s p (m, n) and image i p The transformation relationship between (u, v) is expressed as follows:
wherein d 1 The distance vector represents a vector with length M obtained by randomly downsampling a vector with original length U under the condition of sparse observation, and d 1m Is d 1 The m-th element of (a); d, d 2 Is a direction defect vector, which means that under the condition of sparse observation, the vector with the original length V is subjected to random downsamplingThe resulting vector of length N, d 2n Is d 2 N-th element of (a);
1.5 Instruction) commandIs a complex domain distance dictionary, phi 1 The mth row and the mth column of the complex field elements are Is complex domain Doppler dictionary, phi 2 The plural field element of the nth row and the nth column of (a) is +.>Obtaining a sparse observation model Y under the condition of echo defect:
Y＝Φ 1 XΦ 2 +E
wherein,for ISAR image, ++>For echo signals +.>For a complex domain noise matrix, when sparse observation conditions exist in a frequency band or azimuth, M is smaller than U, and N is smaller than V;
in the present embodiment, the size of the ISAR image is set to, but not limited to, u=128, v=128.
And secondly, constructing an objective function which accords with an L1 norm optimization criterion according to the sparse observation model Y under the echo defect condition.
2.1 Vectorization operation Vect (·) is performed on the sparse observation model under the echo defect condition, so that y=vect (Y), x=vect (X), e=vect (E), and a vectorized sparse observation model Y is obtained:
y＝Φ 0 x+e
wherein,is a complex domain distance dictionary->Is a complex-domain doppler dictionary which,
t represents the transpose operation and,represents Kronecker product;
2.2 Constructing an objective function meeting the L1 norm optimization criterion according to the vectorized sparse observation model
Wherein,representing the Euclidean norm of the vector, < >>Representing an ISAR scene to be solved, lambda representing a regularization coefficient, and x sparsity being constrained by an L1 norm.
And thirdly, constructing an ISAR imaging network based on depth expansion.
Referring to fig. 2, the implementation of this step includes the following:
3.1 Developing a 2D-ISTA iterative algorithm into an imaging network comprising a gradient descent layer and a proximal mapping layer:
the 2D-ISTA algorithm is an algorithm for solving the convex optimization problem, and the flow is as follows:
input echo signal Y, complex domain distance dictionary Φ 1 Complex domain Doppler dictionary phi 2 Initializing parameter iteration step lengthρ, regularization coefficient λ, initial imaging result X (0) And an auxiliary variable R for the first iteration (1) ；
According to iterative stepsAnd X (k) ＝S soft (R (k) Alpha) pair imaging result X (k) Performing iterative update until the iteration stop condition is satisfied>Stopping iteration to obtain the final imaging result X (N) ；
Where ε is a small value, usually ε=1×10 -6 K represents the iteration number of the 2D-ISTA, ρ represents the iteration step size, S soft (R (k) ,α (k) ) Representing a soft threshold shrink function, defined as S soft (R (k) ,α (k) )＝sgn(R (k) )max{|R (k) |-α (k) 0, α represents the contraction threshold of the iteration.
The present example is developed into an imaging network by using the 2D-ISTA algorithm, and is used for solving the objective function which is constructed in the second step and meets the L1 norm optimization criterion, and the implementation steps include the following steps:
3.1.1 Intercepting N iterations in the 2D-ISTA iterative algorithm, and adding the auxiliary variable R contained in the nth iteration (n) Reconstruction result X (n) Two different network layers, namely a gradient descent layer and a near-end mapping layer, are correspondingly adopted;
3.1.2 Sequentially connecting the gradient descent layer and the near-end mapping layer to obtain an nth layer of sub-network, and repeatedly cascading a plurality of sub-networks to obtain an ISTA unfolding imaging network;
3.2 Constructing a controllable proximal mapping module comprising a proximal mapping unit and a controllable unit;
3.2.1 Design of a residual network based near-end mapping unit:
the unit comprises two convolution modules c 1 、c 2 And a jump connection, wherein: a first convolution module c 1 By having a convolution kernel size of 3 x 3,the convolution layer with the number of convolution kernel channels being 32 is formed by cascading with a ReLU activation layer; a second convolution module c 2 The method consists of a convolution layer with a convolution kernel size of 3 multiplied by 3 and a convolution kernel channel number of 32; the output of each convolution module is represented asx is the input of the near-end mapping unit;
the output of the entire near-end mapping unit is denoted as z=relu (y 2 +x), wherein ReLU (·) represents a ReLU activation function;
3.2.2 Design a controllable unit:
the controllable unit is formed by a full connection layer W CU The implementation is that the input is a conditional vector z containing echo defect information, and a controllable vector with dimension of C is generated
3.2.3 Embedding the controllable unit into the near-end mapping unit to obtain a controllable near-end mapping moduleIts output F (n) The method comprises the following steps:
wherein G is (n) Is the output in the front layer module, z isInput of->And->Respectively represent controllable near-end mapping modules +.>The first and last convolutional layers of (a), in this embodiment, vector dimension c=32;
3.3 Embedding a controllable near-end mapping module into a near-end mapping layer in an imaging network and with a high-frequency feature extractorResidual reconstructor->Sequentially connecting, and jump connecting the output of the gradient descent layer with the output of the residual error reconstructor to form a controllable near-end mapping layer, wherein the high-frequency characteristic extractor is->And residual reconstructor->Structural parameters, and output X of the entire controllable near-end mapping layer (n) The method comprises the following steps:
the high frequency feature extractorThe method consists of a convolution layer with a convolution kernel size of 3 multiplied by 3 and a convolution kernel channel number of 32;
said residual reconstructorThe method consists of convolution layers with the convolution kernel size of 3 multiplied by 32 and the convolution kernel channel number of 1;
the output X (n) Is that
Wherein,and->Respectively soft threshold shrink function S soft (-) front-to-back controllable proximal mapping module; />High-frequency feature extractor for the nth layer of the network,>residual reconstructor for network n-th layer, alpha (n) A learnable shrink threshold for the nth layer of the network;
3.4 The gradient descent layer and the controllable near-end mapping layer are sequentially connected to obtain an nth layer of sub-network, and then a plurality of sub-networks are cascaded to obtain a sparse ISAR high-resolution imaging network, and the structure is as follows:
network input- & gt layer 1 gradient descent layer, controllable near-end mapping layer- & gt … - & gt layer N gradient descent layer, controllable near-end mapping layer- & gt … - & gt layer N gradient descent layer, controllable near-end mapping layer- & gt network output.
The learnable parameters in the network are denoted by Θ:
wherein ρ is (n) For gradient descent layer R (n) In (a) the iteration step of (c) is,is a high-frequency characteristic extractor parameter,For residual reconstructor parameters, α (n) For the contraction threshold value->For two convolution layer parameters in the controllable near-end mapping module, W CU N is the total stage number of the network, which is the parameter of the full connection layer, and the parameters can be obtained through the neural network learning;
in the present embodiment, the iteration step ρ (n) And regularization parameter alpha (n) Initialized to 1.0 and 0.01, respectively, parametersThen random initialization is used and the number of network layers is set, but not limited to, n=9.
And step four, generating a training set.
4.1 Scattering points with random distribution of positions and amplitude conforming to Gaussian distribution are used as training labels X;
4.2 Substituting the label image X into the sparse observation model y=Φ 1 XΦ 2 +E obtaining echo data with different degree of defect and random signal-to-noise ratio noise, and initializing to calculateObtain the original image +.>And takes the data as network input data;
4.3 The original image and the label image are formed into an image pair, and the number of construction samples is N b Image pair of (a)As a training set;
in the embodiment, the size of the training set image is set to be 128 multiplied by 128, the defect rate range of the echo is set to be 25 percent to 75 percent, the signal to noise ratio range of the echo is set to be 0 to 20dB, and the number of training set samples is N b Let us say but not be limited to 800.
And fifthly, training a sparse ISAR high-resolution imaging network.
5.1 Each time data of the batch size is selected from the training set, the loss value in each batch is calculatedUpdating network parameters through a random gradient descent algorithm until all data in the training set are selected;
5.1.1 Data of the batch size is selected from the training set each time and is input into a sparse ISAR high-resolution imaging network to calculate a loss value in each batch
Wherein,is F norm, N batch For the number of samples per batch, +.>Representing a reconstructed image output by an imaging network, and X represents an input label image;
5.1.2 Calculating gradient of loss function with respect to arbitrary parameters in sparse ISAR high-resolution imaging network by using complex domain back propagation algorithmAccording to the solved gradient->Updating network parameters of the sparse ISAR high-resolution imaging network to obtain network parameters theta' of the current training stage:
wherein Θ is 0 For the network parameters of the sparse ISAR high-resolution imaging network training phase, lr represents the learning during trainingA learning rate;
5.2 Repeating the step 5.1) until the network converges to obtain the final trained network parameters, and completing the training of the sparse ISAR high-resolution imaging network.
In this embodiment, the single Batch processing size Batch is set to 1 during training, and Adam optimizer with learning rate of 0.005 is used for training.
And step six, obtaining an ISAR target imaging result.
And inputting the actually measured Yak-42 aircraft data into a trained imaging network, and obtaining a final ISAR imaging result through forward propagation of the network.
The sequence numbers of the steps are for more clearly describing the implementation scheme of the invention, and the sequence numbers are not limited.
The effect of the invention can be further illustrated by the following simulation experiments:
1. simulation experiment condition
The software platform of the simulation experiment is Windows10 operating system and Pytorch 3.7, and the hardware is configured as follows: core i5-9300H CPU and NVIDIA GeForce 1650GPU.
The simulation experiment of the invention uses point simulation data with random distribution of positions and amplitude compliant with Gaussian distribution, the sizes of images in a training set are 128 multiplied by 128, and the number of images is 800.
2. Simulation content and result analysis
Under the simulation conditions, the invention and the existing structured sparse aperture ISAR imaging method based on C-ADMN are used for imaging the actually measured ISAR data. In the prior art, the imaging network images measured data in corresponding scenes after training in each fixed ISAR scene respectively, namely one network corresponds to one test scene, so as to obtain an ISAR image with good target focusing; in the invention, the imaging network can obtain ISAR images with good target focusing for the measured data in different ISAR scenes after training is completed, namely one network corresponds to a plurality of test scenes, and the imaging result is shown in figure 3. Wherein:
FIG. 3 (a) is an image of measured Yak-42 aircraft data obtained in the prior art at an echo defect rate of 25% and an echo signal to noise ratio of 5dB,
FIG. 3 (b) is a graph showing the imaging result obtained from the measured Yak-42 aircraft data under the condition that the echo defect rate is 25% and the echo signal-to-noise ratio is 10dB,
FIG. 3 (c) is an image of measured Yak-42 aircraft data obtained with an echo defect rate of 25% and an echo signal to noise ratio of 5dB,
FIG. 3 (d) is an image of measured Yak-42 aircraft data obtained with an echo defect rate of 25% and an echo signal to noise ratio of 10dB,
FIG. 3 (e) is a graph showing the imaging result obtained from the measured Yak-42 aircraft data under the condition that the echo defect rate is 50% and the echo signal-to-noise ratio is 5dB,
FIG. 3 (f) is an image of measured Yak-42 aircraft data obtained in the prior art at an echo defect rate of 50% and an echo signal to noise ratio of 10dB,
FIG. 3 (g) is an image of measured Yak-42 aircraft data obtained with an echo defect rate of 50% and an echo signal to noise ratio of 5dB,
FIG. 3 (h) is an image of measured Yak-42 aircraft data obtained with an echo defect rate of 50% and an echo signal to noise ratio of 10 dB.
As can be seen from fig. 3, under the condition of different echo defect rates and echo signal-to-noise ratios, the imaging result obtained by the method of the invention retains more information in the aircraft image, especially the aircraft head information, and has better imaging quality compared with the imaging result obtained by the prior art.
Two imaging evaluation indexes, namely a normalized mean square error and a peak signal to noise ratio, are respectively calculated for the imaging results, and the results are shown in table 1,
table 1 comparison of measured data imaging results evaluation index
The evaluation index normalized mean square error and peak signal to noise ratio in table 1 are specifically defined as follows:
the normalized mean square error is a measurement method based on energy normalization, and the smaller the value is, the better the image quality is, and the method is defined as:
where g (i, j) is each pixel of the original image,for processing each pixel of the image, M and N are the length and width of the image;
the peak signal-to-noise ratio is an objective standard for measuring image distortion or noise level, the larger the peak signal-to-noise ratio between two images is, the more similar the peak signal-to-noise ratio is, the more obvious the image degradation below 30dB is, and the definition is that:
wherein,representing the mean square error between the mxn monochromatic image and the processed image, MAX representing the maximum value of the image color.
Compared with the prior art, the method has the advantages that compared with the prior art, under the condition of training completion, the imaging evaluation index normalization mean square error of the imaging network is smaller under the condition of different echo defect rates and echo signal to noise ratios, and the peak signal to noise ratio is higher, so that the imaging quality of the method is better.
In summary, the imaging results and imaging evaluation indexes of the invention under the conditions of different echo defect rates and echo signal-to-noise ratios are superior to those of the existing scheme, and the advantages of improving the robustness of the network to the echo signal-to-noise ratio and the defect rate and reducing the time and space complexity are verified.
Claims (8)
1. The sparse ISAR high-resolution imaging method based on depth expansion is characterized by comprising the following steps of:
(1) Establishing a sparse observation model under the condition of echo defect:
Y＝Φ 1 XΦ 2 +E
wherein,for echo signals +.>For ISAR image, ++>Is a complex domain distance dictionary which is used for the data processing,is complex domain Doppler dictionary->M and N are respectively the number of sampling points and the total number of pulses, U and V are respectively the length and the width of the image, and when sparse observation conditions exist in the frequency band or azimuth, M is less than U, and N is less than V;
(2) Constructing an objective function conforming to an L1 norm optimization criterion according to a sparse observation model Y under the echo defect condition:
where y=vect (Y), x=vect (X), vect (·) represents the vectorization operation,representing the observation matrix, T representing the transpose operation, +.>Represents Kronecker product;
(3) Constructing a sparse ISAR high-resolution imaging network based on depth expansion:
3a) Expanding the iterative algorithm into an imaging network comprising a gradient descent layer and a proximal mapping layer;
3b) Constructing a controllable near-end mapping module comprising a near-end mapping unit and a controllable unit;
3c) Embedding a controllable near-end mapping module into a near-end mapping layer in an imaging network and combining the controllable near-end mapping module with a high-frequency feature extractorResidual reconstructor->Sequentially connecting, and jumping and connecting the output of the gradient descent layer with the output of the residual error reconstructor to form a controllable near-end mapping layer;
3d) Sequentially connecting the gradient descent layer and the controllable near-end mapping layer to obtain an nth layer of sub-network, and cascading a plurality of sub-networks to obtain a sparse ISAR high-resolution imaging network;
(4) Generating a training set:
4a) Scattering points with random positions and amplitude conforming to Gaussian distribution are used as training labels, different echo defect rates and signal to noise ratios are set to obtain different ISAR scenes, and then original images in different scenes are generated;
4b) Forming image pairs by the original image and the label image, and generating n image pairs as training sets, wherein n is more than or equal to 600;
(5) Training a sparse ISAR high-resolution imaging network:
5a) Using normalized mean square error NMSE as loss function of network training, selecting batch size data from training set each time, inputting into sparse ISAR high resolution imaging network, calculating loss value in each batchBy random gradientThe descent algorithm updates network parameters until all data in the training set are selected, and sums and averages the loss values obtained by all latches to obtain the loss after one-time network training;
5b) Repeating the step 5 a) until the network converges to obtain a trained sparse ISAR high-resolution imaging network;
(6) Inputting the measured data into a trained imaging network, and obtaining a final ISAR imaging result through forward propagation of the network.
2. The method of claim 1, wherein the step (1) of establishing a sparse observation model for echo defects comprises the steps of:
1a) Assuming that the radar transmits a linear frequency modulation pulse signal, the target is represented by Q scattering points, and an echo signal of a p-th target scattering point in a range fast time-azimuth slow time domain is obtained
Wherein f c Is the carrier frequency, T p Pulse width, gamma frequency modulation, t n =nt represents slow time, T is pulse repetition interval, n=1, 2,..n represents pulse number;representing a fast time; rect (·) represents a unit rectangular functionR p (t n ) Represents the distance between the p-th scattering point and the radar, A p A represents the back reflection coefficient of the p-th scattering point, a a (t n ) Is a function of azimuth window and->T a For the observation time, exp (·) represents an exponential operation with the base of the natural constant e, j represents an imaginary unit symbol;
1b) For echo signalsPerforming line-separating tone processing to obtain an echo s of the p-th scattering point in a distance frequency-slow time domain p (f,t n )：
Wherein,distance frequency, B is bandwidth; r is R ps For the instantaneous skew between the scattering point p and the reference point s, according to the turntable model, the instantaneous skew is expressed as: r is R ps (θ)≈x p sinθ+y p cosθ≈x p θ+y p ，(x p ,y p ) Representing the coordinates of the p-th scattering point of the target, θ=ωt n Omega represents the ISAR target rotational angular velocity; since the ISAR imaging integration time is short, the rotation angle ω of the target relative to the radar is small in the imaging integration time, so sin (ωt) n )≈ωt n ＝θ、cos(ωt n )≈1；
R is R ps Substitution of (θ) into s p (f,t n ) Can be obtained by:
1c) The distance and azimuth directions are discretized to f=mΔf and T, respectively, given the frequency interval Δf and the pulse repetition interval T n =nt, where m=1, 2,..m, n=1, 2,., N, M and N represent the number of sampling points and the total number of pulses, respectively; s is then p (f,t n ) Can be further converted into a discrete form s p (m,n)：
1d) Let ISAR image size be U×V, and element in image be i p (u,v)，u∈[1,U]，v∈[1,V]The method comprises the steps of carrying out a first treatment on the surface of the Echo signal s by fourier transformation of distance and azimuth directions p (m, n) is represented as follows:
wherein d 1 The distance vector represents a vector with length M obtained by randomly downsampling a vector with original length U under the condition of sparse observation, and d 1m Is d 1 The m-th element of (a); d, d 2 The azimuth defect vector is a vector with the length N obtained by randomly downsampling a vector with the original length V under the sparse observation condition, and d 2n Is d 2 N-th element of (a);
1e) Order theIs a complex domain distance dictionary, phi 1 The mth row and the mth column of the plural field elements are +.> Is complex domain Doppler dictionary, phi 2 The plural field element of the nth row and the nth column of (a) is +.>Obtaining a sparse observation model Y under the condition of echo defect:
Y＝Φ 1 XΦ 2 +E
wherein,for ISAR image, ++>For echo signals +.>Is a complex domain noise matrix.
3. The method of claim 1, wherein in step (2) an objective function meeting the L1 norm optimization criterion is constructed based on the sparse observation model Y in the echo defect case, and the implementing step includes:
2a) Vectorization operation Vect (·) is performed on the sparse observation model under the echo defect condition, so that y=Vect (Y), x=Vect (X), e=Vect (E), and a vectorized sparse observation model Y is obtained:
y＝Φ 0 x+e
wherein,is a complex domain distance dictionary->For a complex domain Doppler dictionary, T represents a transpose operation, < >>Represents Kronecker product;
2b) Constructing an objective function conforming to an L1 norm optimization criterion according to the vectorized sparse observation model
Wherein,representing the Euclidean norm of the vector, < >>Representing an ISAR scene to be solved, lambda representing a regularization coefficient, and x sparsity being constrained by an L1 norm.
4. The method according to claim 1, wherein the step 3 a) of expanding the iterative algorithm into an imaging network comprising a gradient descent layer and a proximal mapping layer comprises the steps of:
3a1) Intercepting N iterations in the 2D-ISTA iterative algorithm, and adding the auxiliary variable R contained in the nth iteration (n) Reconstruction result X (n) Corresponding to two different network layers, namely a gradient descent layer and a near-end mapping layer, wherein:
X (n) ＝S soft (R (n) ,α (n) )，
y is the input measurement value ρ (n) For the iteration step of the nth iteration,is a complex domain distance dictionary which is used for the data processing,is a complex domain Doppler dictionary; s is S soft (R (n) ,α (n) ) Representing a soft threshold shrink function, defined as S soft (R (n) ,α (n) )＝sgn(R (n) )max{|R (n) |-α (n) ,0}，α (n) A shrink threshold for the nth iteration; n is E [1, N]；
3a2) And sequentially connecting the gradient descent layer and the near-end mapping layer to obtain an nth layer of sub-network, and repeatedly cascading a plurality of sub-networks to obtain the ISTA unfolding imaging network.
5. The method according to claim 1, wherein constructing a controllable proximal mapping module comprising a proximal mapping unit and a controllable unit in step 3 b) comprises the steps of:
3b1) Designing a near-end mapping unit based on a residual network:
the unit comprises two convolution modules c 1 、c 2 And a jump connection, wherein: a first convolution module c 1 Consists of a 3X 3 convolution layer and a ReLU activation layer cascade; a second convolution module c 2 Consists of 3 x 3 convolutional layers; the output of each convolution module is represented asx is the input of the near-end mapping unit;
the output of the entire near-end mapping unit is denoted as z=relu (y 2 +x), wherein ReLU (·) represents a ReLU activation function;
3b2) Designing a controllable unit:
the controllable unit is formed by a full connection layer W CU The implementation is that the input is a conditional vector z containing echo defect information, and a controllable vector with dimension of C is generated
3b3) Embedding the controllable unit into the near-end mapping unit to obtain a controllable near-end mapping moduleIts output F (n) The method comprises the following steps:
wherein G is (n) Is the output in the front layer module, z isInput of->And->Respectively represent controllable near-end mapping modules +.>Is the first and last convolutional layer of the (c).
6. The method of claim 1, wherein the high frequency feature extractor comprising the near-end mapping layer in step 3 c) is configured toAnd residual reconstructor->Structural parameters, and output X of the entire controllable near-end mapping layer (n) The method comprises the following steps:
the high frequency feature extractorThe method consists of a convolution layer with a convolution kernel size of 3 multiplied by 3 and a convolution kernel channel number of 32;
said residual reconstructorThe method consists of convolution layers with the convolution kernel size of 3 multiplied by 32 and the convolution kernel channel number of 1;
the output X (n) Is that
Wherein,and->Respectively soft threshold shrink function S soft (-) front-to-back controllable proximal mapping module;high-frequency feature extractor for the nth layer of the network,>residual reconstructor for network n-th layer, alpha (n) A learnable shrink threshold for the nth layer of the network.
7. The method of claim 1, wherein the sparse ISAR high resolution imaging network obtained in step 3 d) has the following structure:
network input- & gt layer 1 gradient descent layer, controllable near-end mapping layer- & gt … - & gt layer N gradient descent layer, controllable near-end mapping layer- & gt … - & gt layer N gradient descent layer, controllable near-end mapping layer- & gt network output.
8. The method of claim 1, wherein the loss value in each batch is calculated in step 5 a)By random gradientThe network parameters are updated by the descent algorithm, and the implementation steps comprise the following steps:
5a1) Selecting batch-sized data from the training set each time, inputting the data into a sparse ISAR high-resolution imaging network, and calculating a loss value in each batch
Wherein,is F norm, N batch For the number of samples per batch, +.>Representing a reconstructed image output by an imaging network, and X represents an input label image;
5a2) Calculating gradient of loss function with respect to arbitrary parameter in sparse ISAR high-resolution imaging network by using complex domain back propagation algorithmAccording to the solved gradient->Updating network parameters of the sparse ISAR high-resolution imaging network to obtain network parameters theta' of the current training stage:
wherein Θ is 0 For the network parameters of the sparse ISAR high-resolution imaging network training phase, lr represents the learning rate during training.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311093320.1A CN117192548A (en) | 2023-08-28 | 2023-08-28 | Sparse ISAR high-resolution imaging method based on depth expansion |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311093320.1A CN117192548A (en) | 2023-08-28 | 2023-08-28 | Sparse ISAR high-resolution imaging method based on depth expansion |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117192548A true CN117192548A (en) | 2023-12-08 |
Family
ID=88982766
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311093320.1A Pending CN117192548A (en) | 2023-08-28 | 2023-08-28 | Sparse ISAR high-resolution imaging method based on depth expansion |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117192548A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114140325A (en) * | 2021-12-02 | 2022-03-04 | 中国人民解放军国防科技大学 | C-ADMMN-based structured sparse aperture ISAR imaging method |
| CN117892068A (en) * | 2024-03-15 | 2024-04-16 | 江南大学 | Flip chip ultrasonic signal denoising method and device |
-
2023
- 2023-08-28 CN CN202311093320.1A patent/CN117192548A/en active Pending
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114140325A (en) * | 2021-12-02 | 2022-03-04 | 中国人民解放军国防科技大学 | C-ADMMN-based structured sparse aperture ISAR imaging method |
| CN117892068A (en) * | 2024-03-15 | 2024-04-16 | 江南大学 | Flip chip ultrasonic signal denoising method and device |
| CN117892068B (en) * | 2024-03-15 | 2024-06-07 | 江南大学 | Flip chip ultrasonic signal denoising method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN117192548A (en) | Sparse ISAR high-resolution imaging method based on depth expansion | |
| CN107462887B (en) | Compressed sensing based wide cut satellite-borne synthetic aperture radar imaging method | |
| CN110068805B (en) | High-speed target HRRP reconstruction method based on variational Bayesian inference | |
| CN112099008B (en) | SA-ISAR imaging and self-focusing method based on CV-ADMMN | |
| CN113567985B (en) | Inverse synthetic aperture radar imaging method, device, electronic equipment and storage medium | |
| CN110161499B (en) | Improved sparse Bayesian learning ISAR imaging scattering coefficient estimation method | |
| CN110244303B (en) | SBL-ADMM-based sparse aperture ISAR imaging method | |
| CN111580104B (en) | Maneuvering target high-resolution ISAR imaging method based on parameterized dictionary | |
| CN110333489B (en) | Processing method for SAR echo data sidelobe suppression by adopting CNN and RSVA combination | |
| CN111948652B (en) | SAR intelligent parameterized super-resolution imaging method based on deep learning | |
| Li et al. | A computational efficient 2-D block-sparse ISAR imaging method based on PCSBL-GAMP-Net | |
| CN110109114B (en) | Scanning radar super-resolution imaging detection integrated method | |
| CN112147608A (en) | Rapid Gaussian gridding non-uniform FFT through-wall imaging radar BP method | |
| CN113466864B (en) | Rapid combined inverse-free sparse Bayes learning super-resolution ISAR imaging algorithm | |
| Mao et al. | Target fast reconstruction of real aperture radar using data extrapolation-based parallel iterative adaptive approach | |
| CN107193002A (en) | A kind of one-dimensional range profile high-resolution imaging method for suppressing wideband phase noise | |
| CN108562901B (en) | ISAR high-resolution imaging method based on maximum signal-to-noise-and-noise ratio criterion | |
| CN113030964B (en) | Bistatic ISAR (inverse synthetic aperture radar) thin-aperture high-resolution imaging method based on complex Laplace prior | |
| Hu et al. | Inverse synthetic aperture radar imaging using complex‐value deep neural network | |
| Nazari et al. | High‐dimensional sparse recovery using modified generalised sl0 and its application in 3d ISAR imaging | |
| CN116027293A (en) | Rapid sparse angle super-resolution method for scanning radar | |
| CN116577749A (en) | Scanning radar super-resolution method under unknown broadening of antenna pattern | |
| CN113030963B (en) | Bistatic ISAR sparse high-resolution imaging method combining residual phase elimination | |
| CN112946644B (en) | Based on minimizing the convolution weight l1Norm sparse aperture ISAR imaging method | |
| Zha et al. | An iterative shrinkage deconvolution for angular superresolution imaging in forward-looking scanning radar |
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 |