CN116243313A

CN116243313A - SAR rapid intelligent sparse self-focusing technology based on distance partition

Info

Publication number: CN116243313A
Application number: CN202310251820.7A
Authority: CN
Inventors: 丁泽刚; 卫扬铠; 汪子文; 李凌豪; 袁跳跳; 李埔丞
Original assignee: Beijing Institute of Technology BIT
Current assignee: Beijing Institute of Technology BIT
Priority date: 2023-03-08
Filing date: 2023-03-08
Publication date: 2023-06-09

Abstract

In order to solve the problems of high computational complexity and poor focusing effect in the traditional SAR imaging self-focusing method based on sparsity, a SAR rapid intelligent sparse self-focusing technology based on distance partitioning is provided. Firstly, reconstructing distance data of radar echoes, then carrying out distance partitioning, accurately reconstructing a partitioned observation matrix through oblique distance resampling, and reducing calculation and storage burden of subsequent sparse imaging through a distance partitioning strategy. Then, based on the joint constraint of the minimum entropy and the sparsity of the image, a network module is constructed, wherein the network module is formed by mutually and circularly iterating a sparse imaging module and a sparse self-focusing module, and the sparse imaging module is developed into an iterative neural network by an cross direction multiplication method (ADMM) under the motion error. The core module of sparse self-focusing is to estimate the motion error under minimum entropy constraints using a depth network. The method aims at providing a high-intelligent and high-efficiency SAR rapid intelligent sparse self-focusing scheme, and is expected to be applied to the fields of airborne, foundation SAR self-focusing and the like.

Description

SAR rapid intelligent sparse self-focusing technology based on distance partition

Technical Field

The invention aims to provide a high-intelligence and high-efficiency SAR intelligent sparse self-focusing solution, which can be expected to be used in the fields of airborne radar SAR imaging, ground-based radar SAR imaging and the like, can greatly improve the self-focusing effect and the operation speed of SAR sparse imaging, and is beneficial to improving the quality of SAR sparse images.

Background

The synthetic aperture radar (Synthetic Aperture Radar, SAR) is a high-resolution radar which can work around the clock in the whole day without being limited by weather, illumination and other conditions, and a two-dimensional high-resolution image [1] of a target imaging area is obtained in the distance direction and the azimuth direction through pulse compression and a synthetic aperture technology respectively.

SAR imaging actually aims at the problem of estimating target parameters in an imaging scene from the essence, namely, the radar actively transmits electromagnetic waves, then receives echoes of targets with different transmission coefficients in the imaging scene, and the reflection coefficients of the targets are inverted from the echoes by adopting a signal processing method. At present, the traditional SAR imaging method basically adopts a matched filtering (Matched Filtering, MF) mode to image, and energy accumulation can be obtained in a target area by constructing a specific matched signal form to perform correlation processing with radar received echoes, so that the effect of inverting a target scene is achieved, but a SAR imaging result obtained by the linear matching mode can generate a target main lobe and a side lobe, the side lobe can 'blur' real target information, and the readable information quantity of an image is reduced. In order to solve the problems, researchers combine the compressed sensing (Compressed Sensing, CS) theory with the SAR imaging theory to establish a brand new SAR imaging method: SAR sparse imaging method. Compared with the traditional SAR imaging method, the SAR sparse imaging method utilizes the sparse priori information of the target scene, solves the SAR imaging inverse problem by combining nonlinear processing knowledge, and can break through the performance limit of the traditional SAR imaging method to obtain a high-quality SAR imaging result without side lobes.

However, the SAR sparse imaging method faces two major problems: traditional SAR sparse solutions, for example: the cross direction multiplier method (Alternating Direction Method of Multipliers, ADMM) has the problem of high computational complexity and storage complexity due to the need of performing repeated iterative inversion operation; the SAR sparse imaging method is influenced by motion errors introduced by a platform, so that a reconstructed imaging result is defocused, and the readability of an image is influenced. The traditional sparse self-focusing method cannot adapt to complex imaging scenes through the idea of echo matching. Therefore, how to combine the sparse imaging method to perform self-focusing processing to quickly obtain the high-quality SAR sparse image becomes a great difficulty in the SAR sparse imaging field.

Aiming at the problems, limitations of the traditional SAR sparse imaging self-focusing method need to be improved from the precision direction of solving speed, and the invention provides a SAR rapid intelligent sparse self-focusing technology based on distance partitioning. In the invention, firstly, pulse reconstruction is carried out on distance data of radar echo, and partitioning is carried out, the observation matrix after partitioning is accurately reconstructed through inclined distance resampling, and the calculation and storage burden of subsequent sparse imaging is lightened through a distance partitioning strategy. Then, based on the joint constraint of the minimum entropy and the sparsity of the image, a network module is constructed, wherein the network module is formed by mutually and circularly iterating a sparse imaging module and a sparse self-focusing module, and the sparse imaging module is developed into an iterative neural network by an cross direction multiplication method (ADMM) under the motion error. The sparse self-focusing core module utilizes a depth network to estimate a motion error under the minimum entropy constraint, and finally obtains a SAR sparse imaging result with good focusing. The invention aims to provide a high-intelligence and high-efficiency SAR intelligent sparse self-focusing solution, which can be expected to be used in the fields of airborne radar SAR imaging, ground-based radar SAR imaging and the like, can greatly improve the self-focusing effect and the operation speed of SAR sparse imaging, and is beneficial to improving the quality of SAR sparse images.

The relevant documents retrieved are given below:

disclosure of Invention

The invention aims to overcome the defects of the prior art, solve the problems of high computational complexity and poor focusing effect in the traditional SAR sparse imaging and self-focusing method, and provides a SAR rapid intelligent sparse self-focusing technology based on distance partitioning. The specific implementation flow chart is shown in fig. 1.

The method is realized through the following steps:

step one, sparse reconstruction is carried out on the distance direction of a radar echo by a sparse reconstruction method to obtain a distance reconstruction result, and distance partitioning is carried out on the distance reconstruction result to obtain a plurality of distance partition data.

And secondly, accurately establishing a mapping relation between the distance partition data and the imaging scene partition through an oblique distance resampling strategy based on the distance partition data.

And thirdly, establishing a sparse imaging network module based on an ADMM iteration strategy, expanding an ADMM iteration process into a neural network iteration module, and learning super parameters in the iteration process to obtain a sparse imaging result.

And step four, based on the sparse imaging network module, establishing a sparse self-focusing network module based on minimum entropy constraint of the image.

And fifthly, training the iterative network module to obtain a SAR sparse imaging result with good focusing.

The invention has the advantages that:

(1) The radar data distance partition idea is established, and the mapping relation between the distance partition data and the imaging scene partition is accurately established through the oblique distance resampling strategy, so that the calculated amount and the storage amount of the follow-up SAR sparse imaging module can be greatly reduced.

(2) Aiming at the problems that the super-parameter selection is difficult and the echo matching cannot be applied to complex scenes in the traditional SAR sparse imaging and self-focusing method, a neural network based on expansion iteration and a method based on minimum image entropy are provided, the network unsupervised learning is performed through the ideas of echo loss and minimum image entropy, and the SAR sparse result with good focusing is obtained.

Drawings

FIG. 1 is a block flow diagram of a distance-partition-based SAR rapid intelligent sparse self-focusing technique

FIG. 2 is a schematic diagram of the geometrical relationship of a radar platform and a point target

FIG. 3 is a schematic diagram of a mapping relationship between distance zone data and imaging scene zones

FIG. 4 is a schematic diagram of a sparse imaging network module

FIG. 5 sparse self-focusing module schematic

FIG. 6 is a schematic view of a scene in an example of implementation

Imaging results of different methods and truth values in the example of FIG. 7

Detailed Description

Embodiments of the method of the present invention will be described in detail below with reference to the accompanying drawings and examples.

The invention relates to a SAR rapid intelligent sparse self-focusing technology based on distance partitioning, wherein a flow chart is shown in figure 1, and the specific steps comprise:

As shown in fig. 2, which is a schematic diagram of the geometrical relationship between a radar platform and a point target, the radar receives an echo of a point p in an imaging scene as follows:

wherein sigma _p (x, y) is the backscattering coefficient of point p, R (t) _a X, y) is the distance from the radar to point p, f is the radar distance frequency point, c is the speed of light, t _a Is azimuth time. Considering that there are many point targets in the scene imaging grid, the following radar echoes can be obtained:

where σ (x, y) represents the backscattering coefficient of the point object in the scene, R ₁ And R is ₂ Respectively the range of the slant distance of the imaging scene S, f ₀ Is the radar start frequency. D (t) _a R) represents azimuth time t _a The sum of all the point targets with the distance r in the following scene can be expressed as:

where δ (·) is the impact function. Discretizing the imaging region and radar echo data in equation (2) can result in:

wherein D (t) _a ,r _i ) For the result of the oblique distance discrete, L is the number of the oblique distance discrete points, and the above formula is written into a matrix form:

s＝A·D (5)

s is an echo receiving signal, A is a radar observation matrix, and D is a distance reconstruction result of a target scene. The distance reconstruction result D can be obtained by solving through ADMM. Partitioning the distance reconstruction result D to obtain D= [ D ] ₁ ,L,D _K ]K is the number of partitions.

At a certain azimuth time, the data of each range direction r is the superposition of the targets with the radar skew distance r in the imaging scene, but because the imaging grid spacing cannot be infinitely small, grid point targets with the radar skew distance r are not existed in most cases. As shown in FIG. 3, there is only one matching target along each azimuth row of the distance dimension in the scene for a certain azimuth moment and a certain skew data, and the skew r _k The matching targets for the following are not necessarily on grid points, and can be expressed as:

wherein sigma _k (p,q(p,t _a,m ,r _k ) If p and q are the grid azimuth and distance points, respectively, the slant distance is r _k Scattering coefficient of (2). The corresponding skew on the grid points is not necessarily skew data D _k (t _a,m ,r _k ) Instead, the component parts of (1) are mostly composed of non-grid points, and the skew information r _k For our needs, we need to acquire the backscattering coefficient of the corresponding slope under the non-grid point by the slope resampling method, we use the grid point slope information around the non-grid point, and we can get the following expression:

where dr is the skew resolution, and the above formula is rewritten into a matrix form to obtain:

D _k ＝A _k x _k (8)

so far, the accurate mapping relation between the partition data and the imaging scene partition is obtained.

Sparse imaging module is intended to use range-zone echo D _k And distance partition measurement matrix A _k Acquiring SAR image x _k . In the traditional SAR sparse imaging method, such as ADMM, the over-parameter selection in the iteration process very influences the imaging result, so that the applicability of the SAR sparse imaging method in SAR imaging is limited.

Thus, by expanding the iterative process of ADMM into the form of a neural network, we can learn the hyper-parameters in the iterative process to find the optimal hyper-parameters. As shown in fig. 4, which is a schematic diagram of a sparse imaging network module, we convert the three-step iterative process in ADMM into three network modules: x, Z, M can be expressed as:

wherein F is _x (·)、F _z (. Cndot.) and F _m Partition of (The mapping functions of the network modules X, Z, M are shown, respectively. ρ ^k And lambda (lambda) ^k The network may learn super parameters, but these parameters are fixed in the ADMM. To form a self-supervised learning network, the echo reconstruction loss is used to learn network super-parameters, which can be expressed as:

wherein the method comprises the steps of

Representing reconstructed echo->

Is the reconstruction result. So far, a sparse reconstruction result can be obtained, but the result has a defocusing phenomenon, and a focusing module of the next step is required to process.

Self-focusing network module utilizing sparse imaging network output

And an observation matrix A _k Estimating a distance partition error matrix E _k As a result of introducing motion errors, equation (8) becomes:

D _k ＝E _k A _k x _k (11)

the proposed sparse self-focusing module for estimating the motion error matrix consists of a series of focusing network layers. As shown in fig. 5, the network may be represented as an input (estimated SAR image x _k And distance partition measurement matrix A _k ) To output (distance zone motion error matrix E _k ) The mapping of (1) uses a convolutional layer and a pooling layer as follows:

wherein F is ^(m) (·) denotes the mth network module,

representing a motion error matrix for the nth iteration and

sparse imaging network output image +.>

As a loss function, can be expressed as:

wherein Net _image (. Cndot.) represents the forward process of the imaging module network, en (-) represents entropy of the image, and the smaller the entropy is, the better the image focusing effect is. In each iteration process, the observation matrix is iterated, namely:

is used for ensuring that the final solving result can be converged. The optimal solution is realized through gradient descent:

all network operations take complex calculations into account and convert complex operations into real operations, such as quadrature demodulation, which can be expressed as:

wherein the method comprises the steps of

And->

The operations of taking out the real and imaginary parts, respectively. The sparse SAR result with good focusing can be obtained through multiple iterations of the sparse imaging module and the self-focusing module.

Thus, SAR parameterized self-focusing technology based on the generation of the countermeasure intelligent network is realized.

Examples of the embodiments

As shown in fig. 7, a schematic diagram of a simulation scene is shown, a lattice target is placed in the scene, the lattice target comprises 16 points, a real radar track is obtained by adding a motion phase error, and in order to improve universality, additive white gaussian noise is added to an echo to an imaging result signal-to-noise ratio snr=23 dB, and simulation radar parameters are shown in table 1.

Table 1 radar parameters in the example of implementation

Radar parameters	Numerical value	Unit (B)
			Carrier frequency	10.0	GHz
Bandwidth of a communication device	1.0	GHz
			Pitch angle	30	deg
Synthetic aperture length	200	m

As can be seen in fig. 7 (b), the conventional ADMM method results in defocusing of the imaging result due to motion errors. As shown in fig. 7 (c), although the conventional sparse self-focusing method has a point focusing effect, the effect is not very good and contains many noise points. In contrast, as shown in fig. 7 (d), the proposed algorithm can obtain a well-focused result with a distance block number k=8. For quantitative evaluation of the proposed algorithm index, the difference between the reconstructed result and the true result is measured by using normalized mean square error (Normalized mean square error, NMSE), defined as:

wherein g (i, j) and

the true image and the reconstruction result are respectively represented, and M and N are respectively the image sizes. In the implementation example, the evaluation indexes are shown in table 2, so that the proposed algorithm can reconstruct the target real scene faster and more accurately than other algorithms.

Table 2 example evaluation index

Method	ADMM method	Traditional sparse self-focusing method	The method
				NMSE	2.49	0.25	0.04
Run time	45s	201s	201s

Therefore, the technology provided by the invention can solve the problem that the conventional sparse self-focusing method cannot quickly and accurately focus sparse, and can well process the quick self-focusing problem in a sparse scene. In summary, the above embodiments are only preferred embodiments of the present invention, and are not intended to limit the scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. The SAR rapid intelligent sparse self-focusing method based on the distance partition is characterized by comprising the following steps of:

2. The method for rapid intelligent sparse self-focusing on SAR based on distance partitioning according to claim 1, wherein in the first step, the geometrical relationship between the radar platform and the point target is schematically shown, and the echo of the point p in the imaging scene received by the radar is:

wherein sigma _p (x, y) is the backscattering coefficient of point p, R (t) _a X, y) is the distance from the radar to point p, f is the radar distance frequency point, c is the speed of light, t _a Is azimuth time.

3. The distance-partition-based SAR rapid intelligent sparse self-focusing method of claim 1, wherein in the second step, a certain azimuth moment and a certain skew data have only one matching target along each azimuth row of a distance dimension in a scene, and the skew r _k The matching targets for the following are not necessarily on grid points, denoted as:

wherein sigma _k (p,q(p,t _a,m ,r _k ) If p and q are the grid azimuth and distance points, respectively, the slant distance is r _k Is a scattering coefficient of (a) is provided.

4. The distance-partition-based SAR rapid intelligent sparse self-focusing method according to claim 1, wherein the three-step iterative process in ADMM in step three is converted into three network modules: x, Z, M, expressed as:

wherein F is _x (·)、F _z (. Cndot.) and F _m (. Cndot.) represents the mapping functions of the network modules X, Z, M, respectively. ρ ^k And lambda (lambda) ^k The network may learn super parameters, but these parameters are fixed in the ADMM. To form a self-supervised learning network, echo reconstruction loss is used to learn network superparameters, expressed as:

wherein the method comprises the steps of

Representing reconstructed echo->

Is the reconstruction result.

5. The distance-partition-based SAR rapid intelligent sparse self-focusing method according to claim 1, wherein the mapping of input to output in step four uses a convolutional layer and a pooling layer as follows:

wherein F is ^(m) (·) denotes the mth network module,

represents the motion error matrix of the nth iteration and +.>

Sparse imaging network output image +.>

As a loss function, expressed as:

wherein Net _image (. Cndot.) represents the forward process of the imaging module network and En (-) represents entropy of the image.