CN112184643A

CN112184643A - Non-parametric SAR image self-adaptive resampling method

Info

Publication number: CN112184643A
Application number: CN202010997154.8A
Authority: CN
Inventors: 王岩; 丁泽刚; 李凌豪; 曾涛; 赵祎昆
Original assignee: Beijing Institute of Technology BIT
Current assignee: Beijing Institute of Technology BIT
Priority date: 2020-09-21
Filing date: 2020-09-21
Publication date: 2021-01-05
Anticipated expiration: 2040-09-21
Also published as: CN112184643B

Abstract

The invention discloses a non-parametric SAR image self-adaptive resampling method. By using the method, the SAR image resampling under the condition of spectrum folding under the large forward-tilted SAR geometry can be realized. The invention adaptively constructs a matched two-dimensional interpolation kernel by using a morphological method, and realizes the resampling of any SAR image by the two-dimensional convolution of finite points. The method is independent of imaging parameters and a specific imaging plane and has high robustness.

Description

Non-parametric SAR image self-adaptive resampling method

Technical Field

The invention relates to the technical field of synthetic aperture radars, in particular to a non-parametric SAR image self-adaptive resampling method.

Background

Synthetic Aperture Radar (SAR) can obtain two-dimensional high-resolution radar images of a target area all day long. The SAR image can be used for scene matching, elevation inversion, target identification, terrain classification and other application scenes. Resampling of SAR complex images is a key technology in SAR data processing, and is widely applied to SAR imaging and related processing. For example, in SAR imaging, a slant range image is required to obtain a ground range image without geometric distortion by resampling; in interferometric SAR and tomographic SAR processing, multiple complex images require image registration by resampling. Therefore, a robust method for resampling the SAR complex image needs to be provided.

However, resampling of SAR images mainly faces the following two challenges. Firstly, when the SAR operates in a large forward-bias mode, a spectrum folding phenomenon (spectrum folding) across an observation period occurs due to serious two-dimensional coupling of the spectrum of the large forward-bias SAR, so that the baseband resampling method is invalid. Secondly, the complexity of the parameterized SAR image spectrum model limits the application of the parameterized SAR image resampling method, i.e., a parameterized SAR image spectrum model applicable to any mode and any imaging plane cannot be established, so that a parameterized two-dimensional interpolation kernel applicable to any mode and any imaging plane cannot be constructed.

Currently, there are two main types of methods for resampling an SAR image. The first method is a parametric image resampling method, and the method constructs a matched image interpolation kernel by pushing an SAR image spectrum model under a specific parameter and a specific imaging plane, so that the spectrum folding phenomenon under a specific condition can be adapted, and a distortion-free resampling result is obtained; however, due to the complexity of the SAR image spectrum model, the method cannot be used for SAR image resampling of any mode and any imaging plane, and the universality and the robustness are low. The second type is a non-parametric method, including bilinear interpolation, bicubic convolution interpolation, bicubic spline interpolation, baseband truncation sinc interpolation, frequency domain zero filling method, etc., which realizes SAR image resampling by constructing a baseband interpolation kernel or zero filling at a high frequency of a frequency domain, however, when the method resamples an SAR image with a spectrum folding phenomenon, the spectrum truncation of the SAR image causes the image quality to be seriously reduced, so the method is invalid under the condition of large forward slope mode spectrum folding.

Disclosure of Invention

In view of the above, the present invention provides a non-parametric SAR image adaptive resampling method, which adaptively generates a matched two-dimensional interpolation kernel by using morphological operations and realizes resampling of any SAR image by two-dimensional convolution, and is applicable to resampling of SAR images under the condition of spectrum folding in the geometry of large forward-oblique SAR.

The invention relates to a non-parametric SAR image self-adaptive resampling method, which comprises the following steps:

step 1, performing two-dimensional fast Fourier transform on an SAR complex image to obtain an original frequency spectrum; carrying out binarization and expansion operations on the original frequency spectrum to obtain a binarized frequency spectrum;

step 2, carrying out spectrum copying on the binary spectrum to obtain a spectrum containing a complete baseband spectrum passband region;

step 3, refining the stop band of the frequency spectrum obtained in the step 2;

step 4, performing connected domain detection and morphological closing operation on the frequency spectrum subjected to stopband refinement to obtain a baseband frequency spectrum passband region;

and 5, performing IFFT conversion on the frequency spectrum obtained in the step 4, performing truncation and up-sampling to obtain an interpolation kernel, and resampling the SAR image by using the following formula:

wherein (m)_new,n_new) For resampling the coordinates, s, of the resulting image_new(m_new,n_new) The images are resampled; c. d is the relative subscript of the adjacent pixel used in the resampling process of each pixel point respectively; f. of_map1、f_map2The mapping relation between the coordinate to be interpolated and the original image coordinate is obtained; p is a two-dimensional up-sampling multiple; t is the number of the truncation points; tau is_mIs composed of

And pair of

Take downDifference between integers, τ_nIs composed of

And pair of

The difference between the rounded-down; (x, y) are two-dimensional coordinates of the original image;_m，_npixel spacing in x, y axes, s is the original image, s_{Kernel_Cut}Is the constructed interpolation kernel.

Preferably, in step 5, a sharpening window is added to the upsampled interpolation kernel, and then the SAR image is resampled.

Preferably, in the step 1, the binarized threshold is a spectrum average of the original spectrum.

Preferably, in the step 2, the binarized spectrum is copied by 3 times.

Has the advantages that:

the invention provides a non-parametric SAR image self-adaptive resampling method, which is characterized in that a morphological method is used for self-adaptively constructing a matched two-dimensional interpolation kernel, and the resampling of any SAR image is realized through two-dimensional convolution of finite points. The method is independent of imaging parameters and a specific imaging plane and has high robustness.

Drawings

Fig. 1(a), (b), and (c) are respectively common frequency spectrum diagrams of SAR images.

Fig. 2(a), (b), and (c) are schematic diagrams of common frequency spectrums after up-sampling by the conventional method, respectively.

Fig. 3(a), (b), and (c) are graphs of simulation results of the original spectrum of the image, the binarized spectrum, and the dilation operation, respectively.

Fig. 4(a), (b), and (c) are graphs of simulation results of image spectrum copying, morphological closing operation, and baseband connected domain detection, respectively.

Fig. 5(a), (b), and (c) are respectively a spectrum, an image domain, and an image domain contour map of the constructed interpolation kernel.

FIG. 6 is a flowchart of the overall method.

Fig. 7 is a simulated geometric relationship diagram.

Fig. 8(a), (b) are the resampling results of the conventional method and the proposed method, respectively.

Fig. 9(a), (b) are graphs of the coherence coefficients after interferometric registration using the conventional method and the proposed method, respectively.

Fig. 10 is a coherence coefficient statistical curve after interference registration.

Detailed Description

The invention is described in detail below by way of example with reference to the accompanying drawings.

The invention provides a non-parametric SAR image self-adaptive resampling method, which comprises the following steps:

firstly, obtaining an original frequency spectrum by two-dimensional fast Fourier transform of an SAR complex image, and then preliminarily obtaining a binary frequency spectrum through frequency spectrum binarization and expansion operation.

In particular, the method comprises the following steps of,

s11 two-dimensional fast Fourier transform

Establishing a rectangular coordinate system by taking the rows and columns of the SAR complex image as an X axis and a Y axis respectively, and expressing rho (X, Y) by the envelope of the point spread function, so that the SAR complex image is positioned in (X, Y)₀,y₀) The point spread function of the target can be expressed as:

wherein,_m，_npixel intervals of x and y axes, respectively, m and n are discretized x-axis and y-axis pixel coordinates, respectively,

m and N are respectively the number of sampling points on x and y axes, R₀Is the center slope of the object in the image, λ is the wavelength,

is the initial phase.

The original two-dimensional spectrum can be directly obtained by a two-dimensional Fast Fourier Transform (FFT). Due to the discreteness of the SAR image domain, thereforeThe spectrum has an implicit periodicity and the spectrum obtained by the FFT is actually the spectrum within one observation period. Thus, the two-dimensional FFT-derived spectrum S' (k, l; R)₀,x₀,y₀) Can be expressed as:

wherein P () represents the spectral envelope in the case of continuous image domain, a, b are integers, k, l are two-dimensional frequency axes after discrete periodicity, and are both integers

As can be seen from the above formula, the SAR image spectrum may be completely in the same observation period, as shown in fig. 1(a), or may be superimposed on other observation periods, and this phenomenon is called spectrum folding, as shown in fig. 1(b), (c). Fig. 2 shows a schematic diagram of a spectrum after baseband sinc up-sampling, and it can be seen that when a spectrum folding phenomenon exists, baseband sinc interpolation can cause the spectrum to be truncated, thereby causing image distortion.

S12, spectrum binarization

In order to preliminarily distinguish a passband region and a stopband region of a frequency spectrum, carrying out binarization judgment on an original frequency spectrum, wherein 1 is a passband and 0 is a stopband; the binarization threshold may be set as a spectral mean. Let the frequency spectrum after binarization be S'_2Value(k, l). The binarization result is shown in FIG. 3 (b).

S13, binary spectrum expansion

Fig. 3(b) shows a significant gap in the binarization result, which is caused by the coherent superposition of the spectrum of multiple scattering points, and needs to be closed by the dilation operation. The expansion operation can be represented by the following steps:

wherein D is an expansion structural element, D_k,lIndicating that the expansion structure element moves to a certain pixel point (k, l). Expansion ofAs shown in fig. 3(c), the subsequent binary spectrum shows that the gaps are substantially eliminated.

And secondly, carrying out spectrum replication to obtain a spectrum containing a complete baseband spectrum passband region.

In general, folding only exists in adjacent observation periods, namely, a complete baseband spectrum passband region can be obtained through 3 times of copying. The binarized spectrum after the spectrum copying can be represented as:

S_Copy(k+aM,l+bN)＝S′_Delation(k,l) (4)

wherein a is-1, 0,1, b is-1, 0, 1. The binarized spectrum after the spectrum copying is shown in fig. 4 (a).

And step three, refining the stop band to maximize the band pass area of the baseband spectrum.

Due to the influence of factors such as windowing and antenna weighting, the passband region obtained in the above steps may not be complete, thereby causing information loss in the interpolation process and distortion of the interpolation result. To solve this problem, the stop band needs to be refined and the center line of the stop band is extracted, so as to maximize the passband region of the baseband spectrum. Marking the frequency spectrum passband region after the stopband is thinned as Q_{Pass_Fat}Then the pass band of each period falls completely within that region as shown in fig. 4 (b).

And step four, performing connected domain detection and morphological closed operation to extract a baseband spectrum passband region.

S41 connected domain detection

After the steps are completed, the spectrum passband region includes a baseband spectrum passband region, however, the spectrum passbands of other periods are also in the region, so that connected domain detection is required to be performed, and the region where the baseband spectrum passband is located is extracted. And carrying out connected domain detection to extract a connected domain where the baseband frequency spectrum passband is located. Connected component detection can be expressed as:

Q_Choose＝X_twhen X is present_t＝X_t-1 (5)

Wherein,

wherein,

for the dilation operation, B is a structuring element.

S42 morphological close operation

After the connected domain detection, the region block where the baseband spectrum passband is located is extracted, and at this time, a narrow gap and a small hole may still exist, so that the narrow gap and the small hole need to be completely eliminated by expansion at this time. In order to keep the zone boundaries and area constant, it is therefore necessary to perform the same degree of erosion after the expansion has ended, i.e. to complete a closing operation. To this end, a passband Q has been extracted that contains the entire unfolded baseband spectrum_PDInner communicating region Q_ChooseI.e. by

The results of the connected component detection and morphological closing operations are shown in fig. 4 (c).

And step five, performing two-dimensional Inverse Fast Fourier Transform (IFFT) on the frequency spectrum in the step four, and performing truncation and upsampling to obtain a matched interpolation kernel.

The frequency spectrum after the fourth step can be used as the frequency spectrum of the interpolation kernel. However, the sampling of the interpolation kernel obtained by directly performing IFFT on the spectrum in the step four is sparse, which may result in a large interpolation error, so that after the spectrum extraction of the interpolation kernel is completed, zero padding in the frequency domain is required to implement upsampling of the interpolation kernel, and the upsampling multiple determines the interpolation precision. After up-sampling, the image domain interpolation kernel of continuous sampling can be approximately considered to be obtained, namely the spectrum of the discrete non-periodic interpolation kernel is recovered. P times of two-dimensional up-sampling is carried out, and two-dimensional IFFT transformation is carried out to obtain an approximate continuous interpolation kernel. In practical application, the interpolation kernel often needs to be truncated to reduce the operation amount of point-by-point interpolation, so that it is possible to realize interpolation in real-time processing, and if T-point two-dimensional truncation interpolation is to be performed, the upsampled interpolation kernel needs to be subjected to M_Cut、N_CutPoint truncation, and M_Cut＝N_CutP · T. The truncated interpolation kernel is

Having a frequency spectrum of

Where ρ is_Kernel() Representing the interpolated kernel envelope, P_Kernel() Representing the envelope of the interpolated kernel spectrum. Denotes convolution operation.

Due to truncation of an image domain, a certain ringing effect can be generated in an interpolation kernel frequency spectrum, so that some side lobes can appear after SAR image resampling, and in order to weaken the influence of truncation of an interpolation kernel, sharpening window processing is carried out on the interpolation kernel after upsampling.

Up to this point, a two-dimensional interpolation kernel adapted to the SAR complex image to be interpolated has been constructed. The frequency domain of the interpolation kernel is shown in fig. 5(a), and the image domain is shown in fig. 5(b) and (c).

And step six, obtaining an interpolation kernel by using the step five, and realizing the resampling of the SAR image through image domain convolution.

The coordinate (pixel) mapping relation between the grid to be interpolated and the original image is as follows:

wherein, tau_mIs composed of

And pair of

The difference between the rounded-down values, τ_nIs composed of

And pair of

The difference between the rounded-down; f. of_map1、f_map2For interpolating coordinates to originThe mapping relation between the image coordinates is determined by a specific resampling geometry and the like; m is_new、n_newTwo coordinate axes of the interpolated image.

And weighting the T multiplied by T neighborhood complex data mapped near the original image coordinates by utilizing an interpolation core to obtain an interpolated image. The interpolation process is shown as follows:

wherein s is_newFor the resampled image, c and d are respectively the relative subscripts of the adjacent pixels adopted in the resampling process of each pixel point. The interpolation kernel is considered approximately as continuous, in practice the nearest neighbor in the upsampled interpolation kernel is taken.

Therefore, the non-parametric SAR image self-adaptive resampling method is realized. Fig. 6 is an overall flow chart of the proposed resampling method.

Example 1

In this embodiment, two forward-tilt-mode SAR image resampling simulation experiments are designed, which are respectively a lattice target resampling experiment and a surface target interference SAR image registration simulation experiment. The geometrical relationship is shown in fig. 7. The simulation parameters are shown in table 1.

TABLE 1 computer simulation parameters

Parameter/unit	Numerical value	Parameter/unit	Numerical value
				Wave band	Ku	Distance/km	10
Height H/km	6	Azimuth of velocity/deg	10
				Number of azimuth points	4096	Pulse repetition frequency/kHz	4
Bandwidth/MHz	220	Sampling rate/MHz	260

Experiment one is to perform 16 times of up-sampling on a 3 x 3 lattice target scene SAR image, evaluate a point spread function, and use baseband truncation sinc interpolation as a comparison method. The processing result of the baseband truncation sinc interpolation is shown in fig. 8(a), and it can be seen that the target point is severely distorted after upsampling; the processing result of the proposed resampling method is shown in fig. 8(b), and it can be seen that the target point is still an ideal two-dimensional sinc function after upsampling, and there is no obvious distortion. The simulation result proves the applicability of the method under the large forward-bias SAR geometry and the high-precision characteristic of the method.

And the second experiment is to perform interference SAR image registration simulation on the opposite target scene. And simulating echoes of the two channels and imaging to obtain SAR complex images of the two channels, wherein the distance between the two channels is 2 meters and the two channels are distributed in the direction vertical to the speed. And in the image registration process, re-sampling operation in image translation is realized by respectively adopting baseband truncation sinc interpolation and the method. Then, the resampling accuracy of the two methods is compared by comparing the coherence coefficients of the registered master-slave two-channel images (the coherence coefficient reflects the image correlation, and the higher the coherence coefficient is, the more similar the images are). Fig. 9(a) is a correlation coefficient diagram of image registration by using truncated sinc interpolation, and it is shown that the visible values are all low, which indicates that truncated sinc interpolation resampling causes obvious distortion of the image, and is not beneficial to subsequent applications such as interferometry. Fig. 9(b) shows the result of image registration using the proposed resampling method, where the visible values are all above 0.9, thus demonstrating the high precision of resampling by the proposed method. Fig. 10 further counts the percentage of each level of the coherence coefficient, and it can be seen that the coherence coefficient after resampling by the method is intensively distributed in an interval slightly smaller than 1, thereby further explaining that image distortion caused by resampling by the method can be ignored, and being beneficial to subsequent high-precision applications such as interferometry.

In summary, the above description is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A non-parametric SAR image self-adaptive resampling method is characterized by comprising the following steps:

And pair of

The difference between the rounded-down values, τ_nIs composed of

And pair of

2. The non-parametric SAR image adaptive resampling method according to claim 1, wherein in said step 5, the SAR image is resampled after adding a sharpening window to the upsampled interpolated kernel.

3. The unparameterized SAR image adaptive resampling method according to claim 1, wherein in the step 1, the threshold value for binarization is a spectrum mean value of an original spectrum.

4. The unparameterized SAR image adaptive resampling method according to claim 1, characterized in that in the step 2, the binarized spectrum is copied by 3 times.