CN104268875A - SAR image registration method based on specific value correlation function - Google Patents

Abstract

The invention discloses an SAR image registration method based on the specific value correlation function. The defect that an existing correlation registration method can not adapt to an SAR image spot multiplicative noise model easily is mainly overcome. As for a given reference image and an image to be registered, the implementation process includes the steps that (1) a reciprocal image of the image to be registered is obtained; (2) fast Fourier transformation is carried out on the reference image and the reciprocal image, and conjugate multiplication is carried out in the frequency domain; (3) inverse transformation is carried out on the multiplication result to obtain the specific value correlation function; (4) minimum value position pixel-level shifting of the specific value correlation function is obtained; (5) the non-overlapping area is cut, and the process from the step (1) to the step (5) is executed once again; (6) frequency domain zero fill operation is carried out on the cut image, the Fourier inverse transformation within a certain range is calculated, and the minimum point position is sub-pixel shifting; (7) all pixel-level shifting estimation and all sub-pixel-level shifting estimation are added to obtain total shifting estimation. By means of the SAR image registration method, compared with other pixel-level and sub-pixel-level shifting estimation methods, the SAR image shifting estimation result is more stable.

Description

Based on the SAR image registration method of ratio related function

Technical field

The invention belongs to technical field of image processing, relate to a kind of method for registering images, specifically a kind of SAR image registration method based on ratio related function, can be used for the bias estimation right to SAR image.

Background technology

Image formed by synthetic-aperture radar SAR has round-the-clock, the feature such as round-the-clock, high resolving power and powerful penetration capacity, and therefore, this image is widely used target identification, change detection and surface surveillance.But, in these application, SAR image usually in pairs or ordering list existing, and due to the disturbance of radar, these original image log ensure the location of pixels in the accurate correspondence image of point in same physical location on same plane coordinate according to being difficult to.

In SAR image registration, due to the corrosion of multiplicative noise, the stepwise derivation that this noise reflects from backscatter radar, this speckle noise has damaged the radiometric resolution of SAR image, image on registration process have very large impact.At present, multiple method for registering is suggested.The typical algorithm of carrying out registration based on correlation technique of Pixel-level has phase correlation method, expansion phase correlation method, maximum mutual information entropy method etc.The method for registering typical operation of feature based has reference mark method, SIFT feature method etc.But these methods all only carry out registration on pixel level.In order to reduce registration bias, must carry out the bias estimation of sub-pixel, its precision will reach 1/10 pixel.The method of the pixel of sub-pixel is also suggested, the bias estimation of the sub-pixel of frequent use all employ by the thick bias estimation strategy to essence, as fast-NCCA, this method makes the skew of pixel control in 0-1 pixel coverage after other Pixel-level method for registering of execution, recycling lower order polynomial expressions carries out carrying out linear interpolation between pixel, utilizes optimized algorithm to find the sub-pixel location of maximal correlation.The method of related function interpolation, this method utilizes related function to process at frequency domain, make use of the character of Fourier transform---frequency domain zero padding is equivalent to spatial domain interpolation operation, this interpolation method is perfect interpolation method when meeting nyquist sampling theorem, and the operation among a small circle proposing sub-pixel is to improve travelling speed.But the method for registering of these Pixel-level registrations and sub-pixel does not all consider that SAR image coherent spot multiplicative noise is on the impact of bias estimation, and such registration accuracy can be reduced bias estimation error by the noise of varying strength.

Summary of the invention

The object of the invention is to the deficiency overcoming above-mentioned prior art, for the multiplicative model of SAR image speckle noise, propose a kind of SAR image registration method based on ratio related function, carry out the bias estimation that SAR image is right, to reduce the impact of noise on registration result, improve registration accuracy.

Technical scheme of the present invention is: based on the SAR image registration method of ratio related function, comprise the steps:

(1) for given reference picture i _mwith image i subject to registration _s, calculate image i subject to registration _sinverse image, be designated as subscript r represents reciprocal;

(2) respectively to reference diagram i _mwith the inverse image of figure subject to registration do fast fourier transform to go forward side by side line frequency territory conjugate multiplication, be shown below:

$I_m=\mathrm{FFT}\left(i_m\right),I_s=\mathrm{FFT}\left(i_s^r\right),R_{i_m,i_s}^R=I_m{I}_s^{*},$

Wherein, I _mfor image i _mfast fourier transform result, I _sfor image fast fourier transform result, FFT () is fast Fourier transform operations, for I _mand I _sconjugate product, subscript R represents the new ratio measure proposed, and * represents conjugation;

(3) right perform inverse fast fourier transform, obtain the result of ratio related function

(4) search for in minimum value coordinate position coordinate (u', v'), this coordinate is ratioing technigue and estimates SAR image Pixel-level skew (Δ x, Δ y), is shown below:

$(\mathrm{\Δx},\mathrm{\Δy})=\arg \underset{u\text{'},v\text{'}}{\min }\left\{r_{i_m,i_s}^R(u\text{'},v\text{'})\right\},$

(5) the nonoverlapping region of cutting obtains the reference picture after cutting with image subject to registration and repeat implement once (1)-(5) process obtain with wherein, subscript numeral cutting number of times;

(6) for the image pair obtained after step (5) twice cutting with perform (1)-(2) step, calculate carry out frequency domain zero padding, after calculating zero padding, the inverse fourier transform of matrix (-2,2) pixel coverage, obtains interpolation related function and search minimum value by following formula

$(\mathrm{\&Delta;x},\mathrm{\&Delta;y})=\arg \underset{u\text{'},v\text{'}}{\min }\left\{r_{i_m,i_s}^R(u\text{'},v\text{'})\right\},-2p<\mathrm{\&Delta;x},\mathrm{\&Delta;y}<2p,$

Wherein, subscript p represents up-sampling rate, p=32, it is then sub-pix bias estimation result;

(7) all Pixel-level and sub-pixel bias estimation are added, obtain the bias estimation of former reference diagram and figure subject to registration.

The inverse image of image subject to registration is calculated in above-mentioned steps (1) for each location of pixels (x, y), the computing formula used is as follows:

$i_s^r(x,y)=\frac{1}{i_s(x,y)}.$

Right in above-mentioned steps (3) the result of performed inverse fast fourier transform also be with reference to figure i _mwith figure i subject to registration _sratio related function, computing formula is as follows:

$r_{i_m,i_s}^R(u,v)=\mathrm{IFFT}\left(R_{i_m,i_s}^R\right)=\underset{x,y}{\mathrm{\&Sigma;}}\frac{i_m(x,y)}{i_s(x-u,y-v)}$

Wherein (u, v) is coordinate, IFFT () be inverse fast fourier transform operation.

Beneficial effect of the present invention: present invention uses by the thick registration strategies to essence, compared with prior art, the present invention has the following advantages:

1, the present invention considers that multiplicative speckle noise proposes a kind of related function based on ratio newly.

2, new in the present invention related function can better suppress SAR image speckle noise, is a kind of more stable method for registering.

3, the present invention has continued to use the efficient performance of correlation technique on frequency domain, and whole algorithm operation quantity is little, fast operation.

Below with reference to accompanying drawing, the present invention is described in further details.

Accompanying drawing explanation

Fig. 1 is process flow diagram of the present invention;

Fig. 2 has for 100,000 times the one-dimensional signal of time delay to add three to look the registration error statistical graph after amplitude noise;

Fig. 3 is used three reference diagrams looking Amplitude Composition SAR image in experiment;

The Analysis of Noise Influence figure of Fig. 4 tradition correlation technique and ratio correlation technique of the present invention;

Fig. 5 is a pair true SAR image---Chinese Yellow river mouth used in experiment.

Embodiment

With reference to accompanying drawing 1, the present invention includes following steps:

Step 1, structure ratio related function:

1.1) different from traditional related function, for given reference picture i _mwith image i subject to registration _sthe ratio related function of two dimensional image can carry out as given a definition:

$r_{i_m,i_s}^R(u,v)=\underset{x,y}{\mathrm{\&Sigma;}}\frac{i_m(x,y)}{i_s(x-u,y-v)}---\left(1\right)$

Wherein, (x, y) is input picture coordinate, and (u, v) is ratio related function coordinate;

1.2) for given reference diagram i _mwith figure i subject to registration _sby calculating i _sinverse image the ratio related function of this stylish definition can be write as the form of traditional related function:

$r_{i_m,i_s}^R(u,v)=\underset{x,y}{\mathrm{\&Sigma;}}\frac{i_m(x,y)}{i_s(x-u,y-v)}=\underset{x,y}{\mathrm{\&Sigma;}}{i}_m(x,y)i_s^r(x-u,y-v)---\left(2\right)$

Step 2, Pixel-level bias estimation:

2.1) related function can utilize Fourier transform to calculate at frequency domain, calculates i respectively _mwith fast fourier transform I _mand I _s

$I_m=\mathrm{FFT}\left(i_m\right),I_s=\mathrm{FFT}\left(i_s^r\right)---\left(3\right)$

Wherein, FFT () is fast Fourier transform operations

2.2) two width image frequency domains carry out conjugate multiplication and obtain

$R_{i_m,i_s}^R=I_m{I}_s^{*},---\left(4\right)$

Wherein, subscript R represents the ratio correlation metric approach in the present invention, and * represents conjugation;

2.3) right inverse fast fourier transform, what this inverse transformation obtained is then ratio related function result:

$r_{i_m,i_s}^R=\mathrm{IFFT}\left(R_{i_m,i_s}^R\right);---\left(5\right)$

Wherein IFFT () represents inverse fast fourier transform operation

2.4) search for the minimum value of middle coordinate (u', v') is ratioing technigue and estimates the skew (Δ x, Δ y) that SAR image is right:

$(\mathrm{\&Delta;x},\mathrm{\&Delta;y})=\arg \underset{u\text{'},v\text{'}}{\min }\left\{r_{i_m,i_s}^R(u\text{'},v\text{'})\right\},---\left(6\right)$

2.5) edge that the two width images that cutting has been estimated do not overlap obtains the reference picture after cutting with image subject to registration in order to the not overlapping edge reducing to have nothing to do is on the impact of registration result.Image after cutting is obtained after new cutting again repeating step 2 process with wherein, subscript numeral cutting number of times.Twice pixel-shift is estimated to be added, completes the bias estimation of Pixel-level;

Step 3, sub-pixel bias estimation

3.1) for the image after twice cutting with to what calculate after carrying out frequency domain zero padding, inverse transformation operates, and calculates the related function after interpolation to carrying out inverse fourier transform in (-2,2) scope

$(\mathrm{\&Delta;x},\mathrm{\&Delta;y})=\arg \underset{u\text{'},v\text{'}}{\min }\left\{r_{i_m,i_s}^R(u\text{'},v\text{'})\right\},-2p<\mathrm{\&Delta;x},\mathrm{\&Delta;y}<2p,---\left(7\right)$

Wherein, subscript p represents up-sampling rate, in the present invention, and p=32.IDFT _prepresent that above sampling rate p carries out inverse fourier transform after zero padding;

3.2) search by following formula the minimum value (Δ x, Δ y) of coordinate (u', v'):

$(\mathrm{\&Delta;x},\mathrm{\&Delta;y})=\arg \underset{u\text{'},v\text{'}}{\min }\left\{r_{i_m,i_s}^{R,p}(u\text{'},v\text{'})\right\},---\left(8\right)$

Now, it is then sub-pix bias estimation result;

3.3) above all registration result are added, obtain the bias estimation that two width SAR image are total.

Effect theory of the present invention proves:

According to the noise model i=ue of amplitude SAR image, and speckle noise u is without to make an uproar pixel e probability density function: noise average is 1, and variance is two noise spots can be obtained to be multiplied the variance after operator and ratio operator mean normalization:

Wherein, M=u ₁u ₂for noise is multiplied operator, R=u ₁u ₂for noise value operator, Var (.) represents variance.Suppose muting reference diagram e _mwith figure e subject to registration _scorrespondence position pixel is identical, then have e during registration _m=e _s=e.For the following two kinds related function:

Wherein C represents the average statistical of multiplication operator or ratio operator, the Section 1 quilt of its band noise

After registration without a peak value normalization stochastic variable of making an uproar be

$f^M(u,v)=\frac{{T_1}^M(u,v)}{{T_2}^M\left(\mathrm{0,0}\right)}$

$f^R(u,v)=\frac{{T_1}^R(u,v)}{{T_2}^R\left(\mathrm{0,0}\right)}$

Two stochastic variables calculate variance D

$D\left(f^M(u,v)\right)=\frac{D\left(M\right)\underset{x,y}{\mathrm{\&Sigma;}}{e}^2(x,y)e^2(x-u,y-v)}{{\left(\underset{x,y}{\mathrm{\&Sigma;}}{e}^2\right)}^2}$

$D\left(f^R(u,v)\right)=\frac{D\left(R\right)\underset{x,y}{\mathrm{\&Sigma;}}\frac{e^2(x,y)}{e^2(x-u,y-v)}}{N^2}$

From above formula, it is not only relevant with two noise operators that its related function often puts variance, also relevant with noise-free picture, and use image shown in Fig. 3 (a) to be example here, related function point span after its peak point mated completely and displacement is designated as S (u, v), defining variable Q=δ (f (u, v))/S (u, v), the ratio of standard deviation and span, the reflection of this variable due to noise effect, the size that each n-point correlation function n affects peak value.Fig. 4 draws two kinds of related functions and looks Q value change under noise effect three, as shown in the figure, the each point tolerance of ratio related function is far smaller than the impact of traditional multiplicative related function to real peak impact, and it is a kind of more stable correlation metric approach that ratio is correlated with under SAR image multiplicative noise.

Effect of the present invention can be confirmed further by following experiment:

1) one-dimensional signal registration experiment: given one dimension continuous signal s (t)=sin (5t)+sin (10t)+sin (20t), with T _s=0.0025 π carries out sampling and obtains sequence s ₁n (), gets s ₂(n)=s ₁(n+3), add three and look amplitude noise, carry out 100000 experiments, traditional correlation technique registration error and registration error of the present invention statistics are as shown in Figure 2;

2) two dimension synthesis SAR image registration experiment: three reference diagrams looking Amplitude Composition SAR image shown in Fig. 3, its image is to acquisition methods: former figure is amplified 10 times, the some pixels of translation, then downscaled images, obtains the sub-pixel displacement that the order of magnitude is 1/10.Its true excursions, the registration result of two kinds of methods, registration square error is listed in table one and table two.Two kinds of related functions to be looked under noise to the influence curve of peak value as shown in Figure 4 three;

3) the true SAR image registration experiment of two dimension: the true SAR image shown in Fig. 5 obtains to by the some pixel down-samplings of a pair high-definition picture translation the sub-pixel displacement that the order of magnitude is 1/10.Because High Resolution SAR Images deviation is at sub-pixel range, after reducing, can think that former figure is registering images.Fast-NCCA is used, traditional related function method of interpolation, new relevant function interpolation method for registering proposed by the invention in this experiment.Image true excursions, three kinds of method registration result and registration square error are listed in table three and table four;

The all experiments of the present invention all complete at matlab platform.

Table one, three look synthesis SAR image to bias estimation (Δ x, Δ y) result

Table two, three look synthesis SAR image to bias estimation square error (MSE) result

Table three, true SAR image is to bias estimation (Δ x, Δ y) result

True excursions	Fast-NCCA	Tradition related function	Ratio related function
				(0.2,0.8)	(0.4215,0.5724)	(0.1250,0.8750)	(0.1250,0.8750)
(0.4,0.6)	(0.4447,0.5243)	(0.1875,0.6875)	(0.1875,0.6875)
				(0.6,0.4)	(0.5196,0.5048)	(0.8125,0.5625)	(0.6875,0.5625)
(0.8,0.2)	(0.5612,0.4551)	(0.9375,0.2500)	(0.8750,0.3125)

Table four, true SAR image is to bias estimation square error (MSE) result

True excursions

Fast-NCCA

Tradition related function

Ratio related function

(0.2,0.8)	0.1588	0.0530	0.0530
				(0.4,0.6)	0.0439	0.1149	0.1149
(0.6,0.4)	0.0661	0.1338	0.0923
				(0.8,0.2)	0.1747	0.0732	0.0676
Square error average	0.1108	0.0937	0.0869

Above experimental result shows, the present invention more meets SAR image characteristic relative to other some classical sub-pixel registration result, and registration result is more accurate.It should be noted that the present invention has continued to use that the calculated amount of Fourier transform is little, the feature of fast operation.

The part that present embodiment does not describe in detail belongs to the known conventional means of the industry, does not describe one by one here.More than exemplifying is only illustrate of the present invention, does not form the restriction to protection scope of the present invention, everyly all belongs within protection scope of the present invention with the same or analogous design of the present invention.

Claims (3)

1. based on the SAR image registration method of ratio related function, it is characterized in that: comprise the steps:

(1) for given reference picture i _mwith image i subject to registration _s, calculate image i subject to registration _sinverse image, be designated as subscript r represents reciprocal;

(2) respectively to reference diagram i _mwith the inverse image of figure subject to registration do fast fourier transform to go forward side by side line frequency territory conjugate multiplication, be shown below:

$I_m=\mathrm{FFT}\left(i_m\right),I_s=\mathrm{FFT}\left(i_s^r\right),R_{i_m,i_s}^R=I_m{I}_s^{*},$

Wherein, I _mfor image i _mfast fourier transform result, I _sfor image fast fourier transform result, FFT () is fast Fourier transform operations, for I _mand I _sconjugate product, subscript R represents the new ratio measure proposed, and * represents conjugation;

(3) right perform inverse fast fourier transform, obtain the result of ratio related function

(4) search for in minimum value coordinate position coordinate (u', v'), this coordinate is ratioing technigue and estimates SAR image Pixel-level skew (Δ x, Δ y), is shown below:

$(\mathrm{\&Delta;x},\mathrm{\&Delta;y})=\arg \underset{u\text{'},v\text{'}}{\min }\left\{r_{i_m,i_s}^R(u\text{'},v\text{'})\right\},$

(5) the nonoverlapping region of cutting obtains the reference picture after cutting with image subject to registration and repeat implement once (1)-(5) process obtain with wherein, subscript numeral cutting number of times;

(6) for the image pair obtained after step (5) twice cutting with perform (1)-(2) step, calculate carry out frequency domain zero padding, after calculating zero padding, the inverse fourier transform of matrix (-2,2) pixel coverage, obtains interpolation related function and search minimum value by following formula

$(\mathrm{\&Delta;x},\mathrm{\&Delta;y})=\arg \underset{u\text{'},v\text{'}}{\min }\left\{r_{i_m,i_s}^R(u\text{'},v\text{'})\right\},-2p<\mathrm{\&Delta;x},\mathrm{\&Delta;y}<2p,$

Wherein, subscript p represents up-sampling rate, p=32, it is then sub-pix bias estimation result;

(7) all Pixel-level and sub-pixel bias estimation are added, obtain the bias estimation of former reference diagram and figure subject to registration.

2. the SAR image registration method based on ratio related function according to claim 1, is characterized in that: the inverse image calculating image subject to registration in step (1) for each location of pixels (x, y), the computing formula used is as follows:

$i_s^r(x,y)=\frac{1}{i_s(x,y)}.$

3. the SAR image registration method based on ratio related function according to claim 1, is characterized in that: right in step (3) the result of performed inverse fast fourier transform also be with reference to figure i _mwith figure i subject to registration _sratio related function, computing formula is as follows:

$r_{i_m,i_s}^R(u,v)=\mathrm{IFFT}\left(R_{i_m,i_s}^R\right)=\underset{x,y}{\mathrm{\&Sigma;}}\frac{i_m(x,y)}{i_s(x-u,y-v)}$

Wherein (u, v) is coordinate, IFFT () be inverse fast fourier transform operation.