CN101901476A - SAR image de-noising method based on NSCT domain edge detection and Bishrink model - Google Patents

Abstract

The invention discloses an SAR (Synthetic Aperture Radar) image de-noising method based on NSCT (Non-Subsampled Contourlet Transform) domain edge detection and a Bishrink model, mainly solving the problems of scratch effect and detail loss caused by carrying out the SAR image de-noising process by means of non-subsampled contourlet transform. The method comprises the following steps of: carrying out the non-subsampled contourlet transform on a selected SAR image and dividing the image into six layers of sub-band coefficients; keeping the first layer and the second layer of sub-band coefficients invariable and contracting the third to six layers of the sub-band coefficients by using the Bishrink model; reconstructing an image by means of non-subsampled contourlet inverse transform, and detecting an edge of the reconstructed image to carry out mean value filtering on the image subjected to the edge detection to obtain a filtered image; and carrying out non-linear anisotropism dispersion on a difference image obtained by subtracting the input image from the filtered image to obtain a de-noised image. The invention can excellently maintain edge information of the image and point target characteristic information, and can be used for interpretation analysis in the SAR image and pre-processing of image understand.

Description

SAR image de-noising method based on NSCT territory rim detection and Bishrink model

Technical field

The invention belongs to digital image processing field, relate to a kind of denoising method of SAR image, this method can be used for the interpretation analysis in the SAR image and the pre-service of image understanding.

Background technology

The SAR image field that is widely used, but during the SAR image imaging, the scatter echo of imaging scatterer has relevant effect, makes image can not effectively react the scattering properties of ground object target, and this interference is called speckle noise in the SAR image.Because the coherence of imaging system, the SAR image is subjected to the pollution of noise inevitably, reduced image resolution, hidden image details, damaged the quality of image, caused great difficulty for Target Recognition and image interpretation.Therefore, it is very important to the subsequent treatment and the decipher of SAR image to suppress coherent speckle noise effectively.The main target that suppresses speckle noise is to keep minutia such as edge of image when effectively suppressing speckle noise again.

The filtering method of spatial domain is the SAR image filtering technology of using the earliest.Representational filtering method has Lee filtering, Kuan filtering, Frost filtering, Gamma-MAP and their enhancement mode filtering, adaptive filter method.These filtering all use partial statistics parameter such as average, variance to wait and describe true picture, and then estimate the actual value of pixel.These methods are applicable to that background is simple, texture information abundant image not, and for complex background, contain the SAR image that enriches texture, then can produce level and smooth phenomenon, can not well keep marginal information.

Frequency field coherent spot filtering method is by frequency domain transform, the SAR image is carried out the method for Filtering Processing.Be typically by behind the wavelet transformation, wavelet coefficient handled the effect that reaches denoising.Based on the filtering method of wavelet transformation, studying more at present is by the research spot wavelet coefficient to be influenced, and according to the model of noisy wavelet coefficient, wavelet coefficient is carried out thresholding shrink process, reconstructed image then.This method can keep the edge when utilizing wavelet technique to suppress speckle noise to a certain extent, but has sacrificed level and smooth performance, causes the target in the image to be lost easily, and subsequent image processing and decipher analytic band are come error result.

Summary of the invention

The objective of the invention is to overcome the deficiency of above-mentioned prior art, a kind of SAR image de-noising method based on NSCT territory rim detection and Bishrink model is proposed, can keep edge of image and target property effectively, for subsequent image processing and interpretation analysis provide accurate target characteristic and edge conservation degree.

Realize that technical scheme of the present invention is is transformation tool with the non-down sampling contourlet, coefficient after adopting the Bishrink model to conversion shrinks, and the non-down sampling contourlet inverse transformation obtains reconstructed image, detects the edge of reconstructed image, homogeneous area is carried out mean filter, obtain image after the filtering; Input picture and filtered error image are carried out the Anisotropic Nonlinear diffusion, obtain the image after coherent spot suppresses.Concrete steps are as follows:

(1) the test pattern I that chooses being carried out non-down sampling contourlet NSCT conversion, is 6 straton band coefficients with testing SA R picture breakdown;

(2) keep the sub-band coefficients of layers 1 and 2 constant;

(3) the 3rd～6 layer sub-band coefficients is shunk with the Bishrink model;

(4) to the sub-band coefficients after shrinking through step (3), carry out the non-down sampling contourlet inverse transformation and obtain reconstructed image R;

(5) reconstructed image is carried out following rim detection and mean filter;

(5a) each pixel in the reconstructed image is got 3 * 3 fields, in this neighborhood, set 12 directions, note P ₁And P ₂Be respectively the mean value of both sides pixel on 12 directional rays, establish p=P ₁/ P ₂, if p≤1, then ratio detects factor r=p, otherwise r=p ^-1, remember that the minimum value of the ratio detection factor on 12 directions is r _MinIf edge detection threshold T ₀: 0.5≤T ₀≤ 1, if r _Min＞T ₀, think that then current point is a marginal point, be labeled as 1, otherwise be labeled as 0; After all element markings are intact, if pixel is labeled as 1, be that 3 * 3 neighborhood window is got at the center with it then, is that 1 number is less than 4 as if this neighborhood window internal labeling, then this pixel is labeled as 0; Realization is to the rim detection of reconstructed image;

(5b) detect the edge of reconstructed image after, (i j) gets its 5 * 5 neighborhood W for the center to be labeled as 0 pixel R with each ₁, if W ₁Internal labeling is that 1 pixel number is less than 4, and then (i, pixel value j) is taken as W to R ₁The mean value of interior all pixels; (i j) is the center, gets its 3 * 3 neighborhood W otherwise with R ₂, if W ₂Internal labeling is that 1 pixel number is less than 4, and then (i, pixel value j) is taken as W to R ₂In the mean value of all pixels, otherwise R (i, pixel value j) is taken as W ₂In all are labeled as the mean value of 0 pixel, finish the mean filter of image after the rim detection, obtain image U after the filtering;

(6) original image I and filtered image U are subtracted each other, and the error image v that obtains is carried out Anisotropic Nonlinear diffusion k time, k=140-230;

(7) Anisotropic Nonlinear is spread the image v that k iteration obtains ^kWith image U addition after the filtering, obtain the image after coherent spot suppresses;

The present invention compared with prior art has following advantage:

1, the present invention can effectively keep the detailed information such as point target of image.

Because the direction of non-down sampling contourlet itself is more, energy to point target behind the point target rarefaction representation in the image can be dispersed, therefore make the point target of image fuzzy after the non-down sampling contourlet inverse transformation, and the used error image of the present invention carry out the point target that the method for Anisotropic Nonlinear diffusion can effectively keep image.

2, the present invention can keep edge of image preferably.

Because non-down sampling contourlet transform has multiple dimensioned, multi-direction and translation invariance, and its odd function satisfies anisotropy, make non-down sampling contourlet transform have very strong directivity and anisotropy, the edge that can optimum rarefaction representation has straight line and curve singularity, therefore, the present invention can more effective maintenance edge of image than existing method such as Lee filtering and small echo, make that image border after coherent spot suppresses keeps better.

3, the present invention can eliminate the cut effect.

The present invention carries out mean filter to the homogeneous area of image after by rim detection, can eliminate the not enough cut effect of being brought of non-down sampling contourlet itself.

Description of drawings

Fig. 1 is a schematic flow sheet of the present invention;

Fig. 2 is 12 directional ray synoptic diagram setting during rim detection among the present invention;

Fig. 3 is with the present invention and the existing method coherent spot inhibition effect contrast figure to test pattern Horse track one 256 * 256;

Fig. 4 is with the present invention and the existing method coherent spot inhibition effect contrast figure to test pattern Horse track two 256 * 256;

Fig. 5 is with the present invention and the existing method coherent spot inhibition effect contrast figure to test pattern Bedfordshire 256 * 256.

Embodiment

With reference to Fig. 1, specific implementation step of the present invention is as follows:

Step 1: choose testing SA R image I, it is carried out non-down sampling contourlet transform.

The mathematical model that the present invention chooses testing SA R image is: y=x+n

Wherein y={y (i, j) | i, j=1,2 ... N} represents the SAR image, x={x (i, j) | i, j=1,2 ... N} represents the backscatter intensity of SAR image real scene, n={n (i, j) | i, j=1,2 ... N} represents that zero-mean and variance equal σ ²Gaussian noise, the size of N presentation video;

The testing SA R image of choosing is carried out non-down sampling contourlet transform, is 6 straton band coefficients with testing SA R picture breakdown; y ₁Be the coefficient of S layer, y ₂Be the coefficient of S-1 layer, S=3,4,5,6.

Step 2: the coefficient of the layers 1 and 2 behind the maintenance SAR image non-down sampling contourlet transform is constant, and employing Bishrink model shrinks the 3rd～6 layer coefficient.

The Bishrink model is used in the wavelet coefficient shrink process at first, and the present invention uses it for the non-profile ripple that adopts down.Bishrink model coefficient contraction process is as follows:

Estimate current coefficient y with the prior estimate form ₁Noise variance

${\mathrm{\&sigma;}}_{\mathrm{<figure-callout\; id="2"\; label="n"\; filenames="HSA00000186146500041.png"\; state="\{\{state\}\}">n</figure-callout>}}^2=\frac{\mathrm{median}\left(\right|{\mathrm{<figure-callout\; id="1"\; label="y"\; filenames="HSA00000186146500021.png"\; state="\{\{state\}\}">y</figure-callout>}}_1\left|\right)}{0.6745}$

Estimate current coefficient y with the method for local auto-adaptive ₁The edge of model variances sigma ²:

Wherein

Expression is worked as

The time σ ²=0, when The time

Wherein N (k) represents the coefficient y of S layer ₁Be the square window at center, S=3,4,5,6; M is the size of window coefficient, and value is 11;

Calculate the coefficient y of S layer ₁Estimated value

S=3,4,5,6:

${\hat{y}}_1=\frac{{[\sqrt{y_1^2+y_2^2}-\frac{\sqrt{3}{\mathrm{\&sigma;}}_{\mathrm{<figure-callout\; id="2"\; label="n"\; filenames="HSA00000186146500041.png"\; state="\{\{state\}\}">n</figure-callout>}}^2}{\mathrm{\&sigma;}}]}_{+}}{\sqrt{y_1^2+y_2^2}}{y}_1$

In the formula

y ₁Be the coefficient of S layer, y ₂Be the coefficient of S-1 layer, S=3,4,5,6;

Wherein

Expression is worked as

The time result be zero, when The time result remain unchanged.

Step 3: the coefficient after the shrink process is carried out the non-down sampling contourlet inverse transformation, obtain reconstructed image R.

Step 4: detect the edge of reconstructed image R, the image after the rim detection is carried out mean filter, obtain image U after the filtering.

(4a) (i j) gets 3 * 3 neighborhood window, calculates the pixel average P on 12 directional ray both sides in this window respectively to each pixel R of reconstructed image ₁And P ₂, 12 directional rays that the present invention adopts as shown in Figure 2; If p=P ₁/ P ₂, if P≤1, then ratio detects factor r=p, otherwise r=p ^-1, find out 12 minimum rates on the direction and detect factor r _MinEdge detection threshold T is set ₀: 0.5≤T ₀≤ 1, if r _Min＞T ₀, think that then (i j) is marginal point to R, is labeled as 1, otherwise be labeled as 0, after all element markings are finished, if pixel is labeled as 1, be that 3 * 3 neighborhood window is got at the center then, be less than 4, then this pixel is labeled as 0 if this neighborhood window internal labeling is 1 number with it; Finish rim detection to reconstructed image;

If (4b) (i j) is labeled as 0 to pixel R, and then (i j) gets its 5 * 5 neighborhood W for the center with R ₁, if W ₁Internal labeling is that 1 pixel number is less than 4, and then (i, pixel value j) is taken as W to R ₁The mean value of interior all pixels; (i j) is the center, gets its 3 * 3 neighborhood W otherwise with R ₂, if W ₂Internal labeling is that 1 pixel number is less than 4, and then (i, pixel value j) is taken as W to R ₂In the mean value of all pixels, otherwise R (i, pixel value j) is taken as W ₂In all are labeled as the mean value of 0 pixel, finish the mean filter of image after the rim detection, obtain image U after the filtering;

Step 5: original image and filtered image subtraction are obtained error image, error image is carried out k iteration of Anisotropic Nonlinear diffusion:

$v_{i,j}^{k+1}=v_{i,j}^k+\mathrm{\&tau;}\underset{r,s=-1}{\overset{1}{\mathrm{\&Sigma;}}}\frac{g\left({\sqrt{2}}^{1-\left|r\right|-\left|s\right|}\right|v_{i+r,j+s}-v_{i,j}\left|\right)(v_{i+r,j+s}-v_{i,j})}{({\mathrm{<figure-callout\; id="2"\; label="r"\; filenames="HSA00000186146500041.png"\; state="\{\{state\}\}">r</figure-callout>}}^2+s^2)}$

Wherein v is that original image I and filtered image U subtract each other the error image that obtains, v _{I, j}The i that is v is capable, the pixel of j row, v _{I+r, j+s}The i+r that is v is capable, the pixel of j+s row; R and s are the neighborhood of pixels value, (r, s) ≠ (0,0); Be to carry out anisotropy to spread error image before,

Be to carry out the error image after the anisotropy diffusion one time; τ is a scale parameter,

K=140-230; G is the diffusivity function, $g\left({\sqrt{2}}^{1-\left|r\right|-\left|s\right|}\right|v_{i+r,j+s}-v_{i,j}\left|\right)=\frac{1}{(1+\frac{{\left({\sqrt{2}}^{1-\left|r\right|-\left|s\right|}\right|v_{i+r,j+s}-v_{i,j}\left|\right)}^2}{{\mathrm{rr}}^2})},$

Wherein rr is a reduced parameter, rr=6,

Original image I and filtered image U can obtain nonlinear diffusion image v after subtracting each other k iteration of the above-mentioned Anisotropic Nonlinear diffusion of the error image v process that obtains ^k

Step 6: obtain nonlinear diffusion image v with after iteration k time ^kWith filtered image U addition, obtain the image after coherent spot suppresses.

Effect of the present invention further specifies by following emulation:

1, simulated conditions

Adopt image commonly used in the experiment of SAR image denoising: 1. size is 256 * 256, resolution is that 1m, equivalent number are 4 Ku wave band New Mexico Horse track one area SAR intensity image, shown in Fig. 3 (a); 2. size is 256 * 256, resolution is that 1m, equivalent number are 4 Ku wave band New Mexico Horse track two area SAR intensity image, shown in Fig. 4 (a); 3. size is 256 * 256, resolution is that 3m, equivalent number are 2 X-band Britain Bedfordshire area SAR magnitude image, shown in Fig. 5 (a); 3 width of cloth SAR images are used Lee filtering method, G-MAP filtering method, stationary wavelet soft-threshold method, non-down sampling contourlet method respectively as tested object altogether, abbreviate NSCT as, and the inventive method experimentize.

2, analysis of simulation result

To above-mentioned image simulation result 1. as shown in Figure 3, wherein Fig. 3 (b) is the simulation result of Lee filtering method, Fig. 3 (c) is the simulation result of G-MAP filtering method, Fig. 3 (d) is the simulation result of stationary wavelet soft-threshold method, Fig. 3 (e) is the simulation result of NSCT method, and Fig. 3 (f) is the simulation result of the inventive method;

To above-mentioned image simulation result 2. as shown in Figure 4, wherein Fig. 4 (b) is the simulation result of Lee filtering method, Fig. 4 (c) is the simulation result of G-MAP filtering method, Fig. 4 (d) is the simulation result of stationary wavelet soft-threshold method, Fig. 4 (e) is the simulation result of NSCT method, and Fig. 4 (f) is the simulation result of the inventive method;

To above-mentioned image simulation result 3. as shown in Figure 5, wherein Fig. 5 (b) is the simulation result of Lee filtering method, Fig. 5 (c) is the simulation result of G-MAP filtering method, Fig. 5 (d) is the simulation result of stationary wavelet soft-threshold method, Fig. 5 (e) is the simulation result of NSCT method, and Fig. 5 (f) is the simulation result of the inventive method;

As can be seen from Figure 3, existing Lee filtering method exists blooming, has lost detailed information such as edge and texture, shown in Fig. 3 (b); Existing G-MAP filtering method exists blooming equally, has occurred the bright spot that do not have originally in some place, and has lost detailed information such as whole weak textures and most point target, shown in Fig. 3 (c); Existing stationary wavelet soft-threshold method has kept the profile of image substantially, and the weak texture in the image has also been kept a part, and that still compares is fuzzy, shown in Fig. 3 (d); Existing NSCT method exists the edge and loses situation with distortion, shown in Fig. 3 (e); Homogeneous area of the present invention is smoother, and detailed information such as edge and point target also keep better, shown in Fig. 3 (f).

From Fig. 4, existing Lee filtering method exists blooming, and edge details is also by fuzzy, shown in Fig. 4 (b); The G-MAP filtering method also exists blooming, the bright spot that do not have originally occurred in some place, shown in Fig. 4 (c); Existing stationary wavelet soft-threshold method is excessively smooth to homogeneous area, shown in Fig. 4 (d); Homogeneous area of the present invention is smoother, and the edge also keep better, shown in Fig. 4 (f).

As can be seen from Figure 5, the white point target that existing Lee filtering, G-MAP filtering, three kinds of methods of stationary wavelet soft-threshold have all been lost test pattern, and also the edge is also by fuzzy, respectively shown in Fig. 5 (b), 5 (c) and 5 (d); The point target of existing NSCT method keeps better, but has lost a part of edge in the square region, and homogeneous area is level and smooth inadequately, shown in Fig. 5 (e); The present invention has kept detailed information such as point target and edge preferably, and has improved the flatness of homogeneous area than additive method, shown in Fig. 5 (f).To sum up, the present invention can not only effectively remove speckle noise, and can effectively keep image detail features such as edge of image and point target.

Denoising performance for concrete comparison the whole bag of tricks, provided the index of estimating Horse track one area SAR image de-noising method performance: the equivalent number of homogeneous area and image average, equivalent number is to weigh a common counter of SAR image de-noising method, equivalent number is big more, illustrates that denoising effect is good more; The mean value of pixel in the image average representative image generally requires the image average of denoising front and back to be consistent, and both differ big more, illustrate that the distortion of backward radiation degree is big more.

Table 1 is the SAR image denoising experimental index contrast of Horse track one area, comprises the contrast of equivalent number and image average; The equivalent number in three white box zones 1 in the table 1 in A, B, the C difference presentation graphs 3 (a), zone 2, zone 3.

The SAR image denoising experimental index contrast of table 1 Horse track one area

Index	A	B	C	The image average
					Former figure	12.327	18.485	10.659	74.2231
Lee filtering	66.511	222.85	69.615	73.7958
					G-MAP filtering	66.511	222.85	69.615	72.7787
The stationary wavelet soft-threshold	84.085	411.04	76.927	66.6424
					NSCT	90.093	354.09	99.258	74.2231
Method of the present invention	131.7823	571.7528	243.0433	74.2085

As seen from Table 1, the equivalent number of the inventive method is all big than existing Lee filtering, G-MAP filtering, stationary wavelet soft-threshold, NSCT method equivalent number, the advantage of the inventive method in denoising is described, though the image average of the inventive method is consistent fully not as the image average and the original image average of NSCT method, but, illustrate that the distortion of backward radiation degree is very little also very near the average of original image.

Table 2 is the SAR image denoising experimental index contrasts of Horse track two area, comprises the contrast of equivalent number and image average; The equivalent number in three white box zones 1 in the table 2 in A, B, the C difference presentation graphs 4 (a), zone 2, zone 3.

The SAR image denoising experimental index contrast of table 2 Horse track two area

Index	A	B	C	The image average
					Former figure	14.692	13.202	9.5038	83.8504
Lee filtering	217.19	122.21	90.753	83.6654
					G-MAP filtering	217.19	122.21	90.753	83.1247
The stationary wavelet soft-threshold	304.56	141.67	108.97	77.7831
					NSCT	210.4	127.77	173.46	83.8504

Method of the present invention

732.8195

233.4959

299.9914

83.8478

As seen from Table 2, the equivalent number of the inventive method is bigger than the equivalent number of existing Lee filtering, G-MAP filtering, stationary wavelet soft-threshold, NSCT method, and the image average illustrates that also very near the average of original image the distortion of backward radiation degree is little.

Table 3 is the SAR image denoising experimental index contrasts of Bedfordshire area, comprises the contrast of equivalent number and image average; The equivalent number in three white box zones 1 in the table 3 in A, B, the C difference presentation graphs 5 (a), zone 2, zone 3.

The SAR image denoising experimental index contrast of table 3 Bedfordshire area

Index	A	B	C	The image average
					Former figure	3.1199	2.937	2.6746	106.8689
Lee filtering	28.153	19.289	38.485	106.3758
					G-MAP filtering	28.153	19.289	38.485	105.547
The stationary wavelet soft-threshold	51.088	26.443	45.795	98.3034
					NSCT	44.792	28.583	71.877	106.8689
Method of the present invention	119.7882	49.7886	91.9971	106.8563

Method of the present invention as seen from Table 3 in trizonal equivalent number all greater than existing Lee filtering, G-MAP filtering, stationary wavelet soft-threshold, NSCT method; Average illustrates that also very near the average of original image the distortion of backward radiation degree is very little.

Claims (3)

1. the SAR image de-noising method based on NSCT territory rim detection and Bishrink model comprises the steps:

(1) the test pattern I that chooses being carried out non-down sampling contourlet NSCT conversion, is 6 straton band coefficients with picture breakdown;

(2) keep the sub-band coefficients of layers 1 and 2 constant;

(3) the 3rd～6 layer sub-band coefficients is shunk with the Bishrink model;

(4) to the sub-band coefficients after shrinking through step (3), carry out the non-down sampling contourlet inverse transformation and obtain reconstructed image R;

(5) reconstructed image is carried out following rim detection and mean filter;

(5a) each pixel in the reconstructed image is got 3 * 3 fields, in this neighborhood, set 12 directions, note P ₁And P ₂Be respectively the mean value of both sides pixel on 12 directional rays, establish p=P ₁/ P ₂, if P≤1, then ratio detects factor r=p, otherwise r=p ^-1, remember that the minimum value of the ratio detection factor on 12 directions is r _MinIf edge detection threshold T ₀: 0.5≤T ₀≤ 1, if r _Min＞T ₀, think that then current point is a marginal point, be labeled as 1, otherwise be labeled as 0; After all element markings are intact, if pixel is labeled as 1, be that 3 * 3 neighborhood window is got at the center with it then, is that 1 number is less than 4 as if this neighborhood window internal labeling, then this pixel is labeled as 0; Realization is to the rim detection of reconstructed image;

(5b) detect the edge of reconstructed image after, (i j) gets its 5 * 5 neighborhood W for the center to be labeled as 0 pixel R with each ₁, if W ₁Internal labeling is that 1 pixel number is less than 4, and then (i, pixel value j) is taken as W to R ₁The mean value of interior all pixels; (i j) is the center, gets its 3 * 3 neighborhood W otherwise with R ₂, if W ₂Internal labeling is that 1 pixel number is less than 4, and then (i, pixel value j) is taken as W to R ₂In the mean value of all pixels, otherwise R (i, pixel value j) is taken as W ₂In all are labeled as the mean value of 0 pixel, finish the mean filter of image after the rim detection, obtain image U after the filtering;

(6) original image I and filtered image U are subtracted each other, and the error image v that obtains is carried out k iteration of Anisotropic Nonlinear diffusion, k=140-230;

(7) Anisotropic Nonlinear is spread the image v that k iteration obtains ^kWith image U addition after the filtering, obtain the image after coherent spot suppresses.

2. the SAR image according to claim 1 speckle suppression method that is concerned with, wherein step (3) is described carries out coefficient with the Bishrink model and shrinks, and carries out as follows:

(2a) establish y ₁Be the coefficient of S layer, y ₂Be the coefficient of S-1 layer, S=3,4,5,6;

(2b) estimate current coefficient y with the prior estimate form ₁Noise variance

${\mathrm{\&sigma;}}_n^2=\frac{\mathrm{median}\left(\right|y_1\left|\right)}{0.6745}$

(2c) method of usefulness local auto-adaptive is estimated the edge of model variances sigma of current each coefficient ²:

${\mathrm{\&sigma;}}^2={({\mathrm{\&sigma;}}_{y_1}^2-{\mathrm{\&sigma;}}_n^2)}_{+},$

Wherein

Expression is worked as The time σ ²=0, when

The time

Wherein N (k) represents the coefficient y of S layer ₁Be the square window at center, S=3,4,5,6, M are the size of window coefficient, and value is 11;

(2d) utilize following formula design factor y ₁Estimated value

${\hat{y}}_1=\frac{{(\sqrt{y_1^2+y_2^2}-\frac{\sqrt{3}{\mathrm{\&sigma;}}_n^2}{\mathrm{\&sigma;}})}_{+}}{\sqrt{y_1^2+y_2^2}}{y}_1,$

In the formula

y ₁Be the coefficient of S layer, y ₂Be the coefficient of S-1 layer, S=3,4,5,6;

Wherein

Expression is worked as

The time its result be zero, when

The time its result remain unchanged.

3. the SAR image according to claim 1 speckle suppression method that is concerned with, wherein step (6) is described carries out k iteration of Anisotropic Nonlinear diffusion to error image, carries out according to following formula:

$v_{i,j}^{k+1}=v_{i,j}^k+\mathrm{\&tau;}\underset{r,s=-1}{\overset{1}{\mathrm{\&Sigma;}}}\frac{g\left({\sqrt{2}}^{1-\left|r\right|-\left|s\right|}\right|v_{i+r,j+s}-v_{i,j}\left|\right)(v_{i+r,j+s}-v_{i,j})}{(r^2+s^2)}$

Wherein v is that original image I and filtered image U subtract each other the error image that obtains, v _{I, j}The i that is v is capable, the pixel of j row, v _{I+r, j+s}The i+r that is v is capable, the pixel of j+s row; R and s are the neighborhood of pixels value, (r, s) ≠ (0,0);

Be to carry out anisotropy to spread error image before, Be to carry out the error image after the anisotropy diffusion one time; τ is a scale parameter,

G is the diffusivity function,

$g\left({\sqrt{2}}^{1-\left|r\right|-\left|s\right|}\right|v_{i+r,j+s}-v_{i,j}\left|\right)=\frac{1}{(1+\frac{{\left({\sqrt{2}}^{1-\left|r\right|-\left|s\right|}\right|v_{i+r,j+s}-v_{i,j}\left|\right)}^2}{{\mathrm{rr}}^2})},$

Wherein rr is a reduced parameter, rr=6.

Images