CN104867120A - Ratio distribution-based sar image non-local despeckling method - Google Patents

Abstract

The invention discloses a ratio distribution-based SAR image non-local despeckling method, which mainly solves the problem that the calculation of a similarity weight value in the existing non-local despeckling method is inaccurate, and a realizing process of the method comprises the steps of: 1, inputting an SAR image and initializing parameters; 2, for each pixel of the input SAR image, calculating ratio distances dt1 of the pixel with neighborhood pixels by ratio distribution; 3, calculating priori distances dt2 between adjacent pixels; 4, calculating a similarity weight value of the pixel point with each neighborhood pixel point according to the distances dt1 and dt2; 5, performing weighted averaging on a central pixel to obtain a central pixel estimation value; 6, repeating steps 2-5, traversing the whole image to obtain an estimation image; 7, repeating steps 3-6 till ending of iteration, thus obtaining a final despeckling image. The similarity measuring accuracy of an image block is improved by an iteration manner through using a multiplicative speckle property of the SAR image and adopting the ratio distances and priori distances, so detail information such as edge and textures are better kept while speckle noise of the SAR image is effectively filtered.

Description

Based on method for reducing speckle more non local than the SAR image of Distribution value

Technical field

The invention belongs to technical field of image processing, relate generally to Technologies Against Synthetic Aperture Radar (Synthetic ApertureRadar, SAR) Speckle Filter of image is a kind of based on method for reducing speckle more non local than the SAR image of Distribution value specifically, can be used for falling spot process to SAR image.

Background technology

Image formed by synthetic-aperture radar 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, SAR image corrode by multiplicative coherent speckle noise, the stepwise derivation that this noise reflects from backscatter radar.This speckle noise has damaged the radiometric resolution of SAR image, has influence on context analyzer and understanding simultaneously.

The target of spot method is gone to be exactly the characteristic information retaining image while removing noise, as texture, edge and point target.But due to the multiplicative coherence of speckle noise, suppress speckle noise very difficult.At present, there is a lot of Speckle Filter methods, be broadly divided into spatial domain and the large class of frequency domain two.The SAR image of spot method as converted based on small echo, Contourlet etc. of going of frequency domain class goes spot algorithm to obtain a wide range of applications due to advantages such as it is multiple dimensioned, many resolutions.But transform domain goes spot algorithm to be still filtering based on stationary window in essence, can produce Gibbs phenomenon in regions such as the edge of image, lines.Its shortcoming is the marginal information that well can not retain image.Spatial domain statistics class goes spot method generally first to suppose the multiplicative model of noise, and the partial statistics characteristic then based on neighborhood of pixels window carries out filtering.Typical algorithm based on airspace filter has Lee filtering, Kuan filtering, Frost filtering and enhanced edition thereof.These methods get average at homogeneous region, and change is taked faster to the strategy retained, its shortcoming is that the noise of perimeter also can be retained, and image structure information such as the targets such as edge, linear body, point can to a certain extent by fuzzy or filtering simultaneously.In recent years, non-local method was used to image filtering.Because non local method can better utilize the redundancy of image, achieve good effect.2009, the people such as Charles propose a kind of based on block of pixels probability distribution (Probability Patch-Based, PPB) method, it is a kind of non local SAR image method for reducing speckle of iteration, is considered to current best SAR image and removes one of spot method.But PPB algorithm is going in spot process to SAR image, still using the similarity between euclidean distance metric block of pixels.Theoretical and experiment all proves, the metric form of this Euclidean distance is improper for the Multiplicative noise model of SAR image, and it can not good restraint speckle.In addition, PPB method is by KL distance as the priori distance between pixel, and evaluated error is very faint relative to the noise of image.KL distance is a kind of representation of two norms, can not well approach faint evaluated error.

Summary of the invention

The object of the invention is to overcome the technical matters that PPB method is not enough on Hemifusus ternatanus when spot falls in SAR image, propose a kind of based on the non local wave filter than Distribution value for the method for reducing speckle of SAR image.

The present invention is a kind of based on method for reducing speckle more non local than the diameter radar image of Distribution value, it is characterized in that, includes following steps:

(1). the SAR image V that an input L looks, and initialization iterations i=1, work as input picture V as initial estimated image

(2). from the SAR image V of input, get its arbitrary pixel v _s, with pixel v _scentered by block of pixels be called center pixel block v _s, with pixel v _sthe pixel v of neighborhood _tcentered by block of pixels be called neighborhood territory pixel block v _t, calculate center pixel block v _swith neighborhood territory pixel block v _tbetween ratio distance d _t1.

Alternatively, the L inputted is looked to the pixel v of SAR image V _s, calculate with pixel v _scentered by block of pixels with v _seach pixel v in neighborhood _tcentered by block of pixels between ratio distance d _t1.

(3). calculate the priori distance d in estimated image between neighbor _t2:

If i=1, from initial estimation image middle capture vegetarian refreshments with pixel centered by block of pixels be called center estimate block with pixel the pixel of neighborhood centered by block of pixels be called neighborhood estimate block block is estimated at the center that calculates with its neighborhood estimation point between priori distance d _t2;

If i>1, from the estimated image that i-1 iteration draws middle capture vegetarian refreshments with pixel centered by block of pixels be called center estimate block with pixel the pixel of neighborhood centered by block of pixels be called neighborhood estimate block block is estimated at the center that calculates with its neighborhood estimation point between priori distance d _t2;

(4). according to ratio distance d _t1with priori distance d _t2calculate each some v in its neighborhood _tthe weight w of i-th iteration _s,t ⁱ;

(5). to the pixel in estimated image after each the some weighted mean in neighborhood

${{\hat{u}}_s}^i=\frac{\underset{t}{\Σ}{\mathrm{\ω}}_{s,t}^i{v}_t}{\underset{t}{\Σ}{w}_{s,t}^i}.$

(6). for each pixel, repeat step (2) ~ (5), the whole SAR image of traversal input obtains the estimated image of i-th iteration

(7). repeat step (3) ~ (6), until iterations i=n, obtain the image after most final decline spot n is given iterations.

The present invention combines the ratio Distribution value of speckle noise under the framework of PPB method, construct an adaptive weight computing formula, to measure the similarity of each point and central point in neighborhood more accurately, it can while suppressing speckle noise preferably, retain the grain details in image, thus improve the effect of falling spot.

Realization of the present invention is also: step (2) calculates center pixel block v _swith neighborhood territory pixel block v _tbetween ratio distance d _t1, include following steps:

(2a). with pixel v _scentered by, choose the Search Area of neighborhood as this pixel of N × N size;

(2b). with pixel v _scentered by, get the block of M × M size, with the gray-scale value matrix v of pixel each in block _srepresent this center pixel block;

(2c). to remove central pixel point v in Search Area _souter each pixel v _tcentered by, get the block of M × M size, with the gray-scale value matrix v of pixel each in block _trepresent this neighborhood territory pixel block;

(2d). calculate above-mentioned two block of pixels v _sand v _tin the ratio of each corresponding pixel points:

$r_{s,k}=min\left\{\frac{v_{s,k}}{v_{t,k}},\frac{v_{t,k}}{v_{s,k}}\right\},r_{s,k}\∈\[0,1\]$

Wherein v _s,kand v _t,krepresent v respectively _sand v _ta kth pixel;

(2e). according to ratio distribution probability formula, calculate two block of pixels v _sand v _tin the ratio r of each corresponding pixel points _s,kthe probability occurred; If input SAR image is intensity image, working strength new probability formula:

$p\left(r_{s,k}\right)=\frac{2\mathrm{\Γ}\left(2L\right)}{\mathrm{\Γ}{\left(L\right)}^2}\frac{{r_{s,k}}^{L-1}}{{(r_{s,k}+1)}^{2L}},$

If input SAR image is magnitude image, then use amplitude new probability formula:

$p\left(r_{s,k}\right)=\frac{4\mathrm{\Γ}\left(2L\right)}{\mathrm{\Γ}{\left(L\right)}^2}\frac{{r_{s,k}}^{2L-1}}{{({r_{s,k}}^2+1)}^{2L}};$

(2f). calculate with pixel v _scentered by center pixel block v _swith with pixel v _tcentered by neighborhood territory pixel block v _tbetween ratio distance d _t1,

$d_{t1}=\underset{k=1}{\overset{M\×M}{\Σ}}log\left(p\left(r_{s,k}\right)\right).$

In PPB method, by the similarity between the euclidean distance metric pixel between block of pixels, this method is not suitable for the multiplicative characteristic of SAR image speckle noise.Therefore, this method is for the multiplicative characteristic of SAR image, and use the ratio distribution probability between pixel to measure its similarity, it is more reasonable to explain in theory.In practice, it can better suppress speckle noise, thus improves the effect of falling spot.

Realization of the present invention is also: step (3) calculates center and estimates block with its neighborhood estimation point between priori distance d _t2, include following steps:

(3a). with estimation point centered by, choose the Search Area of neighborhood as this estimation point of N × N size;

(3b). with estimation point centered by, get the block of M × M size, with the gray-scale value matrix of estimation point each in block represent that block is estimated at this center;

(3c). with Search Area Zhong Chu center estimation point v _seach outer neighborhood estimation point centered by, get the block of M × M size, with the gray-scale value matrix of estimation point each in block represent that this neighborhood estimates block;

(3d). calculate above-mentioned center and estimate block block is estimated with its neighborhood between priori distance:

$d_{t2}=\underset{k=1}{\overset{M\×M}{\Σ}}log\left(min\right\{\frac{{{\hat{u}}_{s,k}}^{i-1}}{{{\hat{u}}_{t,k}}^{i-1}},\frac{{{\hat{u}}_{t,k}}^{i-1}}{{{\hat{u}}_{s,k}}^{i-1}}\left\}\right)$

Wherein with represent respectively with a kth pixel.

In PPB method, by the KL distance estimated between block as the priori distance between pixel, and evaluated error is very faint relative to the noise of image.KL distance is a kind of representation of two norms, and it well can not approach this faint evaluated error.Therefore, this method is for the faint characteristic of evaluated error, and propose a kind of metric form of a norm, it is more reasonable to explain in theory.In practice, the small and weak details in the result after it can make mage retrieval model is retained, thus improves the effect of falling spot.

Realization of the present invention is also: described in step (4) " according to ratio distance d _t1with priori distance d _t2calculate each some v in its neighborhood _tthe weight w of i-th iteration _s,t ⁱ", comprise the steps:

If (4a). be input as strength S AR image, then the number of looking of simulating generation two separate is the noise pattern that the obedience gamma of L distributes, and is designated as the first intensity noise figure R respectively ₁with the second intensity noise figure R ₂, it compares value matrix the ratio matrix r formula of bringing into is below tried to achieve distribution matrix D ₀:

D ₀＝(L-1)logr-2Llog(r+1)；

If be input as amplitude SAR image, then the number of looking of simulating generation two separate is the noise pattern that the obedience Nakagami of L distributes, and is designated as the first amplitude noise figure A respectively ₁with the second amplitude noise figure A ₂, compare value matrix the ratio matrix r formula of bringing into is below tried to achieve distribution matrix D ₀:

D ₀＝(2L-1)logr-2Llog(r ²+1)；

(4b). get distribution matrix D ₀in be distributed in the element q that quantile is α (in the method α=0.92) place, D ₀average be m, be calculated as follows smoothing parameter h:

h＝q-m；

If (4c). be input as strength S AR image, then the weight w of i-th iteration _s,t ⁱfor:

${w_{s,t}}^i=\exp \[\underset{k}{\Σ}(\frac{1}{h}\left(\right(L-1){\mathrm{logr}}_{s,k}-2Llog(r_{s,k}+1))+\frac{1}{T}log\left(min\right\{\frac{{{\hat{u}}_s}^{i-1}}{{{\hat{u}}_t}^{i-1}},\frac{{{\hat{u}}_t}^{i-1}}{{{\hat{u}}_s}^{i-1}}\left\}\right))\]$

If be input as amplitude SAR image, then the weight w of i-th iteration _s,t ⁱfor:

${w_{s,t}}^i=\exp \[\underset{k}{\Σ}(\frac{1}{h}\left(\right(2L-1){\mathrm{logr}}_{s,k}-2Llog({r_{s,k}}^2+1))+\frac{1}{T}log\left(min\right\{\frac{{{\hat{u}}_s}^{i-1}}{{{\hat{u}}_t}^{i-1}},\frac{{{\hat{u}}_t}^{i-1}}{{{\hat{u}}_s}^{i-1}}\left\}\right))\]$

Wherein T is local auto-adaptive parameter.

In the process of image denoising, while restraint speckle, also details will be kept.And restraint speckle is often to sacrifice details for cost, therefore, weighing apparatus of how making even between squelch and Hemifusus ternatanus is a Focal point and difficult point of image denoising.In the present invention, ratio distance is used for restraint speckle, and priori distance is for keeping details.And the two is joined together by ratio Distance geometry priori distance to employ two adaptive parameters, arranging of two parameters is all relevant with the equivalent number L of SAR image.It can be applicable to the SAR image of different noise intensity.

The present invention compared with prior art has following advantage:

1. the ratio Distribution value of nexus spot noise of the present invention, constructs an adaptive weight computing formula.Compared to the Euclidean distance used in PPB method, the ratio distance used in the present invention can better be applicable to the Multiplicative noise model of SAR image.Therefore, can better restraint speckle.

2. the present invention proposes a kind of new priori distance for measuring the evaluated error in spot process.Compared to the KL distance used in PPB, the priori distance based on a norm used in the present invention can better measure the faint evaluated error produced in spot process.Therefore, detailed information can better be kept.

3. the present invention uses adaptive weight computing formula, does not need to arrange different parameters to different images, is easy to realize.

Accompanying drawing explanation

Fig. 1 of the present inventionly falls spot process flow block diagram;

Fig. 2 is the primary standard image that the present invention tests the synthesis SAR image of input;

Fig. 3 is the result figure that the synthesis SAR image to input removes spot by existing PPB method and the inventive method, 3 row from left to right represent that (a) synthesizes noisy SAR image respectively, spot result figure is removed after (b) PPB method and the filtering of (c) the inventive method, be respectively equivalent number L=1 from top to bottom, 2,4, experimental result when 16;

The partial enlarged drawing of experimental result when Fig. 4 is equivalent number L=1 in Fig. 3.Wherein, Fig. 4 (a) and Fig. 4 (b) is the partial enlarged drawing of Fig. 3 (b1) and Fig. 3 (c1) respectively;

The partial enlarged drawing of experimental result when Fig. 5 is equivalent number L=4 in Fig. 3.Wherein, Fig. 5 (a) and Fig. 5 (b) is the partial enlarged drawing of Fig. 3 (b3) and Fig. 3 (c3) respectively;

The partial enlarged drawing of experimental result when Fig. 6 is equivalent number L=16 in Fig. 3.Wherein, Fig. 6 (a) and Fig. 6 (b) is the partial enlarged drawing of Fig. 3 (b4) and Fig. 3 (c4) respectively.

Embodiment

Below in conjunction with accompanying drawing to the detailed description of the invention:

Embodiment 1

In the imaging process of SAR, inevitably there is the interference of speckle noise.This speckle noise can cause serious impact to the understanding of SAR image and decipher.Therefore, before to SAR image process, the pre-service of falling spot usually to be carried out.At present, the spot that falls for SAR image has had a lot of methods, and traditional filtering method is as Lee filtering, Kuan filtering, Frost filtering etc. is generally get average to the homogeneous region of local, and change is taked faster to the strategy retained, its shortcoming is that the noise of perimeter also can be retained.In order to overcome these shortcomings, the people such as Charles proposed a kind of method of PPB in 2009, were considered to current best SAR image and removed one of spot method.But PPB algorithm, in the Euclidean distance of going to use in spot process to SAR image, well can not be applicable to the Multiplicative noise model of SAR image.In addition, the KL distance used in PPB algorithm is also accurate not to the tolerance of the evaluated error in denoising process, and detailed information can be made to lose.

The present invention proposes a kind of based on method for reducing speckle more non local than the SAR image of Distribution value, see Fig. 1, includes following steps:

(1). the SAR image V that looks of an input L, the SAR image that this example inputs can see Fig. 3 (a1), and the noise image in Fig. 3 (a1) is on the noise-free picture of Fig. 2, added the Nakagami speckle noise that L=1 looks.And initialization iterations i=1, input picture V is worked as initial estimation image

(2). from the SAR image V of input, get its arbitrary pixel v _s, with pixel v _scentered by block of pixels be called center pixel block v _s, with pixel v _sthe pixel v of neighborhood _tcentered by block of pixels be called neighborhood territory pixel block v _t, calculate center pixel block v _swith neighborhood territory pixel block v _tbetween ratio distance d _t1.

That is, the L=1 inputted is looked to the pixel v of SAR image V _s, with pixel v _scentered by block of pixels with v _seach pixel v in neighborhood _tcentered by block of pixels between ratio distance d _t1, namely with pixel v _scentered by block of pixels with v _seach pixel v in neighborhood _tcentered by block of pixels between ratio distance d _t1.

(3). calculate the priori distance d in estimated image between neighbor _t2:

If iterations i=1, from initial estimation image middle capture vegetarian refreshments with pixel centered by block of pixels be called center estimate block with pixel the pixel of neighborhood centered by block of pixels be called neighborhood estimate block block is estimated at the center that calculates with its neighborhood estimation point between priori distance d _t2;

If iterations i>1, from the estimated image that i-1 iteration draws middle capture vegetarian refreshments with pixel centered by block of pixels be called center estimate block with pixel the pixel of neighborhood centered by block of pixels be called neighborhood estimate block block is estimated at the center that calculates with its neighborhood estimation point between priori distance d _t2;

(4). according to ratio distance d _t1with priori distance d _t2calculate each some v in its neighborhood _tthe weight w of i-th iteration _s,t ⁱ.

(5). to the pixel in estimated image after each the some weighted mean in neighborhood

${{\hat{u}}_s}^i=\frac{\underset{t}{\Σ}{\mathrm{\ω}}_{s,t}^i{v}_t}{\underset{t}{\Σ}{w}_{s,t}^i}.$

(6). for each pixel, repeat step (2) ~ (5), the whole SAR image of traversal input obtains estimated image

(7) if. iterations i < n, n is given iteration ends number of times, then repeat step (3) ~ (6), by i+1 give i, enter next iteration; If i=n, then iteration ends, and the estimated image exporting i-th=n time iteration as final, spot image falls.Experiment shows, when iterating to the 7th time, falling spot result and tending towards stability, therefore n=7 in this example, obtains the image after most final decline spot

In order to solve the technical matters existing for PPB method, one aspect of the present invention employs ratio distance to replace Euclidean distance, can better be applicable to the Multiplicative noise model of SAR image; On the other hand for the faint characteristic of evaluated error, propose a kind of metric form of a norm, it is more reasonable to explain in theory.In practice, the small and weak details in the result after it can make mage retrieval model is retained, thus improves the effect of falling spot.

Embodiment 2

Based on than the non local method for reducing speckle of the SAR image of Distribution value with embodiment 1, wherein, step (2) calculates center pixel block v _swith neighborhood territory pixel block v _tbetween ratio distance d _t1, include following steps:

(2a). with pixel v _scentered by, choose the Search Area of neighborhood as this pixel of N × N size, get N=21 in this example;

(2b). with pixel v _scentered by, get the block of M × M size, with the gray-scale value matrix v of pixel each in block _srepresent this center pixel block, M=7 in this example;

(2c). to remove central pixel point v in Search Area _souter each pixel v _tcentered by, get the block of M × M size, with the gray-scale value matrix v of pixel each in block _trepresent this neighborhood territory pixel block;

(2d). calculate above-mentioned two block of pixels v _sand v _tin the ratio of each corresponding pixel points:

$r_{s,k}=min\left\{\frac{v_{s,k}}{v_{t,k}},\frac{v_{t,k}}{v_{s,k}}\right\},r_{s,k}\&Element;\&lsqb;0,1\&rsqb;$

Wherein v _s,kand v _t,krepresent v respectively _sand v _ta kth pixel;

(2e). according to ratio distribution probability formula, calculate two block of pixels v _sand v _tin the ratio r of each corresponding pixel points _s,kthe probability occurred; This example input SAR image is magnitude image, then use amplitude new probability formula:

$p\left(r_{s,k}\right)=\frac{4\mathrm{\&Gamma;}\left(2L\right)}{\mathrm{\&Gamma;}{\left(L\right)}^2}\frac{{r_{s,k}}^{2L-1}}{{({r_{s,k}}^2+1)}^{2L}};$

(2f). calculate center pixel block v _swith neighborhood territory pixel block v _tbetween ratio distance d _t1, that is, calculate with pixel v _scentered by center pixel block v _swith with pixel v _tcentered by neighborhood territory pixel block v _tbetween ratio distance d _t1:

$d_{t1}=\underset{k=1}{\overset{M\&times;M}{\&Sigma;}}log\left(p\left(r_{s,k}\right)\right).$

Be directed to the Multiplicative noise model of SAR image, under the hypothesis of homogeneous region, the ratio between adjacent pixel blocks is only relevant with noise profile, and has nothing to do with noise-free picture.Therefore, compared to the Euclidean distance used in traditional algorithm, in the present invention, utilize its similarity of ratio distance metric between adjacent pixel blocks can better be applicable to the Multiplicative noise model of SAR image.In practice, it can better suppress the speckle noise of SAR image.

Embodiment 3

Based on than the non local method for reducing speckle of the SAR image of Distribution value with embodiment 1-2, if in this example in step (2) input SAR image be intensity image, then working strength new probability formula in step (2e):

$p\left(r_{s,k}\right)=\frac{2\mathrm{\&Gamma;}\left(2L\right)}{\mathrm{\&Gamma;}{\left(L\right)}^2}\frac{{r_{s,k}}^{L-1}}{{(r_{s,k}+1)}^{2L}},$

Calculate center pixel block v _swith neighborhood territory pixel block v _tbetween ratio distance d _t1, that is, calculate with pixel v _scentered by center pixel block v _swith with pixel v _tcentered by neighborhood territory pixel block v _tbetween ratio distance d _t1,

$d_{t1}=\underset{k=1}{\overset{M\&times;M}{\&Sigma;}}log\left(p\left(r_{s,k}\right)\right).$

Embodiment 4

Based on than the non local method for reducing speckle of the SAR image of Distribution value with embodiment 1-3, wherein, step (3) calculates center and estimates block with its neighborhood estimation point between priori distance d _t2, include following steps:

(3a). with estimation point centered by, choose the Search Area of neighborhood as this estimation point of N × N size, N=21 in this example;

(3b). with estimation point centered by, get the block of M × M size, with the gray-scale value matrix of estimation point each in block represent that block, in this example M=7 are estimated in this center;

(3c). with Search Area Zhong Chu center estimation point v _seach outer neighborhood estimation point centered by, get the block of M × M size, with the gray-scale value matrix of estimation point each in block represent that this neighborhood estimates block;

(3d). calculate above-mentioned center and estimate block block is estimated with its neighborhood between priori distance:

$d_{t2}=\underset{k=1}{\overset{M\&times;M}{\&Sigma;}}log\left(min\right\{\frac{{{\hat{u}}_{s,k}}^{i-1}}{{{\hat{u}}_{t,k}}^{i-1}},\frac{{{\hat{u}}_{t,k}}^{i-1}}{{{\hat{u}}_{s,k}}^{i-1}}\left\}\right)$

Wherein with represent respectively with a kth pixel.

To in the denoising process of SAR image, there is evaluated error all the time.In order to obtain better denoising result, the present invention carries out the calculating of iteration to this evaluated error, and each iteration all can reduce evaluated error, until evaluated error no longer changes.

Compared to initial noise, evaluated error is fainter.Therefore, be employed herein a kind of metric form of a norm, it is more reasonable to explain in theory.In practice, the small and weak details in the result after it can make mage retrieval model is retained, thus improves the effect of falling spot.

Embodiment 5

Based on than the non local method for reducing speckle of the SAR image of Distribution value with embodiment 1-4, wherein, described in step (4) " according to ratio distance d _t1with priori distance d _t2calculate each some v in its neighborhood _tthe weight w of i-th iteration _s,t ⁱ", comprise the steps:

If (4a). be input as strength S AR image, then the number of looking of simulating generation two separate is the noise pattern that the obedience Gamma of L distributes, and is designated as the first intensity noise figure R respectively ₁with the second intensity noise figure R ₂, it compares value matrix the ratio matrix r formula of bringing into is below tried to achieve distribution matrix D ₀:

D ₀＝(L-1)logr-2Llog(r+1)；

If be input as amplitude SAR image, then the number of looking of simulating generation two separate is the noise pattern that the obedience Nakagami of L distributes, and L=1 in this example, is designated as the first amplitude noise figure A respectively ₁with the second amplitude noise figure A ₂, compare value matrix the ratio matrix r formula of bringing into is below tried to achieve distribution matrix D ₀:

D ₀＝(2L-1)logr-2Llog(r ²+1)。

(4b). get distribution matrix D ₀in be distributed in the element q that quantile is α (in this example α=0.92) place, D ₀average be m, be calculated as follows smoothing parameter h:

h＝q-m；

If (4c). be input as strength S AR image, then the weight w of i-th iteration _s,t ⁱfor:

${w_{s,t}}^i=\exp \&lsqb;\underset{k}{\&Sigma;}(\frac{1}{h}\left(\right(L-1){\mathrm{logr}}_{s,k}-2Llog(r_{s,k}+1))+\frac{1}{T}log\left(min\right\{\frac{{{\hat{u}}_s}^{i-1}}{{{\hat{u}}_t}^{i-1}},\frac{{{\hat{u}}_t}^{i-1}}{{{\hat{u}}_s}^{i-1}}\left\}\right))\&rsqb;$

If be input as amplitude SAR image, then the weight w of i-th iteration _s,t ⁱfor:

${w_{s,t}}^i=\exp \&lsqb;\underset{k}{\&Sigma;}(\frac{1}{h}\left(\right(2L-1){\mathrm{logr}}_{s,k}-2Llog({r_{s,k}}^2+1))+\frac{1}{T}log\left(min\right\{\frac{{{\hat{u}}_s}^{i-1}}{{{\hat{u}}_t}^{i-1}},\frac{{{\hat{u}}_t}^{i-1}}{{{\hat{u}}_s}^{i-1}}\left\}\right))\&rsqb;$

Wherein T is local auto-adaptive parameter, and its span is (0 ,+∞), gets T=0.1 in this example.

In the process of image denoising, weighing apparatus of how making even between squelch and Hemifusus ternatanus is technological difficulties.In the present invention, ratio distance is used for restraint speckle, and the two, for keeping details, is joined together by two adaptive parameters by priori distance.Arranging of two parameters is all relevant with the equivalent number L of SAR image, therefore, and the adaptive SAR image being applicable to various different noise intensity of energy is in use very convenient, quick.

Embodiment 6

Based on than the non local method for reducing speckle of the SAR image of Distribution value with embodiment 1-5,

With reference to accompanying drawing 1, the present invention can also explain with the realization of following steps on the whole again:

(1). the L=2 of input looks the pixel v of SAR image V _s, this example can see Fig. 3 (a2), and the noise image in Fig. 3 (a1) is on the noise-free picture of Fig. 2, added the Nakagami speckle noise that L=2 looks, and initialization iterations i=1, using input picture V as initial estimated image

(2). for the SAR image V of input, calculate each pixel v with its neighborhood as follows _tblock ratio distance d _t1:

(2a). with pixel v _scentered by, choose the Search Area of neighborhood as this pixel of N × N size, N=21 in this example;

(2b). with pixel v _scentered by, get the block of M × M size, in block, the gray-scale value of each pixel is designated as matrix v _s, M=7 in this example;

(2c). to remove central pixel point v in Search Area _souter each pixel v _tcentered by, get the block of M × M size, in block, the gray-scale value of each pixel is designated as matrix and is designated as v _t;

(2d). calculate above-mentioned two block of pixels v _sand v _tratio.Wherein v _s,kand v _t,krepresent v respectively _sand v _ta kth pixel:

$r_{s,k}=min\left\{\frac{v_{s,k}}{v_{t,k}},\frac{v_{t,k}}{v_{s,k}}\right\},r_{s,k}\&Element;\&lsqb;0,1\&rsqb;.$

(2e). according to ratio distribution probability formula, calculate the ratio r of each point _s,kthe probability occurred.If input SAR image is intensity image, probability of use formula:

$p\left(r_{s,k}\right)=\frac{2\mathrm{\&Gamma;}\left(2L\right)}{\mathrm{\&Gamma;}{\left(L\right)}^2}\frac{{r_{s,k}}^{L-1}}{{(r_{s,k}+1)}^{2L}};$

If input SAR image is magnitude image, then probability of use formula:

$p\left(r_{s,k}\right)=\frac{4\mathrm{\&Gamma;}\left(2L\right)}{\mathrm{\&Gamma;}{\left(L\right)}^2}\frac{{r_{s,k}}^{2L-1}}{{({r_{s,k}}^2+1)}^{2L}};$

(2f). calculate central pixel point v _swith the pixel v in its neighborhood _tbetween block ratio distance d _t1.

$d_{t1}=\underset{k=1}{\overset{M\&times;M}{\&Sigma;}}log\left(p\left(r_{s,k}\right)\right).$

(3) if. i=1, from initial estimation image middle capture vegetarian refreshments with pixel centered by block of pixels be called center estimate block with pixel the pixel of neighborhood centered by block of pixels be called neighborhood estimate block block is estimated at the center that calculates with its neighborhood estimation point between priori distance d _t2:

$d_{t2}=log\left(min\right\{\frac{{{\hat{u}}_s}^0}{{{\hat{u}}_t}^0},\frac{{{\hat{u}}_t}^0}{{{\hat{u}}_s}^0}\left\}\right)$

Wherein, with represent estimated image respectively s pixel and s neighborhood in t pixel;

If i>1, from the estimated image that i-1 iteration draws middle capture vegetarian refreshments with pixel centered by block of pixels be called center estimate block with pixel the pixel of neighborhood centered by block of pixels be called neighborhood estimate block block is estimated at the center that calculates with its neighborhood estimation point between priori distance d _t2:

$d_{t2}=log\left(min\right\{\frac{{{\hat{u}}_s}^{i-1}}{{{\hat{u}}_t}^{i-1}},\frac{{{\hat{u}}_t}^{i-1}}{{{\hat{u}}_s}^{i-1}}\left\}\right)$

Wherein, with represent estimated image respectively s pixel and s neighborhood in t pixel.

(4). according to distance d _t1and d _t2calculate each some v in its neighborhood _tthe weight w of i-th iteration _s,t ⁱ.

If (4a). be input as strength S AR image, then simulate generation two and independently look the gamma noise pattern R of number for L ₁and R ₂, compare value matrix the ratio matrix r formula of bringing into is below tried to achieve distribution matrix D ₀:

D ₀＝(L-1)logr-2Llog(r+1)

If be input as amplitude SAR image, then simulate generation two and independently look the Nakagami noise pattern A of number for L ₁and A ₂, L=1 in this example, compares value matrix the ratio matrix r formula of bringing into is below tried to achieve distribution matrix D ₀:

D ₀＝(2L-1)logr-2Llog(r ²+1)；

(4b). get matrix D ₀in to be distributed in quantile be the element at α place is q, D ₀average be m, then smoothing parameter h=q-m, in the method α=0.92.

If (4c). be input as strength S AR image, then

${w_{s,t}}^i=\exp \&lsqb;\underset{k}{\&Sigma;}(\frac{1}{h}\left(\right(L-1){\mathrm{logr}}_{s,k}-2Llog(r_{s,k}+1))+\frac{1}{T}log\left(min\right\{\frac{{{\hat{u}}_s}^{i-1}}{{{\hat{u}}_t}^{i-1}},\frac{{{\hat{u}}_t}^{i-1}}{{{\hat{u}}_s}^{i-1}}\left\}\right))\&rsqb;$

If be input as amplitude SAR image, then

${w_{s,t}}^i=\exp \&lsqb;\underset{k}{\&Sigma;}(\frac{1}{h}\left(\right(2L-1){\mathrm{logr}}_{s,k}-2Llog({r_{s,k}}^2+1))+\frac{1}{T}log\left(min\right\{\frac{{{\hat{u}}_s}^{i-1}}{{{\hat{u}}_t}^{i-1}},\frac{{{\hat{u}}_t}^{i-1}}{{{\hat{u}}_s}^{i-1}}\left\}\right))\&rsqb;$

Here h represents smoothing parameter, and T is local auto-adaptive parameter, in this example T=0.1.

(5). estimate after each the some weighted mean in neighborhood

${{\hat{u}}_s}^i=\frac{\underset{t}{\&Sigma;}{\mathrm{\&omega;}}_{s,t}^i{v}_t}{\underset{t}{\&Sigma;}{w}_{s,t}^i};$

(6). for each pixel, repeat step (2) ~ (5), obtain estimated image

(7). repeat step (3) ~ (6), until iterations i=n.N=7 in this example, obtains the image after falling spot see accompanying drawing 3 (c2).Contrast noise image Fig. 3 (a2) is visible, and speckle noise is obviously suppressed, simultaneously detailed information, as window eaves etc. are all kept preferably.

Accompanying drawing 3 (b2) is that PPB method falls the image after spot to accompanying drawing 3 (a2), contrast accompanying drawing 3 (b2) and (c2), can find the edge such as eaves and window clear and natural more in the result of the inventive method, the squelch of the smooth region such as sky and metope is also better.

Embodiment 7

Based on than the non local method for reducing speckle of the SAR image of Distribution value with embodiment 1-6,

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

One, experiment condition and content

Experiment condition: as shown in Figure 2, Fig. 2 is the primary standard image of the synthesis SAR image of experiment input to the image that experiment uses, noiseless in figure.In order to verify the effect of denoising, add difference in fig. 2 and look several Nakagami noise, as shown in Fig. 3 (a1)-(a4), wherein, Fig. 3 (a1) is for looking the synthesis SAR image of several L=1, Fig. 3 (a2) is for looking the synthesis SAR image of several L=2, and Fig. 3 (a3) is for looking the synthesis SAR image of several L=4, and Fig. 3 (a4) is for looking the synthesis SAR image of several L=16.Fig. 3 (b1)-(b4) is with the filtered spot image that goes of PPB method, Fig. 3 (c1)-(c4) the inventive method is filtered removes spot image, wherein, Fig. 3 (b1) is to Fig. 3 (a1) filtered result by PPB method, Fig. 3 (c1) is that the inventive method is to Fig. 3 (a1) filtered result, the like, Fig. 3 (b2) and Fig. 3 (c2) uses PPB method and the inventive method to result after Fig. 3 (a2) filtering, Fig. 3 (b3) and Fig. 3 (c3) uses PPB method and the inventive method to result after Fig. 3 (a3) filtering, Fig. 3 (b4) and Fig. 3 (c4) uses PPB method and the inventive method to result after Fig. 3 (a4) filtering.

In experiment, various filtering method is all use MATLAB Programming with Pascal Language to realize.

Experiment content:

Under these experimental conditions, use PPB method and the present invention to test respectively, wherein block of pixels size gets 7 × 7, and search window size is 21 × 21.In order to obtain better result, in two kinds of methods, search window and block of pixels are all become large gradually in an iterative process.In PPB method, quantile α gets 0.92, T=0.2, iteration 25 times.And quantile α gets 0.92, T=0.1 in the method, iteration 7 times altogether, n=7.

Synthesis SAR image is fallen to for the experiment of spot, its experimental result is as Fig. 3, Fig. 4, shown in Fig. 5 and Fig. 6, wherein Fig. 4 (a) and Fig. 4 (b) is respectively at noise depending on PPB method during several L=1 and the partial enlarged drawing falling spot result of the present invention, Fig. 5 (a) and Fig. 5 (b) is respectively at noise depending on PPB method during several L=4 and the partial enlarged drawing falling spot result of the present invention, and Fig. 6 (a) and Fig. 6 (b) is respectively at noise depending on PPB method during several L=16 and the partial enlarged drawing falling spot result of the present invention.

Two, experimental result

As can be seen from Figure 3, compared with existing PPB method for reducing speckle, use the inventive method to suppress speckle noise, it is more level and smooth that the homogeneous region such as metope, sky as middle in Fig. 3 (c1) compares Fig. 3 (b1).Meanwhile, the inventive method also can better retain detailed information, as the eaves in the window in Fig. 3 (c1), Fig. 3 (c3) can well retain.

Fig. 4 (a) and Fig. 4 (b) is the partial enlarged drawing of Fig. 3 (b1) and Fig. 3 (c1) respectively, therefrom can find out, the result figure of PPB method process, namely in Fig. 4 (a), window profile is unintelligible, by contrast, the inventive method maintains window profile preferably.

Fig. 5 (a) and Fig. 5 (b) is the partial enlarged drawing of Fig. 3 (b3) and Fig. 3 (c3) respectively, therefrom can find out, the result figure of PPB method process, namely in Fig. 5 (a), eaves have artificial halo effect, and the present invention obtains falls spot result edge details clear and natural more, at eaves and fringe region all closer to former figure.

Similar with the result of Fig. 5, Fig. 6 (a) and Fig. 6 (b) is the partial enlarged drawing of Fig. 3 (b4) and Fig. 3 (c4) respectively, when noise is smaller, the spot result of going of PPB method can produce obvious distortion at edge, and the inventive method can well overcome this phenomenon.

Table 1 gives Y-PSNR (PSNR) value of falling spot experimental result of PPB method and the inventive method in Fig. 3.Also can find out from table 1, the PSNR value of falling spot result that spot effect obtains compared with PPB method of falling of the inventive method all increases under several various.

Table 1 makes PSNR value (dB) contrast of differently falling spot result

Noisy image looks number	PPB method	The inventive method
			L＝1	25.39	26.08
L＝2	27.86	27.95
			L＝4	29.31	29.88
L＝16	32.21	33.37

Above experimental result shows, the present invention has better performance relative to existing PPB method for reducing speckle, overcomes existing PPB filtering method to the not accurate enough shortcoming of weights estimation.Edge and the grain details of SAR image can not only be kept by better smooth speckle noise simultaneously.

To sum up, a kind of iteration disclosed by the invention based on method for reducing speckle more non local than the SAR image of Distribution value, belong to technical field of image processing, mainly solve the inaccurate problem of similarity weight computing in existing non local method for reducing speckle, implementation procedure is: 1. the SAR image of input, and initialization iterations, using input picture as initial estimated image; 2. each pixel of the SAR image of pair input, with calculating the ratio distance d with its neighborhood territory pixel than Distribution value _t1; 3. simultaneously corresponding, calculate the priori distance d between neighbor _t2; 4. according to distance d _t1and d _t2, calculate the similarity weights with each pixel of its neighborhood; 5., according to similarity weights, to center pixel weighted mean, obtain the estimated value of center pixel; 6. repeat step 2 ~ 5, travel through whole image, obtain estimated image; 7. repeat step 3 ~ 6, until iteration ends, obtain final falling spot image.The present invention utilizes the multiplicative speckle characteristics of SAR image, adopt ratio Distance geometry priori distance, and improved the accuracy of the similarity measurement of image block by the mode of iteration, thus while effective filtering SAR image speckle noise, maintain the detailed information such as edge, texture preferably.

Claims (4)

1., based on a method for reducing speckle more non local than the SAR image of Distribution value, it is characterized in that, include following steps:

(1). the SAR image V that an input L looks, makes its initialization iterations i=1, using input picture V as initial estimation image

(2). from the SAR image V of input, get its arbitrary pixel v _s, with pixel v _scentered by block of pixels be called center pixel block v _s, with pixel v _sthe pixel v of neighborhood _tcentered by block of pixels be called neighborhood territory pixel block v _t, calculate center pixel block v _swith neighborhood territory pixel block v _tbetween ratio distance d _t1;

(3). calculate the priori distance d between neighbor in estimated image _t2:

If i=1, from initial estimation image middle capture vegetarian refreshments with pixel centered by block of pixels be called center estimate block with pixel the pixel of neighborhood centered by block of pixels be called neighborhood estimate block block is estimated at the center that calculates with its neighborhood estimation point between priori distance d _t2;

If i>1, from the estimated image that i-1 iteration draws middle capture vegetarian refreshments with pixel centered by block of pixels be called center estimate block with pixel the pixel of neighborhood centered by block of pixels be called neighborhood estimate block block is estimated at the center that calculates with its neighborhood estimation point between priori distance d _t2;

(4). according to ratio distance d _t1with priori distance d _t2calculate each some v in its neighborhood _tthe weight w of i-th iteration _s,t ⁱ;

(5). to the pixel in estimated image after each the some weighted mean in neighborhood

${{\hat{u}}_s}^i=\frac{\underset{t}{\&Sigma;}{w}_{s,t}^i{v}_t}{\underset{t}{\&Sigma;}{w}_{s,t}^i}$

(6). for each pixel in the whole SAR image of input, repeat step (2) ~ (5), travel through the estimated image that whole SAR image obtains i-th iteration

(7) if. iterations i<n (n is given iteration ends number of times), then repeat step (3) ~ (6), enter next iteration; If i=n, then iteration ends, and the estimated image exporting i-th=n time iteration as final, spot image falls.

2. according to claim 1ly it is characterized in that based on method for reducing speckle more non local than the SAR image of Distribution value, step (2) calculates center pixel block v _swith neighborhood territory pixel block v _tbetween ratio distance d _t1, include following steps:

(2a). with pixel v _scentered by, choose the Search Area of neighborhood as this pixel of N × N size;

(2b). with pixel v _scentered by, get the block of M × M size, with the gray-scale value matrix v of pixel each in block _srepresent this center pixel block;

(2c). to remove central pixel point v in Search Area _souter each pixel v _tcentered by, get the block of M × M size, with the gray-scale value matrix v of pixel each in block _trepresent this neighborhood territory pixel block;

(2d). calculate above-mentioned two block of pixels v _sand v _tin the ratio of each corresponding pixel points:

$r_{s,k}=min\left\{\frac{v_{s,k}}{v_{t,k}},\frac{v_{t,k}}{v_{s,k}}\right\},r_{s,k}\&Element;\&lsqb;0,1\&rsqb;$

Wherein v _s,kand v _t,krepresent v respectively _sand v _ta kth pixel;

(2e). according to ratio distribution probability formula, calculate two block of pixels v _sand v _tin the ratio r of each corresponding pixel points _s,kthe probability occurred, if input SAR image is intensity image, working strength new probability formula:

$p\left(r_{s,k}\right)=\frac{2\mathrm{\&Gamma;}\left(2L\right)}{\mathrm{\&Gamma;}{\left(L\right)}^2}\frac{{r_{s,k}}^{L-1}}{{(r_{s,k}+1)}^{2L}},$

If input SAR image is magnitude image, then use amplitude new probability formula:

$p\left(r_{s,k}\right)=\frac{4\mathrm{\&Gamma;}\left(2L\right)}{\mathrm{\&Gamma;}{\left(L\right)}^2}\frac{{r_{s,k}}^{2L-1}}{{({r_{s,k}}^2+1)}^{2L}},$

(2f). calculate center pixel block v _swith neighborhood territory pixel block v _tbetween ratio distance d _t1,

$d_{t1}=\underset{k=1}{\overset{M\&times;M}{\&Sigma;}}log\left(p\right(r_{s,k}\left)\right).$

3. according to claim 1ly it is characterized in that based on method for reducing speckle more non local than the SAR image of Distribution value, step (3) calculates center and estimates block with its neighborhood estimation point between priori distance d _t2, include following steps:

(3a). with estimation point centered by, choose the Search Area of neighborhood as this estimation point of N × N size;

(3b). with estimation point centered by, get the block of M × M size, with the gray-scale value matrix of estimation point each in block represent that block is estimated at this center;

(3c). with Search Area Zhong Chu center estimation point v _seach outer neighborhood estimation point centered by, get the block of M × M size, with the gray-scale value matrix of estimation point each in block represent that this neighborhood estimates block;

(3d). calculate above-mentioned center and estimate block block is estimated with its neighborhood between priori distance:

$d_{t2}=\underset{k=1}{\overset{M\&times;M}{\&Sigma;}}log\left(min\right\{\frac{{{\hat{u}}_{s,k}}^{i-1}}{{{\hat{u}}_{t,k}}^{i-1}},\frac{{{\hat{u}}_{t,k}}^{i-1}}{{{\hat{u}}_{s,k}}^{i-1}}\left\}\right)$

Wherein with represent respectively with a kth pixel.

4. according to claim 1 based on method for reducing speckle more non local than the SAR image of Distribution value, it is characterized in that described in step (4) " according to ratio distance d _t1with priori distance d _t2calculate each some v in its neighborhood _tthe weight w of i-th iteration _s,t ⁱ", comprise the steps:

If (4a). be input as strength S AR image, then the number of looking of simulating generation two separate is the noise pattern that the obedience gamma of L distributes, and is designated as the first intensity noise figure R respectively ₁with the second intensity noise figure R ₂, it compares value matrix the ratio matrix r formula of bringing into is below tried to achieve distribution matrix D ₀:

D ₀＝(L-1)log r-2L log(r+1)

If be input as amplitude SAR image, then the number of looking of simulating generation two separate is the noise pattern that the obedience Nakagami of L distributes, and is designated as the first amplitude noise figure A respectively ₁with the second amplitude noise figure A ₂, compare value matrix the ratio matrix r formula of bringing into is below tried to achieve distribution matrix D ₀:

D ₀＝(2L-1)log r-2L log(r ²+1)

(4b). get distribution matrix D ₀in be distributed in the element q that quantile is α (in the method α=0.92) place, D ₀average be m, be calculated as follows smoothing parameter h:

h＝q-m；

If (4c). be input as strength S AR image, then the weight w of i-th iteration _s,t ⁱfor:

${w_{s,t}}^i=\exp \&lsqb;\underset{k}{\&Sigma;}\left(\frac{1}{h}\right(\left(L-1\right){\mathrm{logr}}_{s,k}-2Llog\left(r_{s,k}+1\right))+\frac{1}{T}log\left(min\right\{\frac{{{\hat{u}}_s}^{i-1}}{{{\hat{u}}_t}^{i-1}},\frac{{{\hat{u}}_t}^{i-1}}{{{\hat{u}}_s}^{i-1}}\left\}\right))\&rsqb;$

If be input as amplitude SAR image, then the weight w of i-th iteration _s,t ⁱfor:

${w_{s,t}}^i=\exp \&lsqb;\underset{k}{\&Sigma;}\left(\frac{1}{h}\right(\left(2L-1\right){\mathrm{logr}}_{s,k}-2Llog\left({r_{s,k}}^2+1\right))+\frac{1}{T}log\left(min\right\{\frac{{{\hat{u}}_s}^{i-1}}{{{\hat{u}}_t}^{i-1}},\frac{{{\hat{u}}_t}^{i-1}}{{{\hat{u}}_s}^{i-1}}\left\}\right))\&rsqb;$

Wherein T is local auto-adaptive parameter.