CN102722879A - SAR (synthetic aperture radar) image despeckle method based on target extraction and three-dimensional block matching denoising - Google Patents

Abstract

The invention discloses an SAR (synthetic aperture radar) despeckle method based on target extraction and three-dimensional block matching denoising, and mainly solves the problem that despeckle effect and detail keeping cannot be considered at one time in the prior art. The implementation steps of the method include: firstly, extracting and removing a target image of an input SAR image; secondly, calculating an iterative initial value, and setting iterative times and iterative step size; thirdly, performing logarithmic transformation to an image after the target image is removed, denoising by the three-dimensional matching denoising algorithm with different standard deviations of noise, and selecting the best noise deviation by evaluating the edge preserving capacity and equivalent number of looks of the denoised image; fourthly, performing three-dimensional matching denoising to a logarithmic image by the best standard deviation of the image; and fifthly, performing exponent transformation to the denoised image, and adding the image with the target image to acquire a final despeckled image. The method has the advantage that image details are saved well, is effective in despeckle, and can be used for SAR image target identification and target detection.

Description

The SAR image speckle method of based target extraction and the denoising of three-dimensional bits coupling

Technical field

The invention belongs to technical field of image processing, specifically a kind of spot method that goes can be applicable to identification of SAR image object and target detection.

Background technology

Synthetic-aperture radar Synthetic Aperture Radar, being called for short SAR is a kind of round-the-clock, round-the-clock radar that can under various weather conditions, obtain the high-definition picture of similar optical imagery, belongs to the active remote sensing system.SAR does not receive condition effect such as illumination, weather, can be applicable to the geographical supervision of Nong Ye ﹑ Jun Shi ﹑ Dao Hang ﹑ and waits numerous areas.But because in the SAR imaging process, because the influence of the fading effect that coherent illumination produces, the even matter region list that makes script in the image have identical backscattering coefficient reveals the graininess noise, and this noise is referred to as coherent spot.The existence of coherent spot has increased the complicacy of image interpretation, greatly reduces the validity of Tu as Fen Ge ﹑ target classification and other information extractions.

The primary goal that spot falls in the SAR image is in the filtering speckle noise, keeps the detailed information of image as much as possible.In recent years, SAR image coherent speckle noise suppresses technological develop rapidly, can be divided into looking smoothing technique and imaging back filtering technique two class greatly before the imaging more.In the early stage SAR imaging processing, mostly adopt and look treatment technology more, but along with the continuous expansion of SAR image applications, the requirement of its spatial resolution is improved constantly, look treatment technology more and can not meet the demands.Filtering technique after the imaging can be divided into again: airspace filter technology and frequency domain filtering technology.

Method for reducing speckle based on the airspace filter technology all is the supposition noise model usually; On image, get a sliding window then; As the input value of wave filter, estimate the image of no coherent speckle noise with all pixels in the window, it is based on the partial statistics characteristic and carries out Filtering Processing.The airspace filter method has enhanced Lee filtering, Kuan filtering, and Frost filtering, and GammaMAP filtering etc.The filtering method in these spatial domains is difficult to keep the minutia of image, and the quality of its filtering performance depends on the size of selected filter window to a great extent.

Can describe the local feature of image from different resolution space based on the method for reducing speckle of frequency domain filtering technology, make signal and noise in wavelet transformed domain, show different character, thereby be prone to differentiation signal and noise.Through comparative analysis; Method of wavelet has the edge and keeps effect preferably; But because the limited directivity of wavelet transformation; The Shui Ping ﹑ that it only can catch image vertically reaches the information of three directions in diagonal angle, and really powerless for other directions, and this loses seriously with regard to the image detail information after causing handling.

Three-dimensional bits coupling Denoising Algorithm is proposed in 2007 by people such as Dabov, and it can utilize the autocorrelation performance of image to carry out effective noise reduction.Three-dimensional bits coupling Denoising Algorithm belongs to non local multiple spot type algorithm, and it had both comprised the thought of non local denoising, had used the method for transform domain filtering simultaneously again, can obtain effect preferably for removing speckle noise.But when it is used in the SAR image speckle,,, therefore can't obtain the best spot effect of going, and can cause the losing of image information of part so can't know accurately that the noise criteria of SAR image is poor because existing Noise Estimation technology all has certain error.

Summary of the invention

The objective of the invention is to be directed against the shortcoming of above-mentioned existing problem, propose the SAR image speckle method of a kind of based target identification and the denoising of three-dimensional bits coupling, with effect and the reservation image information as much as possible that obtains best removal speckle noise.

Realize that technical scheme of the present invention is: earlier the target image in the SAR image of input is extracted; Calculate then and removed the optimum noise standard deviation of the three-dimensional bits coupling Denoising Algorithm of image after the target; And adopt the optimum noise standard deviation to carry out the denoising of three-dimensional bits coupling to the image that has removed after the target; Last again with in the target image adding denoising result image that extracts, concrete steps comprise as follows:

(1) the SAR image A of input is carried out the target image extraction at details and edge, the target image T that obtains extracting;

(2) the target image T that extracts is removed from the SAR image A of input, obtain removing the SAR image B behind the target image;

(3) the SAR image B that removes behind the target image is carried out logarithm operation, obtain the logarithmic image C of image B;

(4) to logarithmic image C, the optimum noise standard deviation a of Calculation of Three Dimensional piece coupling Denoising Algorithm:

The robust median method that 4a) utilizes Donoho to propose removes the noise criteria difference σ of the logarithmic image C of the SAR image B behind the target image with the estimation of quadrature discrete wavelet decomposition:

$\mathrm{\σ}=\frac{\mathrm{median}\left(\mathrm{abs}\left({\mathrm{HH}}_1\right)\right)}{0.6745},$

Wherein, HH ₁Be the quadrature discrete wavelet conversion coefficient of the high pass-high pass subspace of the two-dimensional wavelet transformation of logarithmic image C, abs representes to take absolute value, and median representes to get intermediate value;

4b) iterations t=20 is set, iteration step length step=0.025, the three-dimensional bits of calculating each iteration is mated the noise criteria difference s (i) of Denoising Algorithm:

s(i)=σ+i×step，i=1,2,…,t；

4c) use noise criteria difference s (i) respectively, adopt three-dimensional bits coupling Denoising Algorithm that logarithmic image C is carried out denoising, obtain image sets D (i) after the three-dimensional bits coupling denoising of different noise criteria differences as the Noise Estimation standard deviation, i=1,2 ..., t;

4d) image sets D (i) after the three-dimensional bits coupling denoising of different noise criteria differences is carried out exponent arithmetic, obtain index image group H (i);

4e) the equivalent number array I (i) of gauge index image sets H (i) and edge keep degree array J (i):

$J\left(i\right)=\frac{\left[\mathrm{\Σ}\left(\right|S_b(l,m)-S_b(l+1,m)|+|S_b(l,m)-S_b(l,m+1)\left|\right)\right]}{\left[\mathrm{\Σ}\left(\right|S_a(l,m)-S_a(l+1,m)|+|S_a(l,m)-S_a(l,m+1)\left|\right)\right]}$

In the formula, u and λ are respectively the even value and the standard deviation of a homogeneous region among the index image group H (i), and (l, m) expression removes the coordinate of the edge pixel point in the SAR image B behind the target image, S _a(l, m) expression removes the gray-scale value of the edge pixel point of the SAR image B behind the target image, S _b(l, m) gray-scale value of expression index image group H (i) edge pixel point; The value of I (i) is big more, and picture contrast is more little, and the expression wave filter is strong more to the smoothing capability of speckle noise; The image border keeps degree J (i) big more, and the expression wave filter is good more to the details hold facility of original image;

4f) keep degree array J (i) to carry out standardization equivalent number array I (i) and edge respectively, obtain standardization equivalent number array K (i) and standardization edge and keep degree array L (i);

Be horizontal ordinate 4g) with s (i); K (i) and standardization edge maintenance degree array L (i) with corresponding standardization equivalent number array is that ordinate is made two curves respectively, and then the value of these two intersections of complex curve horizontal ordinates is the value of the three-dimensional bits coupling Denoising Algorithm optimum noise standard deviation a of logarithmic image C;

(5) adopt optimum noise standard deviation a, logarithmic image C is carried out the denoising of three-dimensional bits coupling, obtain image M after the three-dimensional bits coupling denoising of logarithmic image C;

(6) three-dimensional bits is mated denoising after image M carry out exponential transform, obtain the SAR image N behind the spot;

(7) with target image T that extracts and the SAR image N addition of going behind the spot, finally removed spot result images R.

The present invention has the following advantages compared with prior art:

1) the present invention is owing to adopted the method that has combined bilateral filtering and canny operator edge detection that the SAR image of importing is carried out the extraction of target image; Thereby can the point target in the image, line target and grain details be extracted, kept the image detail characteristic in the former SAR image well.

2) the present invention is owing to adopted the three-dimensional bits coupling Denoising Algorithm of using the optimum noise standard deviation to go spot to handle to the SAR image; Keep degree and homogeneous region to go the spot effect through edge of image after the three-dimensional bits coupling denoising of estimating different noise criteria differences; Select the optimum noise standard deviation, make it when having the best to go the spot effect, can not lose too much image information.

Simulation result shows that the inventive method is than other several kinds existing classical SAR image speckle methods, and hold facility and homogeneous region go to spot effect aspect all to increase significantly on the edge of.

Description of drawings

Fig. 1 is realization flow figure of the present invention;

Fig. 2 is that the width of cloth two that the present invention tests input is looked the SAR image;

Fig. 3 is to the SAR image speckle of Fig. 2 comparison diagram as a result with the present invention and existing method;

Fig. 4 is the target image that from Fig. 2, extracts with the present invention;

Fig. 5 is the curve with the optimum noise standard deviation a of Calculation of Three Dimensional block matching algorithm of the present invention.

Embodiment

With reference to Fig. 1, practical implementation process of the present invention is following:

The target image that the SAR image A of step 1. pair input is carried out details and edge extracts the target image T that obtains extracting:

1a) the SAR image A of input is carried out bilateral filtering and handle, it is 120 to 180 that thresholding Y is set, and gray-scale value in the filtered image is made as 0 less than the gray values of pixel points of thresholding Y; Obtain bilateral filtering edge image E, the concrete realization of bilateral filtering algorithm is referring to Tomasi, C.; And Manduchi; R.: " Bilateral filtering for gray and color Images " .Proc.6th Int.Conf.Computer Vision, New Delhi, India; 1998, pp.839 – 846;

1b) use the canny operator to carry out rim detection, obtain canny edge image F the SAR image A of input;

1c) with bilateral filtering edge image E and canny edge image F addition, the total edge image G of the SAR image A that obtains importing;

1d) (x, y) ≠ 0 (x, pixel y) extracts from the SAR image A of input, obtains target image T at the place with gray-scale value g among the total edge image G of SAR image A of input.

Step 2. removes the target image T that extracts from the SAR image A of input, obtain removing the SA image B behind the target image.

The SAR image B that step 3. pair removes behind the target image is carried out logarithm operation, obtains the logarithmic image C of image B.

Step 4. is to logarithmic image C, and the Calculation of Three Dimensional piece matees the optimum noise standard deviation a of Denoising Algorithm:

The robust median method that 4a) utilizes Donoho to propose, estimate the noise criteria difference σ of the logarithmic image C of image B with the quadrature discrete wavelet decomposition:

$\mathrm{\σ}=\frac{\mathrm{median}\left(\mathrm{abs}\left({\mathrm{HH}}_1\right)\right)}{0.6745},$

Wherein, HH ₁Be the quadrature discrete wavelet conversion coefficient of the high pass-high pass subspace of the two-dimensional wavelet transformation of logarithmic image C, abs representes to take absolute value, and median representes to get intermediate value;

4b) iterations t=20 is set, iteration step length step=0.025, the three-dimensional bits of calculating each iteration is mated Denoising Algorithm noise criteria difference s (t):

s(i)=σ+i×step，i=1,2,…,t；

4c) use s (i) respectively, adopt three-dimensional bits coupling Denoising Algorithm that logarithmic image C is carried out denoising, obtain image sets D (i) after the three-dimensional bits coupling denoising of different noise criteria differences as the Noise Estimation standard deviation, i=1,2 ..., t;

Three-dimensional bits is mated the concrete realization of Denoising Algorithm referring to Dabov K, Foi A, Katkovnik V; Egiazarian K. " Image denoising by sparse 3-D transform-domain collaborative filtering; " IEEE Trans.Image Process., vol.16, no.8; Pp.2080-2095, Aug.2007;

4d) image sets D (i) after the three-dimensional bits coupling denoising of different noise criteria differences is carried out exponent arithmetic, obtain index image group H (i);

4e) the equivalent number array I (i) of gauge index image sets H (i) and edge keep degree array J (i):

$J\left(i\right)=\frac{\left[\mathrm{\Σ}\left(\right|S_b(l,m)-S_b(l+1,m)|+|S_b(l,m)-S_b(l,m+1)\left|\right)\right]}{\left[\mathrm{\Σ}\left(\right|S_a(l,m)-S_a(l+1,m)|+|S_a(l,m)-S_a(l,m+1)\left|\right)\right]}$

In the formula, u and λ are respectively the even value and the standard deviation of a homogeneous region among the index image group H (i), and (l, m) expression removes the coordinate of the edge pixel point in the SAR image B behind the target image, S _a(l, m) expression removes the gray-scale value of the edge pixel point of the SAR image B behind the target image, S _b(l, m) gray-scale value of expression index image group H (i) edge pixel point; The value of I (i) is big more, and picture contrast is more little, and the expression wave filter is strong more to the smoothing capability of speckle noise; The value of J (i) is big more, and the expression wave filter is good more to the details hold facility of original image;

4f) keep degree array J (i) to carry out standardization equivalent number array I (i) and edge respectively, obtain standardization equivalent number array K (i) and standardization edge and keep degree array L (i):

$\left\{\begin{array}{c}K\left(i\right)=[I\left(i\right)-\min \left(I\left(i\right)\right)]/[\max \left(I\left(i\right)\right)-\min \left(I\left(i\right)\right)]\\ L\left(i\right)=[J\left(i\right)-\min \left(J\left(i\right)\right)]/[\max \left(J\left(i\right)\right)-\min \left(J\left(i\right)\right)]\end{array}\right.;$

Be horizontal ordinate with s (i) 4g), K (i) and the L (i) with correspondence is that ordinate is made two curves respectively, and then the value of these two intersections of complex curve horizontal ordinates is the value of the three-dimensional bits coupling Denoising Algorithm optimum noise standard deviation a of logarithmic image C.

Step 5. adopts optimum noise standard deviation a, and logarithmic image C is carried out the denoising of three-dimensional bits coupling, obtains image M after the three-dimensional bits coupling denoising of logarithmic image C.

Step 6. with the denoising of the three-dimensional bits of logarithmic image C coupling after image M carry out exponential transform, obtain the SAR image N behind the spot.

Step 7. is finally removed target image T that extracts and the SAR image N addition of going behind the spot to spot result images R.

Effect of the present invention can further specify through following emulation experiment:

1. emulation experiment environment and parameter setting:

In the emulation experiment, various filtering methods all are to use the matlab Programming with Pascal Language to realize.The emulation experiment parameter is set to: bilateral filtering is to get 5 * 5 window, wherein σ _dFor the standard deviation of spatial domain Gaussian function is made as 6, σ _rFor the standard deviation of brightness domain Gaussian function is made as 0.2; Enhanced Lee filtering is to choose 5 * 5 window; When using the present invention to make an experiment, choose step=0.025, t=20.

2. emulation experiment content:

Emulation experiment 1 goes spot to handle simulation result such as Fig. 3 (a) with existing bilateral filtering algorithm to Fig. 2;

Emulation experiment 2 goes spot to handle simulation result such as Fig. 3 (b) with existing enhanced Lee filtering algorithm to Fig. 2;

Emulation experiment 3 goes spot PPB filtering algorithm to go spot to handle simulation result such as Fig. 3 (c) to Fig. 2 with iteration under the existing maximum likelihood framework;

Emulation experiment 4 goes spot to handle simulation result such as Fig. 3 (d) with the inventive method to Fig. 2;

Emulation experiment 5 is carried out the extraction of target image with the present invention to Fig. 2, simulation result such as Fig. 4;

Emulation experiment 6, calculating has removed the three-dimensional bits coupling Denoising Algorithm optimum noise standard deviation of SAR image behind the target image, simulation result such as Fig. 5 to Fig. 2 with the present invention.

3. The simulation experiment result analysis:

Can find out that from Fig. 3 (a) and Fig. 3 (b) existing bilateral filtering algorithm and enhanced Lee filtering algorithm all have certain spot effect of going, but not good especially, and make picture produce certain blooming, can lose most edge and minutia; Can find out from Fig. 3 (c), use when iteration goes spot PPB algorithm to carry out the SAR image speckle under the existing maximum likelihood framework that excessively level and smooth problem can occur to homogeneous area, the distortion of line target also can appear in edge region; From Fig. 3 (d), can find out, better preserve of the present invention target image in the image, can obtain even local and to go the spot effect preferably, and not have to occur excessively level and smooth situation; Fig. 4 therefrom can find out the detail textures information in the original image that extracted that the present invention is more complete for the target image of the present invention to Fig. 2 extraction.

As the quantitative evaluation index of going the spot result, the equivalent number ENL of a homogeneous region is big more with the equivalent number ENL that is labeled as three homogeneous regions of 1,2,3 among Fig. 2 and the edge hold facility ESI that removes the spot result images, the expression algorithm go the spot ability strong more; The edge keeps the big more expression algorithm of degree ESI to keep the ability of edge of image strong more; Above-mentioned several kinds of existing filtering methods go spot result and the spot result that goes of the present invention to keep degree ESI to be listed in the table 1 at the ENL and the edge of the homogeneous region of three marks of Fig. 2.

Table 1 goes the contrast of spot arithmetic result for several kinds

From table 1, can find out at everyways such as Jun Zhi ﹑ Biao Zhun Cha ﹑ equivalent number and edge maintenance degree has preferable performance, and the spot effect of going of this method has demonstrated the ability of strong reflection targets such as homogeneous region spot inhibition ability and certain protection dotted line.Simultaneously, the present invention compares with the airspace filter method based on Bian Yuan ﹑ Tong matter Qu Yu ﹑ point target classification, to detailed information, has shown better effect than other method when recovering as line and texture.The present invention the most significantly advantage is that its level and smooth performance to the homogeneity local is fine, and the tiny texture structure information in the while reservation image as much as possible.The spot method that goes in some spatial domains with respect to other has more performance, can in the better smooth speckle noise, keep SAR edge of image and grain details.

Claims (4)

1. a based target extracts the SAR image speckle method of mating denoising with three-dimensional bits, comprises the steps:

(1) the SAR image A of input is carried out the target image extraction at details and edge, the target image T that obtains extracting;

(2) the target image T that extracts is removed from the SAR image A of input, obtain removing the SAR image B behind the target image;

(3) the SAR image B that removes behind the target image is carried out logarithm operation, obtain the logarithmic image C of image B;

(4) to logarithmic image C, the optimum noise standard deviation a of Calculation of Three Dimensional piece coupling Denoising Algorithm:

The robust median method that 4a) utilizes Donoho to propose removes the noise criteria difference σ of the logarithmic image C of the SAR image B behind the target image with the estimation of quadrature discrete wavelet decomposition:

$\mathrm{\σ}=\frac{\mathrm{median}\left(\mathrm{abs}\left({\mathrm{HH}}_1\right)\right)}{0.6745},$

Wherein, HH ₁Be the quadrature discrete wavelet conversion coefficient of the high pass-high pass subspace of the two-dimensional wavelet transformation of logarithmic image C, abs representes to take absolute value, and median representes to get intermediate value;

4b) iterations t=20 is set, iteration step length step=0.025, the three-dimensional bits of calculating each iteration is mated the noise criteria difference s (i) of Denoising Algorithm:

s(i)=σ+i×step，i=1,2,…,t；

4c) use noise criteria difference s (i) respectively, adopt three-dimensional bits coupling Denoising Algorithm that logarithmic image C is carried out denoising, obtain image sets D (i) after the three-dimensional bits coupling denoising of different noise criteria differences as the Noise Estimation standard deviation, i=1,2 ..., t;

4d) image sets D (i) after the three-dimensional bits coupling denoising of different noise criteria differences is carried out exponent arithmetic, obtain index image group H (i);

4e) the equivalent number array I (i) of gauge index image sets H (i) and edge keep degree array J (i):

$J\left(i\right)=\frac{\left[\mathrm{\Σ}\left(\right|S_b(l,m)-S_b(l+1,m)|+|S_b(l,m)-S_b(l,m+1)\left|\right)\right]}{\left[\mathrm{\Σ}\left(\right|S_a(l,m)-S_a(l+1,m)|+|S_a(l,m)-S_a(l,m+1)\left|\right)\right]}$

In the formula, u and λ are respectively the even value and the standard deviation of a homogeneous region among the index image group H (i), and (l, m) expression removes the coordinate of the edge pixel point in the SAR image B behind the target image, S _a(l, m) expression removes the gray-scale value of the edge pixel point of the SAR image B behind the target image, S _b(l, m) gray-scale value of expression index image group H (i) edge pixel point; The value of I (i) is big more, and picture contrast is more little, and the expression wave filter is strong more to the smoothing capability of speckle noise; The value of J (i) is big more, and the expression wave filter is good more to the details hold facility of original image;

4f) keep degree array J (i) to carry out standardization equivalent number array I (i) and edge respectively, obtain standardization equivalent number array K (i) and standardization edge and keep degree array L (i);

Be horizontal ordinate 4g) with s (i); K (i) and standardization edge maintenance degree array L (i) with corresponding standardization equivalent number array is that ordinate is made two curves respectively, and then the value of these two intersections of complex curve horizontal ordinates is the value of the three-dimensional bits coupling Denoising Algorithm optimum noise standard deviation a of logarithmic image C;

(5) adopt optimum noise standard deviation a, logarithmic image C is carried out the denoising of three-dimensional bits coupling, obtain image M after the three-dimensional bits coupling denoising of logarithmic image C;

(6) three-dimensional bits is mated denoising after image M carry out exponential transform, obtain the SAR image N behind the spot;

(7) with target image T that extracts and the SAR image N addition of going behind the spot, finally removed spot result images R.

2. according to the SAR image speckle method of claims 1 described based target extraction and the denoising of three-dimensional bits coupling, wherein the described SAR image A to input of step (1) is carried out the target image extraction at details and edge, carries out according to following steps:

2a) the SAR image A of input is carried out bilateral filtering and handle, it is the numerical value between 120 to 180 that Lower Threshold Y is set, and gray-scale value in the filtered image is made as 0 less than the gray values of pixel points of Lower Threshold Y, obtains bilateral filtering edge image E;

2b) use the canny operator to carry out rim detection, obtain canny edge image F the SAR image A of input;

2c) with bilateral filtering edge image E and canny edge image F addition, obtain the total edge image G of SAR image A;

2d) with gray-scale value g among the total edge image G (x, y) ≠ 0 the place (x, pixel y) from the input the SAR image A extract, obtain target image T.

3. extract the SAR image speckle method of mating denoising with three-dimensional bits according to the described based target of claims 1; Wherein said step 4c) adopting three-dimensional bits coupling Denoising Algorithm that logarithmic image C is carried out denoising, is earlier image to be divided into onesize rectangular block; According to the similarity of structure between the image block, combine the formation three-dimensional array to two dimensional image piece again with analog structure; At last image is carried out denoising according to the similarity between the image block in same group.

4. extract the SAR image speckle method with the denoising of three-dimensional bits coupling, wherein said step 4f according to the described based target of claims 1) in equivalent number array I (i) and edge maintenance degree array J (i) are carried out standardization, be to carry out through following formula: $\left\{\begin{array}{c}K\left(i\right)=[I\left(i\right)-\mathrm{Min}\left(I\left(i\right)\right)]/[\mathrm{Max}\left(I\left(i\right)\right)-\mathrm{Min}\left(I\left(i\right)\right)]\\ L\left(i\right)=[J\left(i\right)-\mathrm{Min}\left(J\left(i\right)\right)]/[\mathrm{Max}\left(J\left(i\right)\right)-\mathrm{Min}\left(J\left(i\right)\right)]\end{array}\right.;$

Wherein, K (i) is a standardization equivalent number array, and L (i) keeps the degree array for the standardization edge.

Images