CN103927737A - SAR image change detecting method based on non-local mean - Google Patents

Abstract

The invention discloses an SAR image change detecting method based on non-local mean filtering. The SAR image change detecting method comprises the steps that two SAR images which are obtained from the same region at different times are pre-processed; the two pre-processed SAR images are used for constructing a ratio difference shadowgraph; all pixels of the ratio difference shadowgraph are traversed and a smooth index matrix of each pixel point is calculated; after non-local mean filtering is conducted on the two pre-processed SAR images respectively, ratio calculation is conducted, and a non-local mean filtering ratio graph is obtained; the ratio difference shadowgraph and the non-local mean filtering ratio graph are summated with smoothness indexes as weights so as to obtain a final difference shadowgraph; the final difference shadowgraph is divided by using the fuzzy local C mean clustering method to obtain a change detecting result graph. According to the SAR image change detecting method based on non-local mean filtering, difference graph edge information is maintained by using the characteristics of the image smoothness indexes, a non-local mean is used for correcting the pixels in a homogeneous region with a low smoothness index, therefore, noise is effectively restrained, actual change information is better shown and the change detecting result accuracy is improved.

Description

SAR image change detection method based on non-local mean

Technical field

The invention belongs to Remote Sensing Imagery Change Detection field, specifically, relate to a kind of SAR(Synthetic Aperture Radar based on non-local mean thought structural differences image, synthetic-aperture radar) image change detection method.

Background technology

Synthetic aperture radar (SAR) Image Change Detection is from different time, to obtain the multi-temporal remote sensing image of same geographic area, the technology of qualitatively analyze and definite earth's surface change procedure and feature, compare with optical imagery, synthetic aperture radar (SAR) has round-the-clock, round-the-clock obtains the ability of data, has simultaneously and passes through certain vegetation and the ability of overcover, compares with optical imagery, it more easily distinguishes the target on ground, general strong the supplementing using it as optical sensor.SAR Image Change Detection technology is just being widely used in every field, for example environmental monitoring, agricultural research, urban area research, the aspects such as forest monitoring.Many kinds of multidate SAR image change detection methods have been engendered in recent years, the most frequently used SAR Image Change Detection technology is mainly comprised of two steps: first, to multidate SAR image configuration difference image figure, then on difference image figure basis, extract region of variation.

Difference image construction process is the pith of SAR Image Change Detection, SAR image by same region that two width different times are obtained compares to process and obtains difference image hum pattern, and the precision height of difference image hum pattern is the performance of impact variation testing result directly.In SAR Image Change Detection, the SAR image through registration and radiant correction being compared to structural differences striograph is a most important and requisite step, the quality of difference image figure directly determines the degree of accuracy that subsequent analysis is processed, and then has influence on the performance of whole SAR Image Change Detection system.

Existing change detecting method often can not well be removed noise problem intrinsic in SAR image, change detecting method based on ratioing technigue is insensitive to region of variation, loss is higher, and the easy not region of variation of change detecting method based on average ratio is attributed to region of variation, and mistake lapse rate is higher.Often noise is more for the difference image figure constructing with existing difference image building method; using some denoising method denoisings more by often causing image information to lose; the difference image constructing is often of low quality; be difficult to obtain the difference image figure that not only comprises great amount of images variations in detail information but also overcome noise effect; cause variation testing result error very large, the analysis of impact to testing result accuracy.

Generally speaking, need at present the urgent technical matters solving of those skilled in the art to be: outstanding region of variation information and effectively suppress the not information of region of variation how farthest, construct and not only include mass efficient information but also difference image figure substantially not affected by noise, therefrom extract change information, improve the degree of accuracy that changes testing result.

Summary of the invention

In view of this, technical matters to be solved by this invention is, the disparity map producing for existing difference drawing generating method is of low quality, the problem that information dropout is more, proposed based on non-local mean structural differences striograph, method for SAR Image Change Detection, feature for SAR image, with pixel Smoothness Index as weight, the thought of non-local mean is introduced to difference image construction process, structure comprises more effective informations and can suppress to a certain extent the difference image figure of noise, improves SAR Image Change Detection precision.

The invention provides a kind of SAR image change detection method based on non-local mean, concrete steps comprise as follows:

(1) the two width SAR images that same region different time obtained carry out filtering and noise reduction, and the pre-service of radiant correction and geometrical registration obtains pretreated two width SAR image I ₁, I ₂;

(2) utilize pretreated two width SAR image I ₁and I ₂, structure ratio difference striograph

(3) traversal ratio difference striograph D _reach pixel, calculate the Smoothness Index matrix of each pixel wherein, μ (x) is the pixel value average in the neighborhood centered by pixel i, and σ (x) is the variance of pixel value in the neighborhood centered by pixel i;

(4) to pretreated SAR image I ₁and I ₂carry out respectively non-local mean filtering, obtain image NL (I ₁) and NL (I ₂);

(5) will be through the filtered image NL of non-local mean (I ₁) and NL (I ₂) do ratio computing, obtain non-local mean filtering ratio figure D _nR;

(6) using Smoothness Index as weight, by ratio difference striograph D _rwith non-local mean filtering ratio images D _nRcarry out summation operation, obtain final difference image figure

(7) use fuzzy Local C means clustering method to carry out image to final difference image figure DI and cut apart, generate and change testing result figure, complete the final detection to two width SAR image change information.

In described step (4) to pretreated SAR image I ₁and I ₂carry out respectively non-local mean filtering, step is as follows:

To SAR image I ₁carry out non-local mean computing, traversal image I ₁each pixel, calculates wherein refer to the image I at SAR ₁in the radius of take centered by pixel i be r search window, x _pthe pixel value of pixel p, be pixel i and in search window the similarity weight of pixel p, and meet 0≤w _ip≤ 1 He ${\mathrm{\Σ}}_{p\∈W_i^r}{w}_{\mathrm{ip}}=1,w_{\mathrm{ip}}(p\∈W_i^r)$ By formula $w_{\mathrm{ip}}=\frac{1}{Z_i}\exp (-\underset{k=1}{\overset{{(2s+1)}^2}{\mathrm{\Σ}}}\frac{1}{h}\log (\frac{A_{i,k}}{A_{p,k}}+\frac{A_{p,k}}{A_{i,k}}))$ Try to achieve, wherein s is neighborhood windows radius and s=3, parameter h is for the decay of control characteristic function, theoretically, non-local mean will be got 7 * 7 neighborhood pieces all over each point in image, and due in the situation that image is larger, such time complexity is too high, therefore non-local mean computing is carried out in a near region (being search window) larger common selected pixels point, make in the present invention r=10, in the region of 21 * 21, carry out non local computing, A _i,k, A _p,krepresent respectively the pixel value of k pixel centered by pixel i and pixel p. for to SAR image I ₁in non local filtered picture element matrix, the pixel value of i pixel, obtains SAR image I ₁non-local mean filtering image NL (I ₁).

To SAR image I ₂carry out non-local mean computing, traversal image I ₂each pixel, calculates wherein refer to the image I at SAR ₂in the radius of take centered by pixel i be r search window, x _pthe pixel value of pixel p, be pixel i and in search window the similarity weight of pixel p, and meet 0≤w _ip≤ 1 He ${\mathrm{\Σ}}_{p\∈W_i^r}{w}_{\mathrm{ip}}=1,w_{\mathrm{ip}}(p\∈W_i^r)$ By formula $w_{\mathrm{ip}}=\frac{1}{Z_i}\exp (-\underset{k=1}{\overset{{(2s+1)}^2}{\mathrm{\Σ}}}\frac{1}{h}\log (\frac{A_{i,k}}{A_{p,k}}+\frac{A_{p,k}}{A_{i,k}}))$ Try to achieve, wherein s is neighborhood windows radius and s=3, and parameter h, for the decay of control characteristic function, makes r=10, in the region of 21 * 21, carries out non local computing, A _i,k, A _p,krepresent respectively the pixel value of k pixel centered by pixel i and pixel p. for to SAR image I ₂in non local filtered picture element matrix, the pixel value of i pixel, obtains SAR image I ₁non-local mean filtering image NL (I ₂).

The present invention has following beneficial effect compared with prior art:

1, the present invention adopts Smoothness Index as weight, at structural differences, affect in the process of figure, introduce the corrected pixel value of non-local mean, in conjunction with original ratio information and non local thought, image is revised, contribute certain weight, constructed the difference image figure of the better quality that comprises more details information.

2, the present invention combines the method for non-local mean computing in the process of structural differences striograph, has removed preferably the noise in difference image figure, has improved the performance of difference image figure, has increased the degree of accuracy that changes testing result.

3, the present invention adopts the method for fuzzy Local C average to cut apart difference image figure, can suppress greatly the background information of region of variation, strengthens the change information of region of variation, thereby obtains higher accuracy of detection.

4, the method construct that the present invention introduces non-local mean filtering had not only kept the different information figure that changes detailed information but also fully suppress noise, was convenient to later stage disparity map analysis, improved the accuracy of detection in SAR Image Change Detection, reduced error rate.

Accompanying drawing explanation

Fig. 1 is the general flow chart of realizing of the present invention;

Fig. 2 is Bern area two width SAR images and the width canonical reference figure that emulation of the present invention is used;

Fig. 3 is the Bern area two width SAR image difference striographs that the present invention is based on non-local mean method construct;

Fig. 3 changes respectively the simulation result figure of detection with the present invention and existing change detecting method based on ratioing technigue, change detecting method based on average ratio to Bern area two width SAR images.

Embodiment

Referring to accompanying drawing, implementation of the present invention and advantage are described in detail.

With reference to accompanying drawing 1, performing step of the present invention is as follows:

Step 1, the two width SAR images that same region different time is obtained carry out filtering and noise reduction, and the pre-service of radiant correction and geometrical registration obtains pretreated two width SAR image I ₁, I ₂.

Geometric error that can removal of images by pre-service, has reached the coupling to the geographic coordinate of the same area different images, eliminates the radiated noise that noise that sensor self causes and atmosphere radiation cause.

Step 2, utilizes pretreated two width SAR image I ₁and I ₂, structure ratio difference striograph

By image I ₁in be positioned at the gray-scale value I of the pixel (i, j) of the capable j of i row ₁(i, j) and corresponding image I ₂in be positioned at the gray-scale value I of the pixel of the capable j of i row ₂(i, j), by ratio computing D _r(i, j)=I ₁(i, j) I ₂(i, j), obtains ratioing technigue disparity map D _rin be positioned at the gray-scale value D of the pixel (i, j) of the capable j of i row _r(i, j); To image I ₁and image I ₂in each be positioned at the capable j of i row pixel gray-scale value from left to right, all carry out from top to bottom ratio computing, construct ratio difference striograph D _r.

Step 3, traversal ratio difference striograph D _reach pixel, calculate the Smoothness Index matrix of each pixel wherein, μ (x) is the average of the pixel value in the neighborhood centered by pixel i, and σ (x) is the variance of pixel value in the neighborhood centered by pixel i.

Here, wherein, x _iradius centered by represent pixel point is the grey scale pixel value of interior i the pixel of neighborhood of n, and Smoothness Index matrix is the matrix the same with ratio difference striograph scale.

According to Smoothness Index characteristic, in the large place of Smoothness Index, be generally image border, the place that Smoothness Index is little is homogeneous region.The present invention utilizes image smoothing indicial response at difference image figure construction phase, in the large place of Smoothness Index, is generally image border, and the pixel value weight of difference image figure is larger; More redundant information is often contained as homogeneous region in the place that Smoothness Index is little, after with non-local mean, its pixel being revised, can better represent truth, and the performance of the SAR Image Change Detection result obtaining is like this best.

Step 4, to pretreated SAR image I ₁and I ₂carry out respectively non-local mean filtering, obtain image NL (I ₁) and NL (I ₂).

4.1 pairs of SAR image I ₁carry out non-local mean computing, traversal image I ₁each pixel, calculates wherein refer to the image I at SAR ₁in the radius of take centered by pixel i be r search window, x _pthe pixel value of pixel p, be pixel i and in search window the similarity weight of pixel p, and meet 0≤w _ip≤ 1 He ${\mathrm{\Σ}}_{p\∈W_i^r}{w}_{\mathrm{ip}}=1,w_{\mathrm{ip}}(p\∈W_i^r)$ By formula $w_{\mathrm{ip}}=\frac{1}{Z_i}\exp (-\underset{k=1}{\overset{{(2s+1)}^2}{\mathrm{\Σ}}}\frac{1}{h}\log (\frac{A_{i,k}}{A_{p,k}}+\frac{A_{p,k}}{A_{i,k}}))$ Try to achieve, wherein s is neighborhood windows radius and s=3, parameter h is for the decay of control characteristic function, theoretically, non-local mean will be got 7 * 7 neighborhood pieces all over each point in image, and due in the situation that image is larger, such time complexity is too high, therefore non-local mean computing is carried out in a near region (being search window) larger common selected pixels point, make in the present invention r=10, in the region of 21 * 21, carry out non local computing, A _i,k, A _p,krepresent respectively the pixel value of k pixel centered by pixel i and pixel p. for to SAR image I ₁in non local filtered picture element matrix, the pixel value of i pixel, obtains SAR image I ₁non-local mean filtering image NL (I ₁).

4.2 pairs of SAR image I ₂carry out non-local mean computing, traversal image I ₂each pixel, calculates wherein refer to the image I at SAR ₂in the radius of take centered by pixel i be r search window, x _pthe pixel value of pixel p, be pixel i and in search window the similarity weight of pixel p, and meet 0≤w _ip≤ 1 He ${\mathrm{\Σ}}_{p\∈W_i^r}{w}_{\mathrm{ip}}=1,w_{\mathrm{ip}}(p\∈W_i^r)$ By formula $w_{\mathrm{ip}}=\frac{1}{Z_i}\exp (-\underset{k=1}{\overset{{(2s+1)}^2}{\mathrm{\Σ}}}\frac{1}{h}\log (\frac{A_{i,k}}{A_{p,k}}+\frac{A_{p,k}}{A_{i,k}}))$ Try to achieve, wherein s is neighborhood windows radius and s=3, and parameter h, for the decay of control characteristic function, makes r=10, in the region of 21 * 21, carries out non local computing, A _i,k, A _p,krepresent respectively the pixel value of k pixel centered by pixel i and pixel p. for to SAR image I ₂in non local filtered picture element matrix, the pixel value of i pixel, obtains SAR image I ₁non-local mean filtering image NL (I ₂).

Non-local mean is the image de-noising method of commonly using in recent years, because traditional local mean value is only considered the impact of the pixel value of pixel peripheral part on itself, can cause weakening edge, and in bilateral filtering, only relate to the distance between pixel and the effect of similarity to pixel value in image, do not take the Global Information of pixel periphery into account.The non local advantage of comprehensive above two kinds of filtering modes preferably, had both considered the directive function of Image neighborhood piece to itself, considered again the impact of the pixel value of other non-neighborhoods in full figure on itself.Similarity correction pixel value by between neighborhood of pixel points piece and other neighborhood of pixel points pieces, can retain image detail, fully suppresses picture noise.The level and smooth index of non-local mean combining image, can retain more pixel value own at the larger fringe region of Smoothness Index, in the less homogeneous region of Smoothness Index, get non-local mean composition more, like this, both the effective filtering of non-local mean to noise can be utilized, and differential image marginal information can be fully retained again.

Step 5, will be through the filtered image NL of non-local mean (I ₁) and NL (I ₂) do ratio computing, obtain non-local mean filtering ratio images D _nR.

By image NL (I ₁) in be positioned at the gray-scale value NL(I of the pixel (i, j) of the capable j of i row ₁(i, j)) and corresponding image NL (I ₂) in be positioned at the gray-scale value NL(I of the pixel of the capable j of i row ₂(i, j)), by ratio computing D _nR(i, j)=NL(I ₁(i, j))/NL(I ₁(i, j)), obtain non-local mean filtering ratio images D _nRin be positioned at the gray-scale value D of the pixel (i, j) of the capable j of i row _nR(i, j); To image NL (I ₁) and image NL (I ₂) I ₁in each be positioned at the capable j of i row pixel gray-scale value from left to right, all carry out from top to bottom ratio computing, construct non-local mean filtering ratio images D _nR.

Step 6, using Smoothness Index as weight correlative value disparity map with the filtering ratio images summation of non-office, does weighted sum computing to each coordinate corresponding point on two width images $\mathrm{DI}(i,j)=\∂(i,j)*D_R(i,j)+(1-\∂(i,j))*D_{\mathrm{NR}}(i,j),$ Wherein DI (i, j), for the grey scale pixel value of the pixel that in summation disparity map DI, coordinate is (i, j), obtains final difference image figure DI, i.e. SAR image I ₁and image I ₂different information figure.

Smoothness Index is the important indicator of evaluation map picture, is the ratio to the variance in each neighborhood of pixel points and average, and the Smoothness Index of pixel is larger, represents that this pixel is image border part; The Smoothness Index of pixel is less, represents that this pixel is for the homogeneous region at non-edge in image.Comparatively speaking, in homogeneous region, redundant information is more, can be by take Smoothness Index as weight, the corrected pixel value of non-local mean is incorporated in disparity map construction process, in generating the process of disparity map, can to image, revise in conjunction with original ratio information and non local thought, contribute certain weight, produce more rational disparity map.

, step 7, uses fuzzy Local C means clustering method to carry out image to final difference image figure DI and cuts apart, and generates and changes testing result figure, completes the final detection to two width SAR image change information.

Effect of the present invention can further illustrate by following emulation:

1, simulated conditions:

Emulation of the present invention is to carry out under the Intel of dominant frequency 1.87GHZ Pentium CPU P6000, the hardware environment of internal memory 2.00GB and the software environment of MATLAB R2009a.

2, simulation parameter

For the experiment simulation figure with reference diagram, can carry out quantitative variation Analysis of test results, main evaluation index has:

1. undetected survey number: the number of pixels in the region that changes in statistical experiment result figure, contrast with the number of pixels of region of variation in reference diagram, changing in reference diagram but detect as unchanged number of pixels in experimental result picture, be called undetected number;

2. error detection number: the number of pixels in the region that do not change in statistical experiment result figure, contrast with the number of pixels in the region that do not change in reference diagram, detect the number of pixels for changing not changing in reference diagram but in experimental result picture, be called flase drop number;

3. total wrong number: undetected number and flase drop number and;

3, emulation experiment content and interpretation of result

In order to verify the superiority of the SAR image change detection method based on non-local mean filtering, the performance of the inventive method and the traditional change detecting method based on ratioing technigue and the change detecting method based on average ratioing technigue algorithm is made to contrast and quantitative test.Here the change detecting method based on ratioing technigue referred to as LR, the change detecting method based on average ratioing technigue algorithm is referred to as MR, the inventive method is referred to as NLR.

Change detecting method in the present invention is applied in as shown in Figure 2 on the SAR image data set in true Bern area, and is analyzed with existing change detecting method and the change detecting method based on average ratioing technigue algorithm based on ratioing technigue.

As Bern area three width images as shown in Figure 2, wherein Fig. 2 (a) represents the geomorphology information in Bern in April, 1999 area, is the very first time to obtain image.Fig. 2 (b) represents that the geomorphology information in Bern in May, 1999 area was for the second time and obtains image.Fig. 2 (c) represents to change the canonical reference figure detecting.

With two width SAR images shown in Fig. 2 (a) and Fig. 2 (b), use respectively the change detecting method based on ratioing technigue, change detecting method and the inventive method based on average ratioing technigue algorithm to obtain changing testing result figure, as shown in Figure 3, wherein Fig. 3 (a) is change detecting method (LR) the result figure based on ratioing technigue, Fig. 3 (b) is change detecting method (MR) the result figure based on average ratio, and Fig. 3 (c) represents the result figure of the present invention (NLR) method.

The performance index of three kinds of methods shown in Fig. 3 are contrasted, as shown in table 1:

The three kinds of change detecting method result Performance Ratios in table 1Bern area

As can be seen from Table 1, the variation testing result of the inventive method is minimum to the undetected number of pixel in image, and total wrong number is minimum, has improved the precision that changes testing result, has shown the superiority of the inventive method.

Although above-mentioned, by reference to the accompanying drawings the specific embodiment of the present invention is described; but be not limiting the scope of the invention; one of ordinary skill in the art should be understood that; on the basis of technical scheme of the present invention, those skilled in the art do not need to pay various modifications that creative work can make or distortion still in protection scope of the present invention.

Claims (3)

1. the SAR image change detection method based on non-local mean filtering, is characterized in that, comprises the steps:

(1) the two width SAR images that same region different time obtained carry out filtering and noise reduction, and the pre-service of radiant correction and geometrical registration obtains pretreated two width SAR image I ₁, I ₂;

(2) utilize pretreated two width SAR image I ₁and I ₂, structure ratio difference striograph

(3) traversal ratio difference striograph D _reach pixel, calculate the Smoothness Index matrix of each pixel wherein, μ (x) is the pixel value average in the neighborhood centered by pixel i, and σ (x) is the variance of pixel value in the neighborhood centered by pixel i;

(4) to pretreated SAR image I ₁and I ₂carry out respectively non-local mean filtering, obtain image NL (I ₁) and NL (I ₂);

(5) will be through the filtered image NL of non-local mean (I ₁) and NL (I ₂) do ratio computing, obtain non-local mean filtering ratio figure D _nR;

(6) using Smoothness Index as weight, by ratio difference striograph D _rwith non-local mean filtering ratio images D _nRcarry out summation operation, obtain final difference image figure

(7) use fuzzy Local C means clustering method to carry out image to final difference image figure DI and cut apart, generate and change testing result figure, complete the final detection to two width SAR image change information.

2. in step as claimed in claim 1 (3), travel through ratio difference striograph D _reach pixel, calculate the Smoothness Index matrix of each pixel, it is characterized in that, described Smoothness Index matrix is the matrix the same with ratio difference striograph scale

As claimed in claim 1 in step (4) to pretreated SAR image I ₁and I ₂carry out respectively non-local mean filtering, obtain image NL (I ₁) and NL (I ₂) it is characterized in that, described to pretreated SAR image I ₁and I ₂the step of carrying out respectively non-local mean filtering is:

To SAR image I ₁carry out non-local mean computing, traversal image I ₁each pixel, calculates wherein refer to the image I at SAR ₁in the radius of take centered by pixel i be r search window, x _pthe pixel value of pixel p, be pixel i and in search window the similarity weight of pixel p, and meet 0≤w _ip≤ 1 He ${\mathrm{\Σ}}_{p\∈W_i^r}{w}_{\mathrm{ip}}=1,w_{\mathrm{ip}}(p\∈W_i^r)$ By formula $w_{\mathrm{ip}}=\frac{1}{Z_i}\exp (-\underset{k=1}{\overset{{(2s+1)}^2}{\mathrm{\Σ}}}\frac{1}{h}\log (\frac{A_{i,k}}{A_{p,k}}+\frac{A_{p,k}}{A_{i,k}}))$ Try to achieve, wherein s is neighborhood windows radius and s=3, parameter h is for the decay of control characteristic function, theoretically, non-local mean will be got 7 * 7 neighborhood pieces all over each point in image, and due in the situation that image is larger, such time complexity is too high, therefore non-local mean computing is carried out in a near region (being search window) larger common selected pixels point, make in the present invention r=10, in the region of 21 * 21, carry out non local computing, A _i,k, A _p,krepresent respectively the pixel value of k pixel centered by pixel i and pixel p. for to SAR image I ₁in non local filtered picture element matrix, the pixel value of i pixel, obtains SAR image I ₁non-local mean filtering image NL (I ₁);

To SAR image I ₂carry out non-local mean computing, traversal image I ₂each pixel, calculates wherein refer to the image I at SAR ₂in the radius of take centered by pixel i be r search window, x _pthe pixel value of pixel p, be pixel i and in search window the similarity weight of pixel p, and meet 0≤w _ip≤ 1 He ${\mathrm{\Σ}}_{p\∈W_i^r}{w}_{\mathrm{ip}}=1,w_{\mathrm{ip}}(p\∈W_i^r)$ By formula $w_{\mathrm{ip}}=\frac{1}{Z_i}\exp (-\underset{k=1}{\overset{{(2s+1)}^2}{\mathrm{\Σ}}}\frac{1}{h}\log (\frac{A_{i,k}}{A_{p,k}}+\frac{A_{p,k}}{A_{i,k}}))$ Try to achieve, wherein s is neighborhood windows radius and s=3, and parameter h, for the decay of control characteristic function, makes r=10, in the region of 21 * 21, carries out non local computing, A _i,k, A _p,krepresent respectively the pixel value of k pixel centered by pixel i and pixel p. for to SAR image I ₂in non local filtered picture element matrix, the pixel value of i pixel, obtains SAR image I ₁non-local mean filtering image NL (I ₂).