CN106056577A - Hybrid cascaded SAR image change detection method based on MDS-SRM - Google Patents

Abstract

The invention discloses a hybrid cascaded SAR image change detection method based on MDS-SRM, and the method mainly solves the problem in an existing SRM algorithm that the change detection is not precise due to factors such as static order. The method comprises the steps of: 1, inputting two-time-phase single-channel SAR images, and respectively carrying out denoising processing on the images; 2, constructing logarithm ratio graphs and mean value ratio graphs for the images after the denoising, and superposing the two kinds of graphs to form dual-channel difference images; 3, merging the dual-channel difference images, and obtaining a first merging result; 4, carrying out merging again based on the first merging result, and obtaining a second merging result; and 5, carrying out ascending sorting on grey mean value of all regions in the second merging result, carrying out regional merging, and obtaining a final change detection result. According to the invention, the detection precision of change regions in the SAR images is effectively improved, and the hybrid cascaded SAR image change detection method can be used for positioning disaster areas and analyzing the expanding condition of a city.

Description

SAR image change detection based on MDS-SRM Mixed cascading

Technical field

The invention belongs to image processing field, further relate to the detection method in two phase image change regions, available Location and the dilatation analysis in city in disaster region.

Background technology

Along with the fast development of remote sensing technology in recent years, remotely-sensed data quantity is growing, and is widely used in environment prison The fields such as survey, atmospheric analysis and urban planning.Wherein SAR image change-detection is applied to military and civilian field, mainly relates to And as the location in natural disaster region, the dilatation analysis in city and the Military Application such as flood, fire and earthquake are hit The assessment of effect, so research SAR image change-detection is significant.

The pretreatment of phase images when SAR image change-detection mainly comprises, generation disparity map and extraction region of variation three Step.Wherein, extract region of variation part, substantially can be divided into without supervisory detection and have supervisory detection method.Comparatively speaking, Unsupervised approaches directly can process differential image by cluster or partitioning algorithm thus obtain change-detection result, compares and agrees with Change-detection lacks the feature of prior information, is therefore widely used.

Unsupervised approaches based on image segmentation includes: maximum between-cluster variance algorithm OTSU, Threshold Segmentation Algorithm K&I, system Meter partitioning algorithm MRF and statistical regions merge algorithm SRM etc..Wherein SRM algorithm has the probability distribution vacation not relying on data And if have the advantage of preferable anti-noise ability, this makes it be more suitable for SAR image change-detection.But SRM algorithm only considers region All value differences carry out static ordering, and merging error probability can be caused to increase, and are intended for single use SRM and cannot obtain final change and examine Survey result.

Summary of the invention

Present invention aims to the deficiency of above-mentioned prior art, propose a kind of based on multilamellar dynamic order Statistical Area Territory merges the Mixed cascading change detecting method of algorithm MDS-SRM, to realize the detection of region of variation in SAR image, promotes inspection The accuracy rate surveyed.

Realization the technical scheme is that and input picture carries out denoising, uses logarithm ratio and average ratio method to build Dual pathways differential image, has processed the extraction of region of variation by Mixed cascading structure to differential image, and wherein first Level uses SRM algorithm that from pixel space, disparity map is converted to super-pixel space, and the second level utilizes MDS-SRM algorithm to complete difference Image small area, to the merging process in big region, finally uses the SRM algorithm of simplification to obtain final change-detection result, its Realize step to include the following:

(1) the single-channel SAR image X of the two identical regions of width different time is inputted₁、X₂, and utilize non-local mean algorithm to go Except the coherent speckle noise in this two width image, obtain two width image I after denoising₁And I₂；

(2) with the image I after denoising₁And I₂, construct log ratio figure DI₁With average ratio value figure DI₂；

(3) by log ratio disparity map DI of structure₁With average ratio value disparity map DI₂It is overlapped, obtains into a width bilateral Road differential image DI；

(4) utilize statistical regions to merge algorithm the pixel in dual pathways differential image DI is merged, complete difference Figure transformation from pixel space to regional space, obtains merging image DT for the first time₁；

(5) multilamellar dynamic order statistical regions is utilized to merge algorithm to image DT after once merging₁In region carry out two Secondary merging, the secondary obtaining having large area merges image DT₂:

(5a) traversal once merges image DT₁In adjacent region, determine that region is to number M；

(5b) similarity f (R', R) of zoning pair:

$f(R^{\′},R)=\underset{k\∈\{{\mathrm{DI}}_1,{\mathrm{DI}}_2\}}{max}\left|\sqrt{\[H_1-H_2\]*S^{-1}*{\[H_1-H_2\]}^T}\right|$

In formula, R' and R represents two adjacent regions, H₁And H₂Represent the characteristic vector in adjacent two regions respectively,WhereinRepresent the average pixel value in k passage Zone R territory, A (R_k) Representing the statistic histogram matrix in Zone R territory, | R | represents the number of pixels of region R；S^-1Covariance for adjacent area characteristic vector Matrix S (a, inverse matrix b)；

(5c) similarity f (R', R) being sorted from small to large, obtain merging order list Index (i), i merges for traversal The sequentially pointer of list, if i=1, the region of selected similarity f (R, R') minima is to (R, R')；

(5d) judge whether (R, R') is merged by the region of minima according to below equation: if meetingThen merge, calculate adjacent area now and to number M and update adjacent area pair Similarity f of (R, R')_n+1(R, R'):

Wherein,Parameter Q represents Statistical Complexity, obtains segmentation in various degree by regulation Q As a result, R_|R|Indicate the region of the individual pixel of | R |, and have | | R_R|||≤(|R|+1)^min(R|,g), g=256, | R | represent that region contains Some number of pixels, constant| I | is the pixel number that image I comprises,Represent the average picture in k passage Zone R territory Element value；f_n(R, R') represent n-th merge before R' region and the similarity in Zone R territory, φ andRepresent this similarity respectively With the weight of history similarity,Update merging order list Index (i), if i=1, selected f_n+1(R, R') is minimum The region of value, to (R, R'), judges whether (R, R') is merged by the region of minima again；

Otherwise, i=i+1, it is judged that whether i≤M sets up, if setting up, then takes adjacent area pair corresponding to Index (i), weight Multiple step (5d)；If being false, merging terminates, and obtains secondary and merges image DT₂；

(6) secondary is merged image DT₂In the gray average in each region carry out ascending sort and obtain merging order row Table, the region then using the merging criterion in SRM algorithm to be combined in order list carries out region merging technique successively, obtains final Change-detection result.

Present invention have the advantage that

First, the present invention, by utilizing average ratio and the respective advantage of logarithm ratio to build dual pathways disparity map, can make up Owing to SAR image comprises only single-channel data in experiment, the shortcoming of multichannel restriction ability in SRM algorithm cannot be made full use of；

Second, the ranking criteria that the present invention is optimized by employing, add on the basis of equal value difference region histogram and Area discrepancy feature so that sequence is more reasonable, thus reduces the probability merged that makes a mistake；

3rd, the present invention utilizes Mixed cascading to merge change detecting method structure can be effectively improved SAR image change inspection The precision surveyed.

Accompanying drawing explanation

Fig. 1 is the flowchart of the present invention；

Fig. 2 is the result figure that Bern data set carries out SAR image change-detection by the present invention；

Fig. 3 is the present invention with existing RFLICM algorithm, SRM algorithm to Bern data set, Ottawa data set and the Yellow River number Change-detection result figure according to collection.

Specific embodiments

Below in conjunction with the accompanying drawings, the present invention is realized step and effect is described in further detail.

With reference to Fig. 1, the present invention to realize step as follows:

Step 1, inputs the single-channel SAR image X of the two identical regions of width different time₁、X₂, and utilize non-local mean to calculate Method removes the coherent speckle noise in this two width image, obtains the image I after denoising₁、I₂。

(1a) the noise image X before change in data set is chosen₁={ X₁(i) | i ∈ I}, I is image pixel fields, to image In any one pixel i, use non-local mean algorithm denoising, the gray scale estimated value obtaining this point is:

$I_1\left(i\right)=\underset{j\∈I}{\Σ}\mathrm{\ω}(i,j)X_1\left(i\right)$

Wherein, ω (i, j) is weights, the similarity degree between expression ith pixel and jth pixel, and 0≤ω (i, j) ≤ 1, andTraversing graph is as X₁In all of pixel, obtain the image I after the first width denoising₁；

(1b) the noise image X after change in data set is chosen₂={ X₂(i) | i ∈ I}, to any one pixel in image Point i, uses non-local mean algorithm denoising, and the gray scale estimated value obtaining this point is:

$I_2\left(i\right)=\underset{j\∈I}{\Σ}\mathrm{\ω}(i,j)X_2\left(i\right)$

Wherein, weights ω (i, j) represents the similarity degree between ith pixel and jth pixel, and 0≤ω (i, j)≤1 AndTraversing graph is as X₂In all of pixel, obtain the image I after the second width denoising₂。

Step 2, with two width image I after denoising₁、I₂, construct log ratio figure DI₁With average ratio value figure DI₂。

(2a) with image I after denoising₁、I₂Middle coordinate is that (i, the pixel value of pixel j) calculate log ratio disparity map DI₁Middle pixel coordinate be (i, pixel value j):Image I after traversal denoising₁、I₂In all of Pixel, obtains log ratio disparity map DI₁, wherein I₁(i, j), I₂(i j) is respectively image I₁, I₂Middle coordinate be (i, j) as The pixel value of vegetarian refreshments；

(2b) with the image I after denoising₁、I₂Middle coordinate is that (i, the pixel value of pixel j) calculate average ratio value difference Figure DI₂Middle pixel coordinate be (i, pixel value j):Scheme after traversal denoising As I₁、I₂In all of pixel, obtain average ratio value disparity map DI₂, wherein μ₁(i, j), μ₂(i j) is respectively image I₁, I₂ In with coordinate for (i, the 2*2 field pixel average centered by pixel j).

Step 3, log ratio disparity map DI that will obtain₁With average ratio value disparity map DI₂It is superimposed as a width dual pathways difference Image DI.

Step 4, utilizes statistical regions to merge algorithm and merges the pixel in dual pathways differential image DI, and it is poor to complete The transformation from pixel space to regional space of the different figure, obtains merging image DT for the first time₁。

(4a) every pair of pixel in dual pathways differential image DI is calculated similarity weight:

$f(p,p^{\′})=\underset{k\∈\{{\mathrm{DI}}_1,{\mathrm{DI}}_2\}}{max}|p_k-p_k^{\′}|$

Wherein p_kAnd p'_kIt is two pixel values the most adjacent；

(4b) similarity weight is carried out from small to large ascending sort, and according to clooating sequence, selected pixels is to sentencing successively Whether this pixel disconnected is to merging: if meetingThen merge, wherein,Represent k passage The average pixel value in Zone R territory, Zone R territory is p_kPixel affiliated area,Parameter Q represents statistics complexity Degree, obtains segmentation result in various degree, R by regulation Q_|R|Indicate the regional ensemble of the individual pixel of | R |, and have | | R_R||≤ (n+1)^min(|R|,g), g=256, constant| I | is the pixel number that image I comprises；Otherwise, nonjoinder；When each Group pixel, to when all completing this judge process, i.e. obtains merging image DT for the first time₁。

Step 5, utilizes multilamellar dynamic order statistical regions to merge algorithm to merging image DT for the first time₁In region carry out Further merge, obtain second time amalgamation result DT₂。

(5a) traversal merges image DT for the first time₁In adjacent region, determine that region is to number M；

(5b) similarity f (R', R) of zoning pair:

$f(R^{\′},R)=\underset{k\∈\{{\mathrm{DI}}_1,{\mathrm{DI}}_2\}}{max}\left|\sqrt{\[H_1-H_2\]*S^{-1}*{\[H_1-H_2\]}^T}\right|,$

In formula, R' and R represents two the most adjacent regions, H₁And H₂Represent the characteristic vector in adjacent two regions respectively,WhereinRepresent the average pixel value in k passage Zone R territory, A (R_k) Representing the statistic histogram matrix in Zone R territory, | R | represents the number of pixels of region R；S^-1Covariance for adjacent area characteristic vector Matrix S (a, inverse matrix b)；

(5c) similarity f (R', R) being sorted from small to large, obtain merging order list Index (i), i merges for traversal The sequentially pointer of list, if i=1, the region of selected similarity f (R, R') minima is to (R, R')；

(5d) judge whether (R, R') is merged by the region of minima according to below equation: if meetingThen merge, calculate adjacent area now to number M, renewal adjacent area pair Similarity f of (R, R')_n+1(R, R') and merging order list Index (i), if i=1, selected f_n+1The district of (R, R') minima Territory, to (R, R'), judges whether (R, R') is merged by the region of minima again, f_n+1The computing formula of (R, R') is as follows:

Wherein,Parameter Q represents Statistical Complexity, obtains segmentation in various degree by regulation Q As a result, R_|R|Indicate the region of the individual pixel of | R |, and have | | R_|R|||≤(|R|+1)^min(|R|,g), g=256, | R | represent region The number of pixels contained, constant| I | is the pixel number that image I comprises,Represent the average of k passage Zone R territory Pixel value；f_n(R, R') represent n-th merge before R' region and the similarity in Zone R territory, φ andRepresent this similarity respectively With the weight of history similarity,

Otherwise, i=i+1, it is judged that whether i≤M sets up, if setting up, then takes adjacent area pair corresponding to Index (i), weight Multiple step (5d)；If being false, merging terminates, and obtains second time and merges image DT₂。

Step 6, merges image DT to second time₂In the gray average in each region carry out ascending sort, then use The merging criterion that statistical regions merges in algorithm carries out region merging technique, obtains final change-detection result figure.

(6a) calculate second time and merge image DT₂In the gray average in each region

$\mathrm{\μ}=\underset{i=0}{\overset{k-1}{\Σ}}{\mathrm{ip}}_i$

P in formula_iRepresenting the ratio of all pixels in region shared by the pixel that gray value in image is i, k represents image Gray level maximum, and carried out ascending sort；

(6b) according to clooating sequence, successively chosen area to and use following merging criterion to judge whether this region to closing And: if meetingThen merge, wherein,Represent the mean pixel in k passage Zone R territory Value,Parameter Q represents Statistical Complexity, obtains segmentation result in various degree, R by regulation Q_|R|Table It is shown with the regional ensemble of the individual pixel of | R |, and has | | R_|R|||≤(n+1)^min(|R|,g), g=256, constant| I | is image The pixel number that I comprises；Otherwise, nonjoinder.

After traveling through all regions, i.e. obtain final change-detection result figure.

The effect of the present invention combines following emulation experiment and further illustrates:

1. simulated conditions

The present invention is to be that Intel (R) Core i5-34703.2GHZ, internal memory 8G, WINDOWS 7 operate at central processing unit On the PC of system, use the emulation experiment that MATLAB 2013b is carried out.

2. emulation content

Emulation 1, uses the inventive method that Bern data set is changed detection, and testing result is as in figure 2 it is shown, wherein:

Fig. 2 (a) represents image before the change inputted；

Fig. 2 (b) represents image after the change inputted；

Before Fig. 2 (c) represents the change to input, after change, image carries out the log ratio difference that log ratio process obtains Figure；

Before Fig. 2 (d) represents the change to input, after change, image carries out the average ratio value difference that average ratio value process obtains Figure；

The dual pathways disparity map being superimposed as by log ratio disparity map and average ratio value disparity map is carried out by Fig. 2 (e) expression Merge the first time amalgamation result figure obtained；

Fig. 2 (f) expression merges the second time amalgamation result figure obtained on the basis of first time amalgamation result figure；

Fig. 2 (g) expression merges the final change-detection result obtained on the basis of second time amalgamation result figure further Figure.

As seen from Figure 2, the inventive method can effectively by the Changing Area Detection in SAR image out.

Emulation 2, uses RFLICM algorithm, SRM algorithm and the inventive method to Bern data set, Ottawa data set and Huang River data set is changed detection, and result is as it is shown on figure 3, wherein:

Fig. 3 (a) represents the front image of input change of three group data sets；

Fig. 3 (b) represent three group data sets input change after image；

Fig. 3 (c) represents the testing result of the RFLICM algorithm of three group data sets；

Fig. 3 (d) represents the testing result of the SRM algorithm of three group data sets；

Fig. 3 (e) represents the testing result of the MDS-SRM algorithm of three group data sets；

Fig. 3 (f) represents the canonical reference figure of three group data sets.

As seen from Figure 3, compared to RFLICM algorithm and SRM algorithm, the inventive method can be effectively improved SAR image The detection degree of accuracy of middle region of variation.

Above description is only example of the present invention, it is clear that for those skilled in the art, is understanding After present invention and principle, all may carry out in form and details in the case of without departing substantially from the principle of the invention, structure Various corrections and change, but these corrections based on inventive concept and change still scope of the presently claimed invention it In.

Claims (5)

1. a SAR image change detection based on MDS-SRM Mixed cascading, including:

(1) the single-channel SAR image X of the two identical regions of width different time is inputted₁、X₂, and utilize the removal of non-local mean algorithm to be somebody's turn to do Coherent speckle noise in two width images, obtains two width image I after denoising₁And I₂；

(2) with the image I after denoising₁And I₂, construct log ratio figure DI₁With average ratio value figure DI₂；

(3) by log ratio disparity map DI of structure₁With average ratio value disparity map DI₂It is overlapped, obtains into a width dual pathways poor Different image DI；

(4) utilize statistical regions merge algorithm the pixel in dual pathways differential image DI is merged, complete disparity map from Pixel space, to the transformation of regional space, obtains merging image DT for the first time₁；

(5) multilamellar dynamic order statistical regions is utilized to merge algorithm to image DT after once merging₁In region carry out secondary conjunction And, the secondary obtaining having large area merges image DT₂:

(5a) traversal once merges image DT₁In adjacent region, determine that region is to number M；

(5b) similarity f (R', R) of zoning pair:

$f(R^{\′},R)=\underset{k\∈\{{\mathrm{DI}}_1,{\mathrm{DI}}_2\}}{max}\left|\sqrt{\[H_1-H_2\]\×S^{-1}\×{\[H_1-H_2\]}^T}\right|$

In formula, R' and R represents two adjacent regions, H₁And H₂Represent the characteristic vector in adjacent two regions respectively,WhereinRepresent the average pixel value in k passage Zone R territory, A (R_k) table Showing the statistic histogram matrix in Zone R territory, | R | represents the number of pixels of region R；S^-1Covariance square for adjacent area characteristic vector Battle array S (a, inverse matrix b)；

(5c) similarity f (R', R) being sorted from small to large, obtain merging order list Index (i), i is traversal merging order The pointer of list, if i=1, the region of selected similarity f (R, R') minima is to (R, R')；

(5d) judge whether (R, R') is merged by the region of minima according to below equation: if meetingThen merge, calculate adjacent area now and to number M and update adjacent area pair Similarity f of (R, R')_n+1(R, R'):

Wherein,Parameter Q represents Statistical Complexity, obtains segmentation result in various degree by regulation Q, R_|R|Indicate the region of the individual pixel of | R |, and have | | R_|R|||≤(|R|+1)^min(|R|,g), g=256, | R| represents what region was contained Number of pixels, constant| I | is the pixel number that image I comprises,Represent the average pixel value in k passage Zone R territory； f_n(R, R') represent n-th merge before R' region and the similarity in Zone R territory, φ andRepresent this similarity respectively and go through The weight of history similarity,Update merging order list Index (i), if i=1, selected f_n+1(R, R') minima Region, to (R, R'), judges whether (R, R') is merged by the region of minima again；

Otherwise, i=i+1, it is judged that whether i≤M sets up, if setting up, then takes adjacent area pair corresponding to Index (i), repeats step Suddenly (5d)；If being false, merging terminates, and obtains secondary and merges image DT₂；

(6) secondary is merged image DT₂In the gray average in each region carry out ascending sort and obtain merging order list, so The region that merging criterion in rear employing SRM algorithm is combined in order list carries out region merging technique successively, obtains final change Change testing result.

The SAR image change detection merged based on Mixed cascading statistical regions the most according to claim 1, Qi Zhongbu Suddenly (1) utilizes non-local mean algorithm to remove and be originally inputted the coherent speckle noise in two width images, according to being originally inputted Noise image v={v (i) | i ∈ I}, I are image pixel fields, to any one pixel i in image, use non-local mean to calculate Method denoising, the gray scale estimated value obtaining this point is:

$NL\left(v\right)\left(i\right)=\underset{j\∈I}{\Σ}\mathrm{\ω}(i,j)v\left(i\right)$

Wherein, weights ω (i, j) represents the similarity degree between pixel i and j, and meet condition 0≤ω (i, j)≤1 and

The SAR image change detection merged based on Mixed cascading statistical regions the most according to claim 1, Qi Zhongbu Suddenly with the image I after denoising in (2)₁、I₂, construct log ratio figure DI₁With average ratio value figure DI₂, carry out in accordance with the following steps:

(2a) with image I after denoising₁、I₂Middle coordinate is that (i, the pixel value of pixel j) pass through formulaStructure log ratio disparity map DI₁Middle coordinate is (i, pixel DI j)₁(i j), travels through denoising Rear image I₁、I₂In all of pixel, i.e. obtain log ratio disparity map DI₁；

(2b) with the image I after denoising₁、I₂Middle coordinate is that (i, the pixel value of pixel j) pass through formulaStructure average ratio value disparity map DI₂Middle coordinate is (i, pixel DI j)₂ (i, j), image I after traversal denoising₁、I₂In all of pixel, i.e. obtain average ratio value disparity map DI₂.Wherein μ₁(i, j), μ₂ (i j) is respectively image I₁, I₂In with coordinate for (i, the 2*2 field pixel average centered by pixel j).

The SAR image change detection merged based on Mixed cascading statistical regions the most according to claim 1, Qi Zhongbu Suddenly the pixel utilizing statistical regions to merge in the algorithm differential image DI to building in (4) merges, in accordance with the following steps Carry out:

(4a) every pair of neighbor in differential image DI is calculated similarity weight:

$f(p,p^{\′})=\underset{k\∈\{{\mathrm{DI}}_1,{\mathrm{DI}}_2\}}{max}|p_k-p_k^{\′}|$

Wherein p_kAnd p'_kPixel value for neighbor pair；

(4b) similarity weight is carried out from small to large ascending sort, and according to clooating sequence, selected pixels is to judging to be somebody's turn to do successively Whether pixel is to merging: if meetingThen merge, wherein,Represent k passage Zone R territory Average pixel value, Zone R territory is p_kPixel affiliated area,Parameter Q represents Statistical Complexity, logical Overregulate Q and obtain segmentation result in various degree, R_|R|Indicate the regional ensemble of the individual pixel of | R |, and have | | R_|R|||≤(n+1 )^min(|R|,g), g=256, constant| I | is the pixel number that image I comprises；Otherwise, nonjoinder；When each group of picture Element, to when all completing judge process, i.e. obtains merging image DT for the first time₁。

The SAR image change detection merged based on Mixed cascading statistical regions the most according to claim 1, Qi Zhongbu Suddenly covariance matrix S in (5) (a, b), is expressed as follows:

$S(a,b)=\frac{\{\[H_1\left(a\right)-\mathrm{\μ}\left(a\right)\]\×\[H_1\left(b\right)-\mathrm{\μ}\left(b\right)\]+\[H_2\left(a\right)-\mathrm{\μ}\left(a\right)\]\×\[H_2\left(b\right)-\mathrm{\μ}\left(b\right)\]\}}{2}$

Wherein a and b representsWithIn a Individual or the b characteristic vector,Represent the average pixel value in k passage Zone R territory, A (R_k) represent Zone R territory statistic histogram square Battle array, | R | represents the number of pixels of region R, and μ represents the average of adjacent area characteristic vector.