CN105869146A - Saliency fusion-based SAR image change detection method - Google Patents

Abstract

The invention discloses a saliency fusion-based SAR image change detection method, and aims at solving the problem that the SAR image change detection in the conventional method is easy to be influenced by speckle noise and is low in detection precision. The method comprises the following specific steps: (1) inputting an SAR image; (2) carrying out filtering; (3) calculating the logarithmic ratio of a pixel; (4) constructing a global saliency map; (5) calculating contrast ratio saliency values of different scales; (6) constructing a local saliency map; (7) fusing the local saliency map and the global saliency map; (8) carrying out fuzzy clustering; (9) outputting a change detection result. The method has the advantages of being good in noise influence robustness and high in detection precision for SAR image change detection.

Description

The SAR image change detection merged based on significance

Technical field

The invention belongs to technical field of image processing, further relate in the change detection techniques field of remote sensing images A kind of SAR image change detection merged based on significance.The present invention realizes the synthesis hole obtaining two width difference phases The detection in footpath radar SAR (Synthetic Aperture Radar) image change region, can cover and utilization, nature at atural object Disaster monitoring and assessment, urban planning, the field such as map rejuvenation is extensively applied.

Background technology

SAR image change-detection refers to that the two width SAR image utilizing different phase areal detect and analyze ground Situation of change.Owing to SAR technology has round-the-clock, the feature of round-the-clock work compared with ordinary optical remote sensing technology so that SAR image change-detection has a wide range of applications in national economy and Military Application field.In recent years, SAR image is utilized to carry out Change-detection is highly valued at international remote sensing fields, one of main direction of studying having become as image procossing.

Miriam Cha et al. is at its " Two-Stage Change Detection for Synthetic delivered Aperture Radar”(IEEE Transactions on Geoscience&Remote Sensing,2015,53(12): Paper 6547-6560) proposes the SAR image change detection of a kind of two-phase method.The first stage of the method is first The pixel pair with unequal variance that detection is caused by the target of large scale, obtains initial change-detection figure；To have phase again Etc. the pixel of variance or close variance to the input as second stage.Second stage uses Berger correlation estimation to detect SAR image has the region of minor variations, so that it is determined that final change-detection images.Two width images are only considered due to the method The variance of respective pixel, although higher pixel contrast can be obtained in initial detecting figure, but the method yet suffers from Weak point be that the method does not accounts for the spatial information of SAR image, so still suffering from the highest the asking of change-detection precision Topic.

Patent " method for detecting change of remote sensing image based on the image co-registration " (Shen that Xian Electronics Science and Technology University applies at it Please the patent No.: 201210414782.4, publication number: 102968790A) in propose a kind of remote sensing images based on image co-registration Change detecting method.First the method extracts difference disparity map and the ratio difference figure of two width images before and after change, and extracts difference The variance matrix of value image；Use KI thresholding method, obtain optimal threshold T；According to optimal threshold T and variance matrix by differential chart Merge with ratio figure, final disparity map D after being merged；The gray value of final disparity map D of gained is carried out classification number is 2 Fuzzy C-means clustering segmentation, using a class bigger for cluster centre value as change class, another kind of for non-changing class, obtain Change-detection result.Although the method has preferably detection for the disparity map with conspicuous object, but the method is still deposited Weak point be for having the disparity map Detection results of unconspicuous variation targets bad, and can not effectively to suppress The impact of speckle noise.

The patent " SAR image change detection decomposed based on low-rank matrix " that Xian Electronics Science and Technology University applies at it (number of applying for a patent: 201210193347.3, publication number: 102722892A) propose a kind of based on low-rank matrix decomposition SAR image change detection.First two width SAR image to be measured are carried out dropping speckle pretreatment by the method, are more smoothed SAR image；Then the log ratio of two width images after structure fall speckle, then carries out low-rank sparse decomposition by log ratio, Obtain the low-rank part of log ratio and sparse part；By row, sparse Partial Transformation is become sparse matrix again；Finally calculate by K average The sparse matrix obtained is clustered by method, obtains final change-detection result.Although the method can accurately detect To region of variation, but the weak point that the method yet suffers from is, owing to not accounting for detailed information and space neighborhood information, So yet suffering from the problem that accuracy of detection is unsatisfactory.

Summary of the invention

It is an object of the invention to overcome above-mentioned the deficiencies in the prior art, it is proposed that a kind of SAR merged based on significance Image change detection method.The present invention with existing in compared with other SAR image change detection techniques, it is possible to effectively suppression speckle is made an uproar Sound, and improve the accuracy of detection of SAR image change-detection.

The present invention realizes the thinking of above-mentioned purpose: first scheme the SAR of the different phase areals registrated of input As carrying out non-local mean filtering, more filtered SAR image is carried out log ratio operation, then the notable figure of the structure overall situation, Filtered SAR image calculates the contrast saliency value of different scale simultaneously, and then the notable figure in structure local, notable by the overall situation The notable figure fusion of figure and local obtains significance fusion figure, finally significance fusion figure is carried out fuzzy C-means clustering and is become Change testing result.

The concrete steps that the present invention realizes include the following:

(1) input SAR image:

SAR image I that two width of input areal difference phase registrate, correct₁And I₂；

(2) filtering:

Use the non-local mean wave filter two width SAR image I to input₁And I₂It is filtered respectively, obtains filtered SAR image X₁And X₂；

(3) the log ratio value of calculating pixel:

To filtered SAR image X₁And X₂The gray level taking respective pixel carries out log ratio operation, obtains log ratio Disparity map D_L；

(4) according to the following formula, the overall saliency value of the pixel of calculating log ratio disparity map:

$D\left(I_k\right)=\underset{\∀I_i\∈D_L}{\Σ}||I_k-I_i||$

Wherein, D (I_k) represent the overall saliency value of pixel, I in log ratio disparity map_kRepresent in log ratio disparity map The gray value of kth pixel, k=0,1,2 ..., 255, ∑ represents sum operation,Representing optional sign, ∈ represents and belongs to symbol Number, I_iRepresent in log ratio disparity map except I_kOutside the gray value of the pixel, i=0,1,2 ..., 255, D_LIt is right to represent Number ratio difference figure, | | | | represent the distance operation taking gray value；

(5) the contrast saliency value of calculating different scale:

(5a) respectively to filtered SAR image X₁And X₂Take 1 × 1 pixel, use pixel contrast method, construct 1 × 1 The contrast of pixel dimension is significantly schemed；

(5b) with the sliding window of 3 × 3 pixels respectively to filtered image X₁And X₂Take block；

(5c) respectively the pixel in all of piece is all pulled into string by row, obtain three-dimensional matrice L₁And L₂；

(5d) vector contrast method is used, by three-dimensional matrice L₁And L₂The contrast constructing 3 × 3 pixel dimension is notable Value；

(5e) with the sliding window of 5 × 5 pixels respectively to filtered image X₁And X₂Take block；

(5c) respectively the pixel in all of piece is all pulled into string by row, obtain three-dimensional matrice L₃And L₄；

(5g) vector contrast method is used, by three-dimensional matrice L₃And L₄The contrast constructing 5 × 5 pixel dimension is notable Value；

6) using the weighted mean method of following formula, the contrast merging different scale is significantly schemed, and obtains local and significantly schemes:

S'=α₀·S₀+α₁·S₁+α₂·S₂

Wherein, S' represents the notable figure in local, S₀Represent that the contrast of 1 × 1 pixel dimension is significantly schemed, S₁Represent 3 × 3 pixels The notable figure of yardstick, S₂Represent that the contrast of 5 × 5 pixel dimension is significantly schemed, α₀、α₁、α₂Represent weight coefficient, α₀+α₁+α₂=1, 0≤α₀、α₁、α₂≤1；

(7) according to the following formula, amalgamation of global is significantly schemed and the notable figure in local, obtains significance fusion figure:

S (i, j)=D (i, j) exp (-S'(i, j))

Wherein, S represents significance fusion figure, and D represents the notable figure of the overall situation, and S' represents the notable figure in local, and exp represents that index is grasped Make symbol, i and j represent respectively the pixel of same position in the notable figure of the notable figure of significance fusion figure, the overall situation and local row and Row coordinate figure；

(8) fuzzy clustering:

(8a) use fuzzy C-means clustering method, significance fusion figure is clustered, obtain in significance fusion figure every Individual pixel be under the jurisdiction of non-changing class and change class be subordinate to angle value；

(8b) angle value that is subordinate to of non-changing class is more than the pixel being subordinate to angle value of change class, is judged to the ownership of as non-changing class picture Element；The angle value that is subordinate to of non-changing class is less than the pixel being subordinate to angle value of change class, is judged to the ownership of as change class pixel；Obtain SAR figure The change-detection result of picture；

(9) exporting change testing result.

The present invention compared with prior art has the advantage that

First, owing to present invention employing combines the method that the notable figure of contrast of different scale calculates the notable figure in local, Choose the image block of 1 × 1,3 × 3 and 5 × 5 three kinds of different scales respectively to calculate the contrast of center pixel, thus combine The neighborhood information of image, overcomes and does not accounts for neighborhood information in prior art, causes SAR image change-detection easily to be made an uproar by speckle The impact of sound so that the present invention is conducive to suppressing coherent speckle noise, and then decreases the false drop rate of change-detection.

Second, owing to the present invention uses the method that significance merges, the overall situation is significantly schemed and the notable figure in local merges, On the basis of the notable figure of the overall situation highlights variation targets, add the characteristic of local, overcome in prior art and SAR image is become Change accuracy of detection the highest, the problem to noise robustness difference so that the present invention detection essence to improve SAR image change-detection Degree, the robustness of enhanced SAR Image Change Detection.

Accompanying drawing explanation

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

Fig. 2 is that three groups of real SAR image data using of emulation experiment and corresponding change-detection are with reference to figure；

Fig. 3 is the change-detection result figure of the emulation experiment to Ottawa area SAR image；

Fig. 4 is the change-detection result figure of the emulation experiment to area, Sardinia SAR image；

Fig. 5 is the change-detection result figure of the emulation experiment to the Yellow River estuary SAR image.

Detailed description of the invention

The present invention will be further described below in conjunction with the accompanying drawings.

With reference to Fig. 1, what the present invention realized specifically comprises the following steps that

Step 1, inputs SAR image.

SAR image I that two width of input areal difference phase registrate, correct₁And I₂。

Step 2, filtering.

Use the non-local mean wave filter two width SAR image I to input₁And I₂It is filtered respectively, obtains filtered SAR image X₁And X₂。

Step 3, calculates the log ratio value of pixel.

To filtered SAR image X₁And X₂The gray level taking respective pixel carries out log ratio operation, obtains log ratio Disparity map D_L。

Step 4, according to the following formula, the overall saliency value of the pixel of calculating log ratio disparity map:

$D\left(I_k\right)=\underset{\∀I_i\∈D_L}{\Σ}||I_k-I_i||$

Wherein, D (I_k) represent the overall saliency value of pixel, I in log ratio disparity map_kRepresent in log ratio disparity map The gray value of kth pixel, k=0,1,2 ..., 255, ∑ represents sum operation,Representing optional sign, ∈ represents and belongs to symbol Number, I_iRepresent in log ratio disparity map except I_kOutside the gray value of the pixel, i=0,1,2 ..., 255, D_LIt is right to represent Number ratio difference figure, ‖ ‖ represents the distance operation taking gray value；

Step 5, the contrast calculating different scale is significantly schemed.

Respectively to filtered SAR image X₁And X₂Take 1 × 1 pixel, use pixel contrast method, construct 1 × 1 pixel The contrast of yardstick is significantly schemed.

The concrete operation step of pixel contrast method is as follows:

The first step, according to the following formula, the contrast saliency value of the notable figure of contrast of calculating 1 × 1 pixel dimension:

$S_0(x_1,x_2)=\frac{x_1\·x_2}{{||x_1||}^2+{||x_2||}^2-x_1\·x_2}$

Wherein, S₀(x₁,x₂) represent the contrast saliency value of the notable figure of contrast of 1 × 1 pixel dimension, x₁And x₂Table respectively Show the index of current contrast saliency value coordinate position, | | ‖ represents that Euclidean distance operates；

Second step, according to the following formula, the contrast of structure 1 × 1 pixel dimension is significantly schemed:

$S_0=\left[\begin{array}{cccc}{S}_0(1,1)& S_0(1,2)& ...& S_0(1,N)\\ S_0(2,1)& ...& ...& S_0(2,N)\\ ...& ...& S_0(x_1,x_2)& ...\\ S_0(M,1)& S_0(M,2)& ...& S_0(M,N)\end{array}\right]$

Wherein, S₀Represent that the contrast of 1 × 1 pixel dimension is significantly schemed, S₀(x₁,x₂) represent 1 × 1 pixel dimension contrast Spend the contrast saliency value of notable figure, x₁And x₂Represent the position coordinates index of current contrast saliency value, M and N table respectively respectively Show line number and the columns of the notable figure of contrast.

With the sliding window of 3 × 3 pixels respectively to filtered image X₁And X₂Take block.

Respectively the pixel in all of piece is all pulled into string by row, obtain three-dimensional matrice L₁And L₂。

Use vector contrast method, by three-dimensional matrice L₁And L₂Construct the contrast saliency value of 3 × 3 pixel dimension.

The concrete operation step of vector contrast method is as follows:

The first step, according to the following formula, the contrast saliency value of the notable figure of contrast of calculating 3 × 3 pixel dimension:

$S_1(l_1,l_2)=\frac{{l_1}^T\·l_2}{{||l_1||}^2+{||l_2||}^2-{l_1}^T\·l_2}$

Wherein, S₁(l₁,l₂) represent the contrast saliency value of the notable figure of contrast of 3 × 3 pixel dimension, l₁And l₂Table respectively Show L₁And L₂In column vector, T represents that transposition operates, | | | | represent Euclidean distance operation；

Second step, according to the following formula, the contrast of structure 3 × 3 pixel dimension is significantly schemed:

$S_1=\left[\begin{array}{cccc}{S}_1(1,1)& S_1(1,2)& ...& S_1(1,N)\\ S_1(2,1)& ...& ...& S_1(2,N)\\ ...& ...& S_1(l_1,l_2)& ...\\ S_1(M,1)& S_1(M,2)& ...& S_1(M,N)\end{array}\right]$

Wherein, S₁Represent that the contrast under 3 × 3 pixel dimension is significantly schemed, S₁(l₁,l₂) represent the right of 3 × 3 pixel dimension The contrast saliency value of figure more notable than degree, l₁And l₂Represent L respectively₁And L₂In column vector, M and N represents 3 × 3 pixel chis respectively The line number of the notable figure of contrast of degree and columns.

With the sliding window of 5 × 5 pixels respectively to filtered image X₁And X₂Take block.

Respectively the pixel in all of piece is all pulled into string by row, obtain three-dimensional matrice L₃And L₄。

Use vector contrast method, by three-dimensional matrice L₃And L₄Construct the contrast saliency value of 5 × 5 pixel dimension.

The concrete operation step of vector contrast method is as follows:

The first step, according to the following formula, the contrast saliency value of the notable figure of contrast of calculating 5 × 5 pixel dimension:

$S_2(l_3,l_4)=\frac{{l_3}^T\·l_4}{{||l_3||}^2+{||l_4||}^2-{l_3}^T\·l_4}$

Wherein, S₂(l₃,l₄) represent the contrast saliency value of the notable figure of contrast of 3 × 3 pixel dimension, l₃And l₄Table respectively Show L₁And L₂In column vector, T represents that transposition operates, | | | | represent Euclidean distance operation；

Second step, according to the following formula, the contrast under structure 5 × 5 pixel dimension is significantly schemed:

$S_2=\left[\begin{array}{cccc}{S}_2(1,1)& S_2(1,2)& ...& S_2(1,N)\\ S_2(2,1)& ...& ...& S_2(2,N)\\ ...& ...& S_2(l_3,l_4)& ...\\ S_2(M,1)& S_2(M,2)& ...& S_2(M,N)\end{array}\right]$

Wherein, S₂Represent that the contrast under 3 × 3 pixel dimension is significantly schemed, S₂(l₃,l₄) represent the right of 3 × 3 pixel dimension The contrast saliency value of figure more notable than degree, l₃And l₄Represent L respectively₁And L₂In column vector, M and N represents 3 × 3 pixel chis respectively The line number of the notable figure of contrast of degree and columns.

Step 6, uses the weighted mean method of following formula, and the contrast merging different scale is significantly schemed, and obtains local and significantly schemes:

S'=α₀S₀+α₁·S₁+α₂·S₂

Wherein, S' represents the notable figure in local, S₀Represent that the contrast of 1 × 1 pixel dimension is significantly schemed, S₁Represent 3 × 3 pixels The notable figure of yardstick, S₂Represent that the contrast of 5 × 5 pixel dimension is significantly schemed, α₀、α₁、α₂Represent weight coefficient, α₀+α₁+α₂=1, 0≤α₀、α₁、α₂≤1。

Step 7, according to the following formula, amalgamation of global is significantly schemed and the notable figure in local, obtains significance fusion figure:

S (i, j)=D (i, j) exp (-S'(i, j))

Wherein, S represents significance fusion figure, and D represents the notable figure of the overall situation, and S' represents the notable figure in local, and exp represents that index is grasped Make symbol, i and j represent respectively the pixel of same position in the notable figure of the notable figure of significance fusion figure, the overall situation and local row and Row coordinate figure.

Step 8, fuzzy clustering.

Use fuzzy C-means clustering method, significance fusion figure is clustered, obtains each picture in significance fusion figure Element be under the jurisdiction of non-changing class and change class be subordinate to angle value.

The angle value that is subordinate to of non-changing class is more than the pixel being subordinate to angle value of change class, is judged to the ownership of as non-changing class pixel；Will The angle value that is subordinate to of non-changing class is less than the pixel being subordinate to angle value of change class, is judged to the ownership of as change class pixel；Obtain the change of SAR image Change testing result.

Step 9, exporting change testing result.

Below in conjunction with emulation experiment, the effect of the present invention is described further:

1, simulated conditions:

The emulation experiment of the present invention is Inter (R) Core (TM) i5-3470CPU, internal memory 4GB in dominant frequency 3.2GHz Carry out under the software environment of hardware environment and MATLAB R2015b.

The emulation experiment of the present invention employs three groups of real SAR image data and corresponding change-detection with reference to figure, as Shown in Fig. 2.

First group of real SAR image data and corresponding change-detection that emulation experiment of the present invention uses with reference to figure are The SAR image in Ottawa area, image size is 290 × 350.Wherein, Fig. 2 (a) is the Ottawa area in May, 1997 SAR image, Fig. 2 (b) is the SAR image in the Ottawa area in August, 1997, and Fig. 2 (c) is that Ottawa area changes inspection accordingly Survey with reference to figure.

Second group of real SAR image data and corresponding change-detection that emulation experiment of the present invention uses with reference to figure are The SAR image in Sardinia area, image size is 300 × 412.Wherein, Fig. 2 (d) is the Sardinia area of nineteen ninety-five JIUYUE SAR image, Fig. 2 (e) is the SAR image in the Sardinia area in July, 1996, and Fig. 2 (f) is that Sardinia area is corresponding Change-detection is with reference to figure.

The 3rd group of real SAR image data and corresponding change-detection that emulation experiment of the present invention uses are yellow with reference to figure The SAR image in estuary area, river, image size is 291 × 306.Wherein, Fig. 2 (g) is the estuary ground, the Yellow River in June, 2008 The SAR image in district, Fig. 2 (h) is the SAR image in the estuary area, the Yellow River in June, 2009, and Fig. 2 (i) is estuary area, the Yellow River Corresponding change-detection is with reference to figure.

The simulation parameter that emulation experiment of the present invention is used is as follows:

Calculate missing inspection number: the number of pixels in the region that changes in statistical experiment result figure, and with reference to region of variation in figure Number of pixels contrast, with reference to figure changes, experimental result picture being detected as unchanged number of pixels, claim For missing inspection number FN.

Calculate flase drop number: the number of pixels in the region that do not changes in statistical experiment result figure, and with reference to unchanged in figure The number of pixels changing region contrasts, being detected as the pixel of change in experimental result picture with reference to not changing in figure Number, referred to as flase drop number FP.

Total error number/the total pixel number of accuracy PCC:PCC=1-.

Weigh testing result figure with reference to figure conforming Kappa coefficient: Kappa=(PCC-PRE)/(1-PRE) wherein, Accuracy PCC represents actual concordance rate, the concordance rate of PRE representation theory.

2, emulation content and interpretation of result:

The emulation experiment of the present invention use four kinds of prior aries (LK method based on log ratio and K mean cluster, based on The CDI-K method that disparity map merges, fuzzy clustering method based on Markov random field, SM side based on significance measure Method) and use the inventive method, respectively Ottawa area, Sardinia area and estuary area, the Yellow River SAR image are carried out The testing result of change-detection contrasts.

Fig. 3 is the present invention emulation experiment to Ottawa area, and wherein, Fig. 3 (a) is the SAR image change of Ottawa area The reference picture of detection；Fig. 3 (b) is the simulation result figure using LK method based on log ratio and K mean cluster；Fig. 3 (c) It it is the simulation result figure using the CDI-K method merged based on disparity map；Fig. 3 (d) is to use based on Markov random field The simulation result figure of fuzzy clustering method MRFFCM；Fig. 3 (e) is the simulation result using SM method based on significance measure Figure；Fig. 3 (f) is the simulation result figure using the inventive method.

From the visual effect of Fig. 3 it can be seen that use the noise spot the change-detection result figure of the present invention than existing skill Four kinds of methods of art are few, and the edge clear of change-detection result figure.

Table 1 be the emulation experiment of the present invention use four kinds of prior aries and the inventive method to missing inspection number FP, flase drop number FN, Accuracy PCC and Kappa coefficient are added up.In table " LK " represent use LK method based on log ratio and K mean cluster, In table, " CDI-K " represents that using " MRFFCM " in CDI-K method based on disparity map fusion, table to represent uses based on markov In the MRFFCM method of the fuzzy clustering of random field and table, " SM " represents that SM method based on significance measure uses with the present invention Significance fusion method.From table 1 it follows that change-detection accuracy PCC of the inventive method and Kappa coefficient, all Relatively low higher than other four kinds of control methods, missing inspection number FP and flase drop number FN all ratios, Detection results is good.

Table 1 Ottawa area SAR image change-detection result

	Missing inspection number FP	Flase drop number FN	Accuracy PCC	Kappa coefficient
					LK	1219	2198	0.9663	0.8703
CDI-K	2044	367	0.9762	0.9068
					MRFFCM	790	1711	0.9768	0.9151
SM	5862	910	0.9333	0.7135
					Significance merges	1006	634	0.9838	0.9399

Fig. 4 is the present invention emulation experiment to Sardinia area, and wherein Fig. 4 (a) is that Sardinia area SAR image becomes Change the reference picture of detection；Fig. 4 (b) is the simulation result figure using LK method based on log ratio and K mean cluster；Fig. 4 C () is the simulation result figure using the CDI-K method merged based on disparity map；Fig. 4 (d) is to use based on Markov random field The simulation result figure of fuzzy clustering method MRFFCM；Fig. 4 (e) is the simulation result using SM method based on significance measure Figure；Fig. 4 (f) is the simulation result figure using the inventive method.

From the visual effect of Fig. 4 it can be seen that the change-detection result figure of the present invention and four kinds of methods than prior art Compare the detection of fine edge more effectively, and in the change-detection result figure of the present invention, miscellaneous point is less.

Table 2 Sardinia area SAR image change-detection result

	Missing inspection number FP	Flase drop number FN	Accuracy PCC	Kappa coefficient
					LK	836	1229	0.9833	0.8591
CDI-K	913	1083	0.9839	0.8620
					MRFFCM	947	1523	0.9800	0.8333
SM	1967	544	0.9797	0.8078
					Significance merges	823	1126	0.9842	0.8663

Table 2 be the emulation experiment of the present invention use four kinds of prior aries and the inventive method to missing inspection number FP, flase drop number FN, Accuracy PCC and Kappa coefficient are added up.In table " LK " represent use LK method based on log ratio and K mean cluster, In table, " CDI-K " represents that using " MRFFCM " in CDI-K method based on disparity map fusion, table to represent uses based on markov In the MRFFCM method of the fuzzy clustering of random field and table, " SM " represents that SM method based on significance measure uses with the present invention Significance fusion method.As can be seen from Table 2, the missing inspection number FP of testing result of the present invention will than other four kinds of detection methods Low, and flase drop number FN is also above control methods at relatively low level, accuracy PCC and Kappa coefficient, it is seen that the inventive method Improve the effect of SAR image change-detection, and more preferable to the robustness of noise.

Fig. 5 is the present invention emulation experiment to the Yellow River estuary area, and wherein Fig. 5 (a) is estuary area, the Yellow River SAR figure Reference picture as change-detection；Fig. 5 (b) is the simulation result figure using LK method based on log ratio and K mean cluster； Fig. 5 (c) is the simulation result figure using the CDI-K method merged based on disparity map；Fig. 5 (d) be use based on markov with The simulation result figure of the fuzzy clustering method MRFFCM on airport；Fig. 5 (e) is the emulation using SM method based on significance measure Result figure；Fig. 5 (f) is the simulation result figure using the inventive method.

From the visual effect of Fig. 5 it can be seen that the change-detection result figure of the present invention connects most compared with existing four kinds of methods Near with reference to figure, the miscellaneous point of the change-detection result figure of the present invention is minimum, and visual effect is best, and change can effectively be detected Region.

Estuary area, table 3 the Yellow River SAR image change-detection result

	Missing inspection number FP	Flase drop number FN	Accuracy PCC	Kappa coefficient
					LK	440	10170	0.8808	0.4263
CDI-K	803	1816	0.9706	0.7577
					MRFFCM	472	1853	0.9738	0.7903
SM	102	5209	0.9404	0.6317
					Significance merges	712	350	0.9881	0.8893

Table 3 be the emulation experiment of the present invention use four kinds of prior aries and the inventive method to missing inspection number FP, flase drop number FN, Accuracy PCC and Kappa coefficient are added up.In table " LK " represent use LK method based on log ratio and K mean cluster, In table, " CDI-K " represents that using " MRFFCM " in CDI-K method based on disparity map fusion, table to represent uses based on markov In the MRFFCM method of the fuzzy clustering of random field and table, " SM " represents that SM method based on significance measure uses with the present invention Significance fusion method.As can be seen from Table 3, the flase drop number FN of testing result of the present invention is well below other four kinds contrast inspections Survey method, controls the quantity of missing inspection number FP simultaneously, and accuracy PCC and Kappa coefficient is significantly larger than other four kinds contrasts especially Detection method.

Claims (4)

1. the SAR image change detection merged based on significance, comprises the steps:

(1) input SAR image:

SAR image I that two width of input areal difference phase registrate, correct₁And I₂；

(2) filtering:

Use the non-local mean wave filter two width SAR image I to input₁And I₂It is filtered respectively, obtains filtered SAR Image X₁And X₂；

(3) the log ratio value of calculating pixel:

To filtered SAR image X₁And X₂The gray level taking respective pixel carries out log ratio operation, obtains log ratio difference Figure D_L；

(4) according to the following formula, the overall saliency value of the pixel of calculating log ratio disparity map:

$D\left(I_k\right)=\underset{\∀I_i\∈D_L}{\Σ}\left|\right|I_k-I_i\left|\right|$

Wherein, D (I_k) represent the overall saliency value of pixel, I in log ratio disparity map_kRepresent kth in log ratio disparity map The gray value of pixel, k=0,1,2 ..., 255, ∑ represents sum operation,Representing optional sign, ∈ represents and belongs to symbol, I_i Represent in log ratio disparity map except I_kOutside the gray value of the pixel, i=0,1,2 ..., 255, D_LRepresent log ratio Disparity map, | | | | represent the distance operation taking gray value；

(5) the contrast saliency value of calculating different scale:

(5a) respectively to filtered SAR image X₁And X₂Take 1 × 1 pixel, use pixel contrast method, construct 1 × 1 pixel The contrast of yardstick is significantly schemed；

(5b) with the sliding window of 3 × 3 pixels respectively to filtered image X₁And X₂Take block；

(5c) respectively the pixel in all of piece is all pulled into string by row, obtain three-dimensional matrice L₁And L₂；

(5d) vector contrast method is used, by three-dimensional matrice L₁And L₂Construct the contrast saliency value of 3 × 3 pixel dimension；

(5e) with the sliding window of 5 × 5 pixels respectively to filtered image X₁And X₂Take block；

(5f) respectively the pixel in all of piece is all pulled into string by row, obtain three-dimensional matrice L₃And L₄；

(5g) vector contrast method is used, by three-dimensional matrice L₃And L₄Construct the contrast saliency value of 5 × 5 pixel dimension；

(6) using the weighted mean method of following formula, the contrast merging different scale is significantly schemed, and obtains local and significantly schemes:

S'=α₀·S₀+α₁·S₁+α₂·S₂

Wherein, S' represents the notable figure in local, S₀Represent that the contrast of 1 × 1 pixel dimension is significantly schemed, S₁Represent 3 × 3 pixel dimension Notable figure, S₂Represent that the contrast of 5 × 5 pixel dimension is significantly schemed, α₀、α₁、α₂Represent weight coefficient, α₀+α₁+α₂=1,0≤ α₀、α₁、α₂≤1；

(7) according to the following formula, amalgamation of global is significantly schemed and the notable figure in local, obtains significance fusion figure:

S (i, j)=D (i, j) exp (-S'(i, j))

Wherein, S represents significance fusion figure, and D represents the notable figure of the overall situation, and S' represents the notable figure in local, and exp represents that index operation accords with Number, i and j represents that in the notable figure of the notable figure of significance fusion figure, the overall situation and local, the row and column of the pixel of same position is sat respectively Scale value；

(8) fuzzy clustering:

(8a) use fuzzy C-means clustering method, significance fusion figure is clustered, obtains each picture in significance fusion figure Element be under the jurisdiction of non-changing class and change class be subordinate to angle value；

(8b) angle value that is subordinate to of non-changing class is more than the pixel being subordinate to angle value of change class, is judged to the ownership of as non-changing class pixel；Will The angle value that is subordinate to of non-changing class is less than the pixel being subordinate to angle value of change class, is judged to the ownership of as change class pixel；Obtain the change of SAR image Change testing result；

(9) exporting change testing result.

The SAR image change detection merged based on significance the most according to claim 1, it is characterised in that: step (5a) specifically comprising the following steps that of pixel contrast method described in

The first step, according to the following formula, the contrast saliency value of the notable figure of contrast of calculating 1 × 1 pixel dimension:

$S_0(x_1,x_2)=\frac{x_1\·x_2}{\left|\right|x_1|{|}^2+|\left|x_2\right|{|}^2-x_1\·x_2}$

Wherein, S₀(x₁,x₂) represent the contrast saliency value of the notable figure of contrast of 1 × 1 pixel dimension, x₁And x₂Represent respectively and work as The index of front contrast saliency value coordinate position, | | | | represent Euclidean distance operation；

Second step, according to the following formula, the contrast of structure 1 × 1 pixel dimension is significantly schemed:

$S_0=\left[\begin{array}{cccc}{S}_0(1,1)& S_0(1,2)& ...& S_0(1,N)\\ S_0(2,1)& ...& ...& S_0(2,N)\\ ...& ...& S_0(x_1,x_2)& ...\\ S_0(M,1)& S_0(M,2)& ...& S_0(M,N)\end{array}\right]$

Wherein, S₀Represent that the contrast of 1 × 1 pixel dimension is significantly schemed, S₀(x₁,x₂) represent that the contrast of 1 × 1 pixel dimension is notable The contrast saliency value of figure, x₁And x₂Representing the position coordinates index of current contrast saliency value respectively, M and N represents contrast respectively Spend line number and the columns of notable figure.

The SAR image change detection merged based on significance the most according to claim 1, it is characterised in that: step (5d) described in, vector contrast method specifically comprises the following steps that

The first step, according to the following formula, the contrast saliency value of the notable figure of contrast of calculating 3 × 3 pixel dimension:

$S_1(l_1,l_2)=\frac{{l_1}^T\·l_2}{\left|\right|l_1|{|}^2+|\left|l_2\right|{|}^2-{l_1}^T\·l_2}$

Wherein, S₁(l₁,l₂) represent the contrast saliency value of the notable figure of contrast of 3 × 3 pixel dimension, l₁And l₂Represent L respectively₁ And L₂In column vector, T represents that transposition operates, | | | | represent Euclidean distance operation；

Second step, according to the following formula, the contrast of structure 3 × 3 pixel dimension is significantly schemed:

$S_1=\left[\begin{array}{cccc}{S}_1(1,1)& S_1(1,2)& ...& S_1(1,N)\\ S_1(2,1)& ...& ...& S_1(2,N)\\ ...& ...& S_1(l_1,l_2)& ...\\ S_1(M,1)& S_1(M,2)& ...& S_1(M,N)\end{array}\right]$

Wherein, S₁Represent that the contrast under 3 × 3 pixel dimension is significantly schemed, S₁(l₁,l₂) represent that the contrast of 3 × 3 pixel dimension shows Write the contrast saliency value of figure, l₁And l₂Represent L respectively₁And L₂In column vector, M and N represents the right of 3 × 3 pixel dimension respectively The line number of figure more notable than degree and columns.

The SAR image change detection merged based on significance the most according to claim 1, it is characterised in that: step (5g) described in, vector contrast method specifically comprises the following steps that

The first step, according to the following formula, the contrast saliency value of the notable figure of contrast of calculating 5 × 5 pixel dimension:

$S_2(l_3,l_4)=\frac{{l_3}^T\·l_4}{\left|\right|l_3|{|}^2+|\left|l_4\right|{|}^2-{l_3}^T\·l_4}$

Wherein, S₂(l₃,l₄) represent the contrast saliency value of the notable figure of contrast of 3 × 3 pixel dimension, l₃And l₄Represent L respectively₁ And L₂In column vector, T represents that transposition operates, | | | | represent Euclidean distance operation；

Second step, according to the following formula, the contrast under structure 5 × 5 pixel dimension is significantly schemed:

$S_2=\left[\begin{array}{cccc}{S}_2(1,1)& S_2(1,2)& ...& S_2(1,N)\\ S_2(2,1)& ...& ...& S_2(2,N)\\ ...& ...& S_2(l_3,l_4)& ...\\ S_2(M,1)& S_2(M,2)& ...& S_2(M,N)\end{array}\right]$

Wherein, S₂Represent that the contrast under 3 × 3 pixel dimension is significantly schemed, S₂(l₃,l₄) represent that the contrast of 3 × 3 pixel dimension shows Write the contrast saliency value of figure, l₃And l₄Represent L respectively₁And L₂In column vector, M and N represents the right of 3 × 3 pixel dimension respectively The line number of figure more notable than degree and columns.