CN105844279A - Depth learning and SIFT feature-based SAR image change detection method - Google Patents

Abstract

The invention discloses a depth learning and SIFT feature-based SAR image change detection method. The main objective of the invention is to solve the problem of low accuracy of change detection results which is caused by the sensitivity of a method in the prior art to speckle noises of an SAR image. The method of the invention includes the following steps that: (1) SAR images are read in; (2) normalization is carried out; (3) training features are constructed; (4) a deep neural network is trained; (5) logarithmic ratio operation is performed on the two read-in SAR images, so that a logarithmic ratio difference image is obtained; (6) the neighborhood feature sample matrix of the logarithmic ratio difference image is constructed; (7) the logarithmic ratio difference image is detected; and (8) a change detection result graph is outputted. According to the depth learning and SIFT feature-based SAR image change detection method of the invention, the stability of SIFT features to SAR image speckle noises is fully utilized, and therefore, the influence of the SAR image speckle noises can be eliminated, and the accuracy rate of SAR image change detection can be improved.

Description

Based on degree of depth study and the SAR image change detection of SIFT feature

Technical field

The invention belongs to technical field of image processing, further relate to a kind of base in Remote Sensing Imagery Change Detection technical field In degree of depth study and the conjunction of shift-invariant operator conversion SIFT (Scale-invariant Feature Transform, SIFT) feature Become aperture radar SAR (SyntheticAperture Radar, SAR) image change detection method.The present invention can be used for same The region of variation of one SAR image of period from different places detects.

Background technology

Radar imaging technology is to grow up the fifties in last century, and this technology has been achieved for the development of great-leap-forward so far. This technology is widely used at aspects such as military affairs, agricultural, ocean, geology, mappings at present.

Synthetic aperture radar, as a kind of active microwave sensor, has that resolution is high, round-the-clock, round-the-clock work and wearing The feature that power is strong thoroughly, therefore SAR is not affected by the correlated condition such as atmospheric condition and cloud cover.SAR accident, The detection of natural disaster has, with aspects such as assessments, the advantage that remote sensing image is incomparable.This technology has been widely used In fields such as military, agricultural and scientific researches.SAR image change-detection is by the two width SAR to areal different times Image is analyzed, thus obtains the change information of atural object or target.SAR image change detection techniques is answered many at present Play the most important effect in, as detection and the assessment of natural disaster, the monitoring of environment, city management planning and Military investigation etc..

At present about method rough classification two class of SAR image change-detection: (1) classification and predicting method, classification ratio after being also referred to as Relatively method.First the method carries out independent sorting respectively to two width original images, then the image having classified two width is carried out pixel by pixel Compare, and then obtain final change-detection result；(2) classification method after comparing, also referred to as disparity map sorting technique.The party is first Obtained the disparity map of two width images by differential technique, ratio method or log ratio method etc., then this disparity map is analyzed. Wherein Equations of The Second Kind method is the most popular method, and therefore the structure of multidate SAR image disparity map is the most crucial.Mesh Front with relatively more methods include image differential technique, image ratio method and log ratio method etc., then to drawing more afterwards Disparity map is further analyzed, and including conversion, probability distribution etc., obtains final change-detection result.Equations of The Second Kind method Simple, intuitive, the change details obtained is more significantly.

" SAR image change detection based on gauss hybrid models " that Nanjing electronic technology institute is delivered at it is (existing For radar, 2014,36 (9): 34-36) paper proposes a kind of SAR image change-detection based on gauss hybrid models Method.The method first passes through the disparity map of log ratio method structure SAR image, then analyzes the probability statistics of disparity map, Probability distribution is fitted by the mixed model using four Gaussian functions again, obtains final change-detection result.The method The weak point existed is, first, the method is sensitive to the speckle noise of SAR image, causes the precision of final change-detection The highest.Secondly as the probability distribution of different SAR image is different, thus the method is for the change of different SAR image Detection robustness is the highest.

The patent " the SAR image change-detection based on without supervision deep neural network " that Xian Electronics Science and Technology University applies at it (number of patent application: 201410818305.3, publication number: CN104517124A) propose a kind of based on without the supervision degree of depth The SAR image change-detection of neutral net.First the SAR image of areal difference phase is combined by the method FCM classification obtains coarse change-detection result；Then select, according to initial change-detection result, the non-noise that probability is big Point is as the training sample of deep neural network；It is input in deep neural network be trained by these samples again；Finally will Two images to be detected input in the deep neural network trained and obtain final change-detection result.The method exists Weak point is, the method does not considers the impact of SAR image speckle noise joint classification when, causes selected instruction Practice sample point reliability inadequate.

Summary of the invention

It is an object of the invention to the shortcoming overcoming above-mentioned prior art, it is proposed that a kind of based on degree of depth study and SIFT feature SAR image change detection, to realize the accurate detection to SAR image change region.The method combines the degree of depth Study and two kinds of methods of SIFT feature, directly train deep neural network by SIFT feature, owing to SIFT feature is permissible The local feature of reflection image, is respectively provided with invariance to image rotation, scaling and brightness flop, and to affine change Change and also keep a certain degree of stability with noise, thus can be as the reliable training sample of deep neural network.The method Thinking is simply clear and definite, is improve the precision of change-detection by the feature effectively utilizing original image.

The present invention realizes the thinking of above-mentioned purpose: first use Scale invariant features transform method to extract the SIFT of original image Feature, using this as training sample, trains a deep neural network.Recycling log ratio method obtains the difference of original image Different figure, extracts the domain features of each pixel of this disparity map, in this, as test data, is input to the degree of depth god trained Test in network, export final change-detection result.

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

(1) SAR image is read in:

Read in SAR image I and J that two width of areal difference phase have registrated and corrected；

(2) normalization:

According to the following formula, SAR image I and J are normalized, obtain the SAR image after normalization:

$I^{\′}=\frac{I-min\left(I\right)}{\max \left(I\right)}$

$J^{\′}=\frac{J-min\left(J\right)}{max\left(J\right)}$

Wherein, I' represents the SAR image after the normalization of SAR image I, and min represents that taking minima operates, and max represents Taking maxima operation, J' represents the SAR image after the normalization of SAR image J；

(3) structure training characteristics:

(3a) use shift-invariant operator conversion SIFT method, extract SAR image I' and J' after two width normalization respectively Shift-invariant operator conversion SIFT feature S₁And S₂；

(3b) to two groups of shift-invariant operator conversion SIFT feature S₁And S₂Carry out cascade operation, the feature after being cascaded S；

(3c) to feature S after cascade, use principal component analysis PCA algorithm to carry out dimensionality reduction, obtain feature S' after dimensionality reduction；

(4) feature S' after dimensionality reduction is input in deep neural network, trains deep neural network；

(5) according to the following formula, the log ratio differential image of the two width SAR image that calculating is read in:

$D=\left|log\left(\frac{I+1}{J+1}\right)\right|$

Wherein, D represents the log ratio differential image of two width SAR image of reading, and log represents and takes from right log operations, | | Representing the operation that takes absolute value, I and J represents the SAR image of reading respectively；

(6) the neighborhood characteristics sample matrix of structure log ratio differential image D:

(6a) use neighborhood characteristics extracting method, extract every from the pixel matrix that log ratio differential image D is constituted The neighborhood characteristics vector of individual pixel；

(6b) the neighborhood characteristics vector of the log ratio all pixels of differential image D is formed the neighborhood of a M × N-dimensional Feature samples matrix, wherein, M represents the sum of all pixels in log ratio differential image D, and N represents logarithm ratio The dimension of the neighborhood characteristics vector of each pixel in value differential image D；

(7) detection log ratio differential image:

Being input in the deep neural network trained by the neighborhood characteristics sample matrix of log ratio differential image D, it is right to detect Number ratio difference image D, obtaining each pixel detection in log ratio differential image D is change class or the detection of non-changing class Classification；

(8) output detections classification.

The present invention compared with prior art has the advantage that

First, owing to present invention employs Scale invariant features transform SIFT algorithm, it is extracted the SIFT reading in SAR image Feature, and utilize this feature to be trained deep neural network, overcomes choosing of training sample in existing method unreliable Problem so that the present invention improves the precision of SAR image change-detection.

Second, owing to the present invention is extracted the SIFT feature of reading SAR image, this feature can reflect that the local of image is special Levy, and affine transformation and noise are also kept a certain degree of stability, overcome affected by noise in existing method causing The problem that can not effectively detect region of variation so that the present invention improves the precision of SAR image change-detection.

3rd, owing to the present invention is extracted the SIFT feature of reading SAR image, this feature is to image rotation, scaling And brightness flop is respectively provided with invariance, thus the feature extraction to different images has stability to a certain extent, gram Take problem the highest for different SAR image change-detection robustness in existing method so that the present invention is for different SAR image information has higher adaptability.

Accompanying drawing explanation

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

Fig. 2 is the change-detection result figure in emulation experiment of the present invention to Bern area SAR image；

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

Fig. 4 is the change-detection result figure in emulation experiment of the present invention to estuary area, the Yellow River SAR image.

Detailed description of the invention

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

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

Step 1, reads in SAR image.

Read in SAR image I and J that two width of areal difference phase have registrated and corrected.

Step 2, normalization.

According to the following formula, SAR image I and J are normalized, obtain the SAR image after normalization:

$I^{\′}=\frac{I-min\left(I\right)}{\max \left(I\right)}$

$J^{\′}=\frac{J-min\left(J\right)}{max\left(J\right)}$

Wherein, I' represents the SAR image after the normalization of SAR image I, and min represents that taking minima operates, and max represents Taking maxima operation, J' represents the SAR image after the normalization of SAR image J.

Step 3, constructs training characteristics.

Use shift-invariant operator conversion SIFT method, extract after two width normalization the translation of SAR image I' and J' respectively not Become eigentransformation SIFT feature S₁And S₂。

To two groups of shift-invariant operator conversion SIFT feature S₁And S₂Carry out cascade operation, feature S after being cascaded.

To feature S after cascade, use principal component analysis PCA algorithm to carry out dimensionality reduction, obtain feature S' after dimensionality reduction.

Step 4, is input to feature S' after dimensionality reduction in deep neural network, trains deep neural network.

The concrete operation step of training deep neural network is as follows:

The first step, initializes the parameter of limited Boltzmann machine (RBM)；

Second step, feature S ' to be trained uses limited Boltzmann machine (RBM) to be trained, obtains weight and biasing, The network number of plies is set to 3 hidden layers, and each node layer number is respectively 250, and 150,100, limited Boltzmann machine (RBM) Each layer of 50 generation of training；

3rd step, uses BP neutral net based on minimum cross entropy to be finely adjusted RBM training network, and training algebraically is 50 generations；

4th step, obtains the deep neural network trained.

Step 5, according to the following formula, the log ratio differential image of the two width SAR image that calculating is read in:

$D=\left|log\left(\frac{I+1}{J+1}\right)\right|$

Wherein, D represents the log ratio differential image of two width SAR image of reading, and log represents and takes from right log operations, | | Representing the operation that takes absolute value, I and J represents the SAR image of reading respectively.

Step 6, the neighborhood characteristics sample matrix of structure log ratio differential image D.

Use neighborhood characteristics extracting method, from the pixel matrix that log ratio differential image D is constituted, extract each pixel Neighborhood characteristics vector.

The concrete operation step of neighborhood characteristics extracting method is as follows:

The first step, chooses the sliding window that size is n × n-pixel, by selected window on log ratio differential image D The value of all pixels pulls into the characteristic vector of a 1 × N-dimensional, and wherein, n is the size of sliding window, N=n × n；

Second step, from left to right, sliding window the most successively, obtains all pixels on log ratio differential image D Neighborhood characteristics vector.

The neighborhood characteristics vector of the log ratio all pixels of differential image D is formed the neighborhood characteristics sample of a M × N-dimensional Matrix, wherein, M represents the sum of all pixels in log ratio differential image D, and N represents log ratio disparity map As the dimension of the neighborhood characteristics vector of each pixel in D.

Step 7, detects log ratio differential image D.

Being input in the deep neural network trained by the neighborhood characteristics sample matrix of log ratio differential image D, it is right to detect Number ratio difference image D, obtaining each pixel detection in log ratio differential image D is change class or the detection of non-changing class Classification.

Step 8, output detections classification.

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 Intel Pentium (R) Dual-Core CPU, internal memory 5GB in dominant frequency 2.30GHz Carry out under the software environment of hardware environment and MATLAB R2015a.

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

Missing inspection number: the number of pixels in the region that changes in statistical experiment result figure, individual with reference to the pixel of region of variation in figure Number contrasts, with reference to figure changes, experimental result picture being detected as unchanged number of pixels, and referred to as missing inspection number FN。

Flase drop number: the number of pixels in the region that do not changes in statistical experiment result figure, with the picture with reference to region of variation non-in figure Element number contrasts, and being detected as the number of pixels of change in experimental result picture with reference to not changing in figure, is referred to as by mistake Inspection number FP.

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

Weigh testing result figure and with reference to figure conforming Kappa coefficient:

$Kappa=\frac{(PCC-PRE)}{(1-PRE)}$

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 employs three groups of real SAR image data and change-detection is with reference to figure accordingly, and emulation is real The experimental image data testing middle employing are as follows:

First group of real SAR image data and corresponding change-detection that emulation experiment of the present invention uses are Bern with reference to figure The SAR image in area, as in figure 2 it is shown, image size is 301 × 301, Fig. 2 (a) is in April, 1999 Bern area SAR image, Fig. 2 (b) is the SAR image in May, 1999 Bern area, and Fig. 2 (c) is the corresponding change in Bern area Detection is 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 Ottawa area, as it is shown on figure 3, image size is 290 × 350, Fig. 3 (a) is in May, 1997 Ottawa The SAR image in area, Fig. 3 (b) is the SAR image in August, 1997 Ottawa area, and Fig. 3 (c) is Ottawa area 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 the Yellow River with reference to figure The SAR image in estuary area, as shown in Figure 4, image size is 306 × 291, and Fig. 4 (a) is in June, 2008 the Yellow River The SAR image in estuary area, Fig. 4 (b) is the SAR image in June, 2009 the Yellow River estuary area, and Fig. 4 (c) is yellow The corresponding change-detection in estuary area, river is with reference to figure.

The emulation experiment of the present invention uses broad sense KI threshold value GKI method, fuzzy local message C mean cluster FLICM side Method and employing the inventive method, become Bern area, Ottawa area and estuary area, the Yellow River SAR image respectively The testing result changing detection contrasts.

Emulation experiment one: use broad sense KI threshold value GKI method, fuzzy local message C mean cluster FLICM method and Use the change-detection result that obtains of the inventive method respectively as shown in Fig. 2 (d) to 2 (f), to testing result specifically to score Analysis is shown in Table 1.From the visual effect of Fig. 2 it can be seen that use the present invention testing result figure with reference to figure closest to.By Table 1 is it can be seen that the pixel count of false retrieval of the present invention is fewer 401 and 497 than GKI and FLICM respectively, and total Erroneous pixel number is fewer 183 and 278 than both the most respectively, and Kappa coefficient is also than the two difference high 2.69% and 5.36%.

Table 1 Bern area change-detection result

Method	Missing inspection pixel count	False retrieval pixel count	Total erroneous pixel number	Detection accuracy (%)	Kappa coefficient (%)
						GKI	56	513	569	99.37	79.13
FLICM	55	609	664	99.27	76.46
						The present invention	274	112	386	99.57	81.82

Emulation experiment two: use broad sense KI threshold value GKI method, fuzzy local message C mean cluster FLICM method and Use the change-detection result that obtains of the inventive method respectively as shown in Fig. 3 (d) to 3 (f), to testing result specifically to score Analysis is shown in Table 2.From the visual effect of Fig. 3 it can be seen that use the present invention testing result figure with reference to figure closest to.By Table 2 is it can be seen that the pixel count of false retrieval of the present invention is fewer 904 than GKI, and the pixel count of missing inspection is respectively than GKI and FLICM Lack 1814 and 1493, and total erroneous pixel number is fewer 3003 and 415 than both respectively, Kappa system Number is also than the two difference high 11.00% and 3.98%.

Table 2 Ottawa area change-detection result

Method	Missing inspection pixel count	False retrieval pixel count	Total erroneous pixel number	Detection accuracy (%)	Kappa coefficient (%)
						GKI	2962	2391	5353	94.73	80.29
FLICM	2641	124	2765	97.28	87.31
						The present invention	1148	1202	2350	97.68	91.29

Emulation experiment three: use broad sense KI threshold value GKI method, fuzzy local message C mean cluster FLICM method and Use the change-detection result that obtains of the inventive method respectively as shown in Fig. 4 (d) to 4 (f), to testing result specifically to score Analysis is shown in Table 3.From the visual effect of Fig. 4 it can be seen that use the present invention testing result figure with reference to figure closest to.By Table 3 it can be seen that the pixel count of false retrieval of the present invention is fewer 2392 and 30 than GKI and FLICM respectively, missing inspection Pixel count is fewer 2037 and 88 than GKI and FLICM respectively, and total erroneous pixel number respectively than both few 4429 Individual and 118, Kappa coefficient also than the two respectively high 44.27% and 0.79%.

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

Method	Missing inspection pixel count	False retrieval pixel count	Total erroneous pixel number	Detection accuracy (%)	Kappa coefficient (%)
						GKI	2988	2836	5824	93.46	41.00
FLICM	1039	474	1513	98.30	84.48
						The present invention	951	444	1395	98.43	85.27

Claims (3)

1., based on degree of depth study and a SAR image change detection for SIFT feature, comprise the steps:

(1) SAR image is read in:

Read in SAR image I and J that two width of areal difference phase have registrated and corrected；

(2) normalization:

According to the following formula, SAR image I and J are normalized, obtain the SAR image after normalization:

$I^{\′}=\frac{I-\min \left(I\right)}{\max \left(I\right)}$

$J^{\′}=\frac{J-\min \left(J\right)}{max\left(J\right)}$

Wherein, I' represents the SAR image after the normalization of SAR image I, and min represents that taking minima operates, max Representing and take maxima operation, J' represents the SAR image after the normalization of SAR image J；

(3) structure training characteristics:

(3a) use shift-invariant operator conversion SIFT method, extract SAR image I' after two width normalization respectively SIFT feature S is converted with the shift-invariant operator of J'₁And S₂；

(3b) to two groups of shift-invariant operator conversion SIFT feature S₁And S₂Carry out cascade operation, after being cascaded Feature S；

(3c) to feature S after cascade, principal component analysis PCA algorithm is used to carry out dimensionality reduction, after obtaining dimensionality reduction Feature S'；

(4) feature S' after dimensionality reduction is input in deep neural network, trains deep neural network；

(5) according to the following formula, the log ratio differential image of the two width SAR image that calculating is read in:

$D=\left|log\left(\frac{I+1}{J+1}\right)\right|$

Wherein, D represents the log ratio differential image of two width SAR image of reading, and log represents and takes natural logrithm Operation, | | representing the operation that takes absolute value, I and J represents the SAR image of reading respectively；

(6) the neighborhood characteristics sample matrix of structure log ratio differential image D:

(6a) neighborhood characteristics extracting method is used, from the pixel matrix that log ratio differential image D is constituted Extract the neighborhood characteristics vector of each pixel；

(6b) the neighborhood characteristics vector of the log ratio all pixels of differential image D is formed a M × N-dimensional Neighborhood characteristics sample matrix, wherein, M represents the sum of all pixels, N in log ratio differential image D Represent the dimension of the neighborhood characteristics vector of each pixel in log ratio differential image D；

(7) the change-detection result of acquisition log ratio differential image D:

The neighborhood characteristics sample matrix of log ratio differential image D is input in the deep neural network trained, Obtaining each pixel detection in log ratio differential image D is change class or the detection classification of non-changing class；

(8) output detections classification.

The most according to claim 1 based on degree of depth study and the SAR image change-detection side of SIFT feature Method, it is characterised in that: specifically comprising the following steps that of the training deep neural network described in step (4)

The first step, initializes deep neural network, and the hidden layer number of plies of deep neural network is 4, degree of depth nerve net The interstitial content of network hidden layer is respectively 250, and 150,100,1, each hidden layer of deep neural network is one Limited Boltzmann machine RBM；

Second step, uses deep neural network to be trained the feature S ' after dimensionality reduction, obtains each hidden layer limited The weight of Boltzmann machine RBM and biasing；

3rd step, trains 100 times the limited Boltzmann machine RBM of each hidden layer；

4th step, uses back-propagation algorithm based on minimum cross entropy, is finely adjusted deep neural network, To the deep neural network trained.

The most according to claim 1 based on degree of depth study and the SAR image change-detection side of SIFT feature Method, it is characterised in that: described in step (6a), neighborhood characteristics extracting method specifically comprises the following steps that

The first step, chooses the sliding window that size is n × n-pixel in disparity map D, is owned by selected window The value of pixel pulls into the characteristic vector of a 1 × N-dimensional, and wherein, n is the size of sliding window, N=n × n；

Second step, from left to right, sliding window the most successively, obtains the neck of each pixel of differential image D Characteristic of field vector.