CN106204608A - On-line talking SAR image change detection based on sample local density - Google Patents

Abstract

The invention discloses a kind of on-line talking SAR image change detection based on sample local density.Its scheme is: 1. image to be detected carries out medium filtering and seeks logarithm ratio disparity map；2. normalization disparity map, using its gray value as data set；3. calculate the local density of sample in data set, rearrangement data set piecemeal；4. pair the first blocks of data clusters；5. update the weights of current cluster centre；6. input next blocks of data, by its initial weight and current cluster centre weight number combining；7. this blocks of data is merged with current cluster centre, and give the new weights that step 6 obtains；8. the new data that pair step 7 obtains clusters；9. repeat step 58 until all data are disposed, determine sample class mark；10. according to sample class indicated weight structure change-detection result figure.Present invention reduces the error rate of change-detection, improve Detection results, can be used for the variation monitoring of urban construction and forest.

Description

On-line talking SAR image change detection based on sample local density

Technical field

The invention belongs to image processing field, relate to the change detecting method of multidate SAR image, can be used for commenting of the condition of a disaster Estimate and predict, the construction in city and the variation monitoring of forest.

Background technology

Synthetic aperture radar SAR is a kind of high-resolution imaging radar, along with the development of satellite remote sensing technology, SAR data Become a kind of important remotely-sensed data.Change-detection is to analyze and determine the spy that earth's surface changes in remotely-sensed data the most of the same period Levy and process.The most never then have particularly significant with the region finding out notable change in the SAR image of phase Meaning.

Finding the research of " non-changing " and " change " two classification in SAR image, more common route is disparity map analysis Method, the method first image to two phases carries out the comparison by pixel, then makees further comparing the differential image drawn Process to reach two classification, finally obtain change-detection figure.Disparity map sorting technique can reduce owing to a variety of causes causes Pseudo-change information, it is not necessary to the complicated pretreatment of data, method simple, intuitive.

In the change detecting method analyzed based on differential image, change in disparity map and the accurate of non-changing region are divided Class is a step of most critical, and research direction currently mainly is change detecting method based on unsupervised segmentation.Unsupervised point Class method is the most again clustering method, obtains changing class and non-changing class, SAR image by differential image carries out cluster segmentation Change-detection is i.e. converted into the clustering problem to disparity map.Fuzzy C-means clustering FCM is one of most popular clustering algorithm, passes The FCM algorithm of system is a kind of clustering algorithm based on gradation of image, and in cluster process, each pixel is separate, not In view of the relation of the gray feature of each pixel in image Yu other pixel, for the figure that not Noise or noise are the lowest Picture, its segmentation effect is preferable.But generally, image is inevitably done by different noises in imaging process Disturb.

Summary of the invention

Present invention aims to the deficiency of above-mentioned prior art fuzzy C-means clustering method, propose a kind of based on The on-line talking SAR image change detection of sample local density, to reduce change-detection error rate, improves Detection results.

For achieving the above object, technical scheme includes the following:

(1) two phase images to be detected of input are carried out medium filtering process, obtain filtered two phases to be detected Image I₁And I₂；

(2) filtered two phase images to be detected are sought logarithm ratio differential image gray value X₃And normalization, by normalizing Differential image gray value X after change₃' as cluster data collection；

(3) calculate cluster data and concentrate the local distribution density of all samples, according to the local distribution density of sample from greatly To little to the rearrangement of cluster data collection piecemeal；

(4) inputting the first blocks of data, arranging initial weight is 1, clusters it by Weighted Fuzzy C-Means method；

(5) weights of current cluster centre are updated；

(6) inputting next blocks of data, arranging initial weight is 1, is closed by the weights of this initial weight with current cluster centre And, obtain new weights；

(7) newly inputted data are merged with current cluster centre, obtain new data, then will step (6) obtain New weights are assigned to the data that this is new；

(8) new data obtained step (7) by Weighted Fuzzy C-Means method clusters；

(9) repeat step (5) to (8) until all data are disposed, export final cluster centre；

(10) the cluster data all samples of the concentration degree of membership to the cluster centre of final output is calculated, true according to degree of membership Random sample this class mark；

(11) according to the sample class indicated weight structure change-detection result figure obtained in step (10)..

The present invention compared with prior art has the advantage that

Traditional FCM Algorithms each pixel in cluster process is separate, and does not takes into account in image The gray feature of each pixel and the relation of other pixel, the error rate causing change-detection is higher.

The present invention first calculates the local density of pixel in cluster process, then according to the size of local density to data Rearrangement, clusters the most successively by deblocking after sequence, and this clustering method is due to sample in the biggest region of density Point is closer to real cluster centre, and therefore the present invention compared with prior art reduces the error rate of change-detection, improves Detection results.

Accompanying drawing explanation

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

Fig. 2 is the change-detection result schematic diagram regional to Bern by the inventive method；

Fig. 3 is the change-detection result schematic diagram regional to Ottawa by the inventive method.

Detailed description of the invention

With reference to Fig. 1, the change-detection process of the present invention is as follows:

Step 1, inputs two phase images to be detected and is smoothed.

Inputting two phase images to be detected, selected window size is that the median filter of 3 × 3 two phases to inputting are to be checked Altimetric image carries out medium filtering, removes little noise jamming, obtains filtered two phase image to be detected I₁And I₂。

Filtered two phase images to be detected are sought logarithm ratio differential image gray value X by step 2₃And normalization, will return Differential image gray value X after one change₃' as cluster data collection.

(2a) the pixel gray value X of logarithm ratio differential image is sought₃:

X₃=| log (X₂+1)-log(X₁+1)|

Wherein X₁、X₂Represent filtered two phase image to be detected I respectively₁And I₂Pixel gray value；

(2b) the pixel gray value X to differential image₃It is normalized, obtains the image intensity value after normalization X₃':

X₃'=255* (X₃-X_min)/(X_max-X_min)

Wherein X_max=max (X₃),X_min=min (X₃)；

(2c) using the image intensity value that obtains after normalized as cluster data collection.

Step 3, calculates cluster data and concentrates the local distribution density of all samples, according to the local distribution density of sample from Greatly to little to the rearrangement of cluster data collection piecemeal.

(3a) i-th sample point x is calculated_iWith jth sample point x_jBetween Euclidean distance:

d_ij=| | x_i-x_j||,1≤i≤n,1≤j≤n

(3b) i-th sample point x is calculated_iLocal distribution density value:

$z_i=\underset{j=1,j\≠i}{\overset{n}{\Σ}}\frac{1}{d_{ij}},d_{ij}\≤e$

Wherein, e is the scope limit value of local distribution density, is chosen by experiment；

(3c) from big to small data set resequenced according to calculated local distribution density value and be divided into S block, S's Depending on concrete value is according to experimental data, span is 10≤S≤50, and the sample point number of every blocks of data is n_s。

Step 4, inputs the first blocks of data, and arranging initial weight is 1, by Weighted Fuzzy C-Means method to the first blocks of data Cluster.

(4a) initialize two cluster centres, choose two samples of gray value minimum and maximum in all samples the most respectively O'clock as two initial cluster centers；

(4b) object function of selection FCM Algorithms:

$J_m(U,V)=\underset{i=1}{\overset{2}{\Σ}}\underset{j=1}{\overset{n_s}{\Σ}}{u}_{ij}^m{d}_{ij}^2$

Wherein, U is subordinated-degree matrix, u_ijRepresent that jth sample point is under the jurisdiction of the degree of membership of the i-th class, meet u_ij∈[0, 1] and certain sample point to belong to the degree of membership sum of each fuzzy subset be 1, V is cluster centre, d_ij=| | x_j-v_i| | represent jth Individual sample point is to the Euclidean distance at ith cluster center, and m is fuzziness, and span is 1.5≤m≤2.5；

(4c) degree of membership u of each sample point in the first blocks of data is calculated_ij:

$u_{ij}={\[\underset{k=1}{\overset{2}{\Σ}}{\left(\frac{\left|\right|x_j-v_i\left|\right|}{\left|\right|x_j-v_k\left|\right|}\right)}^{\frac{2}{m-1}}\]}^{-1},\∀i,j$

Wherein, v_iAnd v_kRepresent i-th and kth cluster centre respectively, | | x_j-v_i| | and | | x_j-v_k| | represent jth respectively Individual sample point is to i-th and the Euclidean distance of kth cluster centre；

(4d) according to degree of membership u_ijCalculate new cluster centre v_i:

$v_i=\frac{\underset{j=1}{\overset{n_s}{\Σ}}{w}_j{\left(u_{ij}\right)}^m{x}_j}{\underset{j=1}{\overset{n_s}{\Σ}}{w}_j{\left(u_{ij}\right)}^m},i=1,2$

Wherein, w_jFor the weights of jth sample point, initial weight is set to 1, u_ijRepresent that jth sample point is under the jurisdiction of i-th The degree of membership of class；

(4e) cluster centre and degree of membership are constantly updated by iteration so that the value of object function is constantly leaned on to minima Closely, when meeting condition max{ | | v_k,new-v_k,old||²}≤ε, during 1≤k≤2, stops iteration, exports the most up-to-date cluster centre V, wherein v_k,newAnd v_k,oldKth cluster centre before representing respectively after updating and updating, ε=10^-3。

Step 5, updates the weights of current cluster centre.

Degree of membership according to sample point and weights, calculate the new weight w of the current cluster centre after updating_i':

${w_i}^{\′}=\underset{j=1}{\overset{n_s}{\Σ}}\left(u_{ij}\right)w_j,i=1,2$

Wherein n_sFor the sample point number by blocks of data every after cluster data collection piecemeal, u_ijRepresent that jth sample point is subordinate to In the degree of membership of the i-th class, w_jFor the weights of jth sample point in every blocks of data.

Step 6, inputs next blocks of data, and arranging initial weight is 1, by the weights of this initial weight Yu current cluster centre Merge, obtain new weight w_j′。

Newly inputted data are merged by step 7 with current cluster centre, obtain new data, then will obtain in step 6 New weights are assigned to the data that this is new.

Step 8, the new data obtained step 7 by Weighted Fuzzy C-Means method clusters.

(8a) cluster centre of last round of loop iteration output is chosen as initialized cluster centre；

(8b) degree of membership u of each sample point in the new data after merging is calculated_ij':

${u_{ij}}^{\′}={\[\underset{k=1}{\overset{2}{\Σ}}{\left(\frac{\left|\right|{x_j}^{\′}-v_i\left|\right|}{\left|\right|{x_j}^{\′}-v_k\left|\right|}\right)}^{\frac{2}{m-1}}\]}^{-1},\∀i,j$

Wherein, x_j' represent jth sample point in the new data after merging, v_iAnd v_kRepresent i-th and kth cluster respectively Center, | | x_j′-v_i| | and | | x_j′-v_k| | represent that jth sample point in new data is to i-th and kth cluster centre respectively Euclidean distance；M is fuzziness, and span is 1.5≤m≤2.5；

(8c) according to degree of membership u_ij' calculate new cluster centre v_i':

${v_i}^{\′}=\frac{\underset{j=1}{\overset{n_s+2}{\Σ}}{w_j}^{\′}{\left({u_{ij}}^{\′}\right)}^m{x_j}^{\′}}{\underset{j=1}{\overset{n_s+2}{\Σ}}{w_j}^{\′}{\left({u_{ij}}^{\′}\right)}^m},i=1,2$

Wherein, w_j' for the new weights that obtain after merging in step 6；

(8d) cluster centre and degree of membership are constantly updated by iteration, until meeting condition max{ | | v_k,new-v_k,old||²} ≤ ε, during 1≤k≤2, stops iteration, exports the most up-to-date cluster centre V, wherein v_{K, new}And v_{K, old}Represent respectively after updating With the kth cluster centre before renewal, ε=10^-3。

Step 9, repeats step 5 to 8 until all data are disposed, exports final cluster centre.

Step 10, calculates the cluster data all samples of the concentration degree of membership to the cluster centre of final output, according to being subordinate to Degree determines sample class mark.

Step 11, according to the sample class indicated weight structure change-detection result figure obtained in step 10.

2 cluster centres finally given are respectively v₁And v₂, wherein v₁＜ v₂, and v will be under the jurisdiction of₁The gray scale of pixel Value is set to 0, and remaining is set to 255.

The effect of the present invention can be further illustrated by following emulation:

1. emulation data

First group of SAR data is in April, 1999 and May is obtained near Switzerland Bern city by the SAR entrained by ERS 2 , image size is 301 × 301 pixels, 256 gray levels, and wherein change pixel count is 1155, and not changing pixel count is 89446.

Second group of SAR data be in May, 1997 and August by the SAR entrained by Radarsat in Canada Ottawa city Neighbouring acquisition, image size is 290 × 350 pixels, 256 gray levels, and wherein change pixel count is 11952, does not changes pixel count It is 82871.

2. emulation content

Emulation 1, arranges deblocking number S=20, the scope limit value e=50 of local distribution density, fuzziness m=2.0, It is changed detecting to first group of SAR data by the present invention, result such as Fig. 2.Wherein Fig. 2 (a) is the former of true SAR first phase Figure, Fig. 2 (b) is the artwork of true SAR the second phase, Fig. 2 (c) be the change-detection of two phases with reference to figure, Fig. 2 (d) is this Bright change-detection result figure.

From Fig. 2 (d), with change-detection with reference to compared with figure, the present invention is total to the change-detection of first group of SAR data Body result is good.

The statistics present invention change-detection false retrieval number to first group of SAR data, missing inspection number, total error number and error detection Rate, result such as table 1

The change-detection result of first group of SAR data is added up by table 1 present invention:

	False retrieval number	Missing inspection number	Total error number	False detection rate
					Bern	237	71	308	0.34%

From table 1, the present invention false retrieval number to first group of SAR data, missing inspection number, total error number is all in relatively low water Flat, false detection rate is relatively low.

Emulation 2, arranges deblocking number S=20, the scope limit value e=50 of local distribution density, fuzziness m=2.0, It is changed detecting to second group of SAR data by the present invention, result such as Fig. 3.Wherein Fig. 3 (a) is the former of true SAR first phase Figure, Fig. 3 (b) is the artwork of true SAR the second phase, Fig. 3 (c) be the change-detection of two phases with reference to figure, Fig. 3 (d) is this Bright change-detection result figure.

From Fig. 3 (d), with change-detection with reference to compared with figure, the present invention is total to the change-detection of second group of SAR data Body result is good.

The statistics present invention change-detection false retrieval number to second group of SAR data, missing inspection number, total error number and error detection Rate, result such as table 2

The change-detection result of second group of SAR data is added up by table 2 present invention:

	False retrieval number	Missing inspection number	Total error number	False detection rate
					Ottawa	2266	1111	3377	3.32%

From table 2, the present invention false retrieval number to second group of SAR data, missing inspection number, total error number is all in relatively low water Flat, false detection rate is relatively low.

Claims (5)

1. an on-line talking SAR image change detection based on sample local density, including:

(1) two phase images to be detected of input are carried out medium filtering process, obtain filtered two phase image to be detected I₁ And I₂；

(2) filtered two phase images to be detected are sought logarithm ratio differential image gray value X₃And normalization, after normalization Differential image gray value X₃' as cluster data collection；

(3) calculate cluster data and concentrate the local distribution density of all samples, according to the local distribution density of sample from big to small To the rearrangement of cluster data collection piecemeal；

(4) inputting the first blocks of data, arranging initial weight is 1, clusters it by Weighted Fuzzy C-Means method；

(5) weights of current cluster centre are updated；

(6) inputting next blocks of data, arranging initial weight is 1, by the weight number combining of this initial weight Yu current cluster centre, To new weights；

(7) newly inputted data are merged with current cluster centre, obtain new data, newer by what step (6) obtained Weights are assigned to the data that this is new；

(8) new data obtained step (7) by Weighted Fuzzy C-Means method clusters；

(9) repeat step (5) to (8) until all data are disposed, export final cluster centre；

(10) calculating cluster data concentrates all samples to the degree of membership of the cluster centre of final output, determines sample according to degree of membership This class mark；

(11) according to the sample class indicated weight structure change-detection result figure obtained in step (10).

Method the most according to claim 1, wherein calculates the local distribution density of sample in step (3), according to following step Suddenly carry out:

(3a) i-th sample point x is calculated_iWith jth sample point x_jBetween Euclidean distance:

d_ij=| | x_i-x_j||,1≤i≤n,1≤j≤n

(3b) i-th sample point x is calculated_iLocal distribution density:

$z_i=\underset{j=1,j\≠i}{\overset{n}{\Σ}}\frac{1}{d_{ij}},d_{ij}\≤e$

Wherein, e is the scope limit value of local distribution density, is chosen by experiment.

Method the most according to claim 1, wherein uses first piece to input of Weighted Fuzzy C-Means method in step (4) Data cluster, and carry out in accordance with the following steps:

(4a) initialize two cluster centres, choose two sample point conducts of gray value minimum and maximum in all samples respectively Two initial cluster centers；

(4b) object function of selection FCM Algorithms:

$J_m(U,V)=\underset{i=1}{\overset{2}{\Σ}}\underset{j=1}{\overset{n_s}{\Σ}}{u}_{ij}^m{d}_{ij}^2$

Wherein, U is subordinated-degree matrix, u_ijRepresent that jth sample point is under the jurisdiction of the degree of membership of the i-th class, meet u_ij∈ [0,1] and certain It is 1 that individual sample point belongs to the degree of membership sum of each fuzzy subset；V is cluster centre；d_ij=| | x_j-v_i| | represent jth sample Point is to the Euclidean distance at ith cluster center；M is fuzziness, and span is 1.5≤m≤2.5；

(4c) degree of membership u of each sample point in the first blocks of data is calculated_ij:

$u_{ij}={\[\underset{k=1}{\overset{2}{\Σ}}{\left(\frac{\left|\right|x_j-v_i\left|\right|}{\left|\right|x_j-v_k\left|\right|}\right)}^{\frac{2}{m-1}}\]}^{-1},\∀i,j$

Wherein, x_jRepresent jth sample point, v_iAnd v_kRepresent i-th and kth cluster centre respectively, | | x_j-v_i| | and | | x_j-v_k | | represent that jth sample point is to i-th and the Euclidean distance of kth cluster centre respectively；

(4d) according to degree of membership u_ijCalculate new cluster centre v_i:

$v_i=\frac{\underset{j=1}{\overset{n_s}{\Σ}}{w}_j{\left(u_{ij}\right)}^m{x}_j}{\underset{j=1}{\overset{n_s}{\Σ}}{w}_j{\left(u_{ij}\right)}^m},i=1,2$

Wherein, w_jFor the weights of jth sample point, initial weight is set to 1, u_ijRepresent that jth sample point is under the jurisdiction of the i-th class Degree of membership；

(4e) cluster centre and degree of membership are constantly updated by iteration so that the value of object function is constantly close to minima, when Meet condition max{ | | v_k,new-v_k,old||²}≤ε, during 1≤k≤2, stops iteration, exports the most up-to-date cluster centre V, its Middle v_k,newAnd v_k,oldKth cluster centre before representing respectively after updating and updating, ε=10^-3。

Method the most according to claim 1, wherein updates the weights of current cluster centre in step (5), according to following public Formula calculates:

${w_i}^{\′}=\underset{j=1}{\overset{n_s}{\Σ}}\left(u_{ij}\right)w_j,i=1,2$

Wherein w_i' represent the new weights that ith cluster center obtains in the updated, n_sFor by every piece of number after cluster data collection piecemeal According to sample point number, w_jFor the weights of jth sample point, u in every blocks of data_ijRepresent that jth sample point is under the jurisdiction of the i-th class Degree of membership.

Method the most according to claim 1, the new number after being wherein combined by Weighted Fuzzy C-Means method in step (8) According to clustering, carry out in accordance with the following steps:

(8a) cluster centre of last round of loop iteration output is chosen as initialized cluster centre；

(8b) degree of membership u of each sample point in the new data after merging is calculated_ij':

${u_{ij}}^{\′}={\[\underset{k=1}{\overset{2}{\Σ}}{\left(\frac{\left|\right|{x_j}^{\′}-v_i\left|\right|}{\left|\right|{x_j}^{\′}-v_k\left|\right|}\right)}^{\frac{2}{m-1}}\]}^{-1},\∀i,j$

Wherein, x_j' represent jth sample point in the new data after merging, v_iAnd v_kRepresent respectively in i-th and kth cluster The heart, | | x_j′-v_i| | and | | x_j′-v_k| | represent that in new data, jth sample point is to i-th and the Europe of kth cluster centre respectively Family name's distance；M is fuzziness, and span is 1.5≤m≤2.5；

(8c) according to degree of membership u_ij' calculate new cluster centre v_i':

${v_i}^{\′}=\frac{\underset{j=1}{\overset{n_s+2}{\Σ}}{w_j}^{\′}{\left({u_{ij}}^{\′}\right)}^m{x_j}^{\′}}{\underset{j=1}{\overset{n_s+2}{\Σ}}{w_j}^{\′}{\left({u_{ij}}^{\′}\right)}^m},i=1,2$

Wherein, w_j' for the new weights obtained after merging in step (6)；

(8d) cluster centre and degree of membership are constantly updated by iteration, until meeting condition max{ | | v_k,new-v_k,old||²}≤ε,1 During≤k≤2, stop iteration, export the most up-to-date cluster centre V, wherein v_k,newAnd v_k,oldRepresent respectively after updating and update Front kth cluster centre, ε=10^-3。