CN104680542B - Remote sensing image variation detection method based on on-line study - Google Patents
Remote sensing image variation detection method based on on-line study Download PDFInfo
- Publication number
- CN104680542B CN104680542B CN201510112839.9A CN201510112839A CN104680542B CN 104680542 B CN104680542 B CN 104680542B CN 201510112839 A CN201510112839 A CN 201510112839A CN 104680542 B CN104680542 B CN 104680542B
- Authority
- CN
- China
- Prior art keywords
- image
- testing result
- change
- remote sensing
- width
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Expired - Fee Related
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30181—Earth observation
Abstract
The invention discloses a kind of remote sensing image variation detection method based on on-line study, the problem of mainly solving unstable existing detection technique testing result and low precision.Its implementation process is:Obtain two width remote sensing images;For the width differential image of remote sensing image type structure two;Training sample database is constructed to the first width differential image and the image block set of frame of video form is divided into;Detection is changed to two field picture block by the strategy of on-line study one by one with cascade classifier;Then, the testing result splicing of all two field picture blocks is obtained into the testing result CM1 of the first width differential image;Similar process is carried out to the second width differential image again and obtains testing result CM2;Above-mentioned testing result CM1 and CM2 is subjected to grey scale mapping and merged, the differential image X after being mergedF, then to XFThe final change testing result of cluster generation.The present invention can obtain the Detection results that robustness is good, precision is high to different types of remote sensing image, available for urban planning.
Description
Technical field
The invention belongs to the change inspection of technical field of image processing, more particularly to SAR remote sensing images and remote sensing image
Survey method.Available for the planning of monitoring, urban development, the assessment of Natural Disaster and the ground that utilization power is covered to atural object
The renewal of figure.
Background technology
Remote sensing image change detection is intended to the situation of change occurred between the image of detection areal different time, that is, examines
Measure the change information that the atural object of this area occurs with the time.At present, remote sensing image change detection has been widely used for ground
The renewal of geographical spatial data and violating the regulations built in the assessments of the Natural Disasters such as shake, flood, mud-rock flow and forest fire, mapping
In terms of the investigation built and the planning construction in post-disaster reconstruction city.
The method of remote sensing image change detection, is most commonly based on the change detecting method of differential image analysis.The party
Method mainly includes three processing procedures:1) remote sensing image is pre-processed, including radiant correction and geometrical registration etc.;2) to rectifying
Remote sensing image after just is compared acquisition differential image;3) differential image is analyzed, classified using threshold value or cluster etc.
Differential image is divided into change class and non-changing class by method, obtains final change testing result.
For the sorting technique employed in differential image analysis process, according to whether there is training sample set to examine change
Survey method is divided into unsupervised change and detects and have the change detecting method of supervision.Unsupervised change detecting method need not be first
Test change sample information support, directly by differential image by cluster or partitioning algorithm be can obtain change testing result.It is right
For this method, the selection of differential image segmentation threshold is the key issue of change detection, directly affects the whole of testing result
How body precision, set threshold value to improve one of the problem of precision of testing result is studied as scholars.Compared to unsupervised
Change detecting method, the change detection sorting technique for having supervision can identify the part changed exactly, and to not
Same atmospheric conditions and different illumination conditions have preferable robustness, but the premise for the effect to have obtained is that to have substantial amounts of
Training sample, that is, require to grasp substantial amounts of ground real change information;And it is difficult and expense to collect substantial amounts of ground truth
When.
The content of the invention
It is an object of the invention to propose a kind of remote sensing image variation detection method based on on-line study, to solve no prison
Superintend and direct the shortcoming that accuracy of detection is not high, robustness is low of change detecting method and there is supervision variation detection method to need construction substantial amounts of
The deficiency of training sample.
The present invention is a kind of based on semi-supervised change detecting method, by a small amount of training sample, takes on-line study
Mode be continuously increased Sample Storehouse, lifted grader classification performance, and to differential image carry out analyze obtain change testing result.
Implementation step includes as follows:
(1) it is I × J process radiant correction and the remote sensing image X of geometrical registration to obtain two width sizes1And X2, wherein, I
For the line number of remote sensing image, J is the columns of remote sensing image;
(2) two width remote sensing image X are utilized1And X2Construct two width differential image XLAnd XD;
(3) it is directed to the first width differential image XLSample Storehouse P0 and single pixel Sample Storehouse P1 of the construction based on 2 × 2 image blocks,
And the threshold value TH0 of the first order, i.e. average grader of cascade classifier is initialized according to Sample Storehouse P0;
(4) by the first width differential image XLSize is divided into for N × N frame of video by order from left to right, from top to bottom
The image block set of formWherein N is even number“Z+" it is just whole
Number, "" represent to round up;
(5) initialisation image set of blocksIndex value i=1, start to the 1st two field picture block B1Enter
Row change detection;
(6) using cascade classifier to image block setI-th of two field picture block BiIt is changed
Detection, and renewal is optimized to cascade classifier;
(7) i is from Jia 1, and the cascade classifier after being updated using optimization is changed detection to next frame image block;
(8) repeat step (6)-step (7), untilComplete the change to two field picture set of blocks B
Detection, obtains corresponding change testing result set
(9) the change testing result set C derived above to two field picture set of blocks is pressed from left to right, from top to bottom
Sequential concatenation is completed to the first width differential image X into final change testing result figure CM1LChange detection;
(10) by another width differential image X in step (2)D, according to the first width differential image XLChange detecting step
(3)-step (9) completes the second width differential image XDChange detection, note change testing result figure be CM2;
(11) change the testing result figure CM1 and CM2 of two width differential images are mapped as gray level image A1 and A2, and used
The method of class principal component analysis is merged to gray level image A1 and A2, the differential image X after being mergedF;
(12) with Kmeans clustering algorithms to the differential image X after fusionFClustered, generate final change detection knot
Fruit figure XCD, complete the detection to remote sensing image change information.
The present invention has advantages below compared with prior art:
It is not using view picture differential image as processing pair 1. the present invention has abandoned the method that traditional differential image analyzes process
As carrying out classification processing pixel-by-pixel by a certain fixed criterion;But view picture differential image is divided into the image block of similar frame of video
Form, then by cascade classifier using two field picture block as process object be changed detection frame by frame, due to cascade classifier
Simple in construction, amount of calculation is small, therefore speed is fast, improves the arithmetic speed of change detection.
2. the present invention is drawn, supervised classification robustness is good and accurate advantage of classifying, and overcomes traditional supervision variation detection method
Need to construct the shortcoming of a large amount of training samples, because the present invention using priori passes through the strategy of on-line study one by one to each
Two field picture block, which is changed, to be detected and constantly updates Sample Storehouse, improves constantly the performance of grader, is realized accurately classification, is carried
The accuracy of detection of High variation detection.
3. the present invention uses the adaptive image fusion policy that should determine that weights based on spatial domain, it is to avoid excessive manual intervention
Influence, can Automatic-searching best weight value carry out linear weighted function fusion, the differential image after fusion more comprehensively can truly react
The situation of change of ground atural object emittance, improves the accuracy of detection of change detection;
The present invention fusion method with such as fusion method based on wavelet transformation based on transform domain compared with have simply,
The small advantage of amount of calculation.
Brief description of the drawings
Fig. 1 is the implementation process figure of the present invention;
Fig. 2 is to be changed the sub-process for detecting and being optimized to cascade classifier renewal in the present invention to two field picture block
Figure;
Fig. 3 is the sub-process figure merged to gray level image A1 and A2 in the present invention;
Fig. 4 is the two width SAR remote sensing images and a width canonical reference figure in the Switzerland Berne areas that present invention emulation is used;
The two width differential images constructed when Fig. 5 is present invention emulation by the two width SAR remote sensing images in Berne areas;
Fig. 6 is that the present invention is changed the result figure of detection to Fig. 5 two width differential images respectively;
Fig. 7 is that Fig. 6 two width testing results are mapped as gray level image and merge obtained differential image X by the present inventionF;
Fig. 8 is the simulation result figure that the present invention is changed detection to two width SAR remote sensing images in Fig. 4;
Fig. 9 is the two width remote sensing images and a width canonical reference figure in the Sardinia areas that present invention emulation is used;
The two width differential images constructed when Figure 10 is present invention emulation by the two width remote sensing images in Sardinia areas;
Figure 11 is that the present invention is changed the result figure of detection to Figure 10 two width differential images respectively;
Figure 12 is that Figure 11 two width testing results are mapped as gray level image and merge obtained differential image X by the present inventionF;
Figure 13 is the simulation result figure that the present invention is changed detection to two width remote sensing images in Fig. 9.
Embodiment
Below in conjunction with accompanying drawing, technical scheme and effect are described in further detail.
Reference picture 1, step is as follows for of the invention realizing:
Step 1, two width remote sensing images are obtained.
This two width remote sensing image is the process radiant correction of identical regional different time and the remote sensing image of geometrical registration, is divided
X is not denoted as it1And X2, its image size is I × J, wherein, I is the line number of remote sensing image, and J is the columns of remote sensing image.
Step 2, remote sensing image X is utilized1And X2Construct two width differential image XLAnd XD。
The building method of conventional differential image has differential technique, ratio method, average ratio value method and log ratio method etc..Difference
Method is to be compared two width remote sensing images pixel-by-pixel, and the gray value of corresponding position is made into difference and taken absolute value;Ratio method is
The gray value of two width remote sensing image corresponding positions is made into ratio proccessing;Average ratio value method is to calculate two width remote sensing images pair respectively
The average of the regional area gray value of location of pixels is answered, ratio proccessing is remake;Log ratio method is the difference for obtaining ratio method
The gray value of each pixel of image carries out processing of taking the logarithm.
This example according to SAR remote sensing images it is different with remote sensing image institute Noise the characteristics of, for remote sensing image X1
And X2Type carry out two width differential image XLAnd XDConstruction:
If remote sensing image X1And X2Be SAR remote sensing images, then differential image XLAnd XDRespectively from log ratio method and
Value ratio method is constructed, and corresponding construction expression formula is respectively:
XL=| log X2-log X1|
Wherein, μ1、μ2Respectively X1、X2Regional area gray value average;
If remote sensing image X1And X2Be remote sensing image, then differential image XLAnd XDDifferential technique and logarithm are selected respectively
Ratio method is constructed, referred to as error image and logarithm ratio image, and corresponding construction expression formula is respectively:
XL=| X1-X2|
XD=| log (1+X2)-log(1+X1)|。
Step 3, for the first width differential image XLSample Storehouse P0 and single pixel Sample Storehouse of the construction based on 2 × 2 image blocks
P1, and the threshold value TH0 of the first order, i.e. average grader according to Sample Storehouse P0 initialization cascade classifiers.
3a) in the first width differential image XLThe artificial image block for marking one 2 × 2 is gone up as the block of pixels changed,
It is designated as a positive sample in Sample Storehouse P0, and chooses centered on each pixel of 2 × 2 image block figure of one 3 × 3
As block and be changed into column vector as in single pixel Sample Storehouse P1 positive sample, i.e. Sample Storehouse P0 1 positive sample correspondence Sample Storehouse
4 positive samples in P1;
3b) repeat step 3a) 100 times, obtain in 100 positive samples and single pixel Sample Storehouse P1 in Sample Storehouse P0
400 positive samples;
3c) in the first width differential image XLThe upper artificial image block for marking one 2 × 2 is used as the pixel not changed
Block, is designated as a negative sample in Sample Storehouse P0, and selection one 3 × 3 centered on each pixel of 2 × 2 image block
Image block and be changed into column vector as in single pixel Sample Storehouse P1 negative sample, i.e. Sample Storehouse P0 1 negative sample correspondence sample
4 negative samples in the P1 of this storehouse;
3d) repeat step 3c) 100 times, obtain in 100 negative samples and single pixel Sample Storehouse P1 in Sample Storehouse P0
400 negative samples;
3e) 100 positive samples and 100 negative samples in Sample Storehouse P0, initialize the first order of cascade classifier,
That is the threshold value TH0=(positive sample average+negative sample average)/4 of average grader.
Step 4, by the first width differential image XLThe image of frame of video form is divided into by order from left to right, from top to bottom
Set of blocksThe size of each of which image block is N × N, and N is even number, i ∈ Z+And
Wherein, " Z+" be positive integer, "" represent to round up.
Step 5, initialisation image set of blocksIndex value i=1, start to the 1st two field picture block
B1It is changed detection.
Step 6, using cascade classifier to image block setI-th of two field picture block BiBecome
Change detection, and renewal is optimized to cascade classifier.
Reference picture 2, this step to implement step as follows:
6a) most of non-targeted pixel, i.e. non-changing picture are excluded using the first order average grader of cascade classifier
Vegetarian refreshments, completes the preliminary classification to present frame:
6a1) i-th of two field picture block B is scanned with 2 × 2 overlapping sliding windowsi;
The average of 2 × 2 block of pixels in sliding window 6a2) is calculated, by its threshold with the first order average grader of cascade classifier
Value TH0 is compared, if the pixel that the average of block of pixels is less than in TH0, current 2 × 2 window in sliding window does not include
Target pixel points, otherwise, the pixel in current 2 × 2 window include target pixel points;
The probability that each pixel includes target pixel points 6a3) is counted, if probability is less than 0.5, current pixel is judged
Point is non-targeted pixel, as non-changing pixel, otherwise, performs step 6d);
6b) using the first order average grader of cascade classifier to i-th of two field picture block BiCarry out being based on 2 × 2 pixels
The classification of block, constitutes the first mask MAP0:
I-th of two field picture block B 6b1) is scanned with nonoverlapping 2 × 2 sliding windowi;
The average of 2 × 2 block of pixels in sliding window 6b2) is calculated, by its threshold with the first order average grader of cascade classifier
Value TH0 is compared, if the average of block of pixels is less than TH0 in sliding window, the gray value of the pixel in window is set into 0;
Otherwise, the gray value of the pixel in window is set to 1;
6b3) repeat above-mentioned steps 6b1) -6b2), complete to i-th of two field picture block BiScanning, obtain first and cover
Film MAP0;
2 × 2 block of pixels isolated in the first mask MAP0 6c) are found, and are mapped as original i-th of two field picture block Bi
Block of pixels be added in Sample Storehouse P0, the first order average grader threshold value TH0 of cascade classifier is updated, obtained more
Threshold value TH1=(positive sample average+negative sample average)/4 after new, and 2 × 2 isolated block of pixels are mapped as single pixel sample
It is added in Sample Storehouse P1, the Sample Storehouse P1 ' after being updated;
6d) second level supporting vector machine SVM classifier of cascade classifier is instructed using the Sample Storehouse P1 ' after updating
Practice, and to step 6a) in " disaggregated classification " is carried out by the remaining pixel to be sorted of average grader, i.e., using training
SVM classifier carries out classification processing one by one to remaining pixel to be sorted, further excludes non-targeted pixel, constitutes second and covers
Film MAP1;
6e) the priori that the number of pixels of each connected pixel block in the second mask MAP1 should not be less than K,
Region opening operation processing is done to the second mask MAP1, that is, filters out connected pixel block of the number of pixels less than K=5, obtains i-th
Two field picture block BiChange testing result Ci, and the connected pixel block filtered out is mapped as two field picture block BiThe picture of corresponding position
Vegetarian refreshments is added in Sample Storehouse P1 ', and Sample Storehouse P1 ' is updated again, obtains the Sample Storehouse P1 " after secondary renewal.
Step 7, the average point after the renewal obtained after the change detection completed to previous frame image block is utilized from Jia 1 to i
Sample Storehouse P1 " after class device threshold value TH1 and secondary renewal, detection is changed to current frame image block.
Step 8, repeat the above steps 6- steps 7, constantly updates Sample Storehouse, optimizes the classification performance of cascade classifier, directly
ArriveComplete to two field picture set of blocksChange detection, obtain corresponding change
Testing result set
Step 9, by the change testing result set C derived above to two field picture set of blocks B by from left to right, on to
Under sequential concatenation into final change testing result figure CM1, complete to the first width differential image XLChange detection.
Step 10, by another width differential image X in step 2D, according to the first width differential image XLChange detection
Step 3- steps 9 complete the second width differential image XDChange detection, note change testing result figure be CM2.
Step 11, change the testing result figure CM1 and CM2 of two width differential images are mapped as gray level image A1 and A2, and
Gray level image A1 and A2 are merged using the method for class principal component analysis, the differential image X after being mergedF。
Reference picture 3 steps are implemented as follows:
11a) by the first width differential image XLTesting result figure CM1 change pixel position at pixel value be set to original
Beginning differential image XLThe gray value of corresponding position, and the pixel value at non-changing pixel position is taken 0, obtain the first width poor
Different image XLTesting result figure CM1 mapping after gray level image Y1;
11b) by the second width differential image XDTesting result figure CM2 change pixel position at pixel value be set to original
Beginning differential image XDThe gray value of corresponding position, and the pixel value at non-changing pixel position is taken 0, obtain the second width poor
Different image XDTesting result figure CM2 mapping after gray level image Y2;
Above-mentioned gray level image Y1 and Y2 average 11c) is calculated, and the less image of average is designated as the first average gray-scale map
As A1, the larger image of average is designated as the second average gray level image A2;
11d) respectively by the first average gray level image A1 and the second average gray level image A2 by row major or the preferential side of row
Formula is changed into column vector and calculates covariance matrix;
Characteristic value 11e) is asked for by covariance matrix, the corresponding characteristic vector (x, y) of first principal component is determinedT, wherein " T "
Represent transposition operator;
11f) by the first width differential image XLTesting result figure CM1 and standard change detection contrasted with reference to figure,
The sum of false-alarm number of pixels and missing inspection number of pixels is calculated, E1 is designated as;
11g) by the second width differential image XDTesting result figure CM2 and standard change detection contrasted with reference to figure,
The sum of false-alarm number of pixels and missing inspection number of pixels is calculated, E2 is designated as;
The first average gray level image A1 weight 11h) is determined according to E1 and E2:
Wherein, x be step 11e) in characteristic vector (x, y)TFirst element value, y is characterized vector (x, y)T
Two element values;
11i) calculate the differential image X after being mergedF=w × A1+ (1-w) × A2.
Step 12, with Kmeans clustering algorithms to the differential image X after fusionFClustered, generate final change inspection
Survey result figure XCD, complete the detection to remote sensing image change information.
The effect of the present invention can be further illustrated by following simulation result:
1. experiment condition
Experimental situation is:CPU Intel (R) Core (TM) i5-34703.20GHz, internal memory 8GB, WINDOWS 7 operation system
System, software platform is MATLAB R2013b.
First data set is Switzerland Berne areas SAR remote sensing image data collection, is illustrated in figure 4 ERS-2 images, is schemed
As the shooting time that size is 301 × 301, Fig. 4 (a) and Fig. 4 (b) is respectively in April, 1999 and in May, 1999, Fig. 4 (c)
For canonical reference variation diagram, including 1155 change pixels and 89446 non-changing pixels.
Second data set is the band spectrum images of Landsat-5 satellites TM the 4th in Italy Sardinia areas, figure
As the shooting time that size is 300 × 412, Fig. 9 (a) and Fig. 9 (b) is respectively nineteen ninety-five September and in July, 1996, Fig. 9 (c)
For canonical reference variation diagram, including 7626 change pixels and 115974 non-changing pixels.
2. experimental evaluation index
Quantitative change Analysis of test results can be carried out with reference to the experiment simulation of figure for the change detection with standard, it is main
The evaluation index wanted has:
Flase drop number:Do not change the number of pixels in region in statistical experiment result figure, with not changed with reference in figure
The number of pixels in region is contrasted, with reference to the pixel for not changing and being detected as in experimental result picture change in figure
Number, referred to as flase drop number FP;
Missing inspection number:Change the number of pixels in region in statistical experiment result figure, with the picture with reference to region of variation in figure
Plain number is contrasted, and with reference to being changed in figure but unchanged number of pixels being detected as in experimental result picture, is referred to as leaked
Examine number FN;
Detect total error number OE:Missing inspection number and flase drop number sum;
Correct class probability PCC:
Wherein TP, TN are respectively the number of pixels for being correctly detecting non-changing and change.
Weigh testing result figure and the Kappa coefficients with reference to figure uniformity:Wherein,Here, N represents pixel total number, and Nc and Nu represent actual change respectively
Change pixel count and do not change pixel count.
3. experiment content and experimental result
For SAR remote sensing images, with the inventive method and existing Gong in 2012 in article " Change
Detection in Synthetic Aperture Radar Images based on Image Fusion and Fuzzy
The method proposed in Clustering " is contrasted, and it is Wavelet fusion_RFLICM methods to remember the control methods.
For remote sensing image, with two kinds of change detecting methods in the inventive method and existing 2 patents to second
Individual data set is changed detection, and two kinds of control methods are respectively:The patent application of Xian Electronics Science and Technology University " is based on image
The method for detecting change of remote sensing image of fusion " (number of patent application:201210414782.4, publication number:Carried in 102968790A)
A kind of method for detecting change of remote sensing image based on image co-registration gone out, is designated as Fuse_FCM fusion methods;Xi'an electronics technology is big
Patent application " the remote sensing image change detection based on image co-registration " (number of patent application:201210234076.1, it is public
The number of opening:A kind of remote sensing image change detecting method based on image co-registration proposed in 102750705A), is designated as Fuse_
FLICM fusion methods.
Test 1. the inventive method and Wavelet fusion_RFLICM methods are compared.
Detection is changed to first data set with the inventive method, using the inventive method middle experiment result and
Final testing result is as shown in Figure 5-Figure 8.Wherein, Fig. 5 (a) is the first width differential image that the present invention is constructed, and Fig. 5 (b) is
The second width differential image that the present invention is constructed;Fig. 6 (a) is that the change that detection is obtained is changed to Fig. 5 (a) using the inventive method
Change testing result figure;Fig. 6 (b) is that the change testing result figure that detection is obtained is changed to Fig. 5 (b) using the inventive method;
Fig. 7 is the differential image after the fusion obtained with class PCA, and Fig. 8 is the final change detection knot that the present invention is obtained
Fruit is schemed.
The present invention is compared as shown in table 1 with the change testing result performance indications of Wavelet fusion_RFLICM methods.
The present invention of table 1 is compared with the change testing result performance indications of Wavelet fusion_RFLICM methods
Method | FP | FN | OE | PCC | Kappa |
Wavelet fusion_RFLICM | 133 | 159 | 292 | 99.68% | 0.871 |
The present invention | 120 | 154 | 274 | 99.70% | 0.878 |
As can be seen from Table 1, traditional unsupervised clustering method is better than using the method for the present invention, with good Shandong
Rod;And the accuracy of detection that image co-registration further increases change detection is carried out using class PCA, from final
Change testing result from the point of view of, result figure of the invention has minimum total error number OE, highest accuracy PCC and highest
Kappa coefficients, show it is optimal.
2. the inventive method are tested to be compared with existing Fuse_FCM fusion methods and Fuse_FLICM fusion methods.
Detection, middle experimental result and final testing result are changed to second data set with the inventive method
As shown in Figure 10-Figure 13.Wherein, the first width differential image that Figure 10 (a) constructs for the present invention;Figure 10 (b) constructs for the present invention
The second width differential image;Figure 11 (a) is changed the result figure of detection for the present invention to Figure 10 (a);Figure 11 (b) is the present invention
The result figure of detection is changed to Figure 10 (b);Figure 12 is the differential image after the fusion obtained with class PCA;Figure
13 be the final change testing result figure of the present invention.
Testing result performance indications using of the invention and existing two kinds of change detecting methods are more as shown in table 2.
The present invention of table 2 is compared with the change testing result performance indications in patent
Method | FP | FN | OE | PCC | Kappa |
Fuse_FCM fusion methods | 916 | 1187 | 2103 | 98.30% | 0.8506 |
Fuse_FLICM fusion methods | 1370 | 586 | 1956 | 98.42% | 0.8696 |
The present invention | 907 | 943 | 1850 | 98.50% | 0.8704 |
As can be seen from Table 2, there is good robust using the more traditional unsupervised clustering method of the method for the present invention
Property, FP is relatively balanced with FN, and performance comparision is steady;Change is further increased using the image interfusion method of class principal component analysis
Change the accuracy of detection of detection;From the point of view of final testing result, result figure of the invention has minimum total error number OE, highest
Accuracy PCC and highest Kappa coefficients, show it is optimal.
In summary, the present invention is demonstrated by preferable performance, with institute no matter on subjective effect or in objective indicator
State control methods to compare, total error number is minimum, improve the accuracy of detection of change testing result.
Claims (4)
1. a kind of remote sensing image variation detection method based on on-line study, it is characterised in that:Comprise the following steps:
(1) it is I × J process radiant correction and the remote sensing image X of geometrical registration to obtain two width sizes1And X2, wherein, I is distant
Feel the line number of image, J is the columns of remote sensing image;
(2) two width remote sensing image X are utilized1And X2Construct two width differential image XLAnd XD;
(3) it is directed to the first width differential image XLSample Storehouse P0 and single pixel Sample Storehouse P1 of the construction based on 2 × 2 image blocks, and according to
The threshold value TH0 of the first order, i.e. average grader of Sample Storehouse P0 initialization cascade classifiers;
(4) by the first width differential image XLIt is N × N frame of video forms to be divided into size by order from left to right, from top to bottom
Image block setWherein N is even number, i ∈ Z+And“Z+" it is positive integer,Expression rounds up;
(5) initialisation image set of blocksIndex value i=1, start to the 1st two field picture block B1Become
Change detection;
(6) using cascade classifier to image block setI-th of two field picture block BiIt is changed detection,
And renewal is optimized to cascade classifier;
(7) i is from Jia 1, and the cascade classifier after being updated using optimization is changed detection to next frame image block;
(8) repeat step (6)-step (7), untilThe change to two field picture set of blocks B is completed to examine
Survey, obtain corresponding change testing result set
(9) it is the change testing result set C derived above to two field picture set of blocks B is suitable by from left to right, from top to bottom
Sequence is spliced into final change testing result figure CM1, completes to the first width differential image XLChange detection;
(10) by another width differential image X in step (2)D, according to the first width differential image XLChange detecting step (3)-
Step (9) completes the second width differential image XDChange detection, note change testing result figure be CM2;
(11) change the testing result figure CM1 and CM2 of two width differential images are mapped as gray level image A1 and A2, and use class master
The method of constituent analysis is merged to gray level image A1 and A2, the differential image X after being mergedF;
(12) with Kmeans clustering algorithms to the differential image X after fusionFClustered, generate final change testing result figure
XCD, complete the detection to remote sensing image change information.
2. according to the method described in claim 1, the width remote sensing image X of utilization two wherein described in step (2)1And X2Construct two width
Differential image XLAnd XD, carried out by the type of remote sensing images:
If remote sensing image X1And X2Be SAR remote sensing images, then differential image XLAnd XDStructural formula be respectively:
XL=| logX2-logX1|
<mrow>
<msub>
<mi>X</mi>
<mi>D</mi>
</msub>
<mo>=</mo>
<mn>1</mn>
<mo>-</mo>
<mi>m</mi>
<mi>i</mi>
<mi>n</mi>
<mrow>
<mo>(</mo>
<mfrac>
<msub>
<mi>&mu;</mi>
<mn>1</mn>
</msub>
<msub>
<mi>&mu;</mi>
<mn>2</mn>
</msub>
</mfrac>
<mo>,</mo>
<mfrac>
<msub>
<mi>&mu;</mi>
<mn>2</mn>
</msub>
<msub>
<mi>&mu;</mi>
<mn>1</mn>
</msub>
</mfrac>
<mo>)</mo>
</mrow>
<mo>;</mo>
</mrow>
Wherein, μ1、μ2Respectively X1、X2Regional area gray value average;
If remote sensing image X1And X2Be remote sensing image, then differential image XLAnd XDStructural formula be respectively:
XL=| X1-X2|
XD=| log (1+X2)-log(1+X1)|。
3. according to the method described in claim 1, the utilization cascade classifier wherein described in step (6) is to image block setI-th of two field picture block BiDetection is changed, and renewal is optimized to cascade classifier, by as follows
Step is carried out:
I-th of two field picture block B 6a) is scanned in 2 × 2 overlapping sliding window mode with the first order average grader of cascade classifieriIt is right
It carries out " rough sort ", excludes most of non-targeted pixel, i.e. non-changing pixel, completes the preliminary classification to present frame;
I-th of two field picture block 6b) is scanned in 2 × 2 not overlapping sliding window mode with the first order average grader of cascade classifier
Bi, the classification based on 2 × 2 block of pixels is carried out to it, the first mask MAP0 is constituted;
2 × 2 block of pixels isolated in the first mask MAP0 6c) are found, original i-th of two field picture block B is first mapped asiPicture
Plain block is added in Sample Storehouse P0, and the threshold value for updating the first order average grader of cascade classifier obtains TH1=(positive sample averages
+ negative sample average)/4, then each block of pixels is mapped as single pixel sample addition Sample Storehouse P1;
6d) second level supporting vector machine SVM classifier of cascade classifier is trained using the Sample Storehouse P1 after renewal, and
To step 6a) in " disaggregated classification " is carried out by the remaining pixel to be sorted of average grader, further exclude non-targeted pixel
Point, constitutes the second mask MAP1;
6e) the priori that the number of pixels of each connected pixel block in the second mask MAP1 should not be less than K, to the
Two mask MAP1 do region opening operation processing, filter out the connected pixel block that number of pixels is less than K, obtain i-th of two field picture block Bi
Change testing result Ci, and the connected pixel block filtered out is mapped as two field picture block BiThe pixel of corresponding position is added to
In Sample Storehouse P1, Sample Storehouse P1 is updated again, and wherein K is 5.
4. according to the method described in claim 1, the change testing result by two width differential images wherein described in step (11)
Figure CM1 and CM2 is mapped as gray level image A1 and A2, and gray level image A1 and A2 are melted using the method for class principal component analysis
Close, the differential image X after being mergedF, carry out as follows:
11a) by the first width differential image XLTesting result figure CM1 change pixel position at pixel value be set to original difference
Different image XLThe gray value of corresponding position, and the pixel value at non-changing pixel position is taken 0, obtain the first width disparity map
As XLTesting result figure CM1 mapping after gray level image Y1;
11b) by the second width differential image XDTesting result figure CM2 change pixel position at pixel value be set to original difference
Different image XDThe gray value of corresponding position, and the pixel value at non-changing pixel position is taken 0, obtain the second width disparity map
As XDTesting result figure CM2 mapping after gray level image Y2;
Above-mentioned gray level image Y1 and Y2 average 11c) is calculated, and the less image of average is designated as the first average gray level image
A1, the larger image of average is designated as the second average gray level image A2;
11d) the first average gray level image A1 and the second average gray level image A2 is become by row major or the preferential mode of row respectively
For column vector and calculate covariance matrix;
Characteristic value 11e) is asked for by covariance matrix, the corresponding characteristic vector (x, y) of first principal component is determinedT, wherein " T " is represented
Transposition operator;
11f) by the first width differential image XLTesting result figure CM1 and standard change detection contrasted with reference to figure, calculate void
The sum of alert number of pixels and missing inspection number of pixels, is designated as E1;
11g) by the second width differential image XDTesting result figure CM2 and standard change detection contrasted with reference to figure, calculate void
The sum of alert number of pixels and missing inspection number of pixels, is designated as E2;
The first average gray level image A1 weight 11h) is determined according to described E1 and E2:
<mrow>
<mi>w</mi>
<mo>=</mo>
<mfenced open = "{" close = "">
<mtable>
<mtr>
<mtd>
<mrow>
<mi>y</mi>
<mo>/</mo>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>+</mo>
<mi>y</mi>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
<mtd>
<mrow>
<mi>E</mi>
<mn>1</mn>
<mo>-</mo>
<mi>E</mi>
<mn>2</mn>
<mo>&le;</mo>
<mn>0</mn>
</mrow>
</mtd>
</mtr>
<mtr>
<mtd>
<mrow>
<mi>x</mi>
<mo>/</mo>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>+</mo>
<mi>y</mi>
<mo>)</mo>
</mrow>
</mrow>
</mtd>
<mtd>
<mrow>
<mi>E</mi>
<mn>1</mn>
<mo>-</mo>
<mi>E</mi>
<mn>2</mn>
<mo>></mo>
<mn>0</mn>
</mrow>
</mtd>
</mtr>
</mtable>
</mfenced>
<mo>,</mo>
</mrow>
Wherein, x is characterized vector (x, y)TFirst element value, y is characterized vector (x, y)TSecond element value;
11i) calculate the differential image X after being mergedF=w × A1+ (1-w) × A2.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510112839.9A CN104680542B (en) | 2015-03-15 | 2015-03-15 | Remote sensing image variation detection method based on on-line study |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201510112839.9A CN104680542B (en) | 2015-03-15 | 2015-03-15 | Remote sensing image variation detection method based on on-line study |
Publications (2)
Publication Number | Publication Date |
---|---|
CN104680542A CN104680542A (en) | 2015-06-03 |
CN104680542B true CN104680542B (en) | 2017-10-24 |
Family
ID=53315535
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201510112839.9A Expired - Fee Related CN104680542B (en) | 2015-03-15 | 2015-03-15 | Remote sensing image variation detection method based on on-line study |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN104680542B (en) |
Families Citing this family (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105225227B (en) * | 2015-09-07 | 2018-03-30 | 中国测绘科学研究院 | The method and system of remote sensing image change detection |
CN106056577B (en) * | 2016-05-19 | 2019-02-15 | 西安电子科技大学 | SAR image change detection based on MDS-SRM Mixed cascading |
CN106203521B (en) * | 2016-07-15 | 2019-03-26 | 西安电子科技大学 | The SAR image change detection learnt based on disparity map from step |
CN107248172A (en) * | 2016-09-27 | 2017-10-13 | 中国交通通信信息中心 | A kind of remote sensing image variation detection method based on CVA and samples selection |
CN107368781B (en) * | 2017-06-09 | 2019-08-20 | 陕西师范大学 | Synthetic Aperture Radar images change detecting method based on Subspace partition |
CN108053409B (en) * | 2017-12-11 | 2022-05-13 | 中南大学 | Automatic construction method and system for remote sensing image segmentation reference library |
CN108776772B (en) * | 2018-05-02 | 2022-02-08 | 北京佳格天地科技有限公司 | Cross-time building change detection modeling method, detection device, method and storage medium |
CN109255388B (en) * | 2018-09-28 | 2021-12-31 | 西北工业大学 | Unsupervised heterogeneous remote sensing image change detection method |
CN111091054B (en) * | 2019-11-13 | 2020-11-10 | 广东国地规划科技股份有限公司 | Method, system and device for monitoring land type change and storage medium |
CN111539296B (en) * | 2020-04-17 | 2022-09-23 | 河海大学常州校区 | Method and system for identifying illegal building based on remote sensing image change detection |
CN113033510B (en) * | 2021-05-21 | 2021-10-15 | 浙江大华技术股份有限公司 | Training and detecting method, device and storage medium for image change detection model |
CN113538536B (en) * | 2021-07-21 | 2022-06-07 | 中国人民解放军国防科技大学 | SAR image information-assisted remote sensing optical image dense cloud detection method and system |
CN114973028B (en) * | 2022-05-17 | 2023-02-03 | 中国电子科技集团公司第十研究所 | Aerial video image real-time change detection method and system |
CN116402693B (en) * | 2023-06-08 | 2023-08-15 | 青岛瑞源工程集团有限公司 | Municipal engineering image processing method and device based on remote sensing technology |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101894125A (en) * | 2010-05-13 | 2010-11-24 | 复旦大学 | Content-based video classification method |
CN102789578A (en) * | 2012-07-17 | 2012-11-21 | 北京市遥感信息研究所 | Infrared remote sensing image change detection method based on multi-source target characteristic support |
CN103500450A (en) * | 2013-09-30 | 2014-01-08 | 河海大学 | Multi-spectrum remote sensing image change detection method |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US7929729B2 (en) * | 2007-04-02 | 2011-04-19 | Industrial Technology Research Institute | Image processing methods |
-
2015
- 2015-03-15 CN CN201510112839.9A patent/CN104680542B/en not_active Expired - Fee Related
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101894125A (en) * | 2010-05-13 | 2010-11-24 | 复旦大学 | Content-based video classification method |
CN102789578A (en) * | 2012-07-17 | 2012-11-21 | 北京市遥感信息研究所 | Infrared remote sensing image change detection method based on multi-source target characteristic support |
CN103500450A (en) * | 2013-09-30 | 2014-01-08 | 河海大学 | Multi-spectrum remote sensing image change detection method |
Non-Patent Citations (2)
Title |
---|
"Change Detection in Synthetic Aperture Radar Images based on Image Fusion and Fuzzy Clustering";Maoguo Gong 等;《IEEE TRANSACTIONS ON IMAGE PROCESSING》;20120430;第21卷(第4期);第2141-2151页 * |
"基于小波域Fisher分类器的SAR 图像变化检测";辛芳芳等;《红外与毫米波学报》;20110430;第30卷(第2期);第173-178页 * |
Also Published As
Publication number | Publication date |
---|---|
CN104680542A (en) | 2015-06-03 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104680542B (en) | Remote sensing image variation detection method based on on-line study | |
CN110287932B (en) | Road blocking information extraction method based on deep learning image semantic segmentation | |
CN105303193B (en) | A kind of passenger number statistical system based on single-frame images processing | |
CN106355188A (en) | Image detection method and device | |
CN108710913A (en) | A kind of switchgear presentation switch state automatic identification method based on deep learning | |
CN106875380B (en) | A kind of heterogeneous image change detection method based on unsupervised deep neural network | |
CN104182985B (en) | Remote sensing image change detection method | |
CN109784203A (en) | Weakly supervised x-ray image contraband inspection method based on delamination propagation and activation | |
CN105844279A (en) | Depth learning and SIFT feature-based SAR image change detection method | |
CN102867183B (en) | Method and device for detecting littered objects of vehicle and intelligent traffic monitoring system | |
CN108446616B (en) | Road extraction method based on full convolution neural network ensemble learning | |
CN115731164A (en) | Insulator defect detection method based on improved YOLOv7 | |
CN108447055A (en) | SAR image change detection based on SPL and CCN | |
CN104361351A (en) | Synthetic aperture radar (SAR) image classification method on basis of range statistics similarity | |
CN106022254A (en) | Image recognition technology | |
CN109446894A (en) | The multispectral image change detecting method clustered based on probabilistic segmentation and Gaussian Mixture | |
CN108537790A (en) | Heterologous image change detection method based on coupling translation network | |
CN103093243B (en) | The panchromatic remote sensing image clouds of high-resolution sentences method | |
CN110298410A (en) | Weak target detection method and device in soft image based on deep learning | |
CN115294473A (en) | Insulator fault identification method and system based on target detection and instance segmentation | |
CN109002792B (en) | SAR image change detection method based on layered multi-model metric learning | |
CN105205807B (en) | Method for detecting change of remote sensing image based on sparse automatic coding machine | |
CN107516082A (en) | Based on the SAR image change region detection method from step study | |
CN114821102A (en) | Intensive citrus quantity detection method, equipment, storage medium and device | |
CN111242899A (en) | Image-based flaw detection method and computer-readable storage medium |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant | ||
CF01 | Termination of patent right due to non-payment of annual fee | ||
CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20171024 |