CN107194313A - A kind of parallel intelligent object-oriented classification method - Google Patents
A kind of parallel intelligent object-oriented classification method Download PDFInfo
- Publication number
- CN107194313A CN107194313A CN201710255550.1A CN201710255550A CN107194313A CN 107194313 A CN107194313 A CN 107194313A CN 201710255550 A CN201710255550 A CN 201710255550A CN 107194313 A CN107194313 A CN 107194313A
- Authority
- CN
- China
- Prior art keywords
- mrow
- classification
- msub
- kappa
- svm
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
Abstract
A kind of parallel intelligent object-oriented classification method, step is as follows:(1) high-definition remote sensing image data is read in;(2) MeanShift partitioning parameters change collection Q is determinedSRAnd segmentation image set IMS;(3) class categories number is determined, and reads in training sample SC, step (2) segmentation result is classified using svm classifier algorithm, classification image set I is obtainedSVM;(4) according to classification image set ISVMWith checking sample VCConfusion matrix is calculated, Kappa coefficients are calculated, Kappa coefficient sets K is formedSR;(5) data processing parallel schema is selected, Thread Count T is set using;(6) to Kappa coefficient sets KSRWith spatial domain bandwidth Hs, color gamut bandwidth HrFunctional relation analysis is carried out, maximum Kappa is calculatedmaxCorresponding HSKAnd HRK, and corresponding classification results.Advantage of the present invention:Depart from traditional object oriented classification by man-machine interactively formula parameter adjustment mode, reduce operating personnel's professional standards and rely on;With unified evaluation criterion;MPI parallel processing mechanism is incorporated, parallel efficient distributed treatment is completed, improves efficiency.
Description
Technical field
The present invention relates to a kind of parallel intelligent object-oriented classification method, belong to Remote Image Classification field.
Background technology
Classification of remote-sensing images algorithm is distinguished from processing unit, mainly includes two major classes:Sorting technique based on pixel and
Object-based sorting technique.The nicety of grading of atural object depend on spectrum inside mixed pixel number and classification make a variation two because
Element:The raising of spatial resolution, causes the homogeneous atural object mixed pixel number of texture to reduce, improves the precision of classification;And for light
Spectral response variation increase inside spectral space, its heterogeneous big atural object classification, so as to cause the reduction of classification separability.High-altitude
Between resolution remote sense be following development trend, its remote sensing image information enriches, containing much information feature, and based on pixel
Sorting technique can only start with from single spectral signature, and the precision of classification can not be met, while the spiced salt effect produced can be reduced
Classifying quality.Therefore, for the terrain classification of high spatial resolution remote sense image, not only need to consider the Spectral Properties of remote sensing image
Levy, also need to consider its space characteristics.
Object-based classification processing unit is no longer single pixel, but the homogeney in the class that several pixels are constituted
Heterogeneous maximum " homogeneity " pixel group, i.e. destination object between maximum and class.Using object as processing unit, by the light of remote sensing image
Spectrum signature and space characteristics participate in sorting algorithm as object properties joint, more conform to the foundation of mankind's visual interpretation, improve
Classification results precision.The generation of imaged object comes from Image Segmentation, and the first step of object oriented classification is into one by Image Segmentation
Each and every one " homogeneity " object, the quality of Image Segmentation directly affects the precision of classification results.It is usually used in the shadow in object oriented classification
As dividing method mainly has multi-scale division, watershed segmentation, average drift segmentation, Area generation segmentation.These partitioning algorithms
Although obtaining preferable experimental result, still need to carry out lot of experiments, effect by man-machine interactively in terms of partitioning parameters setting
Rate is low, and relies on the professional standards of operating personnel, unifies reasonably to quantify evaluation criterion, fails to reach practical
Stage.Accordingly, it would be desirable to which a kind of innovatory algorithm of efficient self-optimization, completes partitioning parameters automation and determines, in combination with simultaneously
Row treatment technology, improves intellectuality and the parallelization degree of object oriented classification technology.
The content of the invention
The nicety of grading existed for existing object-oriented classification method depends on treatment people professional standards, quality mark unduly
The problem of standard is difficult to ensure disunity and low treatment effeciency, the present invention proposes a kind of parallel intelligent object oriented classification side
Method, i.e., under parallel environment support, evaluate, automation determines the intelligent face of optimum segmentation parameter according to quantification nicety of grading
To object method.
The present invention technical solution be:For high-resolution remote sensing image, first installation space field parameter and color
Field parameter performs MeanShift (average drifting) to remote sensing image and split, and combining classification sample data is to the shadow after segmentation afterwards
As carrying out SVM (Support Vector Machine SVMs) image classification, finally using verifying that sample is changed surely
The nicety of grading evaluation of amount, calculates confusion matrix and evaluating Kappa.And segmentation ginseng is altered in steps with certain step-length s
Number, continues executing with above-mentioned steps, in a series of evaluating is obtained, optimal evaluating is obtained by statistical analysis
Kappa values optimum segmentation parameter corresponding with its, then obtains optimal classification results.This is one with quantification nicety of grading
For highest, the thought flow of optimal classification result is determined, while parallel using MPI (Message Passing Interface)
Operation rule, will calculate the process distributed treatment of a series of Kappa values, shortens total algorithm and performs the time, improves operation
Efficiency.
A kind of parallel intelligent object-oriented classification method, specifically includes following steps:
(1) high-definition remote sensing image data is read in;
(2) MeanShift partitioning parameters change collection Q is determinedSRAnd segmentation image set IMS:Including spatial domain bandwidth parameter HS
With color gamut bandwidth parameter HRChange integrate as QSAnd QR, and change step StepSAnd StepR, utilize gaussian kernel functionPerform MeanShift partitioning algorithms to split high-resolution data, formed
Split image set IMS;
(3) class categories number is determined, and reads in training sample SC, using svm classifier algorithm to the segmentation in step (2)
As a result classified, obtain classification image set ISVM;
(4) according to the classification image set I in step (3)SVMWith checking sample VCConfusion matrix is calculated, and calculates Kappa systems
Number, forms Kappa coefficient sets KSR;
(5) data processing parallel schema is selected, and is set using Thread Count T;
(6) to Kappa coefficient sets KSRWith spatial domain bandwidth Hs, color gamut bandwidth HrFunctional relation analysis is carried out, is calculated
Maximum KappamaxCorresponding HSKAnd HRK, and corresponding classification results.
Wherein, the distant influence sense data of high-resolution described in step (1) are three wave bands, and by red, green, blue subband order
The true color image data of combination;
Wherein, the MeanShift segmentations in step (2), depend on two-dimentional kernel function
K in formulaNormal(x) it is Gaussian function, xSIt is position vector in spatial domain, xRIt is color gamut chroma vector; HS, HR
It is spatial domain bandwidth and color gamut bandwidth respectively, and HS∈QS, HR∈QR, QSR=QS∪QR, two bandwidth parameter HS, HRChange
Determine MeanShift segmentation results, partitioning parameters change collection QSRWith segmentation image set IMSTo correspond, c is that standardization is normal
Amount so that two-dimensional nucleus function expressionIntegration be 1.
Wherein, segmentation image set I of the svm classifier described in step (3) in step (2)MSOn the basis of perform, core is
It is two parameters for determining classification surface function to determine optimal classification surface function Svm (x)=sgn (ω x+b), wherein ω and b.It is main
Flow is wanted to be divided into the following steps:
A. High Dimensional Mapping:Will segmentation image set IMSIn an image ImsiWith training sample SC, pass through RBF
(Radial Basis Function) carries out High Dimensional Mapping, and wherein RBF RBF mathematical expression is as follows:
S, t represent training sample characteristic vector respectively in formula,It is characterized vector distribution variance;
B. training sample is carried out learning to obtain optimal classification surface:Calculating is learnt by training sample and obtains ω and b, so that
Optimal classification surface Svm (x) is determined;
C. to segmentation image ImsiCarry out classification extraction:Its characteristic vector is substituted into optimal classification surface equation Svm (x), area
It is sub-category, obtain classification image Isvmi;
D. segmentation image set I is corresponded toMSThen there is one-to-one classification image set ISVM。
Kappa is a kind of whether believable measurement of classification of assessment result wherein in step (4), calculates basis and depends on and obscures
Matrix, main calculation expression is:
In formula, k is confusion matrix line number, xiiThe sample points for being i for remote sensing classification and ground reference classification, xi+And x+i
The sum of i-th row and the i-th row respectively in confusion matrix, N is total number of samples;On this basis, checking sample V is combined firstC
With classification image set ISVMIn each classification image IsvmiGenerate corresponding confusion matrix MConi, reuse above-mentioned calculation expression
Formula, calculates Kappa values, finally obtains Kappa coefficient sets KSR, it is known that KSRWith ISVM、IMS、QSRCorrespond;
Wherein, step (5) is parallel processing strategy, and four steps are understood more than, from high resolution image data to
Kappa coefficient sets KSRIt is the flow (such as Fig. 1) of a set of object oriented classification and precision quantitative assessment, due to except changing parameter
Collect QSRThe H of middle partitioning parametersSAnd HROutside, do not occur simultaneously between above-mentioned sub-process and sub-process, condition meets parallel processing will
Ask:Each sub-line journey tiThe above-mentioned flow of isolated operation, obtains Kappa COEFFICIENT K i, finally remerges to form Kappa coefficient sets KSR,
Complete whole processing procedure (such as Fig. 2).Thread Count is set using for T, serial total operand is M, then parallel speedup ratio isM=(Length (Q in formulaS))/StepS×(Length(QR))/StepR, Length (QR)=RC-RF, SCAnd SF、
RCAnd RFIt is partitioning parameters change collection Q respectivelySAnd QRUpper and lower bound, StepS, StepRRespectively partitioning parameters HSAnd HRChange
Change step-length;
Wherein, Kappa coefficient sets K in step (6)SRIt is considered as variable HSAnd HRValue collection, i.e., by formula Kappa=K (HS,
HR) represent, because function analytic expression K (x) can not be known, by experimental data, compared with reference to statistical analysis, obtain maximum
KappamaxH corresponding with itsSKAnd HRK, and from image classification collection ISVMCorresponding optimal classification imaging results ISVMax。
The advantage of the present invention compared with prior art is:Overcome need during traditional object oriented classification it is a large amount of artificial
The limitation such as interaction parameter adjustment, treatment effeciency be low, this method is located parallel using quantification evaluation of classification COEFFICIENT K appa and MPI
Reason mechanism realizes Efficient intelligent object oriented classification.It has advantages below:(1) to optimize quantitative assessment parameter as convergence
Criterion, self-feedback formula determines parameter in optimal classification, is joined so as to depart from traditional object oriented classification by man-machine interactively formula
Number adjustment modes, the professional standards for reducing operating personnel are relied on;(2) classification results are up to target with quantification precision evaluation,
With unified evaluation criterion;(3) MPI parallel processing mechanism is incorporated in whole processing procedure, parallel efficient distribution is completed
Formula processing, improves the whole efficiency of method.
Brief description of the drawings
Fig. 1 is object oriented classification and precision Quantitative Evaluation Algorithm flow chart in the present invention.
Fig. 2 is parallelization object oriented classification innovatory algorithm flow.
Fig. 3 is that the Dujiang weir that resolution ratio is 0.5 meter is taken photo by plane true color remote sensing data.
Fig. 4 (a) is Kappa coefficients and partitioning parameters H in 285 groups of experimental resultsRChange curve.
Fig. 4 (b) is Kappa coefficients and partitioning parameters H in 285 groups of experimental resultsSChange curve.
After Fig. 5 (a) is statistical analysis, Kappa coefficients and partitioning parameters HRChange curve.
After Fig. 5 (b) is statistical analysis, Kappa coefficients and partitioning parameters HSChange curve.
Fig. 6 (a) is optimal segmentation binary map.
Fig. 6 (b) is optimal images segmentation figure, and Fig. 6 (c) is optimal classification result figure.
Embodiment
In order to better illustrate object-oriented classification method of the present invention, the Dujiang weir that resolution ratio is 0.5 meter is utilized
Remotely-sensed data of taking photo by plane carries out the classification of Wenchuan earthquake Earthquake damage information and extracted.A kind of parallel intelligent object oriented classification side of the present invention
Method, implements step as follows:
(1) high-definition remote sensing image data is read in:Read in Dujiang weir Airborne Data Classification after Wenchuan earthquake shake, space point
Resolution is 0.5 meter, and subband order is the true color image (such as Fig. 3) of red, green, blue.
(2) concrete meaning of parameter in Kernel Function is split according to MeanShift remote sensing:
HSRepresent and calculate window size in the vectorial time space domains of MeanShift, and HRFor window size in color gamut.And pass through
Cross experiment and primarily determine that and work as HRFor 90 when, test block image is almost divided into a classification, so HRLower limit since 1, so
H is setRCodomain QRIt is set to [1,90], i.e. RC=1, RF=90, work as HRStep-length Step when=1R4 are set to, other value time steps are taken
Long StepRIt is set to 5;And be building actual size in 0.5 meter and image according to the resolution ratio of image, then HSWith building chi
Very little correspondence window size is slightly less than the value for the upper limit, i.e. HSCodomain QSIt is set to [5,19], i.e. SC=5, SF=19, change step
Long StepS=1, so initial data is performed after MeanShift segmentations, generate Image Segmentation collection IMSThere are 285 components to cut
As a result;
(3) class categories number is determined, and selects training sample, according to referring on the spot, class categories number C=6 is determined
(being respectively shade, road, vegetation, water body, intact building and collapsed building), training sample SCWith test sample VCAccording to
Checking and artificial visual are combined mode and determined on the spot.The Image Segmentation collection I generated to (2) stepMSSvm classifier is carried out one by one,
Generation classification image set ISVM, its dimension has 285 groups of classification results as Image Segmentation collection, that is,;
(4) classify image set I to (3) step resultSVMAdd checking sample VC, counted using confusion matrix algorithm, and one by one
Calculate Kappa coefficients and obtain Ki, Kappa coefficient sets K is obtained by 285 cycle calculationsSR;
(5) exist in experimental procedure (2)~(4) it is orthogonal and can independent operating 285 procedure computings, it is parallel real
Test has 15 using the conode of Distributed Calculation one, and each node has 8 double-core CPU, so to maximally utilize calculating energy
Power, opens up Thread Count T=15*8*2=240 altogether, and parallel runtime and speed-up ratio such as table 1 below (test serial arithmetic and simultaneously
The run time and speed-up ratio of row computing) list:
Table 1
(6) the Kappa coefficient sets K obtained with reference to 285 groups of experimental resultsSR, with corresponding partitioning parameters HSAnd HR, carry out
Statistical analysis, finds the corresponding optimized parameter of maximum Kappa values, so as to find optimal classification result.First from experimental data
Take respectively and list HSRespectively 5,10,15,19 and HRPart Experiment classification results when respectively 5,20,60,90, scheme simultaneously
In corresponding Kappa value lists in table 2 below (test block part classifying result Kappa values):
Table 2
This table reflects hrValue have considerable influence to nicety of grading, work as HSSmaller (HS≤ 15) when, Kappa values are with HR
Increase and increase, with HRConstantly increase, and excessive value (HR>=60) cause serious wrong point result;Work as HRSmaller (HR≤20)
When, Kappa values are to HSChange it is insensitive;With HRIncrease (HR>=60), Kappa is with HSIncrease and increase, but general classification
Precision is relatively low.To obtain precise results, all results that Comprehensive Experiment is obtained are asked and work as HSRespectively 5 to 19 is interval default
During value, Kappa and H are obtainedRChange curve (such as Fig. 4 (a) shown in);Work as HRIt is fixed take 1 to 90 between definite value when, Kappa with
HSChange curve (such as Fig. 4 (b) shown in).It can analyze and learn from figure, Kappa values are with variable HRAnd HSThe obtained rule of change
Rule meets above-mentioned conclusion.Kappa values are to HRVariable is sensitive, with HRIncrease present and first increase the rule and in H of rising afterwardsRReached when=20
To peak value, and work as HRWhen >=50, the change of Kappa values is by HSInfluence, larger Characteristic fluctuation is presented;Kappa values are with HS's
Change is irregular to follow, and works as HRWhen≤60, HSIncrease Kappa is not influenceed, work as HRWhen >=60, Kappa values are with HS's
Incremental variations big rise and fall, without rule.
Average Kappa coefficient values are tried to achieve to the Kappa curves in Fig. 4 (a) with hrThe curve (Fig. 5 (a)) of change, can
Work as H to seeRAverage Kappa coefficients peak in (crosspoint of dotted line in figure) when=20, and by qualitative analysis above
Understand Kappa coefficients mainly by hrChange influence, so HR=20 be the maximum necessary conditions of Kappa.Fig. 5 (b) displays are worked as
HRKappa coefficients are with variable H when=20SSituation of change, it can be deduced that Kappa is to HSChange is insensitive.Due to HSIncrease
A large amount of consumption of operation time can be caused, although certain smooth improvement, but larger H can be played to resultSIt is inadvisable
's.H is taken for thisSChange middle number (dotted line crosspoint in figure) as optimal Kappa another condition.Obtained according to test block
It is H to optimum segmentation parameter combinationS=12, HR=20.Optimal segmentation result and classification results such as Fig. 6 are obtained using parameter combination
(a), Fig. 6 (b) and 6 (c) are shown.
Optimum segmentation ginseng can be automatically determined according to optimal classification evaluating Kappa from that can be drawn using the inventive method
Array is combined into HS=12, HR=20, intelligent determination parameter objectives are reached, and highest Kappa is obtained in whole 285 groups of experiments
It is worth for 0.73.By using paralleling tactic, the speed-up ratio for improving parallel method is 16.47, the time by it is serial when it is 3.13 small
When foreshorten to 11 minutes or so, greatly enhance the efficiency of sequential operation, complete Parallel Algorithm processing.
Claims (7)
1. a kind of parallel intelligent object-oriented classification method, it is characterised in that:This method specifically includes following steps:
(1) high-definition remote sensing image data is read in;
(2) MeanShift partitioning parameters change collection Q is determinedSRAnd segmentation image set IMS:Including spatial domain bandwidth parameter HSAnd face
Colour gamut bandwidth parameter HRChange integrate as QSAnd QR, and change step StepSAnd StepR, utilize gaussian kernel functionPerform MeanShift partitioning algorithms to split high-resolution data, formed
Split image set IMS;
(3) class categories number is determined, and reads in training sample SC, using svm classifier algorithm to the segmentation result in step (2)
Classified, obtain classification image set ISVM;
(4) according to the classification image set I in step (3)SVMWith checking sample VCConfusion matrix is calculated, and calculates Kappa coefficients, shape
Into Kappa coefficient sets KSR;
(5) data processing parallel schema is selected, and is set using Thread Count T;
(6) to Kappa coefficient sets KSRWith spatial domain bandwidth Hs, color gamut bandwidth HrFunctional relation analysis is carried out, maximum is calculated
KappamaxCorresponding HSKAnd HRK, and corresponding classification results.
2. a kind of parallel intelligent object-oriented classification method according to claim 1, it is characterised in that:In step (1)
Described high-definition remote sensing image data is three wave bands, and by the true color image data of red, green, blue subband order combination.
3. a kind of parallel intelligent object-oriented classification method according to claim 1, it is characterised in that:In step (2)
MeanShift segmentation, depend on two-dimentional kernel function
<mrow>
<msub>
<mi>G</mi>
<mrow>
<msub>
<mi>H</mi>
<mi>S</mi>
</msub>
<mo>,</mo>
<msub>
<mi>H</mi>
<mi>R</mi>
</msub>
</mrow>
</msub>
<mo>=</mo>
<mfrac>
<mi>c</mi>
<mrow>
<msubsup>
<mi>H</mi>
<mi>S</mi>
<mn>2</mn>
</msubsup>
<msubsup>
<mi>H</mi>
<mi>R</mi>
<mn>3</mn>
</msubsup>
</mrow>
</mfrac>
<msub>
<mi>K</mi>
<mrow>
<mi>N</mi>
<mi>o</mi>
<mi>r</mi>
<mi>m</mi>
<mi>a</mi>
<mi>l</mi>
</mrow>
</msub>
<mrow>
<mo>(</mo>
<mo>|</mo>
<mo>|</mo>
<mfrac>
<msup>
<mi>x</mi>
<mi>S</mi>
</msup>
<msub>
<mi>H</mi>
<mi>S</mi>
</msub>
</mfrac>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
<mo>)</mo>
</mrow>
<msub>
<mi>K</mi>
<mrow>
<mi>N</mi>
<mi>o</mi>
<mi>r</mi>
<mi>m</mi>
<mi>a</mi>
<mi>l</mi>
</mrow>
</msub>
<mrow>
<mo>(</mo>
<mo>|</mo>
<mo>|</mo>
<mfrac>
<msup>
<mi>x</mi>
<mi>R</mi>
</msup>
<msub>
<mi>H</mi>
<mi>R</mi>
</msub>
</mfrac>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
<mo>)</mo>
</mrow>
</mrow>
K in formulaNormal(x) it is Gaussian function, xSIt is position vector in spatial domain, xRIt is color gamut chroma vector;HS, HRIt is respectively
Spatial domain bandwidth and color gamut bandwidth, and HS∈QS, HR∈QR, QSR=QS∪QR, two bandwidth parameter HS, HRChange determine
MeanShift segmentation results, partitioning parameters change collection QSRWith segmentation image set IMSTo correspond, c is normalization constants, is made
Obtain two-dimensional nucleus function expressionIntegration be 1.
4. a kind of parallel intelligent object-oriented classification method according to claim 1, it is characterised in that:In step (3)
Described svm classifier is the segmentation image set I in step (2)MSOn the basis of perform, core is to determine optimal classification surface function
Svm (x)=sgn (ω x+b), wherein ω and b are two parameters for determining classification surface function, and main flow is divided into following several
Step:
A. High Dimensional Mapping:Will segmentation image set IMSIn an image ImsiWith training sample SC, carried out by RBF high
Dimension mapping, wherein RBF RBF mathematical expression is as follows:
<mrow>
<msub>
<mi>K</mi>
<mrow>
<mi>R</mi>
<mi>B</mi>
<mi>F</mi>
</mrow>
</msub>
<mrow>
<mo>(</mo>
<mi>x</mi>
<mo>,</mo>
<mi>y</mi>
<mo>)</mo>
</mrow>
<mo>=</mo>
<mi>exp</mi>
<mrow>
<mo>(</mo>
<mo>-</mo>
<mfrac>
<mrow>
<mo>|</mo>
<mo>|</mo>
<mi>s</mi>
<mo>-</mo>
<mi>t</mi>
<mo>|</mo>
<msup>
<mo>|</mo>
<mn>2</mn>
</msup>
</mrow>
<mrow>
<mn>2</mn>
<msup>
<mo>&part;</mo>
<mn>2</mn>
</msup>
</mrow>
</mfrac>
<mo>)</mo>
</mrow>
</mrow>
S, t represent training sample characteristic vector respectively in formula,It is characterized vector distribution variance;
B. training sample is carried out learning to obtain optimal classification surface:Calculating is learnt by training sample and obtains ω and b, so that it is determined that
Optimal classification surface Svm (x);
C. to segmentation image ImsiCarry out classification extraction:Its characteristic vector is substituted into optimal classification surface equation Svm (x), region class
Not, classification image set I is obtainedsvmi;
D. segmentation image set I is corresponded toMSThen there is one-to-one classification image set ISVM。
5. a kind of parallel intelligent object-oriented classification method according to claim 1, it is characterised in that:In step (4)
Described Kappa is a kind of whether believable measurement of classification of assessment result, calculates basis and depends on confusion matrix, main computational chart
It is up to formula:
<mrow>
<mi>K</mi>
<mi>a</mi>
<mi>p</mi>
<mi>p</mi>
<mi>a</mi>
<mo>=</mo>
<mfrac>
<mrow>
<mi>N</mi>
<munderover>
<mi>&Sigma;</mi>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>k</mi>
</munderover>
<msub>
<mi>x</mi>
<mrow>
<mi>i</mi>
<mi>i</mi>
</mrow>
</msub>
<mo>-</mo>
<munderover>
<mi>&Sigma;</mi>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>k</mi>
</munderover>
<msub>
<mi>x</mi>
<mrow>
<mi>i</mi>
<mo>+</mo>
</mrow>
</msub>
<mo>&times;</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>+</mo>
<mi>i</mi>
</mrow>
</msub>
</mrow>
<mrow>
<msup>
<mi>N</mi>
<mn>2</mn>
</msup>
<mo>-</mo>
<munderover>
<mi>&Sigma;</mi>
<mrow>
<mi>i</mi>
<mo>=</mo>
<mn>1</mn>
</mrow>
<mi>k</mi>
</munderover>
<msub>
<mi>x</mi>
<mrow>
<mi>i</mi>
<mo>+</mo>
</mrow>
</msub>
<mo>&times;</mo>
<msub>
<mi>x</mi>
<mrow>
<mo>+</mo>
<mi>i</mi>
</mrow>
</msub>
</mrow>
</mfrac>
</mrow>
In formula, k is error matrix line number, xiiThe sample points for being i for remote sensing classification and ground reference classification, xi+And x+iRespectively
For the i-th row in error matrix and the i-th sum arranged, N is total number of samples;On this basis, with reference to checking sample VCTo classification shadow
Image set ISVMKappa calculating is carried out, Kappa coefficient sets K is obtainedSR, it is known that KSRWith ISVM、IMS、QSRCorrespond.
6. a kind of parallel intelligent object-oriented classification method according to claim 1, it is characterised in that:The step
(5) it is parallel processing strategy, four steps are understood more than, from high resolution image data to Kappa coefficient sets KSRIt is one
The flow of object oriented classification and precision quantitative assessment is covered, due to except changing parameter set QSRThe H of middle partitioning parametersSAnd HROutside,
Do not occur simultaneously between above-mentioned sub-process and sub-process, condition meets Parallel processing demands:Each sub-line journey tiIn isolated operation
Flow is stated, Kappa COEFFICIENT K i are obtained, finally remerges to form Kappa coefficient sets KSR, complete whole processing procedure;It is set using
Thread Count is T, and serial total operand is M, then parallel speedup ratio isM=(Length (Q in formulaS))/StepS×
(Length(QR))/StepR, Length (QR)=RC-RF, SCAnd SF、RCAnd RFIt is partitioning parameters change collection Q respectivelySAnd QRIt is upper
Limit and lower limit, StepS, StepRRespectively partitioning parameters HSAnd HRChange step.
7. a kind of parallel intelligent object-oriented classification method according to claim 1, it is characterised in that:In step (6)
Kappa coefficient sets KSRIt is considered as variable HSAnd HRValue collection, i.e., by formula Kappa=K (HS,HR) represent, due to function analytic expression
K (x) can not be known, by experimental data, compared with reference to statistical analysis, obtain maximum KappamaxH corresponding with itsSKAnd HRK,
And from image classification collection ISVMCorresponding optimal classification imaging results ISVMax。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710255550.1A CN107194313A (en) | 2017-04-19 | 2017-04-19 | A kind of parallel intelligent object-oriented classification method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710255550.1A CN107194313A (en) | 2017-04-19 | 2017-04-19 | A kind of parallel intelligent object-oriented classification method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107194313A true CN107194313A (en) | 2017-09-22 |
Family
ID=59870986
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710255550.1A Pending CN107194313A (en) | 2017-04-19 | 2017-04-19 | A kind of parallel intelligent object-oriented classification method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107194313A (en) |
Cited By (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107909039A (en) * | 2017-11-16 | 2018-04-13 | 武汉大学 | The ground mulching sorting technique of high-resolution remote sensing image based on parallel algorithm |
CN109858588A (en) * | 2019-01-08 | 2019-06-07 | 哈尔滨理工大学 | A kind of two dimensional code parallel generation method based on chaotic maps |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6266442B1 (en) * | 1998-10-23 | 2001-07-24 | Facet Technology Corp. | Method and apparatus for identifying objects depicted in a videostream |
CN103426158A (en) * | 2012-05-17 | 2013-12-04 | 中国科学院电子学研究所 | Method for detecting two-time-phase remote sensing image change |
CN103729653A (en) * | 2014-01-21 | 2014-04-16 | 武汉大学 | Supervision change detection method of high-resolution remote sensing images |
-
2017
- 2017-04-19 CN CN201710255550.1A patent/CN107194313A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US6266442B1 (en) * | 1998-10-23 | 2001-07-24 | Facet Technology Corp. | Method and apparatus for identifying objects depicted in a videostream |
CN103426158A (en) * | 2012-05-17 | 2013-12-04 | 中国科学院电子学研究所 | Method for detecting two-time-phase remote sensing image change |
CN103729653A (en) * | 2014-01-21 | 2014-04-16 | 武汉大学 | Supervision change detection method of high-resolution remote sensing images |
Non-Patent Citations (1)
Title |
---|
张伟: "《遥感震害信息智能提取优化方法研究》", 《中国优秀硕士学位论文全文数据库 基础科学辑》 * |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107909039A (en) * | 2017-11-16 | 2018-04-13 | 武汉大学 | The ground mulching sorting technique of high-resolution remote sensing image based on parallel algorithm |
CN107909039B (en) * | 2017-11-16 | 2020-03-10 | 武汉大学 | High-resolution remote sensing image earth surface coverage classification method based on parallel algorithm |
CN109858588A (en) * | 2019-01-08 | 2019-06-07 | 哈尔滨理工大学 | A kind of two dimensional code parallel generation method based on chaotic maps |
CN109858588B (en) * | 2019-01-08 | 2022-10-18 | 哈尔滨理工大学 | Chaos mapping-based two-dimensional code parallel generation method |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN104966085B (en) | A kind of remote sensing images region of interest area detecting method based on the fusion of more notable features | |
CN109671124B (en) | Mineral flotation froth image color correction method and froth color detection system | |
CN104182985B (en) | Remote sensing image change detection method | |
CN104834912A (en) | Weather identification method and apparatus based on image information detection | |
EP2638508B1 (en) | System and method for identifying complex tokens in an image | |
CN109657610A (en) | A kind of land use change survey detection method of high-resolution multi-source Remote Sensing Images | |
CN106294705B (en) | A kind of batch remote sensing image preprocess method | |
CN102974551A (en) | Machine vision-based method for detecting and sorting polycrystalline silicon solar energy | |
CN104966285A (en) | Method for detecting saliency regions | |
CN109829507B (en) | Aerial high-voltage transmission line environment detection method | |
US8559714B2 (en) | Post processing for improved generation of intrinsic images | |
CN108846474A (en) | The satellite cloud picture cloud amount calculation method of convolutional neural networks is intensively connected based on multidimensional | |
US20140294296A1 (en) | Spatially varying log-chromaticity normals for use in an image process | |
CN111667461B (en) | Abnormal target detection method for power transmission line | |
CN116485785B (en) | Surface defect detection method for solar cell | |
CN108898096A (en) | A kind of quick accurate extracting method of the information towards high score image | |
CN114463637A (en) | Winter wheat remote sensing identification analysis method and system based on deep learning | |
CN109829426A (en) | Railway construction temporary building monitoring method and system based on high score remote sensing image | |
CN105205816A (en) | Method for extracting high-resolution SAR image building zone through multi-feature weighted fusion | |
CN114332534B (en) | Hyperspectral image small sample classification method | |
CN107194313A (en) | A kind of parallel intelligent object-oriented classification method | |
US8913829B2 (en) | Automatic processing scale estimation for use in an image process | |
CN104217440A (en) | Method for extracting built-up area from remote sensing image | |
CN111079807B (en) | Ground object classification method and device | |
CN109509188A (en) | A kind of transmission line of electricity typical defect recognition methods based on HOG feature |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170922 |
|
WD01 | Invention patent application deemed withdrawn after publication |