CN105975913A - Road network extraction method based on adaptive cluster learning - Google Patents
Road network extraction method based on adaptive cluster learning Download PDFInfo
- Publication number
- CN105975913A CN105975913A CN201610273017.3A CN201610273017A CN105975913A CN 105975913 A CN105975913 A CN 105975913A CN 201610273017 A CN201610273017 A CN 201610273017A CN 105975913 A CN105975913 A CN 105975913A
- Authority
- CN
- China
- Prior art keywords
- road
- sample
- evidence
- feature
- section
- 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.)
- Granted
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
- G06V20/182—Network patterns, e.g. roads or rivers
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a road network extraction method based on adaptive cluster learning. The method comprises steps of connection network constructing of an extracted road vector; new road detection and extraction. An inference of a road extraction result verifies three aspects of contents. A road connection network constructing process comprehensively considers a geometrical characteristic of the road and a detected road intersection structure constraint so as to guarantee rationality of a road connection result. New-road automatic extraction is a difficult point in a road extraction research field. Under the condition of an existing road-vector semantic mark, problems of high resolution remote sensing image road characteristic heterogeneity and diversity still need to be solved. Fusion of adaptive sample cluster and a multi-classifier road sample classification result is taken as a research thought of the new road extraction. Finally, based on an accuracy requirement of a road extraction result, a D-S evidence theory is introduced to verify a road extraction result, and according to relation between each kind of characteristic and the road, a corresponding verification probability distribution function is defined.
Description
Technical field
The present invention relates to remote sensing image applied technical field, especially relate to a kind of road based on self-adaption cluster study
The method that net extracts.
Background technology
Road network is the netted object being interconnected, and the integrity extracting it and correctness are substantially wanting of road network renewal
Ask.Section under guiding based on navigation data is extracted and crossing is extracted as extracting process for the independence in each section, thus
Cause existing between section fracture.From the point of view of the integrity of road network, existing section is extracted result and is not covered newly added road sections,
Need to detect new added road object according to known roadway characteristic.Extract new added road accurately and need correct and comprehensive sample
Character support.Manual sample labeling workload is big and is difficult to cover all of roadway characteristic, utilizes existing road extraction result energy
Enough provide semantization sample labeling information accurately for new added road.
High score remote sensing image road net extracts task and has its particularity, and extracting method based on sample learning is processing height
Problem of both being primarily present when dividing image data: (1) road shows various in different scenes;On not homology image also
Existing characteristics difference, it is difficult to realize pervasive road extraction task by fixing feature and rule;(2) based on the road having supervision
Road is extracted, and depends on the absolute randomness of sampling, but the randomness realizing sampling is extremely difficult, once exist in sampling process
Any prejudice, extracting result will there are differences.
What new added road extracting method based on sample training classification generally solved is that road network " measures " problem, i.e.
To the location of road object and road " identifies ", i.e. determine whether feature is road object, be only road extraction and faced
Big problem, rational road proof procedure will assist in the correctness improving Road network extraction result.The present invention from road network connect,
New added road is extracted, three aspects of road checking carry out Road network extraction research.
Summary of the invention
A kind of method that the technical scheme is that Road network extraction based on self-adaption cluster study, including following step
Rapid:
Step one, section connects geometric properties.Section extraction process under navigation vector guides is to enter each section respectively
Row extracts, and does not the most consider the annexation between section, therefore extracts result section and in intersection position and is not connected with,
Fracture is there is between end points.Consider from the integrity of road net structure, it is necessary to be attached processing to extracting section.
Step 2, the section under chi structure constraint connects to be revised.The fracture of most sections can be completed based on geometric properties
Connection task.But, complicated road network there is also ambiguous structure and cause the connection of mistake.Therefore, according to geometric properties
After completing section connection, need to utilize known chi structure as constraint, inappropriate section connection result is modified.
Step 3, new added road based on sample learning is extracted.By the networking of known road is connected, obtain vector
The road network labelling to road object corresponding in image.For the newly added road sections beyond existing road network, need according to marked road
Road feature, carries out classification prediction to it and extracts.
Step 4, road based on multi-feature evidence fuzzy reasoning is verified.
Described step one, the geometric properties that section connects includes end-point distances, and linkage section direction is poor with existing direction, section;
According to distance restraint and angle threshold, section is attached.
The detailed process of described step 2 is as follows;
Complete the connection task of most sections fracture based on geometric properties, utilize the section, crossing extracted to connect and carry out
Checking, is modified the section of incorrect link.
The detailed process of described step 3 is as follows;
(1) road sample automatization obtains;
(2) normalization texture sample feature;
First detection gray level co-occurrence matrixes reflects the textural characteristics of different directions, during texture feature extraction,
Select the feature with invariable rotary shape, and for the textural characteristics of orientation-sensitive, travel direction normalized in advance, for
Space scale, selects multiple dimensioned carrying out respectively estimate extraction and analyze;Utilize the detection sample shadow of multi-direction Gabor filtering characteristics
The principal direction of picture;
(3) self adaptation road sample clustering;
Use a set of road sample self-adaption cluster strategy, enable road sample to enter according to feature distribution situation in set
Row restructuring so that respectively organize sample after cluster in feature space in Assembled distribution trend;
(4) non-rice habitats object is slightly rejected.
In described step 3, the detailed process of (1) is as follows;
A () imaged objectization is split;
SLIC is as imaged object dividing method in use, and using segmentation result object as sample characteristics extraction unit;
(b) sample based on known road automatic marking;
Known road is extracted in result and is comprised abundant road semantic information, will be far from the region of road vectors position
As background atural object;Road sample set and background sample set is generated thus according to existing road extraction result.
In described step 3, the detailed process of (3) is as follows;
First, feature is carried out dimensionality reduction: sample characteristics includes the statistical measurement information of spectrum, texture and correspondence, passes through
Feature Dimension Reduction, eliminates unrelated and redundancy sample characteristics;
Then, utilize gauss hybrid models GMM to perform self adaptation road sample clustering: owing to classification number K is unknown,
During real data processes, need by repeatedly testing, the relatively fitting result of multiple compositions carrys out certainly defining K value;In order to it is adaptive
Classification number K, this step two metric of proposition: di and merged index should be obtained in ground;
A () sets initial K value, original sample is performed GMM clustering processing, obtains K Gaussian distribution model;
B () builds K Gaussian distribution model center line set L between any two, and calculate the general of each position in line
Rate valueAs shown in formula (8):
Wherein, j, k ∈ K, pj(x),pkX (), for corresponding Gauss model in the probit of position x, max is for taking maximum letter
Number;
C () definition merged index (Merge Index, MI) defines as shown in formula (9):
If MI > TMI, then it is assumed that line liTwo Gauss models connected have bigger degree of overlapping, need to close
And, classification number that will be total reduces to K-1;
(d) using belong to kth classification sample set as complete or collected works, carry out independent two point GMM clustering processing;Calculate current sample
The di Split Index SI of the corresponding Gauss model of this collection, hasWork as SI > TSITime, it is believed that need working as
Front sample set divides, and cluster classification number that will be total increases to K+1;
E () repeats aforesaid operations, until not meeting division and the Gauss model of the condition of merging, obtain final sample
This cluster numbers K;
Finally, the positive sample set marked by navigation road network according to cluster result is divided into multiple set, and negative sample keeps not
Become;To often organize positive sample and one grader of negative sample combined training, it is achieved the extraction to particular category road;Many group roads carry
The fusion results taking result further will be verified as candidate roads object.
In described step 3, the detailed process of (4) is as follows;
Introduce rectangular degree RrectWith breadth length ratio Rwl, it is defined as follows:
If then there being following formula.
Rrect=mean (A/Amer) (10)
Rwl=W/L (11)
Wherein, A is the area of object, AmerArea for object correspondence minimum enclosed rectangle;It is minimum that W and L is respectively object
The width of boundary rectangle and length;RrectReflect the object filling extent to minimum enclosed rectangle, owing to building is usually
Independent rule objects, the R of its correspondencerectRelatively big, RwlBigger than normal compared with strip section;
Analyzing according to above, the discriminant function that object is rejected in definition is as follows:
Wherein, Trect,TwlFor discrimination threshold, the geometry corresponding when the independent communication object in classification results bianry image is special
Levy and meet threshold value constraint condition, then it is assumed that this object is non-rice habitats object, need to reject;Otherwise it is then road object.
The detailed process of described step 4 is as follows;
This step, using D-S evidence theory as road checking reasoning basis, is different from traditional based on D-S evidence theory
The road geometry of method for extracting roads use and spectral signature, incorporated road contextual feature card in road checking model
According to;
(a) D-S evidence theory basis
As the underlying concept of D-S evidence theory, first the set of all for object to be verified possible outcomes is constituted
Space divides, and is defined as validation framework, is denoted as Θ, and the set of all subsets composition in Θ is denoted as 2Θ, for 2ΘIn
Any hypothesis set A, has m (A) ∈ [0,1], and
Wherein, m is referred to as 2ΘOn probability distribution function (BPAF), m (A) be referred to as A basic probability function;
D-S evidence theory defines belief function Bel and likelihood function Pl and carrys out the uncertainty of problem of representation, it may be assumed that
It is genuine trusting degree, also referred to as lower limit function that belief function Bel (A) represents A;Likelihood function Pl (A) represents
To the trusting degree that A is non-vacation, then [Bel (A), Pl (A)] is a trust interval of A, trusts interval and features letter held to A
Appoint the bound of degree in the presence of having multiple evidence, it is possible to use multiple BPAF are closed by Dempster compositional rule
Become, i.e.
Wherein,For n BPAF;
(b) road checking D-S evidence model;
Owing to the road checking road scene feature having only to according to observing in remote sensing image verifies road identity, root
According to D-S evidence theory, take framework of identification Θ for {, for representing non-rice habitats object, N is road object, then have for Y, N}, YDefinition brief inference function m:P (Y, N}) → [0,1],M (Y, N}+m (Y)+m (N))=1;Its
Middle m (N) represent current signature support road object reliability, m (Y) then represents support non-rice habitats object reliability, and m (Y,
N})=1-m (Y)-m (N) represents the reliability that not can determine that object road identity according to this evidence, i.e. supports the reliability of the unknown;
(c) road multi-feature evidence model
Choose edge feature closely-related with road, spectral signature, context characters of ground object, and these features are carried out
The modelling being suitable for road checking processes;
D () is by analyzing and the definition of corresponding probability distribution function, in navigation data road checking correlated characteristic
Each section is respectively processed, and obtains, according to feature detection result in navigation section, the probability distribution function that feature is corresponding, then,
The BPAF that feature is corresponding is synthesized by the combining evidences rule (formula (16)) utilizing D-S evidence theory, obtains comprehensive multiple features
The probability distribution function of evidence;
According to the D-S evidence theory definition (formula (14)) to belief function Bel, section can be calculated and disappear and exist
Trust probability Bel corresponding under statei(Y),Beli(N);According to maximum of probability distribution principle, definition road checking decision criteria
As follows: for section i, if Beli(Y) > Beli(N), then it is assumed that object is not road;Otherwise, it is believed that existing object is road.
In described step 4, (c) road multi-feature evidence model includes edge evidence model, spectral evidence model, vegetation
Evidence model, shade evidence model, vehicle evidence model, topology evidence model.
The present invention, on the basis of existing section is extracted and intersection extracts result, carries out the connection structure of section vector
Net, the detection of new added road and extraction, the reasoning checking of road extraction result, feature is:
(1) section connects the network forming process synthesis geometric properties that considers road and the intersection structure that detected about
Bundle, thus ensure the reasonability of section connection result.
(2) from the correctness requirement of road extraction result, introduce D-S evidence theory and road extraction result is tested
Card, has considered, in checking model, the evidence including edge, spectrum, context and topological characteristic, and according to each category feature with
The contextual definition of road verifies probability-distribution function accordingly.
(3) Road network extraction method based on self-adaption cluster study is proposed, it is possible to overcome traditional method to be difficult in adapt to
The diversified shortcoming in road, has preferable extraction effect.
(4) from the point of view of correctness, it is less that this research method extracts the non-rice habitats object comprised in result, extracts the whole of result
Weight is higher.
Accompanying drawing explanation
Fig. 1 is the process chart of a kind of Road network extraction method based on self-adaption cluster study.
Fig. 2 is geometric properties section connection diagram.
Fig. 3 is masking-out video generation schematic diagram.
Detailed description of the invention
A kind of method that the technical scheme is that Road network extraction based on self-adaption cluster study, including following step
Rapid:
Step one, section connects geometric properties.Extract the fracture of result section and occur mainly in the road friendship of source navigation road network
Can locate, breaking part section end points is mutually adjacent with section node to be connected.According to general knowledge, same section trend generally in
Gradual change trend, therefore, section needs after connecting to keep direction, section continuous print characteristic.Draw according to above analysis, retrain section
The geometric properties connected specifically includes that end-point distances, and linkage section direction is poor with existing direction, section.
Section based on geometric properties connection diagram is as shown in Figure 2.Tu Zhong branch section E11E12And E21E22For fracture
Section, needs and main road section C1C2It is attached.According to distance restraint, with end points E12Centered by, R is that radius carries out detection and obtains
Candidate connects node C1And C2;By central point E12Node C it is connected with candidate1And C2Connect respectively, and calculate the deflection of linkage section
DegreeWithWhenWith section deflectionDifference less than angle threshold, and at all candidate's linkage sections
During the minimum angle of middle acquirement, then select this node as connecting node and connecting.
Above-mentioned connection procedure and constraint message form are represented, as shown in formula (1):
Wherein,For end points E12,For i-th both candidate nodes,For end points E12And between i-th both candidate nodes
Angle, TθFor angle threshold, mod is modulo operation symbol, with the deflection mathematic interpolation of compatible [-180,180] angular interval;
N is that the candidate meeting present node distance restraint connects number of nodes.
Step 2, the section under chi structure constraint connects to be revised.The fracture of most sections can be completed based on geometric properties
Connection task.But, complicated road network there is also ambiguous structure and cause the connection of mistake.Therefore, according to geometric properties
After completing section connection, need to utilize known chi structure as constraint, inappropriate section connection result be modified,
It is allowed to more tally with the actual situation, improves accuracy.
Step 3, new added road based on sample learning is extracted.By the networking of known road is connected, obtain vector
The road network labelling to road object corresponding in image.For the newly added road sections beyond existing road network, need according to marked road
Road feature, carries out classification prediction to it and extracts.Specifically comprise the following steps that
(1) road sample automatization obtains.
A () imaged objectization is split.Use SLIC as imaged object dividing method, and using segmentation result object as
Sample characteristics extraction unit.
(b) sample based on known road automatic marking.Idiographic flow is as follows:
Rasterizing road vectors data, with radius rroadGenerating structure element SroadAnd perform morphological dilations computing, generate
Road sample masking-out image Xroad, such as formula (2);With radius rgap,rgap>rroadGenerating structure element Sgap, and it is swollen to perform morphology
Swollen computing, generates masking-out image Xgap, the generation of auxiliary background sample masking-out sample, such as formula (3);With radius rbg,rbg>rgapRaw
Become structural element Sbg, and perform morphological dilations computing, by operation result and XgapImage step-by-step after logical inversion performs logic
With computing, obtain background sample masking-out image such as formula (4);
Wherein, X is pending image,For morphological dilations operative symbol.Structural element radius rroad,rgap,rbgNeed
The width information corresponding according to navigation road network line of vector and image resolution are arranged, masking-out video generation schematic diagram such as accompanying drawing 3 institute
Show.
Respectively road sample masking-out image and background sample masking-out image are superposed with road objectification segmentation result, statistics
The pixel quantity n of road masking-out is belonged to inside each cutting objectroadWith quantity n belonging to background masking-out pixelbg, according to formula
(5) judge that object belongs to road object or background object.
Wherein, ObjiFor i-th sample object, 1 mark object is road sample, and-1 mark object is background sample, 0 mark
Know non-sample object;N is object sum, TareaThreshold value is compared for effective area.
(2) normalization texture sample feature.During texture feature extraction, selection has rotation not the most as far as possible
The feature of deformation, it addition, for the textural characteristics of orientation-sensitive, can travel direction normalized in advance.For space chi
Degree, then need to select multiple dimensioned carrying out respectively estimate extraction and analyze.
Gray level co-occurrence matrixes (Gray-Level Co-occurrence Matrix, GLCM) be defined as from gray level be i
Point leave certain fixed position relation d=(Dx,Dy) reach the probability that gray scale is j.From the definition of GLCM, according to input
Position relationship parameter d=(Dx,Dy) difference that reaches, testing result has reflected the textural characteristics of different directions.Real world is anti-
Reflect the grain direction in image different, it is therefore necessary to before generating GLCM, determine the principal direction of current sample image
θ, then determines position relationship parameter according to principal direction, it is ensured that the direction normalization of skin texture detection result, thus different sample it
Between textural characteristics there is comparability.Determine shown in position relationship parameter such as formula (6) according to image principal direction:
Wherein, D is pixel-shift distance.
Gabor filtering transformation is a kind of windowed FFT, and it can extract on frequency domain different scale, different directions
The feature of the correction of image.The present invention utilizes the principal direction of the detection sample image of multi-direction Gabor filtering characteristics.Gabor characteristic
Extraction process: fixed filters parameter, obtains a series of filter template according to different θ;Based on each filter template to sample
This image performs convolution algorithm, obtains multiple filter result image;Add up the gray value sum of each filter result image, by gray scale
What accumulated value was maximum filters directioin parameter corresponding to the image principal direction as current image block.
Wherein, i is the sequence number of present filter, and n is the quantity of wave filter, giAfter filtering based on i-th wave filter
Image, θiFor the angle that wave filter i is corresponding, θIPrincipal direction for image.
(3) self adaptation road sample clustering.The present invention designs a set of road sample self-adaption cluster strategy so that road sample
Originally can recombinate according to feature distribution situation in set so that respectively organize sample after cluster in feature space in Assembled distribution
Trend.
Firstly, it is necessary to feature is carried out dimensionality reduction.The sample characteristics that this research is extracted includes the system of spectrum, texture and correspondence
Measurement degree information, it is contemplated that the wave band number of image, the yardstick etc. of textural characteristics, a final sample characteristic vector necessarily higher-dimension
Characteristic vector.But, in the case of sample number is relatively fewer, high dimensional feature makes sample gradation statistically be subject to
To destroying, it is therefore desirable to by Feature Dimension Reduction, eliminate unrelated and redundancy sample characteristics.The present invention utilizes the vector of proposition similar
Sex index, the feature selection approach proposed according to Huang Xin carries out dimension-reduction treatment.
Then, gauss hybrid models (GMM) is utilized to perform self adaptation road sample clustering.Owing to classification number K is unknown,
In real data processes, need by repeatedly testing, the relatively fitting result of multiple compositions carrys out certainly defining K value.In order to from
Adaptively obtain classification number K, this section two metric of proposition: di and merged index.
A () sets initial K value, original sample is performed GMM clustering processing, obtains K Gaussian distribution model;
B () builds K Gaussian distribution model center line set L between any two, and calculate the general of each position in line
Rate valueAs shown in formula (8):
Wherein, j, k ∈ K, pj(x),pkX (), for corresponding Gauss model in the probit of position x, max is for taking maximum letter
Number.
C () definition merged index (Merge Index, MI) defines as shown in formula (9):
Wherein, MIiFor line liThe merged index of two Gauss models connected, TMIFor merging threshold value, if MIi>TMI,
Then think line liTwo Gauss models connected have bigger degree of overlapping, need to merge, classification number that will be total
Reduce to K-1.
(d) using belong to kth classification sample set as complete or collected works, carry out independent two point GMM clustering processing;Calculate current sample
The di (Split Index, SI) of the corresponding Gauss model of this collection, has SIk=1-MIk, work as SIk>TsITime, it is believed that it is right to need
Current sample set divides, and cluster classification number that will be total increases to K+1.Wherein SIkFor the di of kth sample,
MIkFor the merged index of kth sample, TSIFor division threshold value.
E () repeats aforesaid operations, until not meeting division and the Gauss model of the condition of merging, obtain final sample
This cluster numbers K.
Finally, the positive sample set marked by navigation road network according to cluster result is divided into multiple set, and negative sample keeps not
Become.To often organize positive sample and one grader of negative sample combined training, it is achieved the extraction to particular category road.Many group roads carry
The fusion results taking result further will be verified as candidate roads object.
(4) non-rice habitats object is slightly rejected.Admittedly, image exists the building pair similar with roadway characteristic
As, the SLIC segmentation figure speckle of its correspondence and mileage chart speckle have very much like shape, spectrum, textural characteristics, above-mentioned sorted
Cheng Wufa avoids by mistake carrying of this type of non-rice habitats object, needs that classification results carries out corresponding non-rice habitats object and rejects operation.
Due to the connection characteristic of road object, it presents special geometric shape, and is difficult to that fix, quantitative
Geometric properties is described, and reviews and is mixed into the non-rice habitats object of road extraction result and is then generally of quantifiable geometric properties,
Introduce rectangular degree RrectWith breadth length ratio Rwl, it is defined as follows:
If then there being following formula.
Rrect=mean (A/Amer) (10)
Rwl=W/L (11)
Wherein, A is the area of object, AmerArea for object correspondence minimum enclosed rectangle;It is minimum that W and L is respectively object
The width of boundary rectangle and length.RrectReflect the object filling extent to minimum enclosed rectangle, owing to building is usually
Independent rule objects, the R of its correspondencerectRelatively big, RwlBigger than normal compared with strip section.
Analyzing according to above, the discriminant function that object is rejected in definition is as follows:
Wherein, Trect,TwlFor discrimination threshold, the geometry corresponding when the independent communication object in classification results bianry image is special
Levy and meet threshold value constraint condition, then it is assumed that this object is non-rice habitats object, need to reject;Otherwise it is then road object;
Step 4, road based on multi-feature evidence fuzzy reasoning is verified.Road extraction process completes road in image
The location of road feature object, admittedly, is commonly present some objects close with roadway characteristic, causes mistake in image
Extract result.It is thus desirable to introduce road proof procedure to reject non-rice habitats object.The present invention is using D-S evidence theory as road
Checking reasoning basis, is different from road geometry and Spectral Properties that traditional method for extracting roads based on D-S evidence theory uses
Levying, the present invention has innovatively incorporated road contextual feature evidence in road checking model.
(a) D-S evidence theory basis.
As the underlying concept of D-S evidence theory, first the set of all for object to be verified possible outcomes is constituted
Space divides, and is defined as validation framework, is denoted as Θ, and the set of all subsets composition in Θ is denoted as 2Θ, for 2ΘIn
Any hypothesis set A, has m (A) ∈ [0,1], and
Wherein, m is referred to as 2ΘOn probability distribution function (BPAF), m (A) be referred to as A basic probability function.
D-S evidence theory defines belief function Bel and likelihood function Pl and carrys out the uncertainty of problem of representation, it may be assumed that
It is genuine trusting degree, also referred to as lower limit function that belief function Bel (A) represents A;Likelihood function Pl (A) represents
To the trusting degree that A is non-vacation, then [Bel (A), Pl (A)] is a trust interval of A, trusts interval and features letter held to A
Appoint the bound of degree in the presence of having multiple evidence, it is possible to use multiple BPAF are closed by Dempster compositional rule
Become, i.e.
Wherein,For n BPAF.
(b) road checking D-S evidence model.Owing to road checking has only to according to the road field observed in remote sensing image
Scape feature verifies road identity, according to D-S evidence theory, takes framework of identification Θ for { Y, N}, Y are for representing non-rice habitats object, N
For road object, then haveDefinition brief inference function m:P (Y, N}) → [0,1],m({Y,N}+m
(Y)+m (N))=1.Wherein m (N) represents that current signature supports the reliability of road object, and m (Y) then represents support non-rice habitats object
Reliability, and m ({ Y, N})=1-m (Y)-m (N) represents the reliability that not can determine that object road identity according to this evidence, i.e. supports
Unknown reliability.
(c) road multi-feature evidence model.The evidence of road checking, is i.e. remote sensing roadway characteristic, is the base of road checking
Plinth.Feature is generally of compatibility and uncertainty, and the most a certain feature can be under the jurisdiction of multiple target and these features can be spread out
Bear various ways.For ensureing the reliability of road banknote validation result, it is special that the present invention chooses edge closely-related with road
Levy, spectral signature, context characters of ground object, and these features are carried out be suitable for road checking modelling process.
(I) edge evidence model.
In the case of unobstructed, road is generally of significant dual edge feature, thus, dual edge feature can be as road
The strong evidence of road checking.But, road usually can be blocked by shade, shade tree, and quality of image difference will also result on image
The disappearance of road both sides of the edge feature, thus, the disappearance of ancipital feature not can prove that road disappears, but can be as road
The evidence that road exists.
If a length of L in certain section in navigation data, the road section length that there is dual edge feature is Ledge, then R is utilizededge
Express the edge feature of current road segment, as shown in (18);
Redge=Ledge/L (18)
If TedgeFor efficient frontier characteristic threshold value, if Redge≤Tedge, then the probability of road existence and edge feature value are in just
Linear correlation, corresponding probability distribution function is as follows:
Work as Redge> Tedge, then it is assumed that road exists, the probability that needs imparting is bigger:
Wherein, m1(Y) edge feature support probability to non-rice habitats object is represented;m1(N) represent that edge feature is to road pair
The support probability of elephant;m1(Y, N) represents that edge feature cannot determine whether the probability for road;α1For corresponding general of edge feature
Rate weight;ε is predefined small probability value.
(II) spectral evidence model.
Preferably the change of road surface spectral signature is less, and there is bigger gray scale difference with neighbouring non-rice habitats background
Different.But, road is blocked by other background atural objects or oneself factor causes the situation that SPECTRAL DIVERSITY in section is bigger generally to exist,
Thus the evidence that spectral signature cannot disappear as road.But, when in navigation data corresponding road section, spectral signature difference is less
Time, the strong evidence that spectral signature can exist as road.
Arrange navigation relief area, section in the range of pixel be std (DN in the gray value standard difference of each wave bandseg), Tspectra
For effective spectral signature threshold value, then spectral signature RspectraAs shown in formula (21):
Rspectra=std (DNseg)/Tspectra (21)
If Rspectra≤ 1, then it is assumed that the probability of road existence and RspectraIn negative linear correlation, corresponding probability assignments letter
Number is as follows:
If Rspectra> 1, then it is assumed that road surface SPECTRAL DIVERSITY is bigger, it is impossible to prove whether road exists, corresponding probability
Partition function is as follows:
Wherein, m2(Y) spectral signature support probability to non-rice habitats object is represented;m2(N) represent that spectral signature is to road pair
The support probability of elephant;m2(Y, N) represents that spectral signature cannot determine whether the probability for road;α2For corresponding general of edge feature
Rate weight.
(III) vegetation evidence model.
Road vegetation is atural object common in road scene, and it is made up of shade tree, greenbelt and greenery patches in reality;From
From the point of view of geometric shape, road vegetation is generally close to road, and wherein Road greenbelt is generally along road continuous distribution, and has and road
Consistent direction character is moved towards on road, and greenery patches is then surrounded and the geometric shape of formation rule by road;On image, road vegetation is then
Reflect the special curve of spectrum.By analyzing discovery above, in road scene, the direction character of vegetation object can be as inspection
The evidence of road inspection road presence or absence.The present invention is by the long limit of the major axes orientation of independent vegetation object, i.e. object minimum enclosed rectangle
Direction is as the direction of vegetation object.If DplantFor orientation angle set corresponding to vegetation object in navigation relief area, section,
DroadFor Present navigation section principal direction corresponding angle, TplantFor orientation angle difference limen value, RplantSpecial for the vegetation after normalization
Levy.
Rplant=std (Dplant-Droad)/Tplant (24)
Work as RplantWhen≤1, the probability of road existence and RplantIn negative linear correlation, accordingly, the probability of non-rice habitats object
With RplantRelevant in linear positive, vegetation object orientation angle is the biggest with road principal direction angle difference, then non-rice habitats object is general
Rate is the biggest, and the probability of road object is the least.Corresponding probability distribution function is as follows:
Otherwise, work as RplantDuring > 1, it is believed that vegetation object orientation feature is unrelated with trend of road, it it is now non-rice habitats object
Probability relatively big, corresponding probability distribution function is as follows:
If navigation does not has vegetation object in relief area, section, then vegetation characteristics cannot be the most right as the experimental evidence of road
The probability distribution function answered is as follows:
Wherein, m3(Y) vegetation characteristics support probability to non-rice habitats object is represented;m3(N) represent that vegetation characteristics is to road pair
The support probability of elephant;m3(Y, N) represents that vegetation characteristics cannot determine the probability whether road exists;α3Corresponding for vegetation characteristics
Probability right.
(IV) shade evidence model.
Being affected by sun altitude and road surrounding building height, road surface can be projected the shade got off and be hidden
Lid.Road is usually relatively flat open area, thus is incident upon shape facility and the building phase of the shade of road surface
Close, and the most regular;And in the region of atural object scene complexity, shade is the most broken.Therefore, the shape facility of shade can
Using as road experimental evidence.If A is the area of object, AmerFor the area of object correspondence minimum enclosed rectangle, RrectFor shade
The even rectangular degree of object, then have following formula.
Rrect=mean (A/Amer) (28)
It reflects the object filling extent to minimum enclosed rectangle, object shapes the most regular (as rectangle), corresponding
RrectThe highest, corresponding road object probability is the biggest, and corresponding probability distribution function is as follows:
If not having shadow object in navigation relief area, section, this feature cannot be then corresponding as the experimental evidence of road
Probability distribution function is as follows:
Wherein, m4(Y) shadow character support probability to non-rice habitats object is represented;m4(N) represent that shadow character is to road pair
The support probability of elephant;m4(Y, N) represents that shadow character cannot determine the probability whether road exists;α4Corresponding for shadow character
Probability right.
(V) vehicle evidence model.
Vehicle is the evidence of road object, and in remote sensing image, the position that there is vehicle is usually road or dew
It parking lot.In parking lot, vehicle is intensive neatly to be parked;The vehicle on road surface is the most aobvious mixed and disorderly, but global alignment direction and road
Road trend is consistent, and there is the situation that vehicle is assembled near intersection.Application vehicle detection result is as road object
Evidence, if there is vehicle in navigation relief area, section, then corresponding probability distribution function is as follows:
Otherwise, if not having Vehicle Object in navigation relief area, section, then vehicle characteristics cannot provide evidence to road checking,
Thus, corresponding probability distribution function is as follows:
Wherein, m5(Y) vehicle characteristics support probability to non-rice habitats object is represented;m5(N) represent that vehicle characteristics is to road pair
The support probability of elephant;m5(Y, N) represents that vehicle characteristics cannot determine the probability whether road exists;α5Corresponding for vehicle characteristics
Probability right.
(VI) topology evidence model.
Road function in reality determines its topological characteristic being interconnected, and new added road is typically existing road
Extension extend, or be interconnected composition intersection with existing road.Therefore, new added road extracts result and existing road
Position relationship between the vector of road can be as the evidence of road authentication.
If LnewLine object, L is extracted for new added roadexistFor existing road object set, then there are new added road and existing road
The location point common factor P on roadoverlap, definition is as shown in formula (32):
Poverlap=Lnew∩Lexist (33)
If PoverlapFor nonempty set, then it is assumed that new added road intersects with known road, now think present road section
The probability being road is bigger.
Certainly, undeniable, during classification of road extracts, extraction result that real roads is corresponding is possibly also owing to ground
Thing blocks and interrupts, thus causes the road object isolated to exist, therefore, and the proximity relations between newly added road sections and known road
Also can be as a part for road topology evidence, if the distance between newly added road sections and known road is d (Lnew,Lexist), definition
As follows:
d(Lnew,Lexist)=min [dist (pti,Lexist)]pti∈LnewI=1,2 ..., n (34)
Wherein, ptiFor newly added road sections LnewIn i-th node, n is section LnewIn number of nodes, dist is node
Distance function to line.Then corresponding probability distribution function is as follows:
Wherein, m6(Y) the section topological evidence with the known road support probability to non-rice habitats object is represented;m6(N) represent
The topology evidence support probability to road object;m6(Y, N) represents that current evidence cannot determine that whether object is the probability of road;
α6For the probability right that topology evidence is corresponding.
D () is by analyzing and the definition of corresponding probability distribution function, to the number that navigates road checking correlated characteristic above
According to, each section is respectively processed, and obtains, according to feature detection result in navigation section, the probability distribution function that feature is corresponding,
Then, utilize the combining evidences rule (formula (16)) of D-S evidence theory that the BPAF that feature is corresponding is synthesized, obtain the most
The probability distribution function of feature evidence.
According to the D-S evidence theory definition (formula (14)) to belief function Bel, section can be calculated and disappear and exist
Trust probability Bel corresponding under statei(Y),Beli(N).According to maximum of probability distribution principle, definition road checking decision criteria
As follows: for section i, if Beli(Y) > Beli(N), then it is assumed that object is not road;Otherwise, it is believed that existing object is road.
Claims (9)
1. the method for a Road network extraction based on self-adaption cluster study, it is characterised in that: comprise the following steps;
Step one, the geometric properties connected based on section, it is attached processing to extracting section;
Step 2, the section under chi structure constraint connects to be revised;
After completing section connection according to geometric properties, utilize known chi structure as constraint, inappropriate section is connected
Result is modified;
Step 3, new added road based on sample learning is extracted;
By the networking of known road is connected, obtain the vector road network labelling to road object corresponding in image, for existing
There is the newly added road sections beyond road network, need, according to marked roadway characteristic, it is carried out classification prediction and extracts;
Step 4, road based on multi-feature evidence fuzzy reasoning is verified.
The method of a kind of Road network extraction based on self-adaption cluster study the most according to claim 1, it is characterised in that:
Described step one, the geometric properties that section connects includes end-point distances, and linkage section direction is poor with existing direction, section;According to distance
Section is attached by constraint and angle threshold.
The method of a kind of Road network extraction based on self-adaption cluster study the most according to claim 2, it is characterised in that:
The detailed process of described step 2 is as follows;
Complete the connection task of most sections fracture based on geometric properties, utilize the section, crossing extracted to connect and test
Card, is modified the section of incorrect link.
The method of a kind of Road network extraction based on self-adaption cluster study the most according to claim 3, it is characterised in that:
The detailed process of described step 3 is as follows;
(1) road sample automatization obtains;
(2) normalization texture sample feature;
First detection gray level co-occurrence matrixes reflects the textural characteristics of different directions, during texture feature extraction, selects
There is the feature of invariable rotary shape, and for the textural characteristics of orientation-sensitive, travel direction normalized in advance, for space
Yardstick, selects multiple dimensioned carrying out respectively estimate extraction and analyze;Utilize the detection sample image of multi-direction Gabor filtering characteristics
Principal direction;
(3) self adaptation road sample clustering;
Use a set of road sample self-adaption cluster strategy, enable road sample to carry out weight according to feature distribution situation in set
Group so that respectively organize sample after cluster in feature space in Assembled distribution trend;
(4) non-rice habitats object is slightly rejected.
The method of a kind of Road network extraction based on self-adaption cluster study the most according to claim 4, it is characterised in that:
In described step 3, the detailed process of (1) is as follows;
A () imaged objectization is split;
SLIC is as imaged object dividing method in use, and using segmentation result object as sample characteristics extraction unit;
(b) sample based on known road automatic marking;
Known road is extracted in result and is comprised abundant road semantic information, will be far from the region conduct of road vectors position
Background atural object;Road sample set and background sample set is generated thus according to existing road extraction result.
The method of a kind of Road network extraction based on self-adaption cluster study the most according to claim 5, it is characterised in that:
In described step 3, the detailed process of (3) is as follows;
First, feature is carried out dimensionality reduction: sample characteristics includes the statistical measurement information of spectrum, texture and correspondence, passes through feature
Dimensionality reduction, eliminates unrelated and redundancy sample characteristics;
Then, gauss hybrid models GMM is utilized to perform self adaptation road sample clustering: owing to classification number K is unknown, in reality
Data processing, needs by repeatedly testing, the relatively fitting result of multiple compositions carrys out certainly defining K value;In order to adaptively
Obtain classification number K, this step two metric of proposition: di and merged index;
A () sets initial K value, original sample is performed GMM clustering processing, obtains K Gaussian distribution model;
B () builds K Gaussian distribution model center line set L between any two, and calculate the probit of each position in lineAs shown in formula (8):
Wherein, j, k ∈ K, pj(x),pkX (), for corresponding Gauss model in the probit of position x, max is for taking max-value function;
C () definition merged index (Merge Index, MI) defines as shown in formula (9):
Wherein, MIiFor line liThe merged index of two Gauss models connected, TMIFor merging threshold value, if MI > TMI, then recognize
For line liTwo Gauss models connected have bigger degree of overlapping, need to merge, and classification number that will be total is reduced to
K-1;
(d) using belong to kth classification sample set as complete or collected works, carry out independent two point GMM clustering processing;Calculate current sample set
The di Split Index SI of corresponding Gauss model, hasWork as SI > TSITime, it is believed that need current sample
This collection divides, and cluster classification number that will be total increases to K+1;
E () repeats aforesaid operations, until not meeting division and the Gauss model of the condition of merging, obtaining final sample and gathering
Class number K;
Finally, the positive sample set marked by navigation road network according to cluster result is divided into multiple set, and negative sample keeps constant;Will
Often organize positive sample and one grader of negative sample combined training, it is achieved the extraction to particular category road;Many group road extraction knots
The fusion results of fruit further will be verified as candidate roads object.
The method of a kind of Road network extraction based on self-adaption cluster study the most according to claim 6, it is characterised in that:
In described step 3, the detailed process of (4) is as follows;
Introduce rectangular degree RrectWith breadth length ratio Rwl, it is defined as follows:
If then there being following formula.
Rrect=mean (A/Amer) (10)
Rwl=W/L (11)
Wherein, A is the area of object, AmerArea for object correspondence minimum enclosed rectangle;It is external that W and L is respectively object minimum
The width of rectangle and length;RrectReflect the object filling extent to minimum enclosed rectangle, owing to building is the most independent
Rule objects, the R of its correspondencerectRelatively big, RwlBigger than normal compared with strip section;
Analyzing according to above, the discriminant function that object is rejected in definition is as follows:
Wherein, Trect,TwlFor discrimination threshold, the geometric properties corresponding when the independent communication object in classification results bianry image accords with
Close threshold value constraint condition, then it is assumed that this object is non-rice habitats object, need to reject;Otherwise it is then road object.
The method of a kind of Road network extraction based on self-adaption cluster study the most according to claim 7, it is characterised in that:
The detailed process of described step 4 is as follows;
This step, using D-S evidence theory as road checking reasoning basis, is different from traditional road based on D-S evidence theory
The road geometry of extracting method use and spectral signature, incorporated road contextual feature evidence in road checking model;
(a) D-S evidence theory basis
As the underlying concept of D-S evidence theory, first by the set institute structure of all for object to be verified possible outcomes
The space become divides, and is defined as validation framework, is denoted as Θ, and the set of all subsets composition in Θ is denoted as 2Θ,
For 2ΘIn any assume set A, have m (A) ∈ [0,1], and
Wherein, m is referred to as 2ΘOn probability distribution function (BPAF), m (A) be referred to as A basic probability function;
D-S evidence theory defines belief function Bel and likelihood function Pl and carrys out the uncertainty of problem of representation, it may be assumed that
It is genuine trusting degree, also referred to as lower limit function that belief function Bel (A) represents A;Likelihood function Pl (A) represents
The trusting degree of non-vacation, then [Bel (A), Pl (A)] is a trust interval of A, trusts interval and features degree of belief held to A
Bound in the presence of having multiple evidence, it is possible to use multiple BPAF are synthesized by Dempster compositional rule,
I.e.
Wherein,m1,m2,...,mnFor n BPAF;
(b) road checking D-S evidence model;
Owing to the road checking road scene feature having only to according to observing in remote sensing image verifies road identity, according to D-
S evidence theory, takes framework of identification Θ for {, for representing non-rice habitats object, N is road object, then have for Y, N}, YFixed
Justice brief inference function m:P (Y, N}) → [0,1],M (Y, N}+m (Y)+m (N))=1;Wherein m (N) expression is worked as
Front feature supports the reliability of road object, and m (Y) then represents the reliability of support non-rice habitats object, and m ({ Y, N})=1-m (Y)-m
(N) expression not can determine that the reliability of object road identity according to this evidence, i.e. supports the reliability of the unknown;
(c) road multi-feature evidence model
Choose edge feature closely-related with road, spectral signature, context characters of ground object, and these features are suitable for
The modelling of road checking processes;
D () is by analyzing and the definition of corresponding probability distribution function, to navigation data Zhong Ge road road checking correlated characteristic
Section is respectively processed, and obtains, according to feature detection result in navigation section, the probability distribution function that feature is corresponding, then, utilizes
The BPAF that feature is corresponding is synthesized by the combining evidences rule (formula (16)) of D-S evidence theory, obtains comprehensive multi-feature evidence
Probability distribution function;
According to the D-S evidence theory definition (formula (14)) to belief function Bel, section can be calculated and disappear and existence
Trust probability Bel of lower correspondencei(Y),Beli(N);According to maximum of probability distribution principle, definition road checking decision criteria is as follows:
For section i, if Beli(Y) > Beli(N), then it is assumed that object is not road;Otherwise, it is believed that existing object is road.
The method of a kind of Road network extraction based on self-adaption cluster study the most according to claim 8, it is characterised in that:
In described step 4, (c) road multi-feature evidence model include edge evidence model, spectral evidence model, vegetation evidence model,
Shade evidence model, vehicle evidence model, topology evidence model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610273017.3A CN105975913B (en) | 2016-04-28 | 2016-04-28 | Road network extraction method based on adaptive cluster learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610273017.3A CN105975913B (en) | 2016-04-28 | 2016-04-28 | Road network extraction method based on adaptive cluster learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN105975913A true CN105975913A (en) | 2016-09-28 |
CN105975913B CN105975913B (en) | 2020-03-10 |
Family
ID=56993291
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610273017.3A Active CN105975913B (en) | 2016-04-28 | 2016-04-28 | Road network extraction method based on adaptive cluster learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN105975913B (en) |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106446960A (en) * | 2016-10-10 | 2017-02-22 | 太原理工大学 | Dynamic hot degree road network constructing method with density-based clustering time quantum dividing function |
CN106845396A (en) * | 2017-01-18 | 2017-06-13 | 南京理工大学 | Illegal fishing Activity recognition method based on automated graphics identification |
CN107085074A (en) * | 2017-04-19 | 2017-08-22 | 中国科学技术大学 | A kind of method for monitoring motor-vehicle tail-gas of classifying |
CN107203783A (en) * | 2017-05-24 | 2017-09-26 | 西北工业大学 | Remote sensing images Pixel-level sorting technique based on adaptive convolution kernel convolutional neural networks |
CN107273914A (en) * | 2017-05-17 | 2017-10-20 | 西北工业大学 | Efficient fusion identification method based on the adaptive dynamic select of information source |
CN107480679A (en) * | 2017-04-20 | 2017-12-15 | 北京工业大学 | The road network extracting method of classification and connected component analysis based on convolutional neural networks |
CN108229425A (en) * | 2018-01-29 | 2018-06-29 | 浙江大学 | A kind of identifying water boy method based on high-resolution remote sensing image |
CN108256424A (en) * | 2017-12-11 | 2018-07-06 | 中交信息技术国家工程实验室有限公司 | A kind of high-resolution remote sensing image method for extracting roads based on deep learning |
CN110070109A (en) * | 2019-03-28 | 2019-07-30 | 东南大学 | A method of suitable for building the atural object scattered points automatic cluster in dense arrangement area |
CN110443816A (en) * | 2019-08-08 | 2019-11-12 | 重庆市地理信息和遥感应用中心(重庆市测绘产品质量检验测试中心) | Urban road extracting method on remote sensing image based on intersection detection |
CN110597930A (en) * | 2018-06-11 | 2019-12-20 | 百度在线网络技术(北京)有限公司 | Method, device, equipment and medium for detecting building block gland road in electronic map |
CN110716999A (en) * | 2019-09-05 | 2020-01-21 | 武汉大学 | POI positioning method based on position description containing qualitative position and quantitative distance |
CN111145291A (en) * | 2018-11-05 | 2020-05-12 | 罗伯特·博世有限公司 | System and method for automatic semantic map generation |
CN111191597A (en) * | 2019-12-31 | 2020-05-22 | 武汉中海庭数据技术有限公司 | System and method for extracting road structure based on vector line |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101218486A (en) * | 2005-07-22 | 2008-07-09 | 特拉戈公司 | Method, device and system for modeling a road network graph |
CN104361142A (en) * | 2014-12-12 | 2015-02-18 | 华北水利水电大学 | Detection method for rapid change in multi-source navigation electronic map vector road network |
CN104504718A (en) * | 2015-01-06 | 2015-04-08 | 南京大学 | High-definition aerial remote sensing data automatic road extraction method |
CN104866670A (en) * | 2015-05-25 | 2015-08-26 | 武汉大学 | GPS spatial-temporal trajectory-based road network topological change automatic detection method and system |
EP2958054A2 (en) * | 2014-06-18 | 2015-12-23 | Mobileye Vision Technologies Ltd. | Hazard detection in a scene with moving shadows |
-
2016
- 2016-04-28 CN CN201610273017.3A patent/CN105975913B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101218486A (en) * | 2005-07-22 | 2008-07-09 | 特拉戈公司 | Method, device and system for modeling a road network graph |
EP2958054A2 (en) * | 2014-06-18 | 2015-12-23 | Mobileye Vision Technologies Ltd. | Hazard detection in a scene with moving shadows |
CN104361142A (en) * | 2014-12-12 | 2015-02-18 | 华北水利水电大学 | Detection method for rapid change in multi-source navigation electronic map vector road network |
CN104504718A (en) * | 2015-01-06 | 2015-04-08 | 南京大学 | High-definition aerial remote sensing data automatic road extraction method |
CN104866670A (en) * | 2015-05-25 | 2015-08-26 | 武汉大学 | GPS spatial-temporal trajectory-based road network topological change automatic detection method and system |
Non-Patent Citations (1)
Title |
---|
吴晓燕, 车登科, 戴芬: "影像与矢量结合的道路自动提取及变化检测", 《矿山测量》 * |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106446960B (en) * | 2016-10-10 | 2019-05-21 | 太原理工大学 | A kind of Density Clustering divides the dynamic temperature road network construction method of period |
CN106446960A (en) * | 2016-10-10 | 2017-02-22 | 太原理工大学 | Dynamic hot degree road network constructing method with density-based clustering time quantum dividing function |
CN106845396A (en) * | 2017-01-18 | 2017-06-13 | 南京理工大学 | Illegal fishing Activity recognition method based on automated graphics identification |
CN106845396B (en) * | 2017-01-18 | 2020-05-22 | 南京理工大学 | Illegal fishing behavior identification method based on automatic image identification |
CN107085074A (en) * | 2017-04-19 | 2017-08-22 | 中国科学技术大学 | A kind of method for monitoring motor-vehicle tail-gas of classifying |
CN107085074B (en) * | 2017-04-19 | 2019-07-23 | 中国科学技术大学 | A method of classification monitoring motor-vehicle tail-gas |
CN107480679B (en) * | 2017-04-20 | 2021-08-17 | 北京工业大学 | Road network extraction method based on classification and connected region analysis of convolutional neural network |
CN107480679A (en) * | 2017-04-20 | 2017-12-15 | 北京工业大学 | The road network extracting method of classification and connected component analysis based on convolutional neural networks |
CN107273914A (en) * | 2017-05-17 | 2017-10-20 | 西北工业大学 | Efficient fusion identification method based on the adaptive dynamic select of information source |
CN107273914B (en) * | 2017-05-17 | 2020-06-30 | 西北工业大学 | Efficient fusion identification method based on information source self-adaptive dynamic selection |
CN107203783A (en) * | 2017-05-24 | 2017-09-26 | 西北工业大学 | Remote sensing images Pixel-level sorting technique based on adaptive convolution kernel convolutional neural networks |
CN108256424A (en) * | 2017-12-11 | 2018-07-06 | 中交信息技术国家工程实验室有限公司 | A kind of high-resolution remote sensing image method for extracting roads based on deep learning |
CN108229425A (en) * | 2018-01-29 | 2018-06-29 | 浙江大学 | A kind of identifying water boy method based on high-resolution remote sensing image |
CN110597930A (en) * | 2018-06-11 | 2019-12-20 | 百度在线网络技术(北京)有限公司 | Method, device, equipment and medium for detecting building block gland road in electronic map |
CN111145291A (en) * | 2018-11-05 | 2020-05-12 | 罗伯特·博世有限公司 | System and method for automatic semantic map generation |
CN110070109A (en) * | 2019-03-28 | 2019-07-30 | 东南大学 | A method of suitable for building the atural object scattered points automatic cluster in dense arrangement area |
CN110443816A (en) * | 2019-08-08 | 2019-11-12 | 重庆市地理信息和遥感应用中心(重庆市测绘产品质量检验测试中心) | Urban road extracting method on remote sensing image based on intersection detection |
CN110716999A (en) * | 2019-09-05 | 2020-01-21 | 武汉大学 | POI positioning method based on position description containing qualitative position and quantitative distance |
CN111191597A (en) * | 2019-12-31 | 2020-05-22 | 武汉中海庭数据技术有限公司 | System and method for extracting road structure based on vector line |
Also Published As
Publication number | Publication date |
---|---|
CN105975913B (en) | 2020-03-10 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN105975913A (en) | Road network extraction method based on adaptive cluster learning | |
CN106778605A (en) | Remote sensing image road net extraction method under navigation data auxiliary | |
CN105261109B (en) | A kind of bill crown word number recognition methods | |
CN103049763B (en) | Context-constraint-based target identification method | |
CN104834922B (en) | Gesture identification method based on hybrid neural networks | |
CN101739555B (en) | Method and system for detecting false face, and method and system for training false face model | |
CN101925905B (en) | Identification and verification of unknown document according to eigen image process | |
CN106529508A (en) | Local and non-local multi-feature semantics-based hyperspectral image classification method | |
CN109447994A (en) | In conjunction with the remote sensing image segmentation method of complete residual error and Fusion Features | |
CN101329736B (en) | Method of image segmentation based on character selection and hidden Markov model | |
CN106778835A (en) | The airport target by using remote sensing image recognition methods of fusion scene information and depth characteristic | |
CN110503054B (en) | Text image processing method and device | |
CN106127204A (en) | A kind of multi-direction meter reading Region detection algorithms of full convolutional neural networks | |
CN103400151A (en) | Optical remote-sensing image, GIS automatic registration and water body extraction integrated method | |
CN105138968A (en) | Face authentication method and device | |
Piekarczyk | Hierarchical random graph model for off-line handwritten signatures recognition | |
CN102629321A (en) | Facial expression recognition method based on evidence theory | |
CN111815609B (en) | Pathological image classification method and system based on context awareness and multi-model fusion | |
CN105184225A (en) | Multinational paper money image identification method and apparatus | |
CN106529470A (en) | Gesture recognition method based on multistage depth convolution neural network | |
CN104504391B (en) | A kind of hyperspectral image classification method based on sparse features and markov random file | |
CN106909881A (en) | The method and system of corn breeding base ridge number are extracted based on unmanned aerial vehicle remote sensing images | |
CN106778491A (en) | The acquisition methods and equipment of face 3D characteristic informations | |
CN103729851A (en) | Image segmentation method based on significant transitional region | |
JP5164802B2 (en) | Recognition system, recognition method, and recognition program |
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 |