CN104915636B - Remote sensing image road recognition methods based on multistage frame significant characteristics - Google Patents

Remote sensing image road recognition methods based on multistage frame significant characteristics Download PDF

Info

Publication number
CN104915636B
CN104915636B CN201510178609.2A CN201510178609A CN104915636B CN 104915636 B CN104915636 B CN 104915636B CN 201510178609 A CN201510178609 A CN 201510178609A CN 104915636 B CN104915636 B CN 104915636B
Authority
CN
China
Prior art keywords
road
sample
image
region
road network
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.)
Active
Application number
CN201510178609.2A
Other languages
Chinese (zh)
Other versions
CN104915636A (en
Inventor
张菁
王超
卓力
曹嫣
周倩兰
耿文浩
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beijing University of Technology
Original Assignee
Beijing University of Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Beijing University of Technology filed Critical Beijing University of Technology
Priority to CN201510178609.2A priority Critical patent/CN104915636B/en
Publication of CN104915636A publication Critical patent/CN104915636A/en
Application granted granted Critical
Publication of CN104915636B publication Critical patent/CN104915636B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Abstract

The invention discloses a kind of remote sensing image road recognition methods based on multistage frame significant characteristics, are related to field of image processing, and in particular to remote sensing handles identification field.The method of the present invention makes full use of the spectrum of road, physics and morphological feature in remote sensing image, traditional significant characteristics are improved, it is proposed two kinds of significant characteristics of SPECTRAL DIVERSITY and local linear of remote sensing image, simultaneously, extracted road network is further analyzed, obtains the affiliated type in the section.The significant characteristics of remote sensing image are extracted first, and two kinds of features of SPECTRAL DIVERSITY and local linear are merged, establish multistage frame notable figure, then RPS method, MAT method etc. is taken to optimize road network, further eliminate non-rice habitats region, the feature of road network is finally extracted, and is classified using Increment Learning Algorithm to road, realizes the remote sensing image road identification of multistage frame significant characteristics.The present invention improves the road Identification rate of remote sensing images.

Description

Remote sensing image road recognition methods based on multistage frame significant characteristics
Technical field
The present invention discloses a kind of remote sensing image based on multistage frame significant characteristics using remote sensing image as research object Roads recognition method is related to field of image processing, and in particular to remote sensing handles identification field.
Background technique
In recent years, with the rapid development of the technologies such as sensor, remote sensing platform, data communication, remote sensing technology enters one It is a can dynamic, quickly, it is accurate, in time and provide the new stage of various earth observation data with more means, obtain ring for people Border information, epistemic context provide an important channel.
Remote sensing image is to be carried out from the workbench far from ground by electromagnetic wave information of the sensor to earth surface Detection finally obtains remote sensing image then using information transmission, processing and analysis.It has very high spatial resolution and Spectral resolution, can be at several hundred or even thousands of a continuous spectrum wave bands obtain ground object target image.At the same time, with intelligence The it is proposed in intelligent city, road are the main body of modern traffic system as important artificial atural object, have important geography, politics, Economic significance, road are also main record and mark object in map and GIS-Geographic Information System.In the mid 1970s, by In the needs for digitizing geographical traffic information, road image, which automatically extracts technology, to be occurred therewith and gradually develops.Nowadays, multispectral The appearance of high-resolution remote sensing satellite, imaging radar, UAV, so that earth observation means are more complete, it is geographical Image data becomes increasingly abundant.On the other hand, many applications such as mapping, GIS-Geographic Information System update, city observation and planning The appearance of demand and constantly growth, promote automatic road extractive technique to continue to develop.Therefore, the Road Detection in remote sensing image with Identification becomes one of the research hotspot of remote sensing fields.
Remote sensing image recognition methods based on conspicuousness, the conspicuousness according to presentation content, which is established, is suitable for remote sensing image Conspicuousness model carries out comprehensive analysis to the features such as its spectral signature and texture, structure, can be fast in complicated image environment Fast fixation and recognition region, eliminates the redundancy of image to a certain extent, highlights the main contents of image, reduces shadow As handling the complexity of analysis and reducing semantic gap.It can be detected under conditions of no prior information with ambient background in the presence of poor Different target (man-made target in such as natural background).The priori spectral information of background or target is not needed, practicability is stronger, is Remote sensing image target identification, image classification and retrieval lay the foundation.For this purpose, significant characteristics are introduced remote sensing image by the present invention In road Identification, a kind of remote sensing image road recognition methods based on multistage frame significant characteristics is proposed.
Summary of the invention
The present invention is different from existing remote sensing image road recognition methods, special for road area conspicuousness in remote sensing image Obvious feature is levied, traditional remote sensing image identification technology is improved, proposes the road of multistage frame significant characteristics Recognition methods.The invention improves traditional significant characteristics, carries out the more of SPECTRAL DIVERSITY and local linear track to remote sensing image Grade frame significant characteristics extract, then optimize to obtained road network: using in post-processing technology removal road network Isolated patch and noise eliminate non-rice habitats area using partial region segmentation (Region Part Segmentation, RPS) method Domain removes the region of similar road spectral signature by Medial-Axis Transformation (Medial Axis Transform, MAT) method, such as Parking lot, heavy construction roof.Finally road network is trained using the incremental learning based on neighborhood covering method, is realized The road Identification of remote sensing image.Main process is as shown in Fig. 1, can be divided into following steps: based on multistage frame conspicuousness The road network extraction of feature, the road network optimization based on multistage frame significant characteristics and the road kind based on incremental learning Class identification.
Step 1: the extraction of multistage frame significant characteristics
Step 1.1: being extracted based on the local linear significant characteristics for dominating unusual measure
Picture is calculated using the method for gradient matrix singular value decomposition (Singular Value Decomposition, SVD) Then dominant direction threshold value is arranged in the dominant direction operator of vegetarian refreshments, remove the background area without dominant direction, realizes road The extraction of edge pixel;
Step 1.1.1: gradient matrix calculates
For each pixel in image, the gradient vector of N × N size windows around the pixel, image f are calculated first Point (x in (x, y)k,yk) gradient be calculate by the following formula:
WhereinIndicate one-dimensional convolution;
Step 1.1.2: the singular value decomposition of gradient matrix
After obtaining gradient matrix, the covariance formula for calculating gradient matrix SVD is as follows:
Wherein s1,s2,s3,s4For the singular value component of gradient matrix, s1For principal direction singular value;
Step 1.1.3: the calculating of leading singular operator
Leading singular operator (Dominant Singular Operator, DSO) is used to calculate principal direction singular value and surprise The ratio of different value summation illustrates that the peripheral region of the pixel is elongated shape when the ratio is close to 1;Therefore, DSO is defined such as Under:
As can be seen from the above equation, when all gradient components have the same direction, only one singular value is non-zero, I.e. the value of DSO is 1;If four singular values are equal, the value of DSO is 0.25, therefore the range of DSO value is [0.25,1];This Invention defines a threshold value, when the value of DSO is less than the threshold value, illustrate corresponding image blocks be it is noisy, without dominant direction, The feature for not embodying road local linear is accordingly regarded as background area i.e. non-rice habitats region;It is opposite then be counted as Road region;
Step 1.2: the SPECTRAL DIVERSITY significant characteristics based on probabilistic SVMs extract
Remote sensing image is subjected to piecemeal processing first, then extracts the spectral signature of image block, and is indicated with feature vector, The probability tag of pixel is calculated finally by probabilistic SVMs;
Remote sensing image is pre-processed using Tile partitioned mode, by the target image of Remote Sensing Image Database from left to right, It is divided into rule from top to bottom, is not overlapped and equal-sized subimage block;
The spectral signature of image block is extracted, training sample is chosen, training set is divided into non-rice habitats sub-image and road sub-image Block extracts its histogram as characteristic vector, comprising flat for each sub-image sample in three wave bands of red, green, blue Mean value, energy, standard deviation and entropy, each sample are indicated with characteristic vector, and calculate the later remotely-sensed data of piecemeal, then pass through The characteristic vector of these samples of SVM classifier training;
The probability P of identification sample is obtained by probabilistic SVMs:
Wherein y is two tag along sorts, fsIt is the label obtained from SVM decision function, parameter A and B are estimated by maximum likelihood It counts training set to obtain, for each image blocks, if P value is greater than the threshold alpha of defined, assert that the sample is roadway area Domain, it is on the contrary then regard as non-rice habitats region;
Step 1.3: taking the multistage frame significant characteristics of full neural network to merge based on winner
Road edge notable figure is denoted as S firstl, road area notable figure is denoted as Sg, merging is normalized i.e. to the two Total notable figure S can be obtained, calculation formula is as follows:
Wherein N () is normalization factor, and normalization operator N () is calculated by following three steps: 1) will be every The tonal range of one characteristic pattern is normalized to some particular range [O, M], to eliminate the amplitude difference under different characteristic mode; 2) the maximum value M of each characteristic pattern and the average value of every other local maximum are found out3) by characteristic pattern all multiplied byIt is normalized by using N () operator, so that local the larger value on different characteristic figure is enhanced, and those The unconspicuous characteristic pattern of local maximum is inhibited, so that significant region is protruded, and uniform non-significant area Domain is ignored;
Then pass through the probability tag of each pixel of WTA neural computing, the probability tag of neuron is each It is updated by the way of " winner takes entirely " after secondary iteration, if the kth class of the neuron is winner, to the general of kth class Rate increases a constant;On the contrary then one constant of reduction;
WTA method increases road area pixel and deletes non-rice habitats area pixel with iterating, therefore is substantially one The fusion process of a correction mechanism, the method for the present invention need road edge and provincial characteristics as input, give after fusion Pixel one final label road or non-rice habitats, output of this label as WTA neuron, pixel belong to two classes Probability: road k=1, non-rice habitats k=2 are constrained to P (1)+P (2)=1;Compare the probability of two classes, pixel is confirmed as Road or non-rice habitats, it is final to realize road edge and road area fusion;
Step 2: the optimization of multistage frame significant characteristics
Step 2.1: the post-processing of road network
The method that the present invention takes connected component labeling retains disjoint network line segment, removes non-rice habitats region;It calculates The area of each connected region deletes the region for being less than area threshold, while calculating the eccentricity in each connection region, deletes Less than the region of eccentricity threshold value, the calculation formula of eccentricity is as follows:
E=μ2002
Wherein
Step 2.2: the optimum of road net based on partial region dividing method
Partial region dividing method is taken to eliminate the non-rice habitats region of road edge, steps are as follows:
Step 2.2.1: the internal and external contour line of road network is smoothed;
Step 2.2.2: the curvature of smoothed profile is calculated;One curve is expressed as parametric form, wherein t indicates path length Degree, x and y are the coordinates of profile:
R (t)=(x (t), y (t))
Curvature is defined as the change rate of slope, is calculated by the following formula:
Step 2.2.3: determining local extremum, i.e., the derivative of local curvature be 0 point;
Step 2.2.4: by the outer/inner profile of tracing area, using Local Extremum as convex/recessed leading point (Convex/ Concave Dominant Points, CDPs), wherein the convex leading point of outer profile is set as CDPcx, the recessed leading point of Internal periphery sets For CDPce
Step 2.2.5:CDPcxIntroversive movement, CDP are carried out along its normal directionceMotion outward is carried out along its normal direction, The stopping when encountering another CDP moved in the same contour line;
1) starting point of all CDPs freezed is tracked, and connects corresponding CDPs, which will be removed;
Step 2.3: the optimum of road net based on central axes transformation theory
The present invention extracts the central axes of road, then retains the part for being less than mean breadth in road network, deletes width Excessive non-rice habitats region, specific steps are as follows:
Step 2.3.1: the edge contour in image, smooth edges pixel the determination of central axes skeleton: are determined first;So The each pixel for tracking one edge profile afterwards, finds the relating dot of the another one edge corresponding to it;Relating dot is reconnected, Using the midpoint of connecting line as axial point;All axial points for finally connecting whole edge contour, obtain the Medial Axis Skeleton of image;
Step 2.3.2: the elimination in non-rice habitats region: after obtaining central axes skeleton, the attribute of each section is calculated, includes The width, length and the standard deviation with normal width of this section;Then it begins stepping through, will be unsatisfactory for from the source node of road network again The part of hypothesis removes, i.e. the removal mean breadth region excessive greater than normal width or standard deviation;
Step 3: the road category identification based on incremental learning
Step 3.1: the selection of road network feature
The characteristics of according to highway, urban road, mountain road, extract the characterization factor of road network:
Area reflects the size in region shared by road;Flexibility is the length-width ratio of region of out bound rectangle, is able to reflect road Width;Crosspoint number counts the crosspoint number in road network;The completeness of curvature expression road network;Textural characteristics are Using gray level co-occurrence matrixes, energy, gray scale correlation, 5 local stationary, entropy and the moment of inertia textural characteristics coefficients are calculated;
Then these characterization factor vectorizations are obtained into feature vector, finally using Increment Learning Algorithm to feature vector into Row training and identification;
Step 3.2: the feature identification based on incremental learning
The Increment Learning Algorithm that the present invention is taken based on covering is trained and identifies to characteristics of image, is based on incremental learning The training of method includes: that (1) is trained the feature of initial sample with recognition methods, generates fundamental classifier, and building is initial Disaggregated model;(2) the newly-increased data of current a batch are identified, existing classifier is adjusted according to recognition result, is applied Classifier adjusted increases data newly to next group again and identifies;(3) repeatedly, it is adjusted every time according to current recognition result Existing classifier, until all newly-increased data have identified;
Step 3.2.1: the foundation of preliminary classification model
The present invention is using the initial sample of neighborhood covering method training, just for the road image identification building based on incremental learning Beginning disaggregated model, the method is as follows:
1) initialization of neighborhood covering method
Road network sample is initialized first:
(1) road network sample set K={ (x is set1,y1),(x2,y2),...,(xp,yp), wherein there are p sample, x in Ki For the feature vector of road network, each input sample xi(i=1 2 ..., p) has n dimension attribute, yiFor xiCorresponding road network Type;
(2) y different in the corresponding output of p sample is setiThere is q, enabling I (t), (t=1 2 ..., q) respectively represents order The output of sample is ytAll specimen numbers set;Corresponding input set is denoted as P (t), P (t)={ xi|i∈I (t)};
2) neighborhood covering method constructs preliminary classification model
After to the initialization of road network sample, neighborhood covering method constructs preliminary classification model by following steps:
(1) road network training sample set is set as X={ (xt,yt), t=1,2 ..., p }, sample is divided into s class;
(2) the maximum norm r of sample in sample set X is sought, such as shown in (19);Then all sample points in X are successively pressed into public affairs Formula (20) transformation, and center is projected on origin, the hypersphere that radius is R (R=r+1),
Wherein T (xk) be sample point projection after new coordinate;
R=max | xk||xk∈X}
(3) I (t) corresponding to sample set X after projection is sought, P (t) (t=1,2 ..., s);
(4) a sample point x uncovered in P (t) is randomly selectedk, enable
Wherein < xk,xm> indicates xkAnd xmCarry out inner product operation, d1(k) expression and xkThe maximum value of foreign peoples's point inner product, d2(k) Expression and xkThe minimum value of similar inner product;
(5) make with xkThe covering for being θ=d (k) for spherical field center, radiusIts InIndicate j-th of covering of the i-th class sample;
(6) point being capped in P (t) is marked, whether all points are marked in training of judgement sample set X Note;If so, the method for the present invention terminates;If it is not, continuing to judge whether sample point is all capped and marks, i=i+1, j in P (t) =1, t=t+1 is returned (4), if it is not, enabling j=j+1, is returned (4);
The method of the present invention finally finds out one group of coveringC is The training obtained overlay model of initial road network feature;
Step 3.2.2: the identification of newly-increased sample
Steps are as follows for increment method based on covering:
(1) initial road network sample is learnt, constructs preliminary classification model, obtains one group of covering C;
(2) the road network data R newly-increased to current a batch with existing modelnIt is identified;
(3) according to recognition result, existing disaggregated model is adjusted, obtains new model, the method is as follows: first from RnIn find the sample of identification mistake, and the corresponding covering C of error sample in initial modelmIt deletes;Then by identification mistake Sample and rejection sample re-start training, obtain new covering Ca;The covering C that finally will newly obtainaIt is added to original classification In model, so far, new disaggregated model C is obtainedn, and Cn=C-Cm+Ca
(4) (2) (3) are repeated until all newly-increased specimen discernings are completed.
Compared with prior art, the present invention have following apparent advantage and the utility model has the advantages that
The present invention makes full use of the spectrum of road, physics and form in remote sensing image using remote sensing image as research object Feature improves traditional significant characteristics, proposes that two kinds of the SPECTRAL DIVERSITY of remote sensing image and local linear track are significant Property feature, meanwhile, extracted road network is further analyzed, obtains the affiliated type in the section.Remote sensing is extracted first The significant characteristics of image, and SPECTRAL DIVERSITY and local two kinds of features of linear track are merged, it is significant to establish multistage frame Figure;Then it takes RPS method, MAT method etc. to optimize road network, further eliminates non-rice habitats region;Finally extract The feature of road network, and classified using Increment Learning Algorithm to road, realize the remote sensing of multistage frame significant characteristics Image road identification.
Detailed description of the invention:
Remote sensing image road recognition methods process of the Fig. 1 based on multistage frame significant characteristics;
Fig. 2 road edge extracts process;
Fig. 3 road area extracts process;
Fig. 4 multistage frame significant characteristics merge process;
Fig. 5 Medial-Axis Transformation method schematic diagram;
Identification process of the Fig. 6 based on incremental learning;
Fig. 7 road network sample set and the distribution schematic diagram for being covered on two-dimensional space;
Fig. 8 multistage frame significant characteristics extract;
The optimization of Fig. 9 multistage frame significant characteristics.
Specific embodiment
It is a specific implementing procedure below, but the range that this patent is protected is not limited to this according to foregoing description Implementing procedure.The present invention chooses the multispectral satellite image of 512 × 512 sizes as experiment sample, experimental situation 2.10GHz The PC machine of CPU, 5.00G memory, realization environment are Windows 7+Visual Studio 2010.
The extraction for carrying out multistage frame significant characteristics first, extracts remote sensing image using DSM method and P-SVM respectively Then SPECTRAL DIVERSITY and local linear significant characteristics carry out multistage frame conspicuousness spy using " winner takes entirely " neural network again The extraction of multistage frame significant characteristics is realized in sign fusion.Experimental result is as shown in Fig. 8, wherein (a) is 512 × 512 sizes Original input picture, (b) extracted for DSM road edge as a result, (c) P-SVM road area extracts as a result, (d) neural network Fusion Features result.
Then the optimization for carrying out multistage frame significant characteristics, is respectively adopted post processing of image, partial region dividing method Road network is optimized with central axes transform method, removes the non-rice habitats region in road network, further increases road The topological and integrality of net.Experimental result is as shown in Fig. 9, wherein (a) is post processing of image as a result, (b) RPS method is handled As a result, (c) MAT method processing result.

Claims (1)

1. the remote sensing image road recognition methods based on multistage frame significant characteristics, it is characterised in that include the following steps:
Step 1: the extraction of multistage frame significant characteristics
Step 1.1: being extracted based on the local linear significant characteristics for dominating unusual measure
Pixel is calculated using the method for gradient matrix singular value decomposition (Singular Value Decomposition, SVD) Dominant direction operator, then be arranged dominant direction threshold value, remove the background area without dominant direction, realize road edge The extraction of pixel;
Step 1.1.1: gradient matrix calculates
For each pixel in image, the gradient vector of N × N size windows around the pixel is calculated first, image f (x, Y) point (x ink,yk) gradient be calculate by the following formula:
Wherein
Indicate one-dimensional convolution;
Step 1.1.2: the singular value decomposition of gradient matrix
After obtaining gradient matrix, the covariance formula for calculating gradient matrix SVD is as follows:
Wherein s1,s2,s3,s4For the singular value component of gradient matrix, s1For principal direction singular value;
Step 1.1.3: the calculating of leading singular operator
Leading singular operator (Dominant Singular Operator, DSO) is used to calculate principal direction singular value and singular value The ratio of summation illustrates that the peripheral region of the pixel is elongated shape when the ratio is close to 1;Therefore, DSO is defined as follows:
As can be seen from the above equation, when all gradient components have the same direction, only one singular value is non-zero, i.e., The value of DSO is 1;If four singular values are equal, the value of DSO is 0.25, therefore the range of DSO value is [0.25,1];This hair One threshold value of bright definition, when the value of DSO be less than the threshold value when, illustrate corresponding image blocks be it is noisy, without dominant direction, do not have There is the feature for embodying road local linear, is accordingly regarded as background area i.e. non-rice habitats region;It is opposite then be counted as road Region;
Step 1.2: the SPECTRAL DIVERSITY significant characteristics based on probabilistic SVMs extract
Remote sensing image is subjected to piecemeal processing first, then extracts the spectral signature of image block, and is indicated with feature vector, finally The probability tag of pixel is calculated by probabilistic SVMs;
Remote sensing image is pre-processed using Tile partitioned mode, by the target image of Remote Sensing Image Database from left to right, from upper It is divided into rule under, is not overlapped and equal-sized subimage block;
The spectral signature of image block is extracted, chooses training sample, training set is divided into non-rice habitats sub-image and road sub-image block, right In each sub-image sample, its histogram is extracted in three wave bands of red, green, blue as characteristic vector, comprising average value, Energy, standard deviation and entropy, each sample are indicated with characteristic vector, and calculate the later remotely-sensed data of piecemeal, then pass through svm classifier The characteristic vector of these samples of device training;
The probability P of identification sample is obtained by probabilistic SVMs:
Wherein Y is two tag along sorts, fsIt is the label obtained from SVM decision function, parameter A and B passes through maximal possibility estimation training Collection obtains, and for each image blocks, if P value is greater than the threshold alpha of defined, assert that the sample is road area, instead Then regard as non-rice habitats region;
Step 1.3: taking the multistage frame significant characteristics of full neural network to merge based on winner
Road edge notable figure is denoted as S firstl, road area notable figure is denoted as Sg, merging, which is normalized, to the two to obtain To total notable figure S, calculation formula is as follows:
Wherein N () is normalization factor, and normalization operator N () is calculated by following three steps: 1) by each The tonal range of characteristic pattern is normalized to some particular range [O, M], to eliminate the amplitude difference under different characteristic mode;2) it looks for The maximum value M of each characteristic pattern and the average value of every other local maximum out3) by characteristic pattern all multiplied byIt is normalized by using N () operator, so that local the larger value on different characteristic figure is enhanced, and those The unconspicuous characteristic pattern of local maximum is inhibited, so that significant region is protruded, and uniform non-significant area Domain is ignored;
Then pass through the probability tag of each pixel of WTA neural computing, the probability tag of neuron changes each time It is updated by the way of " winner takes entirely " after generation, if the kth class of the neuron is winner, the probability of kth class is increased Add a constant;On the contrary then one constant of reduction;
WTA method increases road area pixel and deletes non-rice habitats area pixel with iterating, therefore is substantially a school The fusion process of positive mechanism, the method for the present invention need road edge and provincial characteristics as input, pixel are given after fusion One final label road or non-rice habitats, output of this label as WTA neuron, pixel belong to the general of two classes Rate: road k=1, non-rice habitats k=2 are constrained to P (1)+P (2)=1;Compare the probability of two classes, pixel is confirmed as road Or non-rice habitats, it is final to realize road edge and road area fusion;
Step 2: the optimization of multistage frame significant characteristics
Step 2.1: the post-processing of road network
The method that the present invention takes connected component labeling retains disjoint network line segment, removes non-rice habitats region;It calculates each The area of connected region deletes the region for being less than area threshold, while calculating the eccentricity in each connection region, and deletion is less than The region of eccentricity threshold value;
Step 2.2: the optimum of road net based on partial region dividing method
Partial region dividing method is taken to eliminate the non-rice habitats region of road edge, steps are as follows:
Step 2.2.1: the internal and external contour line of road network is smoothed;
Step 2.2.2: the curvature of smoothed profile is calculated;One curve is expressed as parametric form, wherein t indicates path length, x It is the coordinate of profile with y:
R (t)=(x (t), y (t))
Curvature is defined as the change rate of slope, is calculated by the following formula:
Step 2.2.3: determining local extremum, i.e., the derivative of local curvature be 0 point;
Step 2.2.4: by the outer/inner profile of tracing area, using Local Extremum as convex/recessed leading point (Convex/ Concave Dominant Points, CDPs), wherein the convex leading point of outer profile is set as CDPcx, the recessed leading point of Internal periphery sets For CDPce
Step 2.2.5:CDPcxIntroversive movement, CDP are carried out along its normal directionceMotion outward is carried out along its normal direction, works as chance Stop when another CDP moved in the same contour line;
1) starting point of all CDPs freezed is tracked, and connects corresponding CDPs, which will be removed;
Step 2.3: the optimum of road net based on central axes transformation theory
The present invention extracts the central axes of road, then retains the part for being less than mean breadth in road network, it is excessive to delete width Non-rice habitats region, specific steps are as follows:
Step 2.3.1: the edge contour in image, smooth edges pixel the determination of central axes skeleton: are determined first;Then with Each pixel of track one edge profile finds the relating dot of the another one edge corresponding to it;Relating dot is reconnected, it will even The midpoint of wiring is as axial point;All axial points for finally connecting whole edge contour, obtain the Medial Axis Skeleton of image;
Step 2.3.2: the elimination in non-rice habitats region: after obtaining central axes skeleton, calculating the attribute of each section, comprising in this Width, length and the standard deviation with normal width of axis skeleton;Then it is begun stepping through again from the source node of road network, it will not Meet the part removal assumed, i.e. the removal mean breadth region excessive greater than normal width or standard deviation;
Step 3: the road category identification based on incremental learning
Step 3.1: the selection of road network feature
The characteristics of according to highway, urban road, mountain road, extract the characterization factor of road network:
Area reflects the size in region shared by road;Flexibility is the length-width ratio of region of out bound rectangle, is able to reflect the width of road Degree;Crosspoint number counts the crosspoint number in road network;The completeness of curvature expression road network;Textural characteristics are to utilize Gray level co-occurrence matrixes calculate energy, gray scale correlation, 5 local stationary, entropy and the moment of inertia textural characteristics coefficients;
Then these characterization factor vectorizations are obtained into feature vector, finally feature vector is instructed using Increment Learning Algorithm Practice and identifies;
Step 3.2: the feature identification based on incremental learning
The Increment Learning Algorithm that the present invention is taken based on covering is trained and identifies to characteristics of image, is based on Increment Learning Algorithm Training with recognition methods include: that (1) is trained the feature of initial sample, generate fundamental classifier, construct preliminary classification Model;(2) the newly-increased data of current a batch are identified, existing classifier is adjusted according to recognition result, using adjustment Classifier afterwards increases data newly to next group again and identifies;(3) repeatedly, existing according to the adjustment of current recognition result every time Classifier, until all newly-increased data have identified;
Step 3.2.1: the foundation of preliminary classification model
The present invention is using the initial sample of neighborhood covering method training, for initial point of the road image identification building based on incremental learning Class model, the method is as follows:
1) initialization of neighborhood covering method
Road network sample is initialized first:
(1) road network sample set K={ (x is set1,y1),(x2,y2),...,(xp,yp), wherein there are p sample, x in KiFor road The feature vector of road network, each input sample xi(i=1 2 ..., p) has n dimension attribute, yiFor xiThe kind of corresponding road network Class;
(2) y different in the corresponding output of p sample is setiThere is q, enabling I (t), (t=1 2 ..., q) is respectively represented and enabled sample Output is ytAll specimen numbers set;Corresponding input set is denoted as P (t), P (t)={ xi|i∈I(t)};
2) neighborhood covering method constructs preliminary classification model
After to the initialization of road network sample, neighborhood covering method constructs preliminary classification model by following steps:
(1) road network training sample set is set as X={ (xt,yt), t=1,2 ..., p }, sample is divided into s class;
(2) the maximum norm r of sample in sample set X is sought, such as shown in (19);Then all sample points in X are successively pressed into formula (20) it converts, and projects center on origin, the hypersphere that radius is R (R=r+1), wherein T (xk) it is after sample point projects New coordinate;
R=max | xk||xk∈X}
(3) I (t) corresponding to sample set X after projection is sought, P (t) (t=1,2 ..., s);
(4) a sample point x uncovered in P (t) is randomly selectedk, enable
Wherein < xk,xm> indicate xkAnd xmCarry out inner product operation, d1(k) expression and xkThe maximum value of foreign peoples's point inner product, d2(k) it indicates With xkThe minimum value of similar inner product;
(5) make with xkThe covering for being θ=d (k) for spherical field center, radiusWherein Indicate j-th of covering of the i-th class sample;
(6) point being capped in P (t) is marked, whether all points are labeled in training of judgement sample set X; If so, the method for the present invention terminates;If it is not, continuing to judge whether sample point is all capped and marks, i=i+1, j=in P (t) 1, t=t+1 is returned (4), if it is not, enabling j=j+1, is returned (4);
The method of the present invention finally finds out one group of coveringC is to train The obtained overlay model of initial road network feature;
Step 3.2.2: the identification of newly-increased sample
Steps are as follows for increment method based on covering:
(1) initial road network sample is learnt, constructs preliminary classification model, obtains one group of covering C;
(2) the road network data R newly-increased to current a batch with existing modelnIt is identified;
(3) according to recognition result, existing disaggregated model is adjusted, obtains new model, the method is as follows: first from RnIn The sample of identification mistake is found, and the corresponding covering C of error sample in initial modelmIt deletes;Then by the sample of identification mistake This and rejection sample re-start training, obtain new covering Ca;The covering C that finally will newly obtainaIt is added to original classification mould In type, so far, new disaggregated model C is obtainedn, and Cn=C-Cm+Ca
(4) (2) (3) are repeated until all newly-increased specimen discernings are completed.
CN201510178609.2A 2015-04-15 2015-04-15 Remote sensing image road recognition methods based on multistage frame significant characteristics Active CN104915636B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201510178609.2A CN104915636B (en) 2015-04-15 2015-04-15 Remote sensing image road recognition methods based on multistage frame significant characteristics

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201510178609.2A CN104915636B (en) 2015-04-15 2015-04-15 Remote sensing image road recognition methods based on multistage frame significant characteristics

Publications (2)

Publication Number Publication Date
CN104915636A CN104915636A (en) 2015-09-16
CN104915636B true CN104915636B (en) 2019-04-12

Family

ID=54084688

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201510178609.2A Active CN104915636B (en) 2015-04-15 2015-04-15 Remote sensing image road recognition methods based on multistage frame significant characteristics

Country Status (1)

Country Link
CN (1) CN104915636B (en)

Families Citing this family (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108475349A (en) * 2016-02-12 2018-08-31 谷歌有限责任公司 System and method for the study of steady large-scale machines
CN106127144B (en) * 2016-06-21 2019-02-01 沈阳航空航天大学 Using when compose the point source risk source extraction method of empty integral feature model
CN106778547B (en) * 2016-11-29 2020-06-30 中南大学 Method and device for constructing remote sensing image ground feature classification reference library
CN106778605B (en) * 2016-12-14 2020-05-05 武汉大学 Automatic remote sensing image road network extraction method under assistance of navigation data
CN108229261A (en) * 2016-12-22 2018-06-29 北京市商汤科技开发有限公司 Remote sensing images detection method, device and computer equipment
CN106780468B (en) * 2016-12-22 2019-09-03 中国计量大学 The conspicuousness detection method of view-based access control model perception positive feedback
CN108229270B (en) * 2017-01-12 2021-07-30 北京市商汤科技开发有限公司 Method, device and electronic equipment for identifying road from remote sensing image
CN108229273B (en) * 2017-02-27 2021-03-05 北京市商汤科技开发有限公司 Method and device for training multilayer neural network model and recognizing road characteristics
CN108229274B (en) * 2017-02-28 2020-09-04 北京市商汤科技开发有限公司 Method and device for training multilayer neural network model and recognizing road characteristics
CN109427062A (en) * 2017-08-30 2019-03-05 深圳星行科技有限公司 Roadway characteristic labeling method, device, computer equipment and readable storage medium storing program for executing
CN108230421A (en) * 2017-09-19 2018-06-29 北京市商汤科技开发有限公司 A kind of road drawing generating method, device, electronic equipment and computer storage media
CN107578446A (en) * 2017-09-19 2018-01-12 中国人民解放军信息工程大学 A kind of method for extracting remote sensing image road and device
CN108256425B (en) * 2017-12-12 2019-06-25 交通运输部规划研究院 A method of harbour container is extracted using Remote Spectra efficient information rate
CN108446599A (en) * 2018-02-27 2018-08-24 首都师范大学 A kind of high spectrum image wave band fast selecting method of p value statistic modeling independence
CN108920995A (en) * 2018-04-08 2018-11-30 华中科技大学 Intelligent security guard video monitoring method and its system and monitor terminal
CN109035196A (en) * 2018-05-22 2018-12-18 安徽大学 Image local fuzzy detection method based on conspicuousness
CN108804616B (en) * 2018-05-30 2020-12-08 中国科学院空间应用工程与技术中心 Device and method for mining on-orbit image data of space payload
CN109522787B (en) * 2018-09-30 2020-10-09 广州地理研究所 Tiny road identification method based on remote sensing data
CN109493320A (en) * 2018-10-11 2019-03-19 苏州中科天启遥感科技有限公司 Method for extracting remote sensing image road and system, storage medium, electronic equipment based on deep learning
CN111476197A (en) * 2020-04-24 2020-07-31 中科天盛卫星技术服务有限公司 Oil palm identification and area extraction method and system based on multi-source satellite remote sensing image

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP1679639A1 (en) * 2005-01-10 2006-07-12 Delphi Technologies, Inc. Object classification method for a collision warning system
CN1932850A (en) * 2006-10-09 2007-03-21 武汉大学 Remoto sensing image space shape characteristics extracting and sorting method
CN102005034A (en) * 2010-12-01 2011-04-06 南京大学 Remote sensing image segmentation method based on region clustering
CN102496034A (en) * 2011-11-29 2012-06-13 南京师范大学 High-spatial resolution remote-sensing image bag-of-word classification method based on linear words

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
EP1679639A1 (en) * 2005-01-10 2006-07-12 Delphi Technologies, Inc. Object classification method for a collision warning system
CN1932850A (en) * 2006-10-09 2007-03-21 武汉大学 Remoto sensing image space shape characteristics extracting and sorting method
CN102005034A (en) * 2010-12-01 2011-04-06 南京大学 Remote sensing image segmentation method based on region clustering
CN102496034A (en) * 2011-11-29 2012-06-13 南京师范大学 High-spatial resolution remote-sensing image bag-of-word classification method based on linear words

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
遥感图像道路提取研究;滕鑫鹏;《中国博士学位论文全文数据库·信息科技辑》;20140815(第8期);I140-26

Also Published As

Publication number Publication date
CN104915636A (en) 2015-09-16

Similar Documents

Publication Publication Date Title
CN104915636B (en) Remote sensing image road recognition methods based on multistage frame significant characteristics
CN108573276B (en) Change detection method based on high-resolution remote sensing image
CN106682598B (en) Multi-pose face feature point detection method based on cascade regression
CN106909902B (en) Remote sensing target detection method based on improved hierarchical significant model
Song et al. Road extraction using SVM and image segmentation
CN103049763B (en) Context-constraint-based target identification method
CN108304873B (en) Target detection method and system based on high-resolution optical satellite remote sensing image
Shi et al. Road detection from remote sensing images by generative adversarial networks
CN103258203B (en) The center line of road extraction method of remote sensing image
CN103218831B (en) A kind of video frequency motion target classifying identification method based on profile constraint
CN108052966B (en) Remote sensing image scene automatic extraction and classification method based on convolutional neural network
CN105528794A (en) Moving object detection method based on Gaussian mixture model and superpixel segmentation
CN107239759B (en) High-spatial-resolution remote sensing image transfer learning method based on depth features
CN103927511A (en) Image identification method based on difference feature description
Guan et al. Real-time lane-vehicle detection and tracking system
Qin A mean shift vector-based shape feature for classification of high spatial resolution remotely sensed imagery
CN109543632A (en) A kind of deep layer network pedestrian detection method based on the guidance of shallow-layer Fusion Features
Alidoost et al. A CNN-based approach for automatic building detection and recognition of roof types using a single aerial image
Zang et al. Traffic lane detection using fully convolutional neural network
Chen et al. Object-based multi-modal convolution neural networks for building extraction using panchromatic and multispectral imagery
Singh et al. A hybrid approach for information extraction from high resolution satellite imagery
Li et al. Semantic segmentation for high-resolution aerial imagery using multi-skip network and Markov random fields
Joshi et al. Automatic rooftop detection using a two-stage classification
Moussa et al. Manmade objects classification from satellite/aerial imagery using neural networks
Xing et al. Traffic sign recognition using guided image filtering

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