A kind of high-resolution remote sensing image crossing automatic identifying method
Technical field
Field is extracted the present invention relates to remote sensing image information more particularly to a kind of high-resolution remote sensing image crossing is known automatically
Other method.
Background technique
Extracting road information using high-resolution remote sensing image is a kind of economical and efficient means, can not only be compared by visual observation
To very intuitively and easily verifying extract as a result, and its cost be also significantly lower than based on field measured data or motion track
The road information acquisition modes of data.But object spectrum abundant information, complex distribution in remote sensing image, " the different spectrum of jljl " with
The phenomenon that " foreign matter is with spectrum ", is generally existing, and same type of spectral characteristic of ground difference becomes larger, different types of object spectrum phase
Seemingly, the Spectral divisibility of image reduces.Currently, extracted from high-resolution remote sensing image road information still face it is many it is difficult with
Challenge.
Crossing (plane crossing is most commonly seen, and " crossing " mentioned by the present invention all refers to plane crossing) is the weight of road network
Component part is wanted, road intersection is located at.Accurate intersection information is to Image registration, intelligent transportation, GIS data update and road
Variation detection etc. is significant.In general, urban road is more straight, therefore it can determine whether out urban road network using crossing
General configuration, such as: connection adjacent intersection can get the general picture of this area's road network.Due to road construction material and surrounding its
His atural object is similar, there is similar spectral signature, and on the other hand, crossing is easy by Adjacent Buildings, parking lot, shade, plant
The interference such as quilt, vehicle.It is not yet thoroughly solved currently, extracting crossing from high-resolution remote sensing image, using high-definition remote sensing
The business software system that image automatically extracts crossing not yet occurs.
It is less for the research of crossing extraction in known references, not yet occur being suitable for high-resolution remote sensing image crossing
Antihunt means and business software.Need to solve two big technical problems: first, global automatic acquisition high-resolution remote sensing image crossing
Position candidate point;Second, using the local detection crossing of piecemeal rectangle rotating model of design, improve crossing identification accuracy and
Integrality.
Summary of the invention
The purpose of the present invention is to provide a kind of high-resolution remote sensing image crossing automatic identifying methods.
The technical solution adopted by the present invention is that:
A kind of high-resolution remote sensing image crossing automatic identifying method comprising following steps:
S1, road primitives extract: first to remote sensing image carry out road class and non-rice habitats class coarse extraction, then to road class into
Row non-rice habitats noise removal, finally using the form of binary morphology operation repairing road primitive.
S2 is connected to road skeleton: connecting neighbouring road primitives, forms complete road skeleton.
S3 generates candidate crossing coordinate: carrying out Refinement operation to the road skeleton of extraction, is figured using Rutoviz intersection
Method obtains road skeleton crosspoint, and as candidate crossing coordinate.(point methods are intersected by above-mentioned global detection skeleton, can be obtained
To candidate crossing center point coordinate.But the local nonlinearity deformation as caused by thinning algorithm generates the non-crossing crosspoint in part,
In the case where guaranteeing recall ratio, the crosspoint at increased non-crossing need to be filtered out by part detection.)
S4 constructs crossing index similarity, specific steps are as follows:
S4.1 constructs piecemeal rectangle template.The present invention devises a kind of novel piecemeal rectangle rotating model identification crossing,
For the model using piecemeal rectangle as template, template is divided into center and the part of frontier district two, and center counts the letter at crossing center
Breath, marginal zone count the road section information being connected with crossing center.The length L of piecemeal rectangle template center1It is set as road
Mean breadth, boundary section length L2It is set as K*L1, K is threshold value constant.
S4.2, mean value angle.
S4.2.1, coherent filtering enhance image edge.
Coherence enhancing diffusion filtering (hereinafter referred to as coherent filtering) be one is improve image border respectively to
The effective ways of anisotropic non-linear tensor diffusion.Linear character is that the important feature of road is also the weight that road distinguishes other atural objects
Will foundation, to crossing image implement coherent filtering not only can smooth image, but also the contrast and road at edge can be enhanced
Directional information.Coherent filtering is carried out to image to be processed, filtered image road linear edge is enhanced, in road surface
Pixel difference be greatly improved.
S4.2.2 calculates mean value angle.
Enhancing image edge is filtered using coherence enhancing diffusion, counts each item in each piecemeal rectangle template frontier district
The angular distribution situation at edge.In the window W of M*N, the edge for having M angle different, each edge is remembered respectively with X-axis angle
Make: θ1,θ2,...,θM, wherein θi∈[0,π).It is A that current rectangle template i, which rotates angle and the angle of X-axis,i, i in the present invention
Value range is [1,36].Current rectangle template i angle mean value SIIt is defined as follows:
Wherein τ is threshold value constant, AiIt calculates as follows:
S4.3, spatial autocorrelation express homogeney.
Texture is the important feature of remote sensing image, the resolution of the areal image usually obtained from dissimilar sensor
Rate and texture characteristics are not quite similar, and the homogeneous feature of road is relatively stable in image.Spatial autocorrelation indicators
Moran ' I can rationally characterize the homogeneous feature (one of important feature of road) of image texture.Compared to other types of texture
Feature (such as variance, energy, entropy, contrast, the degree of correlation), Moran ' the I index calculating of characterization homogeneity characteristic is simple, the time is multiple
Miscellaneous degree is low, is influenced by sensor type smaller.The present invention is using Moran ' the I Mean value of index of rectangle Statistical Area as local grain spy
Sign, specific calculating are as follows:
Wherein I (i, j) indicates Moran ' the I index value of pixel (i, j), and Ω is statistical regions, and N is picture in statistical regions
The sum of element.
S4.4, color away from.
Color is the most intuitive visual signature of image, and color moment is that one kind of descriptive statistics field color distribution situation is succinct
Effective method.Have that calculating is simple, construction color feature vector dimension is small, table using the color characteristic of color moment expression piecemeal
Up to clear succinct advantage.In the present invention, the color feature vector of piecemeal rectangle template includes nine components, may be expressed as: fi
=[μR,σR,SR,μG,σG,SG,μB,σB,SB], it is assumed that characteristic component meets Gaussian Profile, normalization color characteristic component to [- 1,
1] section.
S4.5 constructs crossing index similarity.
This patent counts crossing index similarity (Intersection Similarity using piecemeal rectangle rotating model
Index, ISI), by ISI, locally detection and identification crossing, ISI are defined as follows:
In formulaIndicate the normalization average value angle factor of the i-th piecemeal rectangle template frontier district,Indicate the i-th piecemeal square
Shape form boundary area and the normalized color moment Euclidean distance in center,Indicate the i-th piecemeal rectangle template frontier district and center
Normalized Moran ' the I Mean value of index in area is poor, w1、w2And w3The respectively weight of above three component.
S5, piecemeal rectangle rotating model identify crossing, and the road width for being generally proximal to crossing can gradually broaden, and crossing is presented
The shape feature of similar round.Meanwhile existing at crossing compared with stable image feature, specific manifestation are as follows: (1) crossing center and therewith
Connected road has similar spectrum, textural characteristics;(2) the section internal edge connecting with crossing center is evenly distributed,
Edge has very strong directionality, and the direction that the direction at edge and road extend is almost the same.In view of above-mentioned image road
Mouthful important feature, the invention proposes a kind of novel piecemeal rectangle rotating model identification crossing specifically includes the following steps:
S5.1 counts ISI using piecemeal rectangle rotating model;Using candidate crossing coordinate as the center of circle, with fixed angle interval
It rotates piecemeal rectangle template and counts ISI, using angle as abscissa, ISI is ordinate, forms ISI figure.
S5.2, during locally identification, some crossings will appear multiple place's approximation ISI point situations, can simply merge
Similitude in ISI figure.ISI figure after merging similitude eliminates partial redundance point and retains important local maximum or minimum letter
Breath, Wave crest and wave trough are more obvious.
S5.3, counts the trough point quantity of ISI figure, and angle corresponding to trough indicates the side of the connected road component in crossing
To the trough number n of extraction is to judge crossing important indicator.
When trough number n=>5 or n<2 judge that candidate point is not crossing.
When trough number n belongs to [2,4], also need through the Distance Judgment between analysis trough, criterion is as follows:
1) work as n=2, corresponding two angles of trough, respectively θiAnd θj, and θi<θj.(θ if it existsj–θi) belong to [π/3,
5 π/6] or [7 π/6,5 π/3], then determine that the candidate point for road turn in the road, can mark when necessary in road-map;
2) work as n=3, corresponding 3 trough angle, θsi, θj, θkThere are a pair of of angular distances to belong to [17 π/18,19 π/18],
Then it is determined as T-shape crossing;Otherwise, it is determined that being Y-shaped crossing;
3) work as n=4, be determined as " four troubles " crossing, corresponding four angles of trough, respectively θi、θj、θmAnd θn, and θi<θj
<θm<θn.In the presence of (θm-θi) belong to [π/18 17 π/18,19] and (θn-θj) belong to [17 π/18,19 π/18], then it is determined as " ten "
Shape crossing.
Further, step S1 the following steps are included:
S1.1, road primitives coarse extraction:
S1.1.1, feature extraction, using NSCT (Non-Subsampled Contourlet Transform, no down-sampling
Little profile transformation) image is successively decomposed, texture is constructed using characteristic parameters such as the mean value of low frequency sub-band, standard deviation, homogeneity degree
Feature vector F1 and band logical subband gradient energy, sub-band coefficients variance construct texture feature vector F2, and utilize color histogram
Figure calculates the characteristic parameters such as standard deviation, mean value, gradient, kurtosis value and energy and constructs color feature vector F3.
The study of S1.1.2, multicore SVM (Support Vector Machine, support vector machines) classifier, selects Gauss
Core Radial basis kernel function, the road sample and about 10% non-rice habitats sample for taking in image about 8% respectively carry out classification learning, obtain
Two discriminant classification function of multicore SVM.
S1.1.3, road primitives coarse extraction, image after scanning whole picture pretreatment as center windowing using pixel utilize
The discriminant classification function that S1.1.2 is obtained judges center pel, is divided into two class of road and non-rice habitats, realizes road primitives coarse extraction.
S1.2, Road Base primordial essence are extracted:
S1.2.1, road in the real world are the reticular structure being formed by connecting by different road primitives, usual roadway area
The area in domain will not be too small, and small patches can be considered as to non-rice habitats noise removal, in conjunction with the long and narrow property and compactness of road shape, leads to
It crosses threshold value and filters out non-long and narrow block.
S1.2.2, using the form of binary morphology operation repairing road primitive.
The invention adopts the above technical scheme.The present invention is not necessarily to manual intervention, can automatically extract crossing, extraction it is accurate
Property and integrality are superior to existing similar research achievement, and main advantage is as follows: (1) present invention can obtain candidate crossing automatically
Point initializes road-center seed point without manual intervention mode.(2) the effective part for remaining crossing of the present invention is thin
Information is saved, erroneous detection or the missing inspection of road primitives are avoided.(3) present invention takes full advantage of the spectral information at image crossing, edge letter
Breath and texture information.(4) present invention proposes that piecemeal rectangle rotating model makes full use of the stable image feature at crossing: (a) leading to
It can gradually broaden very close to the road width at crossing, the shape feature of similar round is presented in crossing;(b) crossing center and therewith phase
Road even has similar spectrum, textural characteristics;(c) the section internal edge connecting with crossing center is evenly distributed, side
Edge has very strong directionality, and the direction that the direction at edge and road extend is almost the same.
Detailed description of the invention
The present invention is described in further details below in conjunction with the drawings and specific embodiments:
Fig. 1 is a kind of flow diagram of high-resolution remote sensing image crossing of the invention automatic identifying method;
Fig. 2 is the piecemeal rectangle template schematic diagram of step S4.1 of the invention;
Fig. 3 is the edge statistics window W schematic diagram of step S4.2.2 of the invention;
Fig. 4 is the piecemeal rectangle rotating model structural schematic diagram of step S5.1 of the invention.
Specific embodiment
Shown in one of picture 1-4, the invention discloses a kind of high-resolution remote sensing image crossing automatic identifying method, packets
Include following steps:
S1, road primitives extract: first to remote sensing image carry out road class and non-rice habitats class coarse extraction, then to road class into
Row non-rice habitats noise removal, finally using the form of binary morphology operation repairing road primitive.
Specifically, comprising the following steps:
S1.1, road primitives coarse extraction, specific steps are as follows:
Feature extraction: S1.1.1 uses NSCT (Non-Subsampled Contourlet Transform, no down-sampling
Little profile transformation) image is successively decomposed, texture is constructed using characteristic parameters such as the mean value of low frequency sub-band, standard deviation, homogeneity degree
Feature vector F1 and band logical subband gradient energy, sub-band coefficients variance construct texture feature vector F2, and utilize color histogram
Figure calculates the characteristic parameters such as standard deviation, mean value, gradient, kurtosis value and energy and constructs color feature vector F3.
The study of S1.1.2, multicore SVM (Support Vector Machine, support vector machines) classifier, selects Gauss
Core Radial basis kernel function, the road sample and about 10% non-rice habitats sample for taking in image about 8% respectively carry out classification learning, obtain
Two discriminant classification function of multicore SVM.
S1.1.3, road primitives coarse extraction, image after scanning whole picture pretreatment as center windowing using pixel utilize
The discriminant classification function that S1.1.2 is obtained judges center pel, is divided into two class of road and non-rice habitats, realizes road primitives coarse extraction.
S1.2, Road Base primordial essence are extracted, specific steps are as follows:
S1.2.1, road in the real world are the reticular structure being formed by connecting by different road primitives, usual roadway area
The area in domain will not be too small, and small patches can be considered as to non-rice habitats noise removal;In conjunction with the long and narrow property and compactness of road shape, lead to
It crosses threshold value and filters out non-long and narrow block.
S1.2.2, using the form of binary morphology operation repairing road primitive.
S2 is connected to road skeleton: connecting neighbouring road primitives, forms complete road skeleton.
S3 generates candidate crossing coordinate: carrying out Refinement operation to the road skeleton of extraction, is figured using Rutoviz intersection
Method obtains road skeleton crosspoint, and as candidate crossing coordinate;Intersect point methods by above-mentioned global detection skeleton, can be obtained
Candidate crossing center point coordinate.But the local nonlinearity deformation as caused by thinning algorithm generates the non-crossing crosspoint in part,
In the case where guaranteeing recall ratio, the crosspoint at increased non-crossing need to be filtered out by part detection.
S4 constructs crossing index similarity:
S4.1 constructs piecemeal rectangle template, devises a kind of novel piecemeal rectangle rotating model identification crossing, the model
Using piecemeal rectangle as template;As shown in Fig. 2, piecemeal rectangle template is divided into center and the part of frontier district two, center statistics
The information at crossing center, frontier district count the road section information being connected with crossing center;The length L of piecemeal rectangle template center1
It is set as the mean breadth of road, boundary section length L2It is set as K*L1, K is threshold value constant.
S4.2 calculates the mean value angle of piecemeal rectangle template.
S4.2.1 carries out coherent filtering to image to be processed, enhances image road linear edge, and coherent filtering is a kind of
It is the effective ways for improving the diffusion of image border Anisotropic Nonlinear tensor.Linear character is that the important feature of road is also
The important evidence of other atural objects is distinguished on road, to crossing image implement coherent filtering not only can smooth image, but also can enhance
The contrast at edge and the directional information of road.
S4.2.2 calculates mean value angle: using coherence enhancing diffusion filtering enhancing image edge, counting each piecemeal
The angular distribution situation at each of rectangle template frontier district edge;As shown in figure 3, in the window W of M*N, the side that has M angle different
Edge, each edge are denoted as respectively with X-axis angle: θ1,θ2,...,θM, wherein θi∈[0,π);Current piecemeal rectangle template i rotation
The angle of angle and X-axis is Ai, wherein the value range of i is [1,36];Current piecemeal rectangle template i angle mean value SIDefinition is such as
Under:
Wherein τ is threshold value constant, AiIt calculates as follows:
S4.3, spatial autocorrelation express homogeney: using Moran ' the I Mean value of index of rectangle Statistical Area as local grain spy
Sign, specific calculating are as follows:
Wherein I (i, j) indicates Moran ' the I index value of pixel (i, j), and Ω is statistical regions, and N is picture in statistical regions
The sum of element.
Specifically, texture is the important feature of remote sensing image, the areal usually obtained from dissimilar sensor
The resolution ratio and texture characteristics of image are not quite similar, and the homogeneous feature of road is relatively stable in image.Space is from phase
The homogeneous feature (one of important feature of road) of image texture can rationally be characterized by closing index M oran ' I.Compared to other types
Textural characteristics (such as variance, energy, entropy, contrast, the degree of correlation), Moran ' the I index of characterization homogeneity characteristic calculate it is simple,
Time complexity is low, is influenced by sensor type smaller.
S4.4, color is away from processing: color is the most intuitive visual signature of image, and color moment is descriptive statistics field color point
A kind of succinct effective method of cloth situation.Have using the color characteristic of color moment expression piecemeal and calculates simple, construction color
Feature vector dimension is small, the clear succinct advantage of expression.In the present invention, the color feature vector of piecemeal rectangle template includes nine
Component may be expressed as: fi=[μR,σR,SR,μG,σG,SG,μB,σB,SB], it is assumed that characteristic component meets Gaussian Profile, normalizes face
Color characteristic component is to [- 1,1] section.
S4.5 constructs crossing index similarity: counting crossing index similarity using piecemeal rectangle rotating model
(Intersection Similarity Index, ISI) is locally detected and is identified crossing, crossing index similarity by ISI
The calculation formula of ISI is as follows:
In formulaIndicate the normalization average value angle factor of the i-th piecemeal rectangle template frontier district,Indicate the i-th piecemeal square
Shape form boundary area and the normalized color moment Euclidean distance in center,Indicate the i-th piecemeal rectangle template frontier district and center
Normalized Moran ' the I Mean value of index in area is poor, w1、w2And w3The respectively weight of above three component.
S5 identifies crossing based on piecemeal rectangle rotating model, and the road width for being generally proximal to crossing can gradually broaden, crossing
The shape feature of similar round is presented.Meanwhile at crossing exist compared with stable image feature, specific manifestation are as follows: (1) crossing center and
The road being attached thereto has similar spectrum, textural characteristics;(2) the section internal edge distribution connecting with crossing center is equal
Even, edge has very strong directionality, and the direction that the direction at edge and road extend is almost the same.In view of above-mentioned image
The important feature at crossing, S5 specifically includes the following steps:
S5.1, as shown in figure 4, counting ISI using piecemeal rectangle rotating model;Using candidate crossing coordinate as the center of circle, with solid
Determine angle interval rotation piecemeal rectangle template statistics ISI, and using angle as abscissa, ISI is ordinate, forms ISI figure.
S5.2, during locally identification, during locally identification, it is approximate that some crossings will appear multiple places
ISI point situation, can simply merge similitude in ISI figure, and eliminates the partial redundance point of the ISI figure after merging similitude and protect
Stay important local maximum or the prominent Wave crest and wave trough of minimum information.
S5.3, counts the trough point quantity of ISI figure, and angle corresponding to trough indicates the side of the connected road component in crossing
To, and the trough number n by extracting judges crossing:
When trough number n=>5 or n<2, then judge that candidate point is not crossing.
When trough number n belongs to [2,4], classification judgement is carried out by the distance between analysis trough:
1) work as n=2, corresponding two angles of trough, respectively θiAnd θj, and θi<θj;When in the presence of (θj–θi) belong to [π/3,
5 π/6] or when belonging to [7 π/6,5 π/3], then determine the candidate point for road turn in the road;
2) work as n=3, corresponding 3 trough angle, θsi、θj、θk;When there are any pair of angular distances to belong to [17 π/18,19
π/18] when, then it is determined as T-shape crossing;Otherwise, it is determined that being Y-shaped crossing;
3) work as n=4, corresponding four angles of trough, respectively θi、θj、θmAnd θn, and θi<θj<θm<θn;When in the presence of (θm-
θi) belong to [π/18 17 π/18,19] and (θn-θj) when belonging to [17 π/18,19 π/18], then it is determined as " ten " shape crossing.
The invention adopts the above technical scheme.The present invention is not necessarily to manual intervention, can automatically extract crossing, extraction it is accurate
Property and integrality are superior to existing similar research achievement, and main advantage is as follows: (1) present invention can obtain candidate crossing automatically
Point initializes road-center seed point without manual intervention mode.(2) the effective part for remaining crossing of the present invention is thin
Information is saved, erroneous detection or the missing inspection of road primitives are avoided.(3) present invention takes full advantage of the spectral information at image crossing, edge letter
Breath and texture information.(4) present invention proposes that piecemeal rectangle rotating model makes full use of the stable image feature at crossing: (a) leading to
It can gradually broaden very close to the road width at crossing, the shape feature of similar round is presented in crossing;(b) crossing center and therewith phase
Road even has similar spectrum, textural characteristics;(c) the section internal edge connecting with crossing center is evenly distributed, side
Edge has very strong directionality, and the direction that the direction at edge and road extend is almost the same.