CN109242759A - Figure based on Density Clustering shrinks grouping method for registering - Google Patents
Figure based on Density Clustering shrinks grouping method for registering Download PDFInfo
- Publication number
- CN109242759A CN109242759A CN201810777473.0A CN201810777473A CN109242759A CN 109242759 A CN109242759 A CN 109242759A CN 201810777473 A CN201810777473 A CN 201810777473A CN 109242759 A CN109242759 A CN 109242759A
- Authority
- CN
- China
- Prior art keywords
- image
- group
- registration
- global center
- density clustering
- 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
- 238000000034 method Methods 0.000 title claims abstract description 53
- 239000011159 matrix material Substances 0.000 claims abstract description 30
- PXFBZOLANLWPMH-UHFFFAOYSA-N 16-Epiaffinine Natural products C1C(C2=CC=CC=C2N2)=C2C(=O)CC2C(=CC)CN(C)C1C2CO PXFBZOLANLWPMH-UHFFFAOYSA-N 0.000 claims abstract description 10
- 238000010586 diagram Methods 0.000 claims abstract description 10
- 230000008602 contraction Effects 0.000 claims abstract description 7
- 238000012935 Averaging Methods 0.000 claims description 4
- 230000008901 benefit Effects 0.000 claims description 4
- 239000000203 mixture Substances 0.000 claims description 4
- 238000003786 synthesis reaction Methods 0.000 claims description 4
- 238000010276 construction Methods 0.000 claims description 2
- 238000012545 processing Methods 0.000 abstract description 5
- 210000004556 brain Anatomy 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 208000014644 Brain disease Diseases 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 230000005856 abnormality Effects 0.000 description 1
- HUTDUHSNJYTCAR-UHFFFAOYSA-N ancymidol Chemical compound C1=CC(OC)=CC=C1C(O)(C=1C=NC=NC=1)C1CC1 HUTDUHSNJYTCAR-UHFFFAOYSA-N 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 239000012141 concentrate Substances 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 230000002950 deficient Effects 0.000 description 1
- 239000003814 drug Substances 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000008030 elimination Effects 0.000 description 1
- 238000003379 elimination reaction Methods 0.000 description 1
- 244000144992 flock Species 0.000 description 1
- 238000003384 imaging method Methods 0.000 description 1
- 238000002595 magnetic resonance imaging Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 230000013011 mating Effects 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/14—Transformations for image registration, e.g. adjusting or mapping for alignment of images
- G06T3/147—Transformations for image registration, e.g. adjusting or mapping for alignment of images using affine transformations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
- G06F18/232—Non-hierarchical techniques
- G06F18/2321—Non-hierarchical techniques using statistics or function optimisation, e.g. modelling of probability density functions
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Probability & Statistics with Applications (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of, and the figure based on Density Clustering shrinks grouping method for registering.Steps are as follows by the present invention: step 1 carries out affine registration;Step 2 obtains Global center image I according to formulaO;The improved Density Clustering DBSCAN algorithm of data acquisition system I is clustered, Ω group G is divided intoαElement e in the symmetrical connection matrix E of one step 3, building N × N, symmetrical connection matrix Ei,j=0,j∈{1,…,N};Step 5 is registrated by continuous iteration diagram contraction process, until convergence;The present invention reduces the computation complexity for the grouping registration Algorithm that classification unbiased graph is shunk;It sets that two parametric techniques are relevant to Global center image, so that the figure contraction process carried out later is more easier, reduces registration the number of iterations.The present invention, which is very suitable to processing, has big and labyrinth variation data set, improves registration accuracy to a certain extent.
Description
Technical field
The invention belongs to 3 d medical images to be registrated field, be applied to large-scale medicine brain image and be registrated, specifically come
Say it is the grouping registration Algorithm for being classified unbiased graph and shrinking based on Density Clustering.
Technical background
Since magnetic resonance imaging (MRI) occurs, many researchs based on imaging are had begun working in group, group
Between and the same group different time structure change.In these researchs, image registration is it is critical that elimination group is poor
It is different, these differences include individual directly by difference or listen an experimenter structure variation change (such as with brain diseases phase
The variation of pass).In particular, delicate variation is likely difficult to distinguish if the execution of image registration is not accurate enough.
There are many pairs of method for registering at present is registrated to template image for one group of image.But stencil-chosen is not
It is an inappreciable task, if being made inappropriate, it will influence subsequent statistical analysis.In order to solve this problem,
Grouping method for registering has been proposed recently to snap to all images on public space jointly, without clearly specified mould
Plate.Objective function in grouping registration is intended to minimize the entropy for combining intensity profile on overall intensity difference or all images.For
Extensive objective function in optimization group registration, proposes the Gauss-Newton optimization algorithm based on effective gradient.Pass through choosing
It selects the key point in image and key point is only allowed to drive entire group's registration, wherein also using layering grouping registration principle.
Many existing grouping method for registering need clearly defined registration target.Such as a kind of effective group's registration is calculated
Method, the algorithm alternately do following two step always:
(1) by all images averagely obtained one initialization registration template, then every image respectively with template
It is registrated two-by-two.
(2) image that all previous steps have been registrated is averaged, obtains new registration template.This group's method for registering
The data set smaller suitable for structure change.
However, and structure change larger for data set is very big, this method there may be a fuzzy class mean image,
To reduce registration accuracy.In order to solve this problem, it is thus proposed that group's method for registering based on sharp-mean, this method
Block-based weighted average method is used.However, these methods it is common limitation be, no matter the structure and group of individual images
Whether the average image has significant difference, requires to be registrated a group the average image.Therefore, these methods processing have it is big and
It is restricted when the data set of labyrinth variation.
On the other hand, the precision of group's registration can be improved using image distribution information.By by the group of pictures in group
Several subgroups are integrated, each subgroup is then registrated by flock mating standard, improved registration accuracy can be obtained.Due in subgroup
Picture appearance is similar, therefore can relatively easily obtain accurate registration.After being registrated in subgroup, the representativeness of each subgroup
Image can be used between subgroup being registrated.However, one of this method the disadvantage is that in subgroup distributed intelligence be not used exclusively for group
Registration.
A method of being known as HUGS (layering unbiased graph is shunk).In HUGS, entire data are characterized by using figure
The image distribution of collection, wherein node table diagram picture, and the similarity between its side representative image.It is only connected on chart similar
Image.All image registrations to a hiding public space are expressed as motion graphics and shrink problem, wherein all nodes
It is close to each other along edge.Since global and local image distribution information obtains in figure, so whole image manifold is opened up
Flutterring structure can more accurately retain in entire registration process.Although more accurate, HUGS is handling different types of number
It is still limited according to collection aspect, because single simple graph is often inadequate in terms of complicated image is distributed modeling.HUGS uses simple
Based on the method for threshold value come structure figures, i.e., given threshold value is less than by the distance of two images of connection, gives a heterogeneous datasets,
Threshold value must relax significantly, and to ensure all image connections on the diagram, this often leads to excessively connect and the lower figure of efficiency.This
Also result in many unnecessary registrations between different images pair, this not only can significant increase calculate the time, but also will lead to
Sizable registration error.
Existing group's registration HUGS method is described as follows:
Assuming that the deformation of single image is the dynamic process changed with time variable t.Ii(t) change in time t is indicated
Shape image Ii.Then as follows using figure defined in brain image manifold.Assuming thatIt is node of graph, E=
{ei,j: i, j=1 ..., N } it is side in figure between two nodes.ei,j=1 indicates Ii(t) and Ij(t) there is connection between, it is no
Then, I in figurei(t) and Ij(t) there is no directly contact between.In addition, N × N weighted adjacent matrix is defined, wherein each member
Plain exp (vi,j(t)) the geodesic curve path between two images is described, if ei,j=1 and vi,j(t)=∞ then vi,j(t)>0。
vi,j(t) velocity vector in geodetic path is indicated, it indicates Ii(t) and IjThe distance between (t).Make deformable figure
As registration is to estimate each velocity vector vi,j(t), and pass through exp (vi,j(t)) I is calculatedi(t) and Ij(t) the deformation road between
Diameter, wherein ' exp' is index mapping.The target of group's registration based on figure is exactly the velocity vector minimized on all figure sides,
It is as defined below:
The principle of F (t) is as shown in Figure 3.First, it is assumed that all images are located in manifold of higher dimension.Then, image distribution
Topological structure is described with figure, and wherein side indicates the part connection between node.Velocity vector vi,jIt is (t) relevant with each edge,
Wherein along vi,j(t) integral is exactly from Ii(t) I is arrivedj(t) geodesic distance.
The minimum of F (t) can be counted as Dynamic Graph and reduce process, this process is by each image from Ii(t)
It is deformed into Ii(t+ Δ t) also keeps the topology of entire figure while reducing overall geodesic distance.With the increase of time t, own
Ii(t) it is finally registrated to center image, and reasonably determines velocity vector vi,j(t) and incremental time Δ t.
As figure is dynamically shunk, group's registration more crucially determines each image Ii(t) time t how
Deformation, so that energy function F (t) is minimized.Assuming that each image is from Ii(t0) become Ii(tk), wherein { tkIt is time t (k=
0,…,K,t0=0, tk→ ∞) discretization.According to its connecting node by Ii(tk) moved along mean direction, so each in figure
Node Ii(tk) local connectivity be reasonable.
Since velocity vector is located at the I in manifoldi(tk) tangent line spatially, calculated by linear averagingWhereinIt is Ii(tk) connection quantity.To being scheduled on time point tk
Each velocity vector direction, if each node Ii(tk) along the direction of velocity vector from tkIt is moved to tk+Δtk, whole
Then strictly monotone is reduced a energy function F (t), wherein incremental time Δ tkIt is obtained by following formula:
Finally, by connecting from t0To tkDeformation section obtain each image IiTo the geodesic curve path at centerI.e.WhereinIndicate deformation composition.In the figure systole phase
Between, all images are gradually deformed to hiding public space, and the topology of reserved graph.T is used based on interval of delta tk(k=0,1,
2 ..., wherein t0=the 0 and t as k → ∞k→ ∞) carry out discretization t.It usesAs time tkStylish deformation pattern, willIt is expressed as time tkLocate imageSynthesis deformation field.
A kind of novel group's method for registering (eHUGS), it has a hierarchical diagram, and each node on behalf of the figure is individually schemed
Picture.Hierarchical diagram captures image distribution manifold to obtain image distribution information, so as to utilize image distribution information.More specifically
It says, the figure of low level describes the image distribution in each group, and high-level figure represents figure representative in subgroup
Relationship as between.In a specific chart representation model, we can by the figure in dynamic downscaled images manifold by
All image registrations are to public space.The topology of whole image distribution is always maintained at constant during scheming to shrink.But this side
Method or defective, eHUGS method use neighbour's propagation clustering method (AP cluster), and this method needs that point of reference is arranged in advance,
And the size of point of reference and the number of cluster centre are positively correlated.Since each iteration of AP algorithm requires to update each data point
Attract angle value and ownership angle value, algorithm complexity is higher, and runing time is longer under big data quantity.Moreover, eHUGS method
A Global center image is chosen, representative image in subgraph, the selection pair of this global image are then chosen according to this image
Entire registration process has great effect, largely influences final registration result.
Summary of the invention
The technical problems to be solved by the present invention are: the figure based on cluster shrinks grouping registration Algorithm, group's registration is improved
Precision and the speed of service.
The present invention is based on the figures of cluster to shrink grouping registration Algorithm, using improved Density Clustering DBSCAN algorithm, favorably
There is big and labyrinth variation data set in processing, improve registration accuracy, the complexity of certain algorithm is declined, and is added
Quasi velosity is matched by fast group.
The technical solution adopted by the present invention to solve the technical problems includes the following steps:
Step 1 carries out affine registration, obtains sets of image data I, the I={ I of N affine alignmenti| i=1 ..., N };
All distances are calculated with SSD Obtain Distance matrix D;
Step 2, according to formulaObtain Global center image IO;By data acquisition system I
It is clustered with improved Density Clustering DBSCAN algorithm, is divided into Ω group Gα=(α ∈ 1 ..., Ω } and
Element in the symmetrical connection matrix E of one step 3, building N × N, symmetrical connection matrix EIf image IiWith image IjIt is connection, the element e of symmetrical connection matrix E is seti,j=
1 otherwise ei,j=0;All diagonal entries in symmetrical connection matrix E are both configured to 0;Determine each group GαIn near complete
Office center image IoImage, and using the image as presentation graphics
Step 4, building group GαInterior linked system;
A threshold value is set, with presentation graphicsThe higher image I of similitudei{Ii∈Gα,i≠iαJust there is connection, it is no
Then all it is set as 0;Linked system between building group, each presentation graphics
All there is connection with this Global center image, it is connectionless between any two representative image, it is set as 0;
Step 5 is registrated by continuous iteration diagram contraction process, until convergence;The mesh of group's registration based on figure
Mark is exactly the velocity vector minimized on all figure sides, as defined below:F(t)
Minimum be counted as Dynamic Graph and reduce process, this process is by each image from Ii(t) it is deformed into Ii(t+ Δ t), subtracts
Also the topology of entire figure is kept while few totality geodesic distance;It is all with the increase of time t with the increase of time t
Ii(t) finally it is registrated to center image.
Density clustering algorithm DBSCAN described in step 2 is to be based on one group of " neighborhood " parameter (∈, MinPts) to portray sample
The tightness degree of distribution, the two parameters need user to be arranged in advance, remain unchanged during continuous iteration, therefore DBSCAN
Algorithm can not processing variation density sample.And the selection of the two parameters influences cluster result largely to influence most
Whole registration result, therefore using a kind of two parametric techniques of setting relevant to Global center image, such that figure is more
Add easy contraction, reduces registration the number of iterations.The improved Density Clustering DBSCAN algorithm description is as follows:
2-1. is determined in data set closest to the Global center image I of hiding public spaceO;
For the data set I={ I of N image constructionsi| i=1 ..., N }, apply affine registration by all images first
A public space is snapped to, N × N similar matrix S, the element s of similar matrix S are subsequently filledijIt is defined the SSD that is negative, i.e. sij
=dij=| | Ii-Ij||2;
2-2. determines radius of neighbourhood ∈;
Itself similar value of N × N similar matrix S is removed, that is, by diiRemove, is left N × (N-1)/2 number, it is first
First, preceding Num is obtained with KD tree nearest neighbor algorithmroughA minimum data, Numrough=k*N × (N-1)/2;
Parameter beta is introduced to limit radius of neighbourhood ∈ to obtain k value:
∈roughIt is the preliminary radius of neighbourhood, disiIndicate the i-th small d of sequenceij,
Num∈MeetThe number of o ≠ j, final radius of neighbourhood ∈ are as follows:
2-3. determines the image number MinPts that representative image sample is included in radius of neighbourhood ∈,
NumPts(o, j) indicates Global center image IoWith image IjDistance whether in the ∈ neighborhood of initial point
A parameter alpha is set to adjust MinPts, 0 < α≤1
2-4. is by Global center figure IOAs seed, all representative image samples reachable by its density are found out, represent figure
Decent constitutes current group, removes the point that current group is included, group number c=c+1 from data set I;
2-5. repeats step 2-1 to step 2-4, until data set I is sky.
Step 4 is implemented as follows:
4-1. is determined in data set closest to the Global center image I of hiding public spaceO;To N image I={ Ii
| i=1 ..., N } it is clustered, it is Ω group G by using the improved Density Clustering DBSCAN algorithm described in step 2α
=(α ∈ 1 ..., Ω } and
4-2. is in each group GαInterior, it is exactly Global center image I that first, which clusters obtained kernel object,O;Ω group
Just there are Ω presentation graphicsTherefore there is Global center image
The presentation graphics in group are selected according to Global center image, the benefit done so can exactly make other figures in group
As can more easily be deformed towards Global center;For remaining Ω -1 groups, each group is selected according to similar matrix S
In each point and Global center image IOPresentation graphics apart from the smallest point, as the group;
4-3. is for each group Gα, presentation graphics IiαConnection in group is carried out between every other image;For appointing
Anticipate a pair of of image IiWith image Ij∈Gα, wherein j ≠ i, if i=iαOr j=iαJust there is ei,j=1, otherwise ei,j=0;
4-4. is for the advanced connection between group, Global center image IOConnect every other presentation graphicsIt is right
In any pair of image IiWithFor, if i=o or j=o, there is ei,j=1, otherwise
ei,j=0.
Step 5 is implemented as follows:
As figure is dynamically shunk, group's registration more crucially determines each image Ii(t) time t how
Deformation, so that energy function F (t) is minimized;Assuming that each image is from Ii(t0) become Ii(tk), wherein { tkIt is time t (k=
0,…,K,t0=0, tk→ ∞) discretization;According to its connecting node by Ii(tk) moved along mean direction, so each in figure
Node Ii(tk) local connectivity be reasonable;
Since velocity vector is located at the I in manifoldi(tk) tangent line spatially, calculated by linear averagingWhereinIt is Ii(tk) connection quantity;To being scheduled on time point tk
Each velocity vector direction, if each node Ii(tk) along the direction of velocity vector from tkIt is moved to tk+Δtk, whole
Then strictly monotone is reduced a energy function F (t), wherein incremental time Δ tkIt is obtained by following formula:
Finally, by connecting from t0To tkDeformation section obtain each image IiTo the geodesic curve path at centerI.e.WhereinIndicate deformation composition;In the figure systole phase
Between, all images are gradually deformed to hiding public space, and the topology of reserved graph;T is used based on interval of delta tk(k=0,1,
2 ..., wherein t0=the 0 and t as k → ∞k→ ∞) carry out discretization t;It usesAs time tkStylish deformation pattern, willIt is expressed as time tkLocate imageSynthesis deformation field.
The method of the present invention has the advantage that and beneficial outcomes are as follows:
(1) computation complexity of neighbour's propagation clustering algorithm is O (N3), and the calculating of density clustering algorithm DBSCAN is complicated
Degree is O (N2), using improved DBSCAN method, computation complexity can be reduced to O by the KD tree established when searching for arest neighbors
(NlogN).It is possible thereby to reduce the computation complexity for the grouping registration Algorithm that classification unbiased graph is shunk;
(2) the parameter preference value of neighbour's propagation algorithm setting is unrelated with Global center figure, largely influences cluster result
To influence final registration result.Density clustering algorithm investigates the connectivity between sample from the angle of sample rate,
It is not very big that two parameters in improved DBSCAN algorithm, which influence cluster result, it is most important that, set two parametric techniques
It is relevant to Global center image, so that the figure contraction process carried out later is more easier, reduce registration the number of iterations.
(3) different types of data set is handled, since initial cluster result and Global center are image-related, and is improved
DBSCAN algorithm can cluster while note abnormalities a little, to data concentrate abnormal point it is insensitive, institute it is non-in this approach
Often it is suitble to processing that there is big and labyrinth variation data set, improves registration accuracy to a certain extent.
Detailed description of the invention
Fig. 1 is the schematic diagram of image of the present invention on manifold of higher dimension, the connection of geodesic curve path.
Fig. 2 is according to the obtained image clustering result figure of improved Density Clustering DBSCAN algorithm;
Fig. 3 is the schematic diagram for group's method for registering that the classification unbiased graph based on Density Clustering is shunk.
Specific embodiment
The present invention will be further explained below with reference to the attached drawings.
As shown in Figure 1-3, the figure based on Density Clustering shrinks grouping method for registering, the specific steps are as follows:
Step 1 carries out affine registration, obtains sets of image data I, the I={ I of N affine alignmenti| i=1 ..., N }.
All distances are calculated with SSD Obtain Distance matrix D;
Step 2, according to formulaObtain Global center image IO.By data acquisition system I
It is clustered with improved Density Clustering DBSCAN algorithm, is divided into Ω group Gα=(α ∈ 1 ..., Ω } and
The improved Density Clustering DBSCAN algorithm description is as follows:
2-1. is determined in data set closest to the Global center image I of hiding public spaceO.For N image structures
At data set I={ Ii| i=1 ..., N }, apply affine registration by all image alignments to a public space first, so
N × N similar matrix S, the element s of similar matrix S are filled afterwardsijIt is defined the SSD that is negative, i.e. sij=dij=| | Ii-Ij||2。
2-2. determines radius of neighbourhood ∈.
Itself similar value of N × N similar matrix S is removed, that is, by diiRemove, is left N × (N-1)/2 number, it is first
First, preceding Num is obtained with KD tree nearest neighbor algorithmroughA minimum data, Numrough=k*N × (N-1)/2;
Parameter beta is introduced to limit radius of neighbourhood ∈ to obtain k value:
∈roughIt is the preliminary radius of neighbourhood, disiIndicate the i-th small d of sequenceij,
Num∈Meet | | Io-Ij||2≤∈rough, the number of o ≠ j, final radius of neighbourhood ∈ is as follows:
2-3. determines image the number MinPts, Num that representative image sample is included in radius of neighbourhood ∈Pts(o, j) is indicated
Global center image IoWith image IjDistance whether in the ∈ neighborhood of initial point
A parameter alpha is set to adjust MinPts, 0 < α≤1
2-4. by Global center figure IOAs seed, all representative image samples reachable by its density are found out, represent figure
Decent constitutes current group, removes the point that current group is included, group number c=c+1 from data set I;
2-5. repeats step 2-1 to step 2-4, until data set I is sky.
Element in the symmetrical connection matrix E of one step 3, building N × N, symmetrical connection matrix EIf image IiWith image IjIt is connection, the element e of symmetrical connection matrix E is seti,j
=1 otherwise ei,j=0.All diagonal entries in symmetrical connection matrix E are both configured to 0;Determine each group GαIn near
Global center image IOImage, and using the image as presentation graphics
Step 4, building group GαInterior linked system;
A threshold value is set, with presentation graphicsThe higher image I of similitudei{Ii∈Gα,i≠iαJust there is connection, it is no
Then all it is set as 0.Linked system between building group, each presentation graphics
All there is connection with this Global center image, it is connectionless between any two representative image, it is set as 0;Specific building process is such as
Under:
4-1. is determined in data set closest to the Global center image I of hiding public spaceO.To N image I={ Ii
| i=1 ..., N } it is clustered, it is Ω group G by using the improved Density Clustering DBSCAN algorithm described in step 2α
=(α ∈ 1 ..., Ω } and
4-2. is in each group GαInterior, it is exactly Global center image I that first, which clusters obtained kernel object,o.Ω group
Just there are Ω presentation graphicsTherefore there is Global center image
The presentation graphics in group are selected according to Global center image, the benefit done so can exactly make other figures in group
As can more easily be deformed towards Global center.For remaining Ω -1 groups, each group is selected according to similar matrix S
In each point and Global center image IOPresentation graphics apart from the smallest point, as the group;
4-3. is for each group G α, presentation graphics IiαConnection in group is carried out between every other image.For appointing
Anticipate a pair of image Ii and image Ij∈Gα, wherein j ≠ i, if i=iαOr j=iαJust there is ei,j=1, otherwise ei,j=0.
4-4. is for the advanced connection between group, Global center image IOConnect every other presentation graphicsIt is right
In any pair of image IiWithFor, if i=o or j=o, there is ei,j=1, otherwise
ei,j=0.
Step 5 is registrated by continuous iteration diagram contraction process, until convergence.The mesh of group's registration based on figure
Mark is exactly the velocity vector minimized on all figure sides, as defined below:F(t)
Minimum be counted as Dynamic Graph and reduce process, this process is by each image from Ii(t) it is deformed into Ii(t+ Δ t), subtracts
Also the topology of entire figure is kept while few totality geodesic distance.It is all with the increase of time t with the increase of time t
Ii(t) it is finally registrated to center image, and reasonably determines velocity vector vi,j(t) and incremental time Δ t.Specific implementation is such as
Under:
As figure is dynamically shunk, group's registration more crucially determines each image Ii(t) time t how
Deformation, so that energy function F (t) is minimized.Assuming that each image is from Ii(t0) become Ii(tk), wherein { tkIt is time t (k=
0,…,K,t0=0, tk→ ∞) discretization.According to its connecting node by Ii(tk) moved along mean direction, so each in figure
Node Ii(tk) local connectivity be reasonable.
Since velocity vector is located at the I in manifoldi(tk) tangent line spatially, calculated by linear averagingWhereinIt is Ii(tk) connection quantity.To being scheduled on time point tk
Each velocity vector direction, if each node Ii(tk) along the direction of velocity vector from tkIt is moved to tk+Δtk, whole
Then strictly monotone is reduced a energy function F (t), wherein incremental time Δ tkIt is obtained by following formula:
Finally, by connecting from t0To tkDeformation section obtain each image IiTo the geodesic curve path at centerI.e.WhereinIndicate deformation composition.In the figure systole phase
Between, all images are gradually deformed to hiding public space, and the topology of reserved graph.T is used based on interval of delta tk(k=0,1,
2 ..., wherein t0=the 0 and t as k → ∞k→ ∞) carry out discretization t.It usesAs time tkStylish deformation pattern, willIt is expressed as time tkLocate imageSynthesis deformation field.
Claims (4)
1. the figure based on Density Clustering shrinks grouping method for registering, it is characterised in that include the following steps:
Step 1 carries out affine registration, obtains sets of image data I, the I={ I of N affine alignmenti| i=1 ..., N };Use SSD
Calculate all distances Obtain Distance matrix D;
Step 2, according to formulaObtain Global center image IO;By data acquisition system I with changing
Into Density Clustering DBSCAN algorithm cluster, be divided into Ω group Gα=(α ∈ { 1 ..., Ω } and);
Element in the symmetrical connection matrix E of one step 3, building N × N, symmetrical connection matrix EIf image IiWith image IjIt is connection, the element e of symmetrical connection matrix E is setI, j=
1 otherwise eI, j=0;All diagonal entries in symmetrical connection matrix E are both configured to 0;Determine each group GαIn near complete
Office center image IOImage, and using the image as presentation graphics
Step 4, building group GαInterior linked system;
A threshold value is set, with presentation graphicsThe higher image I of similitudei{Ii∈Gα, i ≠ iαJust there is connection, otherwise entirely
All it is set as 0;Linked system between building group, each presentation graphicsAll with
This Global center image has a connection, connectionless between any two representative image, is set as 0;
Step 5 is registrated by continuous iteration diagram contraction process, until convergence;The target of group's registration based on figure is just
It is the velocity vector minimized on all figure sides, as defined below:F (t) is most
Smallization is counted as Dynamic Graph and reduces process, this process is by each image from Ii(t) it is deformed into Ii(t+ Δ t) is reduced total
Also the topology of entire figure is kept while body geodesic distance;With the increase of time t, with the increase of time t, all Ii
(t) finally it is registrated to center image.
2. the figure according to claim 1 based on Density Clustering shrinks grouping method for registering, it is characterised in that described in step 2
Improved Density Clustering DBSCAN algorithm description it is as follows:
2-1. is determined in data set closest to the Global center image I of hiding public spaceO;
For the data set I={ I of N image constructionsi| i=1 ..., N }, all image alignments are arrived using affine registration first
One public space is subsequently filled N × N similar matrix S, the element s of similar matrix SijIt is defined the SSD that is negative, i.e. sij=dij=
||Ii-Ij||2;
2-2. determines radius of neighbourhood ∈;
Itself similar value of N × N similar matrix S is removed, that is, by diiRemove, is left N × (N-1)/2 number, firstly, with
KD tree nearest neighbor algorithm obtains preceding NumroughA minimum data, Numrough=k*N × (N-1)/2;
Parameter beta is introduced to limit radius of neighbourhood ∈ to obtain k value:
∈roughIt is the preliminary radius of neighbourhood, disiIndicate the i-th small d of sequenceij,
Num∈Meet | | Io-Ij||2≤∈rough, the number of o ≠ j, final radius of neighbourhood ∈ is as follows:
2-3. determines the image number MinPts that representative image sample is included in radius of neighbourhood ∈,
NumPts(o, j) indicates Global center image IoWith image IjDistance whether in the ∈ neighborhood of initial point
A parameter alpha is set to adjust MinPts, 0 α≤1 <
2-4. is by Global center figure IOAs seed, all representative image samples reachable by its density, representative image sample are found out
Current group is constituted, removes the point that current group is included, group number c=c+1 from data set I;1-5;
2-5. repeats step 2-1 to step 2-4, until data set I is sky.
3. the figure according to claim 2 based on Density Clustering shrinks grouping method for registering, it is characterised in that step 4 is specific
It is accomplished by
4-1. is determined in data set closest to the Global center image I of hiding public spaceO;To N image I={ Ii| i=
1 ..., N } it is clustered, it is Ω group G by using the improved Density Clustering DBSCAN algorithm described in step 2α=(α
∈ { 1 ..., Ω } and);
4-2. is in each group GαInterior, it is exactly Global center image I that first, which clusters obtained kernel object,O;Ω group just has
Ω presentation graphicsTherefore there is Global center imageAccording to
Global center image selects the presentation graphics in group, and the benefit done so can exactly enable other images in group
It is enough more easily to be deformed towards Global center;For remaining Ω -1 groups, selected in each group according to similar matrix S
Each point and Global center image IOPresentation graphics apart from the smallest point, as the group;
4-3. is for each group Gα, presentation graphicsConnection in group is carried out between every other image;For any one
To image IiWith image Ij∈Gα, wherein j ≠ i, if i=iαOr j=iαJust there is eI, j=1, otherwise eI, j=0;
4-4. is for the advanced connection between group, Global center image IOConnect every other presentation graphicsFor appointing
What a pair of of image IiWithFor, if i=o or j=o, there is eI, j=1, otherwise eI, j=
0。
4. the figure according to claim 3 based on Density Clustering shrinks grouping method for registering, it is characterised in that step 5 is specific
It is accomplished by
As figure is dynamically shunk, group's registration more crucially determines each image Ii(t) how to be deformed in time t,
So that energy function F (t) is minimized;Assuming that each image is from Ii(t0) become Ii(tk), wherein { tkBe time t (k=0 ...,
K, t0=0, tk→ ∞) discretization;According to its connecting node by Ii(tk) moved along mean direction, so each node I in figurei
(tk) local connectivity be reasonable;
Since velocity vector is located at the I in manifoldi(tk) tangent line spatially, calculated by linear averagingWhereinIt is Ii(tk) connection quantity;To being scheduled on time point tk
Each velocity vector direction, if each node Ii(tk) along the direction of velocity vector from tkIt is moved to tk+Δtk, whole
Then strictly monotone is reduced a energy function F (t), wherein incremental time Δ tkIt is obtained by following formula:
Finally, by connecting from t0To tkDeformation section obtain each image IiTo the geodesic curve path at centerI.e.WhereinIndicate deformation composition;In the figure systole phase
Between, all images are gradually deformed to hiding public space, and the topology of reserved graph;T is used based on interval of delta tk(k=0,1,
2 ..., wherein t0=the 0 and t as k → ∞k→ ∞) carry out discretization t;It usesAs time tkStylish deformation pattern, willIt is expressed as time tkLocate imageSynthesis deformation field.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810777473.0A CN109242759B (en) | 2018-07-16 | 2018-07-16 | Graph contraction grouping registration method based on density clustering |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810777473.0A CN109242759B (en) | 2018-07-16 | 2018-07-16 | Graph contraction grouping registration method based on density clustering |
Publications (2)
Publication Number | Publication Date |
---|---|
CN109242759A true CN109242759A (en) | 2019-01-18 |
CN109242759B CN109242759B (en) | 2022-12-13 |
Family
ID=65072735
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810777473.0A Active CN109242759B (en) | 2018-07-16 | 2018-07-16 | Graph contraction grouping registration method based on density clustering |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN109242759B (en) |
Cited By (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110988856A (en) * | 2019-12-19 | 2020-04-10 | 电子科技大学 | Target detection trace agglomeration algorithm based on density clustering |
CN112417078A (en) * | 2020-11-30 | 2021-02-26 | 中国地质大学(武汉) | Space point aggregation method for aviation geophysical mapping data |
CN113408569A (en) * | 2021-04-27 | 2021-09-17 | 中国人民解放军陆军军医大学 | Image registration method based on density clustering |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20060034545A1 (en) * | 2001-03-08 | 2006-02-16 | Universite Joseph Fourier | Quantitative analysis, visualization and movement correction in dynamic processes |
US20180012463A1 (en) * | 2016-07-11 | 2018-01-11 | Google Inc. | Methods and Systems for Person Detection in a Video Feed |
-
2018
- 2018-07-16 CN CN201810777473.0A patent/CN109242759B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20060034545A1 (en) * | 2001-03-08 | 2006-02-16 | Universite Joseph Fourier | Quantitative analysis, visualization and movement correction in dynamic processes |
US20180012463A1 (en) * | 2016-07-11 | 2018-01-11 | Google Inc. | Methods and Systems for Person Detection in a Video Feed |
Non-Patent Citations (2)
Title |
---|
刘恒等: "一种基于密度的分布式聚类改进算法", 《微电子学与计算机》 * |
薛小娜等: "结合K近邻的改进密度峰值聚类算法", 《计算机工程与应用》 * |
Cited By (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110988856A (en) * | 2019-12-19 | 2020-04-10 | 电子科技大学 | Target detection trace agglomeration algorithm based on density clustering |
CN110988856B (en) * | 2019-12-19 | 2021-08-03 | 电子科技大学 | Target detection trace agglomeration algorithm based on density clustering |
CN112417078A (en) * | 2020-11-30 | 2021-02-26 | 中国地质大学(武汉) | Space point aggregation method for aviation geophysical mapping data |
CN112417078B (en) * | 2020-11-30 | 2022-06-24 | 中国地质大学(武汉) | Space point aggregation method for aviation geophysical mapping data |
CN113408569A (en) * | 2021-04-27 | 2021-09-17 | 中国人民解放军陆军军医大学 | Image registration method based on density clustering |
Also Published As
Publication number | Publication date |
---|---|
CN109242759B (en) | 2022-12-13 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Wei et al. | Superpixel hierarchy | |
WO2021057744A1 (en) | Positioning method and apparatus, and device and storage medium | |
CN109242759A (en) | Figure based on Density Clustering shrinks grouping method for registering | |
CN109934855A (en) | A kind of livewire work scene power components three-dimensional rebuilding method based on cloud | |
CN110264471A (en) | A kind of image partition method, device, storage medium and terminal device | |
CN108961385B (en) | SLAM composition method and device | |
CN115953535A (en) | Three-dimensional reconstruction method and device, computing equipment and storage medium | |
CN108846404A (en) | A kind of image significance detection method and device based on the sequence of related constraint figure | |
CN109271441A (en) | A kind of visualization clustering method of high dimensional data and system | |
CN112328715A (en) | Visual positioning method, training method of related model, related device and equipment | |
CN107123138B (en) | Based on vanilla-R point to the point cloud registration method for rejecting strategy | |
CN109345536A (en) | A kind of image superpixel dividing method and its device | |
Wolf et al. | The semantic mutex watershed for efficient bottom-up semantic instance segmentation | |
CN103544712A (en) | Method for automatically segmenting human lateral geniculate nucleus through prior knowledge | |
Gan et al. | TiC2D: trajectory inference from single-cell RNA-seq data using consensus clustering | |
CN111862176B (en) | Three-dimensional oral cavity point cloud orthodontic front and back accurate registration method based on palatine fold | |
CN111383353B (en) | Fractured bone model registration method based on Gaussian mixture model and contour descriptor | |
Hemanth et al. | Distance metric-based time-efficient fuzzy algorithm for abnormal magnetic resonance brain image segmentation | |
CN108510010A (en) | A kind of density peaks clustering method and system based on prescreening | |
Liu et al. | Cell population tracking in a honeycomb structure using an IMM filter based 3D local graph matching model | |
CN112417507A (en) | Method for issuing node triangle count of large graph based on privacy protection | |
CN117408913A (en) | Method, system and device for denoising point cloud of object to be measured | |
CN113128617A (en) | Spark and ASPSO based parallelization K-means optimization method | |
CN116842818A (en) | Structural plane attitude grouping method based on pelican optimization algorithm | |
CN115690439A (en) | Feature point aggregation method and device based on image plane and electronic equipment |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |