CN104700401B - A kind of image affine transformation control point choosing method based on K Means clustering procedures - Google Patents
A kind of image affine transformation control point choosing method based on K Means clustering procedures Download PDFInfo
- Publication number
- CN104700401B CN104700401B CN201510047526.XA CN201510047526A CN104700401B CN 104700401 B CN104700401 B CN 104700401B CN 201510047526 A CN201510047526 A CN 201510047526A CN 104700401 B CN104700401 B CN 104700401B
- Authority
- CN
- China
- Prior art keywords
- mrow
- characteristic point
- point
- image
- characteristic
- 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.)
- Expired - Fee Related
Links
Abstract
The present invention relates to a kind of feature point extraction and image matching method based on K Means clustering procedures, the cluster centre by K Means clustering algorithms processing procedures (1) for characteristic point distribution respective numbers to be clustered；(2) each characteristic point is calculated to the distance of cluster centre；(3) calculate in each cluster coordinate average value a little, and repeat step (2) and step (3) untill reaching requirement using this average value as new cluster centre；(4) characteristic point farthest from picture centre is chosen from each feature points clustering；(5) characteristic point of matching is extracted；(6) characteristic point for the matching finally obtained is used for the calculating of affine Transform Model parameter.The present invention realize image zooming-out subject to registration characteristic point and standard picture in matching between obtained reference characteristic point, choose the characteristic point with high-precision matching performance.
Description
Technical field
The present invention relates to characteristic point matching method, especially a kind of image affine transformation control based on K-Means clustering procedures
System point choosing method.
Background technology
Image registration is a basic problem of field of machine vision, is always the focus and difficult point of people's research.Image
Registration refers to find between two width or multiple image of the Same Scene from different time, different visual angles or different sensors
Corresponding relation.Method about image registration is broadly divided into：Method based on gray scale registration and feature based registration.
Based on the method for gray scale registration, also referred to as correlation registration method, image registration is carried out with space two-dimensional sleiding form,
The difference of algorithms of different is mainly reflected in the selection of correlation criterion.Distinguished point based method for registering is to be carried first in original image
Feature is taken, then establishes the registering relation of characteristic point between two images, there is higher robustness, such as SURF algorithm.Mesh
Before, it has been used widely based on SURF algorithm image registration, such as applied to medical figure registration, remote sensing image registration
Deng.The dimensional properties that SURF characteristic points have are that the characteristic point extracted in image carries out Feature Points Matching, and by the feature of matching
Point is to for calculating image registration transformation model parameter, making image subject to registration transform to standard picture position exactly, to reach
Registering purpose.The matching of characteristic point is to realize the key of image registration, and the precision of matching directly affects the precision of subsequent registration.
The characteristic point neighborhood information included in being accorded with algorithm according to SURF feature point descriptions, each spy can be found out using K nearest neighbor methods
Potential two optimal match points of sign point, and optimal match point is preferably gone out apart from optimum value and sub-optimal value ratio by match point
It is right.
According to the characteristic point matching method proposed above, it is possible to achieve the characteristic point and standard picture of image zooming-out subject to registration
In matching between obtained reference characteristic point, the characteristic point that the match is successful is to all having very high precision.But actually solving
Cheng Zhong, how therefrom choosing suitable three matching characteristics point and calculating to carrying out parameter is to realize image registration then another is important
Factor, how effective validity feature point is the crucial problem solved the problems, such as.
The content of the invention
It is an object of the invention to provide a kind of image affine transformation control point selection side based on K-Means clustering procedures
Method, optimal validity feature point is chosen, so as to obtain optimized image registration.
The present invention solves its technical problem and takes following technical scheme to realize：
A kind of image affine transformation control point choosing method based on K-Means clustering procedures, it is characterised in that：Utilize SURF
Feature point detection algorithm detection image characteristic point in registering image, and the requirement extracted according to fact characteristic point, by K-
Means clustering algorithm processing procedures, the screening of characteristic point is divided into following steps and realized：
(1) cluster centre of respective numbers is distributed for characteristic point to be clustered；
(2) each characteristic point is calculated to the distance of cluster centre, and by each point cluster into the cluster nearest from the point；
(3) calculate in each cluster coordinate average value a little, and using this average value as new cluster centre weight
Step (2) and step (3) are performed again until cluster centre no longer moves or clustered on a large scale untill number reaches requirement；
(4) after the cluster of all characteristic points is completed, selection one is farthest from picture centre from each feature points clustering
Characteristic point, form a new set of characteristic points；
(5) three characteristic points are randomly selected in the set of characteristic points extracted again, triangle area is calculated, is asked until finally
Go out the triangle area of maximum, extract the characteristic point of matching；
(6) characteristic point finally obtained is made into image subject to registration exactly to the calculating for affine Transform Model parameter
Standard picture position is transformed to, to reach registering purpose.
Moreover, the SURF feature point detection algorithms are by calculating the local extremum of Hessian matrix determinants come really
Determine the position of characteristic point, yardstick is σ image I midpointsHessian matrixes be defined as：
In formula, LxxIt is that Gauss second order is ledThe result of same I0 (x, y) convolution, wherein
Lxy,LyyWith identical implication.
The advantages and positive effects of the present invention are：
How the present invention chooses effective three groups of characteristic points to carrying out parameter calculating from the distribution of matching characteristic point to study,
So as to realize the matching between the reference characteristic point obtained in the characteristic point of image zooming-out subject to registration and standard picture, choosing has height
The characteristic point of Accuracy Matching performance.
Brief description of the drawings
Fig. 1 is 3 groups of matching characteristic points of various combination to (forming 3 triangle areas)；
Fig. 2 is characterized a triangle area/full images area formed and the relation of PSNR values；
Fig. 3 is characterized a result for cluster；
Fig. 4 is eventually for 3 characteristic points for calculating registration transformation model parameter；
Fig. 5 is 9 × 9 box Filtering Templates；
Fig. 6 is characterized the selection of principal direction a little；
The composition of Fig. 7 feature descriptors.
Embodiment
Below in conjunction with the accompanying drawings and the invention will be further described by specific embodiment, and following examples are descriptive
, it is not limited, it is impossible to which protection scope of the present invention is limited with this.
A kind of image affine transformation control point choosing method based on K-Means clustering procedures, utilizes SURF feature point detections
Algorithm detection image characteristic point in registering image, and the requirement extracted according to fact characteristic point, by K-Means clustering algorithms
Processing procedure, the screening of characteristic point is divided into following steps and realized：
(1) cluster centre of respective numbers is distributed for characteristic point to be clustered；
(2) each characteristic point is calculated to the distance of cluster centre, and by each point cluster into the cluster nearest from the point；
(3) calculate in each cluster coordinate average value a little, and using this average value as new cluster centre weight
Step (2) and step (3) are performed again until cluster centre no longer moves or clustered on a large scale untill number reaches requirement；
(4) after the cluster of all characteristic points is completed, one is chosen from each feature points clustering from image picture centre
Farthest characteristic point, form a new set of characteristic points；
(5) three characteristic points are randomly selected in the set of characteristic points extracted again, triangle area is calculated, is asked until finally
Go out the triangle area of maximum, extract the characteristic point of matching；
(6) characteristic point for the matching finally obtained is used for the calculating of affine Transform Model parameter, makes image subject to registration accurate
Standard picture position really is transformed to, to reach registering purpose.
The acquisition of characteristic point is to detect to obtain in registering image using SURF feature point detection algorithms.
SURF feature point detection algorithms are that a kind of speed for being proposed on the basis of SIFT algorithms is fast, the feature of strong robustness
Extraction algorithm.The algorithm not only has preferable robustness to image rotation, translation, scaling and noise, and to illumination variation
There is preferable processing with visual angle change consistency and image blur.Meanwhile by introducing integral image and cassette filter,
3 times are about improved in arithmetic speed, combination property is more superior.
SURF feature point detection algorithms are to determine characteristic point by calculating the local extremum of Hessian matrix determinants
Position.Wherein, yardstick is σ image I midpointsHessian matrixes be defined as：
In formula, LxxIt is that Gauss second order is ledThe result of same I=(x, y) convolution, wherein
Lxy,LyyWith identical implication.
SURF algorithm is filtered come approximate second gaussian filtering using box, constructs a kind of quick Hessian matrixes, and make
Accelerate convolution with integral image to improve calculating speed.
Wherein, it is illustrated in figure 59 × 9 box Filtering Template.In order that SURF features there is scale invariability, it is necessary to
Tectonic scale space, and Local Extremum is obtained by quick Hessian matrix determinants under each metric space.Here,
SURF algorithm uses different size of cassette filter to carry out convolution with source images to obtain the description of the feature in different scale space.
After box filtering process, the value in x directions is denoted as Dxx, the value in xy directions is denoted as Dxy, the value in y directions is denoted as Dyy.Due to box
Filtering is the approximate evaluation of second order Gauss filtering, therefore introduces scale factor ω (ω approximations take 0.9), further solves and obtains soon
The ranks expression formula of fast Hessian matrixes is：
Δ H=Dxx(x)Dyy(x)-(ωDxy(x))2
Wherein, Δ is the cassette filter response of picture point I (x, y) peripheral region, and the detection of extreme point is carried out with Δ.
, will be each after obtaining scalogram picture extreme point local under each metric space according to quick Hessian matrixes
Each 9 points of Local Extremum and 8 consecutive points of same yardstick and its two yardstick up and down form one 3 × 3 × 3 vertical
Body neighborhood.Compared with by each extreme point of metric space 26 points adjacent with three-dimensional neighborhood, only work as Local Extremum
Value when being all higher than (or less than) all 26 consecutive points, just using this Local Extremum as candidate feature point.In order to right
Candidate feature point carries out sub-pixel positioning, can enter row interpolation in metric space and image space, the characteristic point stablized
Position and the scale-value at place.
The quadratic fit function that interpolation uses is：
Obtaining extreme point to function derivation is：
Extreme value of the function at extreme point be：
D (x) ＜ 0.03 candidate feature point is rejected in experiment.
The description of SURF characteristic points
To keep the rotational invariance of characteristic point, after characteristic point position determination, main side is carried out for each characteristic point
To determination.Therefore, centered on characteristic point, calculating radius is in 6 σ (σ is characterized yardstick a little) border circular areas, to figure
Letter is weighted as carrying out Harr small echos (the Harr small echo length of sides take 4 σ) response computing in x and y directions, and using the Gauss that yardstick is 2 σ
It is several that Gauss weighting is carried out to Harr small echos response so that the response contribution closer to characteristic point is bigger；Then with the size of π/3
Sector region scope travels through whole border circular areas, and the vector mould that horizontal directional response in domain and vertical direction are responded is most
The direction being worth greatly is defined as the principal direction (as shown in Figure 6) of characteristic point.
Then, centered on characteristic point, the construction square window region perpendicular to principal direction and the length of side for 20 σ on one side,
And the window area is divided into 4 × 4 subregion.In each sub-regions, the level side of the sampled point of 5 σ × 5 σ is carried out
To the calculating of the Harr small echos response with vertical direction, d is denoted as respectivelyxAnd dy, it is same to use the gaussian weighing function that yardstick is 2 σ
Gauss weighting is carried out to Harr small echos response, to increase the robustness to geometric transformation.Then by the response of every sub-regions
Value is added to form ∑ d with the absolute value of responsex, ∑ dy, ∑ | dx|, ∑ | dy|.Thus, it is formed one per sub-regions
Four dimensional feature description vectors V:
V=(Σ dx,Σdx,Σ|dx|,Σ|dy|)
For each characteristic point, the characteristic vector that 4 × 4 × 4=64 is tieed up is formed.It is right to ensure the consistency to illumination
Characteristic vector is normalized, and obtains final SURF descriptors, as shown in Figure 7.
The matching of characteristic point
The matching of characteristic point is to realize the key of image registration, and the precision of matching directly affects the precision of subsequent registration.
The characteristic point neighborhood information included in being accorded with algorithm according to SURF feature point descriptions, each spy can be found out using K nearest neighbor methods
Potential two optimal match points of sign point, and optimal match point is preferably gone out apart from optimum value and sub-optimal value ratio by match point
It is right.
If M1、M2Respectively two images I1、I2The set of characteristic points to be matched of SURF extractions is respectively adopted, to M1In appoint
One characteristic point m1i, M2In with m1iMinimum two characteristic points of Euclidean distance be respectively m2j,m'2j, respective distances are respectively dij,
d'ijIf dij≤α*d'ij(experiment takes α=0.65), then it is assumed that m1iWith m2jFor preferable matching double points.
Standard picture and all matching double points of image subject to registration can be obtained using above methodWherein S, D are respectively that standard picture and image subject to registration complete matching
Feature point set.Due in these matching double points there is certain Mismatching point pair, herein using the similar method pair of triangle
It is purified.Appoint in feature point set S, D and take three matching double points P1With Q1、P2With Q2And P3With Q3, form virtual triangle
Shape Δ P1P2P3With Δ Q1Q2Q3, and a triangle pair is formed, thirdly the length on side is set to lp1、lp2、lp3And lq1、lq2、
lq3If meeting the relation in following formula, the matching double points for showing to choose are available point pair, Mismatching point pair otherwise be present,
Give and reject, realize the purification of characteristic point pair：
Control errors in the present embodiment experiment between side ratio are 0.02.
Although disclosing embodiments of the invention and accompanying drawing for the purpose of illustration, those skilled in the art can manage
Solution：Do not departing from the present invention and spirit and scope of the appended claims in, it is various replace, change and modifications all be it is possible,
Therefore, the scope of the present invention is not limited to embodiment and accompanying drawing disclosure of that.
Claims (1)
- A kind of 1. image affine transformation control point choosing method based on K-Means clustering procedures, it is characterised in that：It is special using SURF Sign point detection algorithm detection image characteristic point pair in registering image, and according to requirement of the fact characteristic point to extraction, by K- Means clustering algorithm processing procedures, the screening of characteristic point pair is divided into following steps：(1) cluster centre of respective numbers is distributed for characteristic point to be clustered；(2) each characteristic point is calculated to the distance of cluster centre, and by each point cluster into the cluster nearest from the point；(3) the coordinate average value of institute a little in each cluster is calculated, and is repeatedly held using this average value as new cluster centre Row step (2) and step (3) no longer move or clustered on a large scale untill number reaches requirement until cluster centre；(4) after the cluster of all characteristic points is completed, a spy farthest from picture centre is chosen from each feature points clustering Point is levied, forms a new set of characteristic points；(5) three characteristic points are randomly selected in the set of characteristic points extracted again, triangle area is calculated, is obtained most until finally Big triangle area, extract the characteristic point pair of matching；(6) image subject to registration is made to convert exactly the calculating for affine Transform Model parameter the characteristic point finally obtained To standard picture position, to reach registering purpose；The SURF feature point detection algorithms are to determine characteristic point by calculating the local extremum of Hessian matrix determinants Position, yardstick are σ image I midpointsHessian matrixes be defined as：<mrow> <mi>H</mi> <mo>=</mo> <mfenced open = "[" close = "]"> <mtable> <mtr> <mtd> <mrow> <msub> <mi>L</mi> <mrow> <mi>x</mi> <mi>x</mi> </mrow> </msub> <mrow> <mo>(</mo> <mover> <mi>x</mi> <mo>^</mo> </mover> <mo>,</mo> <mi>&sigma;</mi> <mo>)</mo> </mrow> </mrow> </mtd> <mtd> <mrow> <msub> <mi>L</mi> <mrow> <mi>x</mi> <mi>y</mi> </mrow> </msub> <mrow> <mo>(</mo> <mover> <mi>x</mi> <mo>^</mo> </mover> <mo>,</mo> <mi>&sigma;</mi> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msub> <mi>L</mi> <mrow> <mi>x</mi> <mi>y</mi> </mrow> </msub> <mrow> <mo>(</mo> <mover> <mi>x</mi> <mo>^</mo> </mover> <mo>,</mo> <mi>&sigma;</mi> <mo>)</mo> </mrow> </mrow> </mtd> <mtd> <mrow> <msub> <mi>L</mi> <mrow> <mi>y</mi> <mi>y</mi> </mrow> </msub> <mrow> <mo>(</mo> <mover> <mi>x</mi> <mo>^</mo> </mover> <mo>,</mo> <mi>&sigma;</mi> <mo>)</mo> </mrow> </mrow> </mtd> </mtr> </mtable> </mfenced> </mrow>In formula, LxxIt is that Gauss second order is ledThe result of same I=(x, y) convolution, whereinLxy, LyyWith identical implication.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510047526.XA CN104700401B (en) | 2015-01-30 | 2015-01-30 | A kind of image affine transformation control point choosing method based on K Means clustering procedures |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510047526.XA CN104700401B (en) | 2015-01-30 | 2015-01-30 | A kind of image affine transformation control point choosing method based on K Means clustering procedures |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN104700401A CN104700401A (en) | 2015-06-10 |
| CN104700401B true CN104700401B (en) | 2017-11-17 |
Family
ID=53347483
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510047526.XA Expired - Fee Related CN104700401B (en) | 2015-01-30 | 2015-01-30 | A kind of image affine transformation control point choosing method based on K Means clustering procedures |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN104700401B (en) |
Families Citing this family (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111915645B (en) * | 2020-07-14 | 2021-08-27 | 珠海大横琴科技发展有限公司 | Image matching method and device, computer equipment and computer readable storage medium |
| CN113408569B (en) * | 2021-04-27 | 2022-07-19 | 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 |
|---|---|---|---|---|
| US6975755B1 (en) * | 1999-11-25 | 2005-12-13 | Canon Kabushiki Kaisha | Image processing method and apparatus |
| CN102005047A (en) * | 2010-11-15 | 2011-04-06 | 无锡中星微电子有限公司 | Image registration system and method thereof |
-
2015
- 2015-01-30 CN CN201510047526.XA patent/CN104700401B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6975755B1 (en) * | 1999-11-25 | 2005-12-13 | Canon Kabushiki Kaisha | Image processing method and apparatus |
| CN102005047A (en) * | 2010-11-15 | 2011-04-06 | 无锡中星微电子有限公司 | Image registration system and method thereof |
Non-Patent Citations (3)
| Title |
|---|
| 仿射不变特征配准算法及其优化技术研究;杨泽;《中国优秀硕士学位论文全文数据库 信息科技辑》;20140115;第2015年卷(第1期);I138-1006 * |
| 基于K-means聚类和RANSAC的图像配准算法研究;王天召等;《计算机工程与科学》;20140930;第36卷(第9期);全文 * |
| 多特征融合的图像自动变形;肖红光等;《中国图象图形学报》;20140731;第19卷(第7期);1012-1020 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN104700401A (en) | 2015-06-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103426186B (en) | A kind of SURF fast matching method of improvement | |
| CN107067415B (en) | A kind of object localization method based on images match | |
| CN104200495B (en) | A kind of multi-object tracking method in video monitoring | |
| CN104090972A (en) | Image feature extraction and similarity measurement method used for three-dimensional city model retrieval | |
| CN108010045A (en) | Visual pattern characteristic point error hiding method of purification based on ORB | |
| CN104933434A (en) | Image matching method combining length between perpendiculars (LBP) feature extraction method and surf feature extraction method | |
| CN107248159A (en) | A kind of metal works defect inspection method based on binocular vision | |
| CN104899918B (en) | The three-dimensional environment modeling method and system of a kind of unmanned plane | |
| CN103136525B (en) | A kind of special-shaped Extended target high-precision locating method utilizing Generalized Hough Transform | |
| CN103729654A (en) | Image matching retrieval system on account of improving Scale Invariant Feature Transform (SIFT) algorithm | |
| CN102122359B (en) | Image registration method and device | |
| CN106447704A (en) | A visible light-infrared image registration method based on salient region features and edge degree | |
| CN103400384A (en) | Large viewing angle image matching method capable of combining region matching and point matching | |
| CN102074015A (en) | Two-dimensional image sequence based three-dimensional reconstruction method of target | |
| CN103632137B (en) | A kind of human eye iris segmentation method | |
| CN104123554B (en) | SIFT image characteristic extracting methods based on MMTD | |
| CN104182973A (en) | Image copying and pasting detection method based on circular description operator CSIFT (Colored scale invariant feature transform) | |
| CN109376744A (en) | A kind of Image Feature Matching method and device that SURF and ORB is combined | |
| CN108550165A (en) | A kind of image matching method based on local invariant feature | |
| CN108052912A (en) | A kind of three-dimensional face image recognition methods based on square Fourier descriptor | |
| CN102446356A (en) | Parallel and adaptive matching method for acquiring remote sensing images with homogeneously-distributed matched points | |
| CN107154017A (en) | A kind of image split-joint method based on SIFT feature Point matching | |
| CN108664983A (en) | A kind of scale and the adaptive SURF characteristic point matching methods of characteristic strength | |
| CN108182705A (en) | A kind of three-dimensional coordinate localization method based on machine vision | |
| CN101833763B (en) | Method for detecting reflection image on water surface |
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 | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20201010 Address after: 300193 Qingxin garden, Jiaokou, Nanfeng Road, Changjiang Road, Nankai District, Tianjin Patentee after: Tianjin TEDA Property Management Co., Ltd Address before: 300222 Dagu South Road, Tianjin, No. 1038, No. Patentee before: Tianjin University of Science and Technology |
|
| TR01 | Transfer of patent right | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20171117 Termination date: 20210130 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |