CN103955667B - SIFT human face matching method based on geometrical constraint - Google Patents
SIFT human face matching method based on geometrical constraint Download PDFInfo
- Publication number
- CN103955667B CN103955667B CN201310209140.5A CN201310209140A CN103955667B CN 103955667 B CN103955667 B CN 103955667B CN 201310209140 A CN201310209140 A CN 201310209140A CN 103955667 B CN103955667 B CN 103955667B
- Authority
- CN
- China
- Prior art keywords
- matching
- sift
- point
- feature
- picture
- 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
Landscapes
- Image Analysis (AREA)
Abstract
The invention discloses an SIFT human face matching method based on geometrical constraint and the method increases the matching rate of SIFT human face characteristics and reduces error matching conditions and improves the accurate matching rate and practicality. The realization process is as follows: firstly inputting pictures of to-be-matched human faces so as to perform SIFT characteristic extraction and characteristic point matching; using the characteristic points and the matching result as a training sample and obtaining a predictor through a least-square support vector machine (LS-SVM); inputting SIFT characteristics of a human face picture in the to-be-tested pictures again into the predictor after obtaining the predictor; the predictor performing prediction on the input data and obtaining a prediction matching point on a second picture; and at last, searching for SIFT characteristic points around the prediction matching point of the second picture and finding a correct matching point in the SIFT characteristic points.
Description
Technical field
The invention belongs to technical field of image processing, is related to feature point extraction, image registration can be used for identification,
Video monitoring, security system.
Background technology
In computer vision field, Face datection, recognition of face is belonging to a very wide technology of application prospect,
Recognition of face be inside multimedia technology in study hotspot.Due to the multiformity of facial expression, background mixes and hair style,
Illumination condition, itself is blocked, and not equal many factors of facial image resolution cause recognition of face to become extremely difficult
Problem.Most important two directly influence the factor for setting effect in recognition of face, and one is can to find separating capacity
By force, the face characteristic of high stability, two is to find the good identification of recognition effect or matching algorithm according to these features.
Due to lacking for shooting time, shooting angle, the change of natural environment, the use of multiple sensors and sensor itself
Fall into, make the image of shooting not only affected by noise, and there is serious tonal distortion and geometric distortion.In this condition
Under, how matching algorithm reaches high precision, matching accuracy height, fast speed, robustness and strong interference immunity and Parallel Implementation
Become the target of people's pursuit.In the past few decades, various image matching algorithms occur in succession, and with reference to many mathematics
Theoretical and method, people constantly propose new matching process again.Based on the ultimate principle of matching, image matching algorithm is divided into base
In the related matching of gray scale, feature-based matching, the matching based on model and the matching based on transform domain.Feature based
In matching process, the extraction of characteristics of image is the most key, finds robustness in an environmental change, and the feature of good stability is come
Represent the picture.
SIFT algorithms are local feature description's that David Lowe were proposed in 1999, and have been carried out in 2004 deeper
The research that enters and perfect.SIFT feature matching algorithm can be processed occur between two width images translation, rotation, affine transformation situation
Under matching problem, with very strong adaptability.In Mikolajczyk to including ten kinds of partial descriptions including SIFT operators
In the invariance contrast experiment that son is done, SIFT and its expansion algorithm confirm there is most strong vigorousness in similar description.
But in SIFT matching algorithms, characteristic matching rate is low, there is the situation of mismatching, the threshold value of match point is too low, causes matching rate
It is low, should match to SIFT feature be judged to it is not that matching is right, and some properties, the SIFT feature that geometry has a long way to go is recognized
It is set to matching right.
Least square method supporting vector machine (LS-SVM) is belonging to statistical classification method, by machine learning from a series of training
Data learning obtains a predictor, and input is predicted with the predictor, obtains experience output.Obtain the matching predicted
Accurate match point is found within certain limit around this future position after point.The advantage of algorithm of support vector machine is output result
Stable, effect is good, can effectively improve the high shortcoming of former algorithm error hiding rate.
The content of the invention
It is an object of the present invention to improve the deficiency of above-mentioned prior art, more preferable effect is obtained, it is proposed that one kind is based on
The SIFT face matching process of geometrical constraint, increased the number of match point, reduce the rate that mismatches of images match, but improve
Accurate match rate and its practicality.
The technical scheme is that, face picture to be matched is input into first carries out SIFT feature extraction and characteristic point
Match somebody with somebody, using these characteristic points and matching result as training sample, by least square method supporting vector machine (LS-SVM) one is obtained
Predictor, obtains after predictor being again input in predictor the SIFT feature of a face picture in picture to be measured, predicts
Device is predicted to this input data, and prediction and matching point is obtained on the second pictures.Finally, in the second pictures prediction and matching
Search SIFT feature point around point, and find accurate match point in these SIFT feature points.
Comprise the following steps that：
One, two face pictures to be matched are input into, SIFT feature extraction is carried out to picture, then calculated by SIFT matchings
Method carries out Feature Points Matching.
Wherein, SIFT feature is the local description extracted using SIFT algorithms, is the characteristic vector of 128 dimensions.It is special
Description of the position of SIFT feature point, yardstick and 128 dimensions is obtained when levying extraction.Characteristic matching is referred to be calculated with k_d tree matchings
Method finds matching characteristic point of the characteristic point of the first pictures in the second pictures, and the matching characteristic in the second pictures is also
Individual SIFT feature.In SIFT matching algorithms feature point description son between Euclidean distance as standard, find minimum distance and time
In-plant characteristic point, minimum distance and time in-plant ratio set a threshold value, and ratio meets its threshold condition and is judged as
A pair of matchings are right.
Two, used as training sample, training data learning obtains one for the SIFT feature that upper step is obtained and match point
Predictor.
Wherein, training sample data refer to the position of all SIFT features of the first pictures to be matched and at second
The position of its corresponding match point in picture.
According to least square method supporting vector machine principle, vector (v1,v2,..,vi,...,vn) it is input feature vector, (u1,
u2,..,ui,...,un) corresponding match point.Arbitrary characteristic point meets below equation：
ui=Avi+ b, i=1,2 ... ..n
Wherein
The match point corresponding relation defined by this two groups of matching double points is Km=u
Wherein, K is the matrix that the SIFT feature point on the first pictures is constituted, and u is corresponding feature on the second pictures
Point vector, m=[A11 A12 b1 A21 A22 b2]T, and the m predictor to be calculated that is exactly us.
The m that can be obtained with method of least square is：
Three, the SIFT feature of a face picture in picture to be measured is input in predictor again, obtain prediction and matching
Point.
Wherein, again input refers to that the SIFT feature of the first pictures to be matched, as K, is obtained by formula Km=u
U, u refer to the vector of the SIFT feature in the corresponding prediction and matching point composition of the second pictures of the first pictures.
Four, SIFT feature point is searched for around prediction and matching point.
Wherein, prediction and matching point is the point predicted on second photo by predictor, is as center R using this point
Radius obtains a regional area, is called estimation range, searches in this region either with or without SIFT feature point.
Five, Euclidean distance is calculated, obtain accurate match point.
The SIFT feature of the first pictures is calculated with each SIFT feature in corresponding estimation range in the second pictures
Euclidean distance.By calculating, that minimum SIFT feature point of distance is its accurate match point.
Description of the drawings
A kind of SIFT face matching process flow charts based on geometrical constraint of Fig. 1 present invention；
Fig. 2 finds in future position region of the present invention accurate match point schematic diagram.
Specific embodiment
The specific embodiment of the present invention is described in further detail below.
As shown in figure, a kind of SIFT face matching process flow charts based on geometrical constraint of the present invention, first input are treated
Two pictures of matching, extract SIFT feature in face picture, if, I (x, y) and T (x, y) is face figure to be matched
Piece, niFor the sum of the SIFT feature of picture I (x, y), fi jFor j-th feature of picture I (x, y), j ∈ [1, ni], fi jIt is 128
The characteristic vector of dimension, describes the gradient information of 8 gradient directions of regional area, (xij,yij) it is j-th of picture I (x, y) special
Levy position a little.ntFor the sum of the SIFT feature of picture T (x, y), ft jFor j-th feature of picture T (x, y), j ∈ [1,
nt]。(xtj,ytj) be picture T (x, y) j-th characteristic point position.
Picture I (x, y) and T (x, y) are matched, and Best Bin First (BBF) during matching using Rob Hess is calculated
Method is searching for arest neighbors feature and time neighbour.When being matched, feature f of picture I (x, y)i jWith owning on picture T (x, y)
NtIndividual SIFT feature calculates Euclidean distance, and arest neighbors feature and time neighbour's characteristic point are found on picture T (x, y), by than
It is right that value method finds corresponding matching.If ft jIt is fi jArest neighbors and meet fi jWith arest neighbors distance than secondary neighbour distance
It is less than constant this condition, then fi jAnd ft jIt is that a pair of matchings are right.
Followed by least square method supporting vector machine and its training process.According to least square method supporting vector machine, in step
The corresponding relation be given in rapid two is Km=u, and it is the n on picture I (x, y) to be input into training matrix KiIndividual feature point group into 2ni×
6 matrix.Wherein u is the vector of SIFT feature corresponding match point composition in picture T (x, y) of picture I (x, y).If K
The characteristic point of the inside does not have a corresponding match point, with (0,0) representing its match point, that is to say, that u is 2ni× 1 vector.
Wherein,
By above-mentioned formulaPredictor m is obtained, predictor m is 6 × 1 vector.
Next step is that the predictor by calculating carries out matching work to face picture.
Predictor m is had been obtained for above, input matrix K above is multiplied with predictor m, by u=Km in picture T
N is obtained on (x, y)iIndividual prediction and matching point.
Whether search in estimation range has SIFT feature.As shown in Figure 2, (xtl,ytl) it is fi jIn picture T (x, y)
Corresponding prediction and matching point.With (xtl,ytl) centered on take a regional area, C is (xtl,ytl) centered on estimation range, inspection
Survey whether C the insides in region have the SIFT feature of picture T (x, y), being located at C the insides has s SIFT feature point, fc jIt is the jth inside C
Individual feature, j ∈ [1, s].
Final step is to find accurate match point by calculating Euclidean distance.Such as figure two, fi jWith all of SIFT feature in C
Euclidean distance is calculated, arest neighbors feature is found, that is to say, that the minimum feature of Euclidean distance.If fc jForIt is nearest inside C
Adjacent feature, thenAnd fc jFor a pair of match points.If f when s is 0i jWithout corresponding match point.
Above content is detailed step and embodiment party of the present invention based on the SIFT face matching process of support vector machine
Method.For a person skilled in the art without departing from the design of the present invention on the premise of any change for making belong to the present invention
Protection domain within.
Claims (4)
1. the SIFT face matching process of support vector machine is based on, and the method is comprised the following steps：
1) two face pictures to be matched are input into, SIFT feature extraction is carried out to picture, then entered by SIFT matching algorithms
Row Feature Points Matching,
2), used as training sample, training data learning obtains a prediction for the SIFT feature for upper step being obtained and match point
Device；The match point corresponding relation of two groups of matching double points definition is Km=u, and by least square method supporting vector machine prediction is calculated
Device, K is the matrix that the SIFT feature point on the first pictures is constituted, and m is predictor, and u is corresponding matching in the second pictures
The vector of point composition,
3) SIFT feature of picture to be matched is input in predictor, obtains prediction and matching point,
4) SIFT feature point is searched for around prediction and matching point,
5) Euclidean distance is calculated, obtains accurate match point.
2. method according to claim 1, in step 3, by u=Km prediction and matching point is calculated, and K is to be matched
The matrix of the feature composition of the first pictures, m is predictor, and u is the square of the future position composition on the second pictures to be matched
Battle array.
3. method according to claim 1, in step 4, extracts an estimation range, in estimation range centered on future position
In search whether there is SIFT feature point.
4. method according to claim 1, in step 5, the SIFT feature of first picture to be matched and it is corresponding pre-
The all SIFT features surveyed in region calculate Euclidean distance, find the minimum feature of distance as its accurate match point.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310209140.5A CN103955667B (en) | 2013-05-31 | 2013-05-31 | SIFT human face matching method based on geometrical constraint |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201310209140.5A CN103955667B (en) | 2013-05-31 | 2013-05-31 | SIFT human face matching method based on geometrical constraint |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN103955667A CN103955667A (en) | 2014-07-30 |
| CN103955667B true CN103955667B (en) | 2017-04-19 |
Family
ID=51332942
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201310209140.5A Expired - Fee Related CN103955667B (en) | 2013-05-31 | 2013-05-31 | SIFT human face matching method based on geometrical constraint |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN103955667B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109029365B (en) * | 2018-06-26 | 2021-05-18 | 广东电网有限责任公司 | Method, system, medium and device for extracting different-side image connection points of electric power corridor |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102637251A (en) * | 2012-03-20 | 2012-08-15 | 华中科技大学 | Face recognition method based on reference features |
| CN102880852A (en) * | 2012-07-02 | 2013-01-16 | 北京工业大学 | Face recognition method based on affine scale invariant feature transformation and structural similarity |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR100950776B1 (en) * | 2009-10-16 | 2010-04-02 | 주식회사 쓰리디누리 | Method of face recognition |
-
2013
- 2013-05-31 CN CN201310209140.5A patent/CN103955667B/en not_active Expired - Fee Related
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102637251A (en) * | 2012-03-20 | 2012-08-15 | 华中科技大学 | Face recognition method based on reference features |
| CN102880852A (en) * | 2012-07-02 | 2013-01-16 | 北京工业大学 | Face recognition method based on affine scale invariant feature transformation and structural similarity |
Non-Patent Citations (1)
| Title |
|---|
| 一种基于SIFT算子的人脸识别方法;周志铭等;《中国图象图形学报》;20081031;第1882-1885页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN103955667A (en) | 2014-07-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Ma et al. | A novel two-step registration method for remote sensing images based on deep and local features | |
| CN111126360B (en) | Cross-domain pedestrian re-identification method based on unsupervised combined multi-loss model | |
| CN103530599B (en) | The detection method and system of a kind of real human face and picture face | |
| CN108121991B (en) | Deep learning ship target detection method based on edge candidate region extraction | |
| CN105184801B (en) | It is a kind of based on multi-level tactful optics and SAR image high-precision method for registering | |
| CN109949340A (en) | Target scale adaptive tracking method based on OpenCV | |
| CN102460508B (en) | Image-recognizing method and image recognition apparatus | |
| CN104978567B (en) | Vehicle checking method based on scene classification | |
| CN109800624A (en) | A kind of multi-object tracking method identified again based on pedestrian | |
| CN109697692A (en) | One kind being based on the similar feature matching method of partial structurtes | |
| CN104123554B (en) | SIFT image characteristic extracting methods based on MMTD | |
| CN104978582B (en) | Shelter target recognition methods based on profile angle of chord feature | |
| Logoglu et al. | Cospair: colored histograms of spatial concentric surflet-pairs for 3d object recognition | |
| CN107392187A (en) | A kind of human face in-vivo detection method based on gradient orientation histogram | |
| CN102446356A (en) | Parallel and adaptive matching method for acquiring remote sensing images with homogeneously-distributed matched points | |
| CN105809678B (en) | A kind of line segment feature global registration method between two views under short base line condition | |
| CN103745197A (en) | Detection method of license plate and device thereof | |
| CN103995864B (en) | A kind of image search method and device | |
| CN108182705A (en) | A kind of three-dimensional coordinate localization method based on machine vision | |
| Zou et al. | Microarray camera image segmentation with Faster-RCNN | |
| CN110942473A (en) | Moving target tracking detection method based on characteristic point gridding matching | |
| CN113011359B (en) | Method for simultaneously detecting plane structure and generating plane description based on image and application | |
| CN104392209B (en) | A kind of image complexity evaluation method of target and background | |
| CN103955667B (en) | SIFT human face matching method based on geometrical constraint | |
| CN109740405B (en) | Method for detecting front window difference information of non-aligned similar vehicles |
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 | ||
| CF01 | Termination of patent right due to non-payment of annual fee | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20170419 Termination date: 20190531 |