CN106980812A - Three-dimensional face features' independent positioning method based on concatenated convolutional neutral net - Google Patents
Three-dimensional face features' independent positioning method based on concatenated convolutional neutral net Download PDFInfo
- Publication number
- CN106980812A CN106980812A CN201611154383.3A CN201611154383A CN106980812A CN 106980812 A CN106980812 A CN 106980812A CN 201611154383 A CN201611154383 A CN 201611154383A CN 106980812 A CN106980812 A CN 106980812A
- Authority
- CN
- China
- Prior art keywords
- point
- coordinate system
- face
- dimensional face
- dimensional
- 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.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/161—Detection; Localisation; Normalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- General Physics & Mathematics (AREA)
- Biomedical Technology (AREA)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a kind of three-dimensional face features' independent positioning method based on concatenated convolutional neutral net, described method comprises the following steps:Step 1: being pre-processed to three-dimensional face features' point of training data, SPIDER description are obtained;Step 2: construction concatenated convolutional neural network function, respectively to key point inside facial contour point and face, facial contour feature point and the three-dimensional coordinate of face inter characteristic points are calculated using the process of global restriction iterative respectively.The present invention can make three-dimensional face positioning to reach gratifying result in complex scene with the conditions of.
Description
Technical field
The present invention relates to the research of the key algorithm in three-dimensional face identification technology field, and in particular to one kind is based on concatenated convolutional
Three-dimensional face features' independent positioning method of neutral net.
Background technology
Three-dimensional face identification technology possesses abundant depth information for two-dimension human face identification and can preferably gone back
The original geometry information of protoplast's face, thus be a kind of emerging technology for having a broad based growth prospect.It is widely designed into calculating
Many subject knowledges such as machine vision, image procossing, machine learning, signal transacting and pattern-recognition, current three-dimensional face is known
Other technology has become the study hotspot of international field of face identification.But it is due to the restriction of each side factor, such as computer
Calculated level limitation, three-dimensional acquisition equipment it is immature etc., related research does not obtain very big breakthrough.
Three-dimensional face positioning is to be related to the committed step that three-dimensional face is accurately identified, and algorithm main at present has based on phase
To the algorithm of position, the algorithm based on silhouette lines, the algorithm based on curvature, the algorithm of feature based description, based on two dimensional image
Algorithm of auxiliary etc..The algorithm scope of application based on relative position is narrow, more sensitive to noise;Algorithm based on silhouette lines can not
Suitable for the human face characteristic point forgiven the regions such as eyes, the corners of the mouth more;Algorithm based on curvature will to the quality of three-dimensional face data
Ask comparison high, in addition to interior eyespot, prenasale, the locating effect on other human face characteristic points is poor;Feature based description
Algorithm lacks general applicability;The algorithm aided in based on two dimensional image is not used in the only human face data comprising three-dimensional information.
In a word, existing three-dimensional face Position Research at present without solve well three-dimensional face it is pervasive should and the problem of high efficiency.
The content of the invention
Instant invention overcomes the deficiencies in the prior art, there is provided a kind of three-dimensional face positioning based on concatenated convolutional neutral net
Method, realizes in complex scene with the conditions of, reaches the locating effect of precise and high efficiency.
To solve above-mentioned technical problem, the present invention uses following technical scheme:
A kind of three-dimensional face features' independent positioning method based on concatenated convolutional neutral net, described method includes following step
Suddenly:
Step 1: being pre-processed to three-dimensional face features' point of training data, SPIDER description are obtained;
Step 2: construction concatenated convolutional neural network function, respectively to key point inside facial contour point and face, is utilized
The process of global restriction iterative calculates facial contour feature point and the three-dimensional coordinate of face inter characteristic points respectively.
Further technical scheme is that the step one includes:
Step a, the characteristic point data to the three-dimensional face positioning for training carries out coordinate system conversion, according to formula
(1) by human face characteristic point under rectangular coordinate system in space (x, y, z) Coordinate Conversion under spherical coordinate system
Coordinate
The statistics with histogram of Subspace partition and subspace under step b, spherical coordinate system.
Further technical scheme is that the step b includes:
Step I, three-dimensional face curved surface is made up of a series of summit with tri patch, and the set expression of tri patch is
T, tiRepresent wherein i-th tri patch;For a summit p on curved surface, local surface radius of influence R is given, then pinpoints p
The local surface of surrounding is by tri patch set Ti={ ti|||ci- p | | < R, ciFor tiCenter of gravity represent;Radius of influence R is put down
M parts are divided into, by TrIt is divided into M subsetM=1,2 ..., M, andWherein ciFor tiCenter of gravity, m=1,2 ..., M;
Step II, using p as origin center, is divided into the fan-shaped of multiple sector regions by polar coordinate system and divides;I.e. using p as original
Dot center, sets up polar coordinate system, wherein X-axis is polar coordinate system θ axles, and θ axles are divided into N parts on X/Y plane, is N number of by T points
Subset { Tn, n=1,2 ..., N, andFor tri patch ti, its center of gravity ciProjection on X/Y plane
Ci' is expressed as the coordinate value in polar coordinate systemN-th subset is defined asN=1,
2 ..., N;
Step III, cartesian product step I and step II curved surface divided, the local surface centered on p points is divided
For M × N number of subsurfaceM=1,2 ..., M, n=1,2 ..., N, wherein
Step IV, for the m × n subsurfaceOrderOne of tri patch is represented, is used
RepresentNormal vector;It is that origin sets up spherical coordinate system with coordinate (0,0,0) in normal vector space, wherein three-dimensional right angle is sat
Coordinate Z axis is spherical coordinate system in mark systemAxle, X-axis is θ axles, normal directionBy the coordinate value for seeking coordinate systemTo represent;
In spherical coordinate system, half sphereThe span of axle isThe span of θ axles be [0,2 π);Point
Not willAxle K deciles, θ axle L deciles, K × L unit is divided into by normal vector space;Normal vector is located at kth × l and wished
Dough sheet set expression be
WhereinAnd θiIt is tri patch respectivelyNormal directionCoordinate value in the case where seeking coordinate system;Each unit is accordingly to be regarded as one
Bin, for kth × l unit, statistics set Sk×lIn all tri patch area sum be ak×l, and it is regarded as son
Curved surfaceStatistic histogram hm×nThe bin of middle kth × l value;
Step V, the statistic histogram of all m × n subsurfaces around the p of summit is calculated, they are connected and composed one one by one
Individual higher-dimension histogram H'p, while calculating the area sum A of this m × n subsurfacep, use ApTo H'pAfter normalization, obtain
Histogram HpIt is exactly summit p SPIDER description.
Further technical scheme is that the step 2 includes:Convolutional neural networks learning card side is apart from χ2As similar
The measurement of test function is spent, SPIDER of the measurement with test point describes sub and specified human face characteristic point SPIDER and describe subtemplate
Similarity degreeWherein xiBe characteristic point SPIDER description submodule version in number in i-th of bin
Value, y is that the SPIDER of tested point describes the numerical value in son in i-th of bin.
Further technical scheme is that the step 2 includes:For each characteristic point, the minimum m of card side's distance is solved
It is individual, and the cluster centre of this m point is found using k nearest neighbor algorithms, it is the characteristic point of the three-dimensional face on object module.
Further technical scheme is that the step 2 includes:
Step c, the facial contour feature point and face inter characteristic points on son positioning three-dimensional face are described using SPIDER;
Step d, according to the quantity of human face characteristic point and requirement, sets the convolutional neural networks parameter of facial contour point, instruction
Practice and solve facial contour feature point coordinates;
Step e, according to the quantity of human face characteristic point and requirement, the convolutional neural networks of key point are joined inside setting face
Number, training solves crucial point coordinates inside face;
Step f, is the coordinate under rectangular coordinate system in space by the Coordinate Conversion under spherical coordinate system, that is, obtains space right-angle and sit
The crucial point coordinates of face under mark system, transfer process such as formula (6)
Further technical scheme is that the step d includes:
Step I, makes after face contour feature point coordinate vector, is expressed asUse master
Meta analysis obtains characteristic value collectionAnd solve D using formula (2)m;
Step II, defines a suitable DmaxIf, Dm≤DmaxThen stop iteration;{p1,p2,...,pfIt is exactly face
The result of contour feature point location;Otherwise updated one by one using formula (3)In numerical value obtainWillBe tied to one it is suitable
In the range of, wherein qkIt isIn k-th of element, q'kIt isIn k-th of element, i.e.,
UseWithReconstructA suitable threshold value δ is defined, according to formula (4)
Facial contour feature point position { p is updated one by one1,p2,...,pf};
Step III, defines initial search radius r0And reduced radius rate η, 1 > η > 0, the search radius of iteration j
It is expressed as rj, wherein rj=η × rj-1;By the updated each characteristic point { p of step II1,p2,...,pfAround, with rj
To search for respective characteristic point in the neighborhood of radius.
Further technical scheme is that the step d includes:
Loop iteration step I, step II, step III are until rjStop iteration during less than some threshold value ζ, { p now1,
p2,...,pfBe exactly facial contour feature point positioning result.
Compared with prior art, one of beneficial effect of the embodiment of the present invention is:The present invention can be in complex scene and bar
Under part, three-dimensional face positioning is set to reach gratifying result.
Brief description of the drawings
Fig. 1 is the method flow diagram of one embodiment of the invention.
Fig. 2 is the corresponding relation figure between one embodiment of the invention spatial coordinates system and spherical coordinate system.
Embodiment
All features disclosed in this specification, or disclosed all methods or during the step of, except mutually exclusive
Feature and/or step beyond, can combine in any way.
Any feature disclosed in this specification (including any accessory claim, summary and accompanying drawing), except non-specifically is chatted
State, can alternative features equivalent by other or with similar purpose replaced.I.e., unless specifically stated otherwise, each feature
It is an example in a series of equivalent or similar characteristics.
Below in conjunction with the accompanying drawings and embodiment to the present invention embodiment be described in detail.
As shown in figure 1, according to one embodiment of present invention, the present embodiment discloses a kind of based on concatenated convolutional neutral net
Three-dimensional face features' point location method.The present embodiment devise first it is a kind of based on curved surface divide, statistics with histogram three
Tie up local surface description --- ball divides histogram description, and describes the office that three-dimension curved surface summit is extracted in sub- pointwise using this
Portion's surface information, next calculates its similarity, to position the principal character point on three-dimensional face curved surface;Using different convolution god
The crucial point coordinates of inside key point and the exterior contour point of face is respectively calculated respectively through network model, then passed through
The mode of cascade connects above-mentioned network, then by the transformational relation between rectangular coordinate system in space and spherical coordinate system, obtains
The rectangular space coordinate of required three-dimensional face features' point.By concatenated convolutional neutral net can more precise and high efficiency acquisition three
Tie up the characteristic point of face.
Specifically, the present embodiment closes the coordinate under human face characteristic point rectangular space coordinate by the conversion between coordinate system
System, is transformed into the spherical coordinates under spherical coordinate system, and carrying out fan-shaped curved surface with the method for cartesian product divides, and carries out statistics with histogram
Three-dimensional local surface description is obtained, ball is obtained and divides histogram description.Respectively to facial contour feature point and face inside
Key point builds convolutional neural networks respectively, on the one hand can obtain more accurate facial feature estimation, on the other hand due to using
The mode of parallel cascade, training effectiveness is more efficient.
The convolutional neural networks learning card side is apart from χ2As the measurement of similarity examination function, weigh with test point
SPIDER describes the sub similarity degree that subtemplate is described with specified human face characteristic point SPIDERWherein xiIt is
Numerical value in the SPIDER of characteristic point description submodule version in i-th of bin, y is that the SPIDER of tested point is described in son i-th
Numerical value in bin.For each characteristic point, m minimum point of card side's distance is solved, and this m point is found using k nearest neighbor algorithms
Cluster centre, be the characteristic point p ' of the three-dimensional face on object module1,p′2,...,p′f。
Specifically, three-dimensional face features independent positioning method of the present embodiment based on concatenated convolutional neutral net includes following step
Suddenly:
Step 1. is pre-processed to three-dimensional face features' point of training data, obtains SPIDER description.
The characteristic point data that step 1.1 pair is used for the three-dimensional face positioning trained carries out coordinate system conversion, according to formula (1)
By human face characteristic point under rectangular coordinate system in space (x, y, z) Coordinate Conversion under spherical coordinate systemCoordinate.Space
Corresponding relation between coordinate system and spherical coordinate system is as shown in Figure 2.
The statistics with histogram of Subspace partition and subspace under step 1.2 spherical coordinate system
Step 1.2.1 three-dimensional faces curved surface is made up of a series of summit with tri patch, and the set of tri patch can
To be expressed as T, tiRepresent wherein i-th tri patch.For a summit p on curved surface, the local surface radius of influence is given
R, then the local surface pinpointed around p can be by tri patch set Ti={ ti|||ci- p | | < R, ciFor tiCenter of gravity represent.
Radius of influence R is equally divided into M parts, thus can be by TrIt is divided into M subsetM=1,2 ..., M and
Wherein ciFor tiCenter of gravity, m=1,2 ..., M.
Step 1.2.2 is divided into the fan-shaped of multiple sector regions using p as origin center, by polar coordinate system and divided.I.e. using p as original
Dot center, sets up polar coordinate system, wherein X-axis is polar coordinate system θ axles, and θ axles are divided into N parts on X/Y plane, can be by T points accordingly
N number of subset { Tn, n=1,2 ..., N, andFor tri patch ti, its center of gravity ciProjection c on X/Y planei'
It is represented by the coordinate value in polar coordinate systemN-th subset is defined as
The cartesian product that step 1.2.3 divides step 1.2.1 and step 1.2.2 curved surfaces can be by the office centered on p points
Portion's curved surface is divided into M × N number of subsurfaceWherein
Step 1.2.4 is for the m × n subsurfaceOrderOne of tri patch is represented, is used
RepresentNormal vector.It is that origin sets up spherical coordinate system with coordinate (0,0,0) in normal vector space, wherein three-dimensional right angle is sat
Coordinate Z axis is spherical coordinate system in mark systemAxle, X-axis is θ axles, normal directionCan be by seeking the coordinate value of coordinate systemTo represent;
In spherical coordinate system, half sphereThe span of axle isThe span of θ axles be [0,2 π);Point
Not willNormal vector space, can be divided into K × L unit by axle K deciles, θ axle L deciles.Normal vector be located exactly at kth ×
L if only dough sheet set expressions be
WhereinAnd θiIt is tri patch respectivelyNormal directionCoordinate value in the case where seeking coordinate system.Each unit is accordingly to be regarded as one
Bin, for kth × l unit, statistics set Sk×lIn all tri patch area sum be ak×l, and it is regarded as sub- song
FaceStatistic histogram hm×nThe bin of middle kth × l value.
Step 1.2.5 is normalized.The statistic histogram of all m × n subsurfaces around the p of summit is calculated, by them one by one
Connect and compose a higher-dimension histogram H'p, while calculating the area sum A of this m × n subsurfacep, use ApTo H'pNormalization
Afterwards, the histogram H obtainedpIt is exactly summit p SPIDER description.
Step 2. constructs concatenated convolutional neural network function, respectively to key point inside facial contour point and face, utilizes
The process of global restriction iterative calculates facial contour feature point and the three-dimensional coordinate of face inter characteristic points respectively.
The convolutional neural networks learning card side is apart from χ2As the measurement of similarity examination function, weigh with test point
SPIDER describes the sub similarity degree that subtemplate is described with specified human face characteristic point SPIDERWherein xiIt is
Numerical value in the SPIDER of characteristic point description submodule version in i-th of bin, y is that the SPIDER of tested point is described in son i-th
Numerical value in bin.For each characteristic point, m minimum point of card side's distance is solved, and this m point is found using k nearest neighbor algorithms
Cluster centre, be the characteristic point p ' of the three-dimensional face on object module1,p'2,...,p'f。
Step 2.1 describes facial contour feature point and face internal feature on son positioning three-dimensional face using SPIDER
Point.
Quantity and requirement of the step 2.2 according to human face characteristic point, set the convolutional neural networks parameter of facial contour point, instruction
Practice and solve facial contour feature point coordinates.
Step 2.2.1 makes after face contour feature point coordinate vector, is represented byUse master
Meta analysis can obtain characteristic value collectionAnd solve D using formula (2)m。
Step 2.2.2 defines a suitable DmaxIf, Dm≤DmaxThen stop iteration;{p1,p2,...,pfIt is exactly people
The result of face contour feature point location;Otherwise updated one by one using formula (3)In numerical value obtainWillOne is tied to fit
In the range of conjunction, wherein qkIt isIn k-th of element, q'kIt isIn k-th of element, i.e.,
UseWithReconstructA suitable threshold value δ is defined, according to formula (4)
Facial contour feature point position { p is updated one by one1,p2,...,pf}
Step 2.2.3 defines initial search radius r0And reduced radius rate η, 1 > η > 0, the search half of iteration j
Footpath is expressed as rj, wherein rj=η × rj-1.By each characteristic point { p updated step 2.2.21,p2,...,pfAround,
With rjTo search for respective characteristic point in the neighborhood of radius.Loop iteration step 2.2.1,2.2.2,2.2.3 is until rjIt is less than
Stop iteration during some threshold value ζ, { p now1,p2,...,pfBe exactly facial contour feature point positioning result.
The convolutional neural networks ginseng of key point inside quantity and requirement of the step 2.3 according to human face characteristic point, setting face
Number, the model parameter setting that the present invention is provided is as shown in table 2, and training solves crucial point coordinates inside face, solution procedure and
2.2.1~2.2.3 is similar, and the dimension of input vector need to be only adjusted according to the quantity of characteristic point.
Coordinate Conversion under spherical coordinate system is the coordinate under rectangular coordinate system in space by step 2.4, that is, obtains space right-angle
The crucial point coordinates of face under coordinate system, shown in transfer process such as formula (6).
The method of the present embodiment can reach gratifying result in complex scene with the conditions of.
" one embodiment ", " another embodiment ", " embodiment " for being spoken of in this manual etc., refers to combining
Specific features, structure or the feature of embodiment description are included at least one embodiment of the application generality description.
It is not necessarily to refer to same embodiment that statement of the same race, which occur, in multiple places in the description.Furthermore, it is understood that with reference to any
When individual embodiment describes a specific features, structure or feature, what is advocated is this to realize with reference to other embodiment
Feature, structure or feature are also fallen within the scope of the present invention.
Although reference be made herein to invention has been described for the multiple explanatory embodiments invented, however, it is to be understood that this
Art personnel can be designed that a lot of other modification and embodiment, and these modifications and embodiment will fall in the application
Within disclosed spirit and spirit.More specifically, can be to theme group in the range of disclosure claim
The building block and/or layout for closing layout carry out a variety of variations and modifications.Except the modification carried out to building block and/or layout
Outer with improving, to those skilled in the art, other purposes also will be apparent.
Claims (8)
1. a kind of three-dimensional face features' independent positioning method based on concatenated convolutional neutral net, it is characterised in that:Described method
Comprise the following steps:
Step 1: being pre-processed to three-dimensional face features' point of training data, SPIDER description are obtained;
Step 2: construction concatenated convolutional neural network function, respectively to key point inside facial contour point and face, utilizes the overall situation
The process that constraint iteration is solved calculates facial contour feature point and the three-dimensional coordinate of face inter characteristic points respectively.
2. three-dimensional face features' independent positioning method according to claim 1 based on concatenated convolutional neutral net, its feature
It is that described step one includes:
Step a, the characteristic point data to the three-dimensional face positioning for training carries out coordinate system conversion, according to formula (1) by face
Characteristic point under rectangular coordinate system in space (x, y, z) Coordinate Conversion under spherical coordinate systemCoordinate
The statistics with histogram of Subspace partition and subspace under step b, spherical coordinate system.
3. three-dimensional face features' independent positioning method according to claim 2 based on concatenated convolutional neutral net, its feature
It is that described step b includes:
Step I, three-dimensional face curved surface is made up of a series of summit with tri patch, and the set expression of tri patch is T, tiTable
Show wherein i-th tri patch;For a summit p on curved surface, local surface radius of influence R is given, then pinpoints the office around p
Portion's curved surface is by tri patch set Ti={ ti|||ci- p | | < R, ciFor tiCenter of gravity represent;Radius of influence R is equally divided into M parts,
By TrIt is divided into M subsetAndWherein
ciFor tiCenter of gravity, m=1,2 ..., M;
Step II, using p as origin center, is divided into the fan-shaped of multiple sector regions by polar coordinate system and divides;In i.e. using p as origin
The heart, sets up polar coordinate system, wherein X-axis is polar coordinate system θ axles, and θ axles are divided into N parts on X/Y plane, is N number of subset by T points
{Tn, n=1,2 ..., N, andFor tri patch ti, its center of gravity ciProjection c on X/Y planei' represent
For the coordinate value in polar coordinate systemN-th subset is defined as
Step III, cartesian product step I and step II curved surface divided, by the local surface centered on p points be divided into M ×
N number of subsurfaceWherein
Step IV, for the m × n subsurfaceOrderOne of tri patch is represented, is usedRepresent
Normal vector;It is that origin sets up spherical coordinate system with coordinate (0,0,0) in normal vector space, wherein in three-dimensional cartesian coordinate system
Coordinate Z axis is spherical coordinate systemAxle, X-axis is θ axles, normal directionBy the coordinate value for seeking coordinate systemTo represent in spherical coordinate system
In, half sphereThe span of axle isThe span of θ axles be [0,2 π);Respectively willAxle K deciles, θ axles L
Decile, K × L unit is divided into by normal vector space;Normal vector is located at the kth × l dough sheet set expressions wishedWhereinAnd θi
It is tri patch respectivelyNormal directionCoordinate value in the case where seeking coordinate system;Each unit is accordingly to be regarded as a bin, for
K × l unit, statistics set Sk×lIn all tri patch area sum be ak×l, and it is regarded as subsurfaceSystem
Count histogram hm×nThe bin of middle kth × l value;
Step V, the statistic histogram of all m × n subsurfaces around the p of summit is calculated, they are connected and composed one by one one high
Tie up histogram H'p, while calculating the area sum A of this m × n subsurfacep, use ApTo H'pAfter normalization, obtained Nogata
Scheme HpIt is exactly summit p SPIDER description.
4. three-dimensional face features' independent positioning method according to claim 1 based on concatenated convolutional neutral net, its feature
It is that described step two includes:Convolutional neural networks learning card side is apart from χ2As the measurement of similarity examination function, weigh
SPIDER with test point describes the sub similarity degree that subtemplate is described with specified human face characteristic point SPIDERWherein xiBe characteristic point SPIDER description submodule version in numerical value in i-th of bin, y is tested point
SPIDER description in numerical value in i-th of bin.
5. three-dimensional face features' independent positioning method according to claim 4 based on concatenated convolutional neutral net, its feature
It is that described step two includes:For each characteristic point, m minimum point of card side's distance is solved, and look for using k nearest neighbor algorithms
It is the characteristic point of the three-dimensional face on object module to the cluster centre of this m point.
6. three-dimensional face features' independent positioning method according to claim 1 based on concatenated convolutional neutral net, its feature
It is that described step two includes:
Step c, the facial contour feature point and face inter characteristic points on son positioning three-dimensional face are described using SPIDER;
Step d, according to the quantity of human face characteristic point and requirement, sets the convolutional neural networks parameter of facial contour point, training is asked
Solve facial contour feature point coordinates;
Step e, according to the quantity of human face characteristic point and requirement, the convolutional neural networks parameter of setting face inside key point, instruction
Practice and solve crucial point coordinates inside face;
Step f, is the coordinate under rectangular coordinate system in space by the Coordinate Conversion under spherical coordinate system, that is, obtains rectangular coordinate system in space
Under the crucial point coordinates of face, transfer process such as formula (6)
7. three-dimensional face features' independent positioning method according to claim 6 based on concatenated convolutional neutral net, its feature
It is that described step d includes:
Step I, makes after face contour feature point coordinate vector, is expressed asObtained using pivot analysis
Characteristic value collectionAnd solve D using formula (2)m;
Step II, defines a suitable DmaxIf, Dm≤DmaxThen stop iteration;{p1,p2,...,pfIt is exactly that facial contour is special
Levy the result of point location;Otherwise updated one by one using formula (3)In numerical value obtainWillIt is tied to a suitable scope
It is interior, wherein qkIt isIn k-th of element, q'kIt isIn k-th of element, i.e.,
UseWithReconstructA suitable threshold value δ is defined, according to formula (4)
Facial contour feature point position { p is updated one by one1,p2,...,pf};
Step III, defines initial search radius r0And reduced radius rate η, 1 > η > 0, the search radius of iteration j is expressed as
rj, wherein rj=η × rj-1;By the updated each characteristic point { p of step II1,p2,...,pfAround, with rjFor radius
The respective characteristic point of search in neighborhood.
8. three-dimensional face features' independent positioning method according to claim 7 based on concatenated convolutional neutral net, its feature
It is that described step d includes:
Loop iteration step I, step II, step III are until rjStop iteration during less than some threshold value ζ, { p now1,p2,...,
pfBe exactly facial contour feature point positioning result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611154383.3A CN106980812A (en) | 2016-12-14 | 2016-12-14 | Three-dimensional face features' independent positioning method based on concatenated convolutional neutral net |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611154383.3A CN106980812A (en) | 2016-12-14 | 2016-12-14 | Three-dimensional face features' independent positioning method based on concatenated convolutional neutral net |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106980812A true CN106980812A (en) | 2017-07-25 |
Family
ID=59341168
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611154383.3A Pending CN106980812A (en) | 2016-12-14 | 2016-12-14 | Three-dimensional face features' independent positioning method based on concatenated convolutional neutral net |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106980812A (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107704813A (en) * | 2017-09-19 | 2018-02-16 | 北京飞搜科技有限公司 | A kind of face vivo identification method and system |
CN107808129A (en) * | 2017-10-17 | 2018-03-16 | 南京理工大学 | A kind of facial multi-characteristic points localization method based on single convolutional neural networks |
CN109754467A (en) * | 2018-12-18 | 2019-05-14 | 广州市百果园网络科技有限公司 | Three-dimensional face construction method, computer storage medium and computer equipment |
CN109902616A (en) * | 2019-02-25 | 2019-06-18 | 清华大学 | Face three-dimensional feature point detecting method and system based on deep learning |
CN110059707A (en) * | 2019-04-25 | 2019-07-26 | 北京小米移动软件有限公司 | Optimization method, device and the equipment of image characteristic point |
CN110232438A (en) * | 2019-06-06 | 2019-09-13 | 北京致远慧图科技有限公司 | The image processing method and device of convolutional neural networks under a kind of polar coordinate system |
CN110310287A (en) * | 2018-03-22 | 2019-10-08 | 北京连心医疗科技有限公司 | It is neural network based to jeopardize the automatic delineation method of organ, equipment and storage medium |
CN115512074A (en) * | 2022-09-21 | 2022-12-23 | 中冶南方工程技术有限公司 | Blast furnace slag scale distribution visualization method, terminal equipment and storage medium |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101976359A (en) * | 2010-09-26 | 2011-02-16 | 浙江大学 | Method for automatically positioning characteristic points of three-dimensional face |
CN103824049A (en) * | 2014-02-17 | 2014-05-28 | 北京旷视科技有限公司 | Cascaded neural network-based face key point detection method |
US20150347822A1 (en) * | 2014-05-29 | 2015-12-03 | Beijing Kuangshi Technology Co., Ltd. | Facial Landmark Localization Using Coarse-to-Fine Cascaded Neural Networks |
-
2016
- 2016-12-14 CN CN201611154383.3A patent/CN106980812A/en active Pending
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101976359A (en) * | 2010-09-26 | 2011-02-16 | 浙江大学 | Method for automatically positioning characteristic points of three-dimensional face |
CN103824049A (en) * | 2014-02-17 | 2014-05-28 | 北京旷视科技有限公司 | Cascaded neural network-based face key point detection method |
US20150347822A1 (en) * | 2014-05-29 | 2015-12-03 | Beijing Kuangshi Technology Co., Ltd. | Facial Landmark Localization Using Coarse-to-Fine Cascaded Neural Networks |
CN105981041A (en) * | 2014-05-29 | 2016-09-28 | 北京旷视科技有限公司 | Facial landmark localization using coarse-to-fine cascaded neural networks |
Non-Patent Citations (1)
Title |
---|
张晓博 等: "实时三维人脸特征点定位", 《计算机辅助设计与图形学学报》 * |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107704813A (en) * | 2017-09-19 | 2018-02-16 | 北京飞搜科技有限公司 | A kind of face vivo identification method and system |
CN107704813B (en) * | 2017-09-19 | 2020-11-17 | 北京一维大成科技有限公司 | Face living body identification method and system |
CN107808129A (en) * | 2017-10-17 | 2018-03-16 | 南京理工大学 | A kind of facial multi-characteristic points localization method based on single convolutional neural networks |
CN107808129B (en) * | 2017-10-17 | 2021-04-16 | 南京理工大学 | Face multi-feature point positioning method based on single convolutional neural network |
CN110310287A (en) * | 2018-03-22 | 2019-10-08 | 北京连心医疗科技有限公司 | It is neural network based to jeopardize the automatic delineation method of organ, equipment and storage medium |
CN110310287B (en) * | 2018-03-22 | 2022-04-19 | 北京连心医疗科技有限公司 | Automatic organ-at-risk delineation method, equipment and storage medium based on neural network |
CN109754467A (en) * | 2018-12-18 | 2019-05-14 | 广州市百果园网络科技有限公司 | Three-dimensional face construction method, computer storage medium and computer equipment |
CN109754467B (en) * | 2018-12-18 | 2023-09-22 | 广州市百果园网络科技有限公司 | Three-dimensional face construction method, computer storage medium and computer equipment |
CN109902616A (en) * | 2019-02-25 | 2019-06-18 | 清华大学 | Face three-dimensional feature point detecting method and system based on deep learning |
CN109902616B (en) * | 2019-02-25 | 2020-12-01 | 清华大学 | Human face three-dimensional feature point detection method and system based on deep learning |
CN110059707A (en) * | 2019-04-25 | 2019-07-26 | 北京小米移动软件有限公司 | Optimization method, device and the equipment of image characteristic point |
CN110232438A (en) * | 2019-06-06 | 2019-09-13 | 北京致远慧图科技有限公司 | The image processing method and device of convolutional neural networks under a kind of polar coordinate system |
CN115512074A (en) * | 2022-09-21 | 2022-12-23 | 中冶南方工程技术有限公司 | Blast furnace slag scale distribution visualization method, terminal equipment and storage medium |
CN115512074B (en) * | 2022-09-21 | 2024-02-23 | 中冶南方工程技术有限公司 | Blast furnace slag skin distribution visualization method, terminal equipment and storage medium |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106980812A (en) | Three-dimensional face features' independent positioning method based on concatenated convolutional neutral net | |
CN107657226B (en) | People number estimation method based on deep learning | |
CN107203785A (en) | Multipath Gaussian kernel Fuzzy c-Means Clustering Algorithm | |
CN108920678A (en) | A kind of overlapping community discovery method based on spectral clustering with fuzzy set | |
CN107871106A (en) | Face detection method and device | |
CN105740891B (en) | Target detection based on multi level feature selection and context model | |
CN109614935A (en) | Car damage identification method and device, storage medium and electronic equipment | |
CN107967463B (en) | Virtual face recognition method based on synthetic image and deep learning | |
CN109191455A (en) | A kind of field crop pest and disease disasters detection method based on SSD convolutional network | |
CN111046767B (en) | 3D target detection method based on monocular image | |
CN104504394A (en) | Dese population estimation method and system based on multi-feature fusion | |
CN113205595B (en) | Construction method and application of 3D human body posture estimation model | |
CN109446889A (en) | Object tracking method and device based on twin matching network | |
CN109558902A (en) | A kind of fast target detection method | |
CN104657737B (en) | The method and apparatus being corrected based on cluster to QR image in 2 D code | |
CN107871105A (en) | Face authentication method and device | |
CN110298831A (en) | A kind of magic magiscan and its method based on piecemeal deep learning | |
CN104657986B (en) | A kind of quasi- dense matching extended method merged based on subspace with consistency constraint | |
CN108537145A (en) | Human bodys' response method based on space-time skeleton character and depth belief network | |
CN111027140A (en) | Airplane standard part model rapid reconstruction method based on multi-view point cloud data | |
CN109859101A (en) | The recognition methods of corps canopy thermal infrared images and system | |
CN109920050A (en) | A kind of single-view three-dimensional flame method for reconstructing based on deep learning and thin plate spline | |
CN111488498A (en) | Node-graph cross-layer graph matching method and system based on graph neural network | |
CN103700104B (en) | Method for automatically segmenting human lateral geniculate nucleus (LGN) by using prior knowledge | |
CN103345741A (en) | Non-rigid multimode medical image precise registering method |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170725 |
|
RJ01 | Rejection of invention patent application after publication |