CN106096503A - A kind of based on key point with the three-dimensional face identification method of local feature - Google Patents

Abstract

The invention discloses a kind of based on key point with the three-dimensional face identification method of local feature, step is as follows: three-dimensional face model is carried out pretreatment, including human face region cutting, smoothing processing and attitude normalization, is placed under attitude frame of reference by all of face；According to valuable contour line and average curvature measuring key point；The space structure of the formal construction local feature according to DAISY descriptor；Use shape index rectangular histogram, inclination angle rectangular histogram and deflection rectangular histogram as local feature；Carry out key point coupling, and weigh the similarity of two face curved surfaces with the key point number that the match is successful.Recognition performance of the present invention is relatively strong, and has certain robustness to expression shape change.

Description

A kind of based on key point with the three-dimensional face identification method of local feature

Technical field

The invention belongs to technical field of face recognition, be specifically related to a kind of based on key point with the three-dimensional face of local feature Recognition methods, particularly relates to a kind of face identification method utilizing face face key point local feature, is particularly suited for expression The occasion of large change.

Background technology

In modern society, accurately and reliably personal authentication is more and more subject to people's attention, and authentication exists Many occasions suffer from important application, such as gate control system, video monitoring, man-machine interaction etc..Identification technology mainly include based on The method of the human body biological characteristics such as fingerprint, iris, face, wherein authentication based on fingerprint and iris have the highest accurate Property and reliability, these features all have a uniqueness, but are limited owing to needing the factors such as user's cooperation to make it apply when using System, and recognition of face has broader practice prospect because it has the advantage such as nature, close friend.People began to grind from nineteen sixty Studying carefully Automatic face recognition technology, traditional two-dimension human face identification is that monochrome information based on image carries out identity validation, but because of It is relatively big for being affected by factors such as illumination, attitude and cosmetics, so the research of recognition of face in recent years starts to expand to from two dimensional image On three dimensions.

Three-dimensional face identification utilizes the spatial form information of face to carry out authentication, not by the factor such as illumination, cosmetic Impact, although compare with two-dimension human face identification and have bigger advantage, but still suffer from following both sides challenge: one, expression Change, expression can cause face to produce non-rigid shape deformations, and the difference that recognizer is difficult to differentiate between two faces is by expressing one's feelings The still different people that change causes causes.Two, block, block and the three-dimensional data of face can be caused to lose, affect recognition effect.

Summary of the invention

Goal of the invention: in order to overcome the deficiencies in the prior art, the present invention provides one based on key point and local The three-dimensional face identification method of feature, the method method recognition performance is higher, and has certain robustness to expression shape change.

Technical scheme: for achieving the above object, the technical solution used in the present invention is:

A kind of based on key point with the three-dimensional face identification method of local feature, comprise the following steps:

Step 1: three-dimensional face model is carried out pretreatment, including human face region cutting, smoothing processing and attitude normalization, All of face is placed under attitude frame of reference, constitutes the three-dimensional face point cloud of face curved surface；

Step 2: extract the central area of face, extract valuable contour line, to clicking on described valuable contour line Row uniform sampling, obtains crucial point set P1；

Step 3: screen according to the average curvature of each point and neighborhood thereof and obtain crucial point set P2；

Step 4: according to the space structure of the formal construction local feature of DAISY descriptor；

Step 5: calculate key point and the shape index rectangular histogram of neighborhood, inclination angle rectangular histogram and deflection rectangular histogram are made For local feature；

Step 6: mate the key point on two face curved surfaces, obtains the key point logarithm that the match is successful, and this number Foundation as authentication.

Further, described step 2 method particularly includes: with nose for centre of sphere 90mm from the three-dimensional face point cloud of input Extract human face region for radius, the three-dimensional face point cloud extracted is carried out gridding operation, use based on grid smooth Algorithm carries out smoothing denoising process, the smooth three-dimensional face grid that then will obtain through iterative processing to three-dimensional face model Revert to three-dimensional face point cloud.

Further, described step 2 specifically includes following steps:

Step 2.1: extracting face central area, be with prenasale as the centre of sphere, r=50mm is the region of radius；

Step 2.2: detection facial symmetry face, uses central area face mirror image matching method to determine facial symmetry face；

Step 2.3: extract valuable contour line and obtain crucial point set P1, method particularly includes: described facial symmetry face and people The intersection of face curved surface is central lateral plane contour line, according to the cut-off position of central lateral plane contour line described in nose Y-axis coordinate selection Putting, the nasal portion of extraction central lateral plane contour line, as valuable contour line, selects by uniform in described valuable contour line The crucial point set P1 of some composition after sampling.

Further, in described step 2.2, face mirror image matching method determines specifically comprising the following steps that of facial symmetry face

Step 2.2.1: central area curved surface is F, the initial plane of symmetry is S, determines that the initial plane of symmetry of described S is YOZ plane；

Step 2.2.2: the x-axis coordinate of described central area curved surface F is reverse, obtains the mirror image mould for this symmetrical plane Type；Described mirror image model and former input model being alignd, this step uses ICP matching process；

Step 2.2.3: extract the plane of symmetry, after specially two models alignment, take defeated as attitude updating of both meansigma methodss Enter model, and with PCA process, extract three orthogonal major axes orientations of face, by corresponding for first, second and third main constituent The principal direction YOZ plane respectively as Y-axis, X-axis, Z axis, under the plane of symmetry the most now coordinate.

Further, described step 3 specifically includes following steps:

Step 3.1: calculate the average curvature of and neighborhood at every: for summit p, its average curvature C_meanP () is by maximum Principal curvatures C_M(p) and minimum principal curvatures C_mP () is calculated, such as following formula:

C_mean(p)=(C_M(p)+C_m(p))/2 (1)

If its neighborhood is N (p), then average curvature μ (p) of its neighborhood is calculated as follows:

$\mathrm{\μ}\left(p\right)=\frac{1}{N}\underset{q\∈N\left(p\right)}{\Σ}{C}_{mean}\left(q\right)---\left(2\right)$

Step 3.2: screening key point: if the average curvature on one summit of definition meets following formula, then it is assumed that it is crucial Point, screening the key point obtained obtains crucial point set P2:

|C_mean(p)|≥(1+a)|μ(p)|or|C_mean(p)|≤(1-b)|μ(p)|,b≤1,a≥0 (3)

Wherein, a and b is to limit | C_mean(p) | and two parameters of | μ (p) | difference size, the most also it is used for selecting key Point.

Further, described step 4 specifically includes following steps:

Step 4.1: the establishment in standardization direction: for key point p on face curved surface, its neighborhood N (p) is defined as follows:

N (p)=q | d_g(p, q) ＜ R} (4)

Wherein, d_g(p, q) is the p point geodesic distance to q point, and R is radius；

Step 4.1.1: by all of some q ∈ N (p) and normal vector n (q) thereof, is transformed in following interim local coordinate system:

C={t (p'), t (p') × n (p'), n (p') } (5)

Wherein, p' be conversion after p, unit normal vector n (p') be Z axis, t (p') is the face curved surface incisal plane at p' point The unit vector randomly choosed in T (p')；

Step 4.1.2: the q' after conversion and their normal vector n (q') is mapped to incisal plane T (p'), corresponding angle θ (q') and the long mag of mould (q') be:

$\begin{array}{c}\mathrm{\θ}\left(q^{\′}\right)=\arctan \[n_y\left(q^{\′}\right)/n_x\left(q^{\′}\right)\]\\ mag\left(q^{\′}\right)=\sqrt{n_x^2\left(q^{\′}\right)+n_y^2\left(q^{\′}\right)}\end{array}---\left(6\right)$

Wherein, n_x(q')=t (p') n (q'), n_y(q')=t (p') × n (p') n (q')；

Step 4.1.3: one the Gauss weight histogram of gradients of upper structure of the incisal plane T (p') under temporary coordinate system, wide Being 360, each unit covers 1 °, and described Gauss weight is described as follows:

W (p', q')=mag (q') G_σ(d_g(p',q')) (7)

Wherein, standard deviation sigma is set to the half of radius R；

Step 4.1.4: using the peak value of described Gauss weight histogram of gradients as standardization direction d (p') at p', adjacent In territory be the most all transformed under new coordinate system:

C'={d (p'), d (p') × n (p'), n (p') } (8)

Step 4.2: descriptor structure: the spatial configuration of descriptor is that the incisal plane T (p') under new coordinate system C' is enterprising Row, with key point p' and its 8 neighborhood points as the center of circle, obtain 9 circles, and by sequence counterclockwise, described 8 neighborhood points by At standardization direction, start uniform sampling for the center of circle with key point p' to obtain.

Further, described step 5 specifically includes following steps:

Step 5.1: calculate the local feature of each circle shaped neighborhood region of key point p: the local feature of each border circular areas is by three Part is constituted:

1) shape index rectangular histogram h_s, a length of 8, shape index is calculated by following formula:

$SI\left(q\right)=\frac{1}{2}-\frac{1}{\mathrm{\π}}{\tan }^{-1}\frac{C_M\left(q\right)+C_m\left(q\right)}{C_M\left(q\right)-C_m\left(q\right)}---\left(9\right)$

Wherein q ∈ N (p), SI (q) are with a gaussian kernel as weight, and its standard variance is the q geodesic distance to the center of circle.

2) inclination angle rectangular histogram h_a, a length of 8, tilt angle alpha is defined as every some normal vector n (q) and standardization direction d (p') angle:

$\mathrm{\α}\left(q\right)={\cos }^{-1}\frac{n\left(q\right)\·d\left(p\right)}{\left|n\left(q\right)\right|\×\left|d\left(p\right)\right|}---\left(10\right)$

Wherein q ∈ N (p), α (q) are with the angle between its normal vector and normal vector of key point As weight；

3) deflection rectangular histogram h_d, a length of 8, direction β are defined as any two points q in neighborhood₁And q₂Between line l (q₁,q₂) and key point normal vector n (p) between angle:

$\mathrm{\β}(q_1,q_2)={\cos }^{-1}\frac{l(q_1,q_2)\·n\left(p\right)}{\left|l(q_1,q_2)\right|\×\left|n\left(p\right)\right|}---\left(11\right)$

It is normalized to unit vector after each rectangular histogram described above is all normalized to unit vector, and series connection obtain Feature h of the n-th circle_n:

h_n=(h_sn,h_an,h_dn) (12)

Step 5.2: calculate the local feature of key point p: connect feature h of each circle_nObtain the local feature of key point p F:

F=(h₁,h₂,…,h₉) (13)。

Further, described step 6 specifically includes following steps:

Step 6.1: calculate angle γ: set f_iAnd f_jIt is respectively on storehouse collection and test set face curved surface i-th, j key point Feature, γ is calculated as follows:

${\mathrm{\&gamma;}}_{i,j}={\cos }^{-1}\left(\frac{<f_i,f_j>}{||f_i||\&CenterDot;||f_j||}\right)---\left(14\right)$

Step 6.2: key point is mated, and concrete grammar is to each key point i in test set face curved surface, calculates itself and storehouse Collect the angle γ between each key point j on face curved surface_i,j, by angle γ_i,jSort in descending order, if first and second Ratio less than 0.8, then coupling can be accepted, and otherwise it fails to match；

Step 6.3: recognition methods, concrete grammar is that test set face curved surface concentrates each face curved surface to carry out key with storehouse Point matching, test set with and the most face curved surface of its key point number that the match is successful and for same class.

Beneficial effect: the present invention provide based on key point and the three-dimensional face identification method of local feature, according to there being valency Value contour line and average curvature measuring key point, according to the space structure of the formal construction local feature of DAISY descriptor, use Shape index rectangular histogram, inclination angle rectangular histogram and deflection rectangular histogram, as local feature, carry out key point coupling, and with mating Successfully key point number weighs the similarity of two face curved surfaces.The method has following several advantage.

1) key point that this algorithm obtains is stable and representative；

2) space structure of key point local feature, not only has rotational invariance, meanwhile, it is capable to sufficiently give expression to neighbour Contact between territory；

3) key point local feature contains shape index rectangular histogram, inclination angle rectangular histogram and deflection rectangular histogram, fully The geological information illustrating face curved surface, there is good separability.

Accompanying drawing explanation

Fig. 1 be the present invention provide based on half key point and the overall flow of the three-dimensional face identification method of local feature Figure；

Fig. 2 is original faceform；

Fig. 3 is face central area；

Fig. 4 is the face central lateral plane contour line extracted；

Fig. 5 is valuable contour line；

Fig. 6 is crucial point set P1；

Fig. 7 is crucial point set P2；

Fig. 8 is key point；

Fig. 9 is DAISY descriptor space structure.

Detailed description of the invention

Below in conjunction with the accompanying drawings the present invention is further described.

Embodiment 1

The present invention based on key point and the three-dimensional face identification method of local feature, logical in Windows operating system Cross Matlab R2015b programming tool and realize three-dimensional face identification process.Experimental data is from FRGC v2.0 three-dimensional face data Storehouse, comprises 4007 faceforms for 466 people tested in this data base.

Fig. 1 is the inventive method overall flow figure, specifically comprises the following steps that

Step 1: extract human face region for centre of sphere 90mm for radius with nose from the three-dimensional face point cloud of input.To institute The three-dimensional face point cloud extracted carries out gridding operation, uses smoothing algorithm based on grid to smooth three-dimensional face model Denoising, then reverts to three-dimensional face point cloud by the smooth three-dimensional face grid obtained through iterative processing.

Step 2: with prenasale as the centre of sphere, 50mm is the central area that radius extracts face, utilizes face mirror image matching method Extract the central lateral plane contour line of face, thus obtain valuable contour line, the point on valuable contour line is uniformly adopted Sample, obtains crucial point set P1, specifically comprises the following steps that

Step 2.1: extracting face central area, central area is with nose as the centre of sphere, and r=50mm is the district of radius Territory.

Step 2.2: detection facial symmetry face, uses central area face mirror image matching method to determine facial symmetry face.In if Heart region curved surface is F, and the initial plane of symmetry is S, and S elects YOZ plane as herein, and calculation procedure is as follows:

(1) initial symmetrical plane is determined.ICP method for registering has dependency for initial condition, thus initially symmetrical Plane choose the order of accuarcy determining final symmetrical plane to a certain extent.Initial plane we select smoothed denoising to cut The YOZ plane of rear human face data.

(2) x-axis coordinate is reversely obtained the mirror image model for this symmetrical plane；By mirror image model and former input model pair Together.This step uses ICP matching process.

(3) plane of symmetry is extracted.After two model alignment, take both meansigma methodss input model as attitude updating, and use PCA process, extracts three orthogonal major axes orientations of face.Principal direction corresponding for first, second and third main constituent is made respectively For Y-axis, X-axis, Z axis.The most different faces will have approximately uniform frontal pose under this coordinate system.The plane of symmetry is the most now YOZ plane under coordinate.

Step 2.3: extract valuable contour line and obtain crucial point set P1: facial symmetry face is with the intersection of face curved surface Central lateral plane contour line.According to the rest position of nose Y-axis coordinate selection contour line, extract the nose portion of central lateral plane contour line It is allocated as valuable contour line, selects form key point set P1 by the point after uniform sampling in valuable contour line.

Step 3: screen according to the average curvature of each point and neighborhood thereof and obtain crucial point set P2.Specifically comprise the following steps that

Step 3.1: calculate the average curvature of and neighborhood at every: for summit p, its average curvature C_meanP () is by maximum Main

Curvature C_M(p) and minimum principal curvatures C_mP () is calculated, such as following formula:

C_mean(p)=(C_M(p)+C_m(p))/2 (1)

If its neighborhood is N (p), then average curvature μ (p) of its neighborhood is calculated as follows:

$\mathrm{\&mu;}\left(p\right)=\frac{1}{N}\underset{q\&Element;N\left(p\right)}{\&Sigma;}{C}_{mean}\left(q\right)---\left(2\right)$

Step 3.2: screening key point: if our average curvature on one summit of definition meets following formula, then it is assumed that it is Key point, screening the key point obtained obtains crucial point set P2.

|C_mean(p)|≥(1+a)|μ(p)|or|C_mean(p)|≤(1-b)|μ(p)|,b≤1,a≥0 (3)

Wherein, a and b is to limit | C_mean(p) | and two parameters of | μ (p) | difference size, the most also it is used for selecting key Point.

Step 4: according to the space structure of the formal construction local feature of DAISY descriptor, specifically comprise the following steps that

Step 4.1: the establishment in standardization direction.In order to realize the rotational invariance of DAISY descriptor, need to establish specification Change direction.For key point p on face curved surface, its neighborhood N (p) is defined as follows:

N (p)=q | d_g(p, q) ＜ R} (4)

Wherein, d_g(p, q) is the p point geodesic distance to q point, and R is radius.First, by all of some q ∈ N (p) and method thereof Vector n (q), is transformed in following interim local coordinate system:

C={t (p'), t (p') × n (p'), n (p') } (5)

Wherein, p' be conversion after p, unit normal vector n (p') be Z axis, t (p') is the face curved surface incisal plane at p' point The unit vector randomly choosed in T (p').Then, the q' after conversion and their normal vector n (q') is mapped to incisal plane T (p'), corresponding angle theta (q') and the long mag of mould (q') are:

θ (q')=arctan [n_y(q')/n_x(q')]

$mag\left(q^{\&prime;}\right)=\sqrt{n_x^2\left(q^{\&prime;}\right)+n_y^2\left(q^{\&prime;}\right)}---\left(6\right)$

Wherein, n_x(q')=t (p') n (q'), n_y(q')=t (p') × n (p') n (q').Finally, at temporal coordinate Incisal plane T (p') under Xi one Gauss weight histogram of gradients of upper structure, a width of 360, each unit covers 1 °.Gauss weighs Heavily it is described as follows:

W (p', q')=mag (q') G_σ(d_g(p',q')) (7)

In above formula, standard deviation sigma is set to the half of radius R.Selection multiple directions are as standardization direction herein, with Gauss The peak value of weight histogram of gradients is as standardization direction d (p') at p'.After d (p') determines, the institute in neighborhood is the most all It is transformed under new coordinate system:

C'={d (p'), d (p') × n (p'), n (p') } (8)

Step 4.2: descriptor structure: the spatial configuration of descriptor is that the incisal plane T (p') under new coordinate system C' is enterprising Row, with key point p' and its 8 neighborhood points as the center of circle, obtain 9 circles, and by sequence counterclockwise, as shown in fig. 1.8 Neighborhood point by with key point p' for the center of circle from standardization direction at uniform sampling obtain.

Step 5: calculate key point and the shape index rectangular histogram of neighborhood, inclination angle rectangular histogram and deflection rectangular histogram are made For local feature, specifically comprise the following steps that

Step 5.1: calculate the local feature of each circle shaped neighborhood region of key point p: the local feature of each border circular areas is by three Part is constituted:

1) shape index rectangular histogram h_s, a length of 8, shape index is calculated by following formula:

$SI\left(q\right)=\frac{1}{2}-\frac{1}{\mathrm{\&pi;}}{\tan }^{-1}\frac{C_M\left(q\right)+C_m\left(q\right)}{C_M\left(q\right)-C_m\left(q\right)}---\left(9\right)$

Wherein q ∈ N (p), SI (q) are with a gaussian kernel as weight, and its standard variance is the q geodesic distance to the center of circle.

2) inclination angle rectangular histogram h_a, a length of 8, tilt angle alpha is defined as every some normal vector n (q) and standardization direction d (p') angle:

$\mathrm{\&alpha;}\left(q\right)={\cos }^{-1}\frac{n\left(q\right)\&CenterDot;d\left(p\right)}{\left|n\left(q\right)\right|\&times;\left|d\left(p\right)\right|}---\left(10\right)$

Wherein q ∈ N (p), α (q) are with the angle between its normal vector and normal vector of key point As weight.

3) deflection rectangular histogram h_d, a length of 8, direction β are defined as any two points q in neighborhood₁And q₂Between line l (q₁,q₂) and key point normal vector n (p) between angle:

$\mathrm{\&beta;}(q_1,q_2)={\cos }^{-1}\frac{l(q_1,q_2)\&CenterDot;n\left(p\right)}{\left|l(q_1,q_2)\right|\&times;\left|n\left(p\right)\right|}---\left(11\right)$

It is normalized to unit vector after each rectangular histogram described above is all normalized to unit vector, and series connection obtain Feature h of the n-th circle_n:

h_n=(h_sn,h_an,h_dn) (12)

Step 5.2: calculate the local feature of key point p: connect feature h of each circle_nObtain the local feature of key point p F:

F=(h₁,h₂,…,h₉) (13)

Step 6: mate the key point on two face curved surfaces, obtains the key point logarithm that the match is successful, and this number As the foundation of authentication, specifically comprise the following steps that

Step 6.1: calculate angle γ: set f_iAnd f_jIt is respectively on storehouse collection and test set face curved surface i-th, j key point Feature, γ is calculated as follows:

${\mathrm{\&gamma;}}_{i,j}={\cos }^{-1}\left(\frac{<f_i,f_j>}{||f_i||\&CenterDot;||f_j||}\right)---\left(14\right)$

Step 6.2: key point is mated.To each key point i in test set face curved surface, calculate itself and storehouse collection face curved surface Angle γ between upper each key point j_i,j, the angle obtained is sorted in descending order, if the ratio of first and second is little In certain 0.8, then coupling can be accepted, and otherwise it fails to match.

Step 6.3: recognition methods.Test set face curved surface carries out key point with the storehouse each face curved surface of concentration and mates, and surveys Examination integrate with and the most face curved surface of its key point number that the match is successful and as same class.

In above method, storehouse integrates face as processed offline pattern, and test face is online treatment pattern.

Embodiment 2

The method using embodiment 1, carries out experimental verification.Specifically include following steps:

Step 7: authentication identification is tested, experiment all uses R1RR (Rank-one Recognition Rate) conduct Recognition performance index.

Step 7.1: experiment one, this experiment uses FRGC v2.0 data base, and this storehouse acquires 4 007 of 466 objects Face point cloud, comprises smile, the face of the band such as startled, angry expression, experiment choose everyone first as storehouse collection, remaining As test set, obtain the discrimination of 96.9%.

Step 7.2: experiment two, this experiment acquires 105 objects based on Bosphorus data base, this data base 4666 face point clouds, abundant species of wherein expressing one's feelings, and expression amplitude is bigger.This experiment uses 194 Nature faces as storehouse Collection, espressiove face is as test set.Experiment as test set, is surveyed using Nature face as storehouse collection, the face of different expressions Examination collection expression for angry, detest, fear, glad, sad, surprised time, R1RR is respectively 94.4%, 89.9%, 95.7%, 96.9%, 98.5%, 98.6%, during using institute's espressiove face as test set, R1RR is 95.8%, it follows that in expression In the case of relatively big, the algorithm that the present invention proposes remains to complete well to identify.

The above is only the preferred embodiment of the present invention, it should be pointed out that: for the ordinary skill people of the art For Yuan, under the premise without departing from the principles of the invention, it is also possible to make some improvements and modifications, these improvements and modifications also should It is considered as protection scope of the present invention.

Claims (8)

1. one kind based on key point and the three-dimensional face identification method of local feature, it is characterised in that: comprise the following steps:

Step 1: three-dimensional face model is carried out pretreatment, including human face region cutting, smoothing processing and attitude normalization, by institute Somebody's face is placed under attitude frame of reference, constitutes the three-dimensional face point cloud of face curved surface；

Step 2: extract the central area of face, extract valuable contour line, the point on described valuable contour line is carried out all Even sampling, obtains crucial point set P1；

Step 3: screen according to the average curvature of each point and neighborhood thereof and obtain crucial point set P2；

Step 4: according to the space structure of the formal construction local feature of DAISY descriptor；

Step 5: calculate key point and the shape index rectangular histogram of neighborhood, inclination angle rectangular histogram and deflection rectangular histogram as office Portion's feature；

Step 6: mate the key point on two face curved surfaces, obtains the key point logarithm that the match is successful, and using this number as The foundation of authentication.

It is the most according to claim 1 based on key point with the three-dimensional face identification method of local feature, it is characterised in that: institute State step 2 method particularly includes: from the three-dimensional face point cloud of input, extract face district for centre of sphere 90mm for radius with nose Territory, carries out gridding operation to the three-dimensional face point cloud extracted, uses smoothing algorithm based on grid to three-dimensional face model Carry out smoothing denoising process, then the smooth three-dimensional face grid obtained through iterative processing is reverted to three-dimensional face point Cloud.

It is the most according to claim 1 based on key point with the three-dimensional face identification method of local feature, it is characterised in that: institute State step 2 and specifically include following steps:

Step 2.1: extracting face central area, be with prenasale as the centre of sphere, r=50mm is the region of radius；

Step 2.2: detection facial symmetry face, uses central area face mirror image matching method to determine facial symmetry face；

Step 2.3: extract valuable contour line and obtain crucial point set P1, method particularly includes: described facial symmetry face is bent with face The intersection in face is central lateral plane contour line, according to the rest position of central lateral plane contour line described in nose Y-axis coordinate selection, carries Take the nasal portion of central lateral plane contour line as valuable contour line, after selecting by uniform sampling in described valuable contour line The point crucial point set P1 of composition.

It is the most according to claim 3 based on key point with the three-dimensional face identification method of local feature, it is characterised in that: institute State face mirror image matching method in step 2.2 and determine specifically comprising the following steps that of facial symmetry face

Step 2.2.1: central area curved surface is F, the initial plane of symmetry is S, determines that the initial plane of symmetry of described S is YOZ plane；

Step 2.2.2: the x-axis coordinate of described central area curved surface F is reverse, obtains the mirror image model for this symmetrical plane； Described mirror image model and former input model being alignd, this step uses ICP matching process；

Step 2.2.3: extract the plane of symmetry, after specially two model alignment, take both meansigma methodss input mould as attitude updating Type, and with PCA process, extract three orthogonal major axes orientations of face, by main formula corresponding for first, second and third main constituent To the YOZ plane respectively as Y-axis, X-axis, Z axis, under the plane of symmetry the most now coordinate.

It is the most according to claim 1 based on key point with the three-dimensional face identification method of local feature, it is characterised in that: institute State step 3 and specifically include following steps:

Step 3.1: calculate the average curvature of and neighborhood at every: for summit p, its average curvature C_meanP () is by maximum principal curvatures C_M(p) and minimum principal curvatures C_mP () is calculated, such as following formula:

C_mean(p)=(C_M(p)+C_m(p))/2 (1)

If its neighborhood is N (p), then average curvature μ (p) of its neighborhood is calculated as follows:

$\mathrm{\&mu;}\left(p\right)=\frac{1}{N}\underset{q\&Element;N\left(p\right)}{\&Sigma;}{C}_{mean}\left(q\right)---\left(2\right)$

Step 3.2: screening key point: if the average curvature on one summit of definition meets following formula, then it is assumed that it is key point, by The key point that obtains of screening obtains crucial point set P2:

|C_mean(p)|≥(1+a)|μ(p)|or|C_mean(p)|≤(1-b)|μ(p)|,b≤1,a≥0 (3)

Wherein, a and b is to limit | C_mean(p) | and two parameters of | μ (p) | difference size, the most also it is used for selecting key point.

It is the most according to claim 1 based on key point with the three-dimensional face identification method of local feature, it is characterised in that: institute State step 4 and specifically include following steps:

Step 4.1: the establishment in standardization direction: for key point p on face curved surface, its neighborhood N (p) is defined as follows:

N (p)=q | d_g(p, q) ＜ R} (4)

Wherein, d_g(p, q) is the p point geodesic distance to q point, and R is radius；

Step 4.1.1: by all of some q ∈ N (p) and normal vector n (q) thereof, is transformed in following interim local coordinate system:

C={t (p'), t (p') × n (p'), n (p') } (5)

Wherein, p' be conversion after p, unit normal vector n (p') be Z axis, t (p') is the face curved surface incisal plane T at p' point (p') unit vector randomly choosed in；

Step 4.1.2: the q' after conversion and their normal vector n (q') is mapped to incisal plane T (p'), corresponding angle theta And the long mag of mould (q') is (q'):

$\begin{array}{c}\mathrm{\&theta;}\left(q^{\&prime;}\right)=\arctan \&lsqb;n_y\left(q^{\&prime;}\right)/n_x\left(q^{\&prime;}\right)\&rsqb;\\ mag\left(q^{\&prime;}\right)=\sqrt{n_x^2\left(q^{\&prime;}\right)+n_y^2\left(q^{\&prime;}\right)}\end{array}---\left(6\right)$

Wherein, n_x(q')=t (p') n (q'), n_y(q')=t (p') × n (p') n (q')；

Step 4.1.3: one the Gauss weight histogram of gradients of upper structure of the incisal plane T (p') under temporary coordinate system, a width of 360, each unit covers 1 °, and described Gauss weight is described as follows:

W (p', q')=mag (q') G_σ(d_g(p',q')) (7)

Wherein, standard deviation sigma is set to the half of radius R；

Step 4.1.4: using the peak value of described Gauss weight histogram of gradients as standardization direction d (p') at p', in neighborhood Be the most all transformed under new coordinate system:

C'={d (p'), d (p') × n (p'), n (p') } (8)

Step 4.2: descriptor structure: the spatial configuration of descriptor is to carry out on the incisal plane T (p') under new coordinate system C', With key point p' and its 8 neighborhood points as the center of circle, obtain 9 circles, and by sequence counterclockwise, described 8 neighborhood points are by close Key point p' is the center of circle to start uniform sampling at standardization direction and obtains.

It is the most according to claim 1 based on key point with the three-dimensional face identification method of local feature, it is characterised in that: institute State step 5 and specifically include following steps:

Step 5.1: calculate the local feature of each circle shaped neighborhood region of key point p: the local feature of each border circular areas is by three parts Constitute:

1) shape index rectangular histogram h_s, a length of 8, shape index is calculated by following formula:

$SI\left(q\right)=\frac{1}{2}-\frac{1}{\mathrm{\&pi;}}{\tan }^{-1}\frac{C_M\left(q\right)+C_m\left(q\right)}{C_M\left(q\right)-C_m\left(q\right)}---\left(9\right)$

Wherein q ∈ N (p), SI (q) are with a gaussian kernel as weight, and its standard variance is the q geodesic distance to the center of circle.

2) inclination angle rectangular histogram h_a, a length of 8, tilt angle alpha is defined as every some normal vector n (q) and standardization direction d (p') Angle:

$\mathrm{\&alpha;}\left(q\right)={\cos }^{-1}\frac{n\left(q\right)\&CenterDot;d\left(p\right)}{\left|n\left(q\right)\right|\&times;\left|d\left(p\right)\right|}---\left(10\right)$

Wherein q ∈ N (p), α (q) are with the angle between its normal vector and normal vector of key pointAs Weight；

3) deflection rectangular histogram h_d, a length of 8, direction β are defined as any two points q in neighborhood₁And q₂Between line l (q₁,q₂) And the angle between key point normal vector n (p):

$\mathrm{\&beta;}(q_1,q_2)={\cos }^{-1}\frac{l(q_1,q_2)\&CenterDot;n\left(p\right)}{\left|l(q_1,q_2)\right|\&times;\left|n\left(p\right)\right|}---\left(11\right)$

It is normalized to unit vector after each rectangular histogram described above is all normalized to unit vector, and series connection and obtains n-th Feature h of individual circle_n:

h_n=(h_sn,h_an,h_dn) (12)

Step 5.2: calculate the local feature of key point p: connect feature h of each circle_nObtain the local feature f of key point p:

F=(h₁,h₂,…,h₉) (13)。

It is the most according to claim 1 based on key point with the three-dimensional face identification method of local feature, it is characterised in that: institute State step 6 and specifically include following steps:

Step 6.1: calculate angle γ: set f_iAnd f_jIt is respectively on storehouse collection and test set face curved surface i-th, the spy of j key point Levying, γ is calculated as follows:

${\mathrm{\&gamma;}}_{i,j}={\cos }^{-1}\left(\frac{<f_i,f_j>}{\left|\right|f_i\left|\right|\&CenterDot;\left|\right|f_j\left|\right|}\right)---\left(14\right)$

Step 6.2: key point is mated, and concrete grammar is to each key point i in test set face curved surface, calculates itself and storehouse collection people Angle γ between each key point j on face curved surface_i,j, by angle γ_i,jSort in descending order, if the ratio of first and second Example is less than 0.8, then coupling can be accepted, and otherwise it fails to match；

Step 6.3: recognition methods, concrete grammar is that test set face curved surface concentrates each face curved surface to carry out key point with storehouse Join, test set with and the most face curved surface of its key point number that the match is successful and for same class.