CN108509866A - A kind of facial contour extraction method - Google Patents
A kind of facial contour extraction method Download PDFInfo
- Publication number
- CN108509866A CN108509866A CN201810199612.6A CN201810199612A CN108509866A CN 108509866 A CN108509866 A CN 108509866A CN 201810199612 A CN201810199612 A CN 201810199612A CN 108509866 A CN108509866 A CN 108509866A
- Authority
- CN
- China
- Prior art keywords
- curve
- point
- local configuration
- facial contour
- local
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/44—Local feature extraction by analysis of parts of the pattern, e.g. by detecting edges, contours, loops, corners, strokes or intersections; Connectivity analysis, e.g. of connected components
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/40—Extraction of image or video features
- G06V10/46—Descriptors for shape, contour or point-related descriptors, e.g. scale invariant feature transform [SIFT] or bags of words [BoW]; Salient regional features
- G06V10/462—Salient features, e.g. scale invariant feature transforms [SIFT]
-
- 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
- G06V40/168—Feature extraction; Face representation
- G06V40/169—Holistic features and representations, i.e. based on the facial image taken as a whole
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Health & Medical Sciences (AREA)
- Human Computer Interaction (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of facial contour extraction methods, including step：1) it inputs after an image, the region of wherein face is extracted using the algorithm of Face datection, and find the approximate region of facial contour by key feature point algorithm；2) a series of dense squares are generated in the approximate region sampling along facial contour, includes wherein by entire facial contour region；3) it in each local square region, all extracts a parabola and guides the local configuration curve based on gradient information, constitute the local result set of local configuration curve composition；4) by the global blending algorithm based on PCA, the local configuration curve fusion of dense redundancy is become into a global contour curve as a result, obtaining complete facial contour line.The method of the present invention has precision high, and speed is fast, can automatically can also user's interaction, the characteristics of facial contour curve finally obtained has many advantages, such as the high-precision of pixel scale, meets facial contour parabolic shape.
Description
Technical field
The present invention relates to the technical fields of image procossing, refer in particular to a kind of facial contour extraction method.
Background technology
In the related application of facial image, it is often necessary to according to the facial image of input, be automatically positioned out the pass of face
Key characteristic point, such as eyes, the profile of nose, corners of the mouth point and face.Existing method can be precisely located eyes, nose and
The features visibility point such as corners of the mouth, and the facial contour positioned at cheek and chin is then challenging.For most of at present
Only several characteristic points are arranged at face edge in the face alignment method of feature based point, it is difficult to extract the curved of facial contour line
Bent shape details.But these minutias are extremely important in the application of many computer visions, such as recognition of face, expression
Identification and three-dimensional face reconstruct etc..The accuracy of facial contour is improved for Digital Image Processing and computer vision research field
Have great importance and be worth, but extracts pixel scale at present accurately continuous facial contour curve is one and chooses very much
The problem of war.
The relatively early stage certain methods of facial contour detection are simply to use the chin area of a Parabolic Fit face,
But this method is too simple, is only used for chin area, and a parabola can not show the bending of facial contour
Feature.[M.Kass, A.Witkin, the and D.Terzopoulos, " Snakes such as Kass:Active contour models,”
IJCV, vol.1, no.4, pp.321-331,1988.] propose active contour model, can be used for objective contour curve is described,
It is mainly used in the Target Segmentation based on shape.Some researcher [V.Perlibakas, " Automatical detection
of face features and exact face contour,”Pattern Recognition Letters,vol.24,
No.16, pp.2977-2985,2003.] this method is applied and extracts facial contour line in face field.Active contour model
It is more sensitive for parameter therein, and different locally optimal solutions can be converged to because of the difference of initialized location, and
Obtain undesirable result.So Active contour usually require continuous adjusting parameter and accurately initialization, and this from
It is difficult to accomplish under right environment.Curve is fitted by active contour model by iteration from the position of initialization to objective contour simultaneously,
And this process is usually highly susceptible to noise jamming, so that often stopping in midway or local optimum.
Face segmentation is also a kind of method that can extract facial contour, [B.Wang, X.Chang, the and such as such as Wang
C.Liu,“Skin detection and segmentation of human face in color images,”
International Journal of Intelligent Engineering and Systems,vol.4,no.1,
Pp.10-17,2011.] skin detection and face dividing method proposed, usual such methods are typically that image is divided into face
Region and non-face region, the boundary between such two parts can serve as facial contour.But these methods are for face
The processing on boundary is relatively rough, so that boundary curve is excessively tortuous and is not bonded the actual profile of face, while often will
In the region that a part of neck area of face and the hair portion on top include face, and it is unable to get good people
Face contour curve.
The method of face alignment is then that the face key point pre-defined, such as eyes, nose are found on human face photo
Point, corners of the mouth point, eyebrow and face each section profile point, are a kind of strategies based on supervised learning.In recent years, shape is cascaded
Regression model has good performance, [the V.Kazemi and such as such as Kazemi in key feature point problem
J.Sullivan,“One millisecond face alignment with an ensemble of regression
Trees, " in CVPR, 2014, pp.1867-1874.] regression model is used, directly study is from facial image to face shape mould
Type volume mapping function, is simple and efficient, and good experimental result is all obtained in laboratory scene and natural scene.Face is aligned
Method greatly improve the robustness of face key point location.But key is counted out rareness on facial contour, Er Qieyou
Most complex scenarios under truth, it may appear that in partially or partially outer situation.So several points sparse in this way, it is difficult to show
Go out the different feature of the face mask line of different people.
Above several method or excessively by parameter and initialization limited can not robust normal operation on natural picture,
Or can find facial contour approximate location can not quasi- description facial contour bending features.It is current increasingly break out it is various
In vision application, it is desirable to be able to which robust, the quick method that must obtain high-precision facial contour are that such as face on upper layer is known
Not, facial expression recognition, the application such as 3D human face rebuildings provide crucial profile information.
Invention content
It is an object of the invention to overcome the deficiencies of existing technologies and insufficient, it is proposed that a kind of facial contour extraction method,
In arbitrarily one facial image of input, the complete continuous facial contour curve for being accurate to pixel scale can be quickly found.
To achieve the above object, technical solution provided by the present invention is：A kind of facial contour extraction method, including it is following
Step：
1) facial image pre-processes
It inputs after an image, the region of wherein face is extracted using the algorithm of Face datection, and by crucial special
Sign point location algorithm finds the approximate region of facial contour；
2) local square region samples
A series of dense squares are generated in the approximate region sampling along facial contour, by entire facial contour region
Including wherein；
3) local configuration curve extracts
In each local square region, it is bent all to extract a local configuration of the parabola guiding based on gradient information
Line constitutes the local result set of local configuration curve composition；
4) global contour curve fusion
By the global blending algorithm based on PCA, the local configuration curve fusion of dense redundancy is become into a global wheel
Wide Dependence Results obtain complete facial contour line.
In step 1), the approximate region that facial contour is found by key feature point algorithm, to initial
Changing position has robustness, specific as follows：
User's face detection algorithm extracts region of the face where in picture, obtains the encirclement frame of human face region, simultaneously
Obtain the size information of the position of face and scale in this pictures；Then it is obtained in face mask using face alignment algorithm
Several coarse key points as initialization, be linked to be an initialization curve for indicating facial contour region；
In step 2), the approximate region sampling along facial contour generates a series of dense squares,
By entire facial contour region include wherein, it is specific as follows：
The initialization curve that characterization facial contour region has been obtained in step 1), along this curve by entire people
Face contour area is divided into the square area of multiple overlappings, and the central point of these square areas is located at initialization curve
On, the default multiplying power of the face encirclement frame size for being dimensioned to the extraction of Face datection algorithm of square area, these overlappings
Square area to cover whole facial contour regions, the direction of square area is set as the initialization of Current central point
Curve near tangent direction, obtained local configuration Dependence Results always since at the top of square area, arrive square bottom
Terminate, initialization curve is located near real human face profile, and the square area of these overlappings just covers entirely really
Facial contour；In addition, the part that the method in intensive sampling facial contour region can obtain a large amount of overlapping redundancies is used to take turns herein
Wide Dependence Results meet the needs of the cross validation in the global contour curve extraction step of step 4)；
It is described in each local square region in step 3), it all extracts a parabola guiding and is based on gradient
The local configuration curve of information constitutes the local result set of local configuration curve composition, specific as follows：
The square area being largely overlapped in step 2) includes one section of true people in each square area
Face contour curve, and the direction of each piece of square area, the i.e. vertical direction on any one square side, with initialization
Direction of curve is consistent；Local configuration of the parabola guiding based on gradient information to be obtained according to following methods bent in this step
Line：
Each piece of square direction is consistent with the direction of initialization curve, then part is with the side of this square area
Individually see that this block square area, local facial contour line C must be from the top of square area as rectangular coordinate system
Start, terminates to the bottom of square area；This square area is the picture element matrix of N × N, then local configuration curve C
It is expressed as N number of from the pixel set that do not go together：
C=＜ p1,p2,...,pi,...,pN＞
Wherein, C is local configuration curve, pi=(i, j) is i-th point on this local configuration curve C, and this
The position of point is arranged in the i-th row of this square area, jth, and the range of wherein i and j are all [1, N]；p1It is first point,
pNIt is the last one point；For two point p adjacent on the contour line of local facialiAnd pi+1, it is desired to its adjacent two columns difference is less
In 1, it is ensured that the flatness between pixel；
Parabola guiding determines position based on energy function below the local configuration curve negotiating of gradient information：
Wherein,It is to optimize to obtain optimal local configuration curve, the Grad of G (C) expression local configuration curves C, and S
(C) indicate that the curvature of local configuration curve C, α values are to adjust the smoothness of this local configuration curve C；Energy letter above
Number is difficult to direct solution, therefore solves this problem using the method for Dynamic Programming, while not having the whole part of direct solution
The smoothness S (C) of contour curve, but greedy algorithm is utilized to guide result more like a parabola；C*(i, j) is using dynamic
The local configuration curve for the i-th row that planning algorithm obtains, then it be necessarily from before { C*(i-1, j-1), C*(i-1,
J), C*(i-1, j+1) } one in three contour lines, use di-1,j+Δ(i, j) indicates that point (i, j) arrives parabola C*(i-1,j+
Δ), the distance of Δ={ -1,0,1 } usesIt indicates to deviate parabolical mistake
The process of difference, dynamic programming algorithm indicates as follows：
Wherein, M indicates the matrix of Dynamic Programming, and M (i-1, j+ Δ) is the energy of matrix lastrow corresponding points, g (i, j) table
Show that the Grad at this point, i are line numbers, j is columns, ei-1,j+Δ(i, j) indicates to deviate parabolical error, and α values are control
The parameter of smoothness, different α values have an impact local configuration curve C results, and a suitable value need to be taken by experiment.
It is described by the global blending algorithm based on PCA in step 4), the local configuration curve of dense redundancy is melted
A global contour curve is synthesized as a result, obtaining complete facial contour line, it is specific as follows：
Square area set is obtained by intensive sampling on facial contour region by step 2), then passes through step
3) a local configuration curve is obtained in each square area；Just obtain the local configuration curve being overlapped by multistage now
Set；It enablesIndicate i-th point on kth local configuration curve；M indicates that the number of local facial contour line, N indicate every
The length of local facial contour line；The a series of point of so all local configuration curve negotiatings indicatesBecause step 2) is to the intensive sampling in facial contour region, the set bulk redundancy of P points, no
Only comprising the point on really facial contour, but also there is the point on the local configuration curve of failure, is merged in global contour curve
In find out a single pixel width accurately global contour curve；
Facial contour curve generally comprises a large amount of bending minutia, can not be indicated by simple parameterized form；
Therefore it needs to use series of points Q={ ql}l∈[1,L]Queue indicate the final global contour curve as a result, wherein L is complete
The length of office's contour curve, while the point in Q is all P to keep the bending minutia of global contour curve；
It is empty most starting queue Q, it is clear that q0It is exactly first point p on first local configuration curve0, then
Next point is found by PCA algorithms every time；
In order to calculate a littleThe directions PCA at place, calculate a little firstFor in P thus other point covariance matrix such as
Under：
Wherein,It is exactly a littleCovariance matrix,Indicate i-th point on kth local configuration curve,
Indicate other points, and i' ∈ [1, N] { i }, k' ∈ [1, M] { k }, M indicates that the number of local facial contour line, N indicate
The length of every local facial contour line；Function # () indicatesR is the distance between 2 points, and h is
One fixed threshold；When point the distance between it is remoter, then between influence it is smaller；So PCA is not to act on entirely to put
What collection closed, accelerate to calculate using a threshold value h；It is calculated by this step and just obtains the characteristic value of covariance matrix and correspond to most
The feature vector of big characteristic value is exactly required pointThe principal direction at place
When the last one point is q in queue Ql, the point in corresponding P isIn order to obtain next point ql+1, first use K close
Adjacent algorithm is found from point qlK nearest pointK is set as 7 herein, if all K points are all to belong to kth item office
Contouring curve, thenIt walks downward along this local configuration curve, if there is being not belonging to kth local configuration curve
Point, then needing to find a littlePlace best suits the directions PCAA point, calculated by following formula：
Wherein,It is the inner product between both direction, pointPlace best suits the directions PCAA point be by seeking extreme valueIt obtainsThe continuous iteration of this process, until there is no a little to add in queue Q, and Q all the points structures
At curve be exactly required global contour curve.
Compared with prior art, the present invention having the following advantages that and advantageous effect：
1, the method that the present invention uses intensive sampling on Initial Face contour area, the square of difference sampling gained
Region includes different image information, can find different local configuration Dependence Results.A region can not be bonded very wherein
Perhaps, real facial contour curve can obtain accurate result in another region.The mechanism of this cross validation ensure that calculation
The accuracy and robustness of method.
2, the present invention generates local facial contour line, fully by the way of the local configuration curve of parabola guiding
It is dexterously to find layout optimal solution using dynamic programming algorithm the characteristics of meeting parabolic shape using real human face contour line.
The method of parabola guiding had not only accelerated the speed of algorithm, but also global contour curve is made to meet truth.
3, the present invention is using the continuous face curve for obtaining high quality from part to global three steps, calculating process letter
Single, accuracy rate is higher, the precision with pixel scale, and is not easily susceptible to the interference of picture illumination variation.
4, in practical applications can be full-automatic, can also user interaction, to the position of the initialization of face contour extraction
Robust.
Description of the drawings
Fig. 1 is facial contour extraction method flow chart.
Fig. 2 is each step results exemplary plot of facial contour extraction method.
Fig. 3 is pretreatment and sampling square area exemplary plot.
Fig. 4 is the comparison diagram of different α values local configuration curvilinear motions.
Fig. 5 is the guiding of (a) (b) parabola and the comparison diagram that (c) (d) is guided without parabola.
Fig. 6 is the exemplary plot of the comparison of the methods of the present invention and ERT.
Fig. 7 be of the invention (g) (h) (i) from ACM methods (d) (e) (f) different initialized locations comparative result figure.
Specific implementation mode
The present invention is further explained in the light of specific embodiments.
As depicted in figs. 1 and 2, the facial contour extraction method that the present embodiment is provided, includes the following steps：
1) facial image pre-processes：It inputs after an image, wherein face is extracted using the algorithm of Face datection
Region, and find by key feature point algorithm the approximate region of facial contour；
2) local square region samples, and a series of dense pros are generated in the approximate region sampling along facial contour
Entire facial contour region is included wherein by shape；
3) local configuration curve extracts, and in each local square region, all extracts a parabola guiding based on ladder
The local configuration curve of information is spent, the local result set of local configuration curve composition is constituted；
4) global contour curve fusion, by the global blending algorithm based on PCA, by the local configuration curve of dense redundancy
Fusion becomes a global contour curve as a result, obtaining complete facial contour line.
In step 1), an image is inputted, it is pre-processed first, extracts the area where face part in image
Domain, and the approximate region of facial contour is found, face contour extraction algorithm is initialized.
User's face detection algorithm extracts region of the face where in picture first, obtains the encirclement frame of human face region,
The size of face scale in this pictures can be determined simultaneously.Then it is the people of the present invention to use any one face alignment algorithm
Face contours extract algorithm provides several coarse key points in face mask as initialization, is linked to be an expression facial contour
The initialization curve in region.
In the present invention, [V.Kazemi the and J.Sullivan, " One millisecond such as Kazemi have been used
face alignment with an ensemble of regression trees,”in CVPR,2014,pp.1867–
1874.] Ensembles of Regression Trees (ERT) algorithm proposed, face contour extraction algorithm of the invention is simultaneously
Do not require initialization that there is higher accuracy, it is only necessary to indicate the approximate region where facial contour.It is entire in order to accelerate
Algorithm, the present invention detect the key feature points near 5 facial contours using ERT algorithms.Their position is respectively in eyes
The region on both sides, face both sides and point.Fig. 3 (a) shows an example of 5 key feature points positions, can by figure
See that the present invention only only used the input of very coarse initialized location.
This 5 points are constituted by an initialization curve by curve fitting algorithm later, indicate this facial contour region,
And provide square center position and direction information for the sampling of the local square region of next step.Key point only have 5 very
It is sparse, in order to ensure this matched curve by all key points without departing from human face region, present invention uses
Catmull-Rom algorithms.Fig. 3 (b), which is shown, is become 5 key feature points fittings at the beginning of one using Catmull-Rom curves
Beginningization curve.
In step 2), the method for local square region sampling is specific as follows：
In previous step 1) in obtained an initialization curve in characterization facial contour region, along this curve I
Can be by square area that entire facial contour region division is many smaller overlappings.The central point of these square areas
In initialization curve, square area be dimensioned to Face datection algorithm extraction face encirclement frame size it is certain
Multiplying power (range is between 0~0.5), the present invention are set as 0.2 times of face encirclement frame size.The direction of square area is arranged
For the tangential direction of the initialization curve of Current central point, in order to which obtained local configuration Dependence Results are always from square region
Start at the top of domain, terminates to square bottom.
Initialization curve is located near real human face profile, and the square area of these overlappings can cover entirely
True facial contour.For the accuracy of result, using the method in intensive sampling facial contour region, default value of the present invention is set
It is set to 70 square areas.The local configuration Dependence Results of a large amount of overlapping redundancies can be obtained in this way, met complete in step 4)
The needs of high-precision global contour curve result are obtained in office's contour curve extraction step.Fig. 3 (c) shows intensive sampling
Square area.
In step 3), the method for local configuration curve extraction is specific as follows：
The square area that a large amount of overlappings have been obtained in previous step contains one section very in each square area
Real facial contour curve, and the direction (vertical direction on any one square side) of each piece of square area with just
Beginningization direction of curve is consistent.Local configuration curve of the parabola guiding based on gradient information will be found out in this step.
Each piece of square direction is consistent with the direction of initialization curve, then part is with the side of this square area
Individually see that this block square area, local facial contour line C must be from the top of square area as rectangular coordinate system
Start, terminates to the bottom of square area.This curve not only possesses maximum gradient, while also possessing as much as possible as throwing
Object line is equally smooth.This square area is the picture element matrix of N × N, then local configuration curve C can be expressed as it is N number of come
From the pixel set that do not go together：
C=＜ p1,p2,...,pi,...,pN＞
Wherein pi=(i, j) is i-th point on this local configuration curve C, and the position of this point is at this
The range of the i-th row, the jth row of square area, wherein i and j are all [1, N].p1It is first point, pNIt is first point.For
Two adjacent point p on the contour line of local facialiAnd pi+1, the present invention claims its adjacent two columns differences to be not more than 1, it is ensured that pixel
Between flatness.
Local configuration curve possesses prodigious Grad, and has the flatness as parabola.Throwing in the present invention
The guiding of object line determines position based on energy function below the local configuration curve negotiating of gradient information：
WhereinIt is to optimize to obtain optimal local configuration curve, the Grad of G (C) expression local configuration curves C, and S
(C) indicate that the curvature of local configuration curve C, α values are to adjust the smoothness of this local configuration curve C.Energy letter above
Number is difficult to direct solution, and the present invention solves this problem using the method for Dynamic Programming.While in order to facilitate the present invention not
There is the smoothness S (C) for solving whole local configuration curve, but greedy algorithm is utilized and guides result more like a parabola.
C*(i, j) is the local configuration curve of the i-th row obtained using dynamic programming algorithm, then it is necessarily from it
Preceding { C*(i-1, j-1), C*(i-1, j), C*(i-1, j+1) } one in three contour lines, use di-1,j+Δ(i, j) is indicated
Point (i, j) arrives parabola C*(i-1, j+ Δ), the distance of Δ={ -1,0,1 } useIt indicates
Deviate parabolical error.The process of dynamic programming algorithm indicates as follows：
Wherein M indicates the matrix of Dynamic Programming, and M (i-1, j+ Δ) is the energy of matrix lastrow corresponding points, g (i, j) table
Show that the Grad at this point, i are line numbers, j is columns.ei-1,j+Δ(i, j) indicates to deviate parabolical error.α values are control
The parameter of smoothness, Fig. 4 show influence of the different α values to local configuration curve C results.By the α of smoothing parameter in the present invention
Value is set as 0.7.Fig. 5 shows that parabola guiding and the dynamic programming algorithm for being based purely on gradient of no parabola guiding obtain
Different Results.
In step 4), the method for global contour curve fusion is specific as follows：
Previous step 2) has obtained square area set on facial contour region by intensive sampling, then passes through
Step 3) has obtained a local configuration curve in each square area.It has just been obtained now by much shorter and overlapping
The set of local configuration curve.It enablesIndicate i-th point on kth local configuration curve.M indicates local facial contour line
Number, N indicate the length of every local facial contour line.So all local configuration curves can pass through a series of table
ShowStep 2) includes not only to the intensive sampling in facial contour region, the set bulk redundancy of P points
The really point on facial contour, but also have the point on the local configuration curve of failure.This hair in the fusion of global contour curve
It is bright to find out a single pixel width accurately global contour curve.
Facial contour curve generally comprises a large amount of bending minutia, can not be indicated by simple parameterized form.
Therefore it needs to use series of points Q={ ql}l∈[1,L]Queue indicate the final global contour curve as a result, wherein L is complete
The length of office's contour curve.The point in Q is all P to keep the bending minutia of global contour curve simultaneously.
It is empty most starting queue Q, it is clear that q0It is exactly first point p on first local configuration curve0, then
Next point is found by PCA algorithms every time；
In order to calculate a littleThe directions PCA at place, the present invention calculate a little firstFor in the P so association side of other points
Poor matrix is as follows：
WhereinIt is exactly a littleCovariance matrix.Indicate i-th point on kth local configuration curve,Table
Show other points, and i' ∈ [1, N] { i }, k' ∈ [1, M] { k }, M indicates that the number of local facial contour line, N indicate every
The length of local facial contour line.Function # () indicatesR is the distance between 2 points, and h is one solid
Determine threshold value.When point the distance between it is remoter, then between influence it is smaller.So PCA is not to act on the collection entirely put to close
, the present invention accelerates to calculate using a threshold value h.Then the characteristic value of covariance matrix and corresponding maximum feature are just obtained
The feature vector of value is exactly required pointThe principal direction at place
When the last one point is q in queue Ql, the point in corresponding P isIn order to obtain next point ql+1, first use K close
Adjacent algorithm is found from point qlK nearest pointK is set as 7 in the present invention.If all K points are all to belong to
K local configuration curve, thenIt walks downward along this local configuration curve, if there is being not belonging to kth local configuration
The point of curve, then needing to find a littlePlace best suits the directions PCAA point, calculated by following formula：
WhereinIt is the inner product between both direction.PointPlace best suits the directions PCAA point be by seeking extreme valueIt obtainsThe continuous iteration of this process, until there is no a little to add in queue Q, and Q all the points structures
At curve be exactly required global contour curve.
In conclusion the method for the present invention has precision high, speed is fast, in practical applications can be full-automatic, can also use
Family interacts, and the facial contour curve finally obtained has the precision of pixel scale, and meets the spy of facial contour parabolic shape
Point.The present invention compares experiment with existing method, as a result as Fig. 6 shows that the present invention has superior performance.This hair simultaneously
Bright method carries out some changes in original initialized location and has no effect on to the position robust of the initialization of face contour extraction
As a result, having used 3 initialized locations as shown in Figure 7, result of the invention is still accurate, and the method for the ACM compared does not have
Obtain correct result.
Embodiment described above is only the preferred embodiments of the invention, and but not intended to limit the scope of the present invention, therefore
Change made by all shapes according to the present invention, principle, should all cover within the scope of the present invention.
Claims (3)
1. a kind of facial contour extraction method, which is characterized in that include the following steps：
1) facial image pre-processes
It inputs after an image, the region of wherein face is extracted using the algorithm of Face datection, and pass through key feature points
Location algorithm finds the approximate region of facial contour；
2) local square region samples
A series of dense squares are generated in the approximate region sampling along facial contour, include by entire facial contour region
Wherein；
3) local configuration curve extracts
In each local square region, all extracts a parabola and guide the local configuration curve based on gradient information, structure
The local result set formed at local configuration curve；
4) global contour curve fusion
By the global blending algorithm based on PCA, it is bent that the local configuration curve fusion of dense redundancy is become into a global profile
Line is as a result, obtain complete facial contour line.
2. a kind of facial contour extraction method according to claim 1, which is characterized in that in step 1), described is logical
The approximate region that key feature point algorithm finds facial contour is crossed, robustness is equipped with to initialization bit, it is specific as follows：
User's face detection algorithm extracts region of the face where in picture, obtains the encirclement frame of human face region, obtains simultaneously
The size information of the position of face and scale in this pictures；Then it is obtained using face alignment algorithm several in face mask
A coarse key point is linked to be an initialization curve for indicating facial contour region as initialization；
In step 2), the approximate region sampling along facial contour generates a series of dense squares, will be whole
A face contour area include wherein, it is specific as follows：
The initialization curve that characterization facial contour region has been obtained in step 1), along this curve by entire face wheel
Wide region division is the square area of multiple overlappings, and the central point of these square areas is located in initialization curve, just
The default multiplying power of the face encirclement frame size for being dimensioned to the extraction of Face datection algorithm of square region, the pros of these overlappings
Shape region will cover whole facial contour regions, and the direction of square area is set as the initialization curve of Current central point
Tangential direction, obtained local configuration Dependence Results always since at the top of square area, terminate, just to square bottom
Beginningization curve is located near real human face profile, and the square area of these overlappings just covers entire true face wheel
It is wide；In addition, the local configuration curve of a large amount of overlapping redundancies can be obtained using the method in intensive sampling facial contour region herein
As a result, meeting the needs of the cross validation in the global contour curve extraction step of step 4)；
It is described in each local square region in step 3), it all extracts a parabola guiding and is based on gradient information
Local configuration curve, constitute local configuration curve composition local result set, it is specific as follows：
The square area being largely overlapped in step 2) includes one section of true face wheel in each square area
Wide curve, and the direction of each piece of square area, the i.e. vertical direction on any one square side, and initialization curve
Direction it is consistent；Parabola is obtained according to following methods guide the local configuration curve based on gradient information in this step：
Each piece of square direction is consistent with the direction of initialization curve, then part using the side of this square area as
Rectangular coordinate system individually sees this block square area, local facial contour line C must be since at the top of square area,
Terminate to the bottom of square area；This square area is the picture element matrix of N × N, then local configuration curve C is expressed as
It is N number of from the pixel set that do not go together：
C=＜ p1,p2,...,pi,...,pN＞
Wherein, C is local configuration curve, pi=(i, j) is i-th point on this local configuration curve C, and this point
Position is arranged in the i-th row of this square area, jth, and the range of wherein i and j are all [1, N]；p1It is first point, pNIt is
The last one point；For two point p adjacent on the contour line of local facialiAnd pi+1, it is desirable that its adjacent two columns difference is not more than 1,
Ensure the flatness between pixel；
Parabola guiding determines position based on energy function below the local configuration curve negotiating of gradient information：
Wherein,It is to optimize to obtain optimal local configuration curve, the Grad of G (C) expression local configuration curves C, and S (C) table
Show that the curvature of local configuration curve C, α values are to adjust the smoothness of this local configuration curve C；Energy function above is difficult to
Direct solution, therefore this problem is solved using the method for Dynamic Programming, and meanwhile it is bent without the whole local configuration of direct solution
The smoothness S (C) of line, but greedy algorithm is utilized to guide result more like a parabola；C*(i, j) is calculated using Dynamic Programming
The local configuration curve for the i-th row that method obtains, then it be necessarily from before { C*(i-1, j-1), C*(i-1, j), C*(i-
1, j+1) } one in three contour lines, uses di-1,j+Δ(i, j) indicates that point (i, j) arrives parabola C*(i-1, j+ Δ), Δ=
The distance of { -1,0,1 } usesIt indicates to deviate parabolical error, Dynamic Programming
The process of algorithm indicates as follows：
Wherein, M indicates the matrix of Dynamic Programming, and M (i-1, j+ Δ) is the energy of matrix lastrow corresponding points, and g (i, j) indicates this
Grad at a point, i are line numbers, and j is columns, ei-1,j+Δ(i, j) indicates to deviate parabolical error, and α values are that control is smooth
The parameter of degree, different α values have an impact local configuration curve C results, and a suitable value need to be taken by experiment.
3. a kind of facial contour extraction method according to claim 1, it is characterised in that：In step 4), described is logical
Cross the global blending algorithm based on PCA, the fusion of the local configuration curve of dense redundancy is become into a global contour curve as a result,
Complete facial contour line is obtained, it is specific as follows：
Square area set is obtained by intensive sampling on facial contour region by step 2), is then existed by step 3)
A local configuration curve is obtained in each square area；The collection for the local configuration curve being overlapped by multistage is just obtained now
It closes；It enablesIndicate i-th point on kth local configuration curve；M indicates that the number of local facial contour line, N indicate every office
The length of portion's facial contour line；The a series of point of so all local configuration curve negotiatings indicatesBecause step 2) is to the intensive sampling in facial contour region, the set bulk redundancy of P points, no
Only comprising the point on really facial contour, but also there is the point on the local configuration curve of failure, is merged in global contour curve
In find out a single pixel width accurately global contour curve；
Facial contour curve generally comprises a large amount of bending minutia, can not be indicated by simple parameterized form；Therefore
It needs to use series of points Q={ ql}l∈[1,L]Queue indicate final global contour curve as a result, wherein L is global takes turns
Wide length of a curve, while the point in Q is all P to keep the bending minutia of global contour curve；
It is empty most starting queue Q, it is clear that q0It is exactly first point p on first local configuration curve0, then every time
Next point is found by PCA algorithms；
In order to calculate a littleThe directions PCA at place, calculate a little firstFor in P thus other point covariance matrix it is as follows：
Wherein,It is exactly a littleCovariance matrix,Indicate i-th point on kth local configuration curve,It indicates
It is other, and i' ∈ [1, N] { i }, k' ∈ [1, M] { k }, M indicates the number of local facial contour line, and N indicates every
The length of local facial contour line；Function # () indicatesR is the distance between 2 points, and h is one
Fixed threshold；When point the distance between it is remoter, then between influence it is smaller；So PCA is not to act on the set entirely put
On, accelerate to calculate using a threshold value h；The characteristic value for just obtaining covariance matrix is calculated by this step and corresponding maximum is special
The feature vector of value indicative is exactly required pointThe principal direction at place
When the last one point is q in queue Ql, the point in corresponding P isIn order to obtain next point ql+1, first calculated using k nearest neighbor
Method is found from point qlK nearest pointK is set as 7 herein, if all K points are all to belong to kth item locally to take turns
Wide curve, thenIt walks downward along this local configuration curve, if there is being not belonging to the point of kth local configuration curve,
It so needs to find a littlePlace best suits the directions PCAA point, calculated by following formula：
Wherein,It is the inner product between both direction, pointPlace best suits the directions PCAA point be by seeking extreme valueIt obtainsThe continuous iteration of this process, until there is no a little to add in queue Q, and Q all the points structures
At curve be exactly required global contour curve.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810199612.6A CN108509866B (en) | 2018-03-12 | 2018-03-12 | Face contour extraction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810199612.6A CN108509866B (en) | 2018-03-12 | 2018-03-12 | Face contour extraction method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108509866A true CN108509866A (en) | 2018-09-07 |
| CN108509866B CN108509866B (en) | 2020-06-19 |
Family
ID=63377528
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810199612.6A Active CN108509866B (en) | 2018-03-12 | 2018-03-12 | Face contour extraction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108509866B (en) |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109409262A (en) * | 2018-10-11 | 2019-03-01 | 北京迈格威科技有限公司 | Image processing method, image processing apparatus, computer readable storage medium |
| CN109558880A (en) * | 2018-10-16 | 2019-04-02 | 杭州电子科技大学 | A kind of whole profile testing method with Local Feature Fusion of view-based access control model |
| CN111667400A (en) * | 2020-05-30 | 2020-09-15 | 温州大学大数据与信息技术研究院 | Human face contour feature stylization generation method based on unsupervised learning |
| CN113160223A (en) * | 2021-05-17 | 2021-07-23 | 深圳中科飞测科技股份有限公司 | Contour determination method, contour determination device, detection device and storage medium |
| CN113297893A (en) * | 2021-02-05 | 2021-08-24 | 深圳高通半导体有限公司 | Method for extracting stroke contour point set |
| CN113837067A (en) * | 2021-09-18 | 2021-12-24 | 成都数字天空科技有限公司 | Organ contour detection method and device, electronic equipment and readable storage medium |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101136105A (en) * | 2007-05-11 | 2008-03-05 | 辽宁师范大学 | Freely differences calculus and deformable contour outline extracting system |
| CN101339612A (en) * | 2008-08-19 | 2009-01-07 | 陈建峰 | Face contour checking and classification method |
| CN102063727A (en) * | 2011-01-09 | 2011-05-18 | 北京理工大学 | Covariance matching-based active contour tracking method |
| US20160070973A1 (en) * | 2013-04-09 | 2016-03-10 | Laboratoires Bodycad Inc. | Concurrent active contour segmentation |
| CN106156739A (en) * | 2016-07-05 | 2016-11-23 | 华南理工大学 | A kind of certificate photo ear detection analyzed based on face mask and extracting method |
| CN106156692A (en) * | 2015-03-25 | 2016-11-23 | 阿里巴巴集团控股有限公司 | A kind of method and device for face edge feature point location |
| CN106529437A (en) * | 2016-10-25 | 2017-03-22 | 广州酷狗计算机科技有限公司 | Method and device for face detection |
| CN106650635A (en) * | 2016-11-30 | 2017-05-10 | 厦门理工学院 | Method and system for detecting rearview mirror viewing behavior of driver |
| CN107403144A (en) * | 2017-07-11 | 2017-11-28 | 北京小米移动软件有限公司 | Face localization method and device |
| CN107452030A (en) * | 2017-08-04 | 2017-12-08 | 南京理工大学 | Method for registering images based on contour detecting and characteristic matching |
-
2018
- 2018-03-12 CN CN201810199612.6A patent/CN108509866B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101136105A (en) * | 2007-05-11 | 2008-03-05 | 辽宁师范大学 | Freely differences calculus and deformable contour outline extracting system |
| CN101339612A (en) * | 2008-08-19 | 2009-01-07 | 陈建峰 | Face contour checking and classification method |
| CN102063727A (en) * | 2011-01-09 | 2011-05-18 | 北京理工大学 | Covariance matching-based active contour tracking method |
| US20160070973A1 (en) * | 2013-04-09 | 2016-03-10 | Laboratoires Bodycad Inc. | Concurrent active contour segmentation |
| CN106156692A (en) * | 2015-03-25 | 2016-11-23 | 阿里巴巴集团控股有限公司 | A kind of method and device for face edge feature point location |
| CN106156739A (en) * | 2016-07-05 | 2016-11-23 | 华南理工大学 | A kind of certificate photo ear detection analyzed based on face mask and extracting method |
| CN106529437A (en) * | 2016-10-25 | 2017-03-22 | 广州酷狗计算机科技有限公司 | Method and device for face detection |
| CN106650635A (en) * | 2016-11-30 | 2017-05-10 | 厦门理工学院 | Method and system for detecting rearview mirror viewing behavior of driver |
| CN107403144A (en) * | 2017-07-11 | 2017-11-28 | 北京小米移动软件有限公司 | Face localization method and device |
| CN107452030A (en) * | 2017-08-04 | 2017-12-08 | 南京理工大学 | Method for registering images based on contour detecting and characteristic matching |
Non-Patent Citations (3)
| Title |
|---|
| ZHONG XUE等: "《Facial feature extraction and image warping using PCA based statistic model》", 《PROCEEDINGS 2001 INTERNATIONAL CONFERENCE ON IMAGE PROCESSING》 * |
| 李月龙等: "《人脸特征点提取方法综述》", 《计算机学报》 * |
| 陈鹏飞等: "《基于形状识别的人脸轮廓线提取》", 《计算机工程与设计》 * |
Cited By (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109409262A (en) * | 2018-10-11 | 2019-03-01 | 北京迈格威科技有限公司 | Image processing method, image processing apparatus, computer readable storage medium |
| CN109558880A (en) * | 2018-10-16 | 2019-04-02 | 杭州电子科技大学 | A kind of whole profile testing method with Local Feature Fusion of view-based access control model |
| CN109558880B (en) * | 2018-10-16 | 2021-06-04 | 杭州电子科技大学 | Contour detection method based on visual integral and local feature fusion |
| CN111667400A (en) * | 2020-05-30 | 2020-09-15 | 温州大学大数据与信息技术研究院 | Human face contour feature stylization generation method based on unsupervised learning |
| CN113297893A (en) * | 2021-02-05 | 2021-08-24 | 深圳高通半导体有限公司 | Method for extracting stroke contour point set |
| CN113160223A (en) * | 2021-05-17 | 2021-07-23 | 深圳中科飞测科技股份有限公司 | Contour determination method, contour determination device, detection device and storage medium |
| CN113837067A (en) * | 2021-09-18 | 2021-12-24 | 成都数字天空科技有限公司 | Organ contour detection method and device, electronic equipment and readable storage medium |
| CN113837067B (en) * | 2021-09-18 | 2023-06-02 | 成都数字天空科技有限公司 | Organ contour detection method, organ contour detection device, electronic device, and readable storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108509866B (en) | 2020-06-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108509866A (en) | A kind of facial contour extraction method | |
| CN101499128B (en) | Three-dimensional human face action detecting and tracing method based on video stream | |
| Sobottka et al. | A novel method for automatic face segmentation, facial feature extraction and tracking | |
| Sun et al. | Face detection based on color and local symmetry information | |
| CN101964064B (en) | Human face comparison method | |
| CN101777116B (en) | Method for analyzing facial expressions on basis of motion tracking | |
| CN105869178B (en) | A kind of complex target dynamic scene non-formaldehyde finishing method based on the convex optimization of Multiscale combination feature | |
| CN109241910B (en) | Face key point positioning method based on deep multi-feature fusion cascade regression | |
| Wu et al. | Face and facial feature extraction from color image | |
| CN106709568A (en) | RGB-D image object detection and semantic segmentation method based on deep convolution network | |
| CN106096535A (en) | A kind of face verification method based on bilinearity associating CNN | |
| CN112418095A (en) | Facial expression recognition method and system combined with attention mechanism | |
| CN106778474A (en) | 3D human body recognition methods and equipment | |
| CN103886589A (en) | Goal-oriented automatic high-precision edge extraction method | |
| CN101216882A (en) | A method and device for positioning and tracking on corners of the eyes and mouths of human faces | |
| CN106096560A (en) | A kind of face alignment method | |
| CN106599785A (en) | Method and device for building human body 3D feature identity information database | |
| CN105184802B (en) | A kind of method and device of image procossing | |
| CN104794693A (en) | Human image optimization method capable of automatically detecting mask in human face key areas | |
| CN106611158A (en) | Method and equipment for obtaining human body 3D characteristic information | |
| CN114270417A (en) | Face recognition system and method capable of updating registered face template | |
| CN111931908A (en) | Face image automatic generation method based on face contour | |
| CN105139398A (en) | Multi-feature-based gray uneven image fast segmentation method | |
| CN106156739B (en) | A kind of certificate photo ear detection and extracting method based on face mask analysis | |
| Chen et al. | Eyes localization algorithm based on prior MTCNN face detection |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |