A kind of based on the degree of depth study and dictionary represent across age face verification method
Technical field
The present invention relates to computer vision field, more particularly, to a kind of based on the degree of depth study and dictionary represent across
Age face verification method.
Background technology
The development of science and technology makes picture pick-up device be popularized, and the face image data of enormous amount produces the most therewith.With
Time, many fields are required for applying face verification technology, such as: various meeting-place entrance, customs's transit passage etc..Apply at these
In, it may only be possible to obtain two facial images across the age, with regard to the problem creating the face verification across the age.All all of
The application of face verification technology is all based on same age section, but once needs the facial image verifying two width across the age
Time, these face recognition technologies and system are the most at a loss as to what to do, also cannot apply.Therefore solve across age face verification
Problem, it is possible to significantly widen the range of application of face recognition technology so that it is broadly for the mankind service.
The aging of face appearance is an extremely complex process, and shape and the texture of face are produced on physiological structure by it
Change.In recent years, human perception physics and computer vision field all propose many research sides about face aging
Method.Todd etc. think that the structural model of organism can change, based on this think of because of the change of the size and Orientation of external force suffered by it
Think that they propose a kind of method of hydrostatics model to portray the growth of human face structure.Burt and Perrett proposes " multiple
Close face " concept, strengthen face after conversion by the method that the poor figure information of compound face is transplanted on facial image
The sense organ age of image, thus reach to simulate aging purpose.The method is then extended by Tiddeman etc., it is proposed that one
Plant new method construct based on small echo and go out the aging method of more representative " compound face ".Xu Zhiwei etc. propose one
Aging method based on original Algorithms of Non-Negative Matrix Factorization predicts face image, but the method do not account for sparse constraint this
The condition impact on ageing results.Wang Zhangye etc. propose a kind of based on face outline office based on yellow race's face database
The personalized Algorithm of portion's curvature criteria difference, and achieve preferable Aging simulation result.
For across age recognition of face, current main method is to build the age ageing model of 2D or 3D, then
Reconstruct the facial image across the age for removing the impact at age.Wu etc. propose to use a relevant face model of growth simulation
Across the face shape at age for across age recognition of face, but this method need age information to predict new face shape,
The most infeasible.Ling etc. propose to use based on the pyramidal SVM algorithm of gradient direction for across age face verification.Li etc.
Propose to use multiple features distinctive analysis for closed set face verification.
Said method can not obtain good effect, for across age face verification mostly in age recognition of face
Occasion cannot apply these methods.
Summary of the invention
In order to overcome above-mentioned the deficiencies in the prior art, the invention provides a kind of based on the degree of depth study and dictionary represent across
Age face verification method.Facial image to be verified as input, is extracted multizone by deep neural network by this method
High-level feature, seek these features coding vector on outside reference character dictionary, finally coding vector sought cosine similarity.
In order to achieve the above object, the technical solution used in the present invention is:
A kind of based on the degree of depth study and dictionary represent across age face verification method, comprise the steps:
(1) for image to be verified, the method using face key point location, orient 10 points, extract these 10
The local facial block that point is corresponding;
(2) for the face block of each key point, training the degree of deep learning framework that this position is corresponding, each region is
Independent, extract the high-level feature of these face blocks, the feature of each piece is a M dimensional vector;
(3) from the Internet, utilize web crawlers, crawl the substantial amounts of facial image across the age as external data, to this
A little images also do step (1), (2) operation, it is thus achieved that the high-level characteristic vector of each piece, by the block of each key point of all classes
Same age section feature constitute a dictionary, here set 8 age brackets, i.e. have the dictionary of 8 age brackets;
(4) the high-level feature of facial image to be verified coding vector on each age bracket reference character dictionary is sought;
(5) each face block obtains 8 coding vectors, may make up the coding vector matrix of a M*8, to this matrix
Using one M dimensional vector of maximum mode pond, pond chemical conversion, a line every to this matrix takes that element of maximum as vector
Element, this vector is just as the final coding vector of this face.
(6) 10 coding vectors to 10 face blocks of facial image use cosine similarity to calculate the phase of two width images
Like property checking.
Preferably, in step (1), treat authentication image and use the mode piecemeal of distinguished point based, in external data, often
The age-grade face block of one position builds the age-grade face reference subset of this position, then obtain each position corresponding
The face subset of 8 age brackets.
Preferably, in step (2), the deep neural network that each face block of two width facial images to be verified is inputted
The high-level feature of middle extraction, deep neural network uses structure based on convolutional neural networks, includes 11 layers, 5 volumes
Lamination and 5 pond layers, each convolutional layer is followed by a pond layer, and pond layer uses the mode in maximum pond, is most followed by one
Full articulamentum, output is high-level feature.
Preferably, in step (3), for obtain on the Internet across age face image data, we limit choose M
Individual class, each class probably has the facial image of the different age group of 8 groups, repeats (1) (2) operation and extracts high-level feature, obtains
Obtain the high-level characteristic vector of each piece, the feature of the same age section of the block of each key point of all classes is constituted a word
Allusion quotation, setting 8 age brackets, i.e. each face key point here has the dictionary at 8 dictionary ages.
Preferably, in step (4), ask the training image high-level feature with test image in each age bracket reference word
Coding vector in allusion quotation, detailed process is as follows:
Wherein, xkRepresent kth face block, C(j,k)It is the dictionary of the jth age bracket of kth face block, α(j,k)It is exactly
The kth face block of facial image to be verified coding vector on jth age bracket dictionary.
Preferably, in step (5), each face block obtains 8 coding vectors, may make up the coding vector of a M*8
Matrix, uses one M dimensional vector of maximum mode pond, pond chemical conversion to this matrix, and a line every to this matrix takes that of maximum
Element is as the element of vector, and this vector is just as the final coding vector of this face.
Preferably, in step (6), 10 coding vectors of every facial image to be verified are used cosine similarity meter
Calculate the similarity checking of two width images.
The present invention has such advantages as relative to prior art and effect:
1, the present invention propose a kind of new based on degree of depth study and dictionary represent across age face verification method.
2, the present invention utilizes degree of depth convolutional neural networks to extract high-level feature.
3, the present invention learns multiple reference character dictionaries across the age by introducing external data, by maximum pond coding vector
Reach the effect with age invariance.
4, the present invention uses the mode demarcated based on key point to position local facial block, makes full use of the spy of face detection
Reference ceases.
Accompanying drawing explanation
Fig. 1 is the overview flow chart of the present invention.
Fig. 2 is degree of depth convolutional neural networks structure chart.
Detailed description of the invention
In order to make the purpose of the present invention, technical scheme and advantage clearer, below in conjunction with drawings and Examples, right
The present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, not
For limiting the present invention.As long as additionally, technical characteristic involved in each embodiment of invention described below that
The conflict of not constituting between this just can be mutually combined.
Accompanying drawing gives the operating process of the present invention,
As it is shown in figure 1, a kind of face verification method learning based on the degree of depth to represent with dictionary, comprise the following steps:
(1) for image to be verified, the method using face key point location, orient 10 points, extract this 10 points
Corresponding local facial block;
(2) for the face block of each key point, training the degree of deep learning framework that this position is corresponding, each region is
Independent, extract the high-level feature of these face blocks, the feature of each piece is a M dimensional vector;
(3) from the Internet, utilize web crawlers, crawl the substantial amounts of facial image across the age as external data, to this
A little images also do step (1), (2) operation, it is thus achieved that the high-level characteristic vector of each piece, by the block of each key point of all classes
Same age section feature constitute a dictionary, here set 8 age brackets, i.e. have the dictionary of 8 age brackets;
(4) the high-level feature of facial image to be verified coding vector on each age bracket reference character dictionary is sought;
(5) each face block obtains 8 coding vectors, may make up the coding vector matrix of a M*8, to this matrix
Using one M dimensional vector of maximum mode pond, pond chemical conversion, a line every to this matrix takes that element of maximum as vector
Element, this vector is just as the final coding vector of this face.
(6) 10 coding vectors to 10 face blocks of facial image use cosine similarity to calculate the phase of two width images
Like property checking.
Detailed process in step (1) is as follows: treats authentication image and uses the mode piecemeal of distinguished point based, at external number
According to, the age-grade face block of each position builds the age-grade face reference subset of this position, then obtain each
The face subset of 8 age brackets that position is corresponding.
Detailed process in step (2) is as follows: the degree of depth inputted by each face block of two width facial images to be verified
Extracting high-level feature in neutral net, degree of deep learning network is as shown in Figure 2.
Detailed process in step (3) is as follows: for obtain on the Internet across age face image data, we limit
Surely choosing M class, each class probably has the facial image of the different age group of 8 groups, repeats step (1), (2) operation extraction high level
Secondary feature, it is thus achieved that the high-level characteristic vector of each piece, by the spy of the same age section of the block of each key point of all classes
Levying one dictionary of composition, setting 8 age brackets, i.e. each face key point here has the dictionary at 8 dictionary ages.
Detailed process in step (4) is as follows: ask the training image high-level feature with test image at each age
Coding vector on section reference character dictionary, detailed process is as follows:
Wherein, xkRepresent kth face block, C(j,k)It is the dictionary of the jth age bracket of kth face block, α(j,k)It is exactly
The kth face block of facial image to be verified coding vector on jth age bracket dictionary.
Detailed process in step (5) is as follows: each face block obtains 8 coding vectors, may make up the volume of a M*8
Code vector matrix, uses one M dimensional vector of maximum mode pond, pond chemical conversion to this matrix, and a line every to this matrix takes maximum
That element as vector element, this vector just as the final coding vector of this face.
Detailed process in step (6) is as follows: 10 coding vectors of every facial image to be verified are used cosine
The similarity checking of Similarity Measure two width image.
The embodiment of invention described above, is not intended that limiting the scope of the present invention.Any at this
Amendment, equivalent and improvement etc. done within bright spiritual principles, should be included in the claim protection of the present invention
Within the scope of.