Embodiment
For making the technical problem to be solved in the present invention, technical scheme and advantage clearly, be described in detail below in conjunction with the accompanying drawings and the specific embodiments.
On the one hand, the method for a kind of face authentication of the present invention, as shown in Figure 1, comprising:
Step S101: use the multi-layer degree of depth convolutional network of training through multistratum classification network association in advance to extract the proper vector of multiple level successively to facial image to be certified and facial image template;
Multi-layer degree of depth convolutional network comprises the convolutional network of more than 2, and each convolutional network comprises convolution, activation and down-sampling operation, and order and the quantity of these operations are not fixed, and determine according to actual conditions; Each convolutional network of the present invention all extracts a proper vector, can be designated as fea
1, fea
2, fea
3(only list the proper vector of 1 group of multiple level here, the i.e. proper vector of multiple levels of facial image to be certified or facial image template, formula hereinafter also only writes out the formula of an image), first convolutional network be input as facial image to be certified or facial image template, a rear convolutional network be input as previous convolutional network operation after characteristic pattern;
General degree of depth network also exists gradient disperse problem, and multi-layer degree of depth convolutional network of the present invention carries out joint training by multistratum classification network and obtains, and avoids the problems referred to above.
Step S102: the proper vector of multiple level is mapped as unified dimensional proper vector by unified dimensional linear mapping matrix successively; Unified dimensional linear mapping matrix is through that training in advance obtains, and can be designated as W
1, W
2, W
3..., unified dimensional proper vector can be designated as f
1, f
2, f
3...
Step S103: unified dimensional proper vector is connected into union feature vector; Feature can be designated as
merge.
Step S104: union feature vector is carried out dimensionality reduction mapping by linear dimensionality reduction mapping matrix and obtains multi-feature vector; Linear dimensionality reduction mapping matrix is through that training in advance obtains, and can be designated as W
t, multi-feature vector can be designated as f
t.
Step S105: by linear discriminant analysis, utilizes absolute value normalization cosine value, to the certification of comparing of the multi-feature vector of the facial image to be certified obtained and the multi-feature vector of facial image template.
In the method for face authentication of the present invention, first the multi-layer degree of depth convolutional network of training through multistratum classification network association is in advance used to extract the proper vector of multiple levels of facial image to be certified and facial image template, then the proper vector of multiple level is mapped as unified dimensional proper vector by unified dimensional linear mapping matrix successively, again unified dimensional proper vector is connected into union feature vector, and by union feature vector by linear dimensionality reduction mapping matrix carry out dimensionality reduction map obtain multi-feature vector, finally by linear discriminant analysis, utilize absolute value normalization cosine value, to the certification of comparing of the multi-feature vector of the facial image to be certified obtained and the multi-feature vector of facial image template.
Compared with prior art, the present invention is by multi-layer degree of depth convolutional network automatic learning and extract feature, and compared with going out a proper vector with engineer in prior art, anti-interference energy is strong, and extensibility is good, and certification accuracy rate is high.
Multi-layer degree of depth convolutional network of the present invention carries out joint training by multistratum classification network and obtains, and avoid gradient disperse problem, certification accuracy rate is high.
And the proper vector of multiple level is merged, increases characteristics of image richness, compensate for general degree of depth network insufficient to each hierarchy characteristic process, only utilize high-level characteristic to be not enough to the defect of abundant Description Image; Further increase certification accuracy rate.
Inventor also finds, traditional comparison authentication method, especially cosine similarity method, have ignored the long difference of vector field homoemorphism, thus the description that makes a difference is comprehensive, reduces the accuracy rate of comparison certification; The present invention adopts linear discriminant analysis, compares, further improve certification accuracy rate to the multiple difference characteristics comprising absolute value normalization cosine value.
Therefore the method antijamming capability of face authentication of the present invention is strong, extensibility is good, and certification accuracy rate is high, and avoids gradient disperse problem, makes up the defect utilizing high-level characteristic to be not enough to abundant Description Image.
One as the method for face authentication of the present invention is improved, and also comprises before step S101:
Step S100: carry out pre-service to facial image to be certified and facial image template, pre-service comprises positioning feature point, image rectification and normalized.In fact, facial image template may cross pre-service through prior, can not carry out this step.
The present invention adopts the Face datection algorithm based on cascade Adaboost to carry out Face datection to image, then the facial modeling algorithm based on SDM is utilized to carry out positioning feature point to the face detected, and by image scaling, rotation and translation, face is corrected and normalization alignment, as shown in Figure 3.
The present invention adopts simple gray scale normalization pre-service, and the fundamental purpose of gray scale normalization is convenient to network processes continuous data and avoids processing larger discrete grey value, thus avoid occurring abnormal conditions.
The present invention carries out pre-service to facial image can facilitate follow-up verification process, and avoids the impact of extraordinary image vegetarian refreshments on authentication result.
Another kind as the method for face authentication of the present invention improves, and each convolutional network comprises the operation of convolution operation, activation manipulation and down-sampling, and the proper vector of each level calculates as follows:
Step S1011: use convolution kernel to carry out convolution operation to facial image to be certified and facial image template, obtain convolution characteristic pattern, convolution operation is same convolution operation;
The present invention adopts the convolution operation of same form, carries out zero padding during operation to input picture.The characteristic pattern that the convolution operation of same form obtains is identical with input picture life size.
Step S1012: use activation function to carry out activation manipulation to convolution characteristic pattern, obtain activating characteristic pattern, activation function is ReLU activation function.
Step S1013: use sampling function to carry out down-sampling operation to activation characteristic pattern, obtain characteristic pattern of sampling, down-sampling is operating as maximal value sampling;
The present invention adopts maximal value to sample, maximal value sampling is using the feature of the maximal value of sampling block interior element value as sampling block, and in image procossing, maximal value sampling can extract the texture information of image, and maintain certain unchangeability of image to a certain extent, as rotation, translation, convergent-divergent etc.; In addition, according to statistics experiment, compare average sample, maximal value sampling is insensitive to data changes in distribution, and feature extraction is relatively stable.
Step S1014: above-mentioned steps is repeated to the sampling characteristic pattern obtained, obtains new sampling characteristic pattern, and so repeat several times;
Step S1015: all sampling characteristic patterns obtained are carried out vectorization, obtains the proper vector of each level, all sampling characteristic patterns each step obtained form a vector.
The present invention can extract feature rich and stable proper vector, can describe facial image fully, add certification accuracy rate.
As another improvement of the method for face authentication of the present invention, multi-layer degree of depth convolutional network is obtained by the joint training of softmax sorter network, comprising:
During training, first to there is facial image Sample Storehouse, then use initialized multi-layer degree of depth convolutional network to extract the proper vector of multiple level successively to facial image sample; Be the same with aforesaid step S101, be only verification process above, be training process here, the parameters in multi-layer degree of depth convolutional network now gets initial value;
The proper vector of multiple level is mapped as successively the unified dimensional proper vector of same dimension by unified dimensional linear mapping matrix;
In softmax sorter network, use linear mapping matrix to map unified dimensional proper vector respectively, obtain mapping vector; Linear mapping matrix now gets initial value;
Use softmax function mapping directive amount to activate, obtain network output valve vector;
With the label data of network output valve vector sum facial image sample for input quantity, calculate network error by cross entropy loss function;
Each unified dimensional proper vector is connected into a union feature vector;
Union feature vector is carried out dimensionality reduction mapping by linear dimensionality reduction mapping matrix and obtains multi-feature vector;
For network error assigns weight, and calculate the renewal gradient of linear mapping matrix, unified dimensional linear mapping matrix, linear dimensionality reduction mapping matrix and convolution kernel;
The renewal gradient of linear mapping matrix, unified dimensional linear mapping matrix, linearly dimensionality reduction mapping matrix and convolution kernel is utilized to carry out iteration renewal to linear mapping matrix, unified dimensional linear mapping matrix, linear dimensionality reduction mapping matrix and convolution kernel;
Judge whether network error and iterations meet the requirements, and if so, terminate, otherwise, go to and use initialized multi-layer degree of depth convolutional network to extract the proper vector of multiple level successively to facial image sample.
Network error meets the requirements and refers to network error value minimum (or to a certain extent little), and now the parameters (linear mapping matrix, unified dimensional linear mapping matrix, linear dimensionality reduction mapping matrix and convolution kernel) of multi-layer degree of depth convolutional network and softmax sorter network is the multi-layer degree of depth convolutional network after training and softmax sorter network; Iterations meets the requirements and refers to that iterations reaches setting value.
The present invention carries out joint training by softmax sorter network, further avoid gradient disperse problem, and can further by the flexibility ratio weighting of sorter network error being increased to e-learning.
As another improvement of the method for face authentication of the present invention, step S105 comprises:
Step S1051: carry out cosine similarity operation with the multi-feature vector of the multi-feature vector of the facial image to be certified obtained and facial image template for input quantity, obtain cosine similarity;
Step S1052: carry out the operation of absolute value normalization cosine with the multi-feature vector of the multi-feature vector of the facial image to be certified obtained and facial image template for input quantity, obtain absolute value normalization cosine value;
Step S1053: ask modulo operation to the multi-feature vector of the facial image to be certified obtained and the multi-feature vector of facial image template, obtains the first mould long long with the second mould;
Step S1054: by cosine similarity, absolute value normalization cosine value, the long difference vector four-dimensional with the long composition of the second mould one of the first mould;
Step S1055: difference vector maps by usage variance DUAL PROBLEMS OF VECTOR MAPPING matrix, obtains one-dimensional vector, as comparison score value;
Step S1056: comparison score value and comparison threshold value are compared, if comparison score value is greater than comparison threshold value, then face authentication passes through.
Inventor finds, traditional comparison authentication method, especially cosine similarity method, have ignored the long difference of vector field homoemorphism, thus the description that makes a difference is comprehensive, reduces the accuracy rate of comparison certification; The contrast of absolute value normalization cosine value is responsive to the long difference of vector field homoemorphism, can make up that cosine similarity ignores the long difference of vector field homoemorphism and the difference that causes describes incomplete problem.
Therefore the cosine similarity of comparison feature, absolute value normalization cosine value and two character modules length are combined as a four-dimensional difference vector by the present invention, carry out linear discriminant analysis, further improve certification accuracy rate.
On the other hand, the invention provides a kind of device of face authentication, as shown in right Fig. 2, comprising:
First extraction module 11, for using the multi-layer degree of depth convolutional network of training through multistratum classification network association in advance to extract the proper vector of multiple level successively to facial image to be certified and facial image template;
First mapping block 12, for being mapped as unified dimensional proper vector by unified dimensional linear mapping matrix successively by the proper vector of multiple level;
First serial module structure 13, for being connected into union feature vector by unified dimensional proper vector;
Second mapping block 14, obtains multi-feature vector for union feature vector is carried out dimensionality reduction mapping by linear dimensionality reduction mapping matrix;
First comparing module 15, for by linear discriminant analysis, utilizes absolute value normalization cosine value, to the certification of comparing of the multi-feature vector of the facial image to be certified obtained and the multi-feature vector of facial image template.
The device antijamming capability of face authentication of the present invention is strong, and extensibility is good, and certification accuracy rate is high, and avoids gradient disperse problem, makes up the defect utilizing high-level characteristic to be not enough to abundant Description Image.
One as the device of face authentication of the present invention is improved, and also comprises before the first extraction module:
Pretreatment module, for carrying out pre-service to facial image to be certified and facial image template, pre-service comprises positioning feature point, image rectification and normalized.
The present invention carries out pre-service to facial image can facilitate follow-up verification process, and avoids the impact of extraordinary image vegetarian refreshments on authentication result.
Another kind as the device of face authentication of the present invention improves, and the proper vector of each level is calculated by such as lower unit:
Convolution unit, for using convolution kernel to carry out convolution operation to facial image to be certified and facial image template, obtain convolution characteristic pattern, convolution operation is same convolution operation;
Activate unit, for using activation function to carry out activation manipulation to convolution characteristic pattern, obtain activating characteristic pattern, activation function is ReLU activation function;
Sampling unit, for using sampling function to carry out down-sampling operation to activation characteristic pattern, obtain characteristic pattern of sampling, down-sampling is operating as maximal value sampling;
Cycling element, for repeating above-mentioned steps to the sampling characteristic pattern obtained, obtaining new sampling characteristic pattern, and so repeating several times;
Primary vector unit, for all sampling characteristic patterns obtained are carried out vectorization, obtains the proper vector of each level.
The present invention can extract feature rich and stable proper vector, can describe facial image fully, add certification accuracy rate.
As another improvement of the device of face authentication of the present invention, multi-layer degree of depth convolutional network is obtained by the joint training of softmax sorter network, comprising:
Second extraction module, extracts the proper vector of multiple level successively for using initialized multi-layer degree of depth convolutional network to facial image sample;
3rd mapping block, for being mapped as the unified dimensional proper vector of same dimension successively by unified dimensional linear mapping matrix by the proper vector of multiple level;
4th mapping block, for using linear mapping matrix to map unified dimensional proper vector respectively in softmax sorter network, obtains mapping vector;
Active module, for using softmax function mapping directive amount to activate, obtains network output valve vector;
First computing module, for the label data of network output valve vector sum facial image sample for input quantity, calculate network error by cross entropy loss function;
Second serial module structure, for being connected into a union feature vector by each unified dimensional proper vector;
5th mapping block, obtains multi-feature vector for union feature vector is carried out dimensionality reduction mapping by linear dimensionality reduction mapping matrix;
Second computing module, for assigning weight for network error, and calculates the renewal gradient of linear mapping matrix, unified dimensional linear mapping matrix, linear dimensionality reduction mapping matrix and convolution kernel;
Update module, carries out iteration renewal for utilizing the renewal gradient of linear mapping matrix, unified dimensional linear mapping matrix, linearly dimensionality reduction mapping matrix and convolution kernel to linear mapping matrix, unified dimensional linear mapping matrix, linear dimensionality reduction mapping matrix and convolution kernel;
Judge module, for judging whether network error and iterations meet the requirements, and if so, terminate, otherwise, go to the second extraction module.
The present invention carries out joint training by softmax sorter network, further avoid gradient disperse problem, and can further by the flexibility ratio weighting of sorter network error being increased to e-learning.
As another improvement of the device of face authentication of the present invention, the first comparing module comprises:
First computing unit, for carrying out cosine similarity operation with the multi-feature vector of the multi-feature vector of the facial image to be certified obtained and facial image template for input quantity, obtains cosine similarity;
Second computing unit, for carrying out the operation of absolute value normalization cosine with the multi-feature vector of the multi-feature vector of the facial image to be certified obtained and facial image template for input quantity, obtains absolute value normalization cosine value;
3rd computing unit, for asking modulo operation to the multi-feature vector of the facial image to be certified obtained and the multi-feature vector of facial image template, obtains the first mould long long with the second mould;
Secondary vector unit, for growing the difference vector four-dimensional with the long composition of the second mould one by cosine similarity, absolute value normalization cosine value, the first mould;
Map unit, maps difference vector for usage variance DUAL PROBLEMS OF VECTOR MAPPING matrix, obtains one-dimensional vector, as comparison score value;
Comparing unit, for comparison score value and comparison threshold value being compared, if comparison score value is greater than comparison threshold value, then face authentication passes through.
The cosine similarity of comparison feature, absolute value normalization cosine value and two character modules length are combined as a four-dimensional difference vector by the present invention, carry out linear discriminant analysis, further improve certification accuracy rate.
With a specific embodiment, present invention is described below:
This present invention needed to train before certification, and as shown in Figure 4, training process is as follows for concrete flow process:
First the present invention proposes new convolutional network and extracts image feature vector, power degree of depth convolutional network (multi-layer degree of depth convolutional network) is composed in the accumulation of multi-layer Fusion Features, then utilizes the learning process shown in softmax network and Fig. 3 to carry out feature learning to image.
Network learning procedure mainly comprises the forward calculation of network and the back-propagating of network error.
(A) convolutional network forward calculation
Basic convolutional network (as shown in Figure 5, notice that Fig. 5 is the example of a convolutional network, it not the convolutional network that the present invention uses, convolutional network of the present invention is: convolution, activation, down-sampling,) comprise the operation of convolution operation, activation manipulation and down-sampling, in order to subsequent calculations is convenient, generally also need to carry out vectorization operation.In figure 6, the convolutional network of every one deck all represents a basic convolutional network, and order and the number of times of the various operations wherein comprised all can set according to particular problem.
Convolution operation has different modes, and the present invention adopts the convolution operation of same form, carries out zero padding during operation to input picture.The characteristic pattern that the convolution operation of same form obtains is identical with input picture life size.
According to convolutional calculation formula, can obtain, when input data are two dimensional image, the computing formula of convolution characteristic pattern element, such as formula (2):
Wherein, c
krepresent a convolution operation kth convolution kernel, c
k(i, j) represents c
kthe i-th row, jth row element, s
crepresent the length of side of convolution kernel, M
ckrepresent input picture I and c
kthe convolution characteristic pattern that convolution obtains, M
ck(m, n) represents M
ckm is capable, the element of the n-th row, neighborhood (m, n, s
c) represent centered by (m, n), the length of side is s
cneighborhood,
represent the convolution operation symbol of same form.
And when input data are the characteristic pattern obtained through some operations, the computing formula of convolution characteristic pattern element, such as formula (3):
To the convolution characteristic pattern M obtained by convolution operation
ckcarry out activation manipulation, refer to M
ckeach element be input in activation function f and map, such as formula (4):
M
Ak(m,n)=f(M
Ck(m,n)).(4)
Wherein, M
akrepresent M
ckthrough the activation characteristic pattern that activation function obtains, f represents activation function.
The present invention adopts ReLU activation function.
f(x)=ReLU(x)=max(0,x).(5)
To the activation characteristic pattern M that activation manipulation obtains
akdo down-sampling operation, mainly reduced the dimension of feature by the mode of sampling, further compression and abstract characteristics of image.
First input Data Placement is become the s without overlapping by down-sampling operation
s× s
sfritter, s
srepresent the length of side of sampling core, then the data of each sub-block be input in sampling function and map, map output and be sampled value corresponding to sub-block, such as formula (6):
M
Sk(m,n)=s(M
Ak(s
s·(m-1)+1:s
s·m,s
s·(n-1)+1:s
s·n))(6)
Wherein, M
skrepresent M
akthrough the sampling characteristic pattern that sampling function obtains, M
sk(m, n) represents M
skm is capable, the element of the n-th row, and s represents sampling function.Fig. 8 illustrates the process of the size input data that are 4 × 4 being carried out to down-sampling operation, wherein s
s=2.
The present invention adopts maximal value to sample.
Maximal value is sampled the feature of the maximal value of sampling block interior element value as sampling block, such as formula (7):
s(I)=max(I)(7)
In image procossing, maximal value sampling can extract the texture information of image, and maintains certain unchangeability of image to a certain extent, as rotation, translation, convergent-divergent etc.; In addition, according to statistics experiment, compare average sample, maximal value sampling is insensitive to data changes in distribution, and feature extraction is relatively stable.
After feature extraction terminates, need to carry out vectorization operation to the characteristic pattern obtained, obtain proper vector fea, feature is input in sorter network, and then network parameter is learnt.
Vectorization operation is such as formula (8):
Wherein, v represents scalar data is stretched as a column vector, and concat represents becomes a high dimension vector, total number of K representation feature figure by the series connection of the vector of instruction.
(B) unified dimensional linear mapping
Image can obtain series of features figure after have passed through the several times convolution of convolutional network, activation and down-sampling operation, and the present invention utilizes linear mapping, the feature of each level is all mapped as the feature of same dimension.Such as formula (9), n in formula
frepresent the dimension of unified dimensional proper vector, n
irepresent fea
idimension:
(C) softmax sorter network
Fig. 7 illustrates the basic structure of softmax network, in diagram, and f
irepresent i-th component of input feature value f, N
crepresent categorical measure, W
idrepresent linear mapping matrix.
It should be noted that herein, when utilizing latticed form to realize linear mapping, generally all can adopt the linear mapping with deviation (bias).Because vectorial addition can realize equivalence by rewriting mapping matrix with mapping vector multiplication, therefore, the present invention is in order to write conveniently, involved all linear mapping operation expressions all adopt rewriting form, and directly utilize former name variable represent revised mapping matrix and map vector, and in expression formula, not embodying bias, in formula, o represents the output after linear mapping, o in figure
irepresent i-th component of o.
o=W
id·f(10)
H
irepresent i-th component of the network output valve h that o obtains after softmax function activates:
h=softmax(o).(11)
Wherein, softmax function is the nonlinear activation function that softmax network adopts, and its expression formula is:
Can be drawn by formula (12), softmax function is " non-negative and and normalizing function ", therefore, and can using its function-output as " input data belong to the probability of corresponding class ", namely
h
i=P(lable
i=1)=P(input∈CLASS
i).(13)
Wherein,
for the binary set of data original tag LABEL (representing LABEL people of data centralization), such as formula (14); CLASS
irepresent the i-th class data set, in recognition of face, represent all images of i-th people:
Class is the categorised decision that network provides according to network output h:
Face identity in recognition image is Images Classification problem, and the sorting algorithm that the present invention adopts is softmax sorter network, and its loss function adopted is cross entropy loss function, such as formula (16):
Wherein, h is the network output valve vector through softmax function in sorter network, and label is the binary set of data original tag LABEL.
Because the parameter of network is more, be easy to occur over-fitting problem, therefore adopt regularization to limit network parameter, thus alleviate Expired Drugs to a certain extent, the present invention adopts two norm regularizations.According to above explanation, the error of network can be expressed as formula (17):
J(θ)=loss(h,label)+λΣ||θ||
2.(17)
Wherein, J (θ) represents network error, and λ is regularization coefficient, θ be all in feature learning network can the set of learning parameter, such as formula (18), comprise the linear mapping matrix of the convolution kernel of convolutional network, sorter network:
θ={θ
c,θ
id},θ
c={c
1,c
2,...,c
K},θ
id=W
id(18)
The learning objective of network is for solving the parameter set θ minimizing network error (17)
opt, as shown in (19):
In figure 6, J (Θ
i) represent the network error that i-th layer of convolutional network calculates, wherein, Θ
irepresent by all network parameters of the 1st layer to i-th layer convolutional network and current layer unified dimensional linear mapping matrix W
iset, such as formula (20):
Wherein, θ
iwhat represent i-th layer of convolutional network can learning parameter set, and comprising relate in the operation of convolution operation, activation manipulation and down-sampling all can learning parameter.
(D) multi-layer Fusion Features and dimensionality reduction
As shown in Figure 6, feature
mergerepresent by each level unified dimensional proper vector f
ithe union feature vector that series connection is formed, namely
W
trepresent associating proper vector feature
mergecarry out the mapping matrix of linear dimensionality reduction mapping, f
trepresent by feature
mergemap the multi-feature vector obtained through linear dimensionality reduction, it comprises the eigenvector information of each hierarchical network, such as formula (22), n in formula
trepresent the f of setting
tdimension:
J (Θ
t) represent and distribute to multi-feature vector f
tthe network error of sorter network; Wherein, Θ
trepresent the set of all convolutional network parameters, all unified linear mapping matrixes and linear dimensionality reduction mapping matrix, such as formula (23):
(E) backpropagation of network error
The present invention utilizes BP algorithm to upgrade network parameter.
According to chain rule, network error propagates from back to front.
The differentiate of sorter network linear mapping:
I-th (i=1 ... 4, T) layer sorter network in can learning parameter be W
i, id, according to J (Θ
i) definition can have with chain type Rule for derivation:
J (Θ can be obtained simultaneously
i) about the derivative of f:
The differentiate of unified dimensional linear mapping:
Each unified dimensional linear mapping matrix W
ican to J (Θ
i) and J (Θ
t) two network error generation effects, therefore, utilize BP algorithm to W
iwhen upgrading, W
irenewal gradient be by J (Θ
i) to W
iderivative and J (Θ
t) to W
iderivative combine formation, meanwhile, in the training process, can give each network error a weight, to sum up, can W be obtained
irenewal gradient, such as formula (26):
Can have according to chain type Rule for derivation:
Therefore, can have
The linear dimensionality reduction of comprehensive characteristics layer maps differentiate:
The linear dimensionality reduction mapping matrix W of comprehensive characteristics layer
tonly to J (Θ
t) generation effect, easily obtain according to chain type Rule for derivation:
Meanwhile, the derivative that can calculate the input feature value of each level unified dimensional linear mapping is:
The differentiate of convolutional network parameter:
The parameter that can learn in convolutional network only has the convolution kernel in convolution operation, therefore, needs to calculate J (Θ
i) about the renewal gradient of each level convolution kernel c.Can have according to chain type Rule for derivation:
Wherein,
Location represents M
svalue at M
ain the binaryzation matrix of position, that is:
The above-mentioned principles and methods for utilizing the accumulation of multi-layer Fusion Features tax power degree of depth convolutional network to carry out the process of feature learning, provides concrete algorithm below, as shown in table 1:
Table 1 weighs for utilizing the accumulation of multi-layer Fusion Features to compose the process that degree of depth convolutional network carries out feature learning.
Just can carry out verification process of the present invention below:
(1) Image semantic classification
The present invention adopts the Face datection algorithm based on cascade Adaboost to carry out Face datection to image, then the facial modeling algorithm based on SDM is utilized to carry out positioning feature point to the face detected, and by image scaling, rotation and translation, face is corrected and normalization alignment, finally obtain the facial image being of a size of 100*100, and in the picture, the image coordinate of left eye is (30,30), the image coordinate of right eye is (30,70), as shown in Figure 3.
The present invention adopts simple gray scale normalization pre-service, and as shown in the formula (1), in formula, I (i, j) represents the gray-scale value of image (i, j).The fundamental purpose of gray scale normalization is convenient to network processes continuous data and avoids processing larger discrete grey value, thus avoid occurring abnormal conditions.
(2) feature extraction
Utilize the network extraction characteristics of image of having trained
Complete after the training of power degree of depth convolutional network is composed to the accumulation of multi-layer Fusion Features, just can utilize the feature of the network extraction input picture trained, as shown in table 2:
(3) aspect ratio pair
(I) absolute value normalization cosine value
Absolute value normalization cosine value (cosinenormalizedbyabsolutevalue, cos that the present invention proposes
aN) define such as formula (39):
Wherein,
Experiment shows, the contrast of absolute value normalization cosine value is responsive to the long difference of vector field homoemorphism, can make up that cosine similarity ignores the long difference of vector field homoemorphism and the difference that causes describes incomplete problem.
(II) the many differences based on LDA merge alignment algorithm
The cosine similarity of comparison feature, absolute value normalization cosine value and two character modules length are combined as a four-dimensional difference vector by the present invention, namely
f
diff(f
T1,f
T2)=[cos(f
T1,f
T2),cos
AN(f
T2,f
T2),|f
T2|,|f
T2|]
T(41)
Then utilize LDA (linear discriminant analysis) that four-dimensional difference vector is fused to one dimension analog quantity, namely, difference vector mapping matrix W
lDAfour-dimensional difference vector is needed to be mapped as one-dimensional vector.
sim(f
T1,f
T2)=W
LDAf
diff(42)
Wherein W
lDArepresent the mapping vector utilizing LDA to obtain.
The beneficial effect that embodiment of the present invention technical scheme is brought:
The present embodiment utilizes the accumulated weights degree of depth convolutional network of multi-layer Fusion Features to carry out feature learning and feature extraction, then utilize the many differences based on LDA to merge the feature of alignment algorithm to two width facial images to compare, there are following five advantages: one, the present invention utilizes convolutional network automatic learning and extracts feature, avoids the deficiency of manual features; Two, gradient disperse problem is avoided by the training of multistratum classification network association; Three, by multi-layer Fusion Features, increase characteristics of image richness, compensate for general degree of depth network insufficient to each hierarchy characteristic process, only utilize high-level characteristic to be not enough to the defect of abundant Description Image; Four, by increasing the flexibility ratio of e-learning to the weighting of multistratum classification network error; Five, by solving cosine similarity based on many differences fusion alignment algorithms of LDA, incomplete problem is portrayed to proper vector difference.FERET database is tested, and four word bank Fb, Fc, DupI, DupII achieve 99.9% respectively, 100%, 98.8%, the certification rate (it is 0.1% that mistake is sentenced rate) of 99.6%.
The above is the preferred embodiment of the present invention; it should be pointed out that for those skilled in the art, under the prerequisite not departing from principle of the present invention; can also make some improvements and modifications, these improvements and modifications also should be considered as protection scope of the present invention.