CN108898094A - A kind of face comparison method and system based on series connection integrated form RMML metric learning - Google Patents
A kind of face comparison method and system based on series connection integrated form RMML metric learning Download PDFInfo
- Publication number
- CN108898094A CN108898094A CN201810670839.4A CN201810670839A CN108898094A CN 108898094 A CN108898094 A CN 108898094A CN 201810670839 A CN201810670839 A CN 201810670839A CN 108898094 A CN108898094 A CN 108898094A
- Authority
- CN
- China
- Prior art keywords
- face characteristic
- matrix
- metric
- characteristic set
- face
- 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
- G06V40/172—Classification, e.g. identification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- 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
Abstract
The invention discloses a kind of face comparison methods and system based on series connection integrated form RMML metric learning, including:Characteristic extracting module, for extracting feature to facial image training set and test sample facial image;L concatenated integration modules, each integration module are used to upset face characteristic set grouping, by being spliced into new face characteristic set after group progress RMML metric learning, Feature Mapping;Metric module, for calculating the distance matrix metric for the face characteristic set that l-th integration module obtains based on RMML metric learning;Whether comparison module for calculating the distance of test sample facial image using distance matrix metric, and is the same person according to range estimation facial image.There are metric learning method provided by the invention closed solutions to be not required to invert, and effectively improve robustness;It by the multiple Linear Mapping and multiple Nonlinear Mapping of cascade process, promotes the separating capacity of metric matrix and avoids over-fitting, it is higher that face compares accuracy rate.
Description
Technical field
The invention belongs to computer visions and mode identification technology, integrated based on series connection more particularly, to one kind
The face comparison method and system of formula RMML (Robust Mahlanobis Metric learning) metric learning.
Background technique
Face alignment is the recognition of face form of one kind one to one, and the purpose is to whether judge two given facial images
It is the same person.Face alignment is in public safety system, human-computer interaction, e-commerce field extensive application.However in reality
In the scene of border, since face is influenced by external factors such as illumination, age, postures, same people is obtained in different environments
Face might have very big difference, it is therefore desirable to learn a kind of measurement that can reflect people to similarity definition, i.e., apart from degree
Amount study, learning distance metric are intended to learn mahalanobis distance measurement so that between the facial image of the same person away from
From becoming closer, and distance becomes remote between the facial image from different people, whether preferably judges two facial images with this
Belong to a people.
Although currently existing many metric learning methods, in face alignment field, existing data set is all very
Greatly, the metric learning method without closed solutions tends not to be applicable in, such as LMNN metric learning method.For there is closed solutions
Method, such as KISSme and XQDA, both methods requires to invert to covariance matrix.In face alignment, face characteristic is past
Toward be it is highly relevant, covariance matrix is usually irreversible.Both methods there are poor robustness, separating capacity is insufficient asks
Topic, it is difficult to be applied to face alignment field.
Summary of the invention
In view of the drawbacks of the prior art, it is an object of that present invention to provide a kind of face comparison method based on metric learning,
This method has higher robustness to face characteristic, has good adaptability for large-scale dataset, can adapt to extensive people
The requirement that face compares, and can effectively enhance the separating capacity of distance metric.
To achieve the above object, on the one hand, the present invention provides a kind of people based on series connection integrated form RMML metric learning
Face comparison method, this approach includes the following steps:
S1. to each sample x in facial image training setiExtract feature fi 0, obtain face characteristic setN is sample size;
S2. face characteristic in face characteristic set is upset into grouping, by group carrying out RMML metric learning, spell after Feature Mapping
It is connected in new face characteristic set, this sequence of operations is repeated L times, face characteristic set is obtained
S3. face characteristic set is calculated based on RMML metric learningDistance matrix metric
Mfinal;
S4. test sample facial image x is extractedIAnd xJFeatureWithUtilize distance matrix metric MfinalCalculate this
The distance d of two facial images;
S5. judge whether distance d is less than threshold value, if so, determining facial image xIAnd xJFor the same person, otherwise, it is determined that people
Face image xIAnd xJIt is not a people.
Specifically, the l times face characteristic set is upset into grouping, included the following steps:
S201. the face characteristic set before the l times being upsetIn each face characteristic fi l-1
Upsetting by dimension becomesObtain the face characteristic set after upsetting afterwards the l times
S202. feature after upsettingDimension be not 2 power, operated by zero padding and its dimension be extended to 2
Power after obtain zero padding after face characteristic set
S203. by the face characteristic after each zero paddingIt is divided into KlK is obtained after grouplA face characteristic set;Wherein, l
=1 ..., L, k=1,2 ..., Kl, Kl=2L-l, kth group face characteristic collection is combined into
For the face characteristic of i-th of sample in kth group face characteristic set.
Specifically, the input of the RMML metric learning is face characteristic set { F1,...,Fi,...,FN, FiIt is i-th
The face characteristic of a sample, the RMML metric learning include the following steps:
(1) judging characteristic is to (Fi,Fj) in feature FiAnd FjWhether be the same person feature, if so, yij=1, otherwise,
yij=0, i, j=1 ..., N;
(2) y is countedij=1 and yij=0 quantity, is denoted as n respectively1And n2, and
(3) feature is calculated to (Fi,Fj) Differential Characteristics dij=Fi-Fj;
(4) pass through objective functionSolve metric matrix M;
Wherein, λ g1Weight, the Parameter adjustable section, tr () is trace function, and the column vector of matrix A is that the same person is special
Levy differential pair dij, the column vector of matrix B is different people signature differential to dij, | | | |FFor F norm, I is unit matrix.
Specifically, it is spliced into new face characteristic set after the l times Feature Mapping, includes the following steps:
(1) for each metric matrix MlkCarry out Schur decomposition, Mlk=Q Λ QT, wherein MlkIt is carried out for the l times by group
Kth group face characteristic set after RMML metric learningMetric matrix;
(2) matrix Λ negative zero setting is obtainedIt willWith matrix Q, QTMultiplication obtains matrix
(3) to each matrixCholesky decomposition is carried out, is obtained
(4) matrix P is utilizedlkTo each featureLinear Mapping is carried out, kth group characteristic set after Linear Mapping is obtained
(5) nonlinear function is utilizedTo the feature after each Linear MappingIt obtains non-linear
Kth group characteristic set after mappingWherein, sign () is sign function;
(6) will It is spliced into characteristic setByIt is spliced.
Specifically, step S4 includes the following steps:
(1) test sample facial image x is extractedIAnd xJFeatureWith
(2) according in face characteristic set face characteristic upset sequence, by face characteristicWithUpset grouping, by
Group is spliced into new face characteristic set after carrying out RMML metric learning, Feature Mapping, this sequence of operations is repeated L times, is obtained
To face characteristicWith
(3) distance matrix metric M is utilizedfinalThe distance d of this two facial images is calculated, calculation formula is:D=
fI LMfinalfJ L。
To achieve the above object, on the other hand, the present invention provides a kind of based on series connection integrated form RMML metric learning
Face alignment system, the system comprise the following modules:
Characteristic extracting module, for each sample x in facial image training setiExtract feature fi 0Obtain face characteristic
SetAnd to test sample facial image xIAnd xJExtract feature fIAnd fJ, N is sample size;
L concatenated integration modules, each integration module are used to upset face characteristic grouping, by a group progress RMML measurement
New face characteristic set is spliced into after study, Feature Mapping;
Metric module, for calculating the face characteristic set that l-th integration module obtains based on RMML metric learningDistance matrix metric Mfinal;
Comparison module, for utilizing distance matrix metric MfinalCalculate test sample facial image distance d, and judge away from
Whether it is less than threshold value from d, if so, determining facial image xIAnd xJFor the same person, otherwise, it is determined that facial image xIAnd xJIt is not
One people.
Specifically, input is upset grouping by each integration module, is included the following steps:
(1) the face characteristic set before upsetting first of integration moduleIn each face it is special
Levy fi l-1Upsetting by dimension becomesThe face characteristic set after first of integration module is upset is obtained afterwards
(2) feature after upsettingDimension be not 2 power, operated by zero padding and its dimension be extended to 2
The face characteristic set after zero padding is obtained after power
(3) by the face characteristic after each zero paddingIt is divided into KlK is obtained after grouplA face characteristic set;
Wherein, l=1 ..., L, k=1,2 ..., Kl, Kl=2L-l, kth group face characteristic collection is combined intoFor the face characteristic of i-th of sample in kth group face characteristic set.
Specifically, the input of the RMML metric learning is face characteristic set { F1,...,Fi,...,FN, FiIt is i-th
The face characteristic of a sample, the RMML metric learning include the following steps:
(1) judging characteristic is to (Fi,Fj) in feature FiAnd FjWhether be the same person feature, if so, yij=1, otherwise,
yij=0, i, j=1 ..., N;
(2) y is countedij=1 and yij=0 quantity, is denoted as n respectively1And n2, and
(3) feature is calculated to (Fi,Fj) Differential Characteristics dij=Fi-Fj;
(4) pass through objective functionSolve metric matrix M;
Wherein, λ g1Weight, the Parameter adjustable section, tr () is trace function, and the column vector of matrix A is that the same person is special
Levy differential pair dij, the column vector of matrix B is different people signature differential to dij, | | | |FFor F norm, I is unit matrix.
Specifically, each integration module will be spliced into new face characteristic set, including following step after Feature Mapping
Suddenly:
(1) for each metric matrix MlkCarry out Schur decomposition, Mlk=Q Λ QT, wherein MlkIt is carried out for the l times by group
Kth group face characteristic set after RMML metric learningMetric matrix;
(2) matrix Λ negative zero setting is obtainedIt willIt is multiplied to obtain matrix with matrix Q, QT
(3) to each matrixCholesky decomposition is carried out, is obtained
(4) matrix P is utilizedlkTo each featureLinear Mapping is carried out, kth group characteristic set after Linear Mapping is obtained
(5) nonlinear function is utilizedTo the feature after each Linear MappingIt obtains non-linear
Kth group characteristic set after mappingWherein, sign () is sign function;
(6) will It is spliced into characteristic setByIt is spliced.
Specifically, the comparison module utilizes the distance matrix metric M learntfinalCalculate this two facial images
Distance d, include the following steps:
(1) upset sequence according to face characteristic in face characteristic set, using L concatenated integration modules by face characteristicWithUpset grouping, by new face characteristic set is spliced into after group progress RMML metric learning, Feature Mapping, obtains people
Face featureWith
(2) distance matrix metric M is utilizedfinalThe distance d of this two facial images is calculated, calculation formula is:
In general, through the invention it is contemplated above technical scheme is compared with the prior art, have below beneficial to effect
Fruit:
(1) present invention proposes a kind of RMML metric learning method, which has closed solutions, in metric matrix
It does not need to invert in solution procedure, so that this method has stronger robustness to face characteristic, have very for large-scale dataset
Good adaptability.
(2) present invention passes through series connection integrated learning mechanism, integrates multiple metric matrixs in each integration module, then will not
Same integration module is connected, and is carried out multiple Linear Mapping to face characteristic in cascade process, can effectively be enhanced metric matrix
Separating capacity;
(3) present invention carries out RMML study to every group of feature after upsetting by upsetting and being grouped face characteristic, and
Multiple Nonlinear Mapping is carried out to face characteristic in cascade process, effectively avoids data over-fitting, so that face comparison is quasi-
True rate is higher.
Detailed description of the invention
Fig. 1 is a kind of face alignment system knot based on series connection integrated form RMML metric learning provided in an embodiment of the present invention
Structure schematic diagram.
Fig. 2 is step S2 flow diagram provided in an embodiment of the present invention.
Fig. 3 is face characteristic provided in an embodiment of the present invention to distribution schematic diagram.
Fig. 4 (a) is the face characteristic visualization distribution signal of the 1st integration module provided in an embodiment of the present invention output
Figure;
Fig. 4 (b) is the face characteristic visualization distribution signal of l-th integration module provided in an embodiment of the present invention output
Figure.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.
Fig. 1 is a kind of face alignment system knot based on series connection integrated form RMML metric learning provided in an embodiment of the present invention
Structure schematic diagram.As shown in Figure 1, the system comprises the following modules:
Characteristic extracting module, for each sample x in facial image training setiExtract feature fi 0Obtain face characteristic
SetAnd to test sample facial image xIAnd xJExtract feature fIAnd fJ, N is sample size;
L concatenated integration modules, each integration module are used to upset face characteristic grouping, by a group progress RMML measurement
New face characteristic set is spliced into after study, Feature Mapping;
Metric module, for calculating the face characteristic set that l-th integration module obtains based on RMML metric learningDistance matrix metric Mfinal;
Comparison module, for utilizing distance matrix metric MfinalCalculate test sample facial image distance d, and judge away from
Whether it is less than threshold value from d, if so, determining facial image xIAnd xJFor the same person, otherwise, it is determined that facial image xIAnd xJIt is not
One people.
The method of characteristic extracting module can be prior art any particular algorithms, for example, convolutional neural networks feature, Fei Xuexiang
Measure feature etc..
Utilize distance matrix metric MfinalCalculate training set sample to the distance between, choose different threshold calculations TPR
(true positive rate, real class rate) and FNR (false positive rate, false positive class rate) draw ROC curve
((receiver operating characteristic curve, receiver operating characteristic curve), ERR point in ROC curve
Corresponding threshold value, the threshold value as comparison module.
Accordingly, a kind of face comparison method based on series connection integrated form RMML metric learning, this method includes following step
Suddenly:
S1. to each sample x in facial image training setiExtract feature fi 0, obtain face characteristic setN is sample size;
S2. face characteristic in face characteristic set is upset into grouping, by group carrying out RMML metric learning, spell after Feature Mapping
It is connected in new face characteristic set, this sequence of operations is repeated L times, face characteristic set is obtained
S3. face characteristic set is calculated based on RMML metric learningDistance matrix metric
Mfinal;
S4. test sample facial image x is extractedIAnd xJFeatureWithUtilize distance matrix metric MfinalCalculate this
The distance d of two facial images;
S5. judge whether distance d is less than threshold value, if so, determining facial image xIAnd xJFor the same person, otherwise, it is determined that people
Face image xIAnd xJIt is not a people.
Fig. 2 is step S2 flow diagram provided in an embodiment of the present invention.As shown in Fig. 2, being divided into four parts:To input
Face characteristic, successively carry out upsetting grouping, metric learning, Feature Mapping;L integration module is sequentially connected in series;The last one
Integration module connects metric module, directly progress metric learning;And export face characteristic to the distance between.
Input is upset grouping by each integration module described in step S2, is included the following steps:
The input of (1) first of integration module is face characteristic setEach face is special
Levy fi l-1Upsetting at random by dimension becomesCharacteristic set after being upset
(2) if feature after upsettingDimension be not 2 power, operated using zero padding and its dimension be extended to 2
Power, the characteristic set after obtaining zero padding
For example, face characteristic fiFor 598 dimensions, it need to guarantee that every group of characteristic length is consistent in grouping, for cannot uniformly divide
It the case where group, is operated using zero padding, it need to be by its zero padding at 640 dimensions.It is divided into 4 groups, every group of characteristic length is 160 dimensions.
(3) by the feature after each zero paddingIt is divided into KlGroup obtains as KlA face characteristic set, wherein l=
1 ..., L, k=1,2 ..., Kl, Kl=2L-l, kth group face characteristic collection is combined into
For the face characteristic of i-th of sample in kth group face characteristic set.
The characteristic set of input, which is upset grouping, can effectively prevent over-fitting.
RMML metric learning is that the present invention proposes that a kind of metric learning method with closed solutions, motivation are desirable to study one
A metric matrix is distributed near origin the signature differential of same people under this measurement as far as possible;For the spy of different people
Levy difference, it is desirable to which it is away as far as possible origin.
The input of the RMML metric learning is face characteristic set { F1,...,Fi,...,FN, FiFor i-th sample
Face characteristic, the RMML metric learning include the following steps:
(1) judging characteristic is to (Fi,Fj) in feature FiAnd FjWhether be the same person feature, if so, yij=1, otherwise,
yij=0, i, j=1 ..., N;
(2) y is countedij=1 and yij=0 quantity, is denoted as n respectively1And n2, and
(3) feature is calculated to (Fi,Fj) Differential Characteristics dij=Fi-Fj;
(4) pass through objective functionSolve metric matrix M;
Wherein, λ g1Weight, the Parameter adjustable section, tr () is trace function, and the column vector of matrix A is that the same person is special
Levy differential pair dij, the column vector of matrix B is different people signature differential to dij, | | | |FFor F norm, I is unit matrix.
Distinguish item g1Keep the characteristic distance of the same person and different people as separated as possible.
Item g is lost in regularization2Make metric matrix M should not be too wide in the gap with unit matrix I.When M is equal to I, distance metric is sloughed off
Turn to Euclidean distance.
Matrix M is the metric matrix that we learn.The metric learning matrix has closed solutions, and does not need to matrix
It inverts, there is stronger adaptability and robustness.
It is spliced into new face characteristic set after the l times Feature Mapping, includes the following steps:
(1) for each metric matrix MlkCarry out Schur decomposition, Mlk=Q Λ QT, wherein MlkIt is carried out for the l times by group
Kth group face characteristic set after RMML metric learningMetric matrix;
(2) matrix Λ negative zero setting is obtainedIt willWith matrix Q, QTMultiplication obtains matrix
(3) to each matrixCholesky decomposition is carried out, is obtained
(4) matrix P is utilizedlkTo each featureLinear Mapping is carried out, kth group characteristic set after Linear Mapping is obtained
(5) nonlinear function is utilizedTo the feature after each Linear MappingIt obtains non-linear
Kth group characteristic set after mappingWherein, sign () is sign function;
(6) will It is spliced into characteristic setByIt is spliced.
Step S4 includes the following steps:
(1) test sample facial image x is extractedIAnd xJFeatureWith
(2) according in face characteristic set face characteristic upset sequence, by face characteristicWithUpset grouping, by
Group is spliced into new face characteristic set after carrying out RMML metric learning, Feature Mapping, this sequence of operations is repeated L times, is obtained
To face characteristicWith
(3) distance matrix metric M is utilizedfinalThe distance d of this two facial images is calculated, calculation formula is:D=
fI LMfinalfJ L。
Fig. 3 is face characteristic provided by the invention to distribution schematic diagram.As shown in figure 3,0 represents feature FiAnd FjFrom not
Same people, 1 represents feature FiAnd FjFrom the same person.The purpose of RMML metric learning is most for the signature differential of same people
It is likely distributed near origin;For the signature differential of different people, it is desirable to which it is away as far as possible origin.
Fig. 4 (a) is the face characteristic visualization distribution signal of the 1st integration module provided in an embodiment of the present invention output
Figure, Fig. 4 (b) are that the face characteristic of l-th integration module provided in an embodiment of the present invention output visualizes distribution schematic diagram.Such as figure
Shown in 4, after the processing of L-1 integration module, the feature of the same person is furthered, and the feature of different people is zoomed out.
Training set includes 2576 pairs of positive samples (same people), 2424 pairs of negative samples (different people) in embodiment.The spy of extraction
Sign uses CNN feature.Test set includes 3235 pairs of positive samples, 1765 pairs of negative samples.
Based on the training set and test set, take respectively series connection integrated form RMML (ECRMML) proposed by the present invention,
LMNN, KISSme, XQDA method carry out face alignment, and accuracy rate is as described in Table 1.Wherein, it is integrated in series connection integrated form RMML
The number of module is 2.
Method | LMNN | KISSme | XQDA | ECRMML |
EER | 20.34% | 30.57% | 33.04% | 10.20% |
Table 1
EER (Equal Error Rate) is the index of evaluation algorithms effect.The smaller algorithm of EER value is better.Comparison sheet 1 can
Know, compared to the prior art, it is higher that face compares accuracy rate to the present invention.
More than, the only preferable specific embodiment of the application, but the protection scope of the application is not limited thereto, and it is any
Within the technical scope of the present application, any changes or substitutions that can be easily thought of by those familiar with the art, all answers
Cover within the scope of protection of this application.Therefore, the protection scope of the application should be subject to the protection scope in claims.
Claims (10)
1. a kind of face comparison method based on series connection integrated form RMML metric learning, which is characterized in that this method includes following
Step:
S1. to each sample x in facial image training setiExtract feature fi 0, obtain face characteristic setN is sample size;
S2. face characteristic in face characteristic set is upset into grouping, by group carrying out RMML metric learning, be spliced into after Feature Mapping
This sequence of operations is repeated L times, obtains face characteristic set by new face characteristic set
S3. face characteristic set is calculated based on RMML metric learningDistance matrix metric Mfinal;
S4. test sample facial image x is extractedIAnd xJFeatureWithUtilize distance matrix metric MfinalCalculate this two
The distance d of facial image;
S5. judge whether distance d is less than threshold value, if so, determining facial image xIAnd xJFor the same person, otherwise, it is determined that face figure
As xIAnd xJIt is not a people.
2. face comparison method as described in claim 1, which is characterized in that the l times face characteristic set is upset grouping, wrapped
Include following steps:
S201. the face characteristic set before the l times being upsetIn each face characteristic fi l-1By dimension
Degree, which is upset, becomes fi′(l-1)Obtain the face characteristic set after upsetting afterwards the l times
S202. feature f after upsettingi′(l-1)Dimension be not 2 power, operated by zero padding and its dimension be extended to 2
The face characteristic set after zero padding is obtained after power
S203. by the face characteristic f after each zero paddingi *(l-1)It is divided into KlK is obtained after grouplA face characteristic set;
Wherein, l=1 ..., L, k=1,2 ..., Kl, Kl=2L-l, kth group face characteristic collection is combined into For the face characteristic of i-th of sample in kth group face characteristic set.
3. face comparison method as claimed in claim 1 or 2, which is characterized in that the input of the RMML metric learning is people
Face characteristic set { F1,...,Fi,...,FN, FiFor the face characteristic of i-th of sample, the RMML metric learning includes following
Step:
(1) judging characteristic is to (Fi,Fj) in feature FiAnd FjWhether be the same person feature, if so, yij=1, otherwise, yij=
0, i, j=1 ..., N;
(2) y is countedij=1 and yij=0 quantity, is denoted as n respectively1And n2, and
(3) feature is calculated to (Fi,Fj) Differential Characteristics dij=Fi-Fj;
(4) pass through objective functionSolve metric matrix M;
Wherein, λ g1Weight, the Parameter adjustable section, tr () is trace function, and the column vector of matrix A is that same person's feature is poor
Divide to dij, the column vector of matrix B is different people signature differential to dij, | | | |FFor F norm, I is unit matrix.
4. face comparison method as claimed in claim 1 or 2, which is characterized in that be spliced into new people after the l times Feature Mapping
Face characteristic set, includes the following steps:
(1) for each metric matrix MlkCarry out Schur decomposition, Mlk=Q Λ QT, wherein MlkRMML degree is carried out by group for the l times
Kth group face characteristic set after amount studyMetric matrix;
(2) matrix Λ negative zero setting is obtainedIt willWith matrix Q, QTMultiplication obtains matrix
(3) to each matrixCholesky decomposition is carried out, is obtained
(4) matrix P is utilizedlkTo each featureLinear Mapping is carried out, kth group characteristic set after Linear Mapping is obtained
(5) nonlinear function is utilizedTo the feature after each Linear MappingObtain Nonlinear Mapping
Kth group characteristic set afterwardsWherein, sign () is sign function;
(6) will It is spliced into characteristic setfi lByIt is spliced.
5. face comparison method as claimed in claim 1 or 2, which is characterized in that step S4 includes the following steps:
(1) test sample facial image x is extractedIAnd xJFeatureWith
(2) according in face characteristic set face characteristic upset sequence, by face characteristicWithUpset grouping, by group into
It is spliced into new face characteristic set after row RMML metric learning, Feature Mapping, this sequence of operations is repeated L times, people is obtained
Face featureWith
(3) distance matrix metric M is utilizedfinalThe distance d of this two facial images is calculated, calculation formula is:D=fI LMfinalfJ L。
6. a kind of face alignment system based on series connection integrated form RMML metric learning, which is characterized in that the system includes following
Module:
Characteristic extracting module, for each sample x in facial image training setiExtract feature fi 0Obtain face characteristic setAnd to test sample facial image xIAnd xJExtract feature fIAnd fJ, N is sample size;
L concatenated integration modules, each integration module are used to upset face characteristic grouping, by a group progress RMML tolerance
It practises, be spliced into new face characteristic set after Feature Mapping;
Metric module, for calculating the face characteristic set that l-th integration module obtains based on RMML metric learningDistance matrix metric Mfinal;
Comparison module, for utilizing distance matrix metric MfinalThe distance d of test sample facial image is calculated, and judges distance d
Whether threshold value is less than, if so, determining facial image xIAnd xJFor the same person, otherwise, it is determined that facial image xIAnd xJIt is not one
People.
7. face alignment system as claimed in claim 6, which is characterized in that each integration module upsets input point
Group includes the following steps:
(1) the face characteristic set before upsetting first of integration moduleIn each face characteristic
fi l-1Upsetting by dimension becomes fi′(l-1)The face characteristic set after first of integration module is upset is obtained afterwards
(2) feature f after upsettingi′(l-1)Dimension be not 2 power, operated by zero padding its dimension is extended to 2 power
The face characteristic set after zero padding is obtained after secondary
(3) by the face characteristic f after each zero paddingi *(l-1)It is divided into KlK is obtained after grouplA face characteristic set;
Wherein, l=1 ..., L, k=1,2 ..., Kl, Kl=2L-l, kth group face characteristic collection is combined into For the face characteristic of i-th of sample in kth group face characteristic set.
8. face alignment system as claimed in claims 6 or 7, which is characterized in that the input of the RMML metric learning is people
Face characteristic set { F1,...,Fi,...,FN, FiFor the face characteristic of i-th of sample, the RMML metric learning includes following
Step:
(1) judging characteristic is to (Fi,Fj) in feature FiAnd FjWhether be the same person feature, if so, yij=1, otherwise, yij=
0, i, j=1 ..., N;
(2) y is countedij=1 and yij=0 quantity, is denoted as n respectively1And n2, and
(3) feature is calculated to (Fi,Fj) Differential Characteristics dij=Fi-Fj;
(4) pass through objective functionSolve metric matrix M;
Wherein, λ g1Weight, the Parameter adjustable section, tr () is trace function, and the column vector of matrix A is that same person's feature is poor
Divide to dij, the column vector of matrix B is different people signature differential to dij, | | | |FFor F norm, I is unit matrix.
9. face alignment system as claimed in claims 6 or 7, which is characterized in that each integration module is by Feature Mapping
It is spliced into new face characteristic set afterwards, includes the following steps:
(1) for each metric matrix MlkCarry out Schur decomposition, Mlk=Q Λ QT, wherein MlkRMML degree is carried out by group for the l times
Kth group face characteristic set after amount studyMetric matrix;
(2) matrix Λ negative zero setting is obtainedIt willWith matrix Q, QTMultiplication obtains matrix
(3) to each matrixCholesky decomposition is carried out, is obtained
(4) matrix P is utilizedlkTo each featureLinear Mapping is carried out, kth group characteristic set after Linear Mapping is obtained
(5) nonlinear function is utilizedTo the feature after each Linear MappingObtain Nonlinear Mapping
Kth group characteristic set afterwardsWherein, sign () is sign function;
(6) will It is spliced into characteristic setfi lByIt is spliced.
10. face alignment system as claimed in claims 6 or 7, which is characterized in that the comparison module, which utilizes, have been learnt to arrive
Distance matrix metric MfinalThe distance d for calculating this two facial images, includes the following steps:
(1) upset sequence according to face characteristic in face characteristic set, using L concatenated integration modules by face characteristicWithUpset grouping, by new face characteristic set is spliced into after group progress RMML metric learning, Feature Mapping, it is special to obtain face
SignWith
(2) distance matrix metric M is utilizedfinalThe distance d of this two facial images is calculated, calculation formula is:D=fI LMfinalfJ L。
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810670839.4A CN108898094B (en) | 2018-06-26 | 2018-06-26 | Face comparison method and system based on series integrated RMML metric learning |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810670839.4A CN108898094B (en) | 2018-06-26 | 2018-06-26 | Face comparison method and system based on series integrated RMML metric learning |
Publications (2)
Publication Number | Publication Date |
---|---|
CN108898094A true CN108898094A (en) | 2018-11-27 |
CN108898094B CN108898094B (en) | 2020-06-05 |
Family
ID=64346003
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810670839.4A Active CN108898094B (en) | 2018-06-26 | 2018-06-26 | Face comparison method and system based on series integrated RMML metric learning |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108898094B (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104123560A (en) * | 2014-07-03 | 2014-10-29 | 中山大学 | Phase encoding characteristic and multi-metric learning based vague facial image verification method |
US9514356B2 (en) * | 2014-09-05 | 2016-12-06 | Huawei Technologies Co., Ltd. | Method and apparatus for generating facial feature verification model |
CN106682606A (en) * | 2016-12-23 | 2017-05-17 | 湘潭大学 | Face recognizing method and safety verification apparatus |
KR101817773B1 (en) * | 2016-09-26 | 2018-01-12 | 동의대학교 산학협력단 | An Advertisement Providing System By Image Processing of Depth Information |
-
2018
- 2018-06-26 CN CN201810670839.4A patent/CN108898094B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104123560A (en) * | 2014-07-03 | 2014-10-29 | 中山大学 | Phase encoding characteristic and multi-metric learning based vague facial image verification method |
US9514356B2 (en) * | 2014-09-05 | 2016-12-06 | Huawei Technologies Co., Ltd. | Method and apparatus for generating facial feature verification model |
KR101817773B1 (en) * | 2016-09-26 | 2018-01-12 | 동의대학교 산학협력단 | An Advertisement Providing System By Image Processing of Depth Information |
CN106682606A (en) * | 2016-12-23 | 2017-05-17 | 湘潭大学 | Face recognizing method and safety verification apparatus |
Non-Patent Citations (4)
Title |
---|
CHEN Y, CHEN Y, WANG X: ""Deep learning face representation by joint identification-verification"", 《INTERNATIONAL CONFERENCE ON NEURAL INFORMATION PROCESSING SYSTEMS. MIT PRESS》 * |
JIANHONG MA, HAN ZHANG AND WEI SHE: "Research on robust face recognition based on depth image sets", 《2017 2ND INTERNATIONAL CONFERENCE ON IMAGE, VISION AND COMPUTING (ICIVC)》 * |
景晨凯,宋涛,庄雷,刘刚,王乐,刘凯伦: ""基于深度卷积神经网络的人脸识别技术综述"", 《计算机应用与软件》 * |
杨飞,苏剑波: "人脸显性特征的融合构造方法及识别", 《电子学报》 * |
Also Published As
Publication number | Publication date |
---|---|
CN108898094B (en) | 2020-06-05 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
EP3074918B1 (en) | Method and system for face image recognition | |
CN108564129B (en) | Trajectory data classification method based on generation countermeasure network | |
WO2019228317A1 (en) | Face recognition method and device, and computer readable medium | |
CN111814661B (en) | Human body behavior recognition method based on residual error-circulating neural network | |
CN107103281A (en) | Face identification method based on aggregation Damage degree metric learning | |
CN106228142A (en) | Face verification method based on convolutional neural networks and Bayesian decision | |
CN105205501B (en) | A kind of weak mark image object detection method of multi classifier combination | |
US20180038909A1 (en) | Analog circuit fault diagnosis method using single testable node | |
CN104408405B (en) | Face representation and similarity calculating method | |
CN110503000B (en) | Teaching head-up rate measuring method based on face recognition technology | |
CN114664452B (en) | General multi-disease prediction system based on causal verification data generation | |
CN110097033B (en) | Single-sample face recognition method based on feature expansion | |
Zhang et al. | Learning upper patch attention using dual-branch training strategy for masked face recognition | |
CN105893947A (en) | Bi-visual-angle face identification method based on multi-local correlation characteristic learning | |
CN112257741B (en) | Method for detecting generative anti-false picture based on complex neural network | |
CN105868711B (en) | Sparse low-rank-based human behavior identification method | |
CN106780422A (en) | A kind of notable figure fusion method based on Choquet integrations | |
CN103714340A (en) | Self-adaptation feature extracting method based on image partitioning | |
CN114492634A (en) | Fine-grained equipment image classification and identification method and system | |
CN102289679B (en) | Method for identifying super-resolution of face in fixed visual angle based on related characteristics and nonlinear mapping | |
US11587345B2 (en) | Image identification device, method for performing semantic segmentation, and storage medium | |
CN108898094A (en) | A kind of face comparison method and system based on series connection integrated form RMML metric learning | |
Tran et al. | Baby learning with vision transformer for face recognition | |
Tang et al. | Learning Hough regression models via bridge partial least squares for object detection | |
TW202119354A (en) | Classification and recognition method for pills capable of precisely detecting various locations for one or multiple pills during a speedy image inputting and performing as classifying and recognizing accordingly |
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 |