CN106503648A - Face identification method and device based on sparse projection binary-coding - Google Patents
Face identification method and device based on sparse projection binary-coding Download PDFInfo
- Publication number
- CN106503648A CN106503648A CN201610917123.0A CN201610917123A CN106503648A CN 106503648 A CN106503648 A CN 106503648A CN 201610917123 A CN201610917123 A CN 201610917123A CN 106503648 A CN106503648 A CN 106503648A
- Authority
- CN
- China
- Prior art keywords
- vector
- object function
- matrix
- pixel
- formula
- 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.)
- Pending
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
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of face identification method and device based on sparse projection binary-coding.The face identification method based on sparse projection binary-coding that the present invention is provided, the corresponding first pixel value difference vector of each pixel including obtaining each training sample in training set;The corresponding first binary feature vector of each pixel of the satisfaction based on the first object function of sparse projection matrix is obtained according to the first pixel value difference vector;All of first binary feature vector is clustered, multiple cluster centre words are obtained;Obtain the corresponding primary vector of each training sample;The corresponding secondary vector of image of face to be detected is obtained, according to primary vector and secondary vector, face recognition result is obtained.The face identification method based on sparse projection binary-coding of the present embodiment and device, recognition speed are fast, solve the problems, such as training sample overfitting, and good to the adaptability of data, improve accuracy and the rapidity of identification face.
Description
Technical field
A kind of the present invention relates to biological identification technology, more particularly to face identification method based on sparse projection binary-coding
And device.
Background technology
From primitive society to information-intensive society now, during society is from publicly-owned to privately owned, for individual privacy
Protection demand grows with each passing day.Increasing field needs to use reliable biological identification technology, and face is special as one kind
Biological characteristic, with feature unique, relatively stable, obtain easy and possess the remarkable advantage such as untouchable and extensively closed
Note.
Face recognition algorithms of the prior art mainly include the face identification method for representing face characteristic based on real number value
With the face identification method based on dense binary feature vector representation face characteristic.Wherein, face characteristic is represented based on real number value
Face identification method refer to extract from the facial image of input using the object function that is pre-designed and represented based on real number value
Face characteristic, then Land use models recognition methodss judge which people input picture belongs to.It is based on dense binary feature vector table
Show that the face identification method of face characteristic is referred to be converted into original based on dense two based on the face characteristic that real number value is represented
The face characteristic of value tag vector representation, the projection matrix for using are dense projection matrix (the most elements i.e. in matrix
For non-zero element), finally judge which people input picture belongs to using mode identification method.
But, for the face identification method for representing face characteristic based on real number value, on the one hand need substantial amounts of storage single
Unit, calculating speed are slow, and general computer or mobile device are difficult to meet its storage and calculating demand;On the other hand, the method
Localized variation in for facial image is sensitive, when the change degree of change of same facial image exceedes a certain threshold value
When, the feature identification that grader will originally belong to same person by two is different people.Based on dense binary feature vector
Represent in the face identification method of face characteristic, due to the impact that there is quantization error in two-value quantizing process all the time, equivalent
When the dimension of the binary feature vector after change is less than the dimension of original feature vector, it is likely that cause losing for a large amount of discriminant informations
Lose, so as to affect the ability of binary feature vector description facial image feature;Additionally, being leted others have a look at based on dense binary feature vector table
Also there is the over-fitting of training sample in the face identification method of face feature, and over-fitting can cause face identification method in meter
Bad adaptability during calculation to data, so that affect the performance of face identification method.
Content of the invention
The present invention provides a kind of face identification method and device based on sparse projection binary-coding, to overcome prior art
In face identification method recognition speed is slow, the technical problem of training sample overfitting and the bad adaptability to data.
The present invention provides a kind of face identification method based on sparse projection binary-coding, including:
The corresponding first pixel value difference vector of each pixel of each training sample in acquisition training set;Wherein, the training
Sample is facial image, and the training set includes several different facial images;
The each institute for meeting the first object function based on sparse projection matrix is obtained according to first pixel value difference vector
State the corresponding first binary feature vector of pixel;
The all of first binary feature vector is clustered, multiple cluster centre-words are obtained;
By the first binary feature vector the first matrix of composition corresponding for each training sample, using the word to institute
State the first matrix linearly to be rebuild, obtain the first linear reconstructed results, first is obtained according to the described first linear reconstructed results
Vector;Wherein, the corresponding primary vector of each training sample;
Obtain each pixel corresponding second binary feature vector of the image of face to be detected, and by each described 2nd 2
Value tag vector the second matrix of composition, is linearly rebuild to second matrix using the word, is obtained the second linear weight
Result is built, and secondary vector is obtained according to the described second linear reconstructed results;
According to the primary vector and the secondary vector, face recognition result is obtained.
Method as above, the first object function is as shown in formula one:
Wherein, R be sparse projection matrix, | R |0≤ m represents that the number of nonzero element in sparse projection matrix R is less than or equal to
M, B is that the first binary feature is vectorial, and parameter m has incidence relation with the sparse degree of the sparse projection matrix R, and m is for just
Integer, X are first pixel value difference vector.
Method as above, described by the first binary feature vector the first square of composition corresponding for each training sample
Battle array, is linearly rebuild to first matrix using the word, is obtained the first linear reconstructed results, according to the First Line
Property reconstructed results obtain primary vector;Wherein, the corresponding primary vector of each training sample, including:
First matrix is linearly rebuild using the word, obtained the first linear reconstructed results, described first
The expression formula of linear reconstructed results is as shown in formula two:
Bi=ai1S1+ai2S2+...+aikSkFormula two;
Wherein, BiFor the first matrix of each first binary feature vector composition of i-th training sample, SkSingle for k-th
Word, aikIt is k-th word when corresponding described first matrix of i-th training sample is linearly rebuild using the word
Weight;
Corresponding for each described word in the expression formula of the described first linear reconstructed results weight is constituted primary vector
(ai1, ai2... ..., aik).
Method as above, corresponding second binary feature of each pixel of the image of acquisition face to be detected to
Amount, and by each second binary feature vector, second matrix of composition, second matrix is carried out linearly using the word
Rebuild, obtain the second linear reconstructed results, secondary vector is obtained according to the described second linear reconstructed results, including:
Each pixel corresponding second pixel value difference vector of the image of face to be detected is obtained, and according to second picture
Plain difference value vector obtains each pixel corresponding described second of the image of the face to be detected for meeting the first object function
Binary feature vector;
Each second binary feature vector is constituted second matrix, and second matrix is adopted the word
Linearly rebuild, obtained the described second linear reconstruction structure, expression formula such as three institute of formula of the second linear reconstructed results
Show:
BTest=b1S1+b2S2+...+bkSkFormula three;
Wherein, BTestFor the second matrix of each second binary feature vector composition, SkFor k-th word, bkFor described
The weight of k-th word when second matrix is linearly rebuild using the word;
By corresponding for each described word in the expression formula of the described second linear reconstructed results weight composition described second to
Amount (b1, b2... ..., bk).
Method as above, described according to the primary vector and the secondary vector, face recognition result is obtained, is wrapped
Include:
By the primary vector and secondary vector input grader;
Obtain the secondary vector of the grader return and the Euclidean distance of each primary vector;
Facial image corresponding to training sample corresponding with the primary vector that the Euclidean distance of secondary vector is most short is known
Not Wei face to be detected image.
Method as above, obtain corresponding first pixel value difference of each pixel of each training sample in training set to
Amount, including:
Each training sample is divided into multiple pieces;
With each pixel in each block as the first central pixel point, with r as radius, first center pixel is obtained
First neighborhood territory pixel point of point, and according to clockwise, by the pixel value and described first of the first neighborhood territory pixel point
The pixel value of imago vegetarian refreshments carries out difference operation, obtain length for (2 × r+1) × (2 × r+1) -1 dimension the first pixel value difference to
Amount;
Each pixel corresponding second pixel value difference vector of the image of face to be detected is obtained, including:
It it is multiple pieces by the image division of the face to be detected;
With each pixel in each block as the second central pixel point, with r as radius, second center pixel is obtained
Second neighborhood territory pixel point of point, and according to clockwise, by the pixel value and described second of the second neighborhood territory pixel point
The pixel value of imago vegetarian refreshments carries out difference operation, obtain length for (2 × r+1) × (2 × r+1) -1 dimension the second pixel value difference to
Amount;
Wherein, r is positive integer.
Method as above, described acquisition according to first pixel value difference vector are met based on sparse projection matrix
The corresponding first binary feature vector of each described pixel of first object function, including:
The agent matrix S of R is introduced, by the first object functionIt is transformed to
Second object function, second object function is as shown in formula four:
Wherein, variable α represents the penalty factor of agent matrix S, in order to equation of equilibrium four inWithThis two variable, | S |0Represent the number of non-zero element in agent matrix S, parameter m and the sparse projection square
The sparse degree of battle array R has incidence relation;The quantization error between agent matrix S and binary-coding B is represented,Represent the error between agent matrix S and R;
Second object function is solved, the first binary feature vector is obtained.
Method as above, described second object function is solved, obtain first binary feature to
Amount, including:
Initial value S is given at random by the S and B in second object function0And B0, and fixation second object function
In S and B, update the R in second object function, obtain R1, specially:
Second object function is write the first expression formula instead, first expression formula is as shown in formula five:
Wherein, C1=SX, is fixed value;
The first expression formula is solved, R is obtained1, specially:
R is obtained by formula six or formula seven1:
Wherein, thrm represents that m element of the maximum in the matrix that will be obtained retains, and remaining element sets to 0;Rt+1Represent
The solution obtained using formula six (t+1) secondary iteration, RtThe solution using six last (i.e. (t) is secondary) iteration of formula is represented, when
When iterationses are N, the solution obtained using formula six is restrained, and stops iteration, the R for now obtainingN=R1, wherein R1Represent to institute
When stating the solution of the second object function, the R that first time iteration is obtained;
R1=thrm (S) formula seven;
Wherein, R1When expression is solved to second object function, the R that first time iteration is obtained.
Initial value B is given by the B in second object function0, the R in second object function is assigned to R1, and solid
B and R in fixed second object function, updates the S in second object function, obtains S1, specially:
Second object function is write the second expression formula instead, second expression formula is as shown in formula eight:
Wherein, C2=(B+ α RX)/(1+ α) is fixed value;
The second expression formula is solved, S is obtained1, specially:
S is solved by formula nine1:
RightCarry out singular value decomposition and obtain U, V:S1=VUT... formula nine;
Wherein, S1When expression is solved to the second object function, the S that first time iteration is obtained;
S in second object function is assigned to S1, the R in second object function is assigned to R1, and fixed described
S and R in second object function, updates the B in second object function, obtains B1, specially:
Second object function is write the 3rd expression formula instead, the 3rd expression formula is as shown in formula ten:
Wherein, C3=SX, is fixed value;
The 3rd expression formula is solved, B is obtained1, specially:
B is obtained by formula 111:
B=sign (C3)=sign (SX) formula 11;
Wherein, sign (*) is sign function, when independent variable * is more than 0, sign (*) function output 1, otherwise, sign (*)
Function output 0, B1When expression is solved to second object function, the B that first time iteration is obtained, independent variable are Matrix C3In
Element;
S in second object function is assigned to S1, the B in second object function is assigned to and B1, and fixed institute
The S and B in the second object function is stated, the R in second object function is updated, is obtained R2;
B in second object function is assigned to B1, the R in second object function is assigned to R2, and fixed described
B and R in second object function, updates the S in second object function, obtains S2;
S in second object function is assigned to S2, the R in second object function is assigned to R2, and fixed described
S and R in second object function, updates the B in second object function, obtains B2;
Repeat and the S in second object function is assigned to Sm-1, the B in second object function is assigned to and
Bm-1, and the S and B in fixation second object function, the R in second object function is updated, R is obtainedm;By described
B in two object functions is assigned to Bm-1, the R in second object function is assigned to Rm, and in fixation second object function
B and R, update the S in second object function, obtain Sm;S in second object function is assigned to Sm, by described
R in two object functions is assigned to Rm, and the S and R in fixation second object function, update in second object function
B, obtains BmOperation, until complete M iteration, the B that the M time iteration is obtainedMThe as described each institute for meeting first object function
State the corresponding first binary feature vector of pixel;Wherein, m represents the m time iteration, and described m, M are positive integer.
The present invention also provides a kind of face identification device based on sparse projection binary-coding, including:
Computing unit, the computing unit are used for corresponding first picture of each pixel for obtaining each training sample in training set
Plain difference value vector;Wherein, the training sample is facial image, and the training set includes several different facial images;
The computing unit is additionally operable to be obtained according to first pixel value difference vector and meets based on sparse projection matrix
The corresponding first binary feature vector of each described pixel of first object function;
Cluster cell, the cluster cell are clustered to all of first binary feature vector, are obtained multiple poly-
Class center-word;
Primary vector acquiring unit, the primary vector acquiring unit are used for corresponding for each training sample first
Binary feature vector the first matrix of composition, is linearly rebuild to first matrix using the word, is obtained first linear
Reconstructed results, obtain primary vector according to the described first linear reconstructed results;Wherein, each training sample corresponding one first to
Amount;
Secondary vector acquiring unit, the secondary vector acquiring unit are used for each pixel of the image for obtaining face to be detected
The corresponding second binary feature vector of point, and by each second binary feature vector, second matrix of composition, using the word
Second matrix is linearly rebuild, is obtained the second linear reconstructed results, obtained according to the described second linear reconstructed results
Secondary vector;
Recognition unit, the recognition unit are used for according to the primary vector and the secondary vector, obtain recognition of face
As a result.
The present invention provides a kind of face identification method and device based on sparse projection binary-coding.The base that the present invention is provided
In the face identification method of sparse projection binary-coding, each pixel including obtaining each training sample in training set corresponding the
One pixel value difference vector;Wherein, training sample is facial image, and training set includes several different facial images;According to
One pixel value difference vector obtains corresponding first two-value of each pixel of the satisfaction based on the first object function of sparse projection matrix
Characteristic vector;All of first binary feature vector is clustered, multiple cluster centre-words are obtained;By each training sample
Corresponding the first binary feature vector the first matrix of composition, is linearly rebuild to the first matrix using word, is obtained first
Linear reconstructed results, obtain primary vector according to the first linear reconstructed results;Wherein, each training sample corresponding one first to
Amount;Obtain each pixel corresponding second binary feature vector of the image of face to be detected, and by each second binary feature to
Amount the second matrix of composition, is linearly rebuild to the second matrix using word, is obtained the second linear reconstructed results, according to the second line
Property reconstructed results obtain secondary vector, according to primary vector and secondary vector, obtain face recognition result.The present embodiment based on
The face identification method of sparse projection binary-coding is using being entered to the pixel value difference vector of real value representation using sparse projection matrix
Row coding obtains binary feature vector, and carries out cluster to binary feature vector and obtain word, is linearly rebuild per width using word
The matrix of the corresponding binary feature vector composition of facial image, recognition speed are fast, solve the problems, such as training sample overfitting,
And good to the adaptability of data, improve accuracy and the rapidity of the identification face of face identification method.
Description of the drawings
In order to be illustrated more clearly that the embodiment of the present invention or technical scheme of the prior art, below will be to embodiment or existing
Accompanying drawing to be used needed for having technology description is briefly described, it should be apparent that, drawings in the following description are these
Some bright embodiments, for those of ordinary skill in the art, without having to pay creative labor, can be with
Other accompanying drawings are obtained according to these accompanying drawings.
The flow chart of the face identification method embodiment one based on sparse projection binary-coding that Fig. 1 is provided for the present invention;
Fig. 2 is that the first pixel value difference vector of the present invention obtains schematic diagram;
The flow chart of the face identification method embodiment two based on sparse projection binary-coding that Fig. 3 is provided for the present invention;
The structural representation of the face identification device based on sparse projection binary-coding that Fig. 4 is provided for the present invention.
Specific embodiment
Purpose, technical scheme and advantage for making the embodiment of the present invention is clearer, below in conjunction with the embodiment of the present invention
In accompanying drawing, to the embodiment of the present invention in technical scheme be clearly and completely described, it is clear that described embodiment is
The a part of embodiment of the present invention, rather than whole embodiments.Embodiment in based on the present invention, those of ordinary skill in the art
The every other embodiment obtained under the premise of creative work is not made, belongs to the scope of protection of the invention.
The flow chart of the face identification method embodiment one based on sparse projection binary-coding that Fig. 1 is provided for the present invention,
Fig. 2 is that the first pixel value difference vector of the present invention obtains schematic diagram.
As shown in Fig. 1~2, the method for the present embodiment can include:
The corresponding first pixel value difference vector of each pixel of each training sample in S101, acquisition training set;Wherein, train
Sample is facial image, and training set includes several different facial images;
S102, each picture for meeting the first object function based on sparse projection matrix according to the vector acquisition of the first pixel value difference
The corresponding first binary feature vector of vegetarian refreshments;
S103, all of first binary feature vector is clustered, obtain multiple cluster centre-words;
S104, by corresponding for each training sample the first binary feature vector the first matrix of composition, using word to first
Matrix is linearly rebuild, and obtains the first linear reconstructed results, obtains primary vector according to the first linear reconstructed results;Wherein,
The corresponding primary vector of each training sample;
S105, the corresponding second binary feature vector of each pixel of the image for obtaining face to be detected, and by each second
Binary feature vector the second matrix of composition, is linearly rebuild to the second matrix using word, is obtained the second linear reconstructed results,
Secondary vector is obtained according to the second linear reconstructed results;
S106, according to primary vector and secondary vector, obtain face recognition result.
For step S101, obtain substantial amounts of different facial images in advance and constitute training set, each face in training set
Image is referred to as training sample.
After training set is built up, the corresponding first pixel value difference vector of each pixel of each training sample in acquisition training set,
Specially:Each training sample is divided into multiple pieces;With each pixel in each block as the first central pixel point, with r it is
Radius, obtains the first neighborhood territory pixel point of the first central pixel point, and according to clockwise, by the picture of the first neighborhood territory pixel point
Plain value carries out difference operation with the pixel value of the first central pixel point, obtain that length is tieed up for (2 × r+1) × (2 × r+1) -1 the
One pixel value difference vector;Preferably, r takes 3.In order to clearly explain clockwise, illustrated using r=2 in the present embodiment suitable
Conterclockwise implication, as shown in Fig. 2 the path with arrow dotted line in Fig. 2 is clockwise, each square represents one
Pixel, A represent the first central pixel point, along the arrow dotted line direction successively by the pixel value and first of the first neighborhood territory pixel point
The pixel value of central pixel point carries out difference operation, the first pixel value difference vector of 24 dimension of composition respectively.Wherein, the first neighborhood picture
The number of vegetarian refreshments is also (2 × r+1) × (2 × r+1) -1.
In the step, it will be appreciated by persons skilled in the art that after the value of r is selected, as training sample is divided into
Multiple pieces, have in each block many edge pixel points the first neighborhood territory pixel point number deficiency (2 × r+1) × (2 ×
R+1) -1, in order that arrive the first pixel value difference vector dimension identical, by each block the first neighborhood territory pixel point
The edge pixel point of number deficiency (2 × r+1) × (2 × r+1) -1 is given up.
For step S102, after the first pixel value difference vector is obtained, obtained according to the first pixel value difference vector and meet base
In the corresponding first binary feature vector of each pixel of the first object function of sparse projection matrix.In the present embodiment,
One object function is as shown in formula one:
Wherein, R be sparse projection matrix, | R |0≤ m represents that the number of nonzero element in sparse projection matrix R is less than or equal to
M, B is that the sparse degree of the first binary feature vector, parameter m and sparse projection matrix R has incidence relation, and m is positive integer,
X is the first pixel value difference vector.
Wherein, the value rule of m is, the 4%~6% of the number of all elements in matrix.
The first pixel value difference vector is encoded using sparse projection matrix, the first binary feature after the quantization for obtaining
Dimension of the dimension of vector far above the first pixel value difference vector, therefore, it is possible to minimize the first pixel value difference vector sum first
Quantization error between binary feature vector, can lift the ability of the first binary feature vector description facial image.And, adopt
The first pixel value difference vector is encoded with sparse projection matrix, so as to get the first binary feature vector have sparse (i.e.
Most elements are property 0), can pass through the degree of rarefication for adjusting projection matrix, by face identification method calculating process
Need the order of magnitude of modulation parameter to be adjusted to the level suitable with training data complexity, solve the data mistake in training sample
The problem of degree fitting.
For step S103, after the corresponding first binary feature vector of each pixel for obtaining training set, to training set
The corresponding first binary feature vector of all pixels point clustered, obtain multiple cluster centre-words;Can in the present embodiment
With using K-Means clustering methods or SGONG clustering methods to corresponding first binary feature of all pixels point of training set
Vector is learnt, and is obtained multiple cluster centre-words, but is not limited to above two clustering method.
For step S104, cluster in the corresponding first binary feature vector of all pixels point to training set, obtain
To after multiple words, primary vector is obtained then, specifically, each training sample has multiple pixels, each pixel pair
The first binary feature vector is answered, each first binary feature vector of a training sample is constituted the first square in order
Battle array, then, is linearly rebuild to the first matrix using word, is obtained the first linear reconstructed results, according to the first linear reconstruction
As a result primary vector is obtained;The corresponding primary vector of each training sample.
By the first binary feature vector the first matrix of composition corresponding for each training sample, the first matrix is entered using word
Line is rebuild, compared with directly being represented with more preferable data adaptability using the first original matrix.
For step S105, after the corresponding primary vector of each training sample in training set is obtained, training process is completed,
Start face recognition process, the image corresponding second that face to be detected is obtained using above-mentioned acquisition primary vector identical method
Vector, specially:Obtain each pixel corresponding second binary feature vector of the image of face to be detected, and by each 2nd 2
Value tag vector the second matrix of composition, is linearly rebuild to the second matrix using word, is obtained the second linear reconstructed results, root
Secondary vector is obtained according to the second linear reconstructed results.
In this step, the method and the method phase for obtaining the first binary feature vector of the second binary feature vector are obtained
With.
For step S106, after primary vector and secondary vector is obtained, just can according to primary vector and secondary vector,
Obtain face recognition result, the method that such as can adopt grader:By primary vector and secondary vector input grader so that point
Class device calculates the Euclidean distance between each primary vector and secondary vector, obtain secondary vector that grader returns with each first to
The Euclidean distance of amount, by the face figure corresponding to training sample corresponding with the primary vector that the Euclidean distance of secondary vector is most short
As being identified as the image of face to be detected.
The face identification method based on sparse projection binary-coding of the present embodiment, including obtaining each training sample in training set
The corresponding first pixel value difference vector of each pixel originally;Wherein, training sample is facial image, and training set includes several not
Same facial image;The each picture for meeting the first object function based on sparse projection matrix is obtained according to the first pixel value difference vector
The corresponding first binary feature vector of vegetarian refreshments;All of first binary feature vector is clustered, obtain multiple cluster centres-
Word;By the first binary feature vector the first matrix of composition corresponding for each training sample, the first matrix is carried out using word
Linear reconstruction, obtains the first linear reconstructed results, obtains primary vector according to the first linear reconstructed results;Wherein, each training
The corresponding primary vector of sample;Each pixel corresponding second binary feature vector of the image of face to be detected is obtained, and
By each second binary feature vector the second matrix of composition, the second matrix is linearly rebuild using word, obtained second linear
Reconstructed results, obtain secondary vector according to the second linear reconstructed results, according to primary vector and secondary vector, obtain recognition of face
As a result.The face identification method based on sparse projection binary-coding of the present embodiment is using sparse projection matrix to real-valued table
The pixel value difference vector for showing carries out coding and obtains binary feature vector, and carries out cluster to binary feature vector and obtain word, adopts
The matrix that the corresponding binary feature vector composition of every width facial image is linearly rebuild with word, recognition speed are fast, solve training
The problem of sample overfitting, and good to the adaptability of data, improve the identification face of face identification method accuracy and
Rapidity.
Below the technical scheme of embodiment of the method shown in Fig. 1 is described in detail.
First to obtaining satisfaction based on each of the first object function of sparse projection matrix according to the first pixel value difference vector
The method of the corresponding first binary feature vector of pixel is illustrated.
Fig. 3 is the stream of the face identification method embodiment two based on sparse projection binary-coding provided in an embodiment of the present invention
Cheng Tu, as shown in figure 3, the method for the present embodiment includes:
S301, the agent matrix S of the sparse projection matrix R being introduced in first object function, by first object functional transformation
For the second object function;
S302, solves to the second object function, obtains the first binary feature vector.
For step S301, the agent matrix S of R is introduced, the size of s-matrix is identical with matrix R, by first object functionThe second object function is transformed to, the second object function is as shown in formula four:
Wherein, variable α represents the penalty factor of agent matrix S, in order to equation of equilibrium four inWithThis two variable, | S |0The number of non-zero element in agent matrix S is represented,Square is acted on behalf of in expression
Quantization error between battle array S and binary-coding B,Represent the error between agent matrix S and R;Square is wherein acted on behalf of
Battle array S is also sparse projection matrix.
For step S302, the second object function is solved, the method for obtaining the first binary feature vector, specifically
For:
Initial value S is given at random by the S and B in the second object function0And B0, and the S that fixes in the second object function and B,
The R in the second object function is updated, R is obtained1, specially:
Second object function is write the first expression formula instead, the first expression formula is as shown in formula five:
Wherein, C1=SX, is fixed value;
The first expression formula is solved, R is obtained1, specially:
R is obtained by formula six or formula seven1:
Wherein, thrm represents that m element of the maximum in the matrix that will be obtained retains, and remaining element sets to 0;Rt+1Represent
The solution obtained using formula six (t+1) secondary iteration, RtThe solution using six last (i.e. (t) is secondary) iteration of formula is represented, when
When iterationses are N, the solution obtained using formula six is restrained, and stops iteration, the R for now obtainingN=R1, wherein R1Represent to the
When two object functions are solved, the R that first time iteration is obtained;
R1=thrm (S) formula seven;
Wherein, R1When expression is solved to the second object function, the R that first time iteration is obtained.
Initial value B is given by the B in the second object function0, the R in the second object function is assigned to R1, and fix the second mesh
B and R in scalar functions, updates the S in the second object function, obtains S1, specially:
Second object function is write the second expression formula instead, the second expression formula is as shown in formula eight:
Wherein, C2=(B+ α RX)/(1+ α) is fixed value;
The second expression formula is solved, S is obtained1, specially:
S is solved by formula nine1:
RightCarry out singular value decomposition and obtain U, V:S1=VUT.... formula nine;
Wherein, S1When expression is solved to the second object function, the S that first time iteration is obtained;
S in second object function is assigned to S1, the R in the second object function is assigned to R1, and fix the second object function
In S and R, update the second object function in B, obtain B1, specially:
Second object function is write the 3rd expression formula instead, the 3rd expression formula is as shown in formula ten:
Wherein, C3=SX, is fixed value;
The 3rd expression formula is solved, B is obtained1, specially:
B is obtained by formula 111:
B=sign (C3)=sign (SX) formula 11;
Wherein, sign (*) is sign function, when independent variable * is more than 0, sign (*) function output 1, otherwise, sign (*)
Function output 0, B1When expression is solved to the second object function, the B that first time iteration is obtained;Independent variable refers to Matrix C3In unit
Element.
S in second object function is assigned to S1, the B in the second object function is assigned to and B1, and fix the second target letter
S and B in number, updates the R in the second object function, obtains R2;
B in second object function is assigned to B1, the R in the second object function is assigned to R2, and fix the second object function
In B and R, update the second object function in S, obtain S2;
S in second object function is assigned to S2, the R in the second object function is assigned to R2, and fix the second object function
In S and R, update the second object function in B, obtain B2;
Repeat and the S in the second object function is assigned to Sm-1, the B in the second object function is assigned to and Bm-1, and solid
S and B in fixed second object function, updates the R in the second object function, obtains Rm;B in second object function is assigned to
Bm-1, the R in the second object function is assigned to Rm, and the B that fixes in the second object function and R, update in the second object function
S, obtains Sm;S in second object function is assigned to Sm, the R in the second object function is assigned to Rm, and fix the second target letter
S and R in number, updates the B in the second object function, obtains BmOperation, until complete M iteration, the M time iteration is obtained
BMAs meet each pixel corresponding first binary feature vector of first object function;Wherein, m represents the m time iteration, m,
M is positive integer.
Wherein, the value of M can determine according to actual needs, be not construed as limiting in the present embodiment.
Below the derivation of the formula six in a upper embodiment, formula 11 is introduced.
First the derivation of formula six is illustrated;
For the solution procedure of expression formula one is relatively difficult, the 3rd object function, the 3rd object function such as formula ten is introduced
Shown in two:
Wherein,
According to the property of F- norms, expression formulaIt is that perseverance is more than 0,
Therefore, the numerical value of the 3rd object function will not be less than the numerical value of expression formula one, and therefore the optimal solution of formula 12 still meets table
Reach formula one;
Using the 3rd object function of solution of iteration, that is, fix S and update R, the 3rd object function is transformed to expression formula four,
Expression four is as shown in formula 13:
Wherein,Const represents the constant for being independent of R;
The vector form of expression formula four is shown below:
Wherein, the solution of the formula can be expressed asThe solution of its matrix form can be generalized for formula ten
Four:
Solution R is iterated using formula 14, iterative formula is formula six.
Secondly, the derivation of formula 11 is illustrated.
The deployable form for formula one of expression formula two:As matrix B is orthogonal
, thereforeFor a constant, therefore, above formula is then further simplified as:
Wherein, n represents the number of training sample, and c represents the dimension of binary feature vector, in order to maximize above formula, needs
As (C3)ijDuring more than or equal to 0, B is allowedij=1, as (C3)ijDuring less than 0, B is allowedij=1, then it is converted into asking to formula 11
Solution.
Secondly, to above-described embodiment in step " by the first binary feature vector composition corresponding for each training sample the
One matrix, is linearly rebuild to the first matrix using word, obtains the first linear reconstructed results, according to the first linear reconstruction knot
Fruit obtains primary vector " illustrate.
By the first binary feature vector the first matrix of composition corresponding for each training sample, the first matrix is entered using word
Line is rebuild, and obtains the first linear reconstructed results, obtains primary vector according to the first linear reconstructed results, including:
The first matrix is linearly rebuild using word, obtained the first linear reconstructed results, the first linear reconstructed results
Expression formula as shown in formula two:
Bi=ai1S1+ai2S2+...+aikSkFormula two;
Wherein, BiFor the first matrix of each first binary feature vector composition of i-th training sample, SkSingle for k-th
Word, aikIt is the power of k-th word when corresponding described first matrix of i-th training sample is linearly rebuild using word
Weight;
Corresponding for each word in the expression formula of the first linear reconstructed results weight is constituted primary vector (ai1,
ai2... ..., aik).
Then, to above-described embodiment in step " obtain each pixel the corresponding 2nd 2 of the image of face to be detected
Value tag vector, and by each second binary feature vector the second matrix of composition, the second matrix is linearly rebuild using word,
The second linear reconstructed results are obtained, and secondary vector are obtained according to the second linear reconstructed results " illustrate.
Each pixel corresponding second binary feature vector of the image of face to be detected is obtained, and will each second two-value spy
Vector the second matrix of composition is levied, the second matrix is linearly rebuild using word, is obtained the second linear reconstructed results, according to the
Bilinear reconstructed results obtain secondary vector, including:
Each pixel corresponding second pixel value difference vector of the image of face to be detected is obtained, and according to the second pixel difference
Value vector obtains each pixel corresponding second binary feature vector of the image of the face to be detected for meeting first object function;
Each second binary feature vector is constituted second matrix, and second matrix is adopted the word
Linearly rebuild, obtained the described second linear reconstruction structure, expression formula such as three institute of formula of the second linear reconstructed results
Show:
BTest=b1S1+b2S2+...+bkSkFormula three;
Wherein, BTestFor the second matrix of each second binary feature vector composition, SkFor k-th word, bkFor described
The weight of k-th word when second matrix is linearly rebuild using the word;
By corresponding for each described word in the expression formula of the described second linear reconstructed results weight composition described second to
Amount (b1, b2... ..., bk).
Wherein, the method for obtaining the second pixel value difference vector is identical with the method for obtaining the first pixel value difference, specially:
It it is multiple pieces by the image division of face to be detected;
With each pixel in each block as the second central pixel point, with r as radius, the second central pixel point is obtained
Second neighborhood territory pixel point, and according to clockwise, by the pixel value of the second neighborhood territory pixel point and the picture of the second central pixel point
Plain value carries out difference operation, obtains second pixel value difference vector of the length for (2 × r+1) × (2 × r+1) -1 dimensions;
The method for obtaining the second binary feature vector is identical with the method for obtaining the first binary feature vector, no longer goes to live in the household of one's in-laws on getting married herein
State.
The structural representation of the face identification device based on sparse projection binary-coding that Fig. 4 is provided for the present invention, such as Fig. 4
Shown, the device of the present embodiment can include:Computing unit 401, cluster cell 402, primary vector acquiring unit 403, second
Vectorial acquiring unit 404, recognition unit 405.
Computing unit 401, corresponding first pixel value difference of each pixel for obtaining each training sample in training set to
Amount;Wherein, training sample is facial image, and training set includes several different facial images;
Computing unit 401 is additionally operable to obtain the first mesh met based on sparse projection matrix according to the first pixel value difference vector
The corresponding first binary feature vector of each pixel of scalar functions;
Cluster cell 402, for clustering to all of first binary feature vector, obtains multiple cluster centres-mono-
Word;
Primary vector acquiring unit 403, for constituting first by the first binary feature vector corresponding for each training sample
Matrix, is linearly rebuild to the first matrix using word, obtains the first linear reconstructed results, according to the first linear reconstructed results
Obtain primary vector;Wherein, the corresponding primary vector of each training sample;
Secondary vector acquiring unit 404, corresponding second two-value of each pixel for obtaining the image of face to be detected
Characteristic vector, and by each second binary feature vector the second matrix of composition, the second matrix is linearly rebuild using word, obtain
To the second linear reconstructed results, secondary vector is obtained according to the second linear reconstructed results;
Recognition unit 405, for according to primary vector and secondary vector, obtaining face recognition result.
The device of the present embodiment, can be used for the technical scheme for executing embodiment of the method shown in Fig. 1, and which realizes principle and skill
Art effect is similar to, and here is omitted.
One of ordinary skill in the art will appreciate that:Realize that all or part of step of above-mentioned each method embodiment can be led to
Cross the related hardware of programmed instruction to complete.Aforesaid program can be stored in a computer read/write memory medium.The journey
Sequence upon execution, executes the step of including above-mentioned each method embodiment;And aforesaid storage medium includes:ROM, RAM, magnetic disc or
Person's CD etc. is various can be with the medium of store program codes.
Finally it should be noted that:Various embodiments above only in order to technical scheme to be described, rather than a limitation;To the greatest extent
Pipe has been described in detail to the present invention with reference to foregoing embodiments, it will be understood by those within the art that:Its according to
So the technical scheme described in foregoing embodiments can be modified, or which part or all technical characteristic are entered
Row equivalent;And these modifications or replacement, do not make the essence of appropriate technical solution depart from various embodiments of the present invention technology
The scope of scheme.
Claims (9)
1. a kind of face identification method based on sparse projection binary-coding, it is characterised in that include:
The corresponding first pixel value difference vector of each pixel of each training sample in acquisition training set;Wherein, the training sample
For facial image, the training set includes several different facial images;
The each described picture for meeting the first object function based on sparse projection matrix is obtained according to first pixel value difference vector
The corresponding first binary feature vector of vegetarian refreshments;
The all of first binary feature vector is clustered, multiple cluster centre-words are obtained;
By corresponding for each training sample the first binary feature vector the first matrix of composition, using the word to described the
One matrix is linearly rebuild, and obtains the first linear reconstructed results, obtains primary vector according to the described first linear reconstructed results;
Wherein, the corresponding primary vector of each training sample;
Each pixel corresponding second binary feature vector of the image of face to be detected is obtained, and will each second two-value spy
Vector the second matrix of composition is levied, second matrix is linearly rebuild using the word, obtain the second linear reconstruction knot
Really, secondary vector is obtained according to the described second linear reconstructed results;
According to the primary vector and the secondary vector, face recognition result is obtained.
2. method according to claim 1, it is characterised in that the first object function is as shown in formula one:
Wherein, R be sparse projection matrix, | R |0≤ m represents the number of nonzero element in sparse projection matrix R less than or equal to m, B
For the first binary feature vector, the sparse degree of parameter m and the sparse projection matrix R has an incidence relation, and m is positive integer, X
For first pixel value difference vector.
3. method according to claim 2, it is characterised in that described by the first corresponding for each training sample two-value
Characteristic vector constitutes the first matrix, first matrix is linearly rebuild using the word, obtains the first linear reconstruction
As a result, primary vector is obtained according to the described first linear reconstructed results;Wherein, the corresponding primary vector of each training sample,
Including:
First matrix is linearly rebuild using the word, obtained the first linear reconstructed results, described first is linear
The expression formula of reconstructed results is as shown in formula two:
Bi=ai1S1+ai2S2+...+aikSkFormula two;
Wherein, BiFor the first matrix of each first binary feature vector composition of i-th training sample, SkFor k-th word, aik
It is the power of k-th word when corresponding described first matrix of i-th training sample is linearly rebuild using the word
Weight;
Corresponding for each described word in the expression formula of the described first linear reconstructed results weight is constituted primary vector (ai1,
ai2... ..., aik).
4. method according to claim 3, it is characterised in that each pixel pair of the image of acquisition face to be detected
The the second binary feature vector that answers, and by each second binary feature vector, second matrix of composition, using the word to institute
State the second matrix linearly to be rebuild, obtain the second linear reconstructed results, second is obtained according to the described second linear reconstructed results
Vector, including:
Each pixel corresponding second pixel value difference vector of the image of face to be detected is obtained, and according to second pixel difference
Value vector obtains corresponding second two-value of each pixel of the image of the face to be detected for meeting the first object function
Characteristic vector;
Each second binary feature vector is constituted second matrix, and second matrix is carried out using the word
Linear reconstruction, obtains the described second linear reconstruction structure, and the expression formula of the second linear reconstructed results is as shown in formula three:
BTest=b1S1+b2S2+...+bkSkFormula three;
Wherein, BTestFor the second matrix of each second binary feature vector composition, SkFor k-th word, bkFor described second
The weight of k-th word when matrix is linearly rebuild using the word;
Corresponding for each described word in the expression formula of the described second linear reconstructed results weight is constituted the secondary vector (b1,
b2... ..., bk).
5. method according to claim 1, it is characterised in that described according to the primary vector and the secondary vector,
Face recognition result is obtained, including:
By the primary vector and secondary vector input grader;
Obtain the secondary vector of the grader return and the Euclidean distance of each primary vector;
Facial image corresponding to training sample corresponding with the primary vector that the Euclidean distance of secondary vector is most short is identified as
The image of face to be detected.
6. method according to claim 4, it is characterised in that each pixel for obtaining each training sample in training set is corresponding
The first pixel value difference vector, including:
Each training sample is divided into multiple pieces;
With each pixel in each block as the first central pixel point, with r as radius, first central pixel point is obtained
First neighborhood territory pixel point, and according to clockwise, by imago in the pixel value and described first of the first neighborhood territory pixel point
The pixel value of vegetarian refreshments carries out difference operation, obtains first pixel value difference vector of the length for (2 × r+1) × (2 × r+1) -1 dimensions;
Each pixel corresponding second pixel value difference vector of the image of face to be detected is obtained, including:
It it is multiple pieces by the image division of the face to be detected;
With each pixel in each block as the second central pixel point, with r as radius, second central pixel point is obtained
Second neighborhood territory pixel point, and according to clockwise, by imago in the pixel value and described second of the second neighborhood territory pixel point
The pixel value of vegetarian refreshments carries out difference operation, obtains second pixel value difference vector of the length for (2 × r+1) × (2 × r+1) -1 dimensions;
Wherein, r is positive integer.
7. method according to claim 2, it is characterised in that described acquisition according to first pixel value difference vector meets
The corresponding first binary feature vector of each described pixel based on the first object function of sparse projection matrix, including:
The agent matrix S of R is introduced, by the first object functionIt is transformed to second
Object function, second object function is as shown in formula four:
Wherein, variable α represents the penalty factor of agent matrix S, in order to equation of equilibrium four inWith
This two variable, S0Represent the number of non-zero element in agent matrix S, the sparse journey of parameter m and the sparse projection matrix R
Degree has incidence relation;The quantization error between agent matrix S and binary-coding B is represented,Table
Show the error between agent matrix S and R;
Second object function is solved, the first binary feature vector is obtained.
8. method according to claim 7, it is characterised in that described second object function is solved, obtain
The first binary feature vector, including:
Initial value S is given at random by the S and B in second object function0And B0, and the S in fixation second object function
And B, the R in second object function is updated, R is obtained1, specially:
Second object function is write the first expression formula instead, first expression formula is as shown in formula five:
Wherein, C1=SX, is fixed value;
The first expression formula is solved, R is obtained1, specially:
R is obtained by formula six or formula seven1:
Wherein, thrm represents that m element of the maximum in the matrix that will be obtained retains, and remaining element sets to 0;Rt+1Represent and utilize
The solution that formula six (t+1) secondary iteration is obtained, RtThe solution using six last (i.e. (t) is secondary) iteration of formula is represented, works as iteration
When number of times is N, the solution obtained using formula six is restrained, and stops iteration, the R for now obtainingN=R1, wherein R1Represent to described the
When two object functions are solved, the R that first time iteration is obtained;
R1=thrm (S) formula seven;
Wherein, R1When expression is solved to second object function, the R that first time iteration is obtained.
Initial value B is given by the B in second object function0, the R in second object function is assigned to R1, and fixed institute
The B and R in the second object function is stated, the S in second object function is updated, is obtained S1, specially:
Second object function is write the second expression formula instead, second expression formula is as shown in formula eight:
Wherein, C2=(B+ α RX)/(1+ α) is fixed value;
The second expression formula is solved, S is obtained1, specially:
S is solved by formula nine1:
RightCarry out singular value decomposition and obtain U, V:
Wherein, S1When expression is solved to the second object function, the S that first time iteration is obtained;
S in second object function is assigned to S1, the R in second object function is assigned to R1, and fixed described second
S and R in object function, updates the B in second object function, obtains B1, specially:
Second object function is write the 3rd expression formula instead, the 3rd expression formula is as shown in formula ten:
Wherein, C3=SX, is fixed value;
The 3rd expression formula is solved, B is obtained1, specially:
B is obtained by formula 111:
B=sign (C3)=sign (SX) formula 11;
Wherein, sign (*) is sign function, when independent variable * is more than 0, sign (*) function output 1, otherwise, sign (*) function
Output 0, B1When expression is solved to second object function, the B that first time iteration is obtained, independent variable are Matrix C3In element;
S in second object function is assigned to S1, the B in second object function is assigned to and B1, and fixation described the
S and B in two object functions, updates the R in second object function, obtains R2;
B in second object function is assigned to B1, the R in second object function is assigned to R2, and fixed described second
B and R in object function, updates the S in second object function, obtains S2;
S in second object function is assigned to S2, the R in second object function is assigned to R2, and fixed described second
S and R in object function, updates the B in second object function, obtains B2;
Repeat and the S in second object function is assigned to Sm-1, the B in second object function is assigned to and Bm-1,
And the S and B in fixation second object function, the R in second object function is updated, R is obtainedm;By second mesh
B in scalar functions is assigned to Bm-1, the R in second object function is assigned to Rm, and the B in fixation second object function
And R, the S in second object function is updated, S is obtainedm;S in second object function is assigned to Sm, by described second
R in object function is assigned to Rm, and the S and R in fixation second object function, the B in second object function is updated,
Obtain BmOperation, until complete M iteration, the B that the M time iteration is obtainedMAs described meet each described of first object function
The corresponding first binary feature vector of pixel;Wherein, m represents the m time iteration, and described m, M are positive integer.
9. a kind of face identification device based on sparse projection binary-coding, it is characterised in that include:
Computing unit, the computing unit are used for corresponding first pixel difference of each pixel for obtaining each training sample in training set
Value vector;Wherein, the training sample is facial image, and the training set includes several different facial images;
The computing unit is additionally operable to obtain first met based on sparse projection matrix according to first pixel value difference vector
The corresponding first binary feature vector of each described pixel of object function;
Cluster cell, the cluster cell are clustered to all of first binary feature vector, are obtained in multiple clusters
The heart-word;
Primary vector acquiring unit, the primary vector acquiring unit are used for the first corresponding for each training sample two-value
Characteristic vector constitutes the first matrix, first matrix is linearly rebuild using the word, obtains the first linear reconstruction
As a result, primary vector is obtained according to the described first linear reconstructed results;Wherein, the corresponding primary vector of each training sample;
Secondary vector acquiring unit, the secondary vector acquiring unit are used for each pixel pair of the image for obtaining face to be detected
The the second binary feature vector that answers, and by each second binary feature vector, second matrix of composition, using the word to institute
State the second matrix linearly to be rebuild, obtain the second linear reconstructed results, second is obtained according to the described second linear reconstructed results
Vector;
Recognition unit, the recognition unit are used for according to the primary vector and the secondary vector, obtain face recognition result.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610917123.0A CN106503648A (en) | 2016-10-20 | 2016-10-20 | Face identification method and device based on sparse projection binary-coding |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610917123.0A CN106503648A (en) | 2016-10-20 | 2016-10-20 | Face identification method and device based on sparse projection binary-coding |
Publications (1)
Publication Number | Publication Date |
---|---|
CN106503648A true CN106503648A (en) | 2017-03-15 |
Family
ID=58318135
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610917123.0A Pending CN106503648A (en) | 2016-10-20 | 2016-10-20 | Face identification method and device based on sparse projection binary-coding |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106503648A (en) |
Cited By (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107273842A (en) * | 2017-06-09 | 2017-10-20 | 北京工业大学 | Selective ensemble face identification method based on CSJOGA algorithms |
CN110874385A (en) * | 2018-08-10 | 2020-03-10 | 阿里巴巴集团控股有限公司 | Data processing method, device and system |
CN113436061A (en) * | 2021-07-01 | 2021-09-24 | 中科人工智能创新技术研究院(青岛)有限公司 | Face image reconstruction method and system |
CN113536974A (en) * | 2021-06-28 | 2021-10-22 | 杭州电子科技大学 | Face binary feature extraction method based on sparse constraint |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102782708A (en) * | 2009-12-02 | 2012-11-14 | 高通股份有限公司 | Fast subspace projection of descriptor patches for image recognition |
CN104978549A (en) * | 2014-04-03 | 2015-10-14 | 北京邮电大学 | Three-dimensional face image feature extraction method and system |
CN105930834A (en) * | 2016-07-01 | 2016-09-07 | 北京邮电大学 | Face identification method and apparatus based on spherical hashing binary coding |
-
2016
- 2016-10-20 CN CN201610917123.0A patent/CN106503648A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102782708A (en) * | 2009-12-02 | 2012-11-14 | 高通股份有限公司 | Fast subspace projection of descriptor patches for image recognition |
CN104978549A (en) * | 2014-04-03 | 2015-10-14 | 北京邮电大学 | Three-dimensional face image feature extraction method and system |
CN105930834A (en) * | 2016-07-01 | 2016-09-07 | 北京邮电大学 | Face identification method and apparatus based on spherical hashing binary coding |
Non-Patent Citations (3)
Title |
---|
LEI TIAN 等: "Learning iterative quantization binary codes for face recognition", 《NEUROCOMPUTING》 * |
夏炎: "大规模图像数据中相似图像的快速搜索", 《中国博士学位论文全文数据库(电子期刊)》 * |
范引娣: "基于高阶结构约束的稀疏人脸识别算法", 《计算机与现代化》 * |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107273842A (en) * | 2017-06-09 | 2017-10-20 | 北京工业大学 | Selective ensemble face identification method based on CSJOGA algorithms |
CN110874385A (en) * | 2018-08-10 | 2020-03-10 | 阿里巴巴集团控股有限公司 | Data processing method, device and system |
CN110874385B (en) * | 2018-08-10 | 2023-11-14 | 阿里巴巴集团控股有限公司 | Data processing method, device and system |
CN113536974A (en) * | 2021-06-28 | 2021-10-22 | 杭州电子科技大学 | Face binary feature extraction method based on sparse constraint |
CN113436061A (en) * | 2021-07-01 | 2021-09-24 | 中科人工智能创新技术研究院(青岛)有限公司 | Face image reconstruction method and system |
CN113436061B (en) * | 2021-07-01 | 2022-08-09 | 中科人工智能创新技术研究院(青岛)有限公司 | Face image reconstruction method and system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Han et al. | Unsupervised generative modeling using matrix product states | |
CN108596248B (en) | Remote sensing image classification method based on improved deep convolutional neural network | |
Zhang et al. | A lightweight and discriminative model for remote sensing scene classification with multidilation pooling module | |
CN104166859B (en) | Based on SSAE and FSALS SVM Classification of Polarimetric SAR Image | |
CN108416755A (en) | A kind of image de-noising method and system based on deep learning | |
CN110852227A (en) | Hyperspectral image deep learning classification method, device, equipment and storage medium | |
CN106503648A (en) | Face identification method and device based on sparse projection binary-coding | |
CN107067367A (en) | A kind of Image Super-resolution Reconstruction processing method | |
CN106845471A (en) | A kind of vision significance Forecasting Methodology based on generation confrontation network | |
CN103295198B (en) | Based on redundant dictionary and the sparse non-convex compressed sensing image reconstructing method of structure | |
CN111681178B (en) | Knowledge distillation-based image defogging method | |
CN104850837B (en) | The recognition methods of handwriting | |
CN111667445B (en) | Image compressed sensing reconstruction method based on Attention multi-feature fusion | |
CN107992891A (en) | Based on spectrum vector analysis multi-spectral remote sensing image change detecting method | |
CN110288526B (en) | Optimization method for improving imaging quality of single-pixel camera by image reconstruction algorithm based on deep learning | |
CN105608690A (en) | Graph theory and semi supervised learning combination-based image segmentation method | |
CN104318214B (en) | A kind of cross-view face identification method shifted based on structuring dictionary domain | |
CN105844635A (en) | Sparse representation depth image reconstruction algorithm based on structure dictionary | |
CN106991355A (en) | The face identification method of the analytical type dictionary learning model kept based on topology | |
Wang et al. | In-context learning unlocked for diffusion models | |
CN109389171A (en) | Medical image classification method based on more granularity convolution noise reduction autocoder technologies | |
CN114332545B (en) | Image data classification method and device based on low-bit pulse neural network | |
CN109165699A (en) | Fine granularity image classification method | |
CN113744136A (en) | Image super-resolution reconstruction method and system based on channel constraint multi-feature fusion | |
CN108171328A (en) | A kind of convolution algorithm method and the neural network processor based on this method |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170315 |
|
RJ01 | Rejection of invention patent application after publication |