CN105550649B - Extremely low resolution ratio face identification method and system based on unity couping local constraint representation - Google Patents
Extremely low resolution ratio face identification method and system based on unity couping local constraint representation Download PDFInfo
- Publication number
- CN105550649B CN105550649B CN201510906586.2A CN201510906586A CN105550649B CN 105550649 B CN105550649 B CN 105550649B CN 201510906586 A CN201510906586 A CN 201510906586A CN 105550649 B CN105550649 B CN 105550649B
- Authority
- CN
- China
- Prior art keywords
- resolution
- image
- low
- coefficient
- dictionary
- 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.)
- Active
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
Abstract
The invention discloses a kind of extremely low resolution ratio face identification method and system based on unity couping local constraint representation, this method includes the following two stage: training stage and test phase；Training stage comprising the following three steps: S1, expression dictionary is updated；S2, its best initial weights coefficient under the conditions of local restriction and linear reconstruction is calculated；S3, optimal neuron number is determined；Test phase comprising the following three steps: S4, acquisition low-resolution face image to be identified calculate its low resolution expression coefficient；S5, local restriction feature is coupled using high-resolution expression coefficient as image；S6, recognition result is obtained.The present invention improves the ability to express of extremely low image in different resolution, improves the discrimination on extremely low facial image；Recognition of face is completed finally by extreme learning machine, keeps recognition result more accurate.
Description
Technical field
The present invention relates to automatic Recognition Technology of Human Face fields, more particularly to one kind to be based on unity couping local constraint representation
Extremely low resolution ratio face identification method and system.
Background technique
In recent years, face recognition technology is succeeded in numerous applications.However conventional face's recognizer is general
It can assume that there is the facial image of input preferable resolution ratio due to the variation of illumination condition, face and to take the photograph in real life
The influence of the factors such as the noise of distance and equipment itself of shadow equipment, facial image often will appear size is small, resolution ratio is low,
The problems such as noise is big, characteristic details information is lost, these when be all that can obtain the very low image of resolution ratio.Especially monitoring
In the case of, video camera and interested object distance are far, and obtained image face area size is small, of poor quality.In addition, different
Posture, expression and illumination make low resolution recognition of face have more challenge.
The conventional method of processing low resolution can be divided into two kinds: a kind of method is to carry out high-resolution image down sampling
Match the image of low resolution；Another method is that the picture up-sampling of low resolution is matched high-resolution image.So
And down-sampling may directly reduce useful information, can especially reduce the high-frequency information for being conducive to identification；For low resolution figure
Picture, up-sampling itself not will increase new information, it is likely that bringing more noises.According to both methods it can be seen that solving
The key of extremely low resolution ratio recognition of face problem is to differentiate and rebuild low-resolution image information.
In order to solve the problems, such as extremely low resolution ratio recognition of face, many scholars propose many different solutions.It is based on
The method of structure is exactly to find between high-low resolution feature space handle extremely low resolution ratio recognition of face problem concern
Relationship promotes the direct the matching analysis between high-resolution baseline sample and low resolution test sample from the angle of classification.
Choi et al. proposes that characteristic dimension mismatches this problem earliest, and utilizes the method (Eigenspase of feature space estimation
Estimation, EE), obtain a new low resolution feature space from high-resolution training sample feature space, and
High-low resolution characteristic matching is carried out in this unified low resolution feature space.Recently, Li et al. people propose one it is more general
Model --- coupled maps (Coupled Mappings, CMs), for describing the unified relationship between high-low resolution, but it
Identification classification performance it is poor.Later, Li, Chang et al. propose the wider model of applicable situation, i.e. coupling part
Mapping (Coupled Locality Preserving Mappings, CLPMs) is kept, this method is by minimizing height point
Difference between resolution image, then by coupled maps, different resolution facial image is projected to learn coupled maps
One unified feature space completes match cognization, however, the performance of CLPMs is by the factors such as the selection of parameter and the variation of posture
It is affected.And occur these problems the most essential is the discriminative information of feature deficiency.
A kind of extremely low resolution ratio face recognition algorithms based on unity couping local constraint representation are proposed based on considerations above,
The advantages of algorithm describes a kind of new method unity couping local constraint representation (LCR), and local constraint representation is utilized is disclosed defeated
Enter test sample and express the manifold relationship between atom, avoids the solution instability problem of traditional sparse expression, utilize coupling
Local restriction expression characteristic high resolution graphics image space carry out characteristic matching, improve the expression energy of extremely low image in different resolution
Power improves the discrimination on extremely low facial image, improves recognition performance.
Summary of the invention
The technical problem to be solved in the present invention is that for sparse expression traditional in image-recognizing method in the prior art
Solve unstable defect, provide it is a kind of it is more accurate identification low-resolution face image based on unity couping local constraint representation
Extremely low resolution ratio face identification method and system.
The technical solution adopted by the present invention to solve the technical problems is:
The present invention provides a kind of extremely low resolution ratio face identification method based on unity couping local constraint representation, including following
Two stages: training stage and test phase；
Training stage comprising the following three steps:
S1, according to the high-resolution human face image in training sample and the low resolution face figure obtained to its down-sampling
Picture forms training sample matrix, is updated according to training sample matrix to expression dictionary according to local constraint representation algorithm, and
Expression dictionary is divided into high-resolution expression dictionary and low resolution expression dictionary；
S2, dictionary is expressed according to high-resolution and low-resolution facial image and high-resolution and low-resolution, calculates it in local restriction and line
Property rebuild under the conditions of best initial weights coefficient, the best initial weights system including high-resolution best initial weights coefficient and low resolution
Number；
S3, the limit is trained according to high-resolution human face image training sample and high-resolution best initial weights coefficient
It practises, the error that the number of automatic adjustment study neuron exports identification is minimum, and neuron number when by error minimum is true
It is set to optimal neuron number, the training stage completes；
Test phase comprising the following three steps:
S4, low-resolution face image to be identified is obtained, obtains its low resolution using low resolution expression dictionary
Express coefficient；
S5, according to manifold consistency it is assumed that by low resolution expression coefficient be maintained at high resolution space, utilize high-resolution
Rate expresses coefficient as image and couples local restriction feature；
S6, according in S3 determine optimal neuron number, construct corresponding limit learning model, and input low resolution
The best initial weights coefficient of facial image, obtains recognition result.
Further, the specific method of step S1 of the invention includes:
S11, to high-resolution training sample facial image, corresponding low resolution facial image is obtained by down-sampling, will
High-low resolution sample image is launched into column vector, and the coupling of corresponding high-low resolution image vector is become a column vector and is done
To express dictionary；
S12, one initial LCR coefficient of acquisition is calculated using following formula:
Wherein, τ is the regularization parameter of Equilibrium fitting error and local restriction, and M indicates training sample number, and X indicates high
The column vector that low-resolution image vector is coupled into, α come indicate all M face sample image reconstructed coefficients composition row to
Amount, " ο " indicate the inner product before two vectors,Indicate European squared-distance,It returns about variable α
Function α when obtaining minimum value value α*, as required LCR coefficient；
diIndicate reconstructed coefficients αiPenalty factor, calculation formula are as follows:
σ indicates the bandwidth of control distribution, diIndicate the part of the distance between measurement input picture and each dictionary atom
Parameter, diValue between zero and one；
S13, the formula for updating expression dictionary are as follows:
Wherein, Y*Expressing dictionary for coupling training sample includes high-resolution and low-resolution dictionary, then according to the resolution of image
Rate is divided into high-resolution to express dictionaryLow resolution expresses dictionary YL∈Rc×M。
Further, the specific method of step S2 of the invention includes:
S21, to high-resolution facial image training sample, it is handled to obtain observation sample matrix, observation chart
As matrix；
S22, calculate low-resolution image best initial weights coefficient formula are as follows:
Wherein, XLFor the low-resolution image in observed image matrix Column vector groups at matrix, YLFor low-resolution table
Up to dictionary, αLTo indicate that all M open the row vector of the reconstructed coefficients composition of low resolution face sample image, αL=(α1,
α2,…,αM),Indicate reconstructed coefficientsPenalty factor, τ is the regularization parameter of Equilibrium fitting error and local restriction,
" ο " indicates the inner product before two vectors, | | | | indicate European squared-distance；
It returns about variable αLFunction obtaining minimum value αLWhen valueAs required
The best initial weights coefficient of low-resolution image
Calculate the formula of the best initial weights coefficient of low-resolution image are as follows:
Wherein, XHFor the high-definition picture Column vector groups in observed image matrix at matrix, YHFor high-resolution expression
Dictionary, as the best initial weights coefficient of required high-definition picture,
Further, step S3 of the invention method particularly includes:
Transposition is carried out to the coefficient matrix of high-resolution human face image and obtains matrix T,Wherein tiIt is feature classification, is the best initial weights coefficient of the category, it is assumed that m=t2C, t are amplification coefficients, and c is one low point
Resolution image size, m are a high-definition picture size, activation primitive g (x) and its formula of hidden layer neuron number
Are as follows:
Wherein, wiIt is the weight in hidden layer between i neuron and the feature of input layer, biIt is in i-th hidden layer
Deviation, βiIt is the weight between i-th of neuron and output layer, ojIt is object vector corresponding to j-th of input,Indicate the interior collection of vector.
Further, step S5 of the invention method particularly includes:
With low-resolution face image test sample, the expression system of low resolution is obtained using low resolution expression dictionary
Number utilizes high-resolution according to manifold consistency it is assumed that low resolution local restriction expression coefficient is maintained at high resolution space
Rate local restriction expresses coefficient as image and couples local restriction feature, then in conjunction with the best initial weights of low-resolution face image
Coefficient obtains prediction output valve.
The present invention provides a kind of extremely low resolution ratio face identification system based on unity couping local constraint representation, including training
Unit and test cell:
Training unit specifically includes:
Dictionary updating unit is low for obtaining according to the high-resolution human face image in training sample and to its down-sampling
Resolution ratio facial image forms training sample matrix, according to local constraint representation algorithm according to training sample matrix to expression word
Allusion quotation is updated, and is divided into high-resolution expression dictionary and low resolution to express dictionary in expression dictionary；
Coefficient calculation unit calculates it for expressing dictionary according to high-resolution and low-resolution facial image and high-resolution and low-resolution
Best initial weights coefficient under the conditions of local restriction and linear reconstruction, including high-resolution best initial weights coefficient and low resolution
Best initial weights coefficient；
Neuron number computing unit, for according to high-resolution human face image training sample and high-resolution optimal power
Value coefficient is trained limit study, and the error that the number of automatic adjustment study neuron exports identification is minimum, by error
Neuron number when minimum is determined as optimal neuron number, and the training stage completes；
Test cell specifically includes:
Image acquisition unit is obtained for obtaining low-resolution face image to be identified using low resolution expression dictionary
Low resolution to it expresses coefficient；
Binding characteristic determination unit, for according to manifold consistency it is assumed that by low resolution expression coefficient be maintained at high score
Resolution space couples local restriction feature using high-resolution expression coefficient as image；
Image identification unit, for constructing corresponding limit learning model according to determining optimal neuron number, and it is defeated
The best initial weights coefficient for entering low-resolution face image, obtains recognition result.
The beneficial effect comprise that: the extremely low resolution ratio face of the invention based on unity couping local constraint representation
Recognition methods the advantages of using local constraint representation, discloses input by the way that identification process is divided into training stage and test phase
Manifold relationship between test sample and expression atom, avoids the solution instability problem of traditional sparse expression；Utilize coupling
Local restriction expression characteristic carries out characteristic matching in high resolution graphics image space, improves the expression energy of extremely low image in different resolution
Power improves the discrimination on extremely low facial image；Recognition of face is completed finally by extreme learning machine, makes recognition result more
It is accurate to add.
Detailed description of the invention
Present invention will be further explained below with reference to the attached drawings and examples, in attached drawing:
Fig. 1 is the stream of the extremely low resolution ratio face identification method based on unity couping local constraint representation of the embodiment of the present invention
Cheng Tu；
Fig. 2 is the another of the extremely low resolution ratio face identification method based on unity couping local constraint representation of the embodiment of the present invention
The flow chart of one embodiment；
Fig. 3 is the word of the extremely low resolution ratio face identification method based on unity couping local constraint representation of the embodiment of the present invention
Allusion quotation renewal process；
Fig. 4 is the knot of the extremely low resolution ratio face identification system based on unity couping local constraint representation of the embodiment of the present invention
Structure block diagram.
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 described herein, specific examples are only used to explain the present invention, not
For limiting the present invention.
As shown in Figure 1, the extremely low resolution ratio recognition of face side based on unity couping local constraint representation of the embodiment of the present invention
Method, including following two stage: training stage and test phase；
Training stage comprising the following three steps:
S1, according to the high-resolution human face image in training sample and the low resolution face figure obtained to its down-sampling
Picture forms training sample matrix, is updated according to training sample matrix to expression dictionary according to local constraint representation algorithm, and
Expression dictionary is divided into high-resolution expression dictionary and low resolution expression dictionary；
S2, dictionary is expressed according to high-resolution and low-resolution facial image and high-resolution and low-resolution, calculates it in local restriction and line
Property rebuild under the conditions of best initial weights coefficient, the best initial weights system including high-resolution best initial weights coefficient and low resolution
Number；
S3, the limit is trained according to high-resolution human face image training sample and high-resolution best initial weights coefficient
It practises, the error that the number of automatic adjustment study neuron exports identification is minimum, and neuron number when by error minimum is true
It is set to optimal neuron number, the training stage completes；
Test phase comprising the following three steps:
S4, low-resolution face image to be identified is obtained, obtains its low resolution using low resolution expression dictionary
Express coefficient；
S5, according to manifold consistency it is assumed that by low resolution expression coefficient be maintained at high resolution space, utilize high-resolution
Rate expresses coefficient as image and couples local restriction feature；
S6, according in S3 determine optimal neuron number, construct corresponding limit learning model, and input low resolution
The best initial weights coefficient of facial image, obtains recognition result.
As shown in Fig. 2, in another embodiment of the present invention, the method for the present invention includes the following steps:
Training stage:
S1, corresponding low-resolution face image is obtained to high-resolution human face image progress down-sampling, then height
Resolution ratio sample image is launched into column vector, the coupling of corresponding high-low resolution image is become a column vector, by whole
High-low resolution sample image forms training sample matrix, expresses according to local constraint representation algorithm sample training matrix update
Dictionary.After substep iteration obtains stable local expression coefficient matrix and corresponding dictionary matrix, by the training sample table of coupling
Two expression dictionaries of high-resolution and low resolution are divided into according to the resolution ratio of image up to dictionary, the process of dictionary updating is as schemed
Shown in 3；
S11, to high-resolution training sample facial image, corresponding low resolution facial image is obtained by down-sampling, will
High-low resolution sample image is launched into column vector, low resolution image Column vector groups at matrix indicated with L, L ∈ Rc×M, c is
One low-resolution image size, M indicate training sample number, high-definition picture Column vector groups at matrix indicated with H,T is amplification coefficient；The coupling of corresponding high-low resolution image vector is become into a column vector as word
Allusion quotation indicates with Y,Y=(y1,y2,…,yM), i=1,2 ..., M use yiEach of representing matrix Y
Column vector.It is identical as training sample number simultaneously using another part facial image as image pattern, using in training sample
Same sample coupling process handles each image, and the coupling of corresponding high-low resolution image vector is become a column vector and is used
X expression,X=(x1,x2,…,xM), use xiEach column vector in representing matrix X.It is indicated with α
The row vector of the reconstructed coefficients composition of all M face sample images,α=(α1,α2,…,αM),
Use αiEach column vector in representing matrix α.
S12, one initial LCR coefficient of acquisition is calculated using following formula:
Wherein, τ is the regularization parameter of Equilibrium fitting error and local restriction, and " ο " indicates the inner product before two vectors,Indicate European squared-distance,Return to the value of function α when obtaining minimum value about variable α
α*, as required LCR coefficient.
Wherein diIndicate reconstructed coefficients αiPenalty factor, can be indicated with following calculation formula,
σ indicates the bandwidth of control distribution, diIndicate the part of the distance between measurement input picture and each dictionary atom
Parameter.For robustness, diValue between zero and one.
S13, dictionary is updated using following formula,
This when obtained dictionary Y*Expressing dictionary for coupling training sample includes high-low resolution dictionary, then
It is divided into high-resolution to express dictionary according to the resolution ratio of imageLow resolution expresses dictionary YL∈Rc×M。
S2, the training stage in recognition of face, for high-resolution face observation sample, using same in step 1
Sample coupling process, obtains observation sample matrix, and observed image matrix is optimal when calculating under local restriction by linearly rebuilding
Weight coefficient obtains the local expression coefficient of high-resolution and low resolution respectively；
S21, the training stage in recognition of face, for high-resolution face observation sample, using same in step 1
Sample coupling process, obtain observation sample matrix, observed image matrix.
S22, the best initial weights coefficient for obtaining low-resolution image is calculated using following formula:
Wherein, XLFor the low-resolution image in observed image matrix Column vector groups at matrix, YLFor low-resolution table
Up to dictionary, αLTo indicate that all M open the row vector of the reconstructed coefficients composition of low resolution face sample image, αL=(α1,
α2,…,αM),Indicate reconstructed coefficientsPenalty factor, τ is the regularization parameter of Equilibrium fitting error and local restriction,
" ο " indicates the inner product before two vectors, | | | | indicate European squared-distance,It returns about variable αL's
Function is obtaining minimum value αLWhen valueThe as best initial weights coefficient of required low-resolution image,
With the best initial weights coefficient method for solving of low-resolution image it is identical, obtain the best initial weights of high-definition picture
Coefficient, as follows using formula:
XHFor the high-definition picture Column vector groups in observed image matrix at matrix, YHDictionary is expressed for high-resolution,The as best initial weights coefficient of required high-definition picture,
S3, the best initial weights coefficient matrix training extreme learning machine using high-resolution training set, learn interior joint to the limit
Number parameter optimizes setting；By the input weight of random initializtion network and the biasing of hidden member, hidden layer output is obtained
Matrix；Automatic adjustment neuron number makes the error of output minimum, determines optimal neuron number, completes extreme learning machine
Training；
Transposition is carried out to the coefficient matrix of high-resolution human face image and obtains matrix T,Wherein tiIt is feature classification, is the best initial weights coefficient of the category, it is assumed that m=t2C, t are amplification coefficients, and c is a low resolution
Rate image size, m are a high-definition picture size, and activation primitive g (x) and its mathematics of hidden layer neuron number are public
Formula expression are as follows:
Wherein wiIt is the weight in hidden layer between i neuron and the feature of input layer, biIt is in i-th of hidden layer
Deviation, βiIt is the weight between i-th of neuron and output layer, ojIt is object vector corresponding to j-th of input,Indicate the interior collection of vector.
The target of neural networks with single hidden layer study is the error minimum so that output, can be indicated are as follows:
There is βi, wiAnd bi, so that
It can be expressed as H β=T, it is output weight that wherein H, which is the output β of hidden node, and T is desired output.
In order to train neural networks with single hidden layer, it is intended that obtainSo that:
Wherein, i=1,2 ..., L, this is equivalent to minimize loss function:
Once above it is found that input weight wiB is biased with hidden layeriIt is determined at random, then the output matrix H of hidden layer is just by only
One determines.Training neural networks with single hidden layer, which can be converted into, solves a linear system H β=T.And exporting weight beta can be by
It determines
Wherein, H÷It is the Moore-Penrose generalized inverse of matrix H.And the provable solution acquiredNorm be the smallest
And it is unique.
Test phase:
S4 inputs low-resolution face image test sample, obtains low resolution using low resolution local expression dictionary
Expression coefficient；
S5, according to manifold consistency it is assumed that low resolution local restriction expression coefficient is maintained at high resolution space, benefit
It uses high-resolution local restriction to express coefficient as image and couples local restriction feature；
S6, the feature vector of input low resolution test facial image, the Optimal Parameters prediction of limit of utilization learning machine are defeated
Enter the category attribute t of facial image, completes face recognition process.
In another embodiment of the present invention, comprising the following steps:
Step 1, to high-resolution human face image, low-resolution face image is obtained by four times of result of down-sampling, it is low
It include low resolution face sample image in resolution ratio training library, high-resolution training includes high-resolution human face sample graph in library
Picture.Used in the examples is AR face database facial image, and all high-resolution image pixel sizes are 32 × 28, all low
The image pixel size of resolution ratio is 8 × 7.
In embodiment, in training sample facial image, a column vector, a picture are converted by high-definition picture
Resulting column vector is 896 × 1, converts a column vector for low-resolution image, the resulting column vector of a picture is 56
× 1, height resolution image is placed in a column vector, for high-definition picture upper, low-resolution image just obtains one under
Then a 952 × 1 matrix is normalized to obtain matrix Y, indicated with Y the row of all M facial images composition to
Amount, Y=(Y1,Y2,…,YM), i=1 ..., M, M are indicated in high-definition picture in training sample number and low-resolution image
Training sample is identical.In image pattern, matrix X, X=(an x is obtained according to the identical method of training sample1,x2,…,
xM), xiIndicate the row vector of M facial image compositions.
Find out an initial LCR coefficient:
Wherein, yiIt is a column vector of high-low resolution training set, all M face sample images is indicated with α
The row vector of reconstructed coefficients composition, α=(α1,α2,…,αM), τ is the regularization parameter of Equilibrium fitting error and local restriction,
" ο " indicates the inner product before two vectors,Indicate European squared-distance,Return to the function about variable α
The value α of α when obtaining minimum value*, as required LCR coefficient.
In the specific implementation, the solution of formula (1) can be obtained by lower formula:
α*=(G+ τ D) 1 (7)
α*By high-low resolution face sample images all in high-low resolution training set to defeated as under local restriction
Enter best initial weights coefficient when high-low resolution image block is linearly rebuild.
Wherein D is the diagonal matrix of a M × M, be may be expressed as:
Dii=di,1≤i≤M (8)
DiiFor the value arranged matrix D the i-th row i, diTo rebuild factor alphaiPenalty factor.
G is a local covariance matrix for inputting high-low resolution facial image X, be may be expressed as:
G=CCT (9)
Matrix C may be defined as:
C=(Xones (1, M)-Y) (10)
Wherein Y is by the corresponding column vector of M high-low resolution face sample images all in high-low resolution training set
Composed matrix, Y=(Y1,Y2,…,YM), ones (1, M) is 1 × M row vector that element is all 1, CTRepresenting matrix C's
Transposed matrix.
Dictionary is updated using following formula:
In the specific implementation, the solution of formula (3) can be obtained by lower formula:
This when obtained dictionary Y*Expressing dictionary for coupling training sample includes high-low resolution dictionary, then
It is divided into high-resolution to express dictionary according to the resolution ratio of imageLow resolution expresses dictionary YL∈Rc×M。
Step 2, in the training stage of recognition of face, for high-resolution face observation sample, using same in step 1
Sample coupling process, obtain observation sample matrix, observed image matrix.For low-resolution face image square in image pattern
Battle array XL∈Rc×M, high-resolution human face image arrayUpdate the low resolution expression dictionary that dictionary obtains
YL, high-resolution expression dictionary YH, calculate separately under local restriction by low resolution face samples all in low resolution sample database
Best initial weights coefficient when this image linearly rebuilds it and under local restriction by all high in high-resolution sample database
Best initial weights coefficient when resolution ratio face sample image linearly rebuilds it.
In the specific implementation, it is obtained respectively with formula the following:
It is low to inputting by low resolution face sample images all in low resolution sample database as under local restriction
Best initial weights coefficient when image in different resolution is linearly rebuildAs by high score under local restriction
Optimal power when all high-resolution human face sample images linearly rebuild input high-definition picture in resolution sample database
Value coefficient
Step 3, transposition is carried out to the coefficient matrix of high-resolution human face image and obtains matrix T,Wherein tiIt is feature classification, is the best initial weights coefficient of the category, it is assumed that m=t2C, t are amplification coefficients, and c is one low point
Resolution image size, m are a high-definition picture size, activation primitive g (x) and its mathematics of hidden layer neuron number
Formula expression are as follows:
Wherein wiIt is the weight in hidden layer between i neuron and the feature of input layer, biIt is in i-th of hidden layer
Deviation, βiIt is the weight between i-th of neuron and output layer, ojIt is object vector corresponding to j-th of input,Indicate the interior collection of vector.
The target of neural networks with single hidden layer study is the error minimum so that output, can be indicated are as follows:
There is βi, wiAnd bi, so that
It can be expressed as H β=T, it is output weight that wherein H, which is the output β of hidden node, and T is desired output.
In order to train neural networks with single hidden layer, it is intended that obtainSo that:
Wherein, i=1,2 ..., L, this is equivalent to minimize loss function.
Once above it is found that input weight wiB is biased with hidden layeriIt is determined at random, then the output matrix H of hidden layer is just by only
One determines.Training neural networks with single hidden layer, which can be converted into, solves a linear system H β=T.And exporting weight beta can be by
It determines
Wherein, H÷It is the Moore-Penrose generalized inverse of matrix H.And the provable solution acquiredNorm be the smallest
And it is unique.
Step 4, low-resolution face image test sample is inputted, the image pixel size of low resolution is 8 × 7, one
It is altogether M facial images, the expression coefficient of low resolution is then obtained using low resolution local expression dictionary.
Step 5, according to manifold consistency it is assumed that low resolution local restriction expression coefficient is maintained at high-resolution sky
Between, local restriction feature is coupled using high-resolution local restriction expression coefficient as image；
Step 6, according to the node number N of the extreme learning machine determined in step 3, and it is pre- to construct corresponding extreme learning machine
Model is surveyed, the best initial weights coefficient of low-resolution face image is inputtedPredict its corresponding category attribute t.
In another embodiment of the present invention, using AR standard faces library, 100 different faces altogether, everyone front
Image 14 is opened, totally 1400 face-images, and every picture size is unified for 32 × 28, will be adopted under every 4 times of high-definition picture
Sample obtains the image of 8 × 7 low resolution.Everyone takes 7 face figures as training image, 7 face figure conducts in experiment
Test image.Then totally 700, training sample picture, totally 700, test sample picture, training sample picture is high resolution graphics
Piece, test picture are low resolution picture.
Data preparation is done with formula (6) (7) (8) (9) (10), with formula (11) Lai Gengxin dictionary, then by the instruction of coupling
Practice sample expression dictionary and is divided into two expression dictionaries of high-resolution and low resolution according to the resolution ratio of image；With formula (13)
The best initial weights coefficient for finding out low resolution picture finds out the best initial weights coefficient of high-resolution pictures with formula (14), will be low
Resolution ratio local restriction expression coefficient is maintained at high resolution space, using high-resolution local restriction expression coefficient as image
Couple local restriction feature.
Formula (15) represents training sample coefficient matrix and activation primitive, the relationship of hidden neuron number.Randomization
After input weight and hidden layer biasing, pass through our available hidden layer output matrixes of formula (18) (19).According to formula (16)
(17), in order to enable the error of output is minimum, optimal value is obtained by automatically adjusting neuron number.So far, learning machine training
It finishes, the prediction process of learning machine is converted into solution linear system.
Then low-resolution face image test sample is used, obtains low resolution using low resolution local expression dictionary
Coefficient is expressed, according to manifold consistency it is assumed that low resolution local restriction expression coefficient is maintained at high resolution space, is utilized
High-resolution local restriction expresses coefficient as image and couples local restriction feature, by the best initial weights of low-resolution face image
CoefficientIt substitutes into above-mentioned linear system, obtains prediction output valve.
The present invention is different from other face identification methods, and Experimental comparison presented below illustrates the validity of this method.
Experiment is by face identification rate as algorithm standard rules.Experimental result comparison is as shown in the table:
| HR | Bicubic interpolation method | Super-resolution restructing algorithm | sparse | Inventive algorithm | |
| Discrimination | 0.921 | 0.314 | 0.756 | 0.716 | 0.795 |
From above table, it is evident that compared with bicubic interpolation method, Super-resolution Reconstruction algorithm and sparse, the present invention is calculated
Method is higher than other algorithms on discrimination.
As shown in figure 4, the extremely low resolution ratio recognition of face system based on unity couping local constraint representation of the embodiment of the present invention
It unites for realizing the extremely low resolution ratio face identification method based on unity couping local constraint representation of the embodiment of the present invention, including instruction
Practice unit and test cell:
Training unit specifically includes:
Dictionary updating unit is low for obtaining according to the high-resolution human face image in training sample and to its down-sampling
Resolution ratio facial image forms training sample matrix, according to local constraint representation algorithm according to training sample matrix to expression word
Allusion quotation is updated, and is divided into high-resolution expression dictionary and low resolution to express dictionary in expression dictionary；
Coefficient calculation unit calculates it for expressing dictionary according to high-resolution and low-resolution facial image and high-resolution and low-resolution
Best initial weights coefficient under the conditions of local restriction and linear reconstruction, including high-resolution best initial weights coefficient and low resolution
Best initial weights coefficient；
Neuron number computing unit, for according to high-resolution human face image training sample and high-resolution optimal power
Value coefficient is trained limit study, and the error that the number of automatic adjustment study neuron exports identification is minimum, by error
Neuron number when minimum is determined as optimal neuron number, and the training stage completes；
Test cell specifically includes:
Image acquisition unit is obtained for obtaining low-resolution face image to be identified using low resolution expression dictionary
Low resolution to it expresses coefficient；
Binding characteristic determination unit, for according to manifold consistency it is assumed that by low resolution expression coefficient be maintained at high score
Resolution space couples local restriction feature using high-resolution expression coefficient as image；
Image identification unit, for constructing corresponding limit learning model according to determining optimal neuron number, and it is defeated
The best initial weights coefficient for entering low-resolution face image, obtains recognition result.
It should be understood that for those of ordinary skills, it can be modified or changed according to the above description,
And all these modifications and variations should all belong to the protection domain of appended claims of the present invention.
Claims (6)
1. a kind of extremely low resolution ratio face identification method based on unity couping local constraint representation, which is characterized in that including following
Two stages: training stage and test phase；
Training stage comprising the following three steps:
S1, according to the high-resolution human face image in training sample and the low-resolution face image obtained to its down-sampling, group
At training sample matrix, expression dictionary is updated according to training sample matrix according to local constraint representation algorithm, and by table
High-resolution expression dictionary and low resolution expression dictionary are divided into up to dictionary；
S2, dictionary is expressed according to high-resolution and low-resolution facial image and high-resolution and low-resolution, calculates it in local restriction and linear weight
Best initial weights coefficient under the conditions of building, the best initial weights coefficient including high-resolution best initial weights coefficient and low resolution；
S3, limit study is trained according to high-resolution human face image training sample and high-resolution best initial weights coefficient,
The error that the number of automatic adjustment study neuron exports identification is minimum, and neuron number when by error minimum is determined as
Optimal neuron number, training stage complete；
Test phase comprising the following three steps:
S4, low-resolution face image to be identified is obtained, is expressed using the low resolution that low resolution expression dictionary obtains it
Coefficient；
S5, according to manifold consistency it is assumed that by low resolution expression coefficient be maintained at high resolution space, utilize high resolution tables
Local restriction feature is coupled as image up to coefficient；
S6, according in S3 determine optimal neuron number, construct corresponding limit learning model, and input low resolution face
The best initial weights coefficient of image, obtains recognition result.
2. the extremely low resolution ratio face identification method according to claim 1 based on unity couping local constraint representation, special
Sign is that the specific method of step S1 includes:
S11, to high-resolution training sample facial image, corresponding low resolution facial image is obtained by down-sampling, will height
Resolution ratio sample image is launched into column vector, and the coupling of corresponding high-low resolution image vector is become a column vector as table
Up to dictionary；
S12, one initial LCR coefficient of acquisition is calculated using following formula:
Wherein, τ is the regularization parameter of Equilibrium fitting error and local restriction, and M indicates training sample number, and X indicates height point
The column vector that resolution image vector is coupled into, α indicate the row vector of the reconstructed coefficients composition of all M face sample image,Indicate the inner product before two vectors,Indicate European squared-distance,Return to the letter about variable α
The value α of number α when obtaining minimum value*, as required LCR coefficient；
diIndicate reconstructed coefficients αiPenalty factor, calculation formula are as follows:
σ indicates the bandwidth of control distribution, diIndicate the local parameter of the distance between measurement input picture and each dictionary atom,
diValue between zero and one；
S13, the formula for updating expression dictionary are as follows:
Wherein, Y*Expressing dictionary for coupling training sample includes high-resolution and low-resolution dictionary, then according to the resolution ratio of image point
It is cut into high-resolution expression dictionaryLow resolution expresses dictionary YL∈Rc×M。
3. the extremely low resolution ratio face identification method according to claim 1 based on unity couping local constraint representation, special
Sign is that the specific method of step S2 includes:
S21, to high-resolution facial image training sample, it is handled to obtain observation sample matrix, observed image square
Battle array；
S22, calculate low-resolution image best initial weights coefficient formula are as follows:
Wherein, XLFor the low-resolution image in observed image matrix Column vector groups at matrix, YLWord is expressed for low resolution
Allusion quotation, αLTo indicate that all M open the row vector of the reconstructed coefficients composition of low resolution face sample image, αL=(α1,α2,…,αM),Indicate reconstructed coefficientsPenalty factor, τ is the regularization parameter of Equilibrium fitting error and local restriction,It indicates
Inner product before two vectors, | | | | indicate European squared-distance；
It returns about variable αLFunction obtaining minimum value αLWhen valueAs required low resolution
The best initial weights coefficient of rate image
Calculate the formula of the best initial weights coefficient of low-resolution image are as follows:
Wherein, XHFor the high-definition picture Column vector groups in observed image matrix at matrix, YHWord is expressed for high-resolution
Allusion quotation,The as best initial weights coefficient of required high-definition picture,
4. the extremely low resolution ratio face identification method according to claim 1 based on unity couping local constraint representation, special
Sign is, step S3's method particularly includes:
Transposition is carried out to the coefficient matrix of high-resolution human face image and obtains matrix T,Wherein tiIt is feature classification,It is the best initial weights coefficient of the category,Assuming that m=t2C, t are amplification coefficients, and c is a low resolution
Rate image size, m are a high-definition picture size, activation primitive g (x) and hidden layer neuron numberIts formula are as follows:
Wherein, wiIt is the weight in hidden layer between i neuron and the feature of input layer, biIt is inclined in i-th of hidden layer
Difference, βiIt is the weight between i-th of neuron and output layer, ojIt is object vector corresponding to j-th of input,
Indicate the interior collection of vector.
5. the extremely low resolution ratio face identification method according to claim 1 based on unity couping local constraint representation, special
Sign is, step S5's method particularly includes:
With low-resolution face image test sample, the expression coefficient of low resolution is obtained using low resolution expression dictionary, according to
According to manifold consistency it is assumed that low resolution local restriction expression coefficient is maintained at high resolution space, high-resolution office is utilized
Portion's constraint expression coefficient couples local restriction feature as image, then in conjunction with the best initial weights system of low-resolution face image
Number obtains prediction output valve.
6. a kind of extremely low resolution ratio face identification system based on unity couping local constraint representation, which is characterized in that including training
Unit and test cell:
Training unit specifically includes:
Dictionary updating unit, for according to the high-resolution human face image in training sample and the low resolution obtained to its down-sampling
Rate facial image, form training sample matrix, according to local constraint representation algorithm according to training sample matrix to expression dictionary into
Row updates, and is divided into high-resolution expression dictionary and low resolution to express dictionary in expression dictionary；
Coefficient calculation unit calculates it in office for expressing dictionary according to high-resolution and low-resolution facial image and high-resolution and low-resolution
Portion's constraint and it is linear rebuild under the conditions of best initial weights coefficient, most including high-resolution best initial weights coefficient and low resolution
Excellent weight coefficient；
Neuron number computing unit, for according to high-resolution human face image training sample and high-resolution best initial weights system
Number is trained limit study, and the error that the number of automatic adjustment study neuron exports identification is minimum, by error minimum
When neuron number be determined as optimal neuron number, the training stage completes；
Test cell specifically includes:
Image acquisition unit obtains it using low resolution expression dictionary for obtaining low-resolution face image to be identified
Low resolution express coefficient；
Binding characteristic determination unit, for according to manifold consistency it is assumed that by low resolution expression coefficient be maintained at high-resolution
Space couples local restriction feature using high-resolution expression coefficient as image；
Image identification unit for constructing corresponding limit learning model according to determining optimal neuron number, and inputs low
The best initial weights coefficient of resolution ratio facial image, obtains recognition result.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510906586.2A CN105550649B (en) | 2015-12-09 | 2015-12-09 | Extremely low resolution ratio face identification method and system based on unity couping local constraint representation |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510906586.2A CN105550649B (en) | 2015-12-09 | 2015-12-09 | Extremely low resolution ratio face identification method and system based on unity couping local constraint representation |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105550649A CN105550649A (en) | 2016-05-04 |
| CN105550649B true CN105550649B (en) | 2019-03-08 |
Family
ID=55829832
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510906586.2A Active CN105550649B (en) | 2015-12-09 | 2015-12-09 | Extremely low resolution ratio face identification method and system based on unity couping local constraint representation |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105550649B (en) |
Families Citing this family (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106407932B (en) * | 2016-09-20 | 2019-05-28 | 中国石油大学(华东) | Handwritten Digit Recognition method based on fractional calculus Yu generalized inverse neural network |
| CN108197641A (en) * | 2017-12-21 | 2018-06-22 | 武汉工程大学 | A kind of spatial pyramid based on interest domain detection matches image classification method |
| CN108198134B (en) * | 2017-12-29 | 2022-01-18 | 天津科技大学 | Single-image super-resolution reconstruction method based on combined average extreme learning machine |
| CN108986059B (en) * | 2018-06-22 | 2021-12-24 | 武汉工程大学 | Face super-resolution reconstruction method based on local constraint expression of manifold regularization group |
| CN109886869B (en) * | 2018-10-15 | 2022-12-20 | 武汉工程大学 | Non-linear expansion face illusion method based on context information |
| CN110046669B (en) * | 2019-04-22 | 2019-12-24 | 广东石油化工学院 | Pedestrian retrieval method based on sketch image half-coupling metric identification dictionary learning |
| CN110046608B (en) * | 2019-04-26 | 2020-01-07 | 广东石油化工学院 | Leaf-shielded pedestrian re-recognition method and system based on semi-coupling identification dictionary learning |
| CN110796022B (en) * | 2019-10-09 | 2023-07-21 | 奥园智慧生活服务(广州)集团有限公司 | Low-resolution face recognition method based on multi-manifold coupling mapping |
| CN111541505B (en) * | 2020-04-03 | 2021-04-27 | 武汉大学 | Time domain channel prediction method and system for OFDM wireless communication system |
| CN112784857B (en) * | 2021-01-29 | 2022-11-04 | 北京三快在线科技有限公司 | Model training and image processing method and device |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101710386A (en) * | 2009-12-25 | 2010-05-19 | 西安交通大学 | Super-resolution face recognition method based on relevant characteristic and non-liner mapping |
| CN102521810B (en) * | 2011-12-16 | 2013-09-18 | 武汉大学 | Face super-resolution reconstruction method based on local constraint representation |
| CN103208109B (en) * | 2013-04-25 | 2015-09-16 | 武汉大学 | A kind of unreal structure method of face embedded based on local restriction iteration neighborhood |
| CN104008538B (en) * | 2014-05-12 | 2017-03-01 | 清华大学 | Based on single image super-resolution method |
| CN104036482B (en) * | 2014-07-07 | 2017-01-11 | 武汉大学 | Facial image super-resolution method based on dictionary asymptotic updating |
-
2015
- 2015-12-09 CN CN201510906586.2A patent/CN105550649B/en active Active
Also Published As
| Publication number | Publication date |
|---|---|
| CN105550649A (en) | 2016-05-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105550649B (en) | Extremely low resolution ratio face identification method and system based on unity couping local constraint representation | |
| US20200126205A1 (en) | Image processing method, image processing apparatus, computing device and computer-readable storage medium | |
| CN108960330B (en) | Remote sensing image semantic generation method based on fast regional convolutional neural network | |
| CN111242841B (en) | Image background style migration method based on semantic segmentation and deep learning | |
| CN105844635B (en) | A kind of rarefaction representation depth image method for reconstructing based on structure dictionary | |
| CN110490252B (en) | Indoor people number detection method and system based on deep learning | |
| CN107067367A (en) | A kind of Image Super-resolution Reconstruction processing method | |
| CN110060286B (en) | Monocular depth estimation method | |
| CN105701504B (en) | Multi-modal manifold embedding grammar for zero sample learning | |
| CN110347857B (en) | Semantic annotation method of remote sensing image based on reinforcement learning | |
| WO2022062164A1 (en) | Image classification method using partial differential operator-based general-equivariant convolutional neural network model | |
| CN112330813A (en) | Wearing three-dimensional human body model reconstruction method based on monocular depth camera | |
| CN117597703A (en) | Multi-scale converter for image analysis | |
| CN107832747A (en) | A kind of face identification method based on low-rank dictionary learning algorithm | |
| CN113591633A (en) | Object-oriented land utilization information interpretation method based on dynamic self-attention Transformer | |
| CN113506333A (en) | Medical image registration network training data set expansion method based on deformable atlas | |
| CN108898568A (en) | Image composition method and device | |
| CN113516693B (en) | Rapid and universal image registration method | |
| CN108038467B (en) | A kind of sparse face identification method of mirror image in conjunction with thickness level | |
| CN112686830B (en) | Super-resolution method of single depth map based on image decomposition | |
| CN109886869A (en) | A kind of unreal structure method of face of the non-linear expansion based on contextual information | |
| Hu et al. | A Noising-Denoising Framework for Point Cloud Upsampling via Normalizing Flows | |
| CN116993639A (en) | Visible light and infrared image fusion method based on structural re-parameterization | |
| CN109191447B (en) | Three-dimensional grid quality evaluation method based on geometric curvature analysis | |
| CN115690497A (en) | Pollen image classification method based on attention mechanism and convolutional neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |