CN106056067B - Low-resolution face image recognition methods based on corresponding relationship prediction - Google Patents
Low-resolution face image recognition methods based on corresponding relationship prediction Download PDFInfo
- Publication number
- CN106056067B CN106056067B CN201610364643.3A CN201610364643A CN106056067B CN 106056067 B CN106056067 B CN 106056067B CN 201610364643 A CN201610364643 A CN 201610364643A CN 106056067 B CN106056067 B CN 106056067B
- Authority
- CN
- China
- Prior art keywords
- resolution
- low
- training sample
- matrix
- image
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses the low resolution face recognition algorithms (CRPFR) predicted under the conditions of a kind of no corresponding relationship based on corresponding relationship: being the determination of one-to-one relationship first, high-low resolution image is trained, in high resolution space with the structure of other class images with the structure in low-resolution spatial be with piece image it is similar, distinguish whether they have corresponding relationship by comparing the feature vector of height fractional diagram picture;Secondly, by high-low resolution image projection to consistent feature space, classified in feature space using nearest neighbor classifier, keep the overall structure gathered, dispersed between class in class, according to the structure of the one-to-one relationship and classification information controlling feature space that have acquired, target is to acquire a pair of of the height projection matrix that can satisfy feature space structure.Compared with traditional low resolution face recognition algorithms, the present invention achieves more preferably effect under conditions of no corresponding relationship.
Description
Technical field
The present invention relates in pattern-recognition field of face identification more particularly to it is a kind of based on corresponding relationship prediction low point
Resolution facial image recognition method.
Background technique
Recognition of face is the important component of pattern-recognition and computer vision neighborhood, as a most important biology
Feature identification technique, in the past few years, recognition of face (FR) receive research circle and widely pay close attention to and achieve fast development.But
Traditional face recognition technology is difficult to apply in real life, especially exist for high-resolution (HR) facial image
The field of the efficient recognition of face of the urgent needs such as monitoring camera-shooting, authentication.In the actual environment, due to light, shooting angle and
The facial image resolution ratio of the influence of the factors such as capture apparatus, acquisition is lower.Typical scene is on street for the wide of monitoring
Angle camera, the pixel of a width facial image may only have tens or even more than ten, and the discriminant information contained by image is with respect to high score
Resolution image wants much less, thereby produces the identification problem of low resolution (LR) facial image.
Traditional face identification method mainly solves the problems, such as dimensionality reduction, wherein principal component analysis PCA and linear discriminant analysis
LDA is most popular dimension reduction method.High dimensional data is transformed to low-dimensional feature by linear projection by PCA, utmostly to retain
The variance energy of data;LDA is as a kind of supervised learning algorithm, and target is to find an optimal axis of projection, so that after projection not
Generic data point is dispersed and the data point of the same category is assembled.The two-dimensional principal component analysis proposed on the basis of PCA
2DPCA is based on data matrix rather than data vector, thus more simple and effective.The non-linear face identification method then proposed is main
There is kernel-based method, such as core principle component analysis KPCA and core linear discriminant analysis KLDA, and the side based on manifold learning
Method, such as Laplacianfaces Laplacianfaces method.However, the above traditional algorithm in low-resolution image identification problem simultaneously
Good recognition effect cannot be shown, main reason is that the quality of image plays vital work to recognition effect
With the pixel of low-resolution image is low, and useful discriminant information is less, is easy to appear error in the training process, causes to identify
Effect reduces.
There are mainly two types of existing low-resolution face image recognizers: " two steps are walked " method and simultaneously by high-low resolution
The method of image projection.
" two steps are walked " method, i.e., be redeveloped into high-definition picture for low-resolution image with super-resolution method first, so
Identification classification is carried out to the reconstruction image in former high resolution space afterwards.It is closed using the one-to-one correspondence of high-low resolution image
System, is reconstructed into high-definition picture for low-resolution image, then with reconstruction high-definition picture and original high-resolution image
It is identified.Specific practice is to find an optimal projection matrix, make low-resolution image by projection after with corresponding high score
The reconstruction error of resolution image is as small as possible, and uses nearest neighbor classifier rebuilding space, gathers similar reconstruction image,
Inhomogeneity dispersion.
It is simultaneously to find a pair of of projection matrix by the thought that high-low resolution image projects, by high-low resolution image
It is projected simultaneously, using the one-to-one relationship and classification information between them, makes high-low resolution image projection to feature
Gather in class when space, disperse between class, that is to say, that low-resolution face image is directly identified.
Low resolution is either redeveloped into high-resolution with super-resolution method, or high-resolution and low-resolution image is same
When project to feature space, require that high-low resolution image has matched in advance, that is, known high-low resolution image
One-to-one relationship.However in the realistic case, this one-to-one relationship needs are manually indicated, are completed this work and are both expended
Manpower expends the time again.Since above-mentioned algorithm is all based on this precondition, if the condition is invalid, then these algorithms
It will be unable to work.
Summary of the invention
The technical problem to be solved by the present invention is to for involved defect in background technique, provide it is a kind of based on pair
The low-resolution face image recognition methods of Relationship Prediction is answered, to solve the problems, such as that high-low resolution image corresponds.
The present invention uses following technical scheme to solve above-mentioned technical problem:
Low-resolution face image recognition methods based on corresponding relationship prediction comprising the steps of:
Step 1) determines the corresponding relationship of training sample middle high-resolution training sample and low resolution training sample;
Step 2), according to corresponding relationship high-resolution training sample and low resolution training sample establish it is unified
Feature space, find out respectively high-resolution and low-resolution image space to feature space projection matrix PHAnd PL;
The high-resolution training image for being used to be compared is passed through projection matrix P by step 3)HProject to feature space;
It is passed through projection matrix P when needing to identify low-resolution image by step 4)LFeature space is projected to, according to
Its category regions fallen into feature space exports the classification of the low-resolution image.
As the further prioritization scheme of low-resolution face image recognition methods predicted the present invention is based on corresponding relationship,
The detailed step of the step 1) is as follows:
It enablesWithFor n width high-resolution human face training sample and low point
Resolution face training sample, whereinIndicate the i-th corresponding d of (1≤i≤n) width high-resolution human face training sampleHDimension
Training sample vector,Indicate the corresponding d of jth (1≤j≤n) width low resolution face training sampleLTie up training sample to
Amount, and have dH> dL, l (xi) ∈ 1,2 ..., and C } it is training sample xiAffiliated class label;
Step 1.1) calculates separately the structure matrix A of high-resolution and low-resolution training sample according to the following formulaH∈Rn×CAnd AL
∈Rn×C:
Wherein, | | | |2Representation vector 2- norm;
Step 1.2), according to the following formula to structure matrix AHAnd ALIt is normalized, obtains new structure matrix BH
∈Rn×CAnd BL∈Rn×C:
Step 1.3) calculates similar between high-resolution training sample and low resolution training sample according to the following formula
Spend matrix M ∈ Rn×n:
Wherein, BL(i :) representing matrix BLThe i-th row, BH(j :) representing matrix BHJth row, t represents Gaussian function
Width parameter.
As the further prioritization scheme of low-resolution face image recognition methods predicted the present invention is based on corresponding relationship,
The detailed step of the step 2) is as follows:
Step 2.1) calculates similarity matrix W ∈ R using high-resolution training sample according to the following formulan×n:
Wherein,Indicate training sampleSimilar local neighbor sample set, nwIt indicatesIn training sample
This number,Indicate training sampleForeign peoples's local neighbor sample set, nbIt indicatesIn training sample
Number, m are the parameter characterized in training sample set class with category difference;
Step 2.2) constructs diagonal matrix D according to similarity matrix WW, the corresponding diagonal matrix of high-resolution training sampleDiagonal matrix corresponding with low resolution training sample
Step 2.3), structural matrix
Wherein, MTThe transposition of representing matrix M;
Step 2.4) solves the following Generalized Characteristic Equation of formula:
ZUZTP=λ ZVZTp
The corresponding feature vector of preceding d maximum eigenvalue of Generalized Characteristic Equation is lined up matrix P=(p by step 2.5)1,
p2,...,pd), and with the preceding d of matrix PLRow constitutes low-resolution image space to the projection matrix P of feature spaceL, with matrix P
Rear dHRow constitutes high resolution graphics image space to the projection matrix P of feature spaceH。
As the further prioritization scheme of low-resolution face image recognition methods predicted the present invention is based on corresponding relationship,
The detailed step of the step 3) is as follows:
For the high-resolution training image for being used to be comparedSquare is projected according to utilizing according to the following formula
Battle array PHIt is calculated in the corresponding feature vector of feature space
As the further prioritization scheme of low-resolution face image recognition methods predicted the present invention is based on corresponding relationship,
The detailed step of the step 4) is as follows:
Step 4.1), for low-resolution image to be identifiedIt calculates according to the following formulaFeature vector
Step 4.2) calculates in feature spaceIt arrivesIt is corresponding to obtain minimum range therein for distance
High-resolution training image, and using classification belonging to the high-resolution training image as the low-resolution image to be identified
Classification output.
The invention adopts the above technical scheme compared with prior art, has following technical effect that
1. under conditions of no high-low resolution image corresponding relationship, by comparing in high-low resolution image space
Image pattern distributed architecture obtains their corresponding relationship, reduces manpower, saves the time.
2. establishing consistent feature space using the method for projecting high-low resolution image simultaneously.
3. utilizing the result gathered, dispersed between class in two constraint condition limited features spatial class in projection.
Compared with the methods of traditional CLPM, CDMMA, the maximum bright spot of the present invention is independent of high-low resolution figure
The corresponding relationship condition of picture, the distributed architecture by comparing image learn to obtain this corresponding relationship.On feature extraction algorithm,
High-low resolution image is projected into feature space simultaneously, is effectively reduced manpower, obtains higher more stable identification effect
Fruit.
Detailed description of the invention
Fig. 1 is the work flow diagram of present invention training and test.
Specific embodiment
Technical solution of the present invention is described in further detail with reference to the accompanying drawing:
The invention discloses the low resolution face recognition algorithms (NCRC) under the conditions of a kind of no corresponding relationship: being first
The determination of high-resolution and low-resolution image one-to-one relationship.It is trained based on high-resolution and low-resolution image, piece image is in high score
It is similar, such distribution pass with the distribution relation in low-resolution spatial with the distribution relation of other images in resolution space
System is portrayed by similarity vector.Whether them are distinguished by comparing the similarity vector of high and low fractional diagram picture has pair
It should be related to.Secondly, a pair of of high-resolution and low-resolution projection matrix is calculated according to the one-to-one relationship and classification information acquired,
High-resolution and low-resolution image projection to a unified feature space is dispersed between class to realize to gather in class accordingly, is then existed
Facial image identification is carried out by nearest neighbor classifier in this feature space.
Low resolution face recognition algorithms under the conditions of no corresponding relationship of the present invention, comprising the following steps:
(1) one-to-one relationship is determined:
It enablesWithFor n width high-resolution human face training image and low point
Resolution face training image, whereinIndicate the i-th corresponding d of (1≤i≤n) width high-resolution human face training imageHDimension
Training sample vector,Indicate the corresponding d of jth (1≤j≤n) width low resolution face training imageLTie up training sample to
Amount, and have dH> dL.Enable l (xi) ∈ 1,2 ..., and C } indicate sample xiAffiliated class label.
The one-to-one relationship of high-low resolution image in order to obtain, it will be assumed that the sample in low-resolution image space
It is distributed with uniformity with the sample distribution of high resolution graphics image space.For high-definition pictureAnd low-resolution imageWhether there is one-to-one relationship to need to compare distribution of this two images in respective space according to assumed condition.
(1) seeks eigenmatrix:
For the ease of comparing, indicate the distribution of piece image in space with the structure vector of C dimension here, this to
K-th of element of amount be in the width image and kth class the square distance of all samples and.Specifically, high-definition pictureIt is right
The structure vector A answeredH(i, k) and low-resolution imageCorresponding structure vector AL(j, k) is defined as
Wherein, the label of l (x) representative sample x.
For high-resolution and low-resolution training image sample, the eigenmatrix A of n × C can be calculated according to (1) formulaHAnd AH, right
Image pattern distribution in high-resolution and low-resolution space is portrayed.However, matrix VHAnd VLThe numerical measure of middle element is usually
Inconsistent, so needing that the element of the two matrixes is normalized, obtain new structure matrix BH∈Rn×CAnd BL
∈Rn×C:
Acquired new structure matrix is in the same order of magnitude, is comparable.
(2) seeks similar matrix:
The corresponding relationship of high-resolution and low-resolution image in order to obtain, need to calculate the structure vector of high-definition picture with it is low
Thus similarity between the structure vector of image in different resolution obtains the similarity matrix M of high-resolution and low-resolution image:
Wherein, BL(i :) representing matrix BLThe i-th row, BH(j :) representing matrix BHJth row, t represents Gaussian function
Width parameter.In obtained similarity M, if two images inhomogeneity, they can not have one-to-one correspondence to close
System, therefore corresponding similarity value M (i, j) is 0;If two images belong to same class, M (i, j) indicates that they are similar
Degree.The value of M (i, j) is bigger, shows that this two images is bigger a possibility that there are one-to-one relationships.
(2) consistent feature space is established
To be conducive to classify, high-resolution and low-resolution image is passed through projection matrix P respectively by usHAnd PLProject to the same spy
Levy space.In this feature space, similar sample should be gathered as far as possible, and foreign peoples's sample should disperse as far as possible.According to phase
Like the definition for spending matrix M it is found that if two images have one-to-one relationship, they project to Ying Jinke after feature space
It can gather, therefore, we are defined as follows function:
M (i, j) numerical value is bigger, indicates thisWithIt is more possible to one-to-one relationship.So, they are in feature sky
Between distance should be smaller.Therefore, optimal projection matrix should make function Js(PH,PL) obtain minimum value.
Meanwhile in order to which using the classification information of training sample, we are defined as follows function:
Wherein, weight matrix W is defined as:
Wherein, m is parameter, nwFor local neighborhood NwThe number of samples for including, nbFor local neighborhood NbThe sample for including
Number,.From (6) formula it is found that matrix W features three kinds of relationships between image:
1. high-definition pictureWithBelong to same class and each other neighbour, then corresponding low-resolution image projects to
The distance of feature space should very little, at this time we take W (i, j) be a negative value
2. high-definition pictureWithBelong to different classes, butWithNeighbour each other, then its corresponding low point
For resolution image after projecting to feature space, distance should be very big, we take the W (i, j) to be at this time
3. high-definition pictureWithWithout neighbor relationships, then W (i, j) is taken as 0, shows in Jc(PH,PL) in this
The sample of sample does not have an impact projection matrix.
(3) algorithm model and solution
Above two constraint can acquire projection matrix PHAnd PLIf both constraints are combined together, to projection
Matrix has stronger constraint.High-low resolution image projection is utilized into their corresponding relationship and classification to special space respectively
Information limits projection matrix, so that reaching ideal distance in feature space, so as to find out projection matrix PHAnd PL, objective function
It is as follows:
Gather after of a sort high-low resolution image projection to feature space, and similarity is higher, distance is closer, different
Disperse after the high-low resolution image projection to feature space of class, in order to reach this target, our algorithm model are as follows:
max J(PH,PL) (8)
Through abbreviation, formula (8) can abbreviation are as follows:
WhereinIt is the diagonal matrix constructed according to matrix W,WithIt is to be respectively corresponded based on the diagonal matrix of matrix M construction
High-resolution and low-resolution space,
According to lagrange multiplier approach, Lagrangian can be constructed are as follows:
L=tr (PTZUZTP)-λtr(PTZVZTP) (10)
To L is about P derivation and to enable be 0:
ZUZTP=λ ZVZTP (11)
Feature decomposition is carried out to formula (11), feature vector corresponding to d maximum eigenvalue before asking forms projection matrix P,
Decomposition obtains PLAnd PH。
(4) high-resolution training image feature extraction
By high-resolution training imagePass through formula (12) projection matrix PHProject to feature sky
Between, obtain the d dimensional feature of high-resolution training image
(5) it identifies
It enablesFor low-resolution image to be identified, in order to predictAffiliated classification first willPass through projection matrix PLIt throws
Shadow obtains its corresponding d dimensional feature to feature space:
Then it is calculated separately in feature spaceWith in Ω at a distance from sample, enable away fromNearest high-resolution sample is special
Sign isThen determineBelong toIdentical classification, i.e.,
By the low resolution face recognition algorithms under the conditions of no corresponding relationship of the present invention in extension Yale-B data
It is tested on library and CMU PIE database, and experimental result is compared to relevant low resolution face recognition algorithms
Analysis.
The 2432 width direct pictures that Yale-B database includes 38 people are extended, everyone there are 64 width or so in different illumination
Under close to positive image.The face of each original image is tailored to 192 × 168 image.As shown in table 1, at us
Experiment in, the pixel of high-definition picture is taken as 32 × 28, low-resolution image capture element is 10 × 10, everyone takes at random
The image of half does test set as training set, remaining half, carries out 20 random experiments, obtains average recognition rate such as table 1:
Table 1 extends the discrimination of various algorithms on Yale-B database
CMU PIE database includes the 41368 width images of 68 people, everyone image has 13 kinds of different postures, 43
Kind different light conditions and 4 kinds of different expressions.In our experiment, the pixel of high-definition picture is taken as 32 × 32,
Low-resolution image capture element is 15 × 15, everyone takes the image of half as training set at random, and remaining half does test set, into
20 random experiments of row obtain average recognition rate such as table 2:
Table 2 extends the discrimination of various algorithms on CMU PIE database
By table 1, table 2 as it can be seen that on different face databases, compared to other three kinds of algorithms, our algorithm does not have
It is supplied to one-to-one known conditions, but there is no influenced recognition effect by great.As can be seen that " two steps with early stage
Walk " method GDAMM compares, develop at present by high-low resolution image while the method (including the CLPM, CDMMA that project
Deng) 35% or so is higher by recognition effect, and the stability of algorithm also substantially increases.NCRC proposed by the present invention is not only
It is better than other methods on discrimination, and stability is higher.
The experimental results showed that the low resolution face recognition algorithms of no corresponding relationship described herein are better than other low point
Resolution algorithm, and maximum bright spot is based on no corresponding relationship, is the algorithm for being more suitable for low-resolution image identification.
Those skilled in the art can understand that unless otherwise defined, all terms used herein (including skill
Art term and scientific term) there is meaning identical with the general understanding of those of ordinary skill in fields of the present invention.Also
It should be understood that those terms such as defined in the general dictionary should be understood that have in the context of the prior art
The consistent meaning of meaning will not be explained in an idealized or overly formal meaning and unless defined as here.
Above-described specific embodiment has carried out further the purpose of the present invention, technical scheme and beneficial effects
It is described in detail, it should be understood that being not limited to this hair the foregoing is merely a specific embodiment of the invention
Bright, all within the spirits and principles of the present invention, any modification, equivalent substitution, improvement and etc. done should be included in the present invention
Protection scope within.
Claims (4)
1. the low-resolution face image recognition methods based on corresponding relationship prediction, which is characterized in that comprise the steps of:
Step 1) determines the corresponding relationship of training sample middle high-resolution training sample and low resolution training sample;
It enablesWithFor n width high-resolution human face training sample and low resolution
Face training sample, whereinIndicate the i-th corresponding d of (1≤i≤n) width high-resolution human face training sampleHDimension training
Sample vector,Indicate the corresponding d of jth (1≤j≤n) width low resolution face training sampleLTraining sample vector is tieed up,
And there is dH> dL, l (xi) ∈ 1,2 ..., and C } it is training sample xiAffiliated class label;
Step 1.1) calculates separately the structure matrix A of high-resolution and low-resolution training sample according to the following formulaH∈Rn×CAnd AL∈Rn ×C:
Wherein, | | | |2Representation vector 2- norm;
Step 1.2), according to the following formula to structure matrix AHAnd ALIt is normalized, obtains new structure matrix BH∈Rn ×CAnd BL∈Rn×C:
Step 1.3) calculates the similarity moment between high-resolution training sample and low resolution training sample according to the following formula
Battle array M ∈ Rn×n:
Wherein, BL(i :) representing matrix BLThe i-th row, BH(j :) representing matrix BHJth row, t represents the width of Gaussian function
Parameter;
Step 2), according to corresponding relationship high-resolution training sample and low resolution training sample establish unified feature
Space, find out respectively high-resolution and low-resolution image space to feature space projection matrix PHAnd PL;
The high-resolution training image for being used to be compared is passed through projection matrix P by step 3)HProject to feature space;
It is passed through projection matrix P when needing to identify low-resolution image by step 4)LProject to feature space, according to its
The category regions that feature space is fallen into export the classification of the low-resolution image.
2. the low-resolution face image recognition methods according to claim 1 based on corresponding relationship prediction, feature exist
In the detailed step of the step 2) is as follows:
Step 2.1) calculates similarity matrix W ∈ R using high-resolution training sample according to the following formulan×n:
Wherein,Indicate training sampleSimilar local neighbor sample set, nwIt indicatesIn training sample
Number,Indicate training sampleForeign peoples's local neighbor sample set, nbIt indicatesIn training sample number, m is
Characterize the parameter in training sample set class with category difference;
Step 2.2) constructs diagonal matrix D according to similarity matrix WW, the corresponding diagonal matrix of high-resolution training sampleWith
The corresponding diagonal matrix of low resolution training sample
Step 2.3), structural matrix
Wherein, MTThe transposition of representing matrix M;
Step 2.4) solves the following Generalized Characteristic Equation of formula:
ZUZTP=λ ZVZTp
The corresponding feature vector of preceding d maximum eigenvalue of Generalized Characteristic Equation is lined up matrix P=(p by step 2.5)1,
p2,...,pd), and with the preceding d of matrix PLRow constitutes low-resolution image space to the projection matrix P of feature spaceL, with matrix P
Rear dHRow constitutes high resolution graphics image space to the projection matrix P of feature spaceH。
3. the low-resolution face image recognition methods according to claim 2 based on corresponding relationship prediction, feature exist
In the detailed step of the step 3) is as follows:
For the high-resolution training image for being used to be comparedAccording to according to the following formula utilize projection matrix PH
It is calculated in the corresponding feature vector of feature space
4. the low-resolution face image recognition methods according to claim 3 based on corresponding relationship prediction, feature exist
In the detailed step of the step 4) is as follows:
Step 4.1), for low-resolution image to be identifiedIt calculates according to the following formulaFeature vector
Step 4.2) calculates in feature spaceIt arrivesDistance obtains the corresponding high score of minimum range therein
Resolution training image, and using classification belonging to the high-resolution training image as the classification of the low-resolution image to be identified
Output.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610364643.3A CN106056067B (en) | 2016-05-27 | 2016-05-27 | Low-resolution face image recognition methods based on corresponding relationship prediction |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201610364643.3A CN106056067B (en) | 2016-05-27 | 2016-05-27 | Low-resolution face image recognition methods based on corresponding relationship prediction |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106056067A CN106056067A (en) | 2016-10-26 |
CN106056067B true CN106056067B (en) | 2019-04-12 |
Family
ID=57175469
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201610364643.3A Active CN106056067B (en) | 2016-05-27 | 2016-05-27 | Low-resolution face image recognition methods based on corresponding relationship prediction |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106056067B (en) |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106686310A (en) * | 2017-01-13 | 2017-05-17 | 浪潮(苏州)金融技术服务有限公司 | Method and system of dynamically improving the practical effect of shooting equipment |
CN108629246B (en) * | 2017-03-23 | 2024-03-01 | 北京行易道科技有限公司 | Vehicle-mounted image processing method and device and vehicle |
CN108665441B (en) * | 2018-03-30 | 2019-09-17 | 北京三快在线科技有限公司 | A kind of Near-duplicate image detection method and device, electronic equipment |
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 |
CN111695455B (en) * | 2020-05-28 | 2023-11-10 | 广西申能达智能技术有限公司 | Low-resolution face recognition method based on coupling discrimination manifold alignment |
WO2022087778A1 (en) * | 2020-10-26 | 2022-05-05 | 深圳大学 | Low-resolution image recognition method based on multi-layer coupled mapping |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101299235A (en) * | 2008-06-18 | 2008-11-05 | 中山大学 | Method for reconstructing human face super resolution based on core principle component analysis |
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 |
JP2011034343A (en) * | 2009-07-31 | 2011-02-17 | Fujifilm Corp | Image processing apparatus and method, data processing device and method, and program |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP4757116B2 (en) * | 2006-06-30 | 2011-08-24 | キヤノン株式会社 | Parameter learning method and apparatus, pattern identification method and apparatus, and program |
US9158992B2 (en) * | 2013-03-14 | 2015-10-13 | Here Global B.V. | Acceleration of linear classifiers |
-
2016
- 2016-05-27 CN CN201610364643.3A patent/CN106056067B/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101299235A (en) * | 2008-06-18 | 2008-11-05 | 中山大学 | Method for reconstructing human face super resolution based on core principle component analysis |
JP2011034343A (en) * | 2009-07-31 | 2011-02-17 | Fujifilm Corp | Image processing apparatus and method, data processing device and method, and program |
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 |
Non-Patent Citations (1)
Title |
---|
"Low-Resolution Face Recognition via Coupled Locality Preserving Mappings";Bo Li et. al;《IEEE SIGNAL PROCESSING LETTERS》;20090904;第2章,图1 |
Also Published As
Publication number | Publication date |
---|---|
CN106056067A (en) | 2016-10-26 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106056067B (en) | Low-resolution face image recognition methods based on corresponding relationship prediction | |
Tabernik et al. | Deep learning for large-scale traffic-sign detection and recognition | |
Chen et al. | Deep learning features at scale for visual place recognition | |
Jin et al. | Uncertainty-aware multi-shot knowledge distillation for image-based object re-identification | |
Zhang et al. | Fully convolutional adaptation networks for semantic segmentation | |
Tao et al. | Smoke detection based on deep convolutional neural networks | |
Zhang et al. | Joint dynamic sparse representation for multi-view face recognition | |
CN109740413A (en) | Pedestrian recognition methods, device, computer equipment and computer storage medium again | |
CN106023257B (en) | A kind of method for tracking target based on rotor wing unmanned aerial vehicle platform | |
Li et al. | Recognizing materials from virtual examples | |
CN107463920A (en) | A kind of face identification method for eliminating partial occlusion thing and influenceing | |
Zhao et al. | Transfer learning with ensemble of multiple feature representations | |
Liao et al. | A deep ordinal distortion estimation approach for distortion rectification | |
Spadotto et al. | Unsupervised domain adaptation with multiple domain discriminators and adaptive self-training | |
Koščević et al. | Deep learning-based illumination estimation using light source classification | |
CN110363218A (en) | A kind of embryo's noninvasively estimating method and device | |
CN110796022A (en) | Low-resolution face recognition method based on multi-manifold coupling mapping | |
Parde et al. | Deep convolutional neural network features and the original image | |
Zhang et al. | The performance research of the data augmentation method for image Classification | |
CN111695455A (en) | Low-resolution face recognition method based on coupling discrimination manifold alignment | |
Wang et al. | Fusion of multiple channel features for person re-identification | |
CN114550110A (en) | Vehicle weight identification method and system based on unsupervised domain adaptation | |
Xi et al. | SR-POD: Sample rotation based on principal-axis orientation distribution for data augmentation in deep object detection | |
Zheng et al. | Overwater image dehazing via cycle-consistent generative adversarial network | |
CN105975921B (en) | Pedestrian detection method based on local feature symbiosis and Partial Least Squares |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
C06 | Publication | ||
PB01 | Publication | ||
C10 | Entry into substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
CB02 | Change of applicant information | ||
CB02 | Change of applicant information |
Address after: No. 66, New Model Road, Gulou District, Nanjing City, Jiangsu Province, 210000 Applicant after: Nanjing Post & Telecommunication Univ. Address before: 210023 9 Wen Yuan Road, Ya Dong new town, Nanjing, Jiangsu. Applicant before: Nanjing Post & Telecommunication Univ. |
|
GR01 | Patent grant | ||
GR01 | Patent grant |