CN109558814A - A kind of three-dimensional correction and weighting similarity measurement study without constraint face verification method - Google Patents
A kind of three-dimensional correction and weighting similarity measurement study without constraint face verification method Download PDFInfo
- Publication number
- CN109558814A CN109558814A CN201811353070.XA CN201811353070A CN109558814A CN 109558814 A CN109558814 A CN 109558814A CN 201811353070 A CN201811353070 A CN 201811353070A CN 109558814 A CN109558814 A CN 109558814A
- Authority
- CN
- China
- Prior art keywords
- face
- image
- correction
- constraint
- target
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/168—Feature extraction; Face representation
-
- 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
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
-
- 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
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/24—Aligning, centring, orientation detection or correction of the image
- G06V10/247—Aligning, centring, orientation detection or correction of the image by affine transforms, e.g. correction due to perspective effects; Quadrilaterals, e.g. trapezoids
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Physics & Mathematics (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Data Mining & Analysis (AREA)
- Human Computer Interaction (AREA)
- General Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The present invention relates to without constraint face verification technical field, and in particular to a kind of three-dimensional correction and weighting similarity measurement study without constraint face verification method.Firstly, proposing a kind of effective three-dimensional face antidote for without influence of the face posture variation to verifying accuracy in constraint facial image.Target human face characteristic point screening need to be carried out for plurality of human faces image by carrying out facial feature points detection to facial image first.Secondly, the human face characteristic point using detection carries out three-dimensional face correction, and cut the image after correction face relevant range to remove complicated image background.Finally, a kind of method for proposing weighting similarity measurement study, two methods of similarity measurement study and mahalanobis distance metric learning are weighted and are combined, the similarity between two images can preferably be learnt, it is more conducive to the differentiation of similar image pair with dissimilar image pair, improves verifying accuracy rate.
Description
Technical field
The present invention relates to without constraint face verification technical field, and in particular to one kind is based on three-dimensional correction and weighting similitude
Metric learning without constraint face verification method.
Background technique
Face is as current one of biometrics identification technology, because its feature directly, conveniently, friendly has obtained extensively
Research and application.The purpose of face verification is to judge whether two images are the same person.Wherein, no constraint face verification is
One very challenging problem.Without constraint face face posture, illumination, background, in terms of variation very greatly,
And it is frequently accompanied by and blocks, and these variations also exert a certain influence to the accuracy rate of verifying.
Face posture variation is a key factor for causing discrimination to decline, and many researchers attempt by studying people
The method of face correction solves the problems, such as that face posture changes this.Fontaine et al. proposes one based on triangle gridding deformation
And the 2D face antidote of affine transformation generates the front face image of input picture.Blanz et al. uses a mark
Quasi- faceform and an illumination model, the feature extraction of facial image is carried out first with the method for principal component analysis,
The parameter in model is being determined by beta function, is finally obtaining the threedimensional model of the facial image, completes face correction.
Kemelmacher et al. restores the relative altitude of its surface each point using the light and shade variation on objects in images surface, completes object
The three-dimensionalreconstruction of body.Although this method is feasible, it is very sensitive for the reflection problems (glasses) of occlusion issue and mirror surface,
And it usually needs in advance to separate facial area from background.
The problem of for without constraint facial image, researcher also proposed some based on metric learning method
Algorithm.How the similarity study that metric learning is namely often said measures image if necessary to calculate the similarity of two images
Between similarity make different classes of image similarity small, and the image similarity of the same category greatly be exactly metric learning
Target.Fu et al. proposes a kind of method for learning relevance metric, which, can be with after carrying out dimensionality reduction to sample
Retain the neighbor relationships between sample, author proposes related insertion analysis (CEA, Correlation also directed to relevance metric
Embedding Analysis) model and related principal component analysis (CPCA, Correlation Principle Component
Analysis) model.Nguyen and Bai et al. propose cosine similarity metric learning (CSML) model, more than the model use
String similarity constructs objective function.Huang et al. proposes the study of broad sense sparse measurement (GSML, Generalized
Sparse Metric Learning) model, this method provides one for many representational sparse measurement learning models
Unified angle, and existing many non-sparse measurement learning models can be expanded into sparse measurement study form.Measurement
Study can be divided into two classes: learning distance metric and similarity measurement study.Most learning distance metric is intended to learn geneva
Distance:Wherein x and y indicates feature vector, and M is the matrix for needing to learn.And similarity measurements
Amount study is intended to learn the similitude of following form: sM(a, b)=aTMb/N (a, b), wherein N (a, b) is a standardization item.
As N (a, b)=1, sM(a, b) is a bilinearity Similarity equations；WhensM(a, b) is one wide
Adopted cosine similarity function.
Summary of the invention
The technical problem to be solved by the present invention is in order to solve to become without face posture, expression, illumination in constraint facial image
Changing big and race, age range greatly influences verifying accuracy rate bring, and the present invention provides one kind to be rectified based on three-dimensional face
Just pass through three-dimensional face correction and the relevant range of face without constraint face verification method with weighting similarity measurement study
Cutting, reduce face posture variation and complexity image background on verifying bring influence, improve verifying accuracy rate.
It can be more efficient by measuring similarity study and learning distance metric weighted combination by weighting similarity measurement study
Study calculate two images between similarity, be conducive to similar image pair with dissmilarity image pair differentiation.
The technical solution adopted by the present invention to solve the technical problems is:
It is a kind of correct and weight based on three-dimensional face similarity measurement study without constraint face verification method, including it is following
Step:
1) input is without constraint facial image；
2) facial feature points detection is carried out with existing Dlib feature detection algorithm；
3) selection of target human face characteristic point is carried out to plurality of human faces image；
4) three-dimensional face correction is carried out using the human face characteristic point detected；
5) the face relevant range of the facial image after correction is cut；
6) feature extraction (such as Gabor, LBP etc.) is carried out to the facial image after cutting, and dimensionality reduction (such as PCA, WPCA
Deng)；
7) similarity measurement study is weighted to the face characteristic after dimensionality reduction, obtains final verification result.
Specifically, described is to be possible to meeting in no constraint image to the progress target human face characteristic point selection of plurality of human faces image
Other faces other than goal task are showed, in order to correct to correct target face, need to carry out target face
Selection.Under normal conditions, as the face of photographer's target person in the picture should in occupation of bigger space, therefore
In the case that one image has multiple faces, the space that the face of target person accounts for for other faces in photo should
It is bigger.The face frame that the Dlib algorithm used herein is positioned does not consider the face size issue in facial image, no
It is appropriate for the screening operation of target face, therefore the characteristic point of all faces occurred in image is detected and retained with Dlib,
VJ algorithm is recycled to carry out Face detection, which can provide the specific location (face frame) of face in image, calculate face frame
Size is simultaneously ranked up, and selects target face of the maximum face of area of corresponding face frame as the image, and retain the mesh
Mark characteristic point detected by face.
Specifically, the three-dimensional face antidote is, using a single, constant 3D faceform, first to use camera
The principle of projection matrix generates the 2D front face image (reference picture) of the 3D face, and saves 3D when generating reference picture
The corresponding relationship between each point between model and 2D image.The reference picture and input picture are detected with Dlib algorithm
Human face characteristic point can find input picture and 3D model by the human face characteristic point and 3D model of two images as intermediary
And its corresponding relationship between 2D reference picture each point.Using the corresponding relationship, by the human face region in the facial image of input
After projecting in the conventional coordinates (i.e. reference picture) generated by 3D faceform, pixel sampling is recycled to generate a correction
Front face image afterwards.
Specifically, the face relevant range for cutting the facial image after correction is complicated back in no constraint image
Scape problem will affect verifying accuracy rate, and after correction, the face of all images can all appear in the central area of image, utilize
This characteristic can carry out the cutting of face relevant range, to remove complicated image background.
Specifically, the method that the weighting similarity measurement learns is, after completing feature extraction and dimensionality reduction, by sample
Originally it is divided into training sample and test sample, learns optimal similar matrix G, distance matrix M and classification out using training sample
Threshold value σ.By similar matrix G and distance matrix M, similar matrix G and distance matrix weighted value can be calculated.It will according to weighted value
The two weighted combination is sentenced in the similarity scores that two test images can be obtained when similarity scores are more than or equal to threshold value σ
Disconnected two images are similar (being the same person), are otherwise judged as dissimilar.
Similarity measurement is corrected and weighted based on three-dimensional face the beneficial effects of the present invention are: the present invention provides one kind
That practises reduces face appearance by three-dimensional face correction and the cutting of the relevant range of face without constraint face verification method
The image background of gesture variation and complexity influences verifying bring, improves verifying accuracy rate.By weighting similarity measurement
Study being capable of two images of significantly more efficient study calculating by measuring similarity study and learning distance metric weighted combination
Between similarity, be conducive to the differentiation of similar image pair with dissimilar image pair.
Detailed description of the invention
Present invention will be further explained below with reference to the attached drawings and examples.
Fig. 1 is of the invention without constraint face verification method flow block diagram；
Fig. 2 is the features of human face images distribution schematic diagram of the invention detected by Dlib；
Fig. 3 is face frame schematic diagram detected by Dlib algorithm and VJ algorithm of the invention；
Fig. 4 is 2D reference picture of the invention and its characteristic point distribution schematic diagram that is detected by Dlib；
Fig. 5 is facial image correction of the invention and cuts front and back lab diagram；
Specific embodiment
In conjunction with the accompanying drawings, the present invention is further explained in detail.
Fig. 1 is of the invention without constraint face verification method flow block diagram, and Fig. 2 is face frame detection signal of the invention
Figure, Fig. 3 are human face characteristic point schematic diagrames of the invention, and Fig. 4 is facial image correction of the invention and cuts front and back lab diagram.
It is a kind of correct and weight based on three-dimensional face similarity measurement study without constraint face verification method, including it is following
Step:
1) input is without constraint facial image；
2) facial feature points detection is carried out with existing Dlib feature detection algorithm；
3) selection of target human face characteristic point is carried out to plurality of human faces image；
4) three-dimensional face correction is carried out using the human face characteristic point detected；
5) the face relevant range of the facial image after correction is cut；
6) feature extraction (such as Gabor, LBP etc.) is carried out to the facial image after cutting, and dimensionality reduction (such as PCA, WPCA
Deng)；
7) similarity measurement study is weighted to the face characteristic after dimensionality reduction, obtains final verification result.
It is described to be using the human face characteristic point detected progress three-dimensional face antidote, it is single for one, constant
3D faceform selectes a camera projection matrix CM=AM[RM tM] (wherein AMFor camera internal parameter matrix, [RM tM] be
By spin matrix RMAnd translation vector tMThe camera external parameter matrix of composition), the 2D front face of the 3D model can be generated
Image (reference picture, while being also reference frame).
Generating reference picture IRWhen, for each pixel p' in image store its corresponding three-dimensional coordinate P=(X, Y,
Z)T:
P'~CMP (1)
For the test image I of inputQ, remember pi=(xi,yi)TFor the two-dimensional coordinate of the characteristic point of test image, to reference
Image IRCharacteristic point detection also equally is carried out using same characteristic detection method, and remembers that its characteristic point coordinate is pi'=(xi',
yi')T, by the available reference picture characteristic point p of equation (1)i' the corresponding 3D coordinate P of characteristic point on 3D faceformi
=(Xi, Yi, Zi)T。
Since the property detector used is identical, so the characteristic point p of test imageiWith the characteristic point p of reference picturei' be
Correspondingly, therefore piWith 3D coordinate PiThere has also been corresponding relationships.Likewise, using camera calibration principle, byBetween corresponding relationship, can approximation obtain test image used camera when being taken
Projection matrix MQ=AQ[RQtQ].There are the projection approximation matrix of test image, each pixel in available test image
With the corresponding relationship of the 3D coordinate of each point of 3D faceform, therefore also just have in test image each pixel with reference to figure
Corresponding relationship as between each pixel, using bilinear interpolation, by pixel value sampling minute of the test image at p point
It is fitted at position p' corresponding to reference picture, thus generates an initial front face of test image.It is correcting
Later, the face part of all images is at the center of image, is cut using this characteristic to face
Face is separated from background.
As shown in Fig. 2, the features of human face images distribution schematic diagram detected for Dlib.As shown in Fig. 2, by Dlib
The human face characteristic point detected shares 68, is respectively distributed on eyes, eyebrow, nose mouth and jaw line.
It as shown in Fig. 3, is face frame schematic diagram detected by Dlib algorithm and VJ algorithm, figure (a) is calculated by Dlib
The face frame of method detection label, figure (b) are by the face frame of VJ algorithm detection label.Scheme a left side for the face frame of the left side (a) face
The coordinate at upper angle is (5,63), and the coordinate in the lower right corner is (79,138), therefore the size for scheming the face frame of the left side (a) face is
(79-5) × (138-63)=5550；Scheme (a) on the right of face (target face) face frame the upper left corner coordinate be (88,
96), the coordinate in the lower right corner is (162,171), therefore the size for scheming the face frame of face on the right of (a) is (162-88) × (171-
96)=5550, two face frame sizes are the same, can not compare.The coordinate in the upper left corner for scheming the face frame of the left side (b) face is
(6,51), the coordinate in the lower right corner are (87,132), therefore the size for scheming the face frame of the left side (a) face is (87-6) × (132-
51)=6561；The coordinate for scheming the upper left corner of the face frame of face (target face) on the right of (a) is (75,72), the coordinate in the lower right corner
For (181,178), therefore the size for scheming the face frame of face on the right of (a) is (181-75) × (178-72)=11236, people from the right
Face frame is greater than the left side, therefore selected the right face is target face.VJ algorithm confines position in face known to above-mentioned calculated result
On performance ratio Dlib algorithm it is more excellent, be more suitable as the screening operation of target face.
It as shown in Fig. 4, is 2D reference picture and its characteristic point distribution schematic diagram detected by Dlib.Such as 4 institute of attached drawing
Show that the 2D reference picture is 3D model 2D front face image generated.
As shown in Fig. 5, front and back lab diagram is corrected and cut for facial image.Wherein, original image size is 250 × 250,
Image size after cutting is 90 × 90.
The method of the weighting similarity measurement study is, by similarity function sMWith mahalanobis distance function dMIt is weighted
In conjunction with obtaining the similarity function f an of broad sense(M,G)To measure two image (ai,aj) between similitude:
f(M,G)(ai,aj)=wsG(ai,aj)-(1-w)dM(ai,aj) (2)
Wherein,It is balance similarity function sMWith mahalanobis distance function dMWeight, (ai,aj) it is by spy
The feature vector of two images after sign extraction and dimensionality reduction operation.Γ=S ∪ D is enabled to indicate all indexed sets constrained in pairs,
If image aiAnd ajIt is similar, export sij=1, otherwise sij=-1.Loss function is for estimating the predicted value of model and true
Inconsistent degree between value can be released using hinge loss:
Minimize the differentiation that the above-mentioned experience error about G is beneficial to similar image pair with dissimilar image pair.It is added
Regularization frame avoids over-fitting, similarity metric function that learn a robust and having distinction, what constraint to be optimized
Parameter:
Wherein, γ is regularization coefficient.The introducing of slack variable can reduce influence of the noise spot to classification:
Using Lagrange duality, multiplier α, β are introduced, objective function (5) is rewritten as
It is available by seeking local derviation to M, G and ξ
Wherein At=(ai-aj)(ai-aj)T,Formula (7) substitution equation (6) can be obtained into objective function finally:
After carrying out the study of above-mentioned weighting similarity measurement using training sample, optimal similar matrix G after being learnt,
Distance matrix M.Using G and M, the weighting similarity scores of training sample are calculated, choose maximum similarity scores ma and
The smallest similarity scores mi, classification thresholds σ range are [mi, ma], step-length (ma-mi)/6000.To each threshold value σ, pass through
By similarity scores compared with the threshold value (when similarity scores be more than or equal to threshold value σ when, judge two images it is similar (be same
It is personal), otherwise it is judged as dissimilar), by the sample number correctly classified divided by the available classification accuracy of total sample number
Accu:
Record the threshold value σ when obtaining highest accuracy rate on training sample*, and by threshold value σ*For test sample
Classification.
Claims (5)
1. a kind of three-dimensional correction and weighting similarity measurement study are without constraint face verification method, comprising the following steps:
1) VJ algorithm is combined with Dlib feature detection algorithm, target facial feature points detection is carried out to plurality of human faces image；
2) three-dimensional face correction is carried out using the human face characteristic point that Dlib is detected, and the face after correction is cut；
3) feature extraction (Gabor) and dimensionality reduction (PCA) are carried out to the face after correction, and face characteristic adds to treated
Similarity measurement study is weighed, final verification result is obtained.
2. specifically, the progress target facial feature points detection that VJ algorithm is combined with Dlib feature detection algorithm is nothing
It is possible to will appear other faces other than goal task out in constraint image, in order to be corrected to correct target face,
Need to carry out the selection of target face.It under normal conditions, in the picture should be in occupation of as the face of photographer's target person
Bigger space, therefore in the case where an image has multiple faces, the face of target person is relative to other in photo
It the space accounted for for face should be bigger.The face frame that the Dlib algorithm used herein is positioned does not consider facial image
In face size issue, be not suitable for the screening operation for carrying out target face, therefore detected and retained in image with Dlib and occurred
All faces characteristic point, recycle VJ algorithm to carry out Face detection, which can provide the specific location of face in image
(face frame) calculates face frame size and is ranked up, select the maximum face of area of corresponding face frame as the image
Target face, and retain characteristic point detected by the target face.
3. specifically, the three-dimensional face antidote is, using a single, constant 3D faceform, first thrown with camera
The principle of shadow matrix generates the 2D front face image (reference picture) of the 3D face, and 3D mould is saved when generating reference picture
The corresponding relationship between each point between type and 2D image.The people of the reference picture and input picture is detected with Dlib algorithm
Face characteristic point, by the human face characteristic point and 3D model of two images as intermediary, can find input picture and 3D model and
Corresponding relationship between its 2D reference picture each point.Using the corresponding relationship, the human face region in the facial image of input is thrown
After shadow is into the conventional coordinates (i.e. reference picture) generated by 3D faceform, after recycling pixel sampling to generate a correction
Front face image.
4. specifically, the face relevant range for cutting the facial image after correction is, complicated background in no constraint image
Problem will affect verifying accuracy rate, and after correction, the face of all images can all appear in the central area of image, utilize this
One characteristic can carry out the cutting of face relevant range, to remove complicated image background.
5. specifically, the method for the weighting similarity measurement study is to complete feature extraction (Gabor) and dimensionality reduction
(PCA) after, sample is divided into training sample and test sample, learns optimal similar matrix G, distance out using training sample
Matrix M and classification thresholds σ.By similar matrix G and distance matrix M, the power of similar matrix G Yu distance matrix M can be calculated
Weight values.The two weighted combination is worked as into similarity scores in the similarity scores that two test images can be obtained according to weighted value
When more than or equal to threshold value σ, judge that two images are similar (being the same person), is otherwise judged as dissimilar.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811353070.XA CN109558814A (en) | 2018-11-14 | 2018-11-14 | A kind of three-dimensional correction and weighting similarity measurement study without constraint face verification method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811353070.XA CN109558814A (en) | 2018-11-14 | 2018-11-14 | A kind of three-dimensional correction and weighting similarity measurement study without constraint face verification method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109558814A true CN109558814A (en) | 2019-04-02 |
Family
ID=65866352
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811353070.XA Pending CN109558814A (en) | 2018-11-14 | 2018-11-14 | A kind of three-dimensional correction and weighting similarity measurement study without constraint face verification method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109558814A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110163369A (en) * | 2019-05-21 | 2019-08-23 | 北京迈格威科技有限公司 | Image recognition and the training method of neural network model, device and system |
| CN110211679A (en) * | 2019-05-24 | 2019-09-06 | 山东海博科技信息***股份有限公司 | A kind of self-service examination machine eyesight detection intelligent processing method and device |
| CN111553253A (en) * | 2020-04-26 | 2020-08-18 | 上海天诚比集科技有限公司 | Standard face image selection method based on Euclidean distance variance algorithm |
| CN111967499A (en) * | 2020-07-21 | 2020-11-20 | 电子科技大学 | Data dimension reduction method based on self-learning |
| CN113139465A (en) * | 2021-04-23 | 2021-07-20 | 北京华捷艾米科技有限公司 | Face recognition method and device |
Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104123560A (en) * | 2014-07-03 | 2014-10-29 | 中山大学 | Phase encoding characteristic and multi-metric learning based vague facial image verification method |
-
2018
- 2018-11-14 CN CN201811353070.XA patent/CN109558814A/en active Pending
Patent Citations (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104123560A (en) * | 2014-07-03 | 2014-10-29 | 中山大学 | Phase encoding characteristic and multi-metric learning based vague facial image verification method |
Non-Patent Citations (1)
| Title |
|---|
| 徐昕等: "基于三维矫正和相似性学习的无约束人脸验证", 《计算机应用》 * |
Cited By (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110163369A (en) * | 2019-05-21 | 2019-08-23 | 北京迈格威科技有限公司 | Image recognition and the training method of neural network model, device and system |
| CN110211679A (en) * | 2019-05-24 | 2019-09-06 | 山东海博科技信息***股份有限公司 | A kind of self-service examination machine eyesight detection intelligent processing method and device |
| CN111553253A (en) * | 2020-04-26 | 2020-08-18 | 上海天诚比集科技有限公司 | Standard face image selection method based on Euclidean distance variance algorithm |
| CN111967499A (en) * | 2020-07-21 | 2020-11-20 | 电子科技大学 | Data dimension reduction method based on self-learning |
| CN111967499B (en) * | 2020-07-21 | 2023-04-07 | 电子科技大学 | Data dimension reduction method based on self-learning |
| CN113139465A (en) * | 2021-04-23 | 2021-07-20 | 北京华捷艾米科技有限公司 | Face recognition method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| JP7482181B2 (en) | Image processing device and image processing method | |
| CN110070074B (en) | Method for constructing pedestrian detection model | |
| CN108549873B (en) | Three-dimensional face recognition method and three-dimensional face recognition system | |
| CN109800648B (en) | Face detection and recognition method and device based on face key point correction | |
| CN110532920B (en) | Face recognition method for small-quantity data set based on FaceNet method | |
| CN109558814A (en) | A kind of three-dimensional correction and weighting similarity measurement study without constraint face verification method | |
| CN108921107B (en) | Pedestrian re-identification method based on sequencing loss and Simese network | |
| CN108520226B (en) | Pedestrian re-identification method based on body decomposition and significance detection | |
| CN110879982B (en) | Crowd counting system and method | |
| CN110674748A (en) | Image data processing method, image data processing device, computer equipment and readable storage medium | |
| JP2017506379A5 (en) | ||
| CN106407911A (en) | Image-based eyeglass recognition method and device | |
| CN106023257A (en) | Target tracking method based on rotor UAV platform | |
| CN105976397B (en) | A kind of method for tracking target | |
| CN108960142B (en) | Pedestrian re-identification method based on global feature loss function | |
| CN104240264A (en) | Height detection method and device for moving object | |
| CN112199983B (en) | Long-time large-range pedestrian re-identification method based on multi-level screening | |
| CN112784782B (en) | Three-dimensional object identification method based on multi-view double-attention network | |
| CN111914643A (en) | Human body action recognition method based on skeleton key point detection | |
| JP6417664B2 (en) | Person attribute estimation device, person attribute estimation method and program | |
| Zhang et al. | Second-and high-order graph matching for correspondence problems | |
| Yu et al. | Sparse 3D directional vertices vs continuous 3D curves: efficient 3D surface matching and its application for single model face recognition | |
| CN116958584A (en) | Key point detection method, regression model training method and device and electronic equipment | |
| CN110751005B (en) | Pedestrian detection method integrating depth perception features and kernel extreme learning machine | |
| CN112329662B (en) | Multi-view saliency estimation method based on unsupervised learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20190402 |
|
| WD01 | Invention patent application deemed withdrawn after publication |