CN109117744A - A kind of twin neural network training method for face verification - Google Patents
A kind of twin neural network training method for face verification Download PDFInfo
- Publication number
- CN109117744A CN109117744A CN201810809219.4A CN201810809219A CN109117744A CN 109117744 A CN109117744 A CN 109117744A CN 201810809219 A CN201810809219 A CN 201810809219A CN 109117744 A CN109117744 A CN 109117744A
- Authority
- CN
- China
- Prior art keywords
- data
- neural network
- sample
- training
- face verification
- 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
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Molecular Biology (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
A kind of twin neural network training method for face verification, comprising: prepare training sample set；Artificial neural network is input to after the picture progress size normalized for concentrating training sample to be trained；Artificial neural network includes two identical sub-neural networks；By treated, training sample is divided into equal the data set data and data_p of quantity, and data set data and data_p are conveyed to two sub-neural networks respectively, carry out sampling feature vectors extraction；By then jumping out iteration until the number of iterations reaches setting value to the iteration optimization for comparing loss function realization neural network, trained artificial neural network is the twin neural network for being used for face verification at this time；The similarity of two groups of feature vectors of loss of contrast function representation.
Description
Technical field
The present invention relates to computer vision correlative technology fields, and in particular to a kind of twin nerve net for face verification
Network training method.
Background technique
Face recognition technology is to pass through computer as the identification based on physiological characteristic a kind of in living things feature recognition field
Face characteristic is extracted, and carries out a kind of technology of authentication according to these features.Compared to other biological identification technologies, face
Identification is a kind of non-contacting identification technology, have quickly, it is easy, accurately and reliably, the advantages that cost performance is high, favorable expandability.Mesh
Before be widely used in the fields such as file administration, security security protection, mobile payment.
But the accuracy rate of recognition of face is subjected to the influence of illumination, four posture, expression and change of age factors, so that base
Series of challenges is still suffered from the recognition of face of image.Wherein change of age is one and continues and complicated process, as shape,
The colour of skin and wrinkle etc.；Meanwhile based on the research across the face identification method under the conditions of the age compared to other factors attention rate compared with
Few, the demand with every field to recognition of face is higher and higher, proposes a kind of efficient face under change of age
Recognition methods is particularly significant.
Face verification is to carry out the technology of identity validation based on facial feature information of people, friendly with it with safety
The domain that is widely used in authentication such as property and reliability.The important content of face verification research is how to obtain effectively
Feature representation, maximum between class distance and minimize inter- object distance.Using traditional Gabor, the artificial selected characteristic such as LBP is insufficient
To capture the substantive characteristics of face, the face verification of pinpoint accuracy is realized.Deep learning method is successfully applied to people in recent years
In face verifying, deep learning method is obtained by the training to extensive truthful data rich in face identity attribute information, real
The essence expression of existing facial image, improves face verification precision.It such as include DeepFace, the face verifications such as DeepID, FaceNet
Special purpose model, the identification that verifying accuracy has reached human eye are horizontal.However the problem of deep learning method maximum is network ginseng
Number is excessive, needs large-scale data mark to be just able to achieve training, often required data make up to million or more, such as DeepFace
With the 400000000 width images of 4000 people.And carry out extensive human face data acquisition and mark is a kind of biggish manpower financial capacity
Consumption.
Summary of the invention
It is big it is an object of the invention to solve above-mentioned extensive human face data acquisition of the existing technology and mark difficulty
The problem of, a kind of twin neural network training method for face verification is provided, can be obtained using less sample data
To the higher identification model of recognition accuracy.
The technical solution adopted by the present invention to solve the technical problems is: a kind of twin neural network for face verification
Training method, comprising: prepare training sample set；People is input to after carrying out size normalized to the picture that training sample is concentrated
Artificial neural networks are trained；Artificial neural network includes two identical sub-neural networks；By treated, training sample divides
It is segmented into quantity equal data set data and data_p, data set data and data_p are conveyed to two sub- nerve nets respectively
Network carries out sampling feature vectors extraction；By realizing the iteration optimization of neural network to comparison loss function, until the number of iterations
Reach setting value, then jump out iteration, trained artificial neural network is the twin nerve net for being used for face verification at this time
Network；The similarity of two groups of feature vectors of loss of contrast function representation.
Further, the loss of contrast function L:
WhereinThe COS distance of two sample characteristics is represented, y is whether matched two samples are true
Label, y=1 represent two training samples as the same person, and it is not the same person, margin that y=0, which then represents two training samples,
For the threshold value of setting；
As y=1, loss function is
As y=0, loss function is
Further, it is corresponded on the picture list order of the data set data and data_p；
The first picture B collectively forms training net in the first picture A and data set data_p in the data set data
First sample of network is to input；If facial image represented by picture A and B come from the same person, the sample pair it is true
Label is 1, as a positive sample pair；If not the same person, then true tag is 0, an as negative sample pair.
Further, the positive negative sample comparative example of the data set data and data_p is 1: 1.
Further, the sub-neural network is ResNet-50 neural network structure, respectively by data set data and
Data_p is inputted as the data Layer of ResNet-50 neural network, carries out characteristic vector pickup respectively.
Substantial effect of the invention: the invention proposes a kind of sample process strategies that binary channels inputs parallel, so that
The treatment process of data Layer is easier；
1. the effect of feature extraction is obviously improved: in terms of the more complicated data set of reply pictorial information, selecting
Deeper and efficient ResNet-50 neural network realize feature extraction, so that the effect of feature extraction be made obviously to be mentioned
It rises, so that the higher identification model of recognition accuracy can be obtained using less sample data, to greatly reduce sample
The workload of acquisition；
2. verify accuracy rate improves to further: selecting cosine phase for the similarity calculation of two pictures in sorting phase
Like degree, COS distance more focuses on difference of two vectors on direction, selects cosine similarity good classification effect.
Detailed description of the invention
Fig. 1 is that the present invention carries out the test flow chart across age face verification.
Fig. 2 is the training process schematic diagram of " twin network " proposed by the present invention.
Fig. 3 is across the age human face data collection of CACD needed for the present invention carries out network training.
Fig. 4 be present invention introduces ResNet-50 neural network structure schematic diagram.
Fig. 5 is the network structure of Conv block in ResNet-50 network.
Fig. 6 is the network structure of ID block in ResNet-50 network.
Specific embodiment
Below by specific embodiment, and in conjunction with attached drawing, technical scheme of the present invention will be further explained in detail.
A kind of twin neural network training method for face verification, comprising: prepare training sample set；To training sample
Artificial neural network is input to after the picture progress size normalized of concentration to be trained；Artificial neural network includes two
Identical sub-neural network；By treated, training sample is divided into equal the data set data and data_p of quantity, by data
Collection data and data_p is conveyed to two sub-neural networks respectively, carries out sampling feature vectors extraction；By losing letter to comparison
Number realizes the iteration optimization of neural network, until the number of iterations reaches setting value, then jumps out iteration, trained at this time artificial
Neural network is to be used for the twin neural network of face verification；The phase of two groups of feature vectors of loss of contrast function representation
Like degree.Its training process is specific as follows:
The accuracy rate of recognition of face is subjected to the influence of illumination, four posture, expression and change of age factors, so that being based on
The recognition of face of image still suffers from series of challenges.Wherein change of age is a lasting and complicated process, such as shape, skin
Color and wrinkle etc., the present embodiment are tested for age factor using the present invention.
One, the sample preparation stage.
Select across age human face data collection CACD2000 as training sample, which is had chosen by internet
2000 public figures amount to 163446 facial images, and the range of age of personage is between 16-62 years old, everyone sample
Quantity is differed from dozens to hundreds of, and quantity variance is larger.Preferably to play effect of the neural network to image procossing, manually
1600 class face samples are had chosen, every one kind (i.e. everyone) is unified to choose 70 samples pictures.Using MTCNN network to screening
Cutting out for good sample set progress face alignment and human face region, ultimately generates really for trained across the age face of 1600 classes
Training set of images.
Two, data Layer processing stage.
For the step of simplifying training twin network, the method that the present invention uses binary channels to input parallel, i.e., by step (1)
The middle CACD1600 class samples pictures screened are divided into equal the data set data and data_p of quantity, then by two data
Collection is conveyed to two identical neural networks respectively.Simultaneously because twin network is according to sample in training to pairs of input
, this requires corresponded on the picture list order of each data set when generating data set data and data_p.Specifically
Realize that steps are as follows:
1. the CACD1600 class samples pictures screened according to step 1 generate the list of file names text under the sample set
Part.
2. the list of file names in step 1 is divided into two, one is used to generate data set data, another is for generating
Data set data_p.Picture number is equal in two datasets, and corresponds on the picture list order of each data set.In full
First sample of trained network is together constituted according to the first picture B in the first picture A and data set data_p in collection data
This is to input.If facial image represented by picture A and B comes from the same person, the true tag of the sample pair is 1, as
One positive sample pair；If not the same person, then true tag is 0, an as negative sample pair.To guarantee training effect,
When generating two datasets, positive negative sample comparative example is designed as 1: 1.According to above-mentioned sample distribution principle, data is finally obtained
With the list of file names (including respective class label) of data_p two datasets, their quantity are equal and interrelated, altogether
With constituting in network structure sample required for data Layer to image data.
3. moreover, it is desirable to carry out resize operation to each picture to realize that size normalizes.The present invention utilizes
ResNet-50 neural network carries out feature extraction, therefore the primary system one of resize is set as 224*224.
4. the CACD1600 class samples pictures as described in step 1 are as data source, according to step 2 data generated and
Data_p listed files is allocated, while being handled according to size specified by step 3, is ultimately generated and is met training net
Two database files of data and data_p that network requires.
Three, feature extraction phases.
Novel twin network proposed by the present invention has selected ResNet-50 residual error network to carry out feature extraction.Based on twin
The thought of network frame constructs two the same ResNet-50 network structures, by described in step 2 to training data
The data Layer for collecting data and data_p respectively as two ResNet-50 neural networks inputs, then using the network respectively into
Row feature extraction.
Four, about the design of loss function.
The present invention selects loss of contrast function, selects COS distance specifically to indicate the similar of two feature vectors
Degree, it is specific as shown in formula 1.
WhereinThe COS distance of two sample characteristics is represented, y is whether matched two samples are true
Label, y=1 represent two training samples as the same person, and it is not the same person that y=0, which is then represented, and margin is the threshold of setting
Value.Specifically, as y=1, the value of cosine similarity should be increased as far as possible, thus the optimization of implementation model, therefore damage at this time
Losing function isOn the contrary as y=0, cosine similarity should be reduced as far as possible
Value, thus the optimization of implementation model, therefore loss function at this time is
So far the twin neural network framework basic building proposed by the invention for realizing face verification is completed.To instruction
Practice parameter constantly to be debugged, then carry out network training and verify accuracy rate, finally obtains ideal neural network
Model.
Existing picture A and B to be measured carries out the operating process of face verification as schemed using the trained neural network of the present invention
Shown in 1:
(1) picture A and B to be verified image pretreatment operation: is inputed into trained classics MTCNN network mould respectively
Type, the network can rapidly and accurately detected the human face region in picture to be detected；Simultaneously respectively in facial image
Right and left eyes, five characteristic points of nose and two corners of the mouths are labeled, and carry out face alignment using these characteristic points.After being aligned
Human face region be cut out, to complete the pretreatment work of image.
(2) picture A ' and B ' size after MTCNN network processes is normalized to the 224*224 of network model requirement,
So that neural network carries out further feature extraction to it.
(3) by the image after size normalized to being conveyed to the trained NewSiamese neural network of the present invention,
Feature extraction is carried out to them respectively and calculates the similarity of image pair.
(4) output of NewSiamese neural network indicates classification prediction result of the former picture A to be measured with B whether similar.
If output is 0, then it represents that original picture A and B dissmilarity, i.e. picture A and B are indicated is not the same person；If output is 1, table
Show that original picture A is similar with B, i.e. that picture A and B are indicated is the same person.
It about NewSiamese neural network mentioned above, trains process as shown in Figure 2: obtaining CACD2000 across year
Age human face data collection, and carry out artificial bulk processing.Specially everyone selects the sample of fixed number, retains as far as possible same
Typical sample of the individual under all ages and classes, while weeding out the error sample that picture personage is not consistent with label.For certain
Sample size is insufficient, and mirror transformation and scale transformation is mainly taken to carry out data amplification.
CACD data set after simple process is inputed into MTCNN neural network by the gross, by the MTCNN net of open source
Network modeling tool realizes the further pretreatment work to training sample, concentrates each picture to available data to realize
Face datection, face alignment and human face region cut out, final output ideal across age facial image training set.
For the data Layer input condition for meeting Siamese network, need to carry out positive and negative sample to the training set that step (2) generate
This pair of production.List of file names corresponding to the training set of step (2) generation is divided into two, data set data is respectively indicated
And data_p.Picture number in data set data and data_p is consistent, and in two data sets each picture sequence
It is stringent to correspond.Two pictures of same position form a sample pair in two data sets, and when the respective sample of two pictures
When this label is identical, the true tag of sample pair is 1；When the respective sample label difference of two pictures, sample pair it is true
Label is 0.To guarantee training effect, positive negative sample comparative example is 1: 1.It generates by above-mentioned requirements for constituting sample to data set
Positive and negative sample list file data and data_p.
To meet training network structure, the size of sample is uniformly normalized to 224*224, generated in conjunction with step (3)
Sample list file, generating can be by efficient data library file data and data_p that Neural Network Data layer is directly read.It will count
It is conveyed to ResNet-50 neural network respectively according to file data and data_p, the overall network structure of ResNet-50 is detailed in Fig. 4.
Feature vector of the layer AVG Pool second from the bottom of ResNet-50 network as training sample is taken, after being conveyed to present invention improvement
Cosine_Similarity_Contrastive Loss.Loss layers by obtain training sample to respective feature vector and
True tag corresponding to them carries out the calculating of cosine similarity, and continues to optimize network model according to the threshold value of setting.Most
Eventually by repeatedly debugging training parameter, ideal be used for across age face verification is obtained after constantly repetitive exercise test
Twin network model.
Fig. 3 is that the part of above-mentioned CACD2000 raw data set is shown, each column represent a kind of sample, and the longitudinal axis represents the sample
This year of birth, horizontal axis represent the time that sample is chosen under particular year.
For above-mentioned ResNet-50 neural network, schematic network structure is as shown in Figure 4:
(1) it is conveyed to convolutional layer CONV as the input of network using the image data collection of 224*224 specification and starts to be rolled up
Product operation.
(2) output of convolutional layer is as Norm layers of Batch of input, and Norm layers of Batch for carrying out normalizing to Batch
Change operation, to solve influence of this data distribution variation to subsequent study.
(3) Norm layers of Batch of output phase is operated after progress ReLU nonlinear transformation and maximum value pondization, thus complete
At the feature learning of first stage.
(4) POOL layers of MAX of output is conveyed to Conv BLOCK convolution block, starts the feature learning of second stage.
Conv BLOCK convolution block is a multilayered structure, is detailed in Fig. 5.The output phase of Conv BLOCK convolution block is the same after being conveyed to two
ID BLOCK convolution block, similar with Conv BLOCK convolution block, ID BLOCK is also a multilayered structure, but its output does not change
The dimension for becoming feature vector, is detailed in Fig. 6.So far the feature learning of second stage is completed.
(5) third and fourth, five stages be by Conv BLOCK series connection ID BLOCK convolution block constitute, unlike third,
Four, the number of five stage ID BLOCK convolution blocks is respectively 3,5 and 2.So far the feature learning in first five stage is completed
Task.
(6) output of the last one ID BLOCK convolution block is conveyed to average value pond layer, the output that Pool layers of AVG
It can be used as the feature vector of the extractor characterized by ResNet-50 network.It can also be added if necessary to carry out classification task
One layer of full articulamentum of FC.
Above-mentioned Conv BLOCK and ID BLOCK convolution block are typical residual error network structure blocks, network structure such as Fig. 5,
Shown in Fig. 6, with conventional convolution, pond operating process is different, parallel link, this new net occurs in residual error network structure
Caused by network structure very good solution is constantly deepened due to network structure " degenerate problem ", while also greatly accelerating training
Time.
In conclusion the present invention combines the preferable ResNet-50 neural network of comprehensive performance instantly, proposes one kind and be used for
The twin neural network training method of face verification, this method have one in terms of the building of neural network model and training method
Fixed innovation, while this method operation possibility is strong and verifying accuracy rate is high.
Embodiment described above is a kind of preferable scheme of the invention, not makees limit in any form to the present invention
System, there are also other variants and remodeling on the premise of not exceeding the technical scheme recorded in the claims.
Claims (5)
1. a kind of twin neural network training method for face verification characterized by comprising
Prepare training sample set；
Artificial neural network is input to after the picture progress size normalized for concentrating training sample to be trained；Artificial mind
It include two identical sub-neural networks through network；
By treated, training sample is divided into equal the data set data and data_p of quantity, by data set data and data_p
It is conveyed to two sub-neural networks respectively, carries out sampling feature vectors extraction；By realizing neural network to comparison loss function
Iteration optimization then jump out iteration until the number of iterations reaches setting value, trained artificial neural network is to use at this time
In the twin neural network of face verification；
The similarity of two groups of feature vectors of loss of contrast function representation.
2. a kind of twin neural network training method for face verification as described in claim 1, which is characterized in that described
Loss of contrast function L
WhereinThe COS distance of two sample characteristics is represented, y is the whether matched true tag of two samples,
Y=1 represents two training samples as the same person, and it is not the same person that y=0, which then represents two training samples, and margin is to set
Fixed threshold value；
As y=1, loss function isAs y=0, loss function is
3. a kind of twin neural network training method for face verification as described in claim 1, which is characterized in that
It is corresponded on the picture list order of the data set data and data_p；
The first picture B collectively forms trained network in the first picture A and data set data_p in the data set data
First sample is to input；If facial image represented by picture A and B comes from the same person, the true tag of the sample pair
It is 1, an as positive sample pair；If not the same person, then true tag is 0, an as negative sample pair.
4. a kind of twin neural network training method for face verification as claimed in claim 3, which is characterized in that
The positive negative sample comparative example of the data set data and data_p is 1:1.
5. a kind of twin neural network training method for face verification as described in claim 1, which is characterized in that
The sub-neural network is ResNet-50 neural network structure, respectively using data set data and data_p as
The data Layer of ResNet-50 neural network inputs, and carries out characteristic vector pickup respectively.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810809219.4A CN109117744A (en) | 2018-07-20 | 2018-07-20 | A kind of twin neural network training method for face verification |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810809219.4A CN109117744A (en) | 2018-07-20 | 2018-07-20 | A kind of twin neural network training method for face verification |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109117744A true CN109117744A (en) | 2019-01-01 |
Family
ID=64863330
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810809219.4A Pending CN109117744A (en) | 2018-07-20 | 2018-07-20 | A kind of twin neural network training method for face verification |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109117744A (en) |
Cited By (28)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109840588A (en) * | 2019-01-04 | 2019-06-04 | 平安科技（深圳）有限公司 | Neural network model training method, device, computer equipment and storage medium |
| CN109919073A (en) * | 2019-03-01 | 2019-06-21 | 中山大学 | A kind of recognition methods again of the pedestrian with illumination robustness |
| CN109977887A (en) * | 2019-01-30 | 2019-07-05 | 四川电科维云信息技术有限公司 | A kind of face identification method of anti-age interference |
| CN110222792A (en) * | 2019-06-20 | 2019-09-10 | 杭州电子科技大学 | A kind of label defects detection algorithm based on twin network |
| CN110309359A (en) * | 2019-05-20 | 2019-10-08 | 北京大学 | Video correlation prediction technique, device, equipment and storage medium |
| CN110365639A (en) * | 2019-05-29 | 2019-10-22 | 中国科学院信息工程研究所 | A kind of malicious traffic stream detection method and system based on depth residual error network |
| CN110414349A (en) * | 2019-06-26 | 2019-11-05 | 长安大学 | Introduce the twin convolutional neural networks face recognition algorithms of sensor model |
| CN110443162A (en) * | 2019-07-19 | 2019-11-12 | 南京邮电大学 | A kind of two-part training method for disguised face identification |
| CN110610230A (en) * | 2019-07-31 | 2019-12-24 | 国家计算机网络与信息安全管理中心 | Station caption detection method and device and readable storage medium |
| CN110781751A (en) * | 2019-09-27 | 2020-02-11 | 杭州电子科技大学 | Emotional electroencephalogram signal classification method based on cross-connection convolutional neural network |
| CN111079785A (en) * | 2019-11-11 | 2020-04-28 | 深圳云天励飞技术有限公司 | Image identification method and device and terminal equipment |
| CN111091144A (en) * | 2019-11-27 | 2020-05-01 | 云南电网有限责任公司电力科学研究院 | Image feature point matching method and device based on depth pseudo-twin network |
| CN111178249A (en) * | 2019-12-27 | 2020-05-19 | 杭州艾芯智能科技有限公司 | Face comparison method and device, computer equipment and storage medium |
| CN111178285A (en) * | 2019-12-31 | 2020-05-19 | 上海昌投网络科技有限公司 | Natural scene user head portrait age prediction method based on deep learning framework |
| CN111274641A (en) * | 2020-01-21 | 2020-06-12 | 江苏艾佳家居用品有限公司 | Twin network-based house layout method and system |
| CN111275137A (en) * | 2020-03-26 | 2020-06-12 | 南京工业大学 | Tea true-checking method based on exclusive twin network model |
| CN111325736A (en) * | 2020-02-27 | 2020-06-23 | 成都航空职业技术学院 | Sight angle estimation method based on human eye difference image |
| CN111368920A (en) * | 2020-03-05 | 2020-07-03 | 中南大学 | Quantum twin neural network-based binary classification method and face recognition method thereof |
| CN111652170A (en) * | 2020-06-09 | 2020-09-11 | 电子科技大学 | Secondary radar signal processing method based on two-channel residual error deep neural network |
| CN111930989A (en) * | 2020-09-22 | 2020-11-13 | 杭州雄迈集成电路技术股份有限公司 | Method for searching picture by picture |
| CN111931758A (en) * | 2020-10-19 | 2020-11-13 | 北京圣点云信息技术有限公司 | Face recognition method and device combining facial veins |
| CN112699811A (en) * | 2020-12-31 | 2021-04-23 | 中国联合网络通信集团有限公司 | Living body detection method, apparatus, device, storage medium, and program product |
| CN112836658A (en) * | 2021-02-08 | 2021-05-25 | 南京林业大学 | Face recognition method based on transfer learning and sparse loss function |
| CN113017630A (en) * | 2021-03-02 | 2021-06-25 | 贵阳像树岭科技有限公司 | Visual perception emotion recognition method |
| JP2021517321A (en) * | 2019-01-15 | 2021-07-15 | 北京市商▲湯▼科技▲開▼▲発▼有限公司Ｂｅｉｊｉｎｇ Ｓｅｎｓｅｔｉｍｅ Ｔｅｃｈｎｏｌｏｇｙ Ｄｅｖｅｌｏｐｍｅｎｔ Ｃｏ．， Ｌｔｄ． | Network optimization methods and devices, image processing methods and devices, storage media and computer programs |
| CN113269903A (en) * | 2021-05-24 | 2021-08-17 | 上海应用技术大学 | Face recognition class attendance system |
| CN113283368A (en) * | 2021-06-08 | 2021-08-20 | 电子科技大学中山学院 | Model training method, face attribute analysis method, device and medium |
| CN115098787A (en) * | 2022-07-21 | 2022-09-23 | 西安电子科技大学 | Article recommendation method based on cosine ranking loss and virtual edge graph neural network |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107506717A (en) * | 2017-08-17 | 2017-12-22 | 南京东方网信网络科技有限公司 | Without the face identification method based on depth conversion study in constraint scene |
| CN107977609A (en) * | 2017-11-20 | 2018-05-01 | 华南理工大学 | A kind of finger vein identity verification method based on CNN |
| CN108171184A (en) * | 2018-01-03 | 2018-06-15 | 南京理工大学 | Method for distinguishing is known based on Siamese networks again for pedestrian |
| CN108198200A (en) * | 2018-01-26 | 2018-06-22 | 福州大学 | The online tracking of pedestrian is specified under across camera scene |
-
2018
- 2018-07-20 CN CN201810809219.4A patent/CN109117744A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107506717A (en) * | 2017-08-17 | 2017-12-22 | 南京东方网信网络科技有限公司 | Without the face identification method based on depth conversion study in constraint scene |
| CN107977609A (en) * | 2017-11-20 | 2018-05-01 | 华南理工大学 | A kind of finger vein identity verification method based on CNN |
| CN108171184A (en) * | 2018-01-03 | 2018-06-15 | 南京理工大学 | Method for distinguishing is known based on Siamese networks again for pedestrian |
| CN108198200A (en) * | 2018-01-26 | 2018-06-22 | 福州大学 | The online tracking of pedestrian is specified under across camera scene |
Non-Patent Citations (1)
| Title |
|---|
| M. KHALIL-HANI 等: ""A convolutional neural network approach for face verification"", 《2014 INTERNATIONAL CONFERENCE ON HIGH PERFORMANCE COMPUTING & SIMULATION (HPCS)》 * |
Cited By (43)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109840588A (en) * | 2019-01-04 | 2019-06-04 | 平安科技（深圳）有限公司 | Neural network model training method, device, computer equipment and storage medium |
| CN109840588B (en) * | 2019-01-04 | 2023-09-08 | 平安科技（深圳）有限公司 | Neural network model training method, device, computer equipment and storage medium |
| JP7074877B2 (en) | 2019-01-15 | 2022-05-24 | 北京市商▲湯▼科技▲開▼▲発▼有限公司 | Network optimization methods and equipment, image processing methods and equipment, storage media and computer programs |
| JP2021517321A (en) * | 2019-01-15 | 2021-07-15 | 北京市商▲湯▼科技▲開▼▲発▼有限公司Ｂｅｉｊｉｎｇ Ｓｅｎｓｅｔｉｍｅ Ｔｅｃｈｎｏｌｏｇｙ Ｄｅｖｅｌｏｐｍｅｎｔ Ｃｏ．， Ｌｔｄ． | Network optimization methods and devices, image processing methods and devices, storage media and computer programs |
| US11416703B2 (en) | 2019-01-15 | 2022-08-16 | Beijing Sensetime Technology Development Co., Ltd. | Network optimization method and apparatus, image processing method and apparatus, and storage medium |
| CN109977887A (en) * | 2019-01-30 | 2019-07-05 | 四川电科维云信息技术有限公司 | A kind of face identification method of anti-age interference |
| CN109919073A (en) * | 2019-03-01 | 2019-06-21 | 中山大学 | A kind of recognition methods again of the pedestrian with illumination robustness |
| CN109919073B (en) * | 2019-03-01 | 2021-04-06 | 中山大学 | Pedestrian re-identification method with illumination robustness |
| CN110309359A (en) * | 2019-05-20 | 2019-10-08 | 北京大学 | Video correlation prediction technique, device, equipment and storage medium |
| CN110309359B (en) * | 2019-05-20 | 2021-06-15 | 北京大学 | Video correlation prediction method, device, equipment and storage medium |
| CN110365639A (en) * | 2019-05-29 | 2019-10-22 | 中国科学院信息工程研究所 | A kind of malicious traffic stream detection method and system based on depth residual error network |
| CN110222792A (en) * | 2019-06-20 | 2019-09-10 | 杭州电子科技大学 | A kind of label defects detection algorithm based on twin network |
| CN110414349A (en) * | 2019-06-26 | 2019-11-05 | 长安大学 | Introduce the twin convolutional neural networks face recognition algorithms of sensor model |
| CN110443162B (en) * | 2019-07-19 | 2022-08-30 | 南京邮电大学 | Two-stage training method for disguised face recognition |
| CN110443162A (en) * | 2019-07-19 | 2019-11-12 | 南京邮电大学 | A kind of two-part training method for disguised face identification |
| CN110610230A (en) * | 2019-07-31 | 2019-12-24 | 国家计算机网络与信息安全管理中心 | Station caption detection method and device and readable storage medium |
| CN110781751A (en) * | 2019-09-27 | 2020-02-11 | 杭州电子科技大学 | Emotional electroencephalogram signal classification method based on cross-connection convolutional neural network |
| CN111079785A (en) * | 2019-11-11 | 2020-04-28 | 深圳云天励飞技术有限公司 | Image identification method and device and terminal equipment |
| CN111091144A (en) * | 2019-11-27 | 2020-05-01 | 云南电网有限责任公司电力科学研究院 | Image feature point matching method and device based on depth pseudo-twin network |
| CN111178249A (en) * | 2019-12-27 | 2020-05-19 | 杭州艾芯智能科技有限公司 | Face comparison method and device, computer equipment and storage medium |
| CN111178285A (en) * | 2019-12-31 | 2020-05-19 | 上海昌投网络科技有限公司 | Natural scene user head portrait age prediction method based on deep learning framework |
| CN111274641A (en) * | 2020-01-21 | 2020-06-12 | 江苏艾佳家居用品有限公司 | Twin network-based house layout method and system |
| CN111274641B (en) * | 2020-01-21 | 2022-07-08 | 江苏艾佳家居用品有限公司 | Twin network-based house layout method and system |
| CN111325736B (en) * | 2020-02-27 | 2024-02-27 | 成都航空职业技术学院 | Eye differential image-based sight angle estimation method |
| CN111325736A (en) * | 2020-02-27 | 2020-06-23 | 成都航空职业技术学院 | Sight angle estimation method based on human eye difference image |
| CN111368920B (en) * | 2020-03-05 | 2024-03-05 | 中南大学 | Quantum twin neural network-based classification method and face recognition method thereof |
| CN111368920A (en) * | 2020-03-05 | 2020-07-03 | 中南大学 | Quantum twin neural network-based binary classification method and face recognition method thereof |
| CN111275137B (en) * | 2020-03-26 | 2023-07-18 | 南京工业大学 | Tea verification method based on exclusive twin network model |
| CN111275137A (en) * | 2020-03-26 | 2020-06-12 | 南京工业大学 | Tea true-checking method based on exclusive twin network model |
| CN111652170A (en) * | 2020-06-09 | 2020-09-11 | 电子科技大学 | Secondary radar signal processing method based on two-channel residual error deep neural network |
| CN111930989A (en) * | 2020-09-22 | 2020-11-13 | 杭州雄迈集成电路技术股份有限公司 | Method for searching picture by picture |
| CN111931758A (en) * | 2020-10-19 | 2020-11-13 | 北京圣点云信息技术有限公司 | Face recognition method and device combining facial veins |
| CN111931758B (en) * | 2020-10-19 | 2021-01-05 | 北京圣点云信息技术有限公司 | Face recognition method and device combining facial veins |
| CN112699811B (en) * | 2020-12-31 | 2023-11-03 | 中国联合网络通信集团有限公司 | Living body detection method, living body detection device, living body detection apparatus, living body detection storage medium, and program product |
| CN112699811A (en) * | 2020-12-31 | 2021-04-23 | 中国联合网络通信集团有限公司 | Living body detection method, apparatus, device, storage medium, and program product |
| CN112836658B (en) * | 2021-02-08 | 2022-03-25 | 南京林业大学 | Face recognition method based on transfer learning and sparse loss function |
| CN112836658A (en) * | 2021-02-08 | 2021-05-25 | 南京林业大学 | Face recognition method based on transfer learning and sparse loss function |
| CN113017630A (en) * | 2021-03-02 | 2021-06-25 | 贵阳像树岭科技有限公司 | Visual perception emotion recognition method |
| CN113269903A (en) * | 2021-05-24 | 2021-08-17 | 上海应用技术大学 | Face recognition class attendance system |
| CN113283368A (en) * | 2021-06-08 | 2021-08-20 | 电子科技大学中山学院 | Model training method, face attribute analysis method, device and medium |
| CN113283368B (en) * | 2021-06-08 | 2023-10-20 | 电子科技大学中山学院 | Model training method, face attribute analysis method, device and medium |
| CN115098787A (en) * | 2022-07-21 | 2022-09-23 | 西安电子科技大学 | Article recommendation method based on cosine ranking loss and virtual edge graph neural network |
| CN115098787B (en) * | 2022-07-21 | 2024-04-19 | 西安电子科技大学 | Article recommendation method based on cosine ranking loss and virtual edge map neural network |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109117744A (en) | A kind of twin neural network training method for face verification | |
| CN104866829B (en) | A kind of across age face verification method based on feature learning | |
| Yuan et al. | Fingerprint liveness detection using an improved CNN with image scale equalization | |
| Yang et al. | Learning face age progression: A pyramid architecture of gans | |
| CN104239858B (en) | A kind of method and apparatus of face characteristic checking | |
| CN106650806B (en) | A kind of cooperating type depth net model methodology for pedestrian detection | |
| CN103258204B (en) | A kind of automatic micro-expression recognition method based on Gabor and EOH feature | |
| CN109325443A (en) | A kind of face character recognition methods based on the study of more example multi-tag depth migrations | |
| CN110443189A (en) | Face character recognition methods based on multitask multi-tag study convolutional neural networks | |
| CN102136024B (en) | Biometric feature identification performance assessment and diagnosis optimizing system | |
| CN105138954A (en) | Image automatic screening, query and identification system | |
| CN109497990B (en) | Electrocardiosignal identity recognition method and system based on canonical correlation analysis | |
| CN110889672A (en) | Student card punching and class taking state detection system based on deep learning | |
| CN106022241B (en) | A kind of face identification method based on wavelet transformation and rarefaction representation | |
| CN108288048B (en) | Facial emotion recognition feature selection method based on improved brainstorming optimization algorithm | |
| CN109977887A (en) | A kind of face identification method of anti-age interference | |
| CN107992783A (en) | Face image processing process and device | |
| CN103984954B (en) | Image combining method based on multi-feature fusion | |
| CN112668486A (en) | Method, device and carrier for identifying facial expressions of pre-activated residual depth separable convolutional network | |
| Dantcheva et al. | Expression recognition for severely demented patients in music reminiscence-therapy | |
| Agbo-Ajala et al. | A lightweight convolutional neural network for real and apparent age estimation in unconstrained face images | |
| Tang et al. | Automatic facial expression analysis of students in teaching environments | |
| CN111368663A (en) | Method, device, medium and equipment for recognizing static facial expressions in natural scene | |
| Pei et al. | Convolutional neural networks for class attendance | |
| CN115719497A (en) | Student concentration degree identification method and system |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190101 |
|
| RJ01 | Rejection of invention patent application after publication |