CN108829900A - A kind of Research on face image retrieval based on deep learning, device and terminal - Google Patents
A kind of Research on face image retrieval based on deep learning, device and terminal Download PDFInfo
- Publication number
- CN108829900A CN108829900A CN201810856269.8A CN201810856269A CN108829900A CN 108829900 A CN108829900 A CN 108829900A CN 201810856269 A CN201810856269 A CN 201810856269A CN 108829900 A CN108829900 A CN 108829900A
- Authority
- CN
- China
- Prior art keywords
- feature vector
- facial image
- attribute
- comparing result
- face
- 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.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; 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
Landscapes
- Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- General Health & Medical Sciences (AREA)
- Human Computer Interaction (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
- Collating Specific Patterns (AREA)
Abstract
The embodiment of the present application discloses a kind of Research on face image retrieval based on deep learning, device and terminal, the method includes：By the facial image to be retrieved input training completion Jing Guo pretreatment operation based in locally shared multitask convolutional neural networks model, the multiple set of properties feature vectors of an identity characteristic vector sum of the facial image to be retrieved are obtained；Identity characteristic vector sum set of properties feature vector by the identity characteristic vector sum set of properties feature vector respectively with the facial image stored in database compares, and obtains identity characteristic vector comparing result, global property group feature vector comparing result and local attribute's feature vector comparing result；According to the identity characteristic vector comparing result, the comparing result and local attribute's feature vector comparing result of global property group feature vector, target facial image is filtered out in the database.Using technical solution provided by the embodiment of the present application, the probability of facial image retrieval misrecognition can effectively reduce.
Description
Technical field
This application involves technical field of image processing, more particularly to a kind of facial image retrieval side based on deep learning
Method, device and terminal.
Background technique
With internet, the fast development of social media, intelligent multimedia equipment, using image and video as more matchmakers of representative
Generation, processing and the acquisition of volume data become increasingly to facilitate, and multimedia application is increasingly extensive, and data volume shows volatile
Increase.How in the image big data of magnanimity interested image to be correctly found with lesser expense, it has also become in recent years
The important research hot spot of multimedia and information retrieval field.
Deep learning is a kind of method based on to data progress representative learning in machine learning, can be from big data
The expression of automatic learning characteristic.In machine learning, convolutional neural networks (Convolutional Neural Network,
It CNN is) a kind of depth feed forward-fuzzy control, with powerful learning ability, efficient feature representation ability, Ke Yicong
Pixel-level initial data successively extracts information to abstract semantic concept, makes it in the global characteristics and context letter for extracting image
Breath aspect has advantage outstanding, is widely applied to every field.
Recognition of face has good as biometrics identification technology due to the feature that its is untouchable and acquisition is convenient
Good development and application prospect, has all played highly important effect in many application scenarios.Existed with face recognition technology
For application in public security system application, some suspects are largely flowed using today's society population and identity card can be pseudo-
It the complex situations such as makes, the true identity of oneself is carried out lying about concealment, is acted as fraudulent substitute for a person, bring phase to public security system detection of handling a case
When big difficulty.And personage's looks are difficult to fake, it in the case can be by the face picture of suspect and face resource
Identification is compared in emphasis population in library, helps the true identity for confirming suspect, improves case handling efficiency.But face figure
As being different from common natural image, feature ga s safety degree relative difficult.Under unconstrained condition, due to illumination, posture, table
The interference of many factors such as feelings, accessories, or even the appearance variation generated due to age range, so that when algorithm comparison face very
The different images similarity for being easy to appear same face is lower, and the different higher problems of face similarity.These phenomenons
More huge difficulty is caused to face retrieval.In addition, with the increase of face database scale, wherein there is similar face
A possibility that also become larger therewith, the probability misidentified occur can greatly increase.
For the problem, a kind of scheme in the prior art is while to carry out face by multitask convolutional neural networks
The extraction of identity information and the classification of several face characters.In this way, traditional based on face identity information sequencing of similarity
On the basis of, with reordering for auxiliary, so as to improve the people of simple identity information for the similarity score of face character classification results
It is misidentified caused by face distinction is undesirable.
But the imaging complexity interfered under the diversity and unconstrained condition due to face itself, at present face category
The still not enough robusts of classification results of property.When the classification results of face character are wrong, may not have to the improvement of face verification
Good effect.In addition, face character in the prior art generallys use the age directly related with face identity, gender, kind
The global properties such as race the case where for face partial occlusion existing under actual environment or other local interference (accessories etc.), do not have
There is good improvement.Therefore, a kind of more preferably Research on face image retrieval urgently occurs.
Summary of the invention
A kind of Research on face image retrieval based on deep learning, device and terminal are provided in the embodiment of the present application, with
It is larger conducive to the probability for solving the problems, such as facial image retrieval misrecognition in the prior art.
In a first aspect, the embodiment of the present application provides a kind of Research on face image retrieval based on deep learning, including：
It is rolled up what the facial image to be retrieved input training Jing Guo pretreatment operation was completed based on locally shared multitask
In product neural network model, obtain the multiple set of properties features of an identity characteristic vector sum of the facial image to be retrieved to
It measures, includes a global property group feature vector and at least one local attribute's group feature in the multiple set of properties feature vector
Vector；
By the identity characteristic vector sum set of properties feature vector of the facial image to be retrieved respectively with stored in database
The identity characteristic vector sum set of properties feature vector of facial image compare, obtain the facial image to be retrieved and data
Identity characteristic vector comparing result, global property group feature vector comparing result and the local attribute of the facial image stored in library
Feature vector comparing result；
It is special according to the identity characteristic vector comparing result, the comparing result of global property group feature vector and local attribute
Vector comparing result is levied, filters out target facial image in the database.
Optionally, described according to the identity characteristic vector comparing result, the comparing result of global property group feature vector
With local attribute's feature vector comparing result, target facial image is filtered out in the database, including：
According to the identity characteristic vector comparing result and the global property group feature vector comparing result, in the number
According to filtering out candidate image set in library；
According to the identity characteristic vector comparing result and local attribute's feature vector comparing result, or according to institute
State identity characteristic vector comparing result, the comparing result of global property group feature vector and local attribute's feature vector comparison knot
Fruit filters out target facial image in the candidate image set.
Optionally, the identity characteristic vector comparing result, global property group feature vector comparing result drawn game subordinate
Property feature vector comparing result, the corresponding identity characteristic vector similarity of facial image stored in the respectively described database obtain
Point, global property group feature vector similarity score and local attribute's feature vector similarity score；
It is described according to the identity characteristic vector comparing result and the global property group feature vector comparing result, in institute
It states and filters out candidate image set in database, including：
The facial image for meeting the first screening conditions is filtered out in the database, obtains scalping face image set,
First screening conditions include that the identity characteristic vector similarity score of facial image is greater than or equal to preset identity characteristic
Vector similarity threshold value, and global property group feature vector similarity score be greater than or equal to preset global property group feature to
Measure similarity threshold；
The facial image for meeting the second screening conditions is filtered out in the scalping face image set, obtains candidate image
Set, second screening conditions are the first fusion maximum N1 of similarity score Score1 in the scalping face image set
A facial image, wherein Score1=(Score_id+Score_gAttrib)/2, Score_id is that identity characteristic vector is similar
Score is spent, Score_gAttrib is global property group feature vector similarity score, and N1≤N0, N0 are scalping face image set
The quantity of facial image in conjunction.
Optionally, according to the identity characteristic vector comparing result and local attribute's feature vector comparing result, or
Person is according to the identity characteristic vector comparing result, the comparing result and local attribute's feature vector of global property group feature vector
Comparing result filters out target facial image in the candidate image set, including：
The second maximum N2 facial image of fusion similarity score Score2 is filtered out in the candidate image set,
As target facial image, N2 >=1；
Wherein, Score2=w1*Score_id+w2*Score_loc, w1, w2 are weight, w1>0, w2>0, w1+w2=1；
If there is the office for meeting threshold value screening conditions at least one corresponding local attribute's group feature vector of facial image
Portion's set of properties feature vector, then by the local attribute's group feature vector for meeting threshold value screening conditions, local attribute's group is special
The maximum value of vector similarity score is levied as Score_loc；If at least one corresponding local attribute's group feature of facial image
There is no the local attribute's group feature vectors for meeting threshold value screening conditions in vector, then by the global property group feature vector phase
Like degree score as Score_loc, the threshold value screening conditions are that local attribute's group feature vector similarity score is greater than or waits
In preset local attribute's group feature vector similarity threshold.
Optionally, at least one described local attribute's group feature vector, including：
Upper face set of properties feature vector, middle face set of properties feature vector and lower face set of properties feature vector.
Optionally, the upper face set of properties feature vector includes eyebrow attribute, eyes attribute, color development for the attribute of characterization
Attribute, hair style attribute and/or upper face adjunct attribute；
Attribute of the middle face set of properties feature vector for characterization includes nose attribute, cheek attribute, cheekbone attribute, temples
Angle attribute and/or middle face adjunct attribute；
Attribute of the lower face set of properties feature vector for characterization includes lip attribute, chin attribute, beard attribute, mouth
Bar attribute and/or lower face adjunct attribute；
Attribute of the global property group feature vector for characterization includes gender attribute, expression attribute, shape of face attribute, face
Color attribute, hair style attribute and/or age attribute.
Specifically, the upper face set of properties feature vector for characterization attribute include heavy eyebrows, arch eyebrow, eye pouch, narrow eye,
Dark hair, white hair, fringe, bald, hair line move back, are branded as and/or wear glasses；
The middle face set of properties feature vector for the attribute of characterization include russian, sharp nose, cheek of blushing, high-malar,
It stays temples and/or wears earrings；
The lower face set of properties feature vector for characterization attribute include big lip, double chin, moustache, beard,
It partly opens one's mouth, have beard and/or use lipstick；
The global property group feature vector for characterization attribute include male, smile, charming, round face, oval face,
Heavy make-up, pale, curly hair and/or youth.
Optionally, the building process of the database includes：
By the facial image input training completion for needing to register based on locally shared multitask convolutional neural networks mould
In type, the multiple attributes of each corresponding identity characteristic vector sum of facial image in the facial image for needing to register are obtained
Group feature vector；
The corresponding multiple set of properties feature vectors of an identity characteristic vector sum of each facial image are stored in number
According in library.
Optionally, the pretreatment operation of the facial image to be retrieved includes：
Detect the face location in facial image to be retrieved and key point position；
According to the face location of the facial image to be retrieved and key point position, the facial image to be retrieved is carried out
Attitude updating and light correction process.
Second aspect, the embodiment of the present application provide a kind of facial image retrieval device based on deep learning, including：
Characteristic extracting module, for will pass through pretreatment operation facial image to be retrieved input training complete based on office
In the shared multitask convolutional neural networks model in portion, an identity characteristic vector sum for obtaining the facial image to be retrieved is more
A set of properties feature vector includes a global property group feature vector and at least one in the multiple set of properties feature vector
Local attribute's group feature vector；
Comparative analysis module, for dividing the identity characteristic vector sum set of properties feature vector of the facial image to be retrieved
Identity characteristic vector sum set of properties feature vector not with the facial image stored in database compares, and obtains described to be checked
The identity characteristic vector comparing result of the facial image stored in rope facial image and database, global property group feature vector pair
Than result and local attribute's feature vector comparing result；
Screening module, for the comparison knot according to the identity characteristic vector comparing result, global property group feature vector
Fruit and local attribute's feature vector comparing result, filter out target facial image in the database.
The third aspect, the embodiment of the present application provide a kind of terminal, including：
Processor；
The memory executed instruction for storage processor；
Wherein, the processor is configured to executing the described in any item methods of above-mentioned first aspect.
Facial image retrieval is carried out using method provided by the embodiments of the present application to have the following advantages that：
1, multitask convolutional neural networks model provided by the embodiments of the present application can extract respectively multiple spies for multitask
Vector is levied, is directly compared with database sample based on multiple feature vectors, rather than attribute point is carried out based on feature vector
Processing result after class, the unreliability for avoiding attributive classification influence, and robustness is higher.
2, the present invention devises similarity score fault tolerant mechanism during multiple feature vectors compare.It is comprehensive first
Identity characteristic similarity score and full face attributive character similarity score carry out coarse sizing, reduce because being based purely on identity characteristic phase
Omission caused by being compared like degree；Secondly, being obtained based on comprehensive identity characteristic similarity score and full face attributes similarity feature
The fusion similarity score divided is ranked up, and is excluded a part with query image and is belonged to the lower candidate figure of identical face probability
Picture；Finally, local notable feature score is found out from the similarity score of several local facial attributive character, as identity characteristic
The strong supplement of similarity score assists face alignment, obtains final candidate face image set.Since similarity score is fault-tolerant
Local attribute's feature of facial image is considered in mechanism, therefore, for face by partial occlusion and other local disturbances
The case where, can produce targetedly improves.
Detailed description of the invention
In order to illustrate the technical solutions in the embodiments of the present application or in the prior art more clearly, to embodiment or will show below
There is attached drawing needed in technical description to be briefly described, it should be apparent that, for those of ordinary skill in the art
Speech, without creative efforts, is also possible to obtain other drawings based on these drawings.
Fig. 1 is a kind of structural schematic diagram of local shared cell provided by the embodiments of the present application；
Fig. 2 is a kind of structure based on locally shared multitask convolutional neural networks model provided by the embodiments of the present application
Schematic diagram；
Fig. 3 is a kind of Research on face image retrieval flow diagram based on deep learning provided by the embodiments of the present application；
Fig. 4 is a kind of facial image screening technique flow diagram provided by the embodiments of the present application；
Fig. 5 is the structural representation that a kind of facial image based on deep learning provided by the embodiments of the present application retrieves device
Figure；
Fig. 6 is a kind of structural schematic diagram of terminal provided by the embodiments of the present application.
Specific embodiment
In order to make those skilled in the art better understand the technical solutions in the application, below in conjunction with the application reality
The attached drawing in example is applied, the technical scheme in the embodiment of the application is clearly and completely described, it is clear that described implementation
Example is merely a part but not all of the embodiments of the present application.Based on the embodiment in the application, this field is common
The application protection all should belong in technical staff's every other embodiment obtained without making creative work
Range.
In the embodiment of the present application, using a kind of based on locally shared multitask convolutional neural networks model, in order to
The technical solution in the application is more fully understood convenient for those skilled in the art, first below to nerve net involved in the application
Network model is simply introduced.
The invention relates to neural network model introduce a kind of local shared structure, Fig. 1 is the embodiment of the present application
The structural schematic diagram of a kind of local shared cell provided, as shown in Figure 1, the part shared cell is that each set of properties is provided with
One convolutional neural networks, referred to as specific network.In addition, being used alone between a shared e-learning different attribute group
Shared information.The feature of all specific network preceding layers is together in series by local shared cell, previous with shared network
Input of the feature of layer together as shared network later layer.And the input of shared network later layer be connected in series to it is each specifically
In the output of property network later layer, as lower layer of the input of specific network.Local shared cell makes shared network at certain
One layer can extract shared information, and specific network can further be extracted using these shared informations it is advantageous mutually to its
Information is mended, the study of corresponding set of properties is promoted.
Further, above-mentioned local shared cell is applied on all levels of neural network, is constituted based on part
Shared multitask convolutional neural networks.Fig. 2 is provided by the embodiments of the present application a kind of based on locally shared multitask convolution
In neural network structure shown in Fig. 2, face character is divided into according to spatial information for the structural schematic diagram of neural network model
4 set of properties constitute 4 subtasks, and carry out spy for one convolutional neural networks of each group of configuration (specific network TSNet)
Sign study.Meanwhile a shared network SNet is separately configured, and by local shared cell shown in FIG. 1, by itself and 4
TSNet is connected.Finally, each TSNet passes through the Classification Loss ALoss of set of properties as loss function, SNet passes through body
Part information differentiates that loss IDLoss instructs the training of network as loss function, and then exports 4 set of properties feature vectors and 1
A identity characteristic vector.Generally speaking, the corresponding set of properties of each TSNet, for learning specific characteristics；And SNet is used for
Learn sharing feature, excavates the relevance between set of properties.
In the embodiment of the present application, what 4 TSNet learnt respectively is upper face set of properties, the middle face set of properties, lower face of face
The feature of set of properties and full face set of properties.Certainly, those skilled in the art can adjust accordingly according to actual needs,
It should fall within the scope of the present application.It is pointed out that multitask convolutional Neural net provided by the embodiments of the present application
Network model can extract respectively multiple feature vectors for multitask；And it is traditional based on entirely shared multitask convolutional neural networks
Model is only capable of providing a feature vector, and identities and attributive classification are carried out based on this feature vector.
In addition, neural network model provided by the embodiments of the present application includes input layer, multiple convolutional layers, multiple concat
Layer, multiple full articulamentums and output layer, details are not described herein.
Before stating network model in use, need to be trained the network model.Specifically, by preset standard size
Facial image be input in the multitask convolutional neural networks, which is trained, until network model restrain, stop
Training.
In Research on face image retrieval provided by the embodiments of the present application, it is also necessary to use for storing the number referring to information
According to library.For the ease of the continuity of subsequent schedule description, first the database is illustrated herein.Wherein, the building of database
Process includes：The multitask convolutional neural networks that the facial image registered will be needed to be sequentially inputted to above-mentioned completion training, execute
Forward calculation obtains corresponding output, and output result is stored in the database.Wherein, the export structure includes identity
Feature vector and multiple set of properties feature vectors (face set of properties feature vector as above, middle face set of properties feature vector, lower face attribute
Group feature vector and full face set of properties feature vector), the multiple set of properties feature vectors of the identity characteristic vector sum are one-dimensional floating
Point number vector.
Fig. 3 is a kind of Research on face image retrieval flow diagram based on deep learning provided by the embodiments of the present application,
As shown in figure 3, it is mainly included the following steps that.
Step S301：Pretreatment operation is carried out to facial image to be retrieved, specifically includes following steps：
A, face location and the key point position in facial image to be retrieved are detected.
It should be pointed out that in the embodiment of the present application, face figure can be detected using existing Face datection algorithm
Face location as in, and according to detected face location, obtain the key point position of face image.Wherein, the people
Face position refers to the information of face present position in the picture, generally includes the picture of the face upper left corner (or central point) in the picture
The length and width of plain coordinate and face.The key point position is the coordinate value of preset some face key points, usually
These key points include eyes, face mask etc. to position important on face characteristic.Therefore, face location information can be with
It is used to refer to the position of face in the picture, key point position can be used to refer to the posture of face, expression, use these information
Facial image is corrected, normalized facial image is obtained, so that the face characteristic in later period extracts.
In the embodiment of the present application, the Face datection algorithm can be human-face detector, such as based on Like-Fenton Oxidation
And the human-face detector or object detector neural network based of AdaBoost, such as R-CNN or FCN.
B, according to the face location of the facial image to be retrieved and key point position, to the facial image to be retrieved into
Row attitude updating and light correction process.
In the embodiment of the present application, key point position and illumination condition possessed by a standard faces are pre-defined, is led to
The key point position that the key point position of the face image is aligned to standard faces by preset image transformation algorithm is crossed, to reach
To the purpose of correction human face posture.
By preset image processing algorithm, light correction is carried out to the face image after registration process, is made
The illumination condition that the illumination condition for obtaining the face image after registration process is converted into the standard faces and has (leads to
Light correction is crossed, so that the face image after registration process is consistent with standard faces, gamma can be used for example
Gamma value is corrected adjustment to image pixel value, so that treated, image has suitable contrast, and facial detail is clear
It is clear visible).
In specific implementation, key point position possessed by a pre-defined standard faces and illumination condition are specially：It can
With the key point location information and illumination condition previously according to multiple face images, is calculated by being averaging, obtain the standard
The key point position and illumination condition that face (i.e. average face) has.
It should be pointed out that the number of operations of the attitude updating and light correction is unlimited, and sequencing can be adjusted
It is whole；The illumination condition refers to the light environment of face image (including facial image and video) when shooting, shows image
In, the as brightness (can simply be interpreted as grey scale pixel value) of image.Preset standard faces image irradiation condition is generally true
It is high-visible to protect human face five-sense-organ, it will not be excessive lightness or darkness.
In specific implementation, the preset image transformation algorithm can be the basic image such as similarity transformation, affine transformation
Transform method is also possible to the combination of these primary images transformation.
In the embodiment of the present application, described " alignment " operation, which refers to, corrects the face image with posture to standard people
Face.Standard faces are usually positive facial image.According to the key point position for getting face image, by being carried out to image
The operation such as similar variation, affine transformation, realizes the alignment of face image, facial image face position and standard faces after alignment
It is almost the same.
Step S302：By the facial image to be retrieved input training Jing Guo pretreatment operation complete based on locally shared
In multitask convolutional neural networks model, the multiple set of properties of identity characteristic vector sum of the facial image to be retrieved are obtained
Feature vector includes a global property group feature vector and at least one local attribute in the multiple set of properties feature vector
Group feature vector.
In the embodiment of the present application, by the facial image by pretreatment operation, it is input to the completion training
In multitask convolutional neural networks, the forward calculation by executing it obtains corresponding multiple outputs, including identity characteristic vector
It include a global property group feature vector and at least one in multiple set of properties feature vector with multiple set of properties feature vectors
A local attribute's group feature vector.Specifically, which can be full face set of properties feature vector；Part
Set of properties feature vector may include upper face set of properties feature vector, middle face set of properties feature vector and lower face set of properties feature to
Amount.Wherein, the definition of set of properties is as shown in table 1, each set of properties may be defined as one in table in corresponding range of attributes,
Multiple or whole attributes.
Table one：
Step S303：By the identity characteristic vector sum set of properties feature vector of the facial image to be retrieved respectively with data
The identity characteristic vector sum set of properties feature vector of the facial image stored in library compares, and obtains the face figure to be retrieved
As with the identity characteristic vector comparing result of the facial image stored in database, global property group feature vector comparing result and
Local attribute's feature vector comparing result.
In the embodiment of the present application, by the identity characteristic vector for each candidate face image being pre-stored in database, by
A identity characteristic vector corresponding with the facial image to be retrieved is compared, and calculates identity characteristic similarity.Specific implementation
In, similarity can be calculated according to the COS distance between feature vector, obtained pre- in the facial image to be retrieved and database
The identity characteristic vector similarity score of each candidate face image of storage, similarity score is higher, shows two face figures
As more similar, the probability that two images belong to same people is larger.
By the multiple set of properties feature vectors for each candidate face image being pre-stored in database, one by one with it is described to be checked
The corresponding multiple set of properties feature vectors of rope facial image are compared, and calculate each set of properties characteristic similarity.
In the specific implementation, since each set of properties feature vector is one-dimensional floating point vector as identity characteristic vector,
Therefore equally similarity can be calculated according to the COS distance between set of properties feature vector, similarity is higher, shows two face figures
As more similar in the set of properties scope.Wherein, upper face set of properties, middle face set of properties, lower face set of properties are local attribute's group, entirely
Face set of properties is global property group；It is corresponding, upper face set of properties feature vector, middle face set of properties feature vector, lower face set of properties
Feature vector is local attribute's group feature vector, and full face set of properties feature vector is global property group feature vector.Scheme when two
Any local attribute's group feature vector similarity of picture is high, shows that the two images are corresponding in local attribute's group feature vector
It is closer in local attribute's group scope.When the global property group feature vector similarity height of two images, show the two figures
As being closer in corresponding global property group scope.If the similarity of each local attribute's group feature vector and global category
The similarity of property group feature vector, all higher, then the probability that can be shown that two images belong to same people is larger.
Step S304：According to the identity characteristic vector comparing result, global property group feature vector comparing result and
Local attribute's feature vector comparing result, filters out target facial image in the database.
In the embodiment of the present application, the identity characteristic vector comparing result, global property group feature vector comparison knot
Fruit and local attribute's feature vector comparing result, the corresponding identity characteristic of facial image stored in the respectively described database to
Measure similarity score, global property group feature vector similarity score and local attribute's feature vector similarity score.It is specific
It may comprise steps of：
Step S401, according to the identity characteristic vector similarity score and the global property group feature vector similarity
Score filters out candidate image set in the database.
In the embodiment of the present application, first according to identity characteristic similarity and global property characteristic similarity (full face attribute
Characteristic similarity) facial image stored in database is screened.For ease of description, by identity characteristic vector similarity
Score is denoted as Score_id, and global property feature vector similarity score is denoted as Score_gAttrib, and identity characteristic vector is similar
Degree threshold value is denoted as T_id, and global property feature vector similarity threshold is denoted as T_gAttrib, wherein T_id>0, T_gAttrib>
0。
Firstly, filtering out the facial image for meeting the first screening conditions in the database, scalping facial image is obtained
Set.First screening conditions are：Score_id >=T_id, and Score_gAttrib >=T_gAttrib, by scalping facial image
Set is denoted as L0, and the facial image number in L0 is N0.
Then, postsearch screening is carried out to scalping face image set L0, is filtered out in the scalping face image set
Meet the facial image of the second screening conditions, obtains candidate image set.Second screening conditions are：To merge similarity score
Score1=(Score_id+Score_gAttrib)/2 is scoring criteria, is sorted according to the sequence of Score1 from big to small, row
Sequence takes maximum N1, as candidate image set L1, wherein N1≤N0.
Generally speaking, the facial image for meeting the first screening conditions is filtered out in the database, obtains scalping face
Image collection, first screening conditions include the identity characteristic vector similarity score of facial image more than or equal to preset
Identity characteristic vector similarity threshold value, and global property group feature vector similarity score is greater than or equal to preset global property
Group feature vector similarity threshold；
The facial image for meeting the second screening conditions is filtered out in the scalping face image set, obtains candidate image
Set, second screening conditions are the first fusion maximum N1 of similarity score Score1 in the scalping face image set
A facial image.
Step S402, it is obtained according to the identity characteristic vector similarity score and local attribute's feature vector similarity
Point, or according to the identity characteristic vector similarity score, global property group feature vector similarity score and local attribute
Feature vector similarity score filters out target facial image in the candidate image set.
In the embodiment of the present application, local attribute's group feature vector includes upper face set of properties feature vector, middle face category
Property group feature vector and lower face set of properties feature vector.Correspondingly, local attribute's group feature vector similarity score includes
Upper face set of properties feature vector similarity score, middle face set of properties feature vector similarity score and lower face set of properties feature vector
Similarity score.
For ease of description, respectively by upper face set of properties feature vector similarity score, middle face set of properties feature vector phase
Score_Attri1, Score_Attrib2, Score_ are denoted as like degree score and lower face set of properties feature vector similarity score
Attrib3, respectively by upper face set of properties feature vector similarity threshold, middle face set of properties feature vector similarity threshold and lower face
Set of properties feature vector similarity threshold is denoted as T_Attri1, T_Attrib2, T_Attrib3, wherein T_Attri1>0, T_
Attri2>0, T_Attri3>0.
Each of candidate image set L1 facial image is traversed, judges whether to meet 3 following conditions.Condition 1：
Score_Attrib1 >=T_Attrib1, condition 2：Score_Attrib2 >=T_Attrib2, condition 3：Score_Attrib3≥
T_Attrib3.If condition 1~3 is all satisfied, the local notable feature score Score_loc=of current candidate facial image
max(Score_Attrib1,Score_Attrib2,Score_Attrib3)；If only there are two conditions to meet, it is assumed that condition i
Meet with condition j, wherein i=1 or 2 or 3, j=1 or 2 or 3, then Score_loc=max (Score_Attribi, Score_
Attribj)；If only a condition i meets, wherein i=1 or 2 or 3, then Score_loc=Score_Attribi；If
Three conditions are not satisfied, then the local notable feature score Score_loc=0 of current candidate image.
If the Score_loc of current candidate image is not equal to 0, the final fusion similarity score of current candidate image
Score2=w1*Score_id+w2*Score_loc；Otherwise Score2=w1*Score_id+w2*Score_gAttrib,
Middle w1, w2 are weight, w1>0, w2>0, w1+w2=1.To all N1 candidate face images in candidate image set L1 into
It after the final fusion similarity score of row calculates, reorders by the sequence of Score2 from big to small, selects wherein that Score2 is most
N2 big image is as final search result image.
Generally speaking, the embodiment of the present application filters out the second fusion similarity score in the candidate image set
The maximum facial image of Score2, as target facial image.Wherein, if at least one corresponding local attribute's group of facial image
Exist in feature vector and meet local attribute's group feature vectors of threshold value screening conditions, then by the threshold value screening conditions that meet
In local attribute's group feature vector, the maximum value of local attribute's group feature vector similarity score is as Score_loc；If face
There is no the local attribute's group features for meeting threshold value screening conditions at least one corresponding local attribute's group feature vector of image
Vector, then using the global property group feature vector similarity score as Score_loc, the threshold value screening conditions are part
Set of properties feature vector similarity score is greater than or equal to preset local attribute's group feature vector similarity threshold.
Facial image retrieval is carried out using method provided by the embodiments of the present application to have the following advantages that：
1, multitask convolutional neural networks model provided by the embodiments of the present application can extract respectively multiple spies for multitask
Vector is levied, is directly compared with database sample based on multiple feature vectors, rather than attribute point is carried out based on feature vector
Processing result after class, the unreliability for avoiding attributive classification influence, and robustness is higher.
2, the present invention devises similarity score fault tolerant mechanism during multiple feature vectors compare.It is comprehensive first
Identity characteristic similarity score and full face attributive character similarity score carry out coarse sizing, reduce because being based purely on identity characteristic phase
Omission caused by being compared like degree；Secondly, being obtained based on comprehensive identity characteristic similarity score and full face attributes similarity feature
The fusion similarity score divided is ranked up, and is excluded a part with query image and is belonged to the lower candidate figure of identical face probability
Picture；Finally, local notable feature score is found out from the similarity score of several local facial attributive character, as identity characteristic
The strong supplement of similarity score assists face alignment, obtains final candidate face image set.Since similarity score is fault-tolerant
Local attribute's feature of facial image is considered in mechanism, therefore, for face by partial occlusion and other local disturbances
The case where, can produce targetedly improves.
Such as when the facial image that same people does not wear glasses in the facial image and database worn glasses compares, face thereon
Attribute group character similarity score may not be high, however middle face attribute group, lower face attribute group are not affected by this but, therefore can be from it
It is middle to generate higher local notable feature score and improve final comparison.For another example, when the facial image and database to wear masks
In the facial image that does not wear masks of same people when comparing, higher local notable feature can be generated from upper face attribute group character and is obtained
Point, so as to improve final comparison.Even in addition, the different age group face alignment of same people, in identity characteristic similarity
In insufficient situation, due to being influenced lesser attribute by change of age there are some, therefore also can be from the higher office of a certain similarity
Higher local notable feature score is generated in subordinate's property group character, or higher obtain is generated from global property group character
Point, secondary identities aspect ratio pair, so as to improve the success rate of the identical face alignment across the age.
Corresponding with above method embodiment, present invention also provides a kind of, and the facial image based on deep learning retrieves dress
It sets, Fig. 5 is the structural schematic diagram that a kind of facial image based on deep learning provided by the embodiments of the present application retrieves device, is such as schemed
Shown in 5, mainly comprise the following modules.
Preprocessing module 501, for carrying out pretreatment operation to facial image to be retrieved.It is specifically used for：
A, face location and the key point position in facial image to be retrieved are detected；
B, according to the face location of the facial image to be retrieved and key point position, to the facial image to be retrieved into
Row attitude updating and light correction process.
Characteristic extracting module 502, the facial image to be retrieved for that will pass through pretreatment operation input the base that training is completed
In locally shared multitask convolutional neural networks model, an identity characteristic vector of the facial image to be retrieved is obtained
It include a global property group feature vector and at least in the multiple set of properties feature vector with multiple set of properties feature vectors
One local attribute's group feature vector.
In the embodiment of the present application, by the facial image by pretreatment operation, it is input to the completion training
In multitask convolutional neural networks, the forward calculation by executing it obtains corresponding multiple outputs, including identity characteristic vector
It include a global property group feature vector and at least one in multiple set of properties feature vector with multiple set of properties feature vectors
A local attribute's group feature vector.
Comparative analysis module 503, for by the identity characteristic vector sum set of properties feature of the facial image to be retrieved to
It measures the identity characteristic vector sum set of properties feature vector respectively with the facial image stored in database to compare, described in acquisition
The identity characteristic vector comparing result of the facial image stored in facial image to be retrieved and database, global property group feature to
Measure comparing result and local attribute's feature vector comparing result.
In the specific implementation, similarity can be calculated according to the COS distance between feature vector, the face to be retrieved is obtained
The identity characteristic vector similarity score for each candidate face image being pre-stored in image and database, similarity score are got over
Height shows that two facial images are more similar, and the probability that two images belong to same people is larger.In addition, since each set of properties is special
Levying vector is all one-dimensional floating point vector as identity characteristic vector, therefore equally can be according to the cosine between set of properties feature vector
Distance calculates similarity, and similarity is higher, shows that two facial images are more similar in the set of properties scope.
Screening module 504, for the comparison according to the identity characteristic vector comparing result, global property group feature vector
As a result with local attribute's feature vector comparing result, target facial image is filtered out in the database.It is specifically used for：
A, according to the identity characteristic vector similarity score and the global property group feature vector similarity score,
Candidate image set is filtered out in the database.
The facial image for meeting the first screening conditions is filtered out in the database, obtains scalping face image set,
First screening conditions include that the identity characteristic vector similarity score of facial image is greater than or equal to preset identity characteristic
Vector similarity threshold value, and global property group feature vector similarity score be greater than or equal to preset global property group feature to
Measure similarity threshold；
The facial image for meeting the second screening conditions is filtered out in the scalping face image set, obtains candidate image
Set, second screening conditions are the first fusion maximum N1 of similarity score Score1 in the scalping face image set
A facial image, wherein Score1=(Score_id+Score_gAttrib)/2, Score_id is that identity characteristic vector is similar
Score is spent, Score_gAttrib is global property group feature vector similarity score, and N1≤N0, N0 are scalping face image set
The quantity of facial image in conjunction.
B, according to the identity characteristic vector similarity score and local attribute's feature vector similarity score, or
According to the identity characteristic vector similarity score, global property group feature vector similarity score and local attribute's feature vector
Similarity score filters out target facial image in the candidate image set.
The second maximum N2 facial image of fusion similarity score Score2 is filtered out in the candidate image set,
As target facial image, N2 >=1, such as N2 can be 1,5 or 10 etc., those skilled in the art can according to actual needs into
The corresponding setting of row；
Wherein, Score2=w1*Score_id+w2*Score_loc, w1, w2 are weight, w1>0, w2>0, w1+w2=1；
If there is the office for meeting threshold value screening conditions at least one corresponding local attribute's group feature vector of facial image
Portion's set of properties feature vector, then by the local attribute's group feature vector for meeting threshold value screening conditions, local attribute's group is special
The maximum value of vector similarity score is levied as Score_loc；If at least one corresponding local attribute's group feature of facial image
There is no the local attribute's group feature vectors for meeting threshold value screening conditions in vector, then by the global property group feature vector phase
Like degree score as Score_loc, the threshold value screening conditions are that local attribute's group feature vector similarity score is greater than or waits
In preset local attribute's group feature vector similarity threshold.
Facial image retrieval is carried out using device provided by the embodiments of the present application to have the following advantages that：
1, multitask convolutional neural networks model provided by the embodiments of the present application can extract respectively multiple spies for multitask
Vector is levied, is directly compared with database sample based on multiple feature vectors, rather than attribute point is carried out based on feature vector
Processing result after class, the unreliability for avoiding attributive classification influence, and robustness is higher.
2, the present invention devises similarity score fault tolerant mechanism during multiple feature vectors compare.It is comprehensive first
Identity characteristic similarity score and full face attributive character similarity score carry out coarse sizing, reduce because being based purely on identity characteristic phase
Omission caused by being compared like degree；Secondly, being obtained based on comprehensive identity characteristic similarity score and full face attributes similarity feature
The fusion similarity score divided is ranked up, and is excluded a part with query image and is belonged to the lower candidate figure of identical face probability
Picture；Finally, local notable feature score is found out from the similarity score of several local facial attributive character, as identity characteristic
The strong supplement of similarity score assists face alignment, obtains final candidate face image set.
Corresponding with above-described embodiment, present invention also provides a kind of for carrying out the terminal of facial image retrieval.Fig. 6 is
The structural schematic diagram of a kind of terminal provided by the embodiments of the present application, as shown in fig. 6, the terminal 600 may include：Processor
610, memory 620 and communication unit 630.These components are communicated by one or more bus, those skilled in the art
It is appreciated that the structure of server shown in figure does not constitute the restriction to the application, it either busbar network,
It can be hub-and-spoke configuration, can also include perhaps combining certain components or different portions than illustrating more or fewer components
Part arrangement.
Wherein, the communication unit 630, for establishing communication channel, so that the storage equipment be allow to set with other
It is standby to be communicated.Receive the user data or send user data to other equipment that other equipment are sent.
The processor 610 utilizes various interfaces and the entire electronic equipment of connection for the control centre for storing equipment
Various pieces, by running or execute the software program and/or module that are stored in memory 620, and call and be stored in
Data in memory, to execute the various functions and/or processing data of electronic equipment.The processor can be by integrated circuit
(Integrated Circuit, abbreviation IC) composition, such as the IC that can be encapsulated by single are formed, can also be by more of connection
The encapsulation IC of identical function or different function and form.For example, processor 610 can only include central processing unit
(Central Processing Unit, abbreviation CPU).In the application embodiment, CPU can be single operation core, can also
To include multioperation core.
The memory 620, for executing instruction for storage processor 610, memory 620 can be by any kind of easy
The property lost or non-volatile memory device or their combination are realized, such as static random access memory (SRAM), electric erasable
Programmable read only memory (EEPROM), Erasable Programmable Read Only Memory EPROM (EPROM), programmable read only memory
(PROM), read-only memory (ROM), magnetic memory, flash memory, disk or CD.
When executing instruction in memory 620 is executed by processor 610, so that terminal 600 is able to carry out the above method
Step some or all of in embodiment.
In the specific implementation, the application also provides a kind of computer storage medium, wherein the computer storage medium can store
There is program, which may include step some or all of in each embodiment provided by the present application when executing.The storage is situated between
Matter can be magnetic disk, CD, read-only memory (English：Read-only memory, referred to as：ROM) or random storage is remembered
Body (English：Random access memory, referred to as：RAM) etc..
It is required that those skilled in the art can be understood that the technology in the embodiment of the present application can add by software
The mode of general hardware platform realize.Based on this understanding, the technical solution in the embodiment of the present application substantially or
Say that the part that contributes to existing technology can be embodied in the form of software products, which can deposit
Storage is in storage medium, such as ROM/RAM, magnetic disk, CD, including some instructions are used so that computer equipment (can be with
It is personal computer, server or the network equipment etc.) execute certain part institutes of each embodiment of the application or embodiment
The method stated.
Same and similar part may refer to each other between each embodiment in this specification.Implement especially for terminal
For example, since it is substantially similar to the method embodiment, so being described relatively simple, related place is referring in embodiment of the method
Explanation.
Above-described the application embodiment does not constitute the restriction to the application protection scope.
Claims (10)
1. a kind of Research on face image retrieval based on deep learning, which is characterized in that including：
Facial image to be retrieved input training Jing Guo pretreatment operation is completed refreshing based on locally shared multitask convolution
Through obtaining the multiple set of properties feature vectors of an identity characteristic vector sum of the facial image to be retrieved, institute in network model
State includes a global property group feature vector and at least one local attribute's group feature vector in multiple set of properties feature vectors；
By the identity characteristic vector sum set of properties feature vector of the facial image to be retrieved respectively with the people that is stored in database
The identity characteristic vector sum set of properties feature vector of face image compares, and obtains in the facial image to be retrieved and database
Identity characteristic vector comparing result, global property group feature vector comparing result and the local attribute's feature of the facial image of storage
Vector comparing result；
According to the identity characteristic vector comparing result, the comparing result of global property group feature vector and local attribute's feature to
Comparing result is measured, filters out target facial image in the database.
2. the method according to claim 1, wherein it is described according to the identity characteristic vector comparing result, it is complete
The comparing result and local attribute's feature vector comparing result of office's set of properties feature vector, filter out target face in the database
Image, including：
According to the identity characteristic vector comparing result and the global property group feature vector comparing result, in the database
In filter out candidate image set；
According to the identity characteristic vector comparing result and local attribute's feature vector comparing result, or according to the body
The comparing result and local attribute's feature vector comparing result of part feature vector comparing result, global property group feature vector,
Target facial image is filtered out in the candidate image set.
3. according to the method described in claim 2, it is characterized in that, the identity characteristic vector comparing result, global property group
The comparing result and local attribute's feature vector comparing result of feature vector, the facial image stored in the respectively described database
Corresponding identity characteristic vector similarity score, global property group feature vector similarity score and local attribute's feature vector phase
Like degree score；
It is described according to the identity characteristic vector comparing result and the global property group feature vector comparing result, in the number
According to filtering out candidate image set in library, including：
The facial image for meeting the first screening conditions is filtered out in the database, obtains scalping face image set, it is described
First screening conditions include that the identity characteristic vector similarity score of facial image is greater than or equal to preset identity characteristic vector
Similarity threshold, and global property group feature vector similarity score is greater than or equal to preset global property group feature vector phase
Like degree threshold value；
The facial image for meeting the second screening conditions is filtered out in the scalping face image set, obtains candidate image collection
It closes, second screening conditions are the first fusion maximum N1 of similarity score Score1 in the scalping face image set
Facial image, wherein Score1=(Score_id+Score_gAttrib)/2, Score_id is identity characteristic vector similarity
Score, Score_gAttrib are global property group feature vector similarity score, and N1≤N0, N0 are scalping face image set
The quantity of middle facial image.
4. according to the method described in claim 3, it is characterized in that, according to the identity characteristic vector comparing result and the office
Portion's attribute feature vector comparing result, or according to the identity characteristic vector comparing result, global property group feature vector
Comparing result and local attribute's feature vector comparing result filter out target facial image in the candidate image set, packet
It includes：
The second maximum facial image of fusion similarity score Score2 is filtered out in the candidate image set, as target
Facial image；
Wherein, Score2=w1*Score_id+w2*Score_loc, w1, w2 are weight, w1>0, w2>0, w1+w2=1；
If there is the part category for meeting threshold value screening conditions at least one corresponding local attribute's group feature vector of facial image
Property group feature vector, then by the local attribute's group feature vector for meeting threshold value screening conditions, local attribute's group feature to
The maximum value of similarity score is measured as Score_loc；If at least one corresponding local attribute's group feature vector of facial image
In there is no local attribute's group feature vectors of threshold value screening conditions is met, then by the global property group feature vector similarity
For score as Score_loc, the threshold value screening conditions are that local attribute's group feature vector similarity score is greater than or equal in advance
If local attribute's group feature vector similarity threshold.
5. method according to claim 1-4, which is characterized in that at least one described local attribute's group feature to
Amount, including：
Upper face set of properties feature vector, middle face set of properties feature vector and lower face set of properties feature vector.
6. according to the method described in claim 5, it is characterized in that,
Attribute of the upper face set of properties feature vector for characterization includes eyebrow attribute, eyes attribute, color development attribute, hair style category
Property and/or upper face adjunct attribute；
Attribute of the middle face set of properties feature vector for characterization includes nose attribute, cheek attribute, cheekbone attribute, temples category
Property and/or middle face adjunct attribute；
Attribute of the lower face set of properties feature vector for characterization includes lip attribute, chin attribute, beard attribute, mouth category
Property and/or lower face adjunct attribute；
Attribute of the global property group feature vector for characterization includes gender attribute, expression attribute, shape of face attribute, face category
Property, hair style attribute and/or age attribute.
7. the method according to claim 1, wherein the building process of the database includes：
By the facial image input training for needing to register complete based in locally shared multitask convolutional neural networks model,
It obtains and described needs the corresponding multiple set of properties of an identity characteristic vector sum of each facial image in the facial image registered special
Levy vector；
The corresponding multiple set of properties feature vectors of an identity characteristic vector sum of each facial image are stored in database
In.
8. the method according to claim 1, wherein the pretreatment operation of the facial image to be retrieved includes：
Detect the face location in facial image to be retrieved and key point position；
According to the face location of the facial image to be retrieved and key point position, posture is carried out to the facial image to be retrieved
Correction and light correction process.
9. a kind of facial image based on deep learning retrieves device, which is characterized in that including：
Characteristic extracting module, the facial image to be retrieved for that will pass through pretreatment operation input the total based on part of training completion
In the multitask convolutional neural networks model enjoyed, the multiple categories of identity characteristic vector sum of the facial image to be retrieved are obtained
Property group feature vector, include a global property group feature vector and at least one part in the multiple set of properties feature vector
Set of properties feature vector；
Comparative analysis module, for by the identity characteristic vector sum set of properties feature vector of the facial image to be retrieved respectively with
The identity characteristic vector sum set of properties feature vector of the facial image stored in database compares, and obtains the people to be retrieved
The identity characteristic vector comparing result of the facial image stored in face image and database, the comparison of global property group feature vector are tied
Fruit and local attribute's feature vector comparing result；
Screening module, for according to the comparing result of the identity characteristic vector comparing result, global property group feature vector and
Local attribute's feature vector comparing result, filters out target facial image in the database.
10. a kind of terminal, which is characterized in that including：
Processor；
The memory executed instruction for storage processor；
Wherein, the processor is configured to perform claim requires the described in any item methods of 1-8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810856269.8A CN108829900B (en) | 2018-07-31 | 2018-07-31 | Face image retrieval method and device based on deep learning and terminal |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810856269.8A CN108829900B (en) | 2018-07-31 | 2018-07-31 | Face image retrieval method and device based on deep learning and terminal |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108829900A true CN108829900A (en) | 2018-11-16 |
| CN108829900B CN108829900B (en) | 2020-11-10 |
Family
ID=64152313
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810856269.8A Expired - Fee Related CN108829900B (en) | 2018-07-31 | 2018-07-31 | Face image retrieval method and device based on deep learning and terminal |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108829900B (en) |
Cited By (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109583403A (en) * | 2018-12-06 | 2019-04-05 | 联想(北京)有限公司 | Image processing method, processor and electronic equipment |
| CN109658194A (en) * | 2018-12-20 | 2019-04-19 | 焦点科技股份有限公司 | A kind of lead referral method and system based on video frequency tracking |
| CN109711357A (en) * | 2018-12-28 | 2019-05-03 | 北京旷视科技有限公司 | A kind of face identification method and device |
| CN109993102A (en) * | 2019-03-28 | 2019-07-09 | 北京达佳互联信息技术有限公司 | Similar face retrieval method, apparatus and storage medium |
| CN110287890A (en) * | 2019-06-26 | 2019-09-27 | 银河水滴科技（北京）有限公司 | A recognition method and device based on gait features and pedestrian re-identification features |
| CN111368101A (en) * | 2020-03-05 | 2020-07-03 | 腾讯科技（深圳）有限公司 | Multimedia resource information display method, device, equipment and storage medium |
| CN111985360A (en) * | 2020-08-05 | 2020-11-24 | 上海依图网络科技有限公司 | Face recognition method, device, equipment and medium |
| CN112241689A (en) * | 2020-09-24 | 2021-01-19 | 北京澎思科技有限公司 | Face recognition method and device, electronic equipment and computer readable storage medium |
| CN112364827A (en) * | 2020-11-30 | 2021-02-12 | 腾讯科技（深圳）有限公司 | Face recognition method and device, computer equipment and storage medium |
| CN112417197A (en) * | 2020-12-02 | 2021-02-26 | 云从科技集团股份有限公司 | Sorting method, sorting device, machine readable medium and equipment |
| CN112417198A (en) * | 2020-12-07 | 2021-02-26 | 武汉柏禾智科技有限公司 | Face image retrieval method |
| CN112949599A (en) * | 2021-04-07 | 2021-06-11 | 广州弘晟计算机系统有限公司 | Candidate content pushing method based on big data |
| CN113269125A (en) * | 2021-06-10 | 2021-08-17 | 北京中科闻歌科技股份有限公司 | Face recognition method, device, equipment and storage medium |
| CN113688764A (en) * | 2021-08-31 | 2021-11-23 | 瓴盛科技有限公司 | Training method and device for face optimization model and computer readable medium |
| CN113947564A (en) * | 2021-08-31 | 2022-01-18 | 国家电网有限公司大数据中心 | Image verification method and system for low-voltage distribution area metering equipment in power industry |
| CN114120386A (en) * | 2020-08-31 | 2022-03-01 | 腾讯科技（深圳）有限公司 | Face recognition method, device, equipment and storage medium |
| CN114385993A (en) * | 2021-12-27 | 2022-04-22 | 阿里巴巴（中国）有限公司 | Identity detection method, device and readable medium |
| CN114724180A (en) * | 2022-03-17 | 2022-07-08 | 重庆中科云从科技有限公司 | Object recognition method, device and computer storage medium |
Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2014229012A (en) * | 2013-05-21 | 2014-12-08 | 沖電気工業株式会社 | Person attribute estimation apparatus, and person attribute estimation method and program |
| CN105404877A (en) * | 2015-12-08 | 2016-03-16 | 商汤集团有限公司 | Face attribute prediction method and device based on deep learning and multi-task learning |
| CN106022317A (en) * | 2016-06-27 | 2016-10-12 | 北京小米移动软件有限公司 | Face identification method and apparatus |
| CN106815566A (en) * | 2016-12-29 | 2017-06-09 | 天津中科智能识别产业技术研究院有限公司 | A kind of face retrieval method based on multitask convolutional neural networks |
| CN106874877A (en) * | 2017-02-20 | 2017-06-20 | 南通大学 | A kind of combination is local and global characteristics without constraint face verification method |
| CN107145857A (en) * | 2017-04-29 | 2017-09-08 | 深圳市深网视界科技有限公司 | Face character recognition methods, device and method for establishing model |
| CN107330359A (en) * | 2017-05-23 | 2017-11-07 | 深圳市深网视界科技有限公司 | A kind of method and apparatus of face contrast |
| US20180025217A1 (en) * | 2016-07-22 | 2018-01-25 | Nec Laboratories America, Inc. | Liveness detection for antispoof face recognition |
-
2018
- 2018-07-31 CN CN201810856269.8A patent/CN108829900B/en not_active Expired - Fee Related
Patent Citations (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2014229012A (en) * | 2013-05-21 | 2014-12-08 | 沖電気工業株式会社 | Person attribute estimation apparatus, and person attribute estimation method and program |
| CN105404877A (en) * | 2015-12-08 | 2016-03-16 | 商汤集团有限公司 | Face attribute prediction method and device based on deep learning and multi-task learning |
| CN106022317A (en) * | 2016-06-27 | 2016-10-12 | 北京小米移动软件有限公司 | Face identification method and apparatus |
| US20180025217A1 (en) * | 2016-07-22 | 2018-01-25 | Nec Laboratories America, Inc. | Liveness detection for antispoof face recognition |
| CN106815566A (en) * | 2016-12-29 | 2017-06-09 | 天津中科智能识别产业技术研究院有限公司 | A kind of face retrieval method based on multitask convolutional neural networks |
| CN106874877A (en) * | 2017-02-20 | 2017-06-20 | 南通大学 | A kind of combination is local and global characteristics without constraint face verification method |
| CN107145857A (en) * | 2017-04-29 | 2017-09-08 | 深圳市深网视界科技有限公司 | Face character recognition methods, device and method for establishing model |
| CN107330359A (en) * | 2017-05-23 | 2017-11-07 | 深圳市深网视界科技有限公司 | A kind of method and apparatus of face contrast |
Non-Patent Citations (3)
| Title |
|---|
| GALEA, CHRISTIAN等: "Forensic Face Photo-Sketch Recognition Using a Deep Learning-Based Architecture", 《IEEE SIGNAL PROCESSING LETTERS》 * |
| QU, ZHI等: "Face Verification Algorithm with Exploiting Feature Distribution", 《PRICAI 2016: TRENDS IN ARTIFICIAL INTELLIGENCE》 * |
| 武亚昆: "融合多种图像特征的人像检索技术研究", 《中国优秀硕士学位论文全文数据库(电子期刊)》 * |
Cited By (23)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109583403A (en) * | 2018-12-06 | 2019-04-05 | 联想(北京)有限公司 | Image processing method, processor and electronic equipment |
| CN109658194A (en) * | 2018-12-20 | 2019-04-19 | 焦点科技股份有限公司 | A kind of lead referral method and system based on video frequency tracking |
| CN109711357A (en) * | 2018-12-28 | 2019-05-03 | 北京旷视科技有限公司 | A kind of face identification method and device |
| CN109993102A (en) * | 2019-03-28 | 2019-07-09 | 北京达佳互联信息技术有限公司 | Similar face retrieval method, apparatus and storage medium |
| CN109993102B (en) * | 2019-03-28 | 2021-09-17 | 北京达佳互联信息技术有限公司 | Similar face retrieval method, device and storage medium |
| CN110287890A (en) * | 2019-06-26 | 2019-09-27 | 银河水滴科技（北京）有限公司 | A recognition method and device based on gait features and pedestrian re-identification features |
| CN111368101A (en) * | 2020-03-05 | 2020-07-03 | 腾讯科技（深圳）有限公司 | Multimedia resource information display method, device, equipment and storage medium |
| CN111368101B (en) * | 2020-03-05 | 2021-06-18 | 腾讯科技（深圳）有限公司 | Multimedia resource information display method, device, equipment and storage medium |
| CN111985360A (en) * | 2020-08-05 | 2020-11-24 | 上海依图网络科技有限公司 | Face recognition method, device, equipment and medium |
| CN114120386A (en) * | 2020-08-31 | 2022-03-01 | 腾讯科技（深圳）有限公司 | Face recognition method, device, equipment and storage medium |
| CN112241689A (en) * | 2020-09-24 | 2021-01-19 | 北京澎思科技有限公司 | Face recognition method and device, electronic equipment and computer readable storage medium |
| CN112364827A (en) * | 2020-11-30 | 2021-02-12 | 腾讯科技（深圳）有限公司 | Face recognition method and device, computer equipment and storage medium |
| CN112364827B (en) * | 2020-11-30 | 2023-11-10 | 腾讯科技（深圳）有限公司 | Face recognition method, device, computer equipment and storage medium |
| CN112417197A (en) * | 2020-12-02 | 2021-02-26 | 云从科技集团股份有限公司 | Sorting method, sorting device, machine readable medium and equipment |
| CN112417198A (en) * | 2020-12-07 | 2021-02-26 | 武汉柏禾智科技有限公司 | Face image retrieval method |
| CN112949599A (en) * | 2021-04-07 | 2021-06-11 | 广州弘晟计算机系统有限公司 | Candidate content pushing method based on big data |
| CN113269125A (en) * | 2021-06-10 | 2021-08-17 | 北京中科闻歌科技股份有限公司 | Face recognition method, device, equipment and storage medium |
| CN113269125B (en) * | 2021-06-10 | 2024-05-14 | 北京中科闻歌科技股份有限公司 | Face recognition method, device, equipment and storage medium |
| CN113688764A (en) * | 2021-08-31 | 2021-11-23 | 瓴盛科技有限公司 | Training method and device for face optimization model and computer readable medium |
| CN113947564A (en) * | 2021-08-31 | 2022-01-18 | 国家电网有限公司大数据中心 | Image verification method and system for low-voltage distribution area metering equipment in power industry |
| CN113947564B (en) * | 2021-08-31 | 2025-04-15 | 国家电网有限公司大数据中心 | A method and system for image verification of low-voltage metering equipment in the power industry |
| CN114385993A (en) * | 2021-12-27 | 2022-04-22 | 阿里巴巴（中国）有限公司 | Identity detection method, device and readable medium |
| CN114724180A (en) * | 2022-03-17 | 2022-07-08 | 重庆中科云从科技有限公司 | Object recognition method, device and computer storage medium |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108829900B (en) | 2020-11-10 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108829900A (en) | A kind of Research on face image retrieval based on deep learning, device and terminal | |
| González-Briones et al. | A multi-agent system for the classification of gender and age from images | |
| CN112418095B (en) | A method and system for facial expression recognition combined with attention mechanism | |
| Ashraf et al. | The painful face: Pain expression recognition using active appearance models | |
| Cament et al. | Fusion of local normalization and Gabor entropy weighted features for face identification | |
| CN105138954B (en) | A kind of image automatic screening inquiry identifying system | |
| CN111310624A (en) | Occlusion recognition method and device, computer equipment and storage medium | |
| WO2021078157A1 (en) | Image processing method and apparatus, electronic device, and storage medium | |
| WO2017107957A1 (en) | Human face image retrieval method and apparatus | |
| CN108509862A (en) | Anti- angle and the fast human face recognition for blocking interference | |
| WO2024109374A1 (en) | Training method and apparatus for face swapping model, and device, storage medium and program product | |
| CN109472206A (en) | Methods of risk assessment, device, equipment and medium based on micro- expression | |
| CN109858375A (en) | Living body faces detection method, terminal and computer readable storage medium | |
| Poornima et al. | Attendance monitoring system using facial recognition with audio output and gender classification | |
| CN105224929A (en) | A kind of method of searching human face photo | |
| CN108182714A (en) | Image processing method and device, storage medium | |
| CN105956570B (en) | Smile recognition method based on lip features and deep learning | |
| CN111178130A (en) | Face recognition method, system and readable storage medium based on deep learning | |
| Paul et al. | Extraction of facial feature points using cumulative histogram | |
| CN103034840A (en) | Gender identification method | |
| CN113468925B (en) | Occlusion face recognition method, intelligent terminal and storage medium | |
| CN109829388A (en) | Video data handling procedure, device and computer equipment based on micro- expression | |
| Becerra-Riera et al. | Age and gender classification using local appearance descriptors from facial components | |
| Maximiano da Silva et al. | Geometrical features and active appearance model applied to facial expression recognition | |
| Alrabiah et al. | Unconstrained Gender Recognition from Periocular Region Using Multiscale Deep Features. |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| CF01 | Termination of patent right due to non-payment of annual fee |
Granted publication date: 20201110 |
|
| CF01 | Termination of patent right due to non-payment of annual fee |