CN107918636A - A kind of face method for quickly retrieving, system - Google Patents
A kind of face method for quickly retrieving, system Download PDFInfo
- Publication number
- CN107918636A CN107918636A CN201710801518.9A CN201710801518A CN107918636A CN 107918636 A CN107918636 A CN 107918636A CN 201710801518 A CN201710801518 A CN 201710801518A CN 107918636 A CN107918636 A CN 107918636A
- Authority
- CN
- China
- Prior art keywords
- image
- face
- hash
- network
- feature
- 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; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
- G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually
- G06F16/583—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content
- G06F16/5838—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually using metadata automatically derived from the content using colour
-
- 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)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Biomedical Technology (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Library & Information Science (AREA)
- Databases & Information Systems (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of face method for quickly retrieving, system, method includes:Obtain the feature vector of image, described eigenvector is inputted into an autoencoder network, full articulamentum weight and corresponding bias term are obtained according to autoencoder network training and renewal, and as the network parameter that feature vector is carried out to binaryzation Hash, the hash index storehouse of image is established by the network parameter and obtains the cryptographic Hash of image to be checked, searches out face result.Extracted in the present invention by using depth convolutional neural networks as face characteristic, efficient face characteristic expression can be obtained.Meanwhile Hash codes are obtained using autoencoder network, overall compact binaryzation expression is obtained based on face characteristic.In addition, the present invention calculates image similarity using the Hamming distance of Hash codes, calculation amount is small can to accelerate retrieval rate.
Description
Technical field
The present invention relates to deep learning field, facial image identification field, more particularly to a kind of face quick-searching side
Method, system, are based primarily upon convolutional neural networks and autoencoder network binaryzation Hash.
Background technology
Facial image database of the prior art includes many types, such as, FERET face databases, CMU-PIE people
Face database, YALE face databases, MIT face databases, ORL face databases etc..And use the mesh of facial image database
Be:Similar facial image is retrieved in facial image database has extensive answer in the recognition of face such as monitoring, security protection direction
Use prospect.
It is known, Hash coding is carried out for original image, the speed of image retrieval can be effectively improved.
Such as in the prior art, Chinese Patent Application No.:CN 201310087561.5 is a kind of to be based on local sensitivity Hash
Similar face method for quickly retrieving, disclose a kind of Research on face image retrieval based on local sensitivity Hash.This method is led to
Cross face region detection, eyes and the detection of face feature and feature extraction, Face Detection, the extraction of face complexion distribution characteristics etc.
Step represents the image as face feature vector, and then face feature vector is built using local sensitivity hash method and is indexed,
So as to improve speed during inquiry.
Analysis understands that deficiency is existing for this method:Eyes, face and features of skin colors can not express whole well
The feature of face, and local sensitivity hash method is a kind of unrelated hash method of data, randomness is strong;In order to ensure preferably
Retrieval precision, it is necessary to coding digit it is very long, retrieval efficiency ratio is relatively low.
Again such as in the prior art, image inspections of the Chinese Patent Application No. CN201410441091 based on semi-supervised Hash
Suo Fangfa, discloses a kind of image search method based on semi-supervised Hash.This method extracts the local space of image about first
The global frequencies feature of beam, with reference to labeled data and unlabeled data Training Support Vector Machines, obtains the coding of data.
Analysis understands that the shortcomings that program is:The data of training image are less, and global frequencies feature is not appropriate for
Recognition of face task.
To sum up, the speed and accuracy rate of facial image retrieval how are improved, is that those skilled in the art have skill to be solved
Art problem.
The content of the invention
The technical problem to be solved in the present invention is, there is provided a kind of face method for quickly retrieving, by using depth convolution god
The feature of facial image is obtained through network, so that feature representation is more efficiently and accurately.In addition, also use own coding net
Network combination facial image feature obtains the binaryzation cryptographic Hash of image, reduces the occupancy of memory space, improves facial image
The speed and accuracy rate of retrieval.
Above-mentioned technical problem is solved, the present invention provides a kind of face method for quickly retrieving, is included the following steps:
The feature vector of image is obtained, described eigenvector is inputted into an autoencoder network,
Full articulamentum weight and corresponding bias term are obtained according to autoencoder network training and renewal, and conduct will
Feature vector carries out the network parameter of binaryzation Hash,
The hash index storehouse of image is established by the network parameter and obtains the cryptographic Hash of image to be checked, is searched for
Go out face result.
Extracted when obtaining the feature vector of image using depth convolutional neural networks as face characteristic, obtain face
The high efficient expression of feature.The binaryzation of feature is obtained using autoencoder network based on face characteristic in above-mentioned autoencoder network
Cryptographic Hash, combines the learning ability of autoencoder network, can obtain greater compactness of binaryzation expression.In addition, searched in index database
Rope goes out the retrieving of face result using the similarity of the Hamming distance calculating image of Hash codes, and calculation amount is small to improve retrieval
Speed.
Further, in the autoencoder network,
Autoencoder network is trained by using loss function and multidimensional characteristic first,
Then corresponding model parameter is updated using stochastic gradient descent method,
Last preservation model parameter, as the network parameter that feature is carried out to binaryzation Hash.
Further, method further includes:The following steps in the hash index storehouse of image are established,
Described eigenvector carries out binaryzation by autoencoder network and obtains cryptographic Hash, utilizes the cryptographic Hash of above-mentioned acquisition
Establish the index of image library.
Further, method further includes:The following steps of query process,
The cryptographic Hash of image to be checked is obtained,
The Hamming distance with Hash codes in index database is calculated,
By the numerical ordering of Hamming distance, and corresponding original image is exported in order, obtain retrieval result.
Further, facial image pre-treatment step is further included before the feature vector for obtaining image:
If the amount of images for participating in the training set of training is Ntrain, for the image I of each inputi, (i=1,
2 ..., Ntrain),
Face datection is carried out to picture to mark with key point, and registration process is carried out to facial image according to key point afterwards,
Image is incorporated into as same scale.
Further, following steps are further included after the feature vector for obtaining image:
By the facial image I ' i after processing, (i=1,2 ..., Ntrain) add convolutional neural networks be trained, obtain
The corresponding network parameter for face characteristic extraction,
After the completion of convolutional neural networks training, facial image and feature vector are carried out by the network parameter of acquisition
Mapping, extracts the feature vector of facial image in training set.
Based on above-mentioned, the present invention provides a kind of face quick retrieval system, including:Feature extraction unit and feature inspection
Cable elements,
The feature extraction unit, to obtain the feature vector of image, an own coding net is inputted by described eigenvector
Network,
Full articulamentum weight and corresponding bias term are obtained according to autoencoder network training and renewal, and conduct will
Feature vector carries out the network parameter of binaryzation Hash,
The feature retrieval unit, to establish the people that index database searches out image to be checked by the network parameter
Face result.
Further, the feature retrieval unit includes:Index establishes unit and index database query unit,
The index establishes unit, and Hash is obtained described eigenvector is carried out binaryzation by autoencoder network
Value, the index of image library is established using the cryptographic Hash of above-mentioned acquisition,
The index database query unit, to obtain the cryptographic Hash of image to be checked, calculates and Hash codes in index database
Hamming distance, by the numerical ordering of Hamming distance, and exports corresponding original image, obtains retrieval result in order.
Further, the feature extraction unit further includes:Pretreatment unit, the pretreatment unit, into pedestrian
Face normalized.
Further, the feature extraction unit includes:Convolutional neural networks and autoencoder network.
Beneficial effects of the present invention:
Face method for quickly retrieving in the present invention, due to including:Obtain image feature vector, by the feature to
Amount one autoencoder network of input, full articulamentum weight and corresponding biasing are obtained according to autoencoder network training and renewal
, and as the network parameter that feature vector is carried out to binaryzation Hash, index database is established by the network parameter and is searched out
Face result.Extract, can obtain efficient as face characteristic by using depth convolutional neural networks in the above-mentioned methods
Face characteristic is expressed.Meanwhile Hash codes are obtained using autoencoder network, overall compact binaryzation is obtained based on face characteristic
Expression.Further, image similarity, the small quickening retrieval rate of calculation amount are calculated using the Hamming distance of Hash codes.
Brief description of the drawings
Fig. 1 is the method flow schematic diagram in one embodiment of the invention;
Fig. 2 is the system structure diagram in one embodiment of the invention;
Fig. 3 is neural network training process schematic diagram;
Fig. 4 is autoencoder network structure diagram;
Fig. 5 is the method flow schematic diagram in one embodiment of the invention;
Fig. 6 is the method flow schematic diagram in one embodiment of the present invention.
Embodiment
The principle of the disclosure is described referring now to some example embodiments.It is appreciated that these embodiments are merely for saying
It is bright and help it will be understood by those skilled in the art that with the purpose of the embodiment disclosure and describe, rather than suggest to the disclosure
Any restrictions of scope.Content of this disclosure described here can in a manner of described below outside various modes implement.
As described herein, term " comprising " and its various variations are construed as open-ended term, it means that
" including but not limited to ".Term "based" is construed as " being based at least partially on ".Term " one embodiment " can be by
It is interpreted as " at least one embodiment ".Term " another embodiment " is construed as " at least one other embodiment ".
Those skilled in the art can understand that the convolutional neural networks in the application are a kind of deep learning algorithms.
Those skilled in the art can understand that the autoencoder network in the application is only one layer of hiding node layer, input
With neutral net of the output with identical number of nodes.
Those skilled in the art can understand that the binaryzation Hash in the application is that the data that will be inquired about pass through Hash
Function Mapping is the binary sequence of certain length, for accelerating to search speed.
It is method flow schematic diagram in one embodiment of the invention to please refer to Fig.1, and a kind of face in the present embodiment is quick
Search method, includes the following steps:
Step S100 obtains the feature vector of image, and described eigenvector is inputted an autoencoder network, obtains image
Feature vector is firstly the need of training convolutional neural networks, and after the completion of convolutional neural networks training, note neural network parameter is to people
The mapping relations of face image and feature vector are Θ, utilize the feature of facial image in the network parameter extraction training set of acquisition
Vector, by taking intrinsic dimensionality is K=128 as an example.Remember training image Ii, (i=1,2 ..., Ntrain) character pair be Fi, (i
=1,2 ..., Ntrain).Then have:
Fi=Θ (Ii)
Step S101 obtains full articulamentum weight and corresponding bias term according to autoencoder network training and renewal,
And as the network parameter that feature vector is carried out to binaryzation Hash, specifically, when establishing autoencoder network, it can utilize upper
Loss function and 800,000 facial images are stated by 800,000 128 dimensional features that convolutional neural networks obtain to train
Autoencoder network, updates corresponding parameter W using stochastic gradient descent method1, b1And W2, b2After a fixed wheel number, mould is preserved
Shape parameter W1, b1, as the network parameter that feature is carried out to binaryzation Hash.
Step S102 establishes the hash index storehouse of image by the network parameter and obtains the Hash of image to be checked
Value, searches out face result.During using this method, it is divided into two parts of Index process and query process.
Please refer to Fig.4, as preferred in the present embodiment, in the autoencoder network, first by using loss function
And multidimensional characteristic trains autoencoder network, corresponding model parameter then is updated using stochastic gradient descent method, is finally protected
Model parameter is deposited, as the network parameter that feature is carried out to binaryzation Hash.
Step S41 face characteristics input Fi
Step S42 model parameters W1, b1
Step S43 is encoded into output hi
Step S44 renewal model parameters W1, b1
Step S45 decoding layers export F '
Step S46 sign functions
Step S47 binaryzation cryptographic Hash
Step S48 loss functions
As shown in Figure 4, in training process, the character pair of N images is randomly selected in training set every time as one
A trained batch, being input to the face characteristic inputted in autoencoder network is
W1, and W2The respectively weight of the full articulamentum of first layer and the full articulamentum of the second layer, b1,b2For corresponding bias term,
Activation primitive isIn training process, N is randomly selected every time and opens image
Character pair is input in autoencoder network as a trained batch.For a trained batch:
The output of first layer network is:
hi=g (W1·Fi+b1), i=1,2 ..., N. is whereinK '=64
The output of second layer network is:
WhereinK=128
For the output of the first layer network, the binary expression for obtaining corresponding cryptographic Hash is constrained using sign function:
bi=sgn (hi), i=1,2 ..., N. is whereinK '=64
Thus, by autoencoder network, 128 dimensional feature vectors of facial image are converted into the binaryzation Hash of 64bit
Value.The loss function of the autoencoder network is as follows:
Wherein, L1For the loss function of autoencoder network itself:
L2For the loss function expression of binaryzation Hash coding.B=[b1, b2..., bN] ∈ { -1,1 }K′×NNote, H=[h1,
h2..., hN]∈RK′×N, then L2It can be represented with following formula:
L2=J1-λ1J2+λ2J3+λ3J4
Wherein,
Distance of the purpose between optimization cryptographic Hash and actual characteristic value;
Purpose is to ensure hash function to the harmony of each
Relaxation orthogonal terms are added, ensure Hash
The independence of intersymbol.
For regularization term
Wherein, λ1, λ2, λ3Hyper parameter respectively corresponding to each several part optimal value.
By experiment, if α=1, λ1=50, λ2=0.001, λ3=0.0005, using min L as optimization aim, own coding
Network can train convergence.
Fig. 5 is refer to, as preferred in the present embodiment, the method in the present embodiment further includes:Establish Index process
Following steps, described eigenvector carry out binaryzation by autoencoder network and obtain cryptographic Hash, utilize the cryptographic Hash of above-mentioned acquisition
Establish the index of image library.
Fig. 5 is refer to, as preferred in the present embodiment, the method in the present embodiment further includes:Query process it is as follows
Step,
The cryptographic Hash of image to be checked is obtained,
The Hamming distance with Hash codes in index database is calculated,
Sort from low to high by the numerical value of Hamming distance, and export corresponding original image in order, obtain retrieval knot
Fruit, Hamming distance is smaller, and the similitude of retrieval is higher.
Fig. 5 is refer to, the step of it includes is:
Image in step S51 facial image databases
Step S52 needs the facial image inquired about
Step S53 is pre-processed
Convolutional neural networks are completed in step S54 training
Step S55 obtains 128 dimension face feature vectors
Step S56 autoencoder network parameters W1、b1,
Step S57 calculates Hamming distance using 64bit cryptographic Hash as index,
Step S58 forms index database by 64bit cryptographic Hash.
During using this method, it is divided into two parts of Index process and query process:In Index process, if common M in search library
Facial image is opened, feature will be extracted by convolutional neural networks after image preprocessing, i.e.,
Fi=Θ (Imagei), i=1,2 ..., M
Obtained feature is subjected to binaryzation by autoencoder network, obtains the cryptographic Hash of final 64bit, i.e.,:
Ci=sgn (g ((W1·Fi+b1))), i=1,2 ..., M
Utilize the cryptographic Hash C of above-mentioned acquisitioni, (i=1,2 ..., M) establishes the index of image library.In query process, according to
The old cryptographic Hash C that image to be checked is obtained by above-mentioned two stepprobe, calculate CprobeWith Hash codes C in index databaseiHamming
Distance.It is smaller by the numerical ordering of Hamming distance, Hamming distance, it was demonstrated that the similarity of image is higher.Export in order corresponding
Original image, obtains retrieval result.
Please refer to Fig.3, further include facial image in certain embodiments, before the feature vector for obtaining image locates in advance
Manage step:
If the amount of images for participating in the training set of training is Ntrain, for the image I of each inputi, (i=1,
2 ..., Ntrain),
Face datection is carried out to picture to mark with key point, and registration process is carried out to facial image according to key point afterwards,
Image is incorporated into as same scale.
As preferred in the present embodiment, following steps are further included after the feature vector for obtaining image:
By the facial image I ' after processingi, (i=1,2 ..., Ntrain) add convolutional neural networks be trained, obtain
The corresponding network parameter for face characteristic extraction,
After the completion of convolutional neural networks training, facial image and feature vector are carried out by the network parameter of acquisition
Mapping, extracts the feature vector of facial image in training set.
As shown in figure 3, it includes the following steps:
Step S31 training set face pictures
Step S32 Face datections are marked with key point
Step S33 face registration process
Step S34 convolutional neural networks
Step S35 network losses functions
In the present invention, substantial amounts of face image data is pre-processed first, afterwards add convolutional neural networks in into
Row training, the network parameter obtained using training obtain the feature vector of image.The network structure design ginseng of convolutional neural networks
Kaiming He, Xiangyu Zhang, Shaoqing Ren etc. have been examined in article " Deep Residual Learning for
Image Recognition, arXiv preprint arXiv:1512.03385 the residual error network structure proposed in 2015 ".
Training process such as Fig. 1 shows.
Original image is pre-processed first.If the amount of images for participating in the training set of training is Ntrain, for each
Open the image I of inputi, (i=1,2 ..., Ntrain), Face datection is carried out to picture and is marked with key point, afterwards according to key
Point carries out registration process to facial image, and image is incorporated into as same scale.By the facial image I ' after processingi, (i=1,
2 ..., Ntrain) add convolutional neural networks be trained, obtain corresponding network parameter, for face characteristic extract.
After the completion of convolutional neural networks training, mapping relations of the note neural network parameter to facial image and feature vector
For Θ, the feature vector of facial image in the network parameter extraction training set of acquisition, intrinsic dimensionality K=128 are utilized.Note instruction
Practice image Ii, (i=1,2 ..., Ntrain) character pair be Fi, (i 1,2 ..., Ntrain.Then have:
Fi=Θ (Ii)
It is that system structure diagram in one embodiment of the invention additionally provides one kind in the present embodiment as shown in Figure 2
Face quick retrieval system, including:Feature extraction unit 1 and feature retrieval unit 2,
The feature extraction unit 1, to obtain the feature vector of image, an own coding is inputted by described eigenvector
Network, full articulamentum weight and corresponding bias term are obtained according to autoencoder network training and renewal, and are used as feature
Vector carries out the network parameter of binaryzation Hash, and the feature retrieval unit 2, indexes to be established by the network parameter
Library searching goes out the face result of image to be checked.
In certain embodiments, the feature retrieval unit 1 includes:Index establishes unit and index database query unit,
The index establishes unit, and Hash is obtained described eigenvector is carried out binaryzation by autoencoder network
Value, the index of image library is established using the cryptographic Hash of above-mentioned acquisition,
The index database query unit, to obtain the cryptographic Hash of image to be checked, calculates and Hash codes in index database
Hamming distance, by the numerical ordering of Hamming distance, and exports corresponding original image, obtains retrieval result in order.
In certain embodiments, the feature extraction unit 1 further includes:Pretreatment unit, the pretreatment unit, is used
To carry out face normalization processing.
In certain embodiments, the feature extraction unit 1 includes:Convolutional neural networks and autoencoder network
Fig. 6 is refer to, is the method flow schematic diagram in one embodiment of the present invention, that includes:Feature extraction unit
Divide and characteristic key part.In characteristic extraction part, the facial image for having marked good person's face identity is pre-processed, it is right
Pretreated image combination identity label adds convolutional neural networks training, obtains network parameter to extract face characteristic.
Using network parameter obtained above, using facial image as inputting, the feature vector of correspondence image is obtained.In characteristic key
Part, by acquired image feature vector, adds in autoencoder network and is trained, and two-value is obtained by the output of network
The cryptographic Hash of change.In implementation phase, an input facial image is given, utilizes trained convolutional neural networks and own coding
Network obtains the binaryzation cryptographic Hash of image as index, and final similar facial image is obtained with reference to Hamming distance.Using upper
Technology is stated, improves accuracy rate and speed during face retrieval.
The a large amount of face pictures for including identity label of input,
Step S61 carries out image preprocessing,
Normalization facial image after step S62 alignment
Step S63 convolutional neural networks are trained
Step S64 extracts the network parameter of feature
The feature vector of step S65 facial images
Step S66 autoencoder networks are trained
Step S67 calculates the network parameter of binaryzation Hash
This part is the network parameter training stage.
Image in step S68 facial image databases
Step S69 inputs image to be retrieved
Step S610 carries out image preprocessing
Face exports after step S611 alignment
Step S612 proposes the network parameter of feature
The feature database of step S613 facial images
The face characteristic of step S614 images to be retrieved
Step S615 calculates the network parameter of binaryzation Hash
The index database that step S616 cryptographic Hash is established
The cryptographic Hash of step S617 images to be retrieved
Step S618 draws retrieval result
This part is test retrieval phase.
Specifically, process can be divided into altogether two stage statements to realizing in the present embodiment:The network parameter training stage
With test retrieval phase.The purpose of network parameter training stage is to obtain the convolutional Neural net for being used for extracting facial image feature
Network parameter and the autoencoder network parameter for obtaining image cryptographic Hash.Testing retrieval phase was obtained using the parameter training stage
The network parameter obtained, establishes the cryptographic Hash index database of face and the cryptographic Hash of picture to be retrieved, calculates Hamming distance and simultaneously sorts
Afterwards, the result retrieved.
For the network parameter training stage, a large amount of facial images comprising face identity label are pre-processed first.
If the amount of images for participating in the training set of training is Ntrain, for the image I of each inputi, (i=1,2 ..., Ntrain),
Face datection is carried out to picture to mark with key point, and registration process is carried out to facial image according to key point afterwards, image is drawn
It is classified as same scale.By the facial image I ' after processingi, (i=1,2 ..., Ntrain) add convolutional neural networks instructed
Practice, obtain corresponding network parameter, extracted for face characteristic.
After the completion of convolutional neural networks training, mapping relations of the note neural network parameter to facial image and feature vector
For Θ, the feature vector of facial image in the network parameter extraction training set of acquisition, intrinsic dimensionality K=128 are utilized.Note instruction
Practice image Ii, (i=1,2 ..., Ntrain) character pair be Fi, (i 1,2 ..., Ntrain).Then have:
Fi=Θ (Ii)
In training process, the character pairs of N images is randomly selected in training set every time as a trained batch,
Being input to the face characteristic inputted in autoencoder network is W1, and W2
The respectively weight of the full articulamentum of first layer and the full articulamentum of the second layer, b1,b2For corresponding bias term, activation primitive isIn training process, the character pair of N images is randomly selected every time as one
Training batch is input in autoencoder network.For a trained batch:
The output of first layer network is:
hi=g (W1·Fi+b1), i=1,2 ..., N. is wherein
The output of second layer network is:
Wherein
For the output of the first layer network, the binary expression for obtaining corresponding cryptographic Hash is constrained using sign function:
bi=sgn (hi), i=1,2 ..., N. is wherein
Thus, by autoencoder network, 128 dimensional feature vectors of facial image are converted into the binaryzation Hash of 64bit
Value.The loss function of the autoencoder network is as follows:
Wherein, L1For the loss function of autoencoder network itself:
L2For the loss function expression of binaryzation Hash coding.B=[b1, b2..., bN] ∈ { -1,1 }K′×NNote, H=[h1,
h2..., hN]∈RK′×N, then L2It can be represented with following formula:
L2=J1-λ1J2+λ2J3+λ3J4
Wherein,
Distance of the purpose between optimization cryptographic Hash and actual characteristic value;
Purpose is to ensure hash function to the harmony of each,
Relaxation orthogonal terms are added, ensure Hash
The independence of intersymbol.
For regularization term
Wherein, λ1, λ2, λ3Hyper parameter respectively corresponding to each several part optimal value.
By experiment, if α=1, λ1=50, λ2=0.001, λ3=0.0005, using min L as optimization aim, own coding
Network can train convergence.
Pass through 800,000 of convolutional neural networks acquisition using above-mentioned loss function and 800,000 facial images
128 dimensional features train autoencoder network, use stochastic gradient descent method to update corresponding parameter W1, b1And W2, b2One is passed through
After fixed wheel number, preservation model parameter W1, b1, as the network parameter that feature is carried out to binaryzation Hash.
In test retrieval phase, it can be divided into and establish two parts of index database process and query process:Establish index database process
In, if common M facial images in search library, feature will be extracted by convolutional neural networks after image preprocessing, i.e.,
Fi=Θ (Imagei), i=1,2, M
Obtained feature is subjected to binaryzation by autoencoder network, obtains the cryptographic Hash of final 64bit, i.e.,:
Ci=sgn (g ((W1·Fi+b1))), i=1,2 ..., M
Utilize the cryptographic Hash C of above-mentioned acquisitioni, (i=1,2 ..., M) establishes the index of image library.In query process, according to
The old cryptographic Hash C that image to be checked is obtained by above-mentioned two stepprobe, calculate CprobeWith Hash codes C in index databaseiHamming
Distance.It is smaller by the numerical ordering of Hamming distance, Hamming distance, it was demonstrated that the similarity of image is higher.Export in order corresponding
Original image, obtains retrieval result.
It should be appreciated that each several part of the present invention can be realized with hardware, software, firmware or combinations thereof.Above-mentioned
In embodiment, multiple steps or method can be performed soft in memory and by suitable instruction execution system with storage
Part or firmware are realized.If, and in another embodiment, can be with well known in the art for example, realized with hardware
Any one of following technology or their combination are realized:With the logic gate for realizing logic function to data-signal
The discrete logic of circuit, the application-specific integrated circuit with suitable combinational logic gate circuit, programmable gate array (PGA),
Field programmable gate array (FPGA) etc..
In the description of this specification, reference term " one embodiment ", " some embodiments ", " example ", " specifically show
The description of example " or " some examples " etc. means specific features, structure, material or the spy for combining the embodiment or example description
Point is contained at least one embodiment of the present invention or example.In the present specification, schematic expression of the above terms is not
Necessarily refer to identical embodiment or example.Moreover, particular features, structures, materials, or characteristics described can be any
One or more embodiments or example in combine in an appropriate manner.
In general, the various embodiments of the disclosure can be with hardware or special circuit, software, logic or its any group
Close and implement.Some aspects can be implemented with hardware, and some other aspect can be with firmware or software implementation, the firmware or soft
Part can be by controller, microprocessor or other computing devices.Although the various aspects of the disclosure are shown and described as
Block diagram, flow chart or using some other drawing represent, but it is understood that frame described herein, equipment, system, technology or
Method can in a non limiting manner with hardware, software, firmware, special circuit or logic, common hardware or controller or its
Its computing device or some combinations are implemented.
Although in addition, operation is described with particular order, this is understood not to require this generic operation with shown suitable
Sequence is performed or performed with generic sequence, or requires all shown operations to be performed to realize expected result.In some feelings
Under shape, multitask or parallel processing can be favourable.Similarly, although the details of some specific implementations is above
By comprising but these are not necessarily to be construed as any restrictions to the scope of the present disclosure, but the description of feature is only in discussion
For specific embodiment.Some features described in some separated embodiments can also be in single embodiment in combination
Perform.Mutually oppose, various features described in single embodiment can also be implemented separately in various embodiments or with
The mode of any suitable sub-portfolio is implemented.
Claims (10)
1. a kind of face method for quickly retrieving, it is characterised in that include the following steps:
The feature vector of image is obtained, described eigenvector is inputted into an autoencoder network,
According to autoencoder network training and update and obtain full articulamentum weight and corresponding bias term, and as by feature to
Amount carries out the network parameter of binaryzation Hash,
The hash index storehouse of image is established by the network parameter and obtains the cryptographic Hash of image to be checked, searches out face
As a result.
2. face method for quickly retrieving according to claim 1, it is characterised in that in the autoencoder network,
Autoencoder network is trained by using loss function and multidimensional characteristic first,
Then corresponding model parameter is updated using stochastic gradient descent method,
Last preservation model parameter, as the network parameter that feature is carried out to binaryzation Hash.
3. face method for quickly retrieving according to claim 1, it is characterised in that further include:Establish the Hash rope of image
Draw the following steps of storehouse process,
Described eigenvector carries out binaryzation by autoencoder network and obtains cryptographic Hash, and figure is established using the cryptographic Hash of above-mentioned acquisition
As the index in storehouse.
4. the face method for quickly retrieving according to claim 1 or 3, it is characterised in that further include:
The cryptographic Hash of image to be checked is obtained,
The Hamming distance with Hash codes in index database is calculated,
By the numerical ordering of Hamming distance, and corresponding original image is exported in order, obtain retrieval result.
5. face method for quickly retrieving according to claim 1, it is characterised in that before the feature vector for obtaining image
Further include facial image pre-treatment step:
If the amount of images for participating in the training set of training is Ntrain, for the image I of each inputi, (i=1,2 ...,
Ntrain),
Face datection is carried out to picture to mark with key point, and registration process is carried out to facial image according to key point afterwards, will be schemed
As incorporating into as same scale.
6. face method for quickly retrieving according to claim 5, it is characterised in that after the feature vector for obtaining image
Further include following steps:
By the facial image I ' after processingi, (i=1,2 ..., Ntrain) add convolutional neural networks and be trained, obtain corresponding
Be used for face characteristic extraction network parameter,
After the completion of convolutional neural networks training, facial image and feature vector are mapped by the network parameter of acquisition,
Extract the feature vector of facial image in training set.
A kind of 7. face quick retrieval system, it is characterised in that including:Feature extraction unit and feature retrieval unit,
The feature extraction unit, to obtain the feature vector of image, an autoencoder network is inputted by described eigenvector,
According to autoencoder network training and update and obtain full articulamentum weight and corresponding bias term, and as by feature to
Amount carries out the network parameter of binaryzation Hash,
The feature retrieval unit, to establish the face knot that index database searches out image to be checked by the network parameter
Fruit.
8. face quick retrieval system according to claim 7, it is characterised in that the feature retrieval unit includes:Rope
Draw and establish unit and index database query unit,
The index establishes unit, obtains cryptographic Hash described eigenvector is carried out binaryzation by autoencoder network, profit
The index of image library is established with the cryptographic Hash of above-mentioned acquisition,
The index database query unit, to obtain the cryptographic Hash of image to be checked, calculates the Hamming with Hash codes in index database
Distance, by the numerical ordering of Hamming distance, and exports corresponding original image, obtains retrieval result in order.
9. face quick retrieval system according to claim 7, it is characterised in that the feature extraction unit further includes:
Pretreatment unit, the pretreatment unit, to carry out face normalization processing.
10. face quick retrieval system according to claim 7, it is characterised in that the feature extraction unit includes:Volume
Product neutral net and autoencoder network.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710801518.9A CN107918636B (en) | 2017-09-07 | 2017-09-07 | Face quick retrieval method and system |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710801518.9A CN107918636B (en) | 2017-09-07 | 2017-09-07 | Face quick retrieval method and system |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107918636A true CN107918636A (en) | 2018-04-17 |
CN107918636B CN107918636B (en) | 2021-05-18 |
Family
ID=61898756
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710801518.9A Active CN107918636B (en) | 2017-09-07 | 2017-09-07 | Face quick retrieval method and system |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107918636B (en) |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107958247A (en) * | 2018-01-17 | 2018-04-24 | 百度在线网络技术(北京)有限公司 | Method and apparatus for facial image identification |
CN108830559A (en) * | 2018-06-01 | 2018-11-16 | 广东蔚海数问大数据科技有限公司 | A kind of Work attendance method and device based on recognition of face |
CN108921911A (en) * | 2018-08-01 | 2018-11-30 | 中国科学技术大学 | The method that structuring picture is automatically converted to source code |
CN109145132A (en) * | 2018-07-12 | 2019-01-04 | 北京奇虎科技有限公司 | The method, apparatus and image search method, device of Hash codes are extracted from image |
CN109146924A (en) * | 2018-07-18 | 2019-01-04 | 北京飞搜科技有限公司 | A kind of method for tracking target and device based on thermodynamic chart |
CN109241315A (en) * | 2018-08-28 | 2019-01-18 | 电子科技大学 | A kind of fast face search method based on deep learning |
CN109325140A (en) * | 2018-07-12 | 2019-02-12 | 北京奇虎科技有限公司 | The method, apparatus and image search method, device of Hash codes are extracted from image |
CN109359516A (en) * | 2018-08-31 | 2019-02-19 | 算丰科技(北京)有限公司 | Data processing method, data processing equipment and electronic equipment |
CN109918528A (en) * | 2019-01-14 | 2019-06-21 | 北京工商大学 | A kind of compact Hash code learning method based on semanteme protection |
CN110895578A (en) * | 2018-09-13 | 2020-03-20 | 武汉真元生物数据有限公司 | Quick retrieval method, equipment and system based on biological characteristics |
CN112424797A (en) * | 2018-05-17 | 2021-02-26 | 弗劳恩霍夫应用研究促进协会 | Concept for the transmission of distributed learning of neural networks and/or parametric updates thereof |
CN113343020A (en) * | 2021-08-06 | 2021-09-03 | 腾讯科技(深圳)有限公司 | Image processing method and device based on artificial intelligence and electronic equipment |
TWI759731B (en) * | 2020-04-27 | 2022-04-01 | 淡江大學 | Machine learning method |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104834748A (en) * | 2015-05-25 | 2015-08-12 | 中国科学院自动化研究所 | Image retrieval method utilizing deep semantic to rank hash codes |
CN105069173A (en) * | 2015-09-10 | 2015-11-18 | 天津中科智能识别产业技术研究院有限公司 | Rapid image retrieval method based on supervised topology keeping hash |
CN106445939A (en) * | 2015-08-06 | 2017-02-22 | 阿里巴巴集团控股有限公司 | Image retrieval, image information acquisition and image identification methods and apparatuses, and image identification system |
CN106778526A (en) * | 2016-11-28 | 2017-05-31 | 中通服公众信息产业股份有限公司 | A kind of extensive efficient face identification method based on Hamming distance |
-
2017
- 2017-09-07 CN CN201710801518.9A patent/CN107918636B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN104834748A (en) * | 2015-05-25 | 2015-08-12 | 中国科学院自动化研究所 | Image retrieval method utilizing deep semantic to rank hash codes |
CN106445939A (en) * | 2015-08-06 | 2017-02-22 | 阿里巴巴集团控股有限公司 | Image retrieval, image information acquisition and image identification methods and apparatuses, and image identification system |
CN105069173A (en) * | 2015-09-10 | 2015-11-18 | 天津中科智能识别产业技术研究院有限公司 | Rapid image retrieval method based on supervised topology keeping hash |
CN106778526A (en) * | 2016-11-28 | 2017-05-31 | 中通服公众信息产业股份有限公司 | A kind of extensive efficient face identification method based on Hamming distance |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107958247A (en) * | 2018-01-17 | 2018-04-24 | 百度在线网络技术(北京)有限公司 | Method and apparatus for facial image identification |
CN112424797A (en) * | 2018-05-17 | 2021-02-26 | 弗劳恩霍夫应用研究促进协会 | Concept for the transmission of distributed learning of neural networks and/or parametric updates thereof |
CN108830559A (en) * | 2018-06-01 | 2018-11-16 | 广东蔚海数问大数据科技有限公司 | A kind of Work attendance method and device based on recognition of face |
CN109325140B (en) * | 2018-07-12 | 2021-07-13 | 北京奇虎科技有限公司 | Method and device for extracting hash code from image and image retrieval method and device |
CN109145132A (en) * | 2018-07-12 | 2019-01-04 | 北京奇虎科技有限公司 | The method, apparatus and image search method, device of Hash codes are extracted from image |
CN109145132B (en) * | 2018-07-12 | 2021-06-18 | 北京奇虎科技有限公司 | Method and device for extracting hash code from image and image retrieval method and device |
CN109325140A (en) * | 2018-07-12 | 2019-02-12 | 北京奇虎科技有限公司 | The method, apparatus and image search method, device of Hash codes are extracted from image |
CN109146924B (en) * | 2018-07-18 | 2020-09-08 | 苏州飞搜科技有限公司 | Target tracking method and device based on thermodynamic diagram |
CN109146924A (en) * | 2018-07-18 | 2019-01-04 | 北京飞搜科技有限公司 | A kind of method for tracking target and device based on thermodynamic chart |
CN108921911B (en) * | 2018-08-01 | 2021-03-09 | 中国科学技术大学 | Method for automatically converting structured picture into source code |
CN108921911A (en) * | 2018-08-01 | 2018-11-30 | 中国科学技术大学 | The method that structuring picture is automatically converted to source code |
CN109241315A (en) * | 2018-08-28 | 2019-01-18 | 电子科技大学 | A kind of fast face search method based on deep learning |
CN109241315B (en) * | 2018-08-28 | 2022-01-04 | 电子科技大学 | Rapid face retrieval method based on deep learning |
CN109359516A (en) * | 2018-08-31 | 2019-02-19 | 算丰科技(北京)有限公司 | Data processing method, data processing equipment and electronic equipment |
CN110895578A (en) * | 2018-09-13 | 2020-03-20 | 武汉真元生物数据有限公司 | Quick retrieval method, equipment and system based on biological characteristics |
CN109918528A (en) * | 2019-01-14 | 2019-06-21 | 北京工商大学 | A kind of compact Hash code learning method based on semanteme protection |
TWI759731B (en) * | 2020-04-27 | 2022-04-01 | 淡江大學 | Machine learning method |
CN113343020A (en) * | 2021-08-06 | 2021-09-03 | 腾讯科技(深圳)有限公司 | Image processing method and device based on artificial intelligence and electronic equipment |
CN113343020B (en) * | 2021-08-06 | 2021-11-26 | 腾讯科技(深圳)有限公司 | Image processing method and device based on artificial intelligence and electronic equipment |
Also Published As
Publication number | Publication date |
---|---|
CN107918636B (en) | 2021-05-18 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107918636A (en) | A kind of face method for quickly retrieving, system | |
CN110689086B (en) | Semi-supervised high-resolution remote sensing image scene classification method based on generating countermeasure network | |
CN107330451B (en) | Clothing attribute retrieval method based on deep convolutional neural network | |
Yang et al. | Efficient image retrieval via decoupling diffusion into online and offline processing | |
CN107122809A (en) | Neural network characteristics learning method based on image own coding | |
CN108960059A (en) | A kind of video actions recognition methods and device | |
CN106682233A (en) | Method for Hash image retrieval based on deep learning and local feature fusion | |
CN107515895A (en) | A kind of sensation target search method and system based on target detection | |
CN109190643A (en) | Based on the recognition methods of convolutional neural networks Chinese medicine and electronic equipment | |
CN106354735A (en) | Image target searching method and device | |
CN104933428B (en) | A kind of face identification method and device based on tensor description | |
CN107506786A (en) | A kind of attributive classification recognition methods based on deep learning | |
CN104281572B (en) | A kind of target matching method and its system based on mutual information | |
CN107291825A (en) | With the search method and system of money commodity in a kind of video | |
KR20200071838A (en) | Face recognition method and apparatus capable of face search using feature vector | |
Altintakan et al. | Towards effective image classification using class-specific codebooks and distinctive local features | |
CN112733602B (en) | Relation-guided pedestrian attribute identification method | |
CN108268890A (en) | A kind of hyperspectral image classification method | |
Buvana et al. | Content-based image retrieval based on hybrid feature extraction and feature selection technique pigeon inspired based optimization | |
KR20210033940A (en) | How to Train Neural Networks for Human Facial Recognition | |
CN107967441A (en) | A kind of video behavior recognition methods based on binary channels 3D-2D RBM models | |
Wang et al. | Hierarchical image segmentation ensemble for objectness in RGB-D images | |
CN108446605A (en) | Double interbehavior recognition methods under complex background | |
CN109919320B (en) | Triplet network learning method based on semantic hierarchy | |
Jia et al. | 2DPCANet: dayside aurora classification based on deep learning |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
TA01 | Transfer of patent application right |
Effective date of registration: 20210426 Address after: 215123 unit 2-b702, creative industry park, No. 328, Xinghu street, Suzhou Industrial Park, Suzhou City, Jiangsu Province Applicant after: SUZHOU FEISOU TECHNOLOGY Co.,Ltd. Address before: 100082 Beijing city Haidian District Xitucheng Road 10, High-Tech Mansion BUPT, 1209 Applicant before: BEIJING FEISOU TECHNOLOGY Co.,Ltd. |
|
TA01 | Transfer of patent application right | ||
GR01 | Patent grant | ||
GR01 | Patent grant |