CN109934197A - Training method, device and the computer readable storage medium of human face recognition model - Google Patents
Training method, device and the computer readable storage medium of human face recognition model Download PDFInfo
- Publication number
- CN109934197A CN109934197A CN201910219576.XA CN201910219576A CN109934197A CN 109934197 A CN109934197 A CN 109934197A CN 201910219576 A CN201910219576 A CN 201910219576A CN 109934197 A CN109934197 A CN 109934197A
- Authority
- CN
- China
- Prior art keywords
- human face
- face recognition
- recognition model
- feature
- training
- 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
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
- Y02T10/00—Road transport of goods or passengers
- Y02T10/10—Internal combustion engine [ICE] based vehicles
- Y02T10/40—Engine management systems
Abstract
The invention discloses a kind of training methods of human face recognition model, the training method of the human face recognition model is the following steps are included: carry out feature extraction to each target facial image using each target network-layer in depth convolutional neural networks, to obtain multiple global characteristics；The global characteristics are split according to the global characteristics corresponding target network-layer to obtain each local feature, and the corresponding each global characteristics of the target facial image are merged to obtain fusion feature；Loss supervised learning is carried out to the global characteristics, the local feature and fusion feature, to be trained to human face recognition model.Invention additionally discloses a kind of training device of human face recognition model and computer readable storage mediums.The present invention improves the precision of recognition of face.
Description
Technical field
The present invention relates to technical field of face recognition more particularly to a kind of training method of human face recognition model, device and
Computer readable storage medium.
Background technique
By means of power in the maturation and the high release for calculating power equipment of depth convolutional neural networks technology, face recognition algorithms are either
The huge revolution all occurred in progress and recognition speed.Face recognition application also becomes most mature widest in AI technology
It represents.Either smart city construction or field of video monitoring, recognition of face become most basic most crucial technological means.
Face recognition technology is to be not required to very important person as a kind of maximum advantage of biological identification technology of contactless unaware
Forced mating can capture the identity information of people.Face identification system is can take the front face of cleaning as base
Plinth, and under actual scene condition, when illumination, shooting angle, mask of wearing glasses, cause the precision of recognition of face compared with
It is low.
Summary of the invention
The main purpose of the present invention is to provide a kind of training method of human face recognition model, device and computer-readable deposit
Storage media, it is intended to solve the problems, such as that the precision of recognition of face is lower.
To achieve the above object, the training method of a kind of human face recognition model provided by the invention, the recognition of face mould
The training method of type the following steps are included:
Feature extraction is carried out to each target facial image using each target network-layer in depth convolutional neural networks, with
Obtain multiple global characteristics；
The global characteristics are split according to the global characteristics corresponding target network-layer to obtain each part spy
Sign, and the corresponding each global characteristics of the target facial image are merged to obtain fusion feature；
Loss supervised learning is carried out to the global characteristics, the local feature and fusion feature, to recognition of face
Model is trained.
In one embodiment, described that loss supervision is carried out to the global characteristics, the local feature and fusion feature
The step of study includes:
Determine the global characteristics, the local feature and the corresponding target loss function of the fusion feature and institute
State the corresponding weighted value of target loss function；
According to the target loss function and the corresponding weighted value of the target loss function, to the objective function pair
The feature answered carries out loss supervised learning.
In one embodiment, the determination global characteristics, the local feature and the fusion feature are corresponding
The step of target loss function and the target loss function corresponding weighted value includes:
Normalization figure penalties function is determined as the corresponding first object loss function of the local feature, and will be described
It normalizes figure penalties function and triple loss function is determined as the global characteristics and the fusion feature is corresponding
Second target loss function；
The first weighted value and the corresponding second target damage of the fusion feature are configured for the first object loss function
It loses normalization exponential function in function and configures the second weighted value, and according to the corresponding target network-layer of the global characteristics to described
Exponential function is normalized in the corresponding second target loss function of global characteristics configures third weighted value.
In one embodiment, using each target network-layer in depth convolutional neural networks to each target facial image into
Row feature extraction, to obtain multiple global characteristics the step of include:
Multiple target network-layers are determined in each network layer of the depth convolutional neural networks, wherein the target
Distance is less than other network layers except the target network-layer between network layer and the output layer of the depth convolutional neural networks
The distance between described output layer；
Target facial image is inputted into each target network-layer, it is corresponding complete to obtain each target network-layer
Office's characteristic pattern；
Pond is carried out to each global characteristics figure, obtains multiple global characteristics.
In one embodiment, described the step of carrying out loss supervised learning to the fusion feature, includes:
Reduce the dimension of the fusion feature；
Loss supervised learning is carried out to the fusion feature for reducing dimension.
In one embodiment, it is described using each target network-layer in depth convolutional neural networks to each target face figure
Before the step of carrying out feature extraction, further includes:
Each test facial image is obtained, and each test facial image is pre-processed, to obtain each target person
Face image, wherein the pretreatment mass image normalization processes pixel and image data enhancing processing.
In one embodiment, described that loss prison is being carried out to the global characteristics, the local feature and fusion feature
After the step of educational inspector practises, further includes:
The corresponding loss function of the depth convolutional neural networks is obtained in real time；
When the corresponding loss function of the depth convolutional neural networks converges to set interval, the recognition of face is completed
The training of model, and save the human face recognition model for completing training.
In one embodiment, after described the step of saving the human face recognition model for completing training, further includes:
The fusion feature for the human face recognition model output for completing training is prestored as the facial image is corresponding
Fusion feature, and fusion feature is prestored described in preservation.
To achieve the above object, the present invention also provides a kind of training device of human face recognition model, the recognition of face moulds
The training device of type includes processor, memory and is stored in the face that can be run on the memory and on the processor
The training program of identification model is realized as described above when the training program of the human face recognition model is executed by the processor
Each step of the training method of human face recognition model.
To achieve the above object, the present invention also provides a kind of computer readable storage medium, the computer-readable storages
Media storage has the training program of human face recognition model, when the training program of the human face recognition model is executed by the processor
Realize each step of the training method of human face recognition model as described above.
Training method, device and the computer readable storage medium of human face recognition model provided by the invention, recognition of face
The training device of model carries out feature to each target facial image using each target network-layer in depth convolutional neural networks
Extraction obtains multiple global characteristics, is split further according to the corresponding target network-layer of global characteristics to the overall situation, to obtain more
A local feature, and then multiple local features of a corresponding facial image and multiple global characteristics are merged and melted
Feature is closed, loss supervised learning finally is carried out to carry out recognition of face to each global characteristics, local feature and fusion feature
The training of model；Due to human face recognition model training device by heterogeneous networks layer export global characteristics, fusion feature with
And the local feature of global characteristics segmentation carries out loss supervised learning, so that device has excavated the subtle spy in facial image
Sign, improves the precision of recognition of face, further, since global characteristics are divided into multiple offices by the training device of human face recognition model
Portion's feature carries out loss supervised learning, without manually blocking the occlusion area of face or to the characteristic processing of face blocked,
Simplify the training process of facial image.
Detailed description of the invention
Fig. 1 be the present embodiments relate to human face recognition model training device hardware structural diagram；
Fig. 2 is the flow diagram of the training method first embodiment of human face recognition model of the present invention；
Fig. 3 is the refinement flow diagram of step S10 in Fig. 2；
Fig. 4 is the refinement flow diagram of step S30 in Fig. 2；
Fig. 5 is the flow diagram of the training method second embodiment of human face recognition model of the present invention.
The embodiments will be further described with reference to the accompanying drawings for the realization, the function and the advantages of the object of the present invention.
Specific embodiment
It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not intended to limit the present invention.
The primary solutions of the embodiment of the present invention are: using each target network-layer in depth convolutional neural networks to each
A target facial image carries out feature extraction, to obtain multiple global characteristics；According to the corresponding target network of the global characteristics
Layer is split the global characteristics to obtain each local feature, and the target facial image is corresponding each described complete
Office's feature is merged to obtain fusion feature；Loss prison is carried out to the global characteristics, the local feature and fusion feature
Educational inspector practises, to be trained to human face recognition model.
Since the training device of human face recognition model is by exporting global characteristics, fusion feature and complete to heterogeneous networks layer
The local feature of office's Image Segmentation Methods Based on Features carries out loss supervised learning, so that device has excavated the fine feature in facial image, mentions
The high precision of recognition of face, further, since that global characteristics are divided into multiple parts is special for the training device of human face recognition model
Sign carries out loss supervised learning, without manually blocking the occlusion area of face or to the characteristic processing of face blocked, simplifies
The training process of facial image.
As a kind of implementation, the training device of human face recognition model can be as shown in Figure 1.
What is involved is the training device of human face recognition model, the training devices of human face recognition model for the embodiment of the present invention
It include: processor 101, such as CPU, memory 102, communication bus 103.Wherein, communication bus 103 is for realizing these components
Between connection communication.In addition, the training device of human face recognition model includes CPU, it is responsible for executing the training of human face recognition model
Program and the storage for carrying out reading data and model；In addition, the training device of human face recognition model includes GPU (graphics process
Device), GPU is used for the training of model；After the training program that CPU calls the human face recognition model stored in memory, CPU is from originally
A collection of training image file decoding is read in local disk and copies GPU video memory at training data format, and in GPU before progress
To calculating and backward reasoning, and model parameter is updated, reads next batch data, constantly repeatedly.
Memory 102 can be high speed RAM memory, be also possible to stable memory (non-
), such as magnetic disk storage volatilememory.As shown in Figure 1, as in a kind of memory 102 of computer storage medium
It may include the training program of human face recognition model；And processor 101 can be used for that the face stored in memory 102 is called to know
The training program of other model, and execute following operation:
Feature extraction is carried out to each target facial image using each target network-layer in depth convolutional neural networks, with
Obtain multiple global characteristics；
The global characteristics are split according to the global characteristics corresponding target network-layer to obtain each part spy
Sign, and the corresponding each global characteristics of the target facial image are merged to obtain fusion feature；
Loss supervised learning is carried out to the global characteristics, the local feature and fusion feature, to recognition of face
Model is trained.
In one embodiment, processor 101 can be used for calling the training of the human face recognition model stored in memory 102
Program, and execute following operation:
Determine the global characteristics, the local feature and the corresponding target loss function of the fusion feature and institute
State the corresponding weighted value of target loss function；
According to the target loss function and the corresponding weighted value of the target loss function, to the objective function pair
The feature answered carries out loss supervised learning.
In one embodiment, processor 101 can be used for calling the training of the human face recognition model stored in memory 102
Program, and execute following operation:
Normalization figure penalties function is determined as the corresponding first object loss function of the local feature, and will be described
It normalizes figure penalties function and triple loss function is determined as the global characteristics and the fusion feature is corresponding
Second target loss function；
The first weighted value and the corresponding second target damage of the fusion feature are configured for the first object loss function
It loses normalization exponential function in function and configures the second weighted value, and according to the corresponding target network-layer of the global characteristics to described
Exponential function is normalized in the corresponding second target loss function of global characteristics configures third weighted value.
In one embodiment, processor 101 can be used for calling the training of the human face recognition model stored in memory 102
Program, and execute following operation:
Multiple target network-layers are determined in each network layer of the depth convolutional neural networks, wherein the target
Distance is less than other network layers except the target network-layer between network layer and the output layer of the depth convolutional neural networks
The distance between described output layer；
Target facial image is inputted into each target network-layer, it is corresponding complete to obtain each target network-layer
Office's characteristic pattern；
Pond is carried out to each global characteristics figure, obtains multiple global characteristics.
In one embodiment, processor 101 can be used for calling the training of the human face recognition model stored in memory 102
Program, and execute following operation:
Reduce the dimension of the fusion feature；
Loss supervised learning is carried out to the fusion feature for reducing dimension.
In one embodiment, processor 101 can be used for calling the training of the human face recognition model stored in memory 102
Program, and execute following operation:
Each test facial image is obtained, and each test facial image is pre-processed, to obtain each target person
Face image, wherein the pretreatment mass image normalization processes pixel and image data enhancing processing.
In one embodiment, processor 101 can be used for calling the training of the human face recognition model stored in memory 102
Program, and execute following operation:
The corresponding loss function of the depth convolutional neural networks is obtained in real time；
When the corresponding loss function of the depth convolutional neural networks converges to set interval, the recognition of face is completed
The training of model, and save the human face recognition model for completing training.
In one embodiment, processor 101 can be used for calling the training of the human face recognition model stored in memory 102
Program, and execute following operation:
The fusion feature for the human face recognition model output for completing training is prestored as the facial image is corresponding
Fusion feature, and fusion feature is prestored described in preservation.
The present embodiment is according to above scheme, and the training device of human face recognition model is using each in depth convolutional neural networks
Target network-layer carries out feature extraction to each target facial image and obtains multiple global characteristics, corresponding further according to global characteristics
Target network-layer is split the overall situation, thus the multiple offices for obtaining multiple local features, and then corresponding to a facial image
Portion's feature and multiple global characteristics are merged to obtain fusion feature, finally to each global characteristics, local feature and are melted
It closes feature and carries out loss supervised learning to carry out the training of human face recognition model；Since the training device of human face recognition model passes through
Loss supervised learning is carried out to heterogeneous networks layer output global characteristics, fusion feature and the local feature of global characteristics segmentation,
So that device has excavated the fine feature in facial image, the precision of recognition of face is improved, further, since recognition of face mould
Global characteristics are divided into multiple local features and carry out loss supervised learning by the training device of type, without manually blocking face
Occlusion area or the characteristic processing blocked to face, simplify the training process of facial image.
The hardware architecture of training device based on above-mentioned human face recognition model proposes the training of human face recognition model of the present invention
The embodiment of method.
It is the first embodiment of the training method of human face recognition model of the present invention, the recognition of face mould referring to Fig. 2, Fig. 2
The training method of type the following steps are included:
Step S10 carries out each target facial image using each target network-layer in depth convolutional neural networks special
Sign is extracted, to obtain multiple global characteristics；
In the present invention, executing subject is the training device of human face recognition model, the training device packet of human face recognition model
CPU, memory and GPU have been included, memory is stored with the training program of human face recognition model, in addition, human face recognition model
Training device is additionally provided with the biggish memory of amount of storage (for example, local disk), which is stored with a large amount of facial image,
For example, memory can store the 30G even facial image for training of larger capacity；CPU is responsible for executing recognition of face mould
The training program of type, and GPU is then the training for model.
When being trained to model, it is multiple to carry out that the face image data in local disk is divided into multiple batches
Model training, the process per a batch of model training is consistent, specifically, CPU calls the recognition of face mould stored in memory
The training program of type, CPU read a certain amount of face image data for training from local disk again, and CPU is again to face
Image data decoding is at training data format, then is copied in GPU video memory, and in GPU carry out forward calculation and after to
Reasoning, and update and preservation model.It is understood that the face that the training device of human face recognition model passes through different batches
Image constantly model is trained repeatedly.
The training device of human face recognition model be equipped with optimizer, device by depth convolutional neural networks to facial image into
Row processing, specifically, referring to figure 3., i.e. step S10 includes:
Step S11 determines multiple target network-layers in each network layer of the depth convolutional neural networks, wherein
Distance is less than its for removing the target network-layer between the target network-layer and the output layer of the depth convolutional neural networks
The distance between its network layer and the output layer；
Target facial image is inputted each target network-layer, to obtain each target network-layer by step S12
Corresponding global characteristics figure；
Step S13 carries out pond to each global characteristics figure, obtains multiple global characteristics.
Depth convolutional neural networks contain multiple network layers, each network layer can carry out feature extraction to facial image,
But the feature that the network layer closer with inputting in depth convolutional neural networks is extracted is more shallow, and therefore, in the present invention, depth is rolled up
Closer several network layers are exported in product neural network as target network-layer namely device only needs to extract facial image each
The feature of target network-layer, it is to be understood that (output layer is deep to output layer in target network-layer and depth convolutional neural networks
The network layer that is output to the outside of degree convolutional neural networks) the distance between be less than other network layers and output except target network-layer
ResNet50 can be used as backbone network in the distance between layer, depth convolutional neural networks, ResNet50 can main decomposition at 5
Network layer, according to network layer and output layer distance sequence from big to small be followed successively by layer0, layer1, layer2, layer3,
Layer4, wherein can be using layer2, layer3, layer4 as target network-layer, and layer0 can be to the facial image of input
The operation such as convolution, normalization, pond is carried out, and layer1 can be used as the extractor of the characteristic pattern of facial image, need to illustrate
Be layer1 to layer4 be four stage, separately include 3,4,6,3 residual error structures, in addition, ResNet50 layer4 it
After further include pond layer and full articulamentum, it is to be understood that layer2, layer3, layer4 can be considered as three levels
Different network structures, and then extractable facial image, in the corresponding feature of these three target network-layers, this category feature is the overall situation
Feature, the face information of different global characteristics characterization different levels.It should be noted that inputting target network when facial image
After network layers, target network-layer exports global characteristics figure, and device carries out pond to global characteristics figure again, and then obtains global characteristics,
The size of the characteristic pattern exported from layer2, layer3, layer4 is respectively 512*28*28,1024*14*14 and 2048*7*
7.The dimension of three global characteristics is respectively 512,1024 and 2048.
It should be noted that needing to pre-process facial image, specifically, obtaining before to facial image training
The data set in model training is taken, multiple test facial images are contained in data set, facial image can be from face image data
Downloading obtains in source；Will in data set test facial image be trained when, processes pixel is normalized to facial image,
Facial image pixel is normalized -1 to 1, so that model training process is easy convergence；Normalization pixel is completed in facial image
After processing, image data enhancing is carried out to facial image using operations such as Random Level overturning, random interception and colour dithers, from
And expand the diversity of training data, and prevent over-fitting；Processes pixel and image is normalized in test facial image
Data enhancing processing after to get arrive target facial image；It is understood that pretreatment mass normalization processes pixel and figure
As data enhancing is handled.
Step S20 is split to obtain each according to the corresponding target network-layer of the global characteristics to the global characteristics
A local feature, and the corresponding each global characteristics of the target facial image are merged to obtain fusion feature；
After extracting global characteristics, global characteristics can be split to obtain local feature, specifically, passing through manually
Mode is repeatedly split the global characteristics figure of same target network-layer output, so that partitioning scheme the most reasonable is obtained,
Then the partitioning scheme and target network-layer are associated with and are stored into device, therefore extracting the complete of target network-layer when device
After office's characteristic pattern, partitioning scheme is obtained according to the binding information of the target network-layer, and then according to partitioning scheme to global special
Sign figure is split, and obtains multiple characteristic blocks, these characteristic blocks are by Chi Huahou, as local feature, it should be noted that feature
The size of block is smaller, and the granularity of face characteristic corresponding to characteristic block is smaller, the description of face more horn of plenty, easier to represent
The fine feature of face, so that device can be easy to excavate the fine feature of face.For example, layer2 can be exported
Characteristic pattern carries out the multi-scale segmentation of 2*2, the local feature block of available 2 28*14,2 14*28 and 4 14*14, simultaneously
The multi-scale segmentation that 2*2 is carried out to layer3 and layer4, respectively obtain 2 14*7,2 7*14 and 4 7*7 characteristic block and
The characteristic block of 2 7*4,2 4*7 and 4 4*4；8 local feature dimensions that layer2 is obtained all are 8 of 512, layer3
Local feature dimension is all that 8 local feature dimensions of 1024, layer4 are all 2048.
In addition, each global characteristics are merged after obtaining multiple global characteristics of same facial image, thus
To fusion feature, the dimension of fusion feature is 512+1024+2048=3588.
Step S30 carries out loss supervised learning to the global characteristics, the local feature and fusion feature, with right
Human face recognition model is trained.
After obtaining the corresponding local feature of facial image, global characteristics and fusion feature, by optimizer to part
Feature, global characteristics and fusion feature carry out loss supervised learning, specifically, referring to Fig. 4, i.e. step S30 includes:
Step S31 determines the global characteristics, the local feature and the corresponding target loss letter of the fusion feature
The corresponding weighted value of several and described target loss function；
Step S32, according to the target loss function and the corresponding weighted value of the target loss function, to the mesh
The corresponding feature of scalar functions carries out loss supervised learning.
In the present embodiment, first object loss function is configured to each local feature, first object loss function is to return
One changes figure penalties function namely Softmax loss, and configures the second target damage to each global characteristics and fusion feature
Function is lost, the second target loss function includes Softmax loss and Triplet loss (triple loss function), can be with
Understand, local feature, fusion feature and global characteristics are provided with Softmax loss, it is therefore desirable to each mesh be arranged
It marks and normalizes figure penalties function respective weights value in loss function, and figure penalties function is normalized in global characteristics needs basis
The target network-layer of output global characteristics is configured, for example, can set the Softmax loss weight of local feature to
0.4, the weight of the corresponding Softmax loss of the local feature of different levels is set as 0.6,0.8 and 1, and fusion feature pair
The weight of the Softmax loss answered is then set as 1.2.In the present embodiment, optimizer can be Adam, will when being trained
The learning rate of optimizer is set as 0.02, epoch and is set as 60, so by the setting parameter of optimizer, loss function and
The corresponding weighted value of loss function carries out loss supervised learning to the corresponding feature of the loss function.
The Softmax loss study of feature is to carry out at Softmax layers, and Triplet loss study is in Triplet
Layer carries out, but each Softmax layers first inputs full articulamentum equipped with corresponding full articulamentum namely feature, then inputs Softmax
Layer.
It should be noted that in the training process, the corresponding loss function of depth convolutional neural networks need to be obtained in real time,
When loss function loss converges to pre-set interval (pre-set interval can be any appropriate range, such as about 0.6), can assert people
The training of face identification model finishes, certainly, can also further progress training, at this point, SGD optimizer is updated, and simultaneously by learning rate
It is set as 0.005, meanwhile, the other parameters of optimizer are constant, in people after all training of each target facial image, obtained
Face identification model is optimal human face recognition model, at this point, the fusion feature of human face recognition model output is corresponding for facial image
Verifying fusion feature, can will verify fusion feature as identify face template to be stored in device or database, this
Outside, since the dimensional comparison of fusion feature is big, dimensionality reduction can be carried out to fusion feature, for example, fusion feature is reduced to 512 dimensions.
When human face recognition model carries out subsequent recognition of face, the facial image of acquisition is subjected to target network-layer
Global characteristics extract, and then each global characteristics are merged, and fusion feature are obtained, finally by the fusion feature and database
In each fusion feature that prestores be compared, and then identify face.
In the present invention, global characteristics are extracted by the heterogeneous networks layer to depth convolutional neural networks, then passes through the overall situation
Image Segmentation Methods Based on Features obtains local feature, and then to local feature, global characteristics and the fusion feature merged by each global characteristics
The supervised training that loss function carries out parameter is set, promotes the local feature finally obtained and global characteristics to final fusion
Feature is all effective, and then the feature for the face that can preferably indicate.
In technical solution provided in this embodiment, the training device of human face recognition model uses depth convolutional neural networks
In each target network-layer to each target facial image carry out feature extraction obtain multiple global characteristics, further according to global characteristics
Corresponding target network-layer is split the overall situation, to obtain multiple local features, and then will correspond to a facial image
Multiple local features and multiple global characteristics are merged to obtain fusion feature, finally to each global characteristics, local feature
And fusion feature carries out loss supervised learning to carry out the training of human face recognition model；Due to the training cartridge of human face recognition model
It sets and loss supervision is carried out by the local feature for exporting global characteristics, fusion feature and global characteristics segmentation to heterogeneous networks layer
Study, so that device has excavated the fine feature in facial image, improves the precision of recognition of face, further, since face
Global characteristics are divided into multiple local features and carry out loss supervised learning by the training device of identification model, without manually blocking
The occlusion area of face or the characteristic processing blocked to face, simplify the training process of facial image.
It is the second embodiment of the training method of human face recognition model of the present invention referring to Fig. 5, Fig. 5, is based on first embodiment,
The step of carrying out loss supervised learning to the fusion feature in the step S30 include:
Step S33 reduces the dimension of the fusion feature；
Step S34 carries out loss supervised learning to the fusion feature for reducing dimension.
In the present embodiment, due to fusion feature be merge to obtain by the corresponding multiple global characteristics of facial image so that
The dimensional comparison of fusion feature is big, and in one embodiment, the dimension of fusion feature is 3588, this meeting is so that the fusion finally exported
Feature also can be relatively high, and the fusion feature finally exported can make people as the feature templates of facial image in database
The speed of face identification increases.In this regard, in the present embodiment, after merging to obtain fusion feature by multiple global characteristics, using volume
Fusion feature is carried out channel reduction by long-pending mode, and fusion feature may make to reduce dimension and can be used in the present embodiment in this way
Fusion feature is reduced to 512 dimensions by the convolution of 1*1.
Convolution is mainly main herein, and there are two the effects of aspect:
1, it realizes the information selection across channel, multiple features is subjected to information integration；
2, dimensionality reduction is carried out to the port number of feature, improves nonlinear characteristic, the high dimensional feature finally merged passes through study
One convolution nuclear parameter realizes de-redundancy, improves recognition time, reduces the storage of feature.
In the technical solution that embodiment provides, device carries out the reduction of dimension by way of convolution to fusion feature,
So that reducing the identification duration of recognition of face when facial image is verified, and then improve the rate of recognition of face, improving
User experience.
To achieve the above object, the present invention also provides a kind of training device of human face recognition model, the recognition of face moulds
The training device of type includes processor, memory and is stored in the face that can be run on the memory and on the processor
The training program of identification model, the training program of the human face recognition model realize embodiment as above when being executed by the processor
Each step of the training method of the human face recognition model.
The present invention also provides a kind of computer readable storage medium, the computer-readable recording medium storage has face knowledge
The training program of other model, the training program of the human face recognition model realize embodiment institute as above when being executed by the processor
Each step of the training method for the human face recognition model stated.
The serial number of the above embodiments of the invention is only for description, does not represent the advantages or disadvantages of the embodiments.
It should be noted that, in this document, the terms "include", "comprise" or its any other variant are intended to non-row
His property includes, so that the process, method, article or the device that include a series of elements not only include those elements, and
And further include other elements that are not explicitly listed, or further include for this process, method, article or device institute it is intrinsic
Element.In the absence of more restrictions, the element limited by sentence "including a ...", it is not excluded that including being somebody's turn to do
There is also other identical elements in the process, method of element, article or device.
Through the above description of the embodiments, those skilled in the art can be understood that above-described embodiment side
Method can be realized by means of software and necessary general hardware platform, naturally it is also possible to by hardware, but in many cases
The former is more preferably embodiment.Based on this understanding, technical solution of the present invention substantially in other words does the prior art
The part contributed out can be embodied in the form of software products, which is stored in one as described above
In storage medium (such as ROM/RAM, magnetic disk, CD), including some instructions are used so that terminal device (it can be mobile phone,
Computer, server, air conditioner or network equipment etc.) execute method described in each embodiment of the present invention.
The above is only a preferred embodiment of the present invention, is not intended to limit the scope of the invention, all to utilize this hair
Equivalent structure or equivalent flow shift made by bright specification and accompanying drawing content is applied directly or indirectly in other relevant skills
Art field, is included within the scope of the present invention.
Claims (10)
1. a kind of training method of human face recognition model, which is characterized in that the training method of the human face recognition model include with
Lower step:
Feature extraction is carried out to each target facial image using each target network-layer in depth convolutional neural networks, to obtain
Multiple global characteristics；
The global characteristics are split according to the global characteristics corresponding target network-layer to obtain each local feature, and
The corresponding each global characteristics of the target facial image are merged to obtain fusion feature；
Loss supervised learning is carried out to the global characteristics, the local feature and fusion feature, to human face recognition model
It is trained.
2. the training method of human face recognition model as described in claim 1, which is characterized in that it is described to the global characteristics,
The step of local feature and fusion feature carry out loss supervised learning include:
Determine the global characteristics, the local feature and the corresponding target loss function of the fusion feature and the mesh
Mark the corresponding weighted value of loss function；
It is corresponding to the objective function according to the target loss function and the corresponding weighted value of the target loss function
Feature carries out loss supervised learning.
3. the training method of human face recognition model as claimed in claim 2, which is characterized in that the determination is described global special
Sign, the local feature and the corresponding target loss function of the fusion feature and the corresponding power of the target loss function
The step of weight values includes:
Normalization figure penalties function is determined as the corresponding first object loss function of the local feature, and by the normalizing
Change figure penalties function and triple loss function is determined as the global characteristics and the fusion feature corresponding second
Target loss function；
The first weighted value and the corresponding second target loss letter of the fusion feature are configured for the first object loss function
Exponential function is normalized in number and configures the second weighted value, and according to the corresponding target network-layer of the global characteristics to the overall situation
Exponential function is normalized in the corresponding second target loss function of feature configures third weighted value.
4. the training method of human face recognition model as described in claim 1, which is characterized in that use depth convolutional neural networks
In each target network-layer feature extraction is carried out to each target facial image, to obtain multiple global characteristics the step of includes:
Multiple target network-layers are determined in each network layer of the depth convolutional neural networks, wherein the target network
Distance is less than other network layers except the target network-layer and institute between layer and the output layer of the depth convolutional neural networks
State the distance between output layer；
Target facial image is inputted into each target network-layer, it is corresponding global special to obtain each target network-layer
Sign figure；
Pond is carried out to each global characteristics figure, obtains multiple global characteristics.
5. the training method of human face recognition model as described in claim 1, which is characterized in that it is described to the fusion feature into
Row loss supervised learning the step of include:
Reduce the dimension of the fusion feature；
Loss supervised learning is carried out to the fusion feature for reducing dimension.
6. the training method of human face recognition model as described in claim 1, which is characterized in that described to use depth convolutional Neural
Before the step of each target network-layer carries out feature extraction to each target facial image in network, further includes:
Each test facial image is obtained, and each test facial image is pre-processed, to obtain each target face figure
Picture, wherein the pretreatment mass image normalization processes pixel and image data enhancing processing.
7. the training method of human face recognition model as claimed in any one of claims 1 to 6, which is characterized in that described to described
After the step of global characteristics, the local feature and fusion feature carry out loss supervised learning, further includes:
The corresponding loss function of the depth convolutional neural networks is obtained in real time；
When the corresponding loss function of the depth convolutional neural networks converges to set interval, the human face recognition model is completed
Training, and save complete training the human face recognition model.
8. the training method of human face recognition model as claimed in claim 7, which is characterized in that the people that training will be completed
After the step of face identification model saves, further includes:
The fusion feature for the human face recognition model output for completing training is prestored into fusion as the facial image is corresponding
Feature, and fusion feature is prestored described in preservation.
9. a kind of training device of human face recognition model, which is characterized in that the training device of the human face recognition model includes place
Manage device, memory and the training journey for being stored in the human face recognition model that can be run on the memory and on the processor
Sequence is realized when the training program of the human face recognition model is executed by the processor as claim 1-8 is described in any item
Each step of the training method of human face recognition model.
10. a kind of computer readable storage medium, which is characterized in that the computer-readable recording medium storage has recognition of face
The training program of model realizes such as claim 1-8 when the training program of the human face recognition model is executed by the processor
Each step of the training method of described in any item human face recognition models.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910219576.XA CN109934197B (en) | 2019-03-21 | 2019-03-21 | Training method and device for face recognition model and computer readable storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910219576.XA CN109934197B (en) | 2019-03-21 | 2019-03-21 | Training method and device for face recognition model and computer readable storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109934197A true CN109934197A (en) | 2019-06-25 |
| CN109934197B CN109934197B (en) | 2023-07-07 |
Family
ID=66987983
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910219576.XA Active CN109934197B (en) | 2019-03-21 | 2019-03-21 | Training method and device for face recognition model and computer readable storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109934197B (en) |
Cited By (26)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110415166A (en) * | 2019-07-29 | 2019-11-05 | 腾讯科技（深圳）有限公司 | Training method, image processing method, device and the storage medium of blending image processing model |
| CN110533184A (en) * | 2019-08-31 | 2019-12-03 | 南京人工智能高等研究院有限公司 | A kind of training method and device of network model |
| CN110569731A (en) * | 2019-08-07 | 2019-12-13 | 北京旷视科技有限公司 | face recognition method and device and electronic equipment |
| CN110688976A (en) * | 2019-10-09 | 2020-01-14 | 创新奇智(北京)科技有限公司 | Store comparison method based on image identification |
| CN110781724A (en) * | 2018-09-11 | 2020-02-11 | 开放智能机器（上海）有限公司 | Face recognition neural network, method, device, equipment and storage medium |
| CN111177469A (en) * | 2019-12-20 | 2020-05-19 | 国久大数据有限公司 | Face retrieval method and face retrieval device |
| CN111339812A (en) * | 2019-06-29 | 2020-06-26 | 北京澎思科技有限公司 | Pedestrian identification and re-identification method based on whole or partial human body structural feature set, electronic equipment and storage medium |
| CN111680701A (en) * | 2020-05-07 | 2020-09-18 | 北京三快在线科技有限公司 | Training method and device of image recognition model and image recognition method and device |
| CN111783607A (en) * | 2020-06-24 | 2020-10-16 | 北京百度网讯科技有限公司 | Training method and device of face recognition model, electronic equipment and storage medium |
| CN111832669A (en) * | 2020-09-21 | 2020-10-27 | 首都师范大学 | Method and device for establishing learning participation degree recognition network model |
| CN111898547A (en) * | 2020-07-31 | 2020-11-06 | 平安科技（深圳）有限公司 | Training method, device and equipment of face recognition model and storage medium |
| CN111898465A (en) * | 2020-07-08 | 2020-11-06 | 北京捷通华声科技股份有限公司 | Method and device for acquiring face recognition model |
| CN112036446A (en) * | 2020-08-06 | 2020-12-04 | 汇纳科技股份有限公司 | Method, system, medium, and apparatus for target recognition feature fusion |
| CN112052792A (en) * | 2020-09-04 | 2020-12-08 | 恒睿(重庆)人工智能技术研究院有限公司 | Cross-model face recognition method, device, equipment and medium |
| CN112052789A (en) * | 2020-09-03 | 2020-12-08 | 腾讯科技（深圳）有限公司 | Face recognition method and device, electronic equipment and storage medium |
| CN112101172A (en) * | 2020-09-08 | 2020-12-18 | 平安科技（深圳）有限公司 | Weight grafting-based model fusion face recognition method and related equipment |
| CN112215066A (en) * | 2020-09-08 | 2021-01-12 | 北京农业信息技术研究中心 | Livestock face image recognition method and device |
| CN112348732A (en) * | 2019-08-08 | 2021-02-09 | 华为技术有限公司 | Model reasoning method and device based on graphics rendering pipeline and storage medium |
| CN112766164A (en) * | 2021-01-20 | 2021-05-07 | 深圳力维智联技术有限公司 | Face recognition model training method, device and equipment and readable storage medium |
| CN112966626A (en) * | 2021-03-16 | 2021-06-15 | 三星（中国）半导体有限公司 | Face recognition method and device |
| CN113128309A (en) * | 2020-01-10 | 2021-07-16 | 中移(上海)信息通信科技有限公司 | Facial expression recognition method, device, equipment and medium |
| CN113610071A (en) * | 2021-10-11 | 2021-11-05 | 深圳市一心视觉科技有限公司 | Face living body detection method and device, electronic equipment and storage medium |
| CN114120430A (en) * | 2022-01-26 | 2022-03-01 | 杭州魔点科技有限公司 | Mask face recognition method based on double-branch weight fusion homology self-supervision |
| CN114909786A (en) * | 2022-04-24 | 2022-08-16 | 青岛海信日立空调系统有限公司 | Air conditioning system and control method thereof |
| CN117576766A (en) * | 2024-01-16 | 2024-02-20 | 杭州魔点科技有限公司 | Cross-space-time compatibility unsupervised self-learning face recognition method and system |
| CN117576766B (en) * | 2024-01-16 | 2024-04-26 | 杭州魔点科技有限公司 | Cross-space-time compatibility unsupervised self-learning face recognition method and system |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107045618A (en) * | 2016-02-05 | 2017-08-15 | 北京陌上花科技有限公司 | A kind of facial expression recognizing method and device |
| CN107330383A (en) * | 2017-06-18 | 2017-11-07 | 天津大学 | A kind of face identification method based on depth convolutional neural networks |
| CN107832672A (en) * | 2017-10-12 | 2018-03-23 | 北京航空航天大学 | A kind of pedestrian's recognition methods again that more loss functions are designed using attitude information |
| US10140544B1 (en) * | 2018-04-02 | 2018-11-27 | 12 Sigma Technologies | Enhanced convolutional neural network for image segmentation |
| CN109033938A (en) * | 2018-06-01 | 2018-12-18 | 上海阅面网络科技有限公司 | A kind of face identification method based on ga s safety degree Fusion Features |
| CN109492560A (en) * | 2018-10-26 | 2019-03-19 | 深圳力维智联技术有限公司 | Facial image Feature fusion, device and storage medium based on time scale |
-
2019
- 2019-03-21 CN CN201910219576.XA patent/CN109934197B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107045618A (en) * | 2016-02-05 | 2017-08-15 | 北京陌上花科技有限公司 | A kind of facial expression recognizing method and device |
| CN107330383A (en) * | 2017-06-18 | 2017-11-07 | 天津大学 | A kind of face identification method based on depth convolutional neural networks |
| CN107832672A (en) * | 2017-10-12 | 2018-03-23 | 北京航空航天大学 | A kind of pedestrian's recognition methods again that more loss functions are designed using attitude information |
| US10140544B1 (en) * | 2018-04-02 | 2018-11-27 | 12 Sigma Technologies | Enhanced convolutional neural network for image segmentation |
| CN109033938A (en) * | 2018-06-01 | 2018-12-18 | 上海阅面网络科技有限公司 | A kind of face identification method based on ga s safety degree Fusion Features |
| CN109492560A (en) * | 2018-10-26 | 2019-03-19 | 深圳力维智联技术有限公司 | Facial image Feature fusion, device and storage medium based on time scale |
Cited By (39)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110781724A (en) * | 2018-09-11 | 2020-02-11 | 开放智能机器（上海）有限公司 | Face recognition neural network, method, device, equipment and storage medium |
| CN111339812A (en) * | 2019-06-29 | 2020-06-26 | 北京澎思科技有限公司 | Pedestrian identification and re-identification method based on whole or partial human body structural feature set, electronic equipment and storage medium |
| CN110415166A (en) * | 2019-07-29 | 2019-11-05 | 腾讯科技（深圳）有限公司 | Training method, image processing method, device and the storage medium of blending image processing model |
| CN110415166B (en) * | 2019-07-29 | 2023-01-06 | 腾讯科技（深圳）有限公司 | Training method for fusion image processing model, image processing method, image processing device and storage medium |
| CN110569731A (en) * | 2019-08-07 | 2019-12-13 | 北京旷视科技有限公司 | face recognition method and device and electronic equipment |
| CN110569731B (en) * | 2019-08-07 | 2023-04-14 | 北京旷视科技有限公司 | Face recognition method and device and electronic equipment |
| CN112348732A (en) * | 2019-08-08 | 2021-02-09 | 华为技术有限公司 | Model reasoning method and device based on graphics rendering pipeline and storage medium |
| CN112348732B (en) * | 2019-08-08 | 2023-11-17 | 华为技术有限公司 | Model reasoning method, device and storage medium based on graphic rendering pipeline |
| US11941724B2 (en) | 2019-08-08 | 2024-03-26 | Huawei Technologies Co., Ltd. | Model inference method and apparatus based on graphics rendering pipeline, and storage medium |
| WO2021022962A1 (en) * | 2019-08-08 | 2021-02-11 | 华为技术有限公司 | Method and device for model inference based on graphics rendering pipelines, and storage medium |
| CN110533184A (en) * | 2019-08-31 | 2019-12-03 | 南京人工智能高等研究院有限公司 | A kind of training method and device of network model |
| CN110688976A (en) * | 2019-10-09 | 2020-01-14 | 创新奇智(北京)科技有限公司 | Store comparison method based on image identification |
| CN111177469A (en) * | 2019-12-20 | 2020-05-19 | 国久大数据有限公司 | Face retrieval method and face retrieval device |
| CN113128309A (en) * | 2020-01-10 | 2021-07-16 | 中移(上海)信息通信科技有限公司 | Facial expression recognition method, device, equipment and medium |
| CN111680701A (en) * | 2020-05-07 | 2020-09-18 | 北京三快在线科技有限公司 | Training method and device of image recognition model and image recognition method and device |
| CN111680701B (en) * | 2020-05-07 | 2023-04-07 | 北京三快在线科技有限公司 | Training method and device of image recognition model and image recognition method and device |
| CN111783607B (en) * | 2020-06-24 | 2023-06-27 | 北京百度网讯科技有限公司 | Training method and device of face recognition model, electronic equipment and storage medium |
| CN111783607A (en) * | 2020-06-24 | 2020-10-16 | 北京百度网讯科技有限公司 | Training method and device of face recognition model, electronic equipment and storage medium |
| CN111898465A (en) * | 2020-07-08 | 2020-11-06 | 北京捷通华声科技股份有限公司 | Method and device for acquiring face recognition model |
| CN111898547B (en) * | 2020-07-31 | 2024-04-16 | 平安科技（深圳）有限公司 | Training method, device, equipment and storage medium of face recognition model |
| CN111898547A (en) * | 2020-07-31 | 2020-11-06 | 平安科技（深圳）有限公司 | Training method, device and equipment of face recognition model and storage medium |
| CN112036446A (en) * | 2020-08-06 | 2020-12-04 | 汇纳科技股份有限公司 | Method, system, medium, and apparatus for target recognition feature fusion |
| CN112036446B (en) * | 2020-08-06 | 2023-12-12 | 汇纳科技股份有限公司 | Method, system, medium and device for fusing target identification features |
| CN112052789A (en) * | 2020-09-03 | 2020-12-08 | 腾讯科技（深圳）有限公司 | Face recognition method and device, electronic equipment and storage medium |
| CN112052792A (en) * | 2020-09-04 | 2020-12-08 | 恒睿(重庆)人工智能技术研究院有限公司 | Cross-model face recognition method, device, equipment and medium |
| CN112052792B (en) * | 2020-09-04 | 2022-04-26 | 恒睿(重庆)人工智能技术研究院有限公司 | Cross-model face recognition method, device, equipment and medium |
| CN112101172B (en) * | 2020-09-08 | 2023-09-08 | 平安科技（深圳）有限公司 | Weight grafting-based model fusion face recognition method and related equipment |
| CN112101172A (en) * | 2020-09-08 | 2020-12-18 | 平安科技（深圳）有限公司 | Weight grafting-based model fusion face recognition method and related equipment |
| CN112215066A (en) * | 2020-09-08 | 2021-01-12 | 北京农业信息技术研究中心 | Livestock face image recognition method and device |
| CN111832669A (en) * | 2020-09-21 | 2020-10-27 | 首都师范大学 | Method and device for establishing learning participation degree recognition network model |
| CN112766164A (en) * | 2021-01-20 | 2021-05-07 | 深圳力维智联技术有限公司 | Face recognition model training method, device and equipment and readable storage medium |
| CN112966626A (en) * | 2021-03-16 | 2021-06-15 | 三星（中国）半导体有限公司 | Face recognition method and device |
| CN113610071B (en) * | 2021-10-11 | 2021-12-24 | 深圳市一心视觉科技有限公司 | Face living body detection method and device, electronic equipment and storage medium |
| CN113610071A (en) * | 2021-10-11 | 2021-11-05 | 深圳市一心视觉科技有限公司 | Face living body detection method and device, electronic equipment and storage medium |
| CN114120430B (en) * | 2022-01-26 | 2022-04-22 | 杭州魔点科技有限公司 | Mask face recognition method based on double-branch weight fusion homology self-supervision |
| CN114120430A (en) * | 2022-01-26 | 2022-03-01 | 杭州魔点科技有限公司 | Mask face recognition method based on double-branch weight fusion homology self-supervision |
| CN114909786A (en) * | 2022-04-24 | 2022-08-16 | 青岛海信日立空调系统有限公司 | Air conditioning system and control method thereof |
| CN117576766A (en) * | 2024-01-16 | 2024-02-20 | 杭州魔点科技有限公司 | Cross-space-time compatibility unsupervised self-learning face recognition method and system |
| CN117576766B (en) * | 2024-01-16 | 2024-04-26 | 杭州魔点科技有限公司 | Cross-space-time compatibility unsupervised self-learning face recognition method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109934197B (en) | 2023-07-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109934197A (en) | Training method, device and the computer readable storage medium of human face recognition model | |
| US11487995B2 (en) | Method and apparatus for determining image quality | |
| CN109657554B (en) | Image identification method and device based on micro expression and related equipment | |
| CN110728330A (en) | Object identification method, device, equipment and storage medium based on artificial intelligence | |
| CN108961675A (en) | Fall detection method based on convolutional neural networks | |
| CN107844784A (en) | Face identification method, device, computer equipment and readable storage medium storing program for executing | |
| CN111738735B (en) | Image data processing method and device and related equipment | |
| CN106548468B (en) | The method of discrimination and device of image definition | |
| CN109840467A (en) | A kind of in-vivo detection method and system | |
| CN112446302B (en) | Human body posture detection method, system, electronic equipment and storage medium | |
| CN112801054B (en) | Face recognition model processing method, face recognition method and device | |
| CN110232326A (en) | A kind of D object recognition method, device and storage medium | |
| CN110458829A (en) | Image quality control method, device, equipment and storage medium based on artificial intelligence | |
| CN113705290A (en) | Image processing method, image processing device, computer equipment and storage medium | |
| CN116110100B (en) | Face recognition method, device, computer equipment and storage medium | |
| CN110070037A (en) | Smooth upgrading method, device and the readable storage medium storing program for executing of human face recognition model | |
| CN112991278A (en) | Method and system for detecting Deepfake video by combining RGB (red, green and blue) space domain characteristics and LoG (LoG) time domain characteristics | |
| CN115035418A (en) | Remote sensing image semantic segmentation method and system based on improved deep LabV3+ network | |
| KR20210100592A (en) | Face recognition technology based on heuristic Gaussian cloud transformation | |
| CN114596608B (en) | Double-stream video face counterfeiting detection method and system based on multiple clues | |
| CN112818899A (en) | Face image processing method and device, computer equipment and storage medium | |
| US20230237630A1 (en) | Image processing method and apparatus | |
| CN116823442A (en) | Prediction method for fraud risk in credit business | |
| CN115936961B (en) | Steganalysis method, equipment and medium based on few-sample comparison learning network | |
| CN109886084A (en) | Face authentication method, electronic equipment and storage medium based on gyroscope |
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 |