CN109063666A - The lightweight face identification method and system of convolution are separated based on depth - Google Patents
The lightweight face identification method and system of convolution are separated based on depth Download PDFInfo
- Publication number
- CN109063666A CN109063666A CN201810921647.6A CN201810921647A CN109063666A CN 109063666 A CN109063666 A CN 109063666A CN 201810921647 A CN201810921647 A CN 201810921647A CN 109063666 A CN109063666 A CN 109063666A
- Authority
- CN
- China
- Prior art keywords
- depth
- neural networks
- convolutional neural
- trained
- convolution
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/172—Classification, e.g. identification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2413—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
- G06F18/24133—Distances to prototypes
- G06F18/24137—Distances to cluster centroïds
- G06F18/2414—Smoothing the distance, e.g. radial basis function networks [RBFN]
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Molecular Biology (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Computing Systems (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention discloses a kind of lightweight face identification method and system for separating convolution based on depth, which includes: to be trained to three levels connection convolutional neural networks, obtains trained three levels connection convolutional neural networks；Convolutional neural networks are separated to depth to be trained, and are obtained depth and are separated convolutional neural networks model；The weight coefficient for separating the convolutional layer in convolutional neural networks model to depth compresses, and obtains trained depth and separates convolutional neural networks model；It inputs facial image to be identified to trained three level connection convolutional neural networks and carries out image preprocessing, then export pretreatment image to trained depth and separate the progress recognition of face of convolutional neural networks model.The present invention separates convolution by using depth and reduces calculation amount, reduces model size, and the weight coefficient for further separating the convolutional layer in convolutional neural networks model to depth compresses, so that model reduces 35 to 49 times.
Description
Technical field
The present invention relates to recognitions of face and deep learning field, in particular to the lightweight for separating convolution based on depth
Face identification method and system.
Background technique
Recognition of face is that a kind of technology of authentication is carried out using human face's characteristic information.Current recognition of face skill
Art has mainly been applied to following aspect: Haikou, airport, the certificate verification of the occasions such as railway station；The archive management of criminal's photo
And criminal investigation and case detection；Face datection in video monitoring, tracking and identification etc., in addition, face recognition technology is in medicine, man-machine friendship
Mutually, the fields such as channel control also have broad application prospects.
Produced by conventional face's recognizer and the feature that uses may be considered that and belong to shallow-layer feature, and cannot be from
More deep high semantic feature and its depth characteristic are obtained in original image, to the recognition effect obtained, these biographies
Face recognition algorithms of uniting must be in conjunction with the help of artificial feature, and the feature being manually set is led in feature extraction and identification process
Not desirable human factor and error can often be brought；Also, under inartificial intervention, conventional face's recognizer is often
Useful identification feature cannot be automatically extracted from original image, so that conventional face's recognition methods discrimination is lower.
Therefore, in recent years, most current optimal face identification methods are all based on convolutional neural networks, convolution
Neural network can learn to the feature with distinction for being better than traditional-handwork feature, know to have better than traditional face
The discrimination of other method.But convolutional neural networks are hierarchical structure, for the recognition effect obtained, the network number of plies is very deep,
It include again wherein multiple convolutional layers, each convolutional layer includes multiple convolution kernels again, so weight parameter is also very much, typically hundred
Ten thousand ranks, to cause convolutional neural networks model very big, identification is slow, and movement can not be widely used in by further resulting in it
End.
Summary of the invention
It is an object of the invention to: in existing face identification method, model is excessive and problem that recognition speed is slow,
A kind of lightweight face identification method and system being separated convolution based on depth is provided, is damaged by training based on AM-Softmax
The depth for losing function separates convolutional neural networks model, the method for carrying out recognition of face, and separates convolutional Neural to depth
The weight coefficient of network is compressed, to reduce model size, while improving discrimination and accuracy rate.
The technical solution adopted by the invention is as follows:
A kind of lightweight face identification method being separated convolution based on depth, is specifically included:
S1, three levels of building join convolutional neural networks, and input the image pyramid of the first facial image data set, to institute
It states three levels connection convolutional neural networks to be trained, obtains trained three levels connection convolutional neural networks；
S2, depth of the building based on AM-Softmax loss function separate convolutional neural networks, and input the second face
Image data set separates convolutional neural networks to the depth and is trained, and obtains depth and separates convolutional neural networks mould
Type；
S3, the weight coefficient that the convolutional layer in convolutional neural networks model is separated to depth compress, and are trained
Good depth separates convolutional neural networks model；
S4, image pyramid to the trained three level connection convolutional neural networks for inputting facial image to be identified carry out figure
As pretreatment, then exports pretreatment image to trained depth and separate the progress recognition of face of convolutional neural networks model.
Further, the depth in step S2 based on AM-Softmax loss function separates convolutional neural networks, knot
Structure includes: 17 convolution modules；
First convolution module is the common convolutional layer of a 3x3；
Second convolution module to the 14th convolution module is that depth separates convolution module；
15th convolution module is the average pond layer of a 7x7；
16th convolution module is a full articulamentum；
17th convolution module is AM-Softmax layers.
Further, first convolution module be into the 16th convolution module, after each convolution module with
Batch normalization and nonlinear activation.
A kind of lightweight face identification system separating convolution based on depth, comprising:
First nerves network training unit (101) joins convolutional neural networks for constructing three levels, and inputs the first face
The image pyramid of image data set is trained three level connection convolutional neural networks, obtains trained three level
Join convolutional neural networks；
Nervus opticus network training unit (102), it is separable for constructing the depth based on AM-Softmax loss function
Convolutional neural networks, and the second face image data collection is inputted, convolutional neural networks are separated to the depth and are trained, are obtained
Convolutional neural networks model is separated to depth；
Model compression unit (103), for separating the weight system of the convolutional layer in convolutional neural networks model to depth
Number is compressed, and is obtained trained depth and is separated convolutional neural networks model；
Face identification unit (104), for inputting the image pyramid of facial image to be identified to trained three level
Join convolutional neural networks and carry out image preprocessing, then exports pretreatment image to trained depth and separate convolutional Neural net
Network model carries out recognition of face.
In conclusion by adopting the above-described technical solution, the beneficial effects of the present invention are:
1, the present invention carries out image preprocessing using three levels connection convolutional neural networks, and trains progress depth that can divide with this
From convolutional neural networks model, the weight coefficient that the convolutional layer in convolutional neural networks model is further separated to depth is carried out
Compression obtains trained depth and separates convolutional neural networks model, and carries out recognition of face to facial image with this；Pass through
Convolution is separated using the depth based on AM-Softmax loss function and reduces calculation amount, reduces model size, and further right
The weight coefficient that depth separates the convolutional layer in convolutional neural networks model is compressed, so that model reduces 35 to 49 times.
2, depth is separated traditional Softmax that convolution uses and damaged by the present invention by introducing a parameter factors l
It loses function and is improved to AM-Softmax loss function, improve discrimination and accuracy rate.
Detailed description of the invention
In order to illustrate the technical solution of the embodiments of the present invention more clearly, below will be to needed in the embodiment attached
Figure is briefly described, it should be understood that the following drawings illustrates only certain embodiments of the present invention, therefore is not construed as pair
The restriction of range for those of ordinary skill in the art without creative efforts, can also be according to this
A little attached drawings obtain other relevant attached drawings.
Fig. 1 is the flow chart that the lightweight face identification method of convolution is separated the present invention is based on depth.
Fig. 2 is that three levels of the invention join convolutional neural networks structural schematic diagram.
Fig. 3 is that the separable convolutional neural networks model structure of the depth of the invention based on AM-Softmax loss function is shown
It is intended to.
Fig. 4 is what the present invention compressed the weight coefficient that depth separates the convolutional layer in convolutional neural networks model
Flow chart.
Fig. 5 is that three levels of utilization of the invention join convolutional neural networks, to the image pyramid of facial image to be identified into
The pretreated flow chart of row.
Fig. 6 is the structural block diagram that the lightweight face identification system of convolution is separated the present invention is based on depth.
Marked in the figure: 101- first nerves network training unit, 102- nervus opticus network training unit, 103- model pressure
Contracting unit, 104- face identification unit.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that described herein, specific examples are only used to explain the present invention, not
For limiting the present invention, i.e., described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is logical
The component for the embodiment of the present invention being often described and illustrated herein in the accompanying drawings can be arranged and be designed with a variety of different configurations.
Therefore, the detailed description of the embodiment of the present invention provided in the accompanying drawings is not intended to limit below claimed
The scope of the present invention, but be merely representative of selected embodiment of the invention.Based on the embodiment of the present invention, those skilled in the art
Member's every other embodiment obtained without making creative work, shall fall within the protection scope of the present invention.
It should be noted that the relational terms of term " first " and " second " or the like be used merely to an entity or
Operation is distinguished with another entity or operation, and without necessarily requiring or implying between these entities or operation, there are any
This actual relationship or sequence.Moreover, the terms "include", "comprise" or its any other variant be intended to it is non-exclusive
Property include so that include a series of elements process, method, article or equipment not only include those elements, but also
Further include other elements that are not explicitly listed, or further include for this process, method, article or equipment it is intrinsic
Element.In the absence of more restrictions, the element limited by sentence "including a ...", it is not excluded that including described
There is also other identical elements in the process, method, article or equipment of element.
A kind of lightweight face identification method being separated convolution based on depth, is specifically included:
1. a kind of lightweight face identification method for separating convolution based on depth, which is characterized in that specifically include:
S1, three levels of building join convolutional neural networks, and input the image pyramid of the first facial image data set, to institute
It states three levels connection convolutional neural networks to be trained, obtains trained three levels connection convolutional neural networks；
S2, depth of the building based on AM-Softmax loss function separate convolutional neural networks, and input the second face
Image data set separates convolutional neural networks to the depth and is trained, and obtains depth and separates convolutional neural networks mould
Type；
S3, the weight coefficient that the convolutional layer in convolutional neural networks model is separated to depth compress, and are trained
Good depth separates convolutional neural networks model；
S4, image pyramid to the trained three level connection convolutional neural networks for inputting facial image to be identified carry out figure
As pretreatment, then exports pretreatment image to trained depth and separate the progress recognition of face of convolutional neural networks model.
A kind of lightweight face identification system separating convolution based on depth, comprising:
First nerves network training unit 101 joins convolutional neural networks for constructing three levels, and inputs the first face figure
As the image pyramid of data set, three level connection convolutional neural networks are trained, obtain trained three levels connection
Convolutional neural networks；
Nervus opticus network training unit 102, for constructing the separable volume of the depth based on AM-Softmax loss function
Product neural network, and the second face image data collection is inputted, convolutional neural networks are separated to the depth and are trained, are obtained
Depth separates convolutional neural networks model；
Model compression unit 103, for separating the weight coefficient of the convolutional layer in convolutional neural networks model to depth
It is compressed, obtains trained depth and separate convolutional neural networks model；
Face identification unit 104, image pyramid to trained three level for inputting facial image to be identified join
Convolutional neural networks carry out image preprocessing, then export pretreatment image to trained depth and separate convolutional neural networks
Model carries out recognition of face.
The present invention carries out image preprocessing using three levels connection convolutional neural networks, and trains progress depth separable with this
Convolutional neural networks model, the weight coefficient for further separating the convolutional layer in convolutional neural networks model to depth are pressed
Contracting obtains trained depth and separates convolutional neural networks model, and carries out recognition of face to facial image with this；By adopting
Convolution is separated with depth and reduces calculation amount, reduces model size, and by separating in convolutional neural networks model to depth
The weight coefficient of convolutional layer compressed so that model reduces 35 to 49 times.
Feature and performance of the invention are described in further detail with reference to embodiments.
Embodiment 1
A kind of lightweight face identification method that convolution is separated based on depth that present pre-ferred embodiments provide such as is schemed
Shown in 1, comprising:
S1, three levels of building join convolutional neural networks, and input the image pyramid of the first facial image data set, to institute
It states three levels connection convolutional neural networks to be trained, obtains trained three levels connection convolutional neural networks.
Specifically, the structure of the three levels connection convolutional neural networks, as shown in Figure 2, comprising:
First layer P-Net, including 3 convolution modules, wherein first convolution module include a 3x3 convolutional layer and
The pond layer of one 2x2, second convolution module and third convolution module respectively include the convolutional layer of a 3x3；
Second layer R-Net, including 4 convolution modules, wherein first convolution module and second convolution module respectively include
The convolutional layer of one 3x3 and the pond layer of a 3x3, third convolution module includes the convolutional layer of a 2x2, the 4th convolution
Module is a full articulamentum；
Third layer O-Net, including 5 convolution modules, wherein first convolution module, second convolution module and third
A convolution module respectively includes the convolutional layer of a 3x3 and the pond layer of a 3x3, and the 4th convolution module includes a 2x2
Convolutional layer, the 5th convolution module are a full articulamentums.
Then the image pyramid for inputting the first facial image data set carries out three level connection convolutional neural networks
Training obtains trained three levels connection convolutional neural networks.Three levels join the convolutional layer in convolutional neural networks and use
Xavier carries out initiation parameter.First facial image data set derives from two disclosed people of Winder_face and CelabA
Face image data set, Winder_face provide Face datection data set, and the coordinate letter of real human face frame is labelled on big figure
Breath, CelabA provide the human face data collection of 5 key points.According to the difference of the task of participation, by the first facial image data set
It is divided into four classes: face positive sample (positives), non-face negative sample (negatives), part face sample
(partfaces), key point (landmark).Face positive sample, non-face negative sample, part face are by the frame and mark that take at random
The size of the overlapping region ratio for the real human face frame set determines, is non-face negative sample less than 0.3, between 0.4 to 0.65
For part face sample, being greater than 0.65 is face positive sample.Face positive sample and non-face negative sample participate in classification task, people
Face positive sample and part face sample participate in recurrence task, and crucial point data participates in key point and returns task.Key point is returned
Return only practical in third layer, the face frame position of crucial point data can be obtained by the model inspection of the first two network at this time,
Or by the coordinate position of key point fit Lai.Sample proportion is as follows, face positive sample: non-face negative sample: part face
Sample: key point=1:3:1:2.Then the image pyramid of the first facial image data set is constructed as three layers of concatenated convolutional mind
Input through network, image pyramid are to carry out echelon to image to obtain set of the image under different scale to down-sampling；
Using cross entropy loss function, the quadratic sum loss function pair of quadratic sum loss function and 5 characteristic points and the data demarcated
Three levels connection convolutional neural networks are trained, and in order to obtain better effect, penalty values maximum 70% are only taken to carry out every time
Simple sample is ignored in backpropagation.
S2, depth of the building based on AM-Softmax loss function separate convolutional neural networks, and input the second face
Image data set separates convolutional neural networks to the depth and is trained, and obtains trained depth and separates convolution mind
Through network model.Second face image data collection derives from two disclosed face image datas of CASIA-Webface and LFW
Collection.
It constructs the depth based on AM-Softmax loss function and separates convolutional neural networks, as shown in figure 3, its structure packet
Include 17 convolution modules；
First convolution module is the common convolutional layer of a 3x3；
Second convolution module to the 14th convolution module is all that depth separates convolution module；Depth separates convolution
Common 3x3 convolutional layer is decomposed the point convolution of one 1x1 of depth convolution sum for a 3x3 by module, and depth convolution will be every
A convolution kernel is applied to each channel, and 1x1 convolution is used to combine the output of channel convolution, and this decomposition can effectively reduce
Calculation amount reduces model size；
15th convolution module is the average pond layer of a 7x7；
16th convolution module is a full articulamentum；
17th convolution module is AM-Softmax layers.
In addition to last full articulamentum, i.e., first convolution module is into the 16th convolution module, each convolution module
All gradient decline is accelerated by batch normalization and nonlinear activation ReLU and is asked most with batch normalization and nonlinear activation afterwards
The speed of excellent solution, while improving precision.
Second face image data collection is finally entered in AM-Softmax loss function and is trained, and is used
RMSprop descent method is optimization method.AM-Softmax loss function improves traditional Softmax loss function, traditional
A parameter factors l, formula are introduced in Softmax loss function are as follows:
Wherein, s is zoom factor, and n, which refers to, n labeling, and i is i-th of labeling, yiIt is i-th dimension output, j
It is to remove yiOther are exported afterwards, and c is that n-1 removes yiThe sum of other outputs afterwards, WyiIt is AM-softmax layers of weight, fi
It is i-th of output of previous layer, cos θyiIt is the weight W in networkyiAnd fiInner product and,It will power
Weight WyiAnd fiCos distance be improved to cos θy- l can adjust the distance between face characteristic, so that people by parameter factors l
Face feature has in bigger class spacing and smaller class away from can obtain better effect in recognition of face.
Step S3, the weight coefficient for separating the convolutional layer in convolutional neural networks model to depth are compressed, are obtained
Trained depth separates convolutional neural networks model, as shown in figure 4, specifically including the weight trimming successively carried out, weight
Shared and quantization and huffman coding:
(1) the weight trimming, specifically: the weight coefficient matrix that depth separates convolutional neural networks model is established,
For any weight coefficient matrix, all weight coefficients lower than preset threshold are set to zero, and are expressed as two matrixes: first
A matrix includes all non-zero weight coefficients in the weight coefficient matrix, and second matrix includes all non-zero weight coefficients
Relative index value, then re -training depth separate the weight of convolutional neural networks model, to retain important weight.Its
In, preset threshold is tested by iteration and is determined.
(2) weight is shared and quantifies, specifically: it is poly- using k-means to the weight of reservation after carrying out weight trimming
Class algorithm is clustered, and all weights in each class share cluster centre, and the weight of cluster is then stored as a code
Book；Then weight is quantified using code book, then re -training code book；
(3) huffman coding (Huffman Cod ing), specifically: it is general to the appearance of the weight stored in code book
Rate coding, the index book of the weight after being finally stored as a coding.The high binary coding shorter with length of probability of occurrence,
The low longer binary coding of use length of probability of occurrence, to solve to encode bring redundancy issue different in size.It is final real
On the basis of keeping original accuracy rate now, model reduces 35 to 49 times.
S4, image pyramid to the trained three level connection convolutional neural networks for inputting facial image to be identified carry out figure
As pretreatment, then exports pretreatment image to trained depth and separate the progress recognition of face of convolutional neural networks model.
Wherein, image pyramid to the trained three level connection convolutional neural networks for inputting facial image to be identified carry out
Image preprocessing, as shown in figure 5, specifically including:
S41, face candidate is obtained using the image pyramid of facial image to be identified as input by first layer P-Net
Window and frame regression vector, while face candidate window is corrected using frame regression vector, then merged using non-maxima suppression
The face candidate window of high superposed；
S42, frame regression vector is utilized using the output image of step S41 as inputting by second layer R-Net again
Face candidate window is finely tuned, then removes overlapping candidate window using non-maxima suppression；
S43, frame regression vector is utilized using the output image of step S42 as inputting by third layer O-Net again
Face candidate window is finely tuned, then removes overlapping candidate window using non-maxima suppression, finally output includes face candidate window and five
The pretreatment image of a characteristic point position；
Then output pretreatment image to trained depth separates convolutional neural networks model and carries out recognition of face:
By two face figures to be identified after trained three level connection convolutional neural networks carry out image preprocessing
Convolutional neural networks are separated as being input to depth, the face characteristic of 1 × 1 × 128 dimensions are extracted respectively, by two facial images
Feature vector cosine similarity as measure two faces between similitude measurement, i.e., the angle of two feature vectors
Embodiment 2
By the lightweight face identification method for separating convolution based on depth of embodiment 1, can establish a kind of based on deep
The lightweight face identification system of separable convolution is spent, as shown in Figure 6, comprising:
A kind of lightweight face identification system separating convolution based on depth, comprising:
First nerves network training unit 101 joins convolutional neural networks for constructing three levels, and inputs the first face figure
As the image pyramid of data set, three level connection convolutional neural networks are trained, obtain trained three levels connection
Convolutional neural networks；
Nervus opticus network training unit 102, for constructing the separable volume of the depth based on AM-Softmax loss function
Product neural network, and the second face image data collection is inputted, convolutional neural networks are separated to the depth and are trained, are obtained
Depth separates convolutional neural networks model；
Model compression unit 103, for separating the weight coefficient of the convolutional layer in convolutional neural networks model to depth
It is compressed, obtains trained depth and separate convolutional neural networks model；
Face identification unit 104, image pyramid to trained three level for inputting facial image to be identified join
Convolutional neural networks carry out image preprocessing, then export pretreatment image to trained depth and separate convolutional neural networks
Model carries out recognition of face.
It is apparent to those skilled in the art that the convenience and letter for description are bought, foregoing description based on
Depth separates the lightweight face identification system of convolution and the specific work process of its each functional unit, can refer to aforementioned
Corresponding process in embodiment of the method, details are not described herein.
Above-mentioned bright each functional unit can integrate in one processing unit, is also possible to the independent physics of each unit and deposits
It can also be integrated in one unit with two or more units.Above-mentioned integrated unit can both use the shape of hardware
Formula is realized, can also be realized in the form of software functional units.
If integrated each functional unit is realized in the form of SFU software functional unit and sells as independent product
Or it in use, can store in a computer readable storage medium.Based on this understanding, technical solution of the present invention
Substantially all or part of the part that contributes to existing technology or the technical solution can be with software product in other words
Form embody, which is stored in a storage medium, including some instructions use so that one
Computer equipment (can be smart phone, tablet computer, personal computer, server or the network equipment etc.) executes this hair
The all or part of the steps of the bright lightweight face identification method.And storage medium above-mentioned include: USB flash disk, mobile hard disk, only
Read memory (ROM, Read-OnlyMemory), random access memory (RAM, Random Access Memory), magnetic disk or
The various media that can store program code such as person's CD.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all in essence of the invention
Made any modifications, equivalent replacements, and improvements etc., should all be included in the protection scope of the present invention within mind and principle.
Claims (4)
1. a kind of lightweight face identification method for separating convolution based on depth, which is characterized in that specifically include:
S1, three levels of building join convolutional neural networks, and input the image pyramid of the first facial image data set, to described three
Level connection convolutional neural networks are trained, and obtain trained three levels connection convolutional neural networks；
S2, depth of the building based on AM-Softmax loss function separate convolutional neural networks, and input the second facial image
Data set separates convolutional neural networks to the depth and is trained, and obtains depth and separates convolutional neural networks model；
S3, the weight coefficient that the convolutional layer in convolutional neural networks model is separated to depth compress, and obtain trained
Depth separates convolutional neural networks model；
S4, the image pyramid of input facial image to be identified are pre- to trained three level connection convolutional neural networks progress image
Then processing exports pretreatment image to trained depth and separates the progress recognition of face of convolutional neural networks model.
2. the lightweight face identification method of convolution is separated based on depth as described in claim 1, which is characterized in that step
Depth based on AM-Softmax loss function in S2 separates convolutional neural networks, and structure includes: 17 convolution modules；
First convolution module is the common convolutional layer of a 3x3；
Second convolution module to the 14th convolution module is that depth separates convolution module；
15th convolution module is the average pond layer of a 7x7；
16th convolution module is a full articulamentum；
17th convolution module is AM-Softmax layers.
3. the lightweight face identification method of convolution is separated based on depth as claimed in claim 2, which is characterized in that described
First convolution module swashs with batch normalization with non-linear after each convolution module into the 16th convolution module
It is living.
4. a kind of lightweight face identification system for separating convolution based on depth characterized by comprising
First nerves network training unit (101) joins convolutional neural networks for constructing three levels, and inputs the first facial image
The image pyramid of data set is trained three level connection convolutional neural networks, obtains trained three levels connection volume
Product neural network；
Nervus opticus network training unit (102) separates convolution for constructing the depth based on AM-Softmax loss function
Neural network, and the second face image data collection is inputted, convolutional neural networks are separated to the depth and are trained, depth is obtained
Spend separable convolutional neural networks model；
Model compression unit (103), for depth is separated the weight coefficient of the convolutional layer in convolutional neural networks model into
Row compression obtains trained depth and separates convolutional neural networks model；
Face identification unit (104), the image pyramid for inputting facial image to be identified joins to trained three level to be rolled up
Product neural network carries out image preprocessing, then exports pretreatment image to trained depth and separates convolutional neural networks mould
Type carries out recognition of face.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810921647.6A CN109063666A (en) | 2018-08-14 | 2018-08-14 | The lightweight face identification method and system of convolution are separated based on depth |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810921647.6A CN109063666A (en) | 2018-08-14 | 2018-08-14 | The lightweight face identification method and system of convolution are separated based on depth |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN109063666A true CN109063666A (en) | 2018-12-21 |
Family
ID=64678390
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810921647.6A Pending CN109063666A (en) | 2018-08-14 | 2018-08-14 | The lightweight face identification method and system of convolution are separated based on depth |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109063666A (en) |
Cited By (39)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109828251A (en) * | 2019-03-07 | 2019-05-31 | 中国人民解放军海军航空大学 | Radar target identification method based on feature pyramid light weight convolutional neural networks |
| CN109858362A (en) * | 2018-12-28 | 2019-06-07 | 浙江工业大学 | A kind of mobile terminal method for detecting human face based on inversion residual error structure and angle associated losses function |
| CN109902603A (en) * | 2019-02-18 | 2019-06-18 | 苏州清研微视电子科技有限公司 | Driver identity identification authentication method and system based on infrared image |
| CN110222841A (en) * | 2019-06-17 | 2019-09-10 | 苏州思必驰信息科技有限公司 | Neural network training method and device based on spacing loss function |
| CN110263956A (en) * | 2019-05-20 | 2019-09-20 | 平安科技（深圳）有限公司 | Dining method, apparatus and computer readable storage medium based on recognition of face |
| CN110335260A (en) * | 2019-06-27 | 2019-10-15 | 华东送变电工程有限公司 | A kind of power cable damage testing method based on light convolutional neural networks |
| CN110458097A (en) * | 2019-08-09 | 2019-11-15 | 软通动力信息技术有限公司 | A kind of face picture recognition methods, device, electronic equipment and storage medium |
| CN110516642A (en) * | 2019-08-30 | 2019-11-29 | 电子科技大学 | A kind of lightweight face 3D critical point detection method and system |
| CN110781896A (en) * | 2019-10-17 | 2020-02-11 | 暨南大学 | Track garbage identification method, cleaning method, system and resource allocation method |
| CN110781784A (en) * | 2019-10-18 | 2020-02-11 | 高新兴科技集团股份有限公司 | Face recognition method, device and equipment based on double-path attention mechanism |
| CN110969089A (en) * | 2019-11-01 | 2020-04-07 | 北京交通大学 | Lightweight face recognition system and recognition method under noise environment |
| CN111204452A (en) * | 2020-02-10 | 2020-05-29 | 北京建筑大学 | Target detection system based on miniature aircraft |
| CN111353577A (en) * | 2018-12-24 | 2020-06-30 | Tcl集团股份有限公司 | Optimization method and device of multi-task-based cascade combination model and terminal equipment |
| CN111368941A (en) * | 2020-04-10 | 2020-07-03 | 浙江大华技术股份有限公司 | Image processing method and device and computer storage medium |
| CN111401292A (en) * | 2020-03-25 | 2020-07-10 | 成都东方天呈智能科技有限公司 | Face recognition network construction method fusing infrared image training |
| CN111401127A (en) * | 2020-01-16 | 2020-07-10 | 创意信息技术股份有限公司 | Human face living body detection joint judgment method |
| CN111400535A (en) * | 2020-03-11 | 2020-07-10 | 广东宜教通教育有限公司 | Lightweight face recognition method, system, computer device and storage medium |
| CN111402143A (en) * | 2020-06-03 | 2020-07-10 | 腾讯科技（深圳）有限公司 | Image processing method, device, equipment and computer readable storage medium |
| CN111428224A (en) * | 2020-04-02 | 2020-07-17 | 苏州杰锐思智能科技股份有限公司 | Computer account login method based on face recognition |
| CN111444884A (en) * | 2020-04-22 | 2020-07-24 | 万翼科技有限公司 | Method, apparatus and computer-readable storage medium for recognizing a component in an image |
| CN111860080A (en) * | 2019-04-30 | 2020-10-30 | 杰创智能科技股份有限公司 | Rapid face detection model based on channel separation type convolution |
| CN111931551A (en) * | 2020-05-26 | 2020-11-13 | 东南大学 | Face detection method based on lightweight cascade network |
| CN111950723A (en) * | 2019-05-16 | 2020-11-17 | 武汉Tcl集团工业研究院有限公司 | Neural network model training method, image processing method, device and terminal equipment |
| CN112215258A (en) * | 2020-09-17 | 2021-01-12 | 九牧厨卫股份有限公司 | Toilet bowl flushing control method and system and toilet bowl |
| CN112233675A (en) * | 2020-10-22 | 2021-01-15 | 中科院微电子研究所南京智能技术研究院 | Voice awakening method and system based on separation convolutional neural network |
| CN112258559A (en) * | 2020-10-26 | 2021-01-22 | 上海萱闱医疗科技有限公司 | Intelligent running timing scoring system and method based on multi-target tracking |
| CN112528830A (en) * | 2020-12-07 | 2021-03-19 | 南京航空航天大学 | Lightweight CNN mask face pose classification method combined with transfer learning |
| CN112560687A (en) * | 2020-04-03 | 2021-03-26 | 上海应用技术大学 | Face recognition method |
| CN112580406A (en) * | 2019-09-30 | 2021-03-30 | 北京中关村科金技术有限公司 | Face comparison method and device and storage medium |
| CN112580772A (en) * | 2019-09-30 | 2021-03-30 | 华为技术有限公司 | Compression method and device of convolutional neural network |
| CN112766176A (en) * | 2021-01-21 | 2021-05-07 | 深圳市安软科技股份有限公司 | Training method of lightweight convolutional neural network and face attribute recognition method |
| CN112800927A (en) * | 2021-01-25 | 2021-05-14 | 北京工业大学 | AM-Softmax loss-based butterfly image fine granularity identification method |
| CN112861602A (en) * | 2020-12-10 | 2021-05-28 | 华南理工大学 | Face living body recognition model compression and transplantation method based on depth separable convolution |
| CN113191220A (en) * | 2021-04-15 | 2021-07-30 | 广州紫为云科技有限公司 | Deep learning-based double-layer license plate recognition method |
| CN113505628A (en) * | 2021-04-02 | 2021-10-15 | 上海师范大学 | Target identification method based on lightweight neural network and application thereof |
| CN113538312A (en) * | 2021-07-22 | 2021-10-22 | 北京工业大学 | Real-time low-illumination image enhancement method, system, equipment and storage medium |
| WO2021253941A1 (en) * | 2020-06-18 | 2021-12-23 | 华为技术有限公司 | Neural network model training method, image classification method, text translation method and apparatus, and device |
| CN113505740B (en) * | 2021-07-27 | 2023-10-10 | 北京工商大学 | Face recognition method based on transfer learning and convolutional neural network |
| CN117593698A (en) * | 2023-12-04 | 2024-02-23 | 成都信息工程大学 | Regional target intrusion detection method, device and system and storage medium |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107609497A (en) * | 2017-08-31 | 2018-01-19 | 武汉世纪金桥安全技术有限公司 | The real-time video face identification method and system of view-based access control model tracking technique |
| CN107944555A (en) * | 2017-12-07 | 2018-04-20 | 广州华多网络科技有限公司 | Method, storage device and the terminal that neutral net is compressed and accelerated |
| US20180137406A1 (en) * | 2016-11-15 | 2018-05-17 | Google Inc. | Efficient Convolutional Neural Networks and Techniques to Reduce Associated Computational Costs |
| CN108229442A (en) * | 2018-02-07 | 2018-06-29 | 西南科技大学 | Face fast and stable detection method in image sequence based on MS-KCF |
| CN108304928A (en) * | 2018-01-26 | 2018-07-20 | 西安理工大学 | Compression method based on the deep neural network for improving cluster |
-
2018
- 2018-08-14 CN CN201810921647.6A patent/CN109063666A/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180137406A1 (en) * | 2016-11-15 | 2018-05-17 | Google Inc. | Efficient Convolutional Neural Networks and Techniques to Reduce Associated Computational Costs |
| CN107609497A (en) * | 2017-08-31 | 2018-01-19 | 武汉世纪金桥安全技术有限公司 | The real-time video face identification method and system of view-based access control model tracking technique |
| CN107944555A (en) * | 2017-12-07 | 2018-04-20 | 广州华多网络科技有限公司 | Method, storage device and the terminal that neutral net is compressed and accelerated |
| CN108304928A (en) * | 2018-01-26 | 2018-07-20 | 西安理工大学 | Compression method based on the deep neural network for improving cluster |
| CN108229442A (en) * | 2018-02-07 | 2018-06-29 | 西南科技大学 | Face fast and stable detection method in image sequence based on MS-KCF |
Non-Patent Citations (6)
| Title |
|---|
| ANDREW G.HOWARD 等: "MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications", 《ARXIV:1704.04861V1》 * |
| FENG WANG 等: "Additive Margin Softmax for Face Verification", 《IEEE SIGNAL PROCESSING LETTERS》 * |
| KAIPENG ZHANG 等: "Joint face detection and alignment using multitask cascaded convolutional networks", 《IEEE SIGNAL PROCESSING LETTERS》 * |
| SONG HAN 等: "DEEP COMPRESSION: COMPRESSING DEEP NEURAL NETWORKS WITH PRUNING, TRAINED QUANTIZATION AND HUFFMAN CODING", 《ARXIV:1510.00149V5》 * |
| WEIYANG LIU 等: "Sphereface: Deep hypersphere embedding for face recognition", 《2017 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION（CVPR）》 * |
| 任梓涵 等: "基于视觉跟踪的实时视频人脸识别", 《厦门大学学报(自然科学版)》 * |
Cited By (55)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111353577A (en) * | 2018-12-24 | 2020-06-30 | Tcl集团股份有限公司 | Optimization method and device of multi-task-based cascade combination model and terminal equipment |
| CN111353577B (en) * | 2018-12-24 | 2024-01-23 | Tcl科技集团股份有限公司 | Multi-task-based cascade combination model optimization method and device and terminal equipment |
| CN109858362A (en) * | 2018-12-28 | 2019-06-07 | 浙江工业大学 | A kind of mobile terminal method for detecting human face based on inversion residual error structure and angle associated losses function |
| CN109902603A (en) * | 2019-02-18 | 2019-06-18 | 苏州清研微视电子科技有限公司 | Driver identity identification authentication method and system based on infrared image |
| CN109828251A (en) * | 2019-03-07 | 2019-05-31 | 中国人民解放军海军航空大学 | Radar target identification method based on feature pyramid light weight convolutional neural networks |
| CN111860080A (en) * | 2019-04-30 | 2020-10-30 | 杰创智能科技股份有限公司 | Rapid face detection model based on channel separation type convolution |
| CN111950723B (en) * | 2019-05-16 | 2024-05-21 | 武汉Tcl集团工业研究院有限公司 | Neural network model training method, image processing method, device and terminal equipment |
| CN111950723A (en) * | 2019-05-16 | 2020-11-17 | 武汉Tcl集团工业研究院有限公司 | Neural network model training method, image processing method, device and terminal equipment |
| CN110263956A (en) * | 2019-05-20 | 2019-09-20 | 平安科技（深圳）有限公司 | Dining method, apparatus and computer readable storage medium based on recognition of face |
| CN110222841A (en) * | 2019-06-17 | 2019-09-10 | 苏州思必驰信息科技有限公司 | Neural network training method and device based on spacing loss function |
| CN110335260A (en) * | 2019-06-27 | 2019-10-15 | 华东送变电工程有限公司 | A kind of power cable damage testing method based on light convolutional neural networks |
| CN110335260B (en) * | 2019-06-27 | 2023-07-14 | 华东送变电工程有限公司 | Power cable damage detection method based on light convolutional neural network |
| CN110458097A (en) * | 2019-08-09 | 2019-11-15 | 软通动力信息技术有限公司 | A kind of face picture recognition methods, device, electronic equipment and storage medium |
| CN110516642A (en) * | 2019-08-30 | 2019-11-29 | 电子科技大学 | A kind of lightweight face 3D critical point detection method and system |
| CN112580772A (en) * | 2019-09-30 | 2021-03-30 | 华为技术有限公司 | Compression method and device of convolutional neural network |
| CN112580406A (en) * | 2019-09-30 | 2021-03-30 | 北京中关村科金技术有限公司 | Face comparison method and device and storage medium |
| CN112580772B (en) * | 2019-09-30 | 2024-04-26 | 华为技术有限公司 | Compression method and device for convolutional neural network |
| CN110781896A (en) * | 2019-10-17 | 2020-02-11 | 暨南大学 | Track garbage identification method, cleaning method, system and resource allocation method |
| CN110781896B (en) * | 2019-10-17 | 2022-07-19 | 暨南大学 | Track garbage identification method, cleaning method, system and resource allocation method |
| CN110781784A (en) * | 2019-10-18 | 2020-02-11 | 高新兴科技集团股份有限公司 | Face recognition method, device and equipment based on double-path attention mechanism |
| CN110969089B (en) * | 2019-11-01 | 2023-08-18 | 北京交通大学 | Lightweight face recognition system and recognition method in noise environment |
| CN110969089A (en) * | 2019-11-01 | 2020-04-07 | 北京交通大学 | Lightweight face recognition system and recognition method under noise environment |
| CN111401127A (en) * | 2020-01-16 | 2020-07-10 | 创意信息技术股份有限公司 | Human face living body detection joint judgment method |
| CN111204452A (en) * | 2020-02-10 | 2020-05-29 | 北京建筑大学 | Target detection system based on miniature aircraft |
| CN111204452B (en) * | 2020-02-10 | 2021-07-16 | 北京建筑大学 | Target detection system based on miniature aircraft |
| CN111400535A (en) * | 2020-03-11 | 2020-07-10 | 广东宜教通教育有限公司 | Lightweight face recognition method, system, computer device and storage medium |
| CN111401292A (en) * | 2020-03-25 | 2020-07-10 | 成都东方天呈智能科技有限公司 | Face recognition network construction method fusing infrared image training |
| CN111428224B (en) * | 2020-04-02 | 2023-10-13 | 苏州杰锐思智能科技股份有限公司 | Face recognition-based computer account login method |
| CN111428224A (en) * | 2020-04-02 | 2020-07-17 | 苏州杰锐思智能科技股份有限公司 | Computer account login method based on face recognition |
| CN112560687A (en) * | 2020-04-03 | 2021-03-26 | 上海应用技术大学 | Face recognition method |
| CN112560687B (en) * | 2020-04-03 | 2024-01-30 | 上海应用技术大学 | Face recognition method |
| CN111368941B (en) * | 2020-04-10 | 2023-09-01 | 浙江大华技术股份有限公司 | Image processing method, device and computer storage medium |
| CN111368941A (en) * | 2020-04-10 | 2020-07-03 | 浙江大华技术股份有限公司 | Image processing method and device and computer storage medium |
| CN111444884A (en) * | 2020-04-22 | 2020-07-24 | 万翼科技有限公司 | Method, apparatus and computer-readable storage medium for recognizing a component in an image |
| CN111931551B (en) * | 2020-05-26 | 2022-04-12 | 东南大学 | Face detection method based on lightweight cascade network |
| CN111931551A (en) * | 2020-05-26 | 2020-11-13 | 东南大学 | Face detection method based on lightweight cascade network |
| CN111402143A (en) * | 2020-06-03 | 2020-07-10 | 腾讯科技（深圳）有限公司 | Image processing method, device, equipment and computer readable storage medium |
| WO2021253941A1 (en) * | 2020-06-18 | 2021-12-23 | 华为技术有限公司 | Neural network model training method, image classification method, text translation method and apparatus, and device |
| CN112215258A (en) * | 2020-09-17 | 2021-01-12 | 九牧厨卫股份有限公司 | Toilet bowl flushing control method and system and toilet bowl |
| CN112215258B (en) * | 2020-09-17 | 2022-10-18 | 九牧厨卫股份有限公司 | Toilet bowl flushing control method and system and toilet bowl |
| CN112233675B (en) * | 2020-10-22 | 2023-10-27 | 中科南京智能技术研究院 | Voice wake-up method and system based on separated convolutional neural network |
| CN112233675A (en) * | 2020-10-22 | 2021-01-15 | 中科院微电子研究所南京智能技术研究院 | Voice awakening method and system based on separation convolutional neural network |
| CN112258559A (en) * | 2020-10-26 | 2021-01-22 | 上海萱闱医疗科技有限公司 | Intelligent running timing scoring system and method based on multi-target tracking |
| CN112258559B (en) * | 2020-10-26 | 2024-05-07 | 上海萱闱医疗科技有限公司 | Intelligent running timing scoring system and method based on multi-target tracking |
| CN112528830A (en) * | 2020-12-07 | 2021-03-19 | 南京航空航天大学 | Lightweight CNN mask face pose classification method combined with transfer learning |
| CN112861602A (en) * | 2020-12-10 | 2021-05-28 | 华南理工大学 | Face living body recognition model compression and transplantation method based on depth separable convolution |
| CN112766176A (en) * | 2021-01-21 | 2021-05-07 | 深圳市安软科技股份有限公司 | Training method of lightweight convolutional neural network and face attribute recognition method |
| CN112766176B (en) * | 2021-01-21 | 2023-12-01 | 深圳市安软科技股份有限公司 | Training method of lightweight convolutional neural network and face attribute recognition method |
| CN112800927A (en) * | 2021-01-25 | 2021-05-14 | 北京工业大学 | AM-Softmax loss-based butterfly image fine granularity identification method |
| CN112800927B (en) * | 2021-01-25 | 2024-03-29 | 北京工业大学 | Butterfly image fine-granularity identification method based on AM-Softmax loss |
| CN113505628A (en) * | 2021-04-02 | 2021-10-15 | 上海师范大学 | Target identification method based on lightweight neural network and application thereof |
| CN113191220A (en) * | 2021-04-15 | 2021-07-30 | 广州紫为云科技有限公司 | Deep learning-based double-layer license plate recognition method |
| CN113538312A (en) * | 2021-07-22 | 2021-10-22 | 北京工业大学 | Real-time low-illumination image enhancement method, system, equipment and storage medium |
| CN113505740B (en) * | 2021-07-27 | 2023-10-10 | 北京工商大学 | Face recognition method based on transfer learning and convolutional neural network |
| CN117593698A (en) * | 2023-12-04 | 2024-02-23 | 成都信息工程大学 | Regional target intrusion detection method, device and system and storage medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109063666A (en) | The lightweight face identification method and system of convolution are separated based on depth | |
| Chen et al. | Convolution neural network for automatic facial expression recognition | |
| Bunrit et al. | Improving the representation of cnn based features by autoencoder for a task of construction material image classification | |
| Sun et al. | Combining multimodal features within a fusion network for emotion recognition in the wild | |
| CN107169117B (en) | Hand-drawn human motion retrieval method based on automatic encoder and DTW | |
| Wang et al. | Soft person reidentification network pruning via blockwise adjacent filter decaying | |
| Lee et al. | Face image retrieval using sparse representation classifier with gabor-lbp histogram | |
| CN113205002B (en) | Low-definition face recognition method, device, equipment and medium for unlimited video monitoring | |
| CN109344720B (en) | Emotional state detection method based on self-adaptive feature selection | |
| US11908222B1 (en) | Occluded pedestrian re-identification method based on pose estimation and background suppression | |
| Balaji et al. | Multi-level feature fusion for group-level emotion recognition | |
| Dang et al. | Facial expression recognition: A survey and its applications | |
| Liu et al. | Image retrieval using CNN and low-level feature fusion for crime scene investigation image database | |
| Sen et al. | Face recognition using deep convolutional network and one-shot learning | |
| Chauhan et al. | Analysis of Intelligent movie recommender system from facial expression | |
| CN105389588A (en) | Multi-semantic-codebook-based image feature representation method | |
| Singh et al. | Age, gender prediction and emotion recognition using convolutional neural network | |
| Schoneveld et al. | Towards a general deep feature extractor for facial expression recognition | |
| Hdioud et al. | Facial expression recognition of masked faces using deep learning | |
| Kumar et al. | One-shot face recognition | |
| Li | Parallel two-class 3D-CNN classifiers for video classification | |
| Jadhav et al. | Content based facial emotion recognition model using machine learning algorithm | |
| Nimbarte et al. | Biased face patching approach for age invariant face recognition using convolutional neural network | |
| Usgan et al. | Deep learning pre-trained model as feature extraction in facial recognition for identification of electronic identity cards by considering age progressing | |
| CN117011219A (en) | Method, apparatus, device, storage medium and program product for detecting quality of article |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181221 |