CN113723238B - Face lightweight network model construction method and face recognition method - Google Patents
Face lightweight network model construction method and face recognition method Download PDFInfo
- Publication number
- CN113723238B CN113723238B CN202110947540.0A CN202110947540A CN113723238B CN 113723238 B CN113723238 B CN 113723238B CN 202110947540 A CN202110947540 A CN 202110947540A CN 113723238 B CN113723238 B CN 113723238B
- Authority
- CN
- China
- Prior art keywords
- face
- face recognition
- network
- lightweight
- 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.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 35
- 238000010276 construction Methods 0.000 title claims abstract description 10
- 238000012549 training Methods 0.000 claims abstract description 58
- 238000004821 distillation Methods 0.000 claims abstract description 37
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 20
- 239000013598 vector Substances 0.000 claims description 30
- 238000012360 testing method Methods 0.000 claims description 18
- 238000004422 calculation algorithm Methods 0.000 claims description 9
- 230000001537 neural effect Effects 0.000 claims 2
- 230000006870 function Effects 0.000 description 20
- 238000004364 calculation method Methods 0.000 description 12
- 238000005516 engineering process Methods 0.000 description 6
- 238000013528 artificial neural network Methods 0.000 description 4
- 238000012795 verification Methods 0.000 description 4
- 238000005457 optimization Methods 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 230000001680 brushing effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000002452 interceptive effect Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- 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
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- 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
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Computing Systems (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Image Analysis (AREA)
- Collating Specific Patterns (AREA)
Abstract
The invention discloses a face recognition lightweight network model construction method and a face recognition method based on embedded space distillation, comprising the following steps: A. acquiring a designated number of face digital images marked with face identity category information as a training data set; B. training a face recognition convolutional neural network of one weight by utilizing the training data set based on a softmax loss function to obtain a trained model; C. based on the embedded space distillation loss function, training the lightweight face recognition convolutional neural network by using the training model and the training data set obtained in the B stage, and obtaining a trained lightweight face recognition model. The invention can fully utilize the priori knowledge obtained by weight model training, effectively improve the accuracy of the lightweight face recognition model, and adapt to the face recognition task of the embedded platform with limited computing resources.
Description
Technical Field
The invention relates to the technical fields of computer vision, face recognition, pattern recognition, machine learning, deep learning and the like, in particular to a construction method of a face lightweight network model and a face recognition method.
Background
The development of Face Recognition technology (Face Recognition) has started from the 70 s of the 20 th century, and is one of the hottest research problems in the Computer Vision field (Computer Vision) so far. As a biological feature recognition technology which is most easy to popularize, the face recognition has a plurality of application scenes, and the most traditional scene of the face recognition technology belongs to the fields of security, access control and monitoring. Whether in the gate of the security check access gate or the safety door lock of the bank safe, the body shadow of the face recognition appears. Compared with the traditional security means such as passwords, certificates, door cards and the like, the face recognition is a natural identity card attached to the face, and has the attribute of extremely difficult counterfeiting. Face recognition based techniques are widely used in the military and civilian fields. In addition, the face recognition technology is widely applied to online and offline face brushing payment systems by using the technology of interactive instruction + continuity judgment +3D judgment. With the wide application of face recognition technology, the demands of algorithms of mobile terminals and embedded devices are increasing, and the current face recognition algorithm is mainly based on a convolutional neural network (Convolutional Neural Networks, CNN for short) model. The CNN model is large in general model, and meanwhile, on a mobile terminal of a mobile phone and an embedded platform with weak computing capability, the reasoning speed is low.
Based on the above problems, a method for constructing a lightweight face convolutional neural network is needed at present so as to realize that real-time calculation requirements can be rapidly inferred and met under the condition of low calculation force such as a mobile terminal, embedded equipment and the like, and face information can be accurately and effectively identified.
Disclosure of Invention
One of the technical problems to be solved by the invention is to provide a face light-weight network model construction method based on embedded space distillation, and the method is used for training a light-weight face convolutional neural network model, so that real-time calculation requirements can be rapidly inferred and met under the condition of low calculation force of a mobile terminal, embedded equipment and the like, the accuracy of the face convolutional network model can be ensured, and face information can be accurately and effectively identified.
The second technical problem to be solved by the invention is to provide a face recognition method which can rapidly infer and meet the real-time calculation requirement under the condition of low calculation power of a mobile terminal, embedded equipment and the like, can also ensure the accuracy of a face convolution network model, and can accurately and effectively recognize face information.
The invention is realized in the following way:
a method for constructing a lightweight network model of a human face comprises the following steps:
step A, acquiring a designated number of face digital images marked with face identity category information as a training data set;
step B, training a face recognition convolutional neural network T of one weight by utilizing the training data set based on a softmax loss function to obtain a trained model;
and C, training the lightweight face recognition convolutional neural network S by using the training model and the training data set obtained in the step B based on the embedded space distillation loss function, and obtaining a trained lightweight face recognition model.
Further, the step B specifically includes:
step B1, training by using the training data set, and then obtaining a face recognition convolutional network T;
step B2, define z T Outputting the last full-connection layer of the face recognition convolution network T; definition p T Output probabilities obtained for the face recognition convolution network T through softmax are then
Wherein,representing the T output probability p of the face recognition neural network T Is the i-th dimensional output of (2); />Representing the last full-connection output z of the face recognition convolution network T T Is the ith dimension of the inputDischarging; />Representing the last full-connection output z of the face recognition convolution network T T Is the j-th dimensional output of (2); k represents the total category number of the training data set; z T 、p T Column vectors of K dimension;
step B3, define f T Output characteristics of the face recognition neural network T are obtained; definition of the definitionFor the face recognition convolution network T corresponds to the y-th i Center vector of feature vector of class, +.>Represented as
Wherein,representing the training data set y i The number of samples of the class; />Representing the output characteristics of the j-th sample corresponding to the face recognition convolutional network T; />Column vectors of D dimension; and D is the dimension of the characteristic vector of the face recognition convolution network T.
Further, the step C includes:
step C1, randomly initializing a lightweight face recognition convolutional network S, and defining z S Outputting the last full-connection layer of the lightweight face recognition convolutional network S; definition p S Acquiring the lightweight face recognition convolutional network S through softmaxOutput probability of (2); definition f S Output characteristics of the lightweight face recognition convolutional network S are obtained;
step C2, defining an objective function 1, wherein the objective function 1 is as follows:
wherein N represents the total number of training data set samples; y is i Representing a class label corresponding to the ith sample;the ith sample is represented to correspond to the y th full connection layer of the last full connection layer of the lightweight face recognition convolutional network S i Outputting dimensions; />The j-th dimension output of the last full-connection layer of the lightweight face recognition convolutional network S corresponding to the i-th sample is represented; k represents the total category number of the training data set; τ and λ represent given constants, respectively;
step C3, defining an objective function 2, wherein the objective function 2 is as follows:
wherein N represents the total number of training data set samples; y is i Representing a class label corresponding to the ith sample; f (f) i S Representing the output characteristics of the ith sample corresponding to the lightweight face recognition convolutional network S;representing that the face recognition convolution network T corresponds to the y-th i A center vector of the class; />Representing the j-th class corresponding to the face recognition convolution network TIs defined by the center vector of (a);
step C4, according to the descriptions of step C2 and step C3, the objective function of the method is expressed as:
L total ＝λ 1 L 1 +λ 2 L 2
wherein lambda is 1 And lambda (lambda) 2 Is a constant;
step C5, incorporating the objective function L of step C4 total And as a loss function, updating the learnable model parameters in the lightweight face recognition convolutional network S according to a back propagation algorithm based on random gradient descent by using the training data set so as to obtain an updated face lightweight network model based on embedded space distillation.
Further, the step C5 further includes:
since the center vector from the face recognition convolution network T is directly fixed when optimizing the update network using a back propagation algorithm with random gradient descentThe model convergence difficulty and even the model convergence failure are easy to be caused, and the face light-weight network model based on the embedded space distillation adopts an approximate update center vector +.>Is a strategy of (1):
wherein,representing the corresponding y-th face recognition convolution network T after the T-th iteration i A center vector of the class;representing the class label y in the training dataset under the batch of the current iteration i Is the number of samples of (a); />And representing the output characteristic of the j-th sample corresponding to the lightweight face recognition convolutional network S.
Further, the method for determining the face lightweight network model based on embedded space distillation in the step C5 is as follows:
acquiring a specified number of face digital images marked with face identity information labels, and constructing a positive sample pair and a negative sample pair to be tested according to the face identity information labels to serve as a test data set; extracting features from the face digital images in all the test data sets by using the face lightweight network model based on embedded space distillation in the step C5, and calculating cosine distances of each sample pair in the test data sets by using the extracted features; and evaluating the face light-weight network model based on embedded space distillation by using a receiver operation characteristic curve (Receiver Operating Characteristic curve) according to the cosine distance of the positive and negative sample pairs.
A face recognition method based on the face light-weight network model construction method comprises the following steps:
step 1, acquiring a face digital image to be recognized;
and 2, inputting the face digital image into the face light-weight network model based on the embedded space distillation, extracting face features through the face light-weight network model based on the embedded space distillation, and carrying out matching recognition with pre-stored face digital image features.
The invention has the advantages that: the method for constructing the face light-weight network model based on embedded space distillation can effectively improve the recognition performance of the light-weight face convolution network. By training the lightweight face convolutional neural network model by using the method, the real-time calculation requirement can be rapidly inferred and met under the condition of low calculation force of a mobile terminal, embedded equipment and the like, the accuracy of the face convolutional neural network model can be ensured, and the face information can be accurately and effectively identified.
Drawings
FIG. 1 is a schematic diagram of an inventive face light-weight network learning construction method based on embedded spatial distillation;
fig. 2 is a flow chart of a method for constructing a face light-weight network learning based on embedded space distillation.
Detailed Description
In order to make the objects, technical embodiments and effects of the present invention more apparent, the present invention will be described in detail with reference to examples. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the present disclosure.
The invention provides a construction method of face light weight network learning based on embedded space distillation in order to improve face recognition performance of a light weight face convolution network under a low-power platform such as a mobile terminal, embedded equipment and the like. According to the method, based on the embedded space distillation loss function, the supervision information provided by the weight face recognition convolutional neural network T is used for constraining the lightweight face recognition convolutional neural network S, so that the accuracy of a lightweight face convolutional network model can be effectively guaranteed.
Example 1
As shown in fig. 1 and fig. 2, the present invention proposes a method for constructing a face lightweight network model based on embedded space distillation, where the method can be used for training a lightweight face convolutional network to complete a face recognition task of a mobile terminal or an embedded platform, and the method includes:
step S101, a specified number of face digital images with face identity information tags are obtained as a training data set. The invention uses a private data set as a face recognition training data set, wherein the data set comprises 11,416,230 face digital images of 979,403 people;
step S102, training a weight face convolution network model by using the training data set constructed in the step S101 based on the softmax loss function, obtaining a trained model T and extracting relevant embedded space information. The weight face convolution network model is formed by arranging 101 convolution layers from shallow to deep, and maxout is used as an activation function;
specifically, the method comprises the following steps:
and step S1021, training the weight face convolution network by using the training data set constructed in the step S101, and obtaining a trained weight face convolution network model T.
Step S1022, extracting embedded space information by using the model T trained in step S1021,
wherein z is T Outputting the last full-connection layer of the face recognition convolution network T; p is p T Output probability p obtained by softmax for face recognition convolution network T T Can be expressed as
Wherein,representing the T output probability p of the face recognition neural network T Is the i-th dimensional output of (2); />Representing the last full-connection output z of the face recognition convolution network T T Is the i-th dimensional output of (2); />Representing the last full-connection output z of the face recognition convolution network T T Is the j-th dimensional output of (2); k represents the total category number of the training data set; z T 、p T Column vectors of K dimension;
step S1023, extracting embedded space information by using the model T trained in step S1021,
wherein f T The output characteristics of the face recognition neural network T are obtained; definition of the definitionFor face recognition convolution network T corresponds to the y i A center vector of the feature vector of the class; />Can be expressed as
Wherein,representing training dataset y i The number of samples of the class; />Representing the output characteristics of the face recognition convolution network T corresponding to the j-th sample; />Column vectors of D dimension; d is the dimension of the characteristic vector of the face recognition convolution network T;
step S103, based on the embedded space distillation loss function, utilizing the training model T and the embedded space information z obtained in the step S102 T 、p T 、Training the lightweight face recognition convolutional neural network S by the training data set to obtain a trained lightweight face recognition model S;
specifically, the method comprises the following steps:
step S1031, defining an objective function 1;
wherein, the objective function 1 is:
wherein N represents the total number of training data set samples; y is i Representing a class label corresponding to the ith sample;y-th representing the last full connection layer of the lightweight face recognition convolutional network S corresponding to the i-th sample i Outputting dimensions; />The j-th dimension output of the last full-connection layer of the lightweight face recognition convolutional network S corresponding to the i-th sample is represented; k represents the total category number of the training data set; τ and λ represent given constants, respectively;
step S1032, defining an objective function 2;
wherein, objective function 2 is:
wherein N represents the total number of training data set samples; y is i Representing a class label corresponding to the ith sample; f (f) i S Representing the output characteristics of the lightweight face recognition convolutional network S corresponding to the ith sample;representing the y-th corresponding to face recognition convolution network T i A center vector of the class; />Representing a center vector representing the j-th class corresponding to the face recognition convolutional network T;
step S1033, according to S1031 and S1032, the objective function of the method may be expressed as:
L total ＝λ 1 L 1 +λ 2 L 2
wherein lambda is 1 And lambda (lambda) 2 Is a constant;
s1034, objective function L total As a loss function, updating the learnable model parameters in the lightweight face recognition convolutional network S by using a training data set according to a back propagation algorithm based on random gradient descent to obtain an updated basisA face light-weight network model embedded with space distillation;
step S104, when optimizing the update network by using the back propagation algorithm of random gradient descent, since the center vector from the face recognition convolution network T is directly fixedThe model convergence difficulty and even non-convergence are easy to cause, and the human face light weight network learning based on the embedded space distillation adopts an approximate updating center vector +.>Is a strategy of (1):
wherein,representing the corresponding y-th face recognition convolutional network T after the T-th iteration i A center vector of the class; />Representing the class label y in the training dataset under the batch of the current iteration i Is the number of samples of (a); />Representing the output characteristics of the lightweight face recognition convolutional network S corresponding to the jth sample;
step S105, judging whether the learnable parameters of the lightweight face convolutional neural network need to be continuously optimized and updated, and if yes, returning to the step S103. If not, executing step S106;
specifically, the method for judging whether the iterative optimization updating needs to be continued is as follows:
step S101, a designated number of face digital images with face identity label information are separated from the training data set obtained in the step S101, and a verification data set is constructed;
using the face light-weight network model S based on embedded space distillation after updating processing in the step S1034 to carry out classification test on the verification data set; and when the test classification accuracy of the verification data set does not reach the specified accuracy, judging that iteration optimization needs to be continued, and repeating the steps S103-S104. When judging that the test classification accuracy of the verification data set reaches the specified accuracy, executing step S106;
and step S106, storing the face light neural network model S based on the embedded space distillation.
In order to more clearly illustrate the remarkable superiority of the face light model constructed by the application and the existing face recognition algorithm, the face recognition model test result of the invention is disclosed as follows:
the invention provides two test data sets in total, wherein one of the two test data sets is used for acquiring 13,233 face digital images from a public data set Labeled Face in the Wild (LFW for short), and the total number of the face digital images is 5,749 people, and the face digital images are taken as a test set 1; secondly, a private data set is constructed, which contains 136,603 digital images of faces of 52,404 persons in total, and the digital images are taken as a test set 2. And testing the trained lightweight face convolutional neural network model by using the two test data sets.
Table 1 shows the comparison results of the face recognition model of the present invention and the other 5 existing face recognition models on test set 1 with a correct recognition rate (tpr@fpr=0.1%) of one thousandth and a detection hit rate (dir@far=1%) of one percent. Table 2 shows the comparison results of the face recognition model of the present invention with 5 other existing face recognition models in terms of the correct recognition rate of one ten thousandth (tpr@fpr=0.01%), the correct recognition rate of one ten thousandth (tpr@fpr=0.001%), and the preferred hit rate (Rank-1). Table 3 shows the results of the speed test of the lightweight face recognition model of the present invention and another 5 existing face recognition models on Intel (R) Core (TM) i9-9820X CPU@3.30GHz, intel (R) Xeon (R) CPU E5-2650v4@2.20GHz, and Hi 3516C. From a comparison of tables 1 and 2, it can be seen that the recognition rate of the method of the present invention on both data sets is better than that of the 5 comparison models selected. From table 3, it can be found that the model speed at which the method of the present invention was trained was faster than that of the 5 comparison models selected. Therefore, the model of the invention can rapidly infer and meet the real-time calculation requirement under the condition of low calculation power of a mobile terminal, embedded equipment and the like, can also ensure the accuracy of the face convolution network model, and accurately and effectively identify the face information.
Table 1:
table 2:
table 3:
example two
The invention further provides a face light-weight network model construction method based on embedded space distillation, which comprises the following steps:
step D1, acquiring a face digital image to be recognized;
and D2, inputting the face digital image into the face light-weight network model based on the embedded space distillation, extracting the face characteristics through the face light-weight network model based on the embedded space distillation, and carrying out matching recognition with the pre-stored face digital image characteristics.
In summary, the method for constructing the face light-weight network model based on the embedded space distillation is based on the embedded space distillation, and the method is used for training the light-weight face convolutional neural network model, so that the real-time calculation requirement can be rapidly inferred and met under the condition of low calculation force of a mobile terminal, embedded equipment and the like, the accuracy of the face convolutional network model can be ensured, and the face information can be accurately and effectively identified. Meanwhile, the innovative method provided by the invention can be applied to the tasks of pedestrian re-identification, vehicle re-identification and other object identification.
The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather to enable any modification, equivalent replacement, improvement or the like to be made within the spirit and principles of the invention.
Claims (4)
1. The method for constructing the face light-weight network model is characterized by comprising the following steps of:
step A, acquiring a designated number of face digital images marked with face identity category information as a training data set;
step B, based on softmax loss function, utilizing the training data set to carry out face recognition convolutional neural network on one weightTraining to obtain a trained model;
the step B specifically comprises the following steps:
step B1, obtaining a face recognition convolutional network after training by using the training data set；
Step B2, definitionConvolutional network for the face recognition>Outputting the last full connection layer; definitions->Convolutional network for the face recognition>Output probability obtained through softmax, then
Wherein,representing the face recognition neural network->Output probability->Is>Outputting dimensions; />Representing the face recognition convolution network->Last full connection output->Is>Outputting dimensions; />Representing the face recognition convolution network->Last full connection output->Is>Outputting dimensions; />Representing a total category number of the training dataset; />、/>All are->Column vectors of dimensions;
step B3, definitionFor the face recognition neural network->Output characteristics of (2); definitions->Convolutional network for the face recognition>Corresponding->Center vector of feature vector of class, +.>Represented as
Wherein,representing the training data set +.>The number of samples of the class; />Representing the face recognition convolution network->Corresponding->Output characteristics of the individual samples; />、/>Column vectors of D dimension; d is the face recognition convolution network +.>The dimension of the feature vector;
step C, based on the embedded space distillation loss function, utilizing the training model and the training data set obtained in the step B to carry out face recognition on the lightweight face recognition convolutional neural networkTraining to obtain a trained face light-weight network model based on embedded space distillation; comprising the following steps:
step C1, randomly initializing a lightweight face recognition convolutional networkDefinitions->Convolutional network for said lightweight face recognition>Outputting the last full connection layer; definitions->Convolutional network for said lightweight face recognition>Output probabilities obtained through softmax; definitions->Convolutional network for said lightweight face recognition>Output characteristics of (2);
step C2, defining an objective function 1, wherein the objective function 1 is as follows:
wherein,representing the total number of training data set samples; />Indicate->Class labels corresponding to the samples; />Represent the firstThe number of samples corresponds to the lightweight face recognition convolutional network +.>Is the last full connection layer +.>Outputting dimensions; />Indicate->The number of samples corresponds to the lightweight face recognition convolutional network +.>First of last full connection layer>Outputting dimensions; />Representing a total category number of the training dataset; />And->Respectively, a given constant;
step C3, defining an objective function 2, wherein the objective function 2 is as follows:
wherein,representing the total number of training data set samples; />Indicate->Class labels corresponding to the samples; />Indicate->The number of samples corresponds to the lightweight face recognition convolutional network +.>Output characteristics of (2); />Representing the face recognition convolution network->Corresponding->A center vector of the class; />Representing the face recognition convolution network->Corresponding->A center vector of the class;
step C4, according to the descriptions of step C2 and step C3, the objective function of the method is expressed as:
wherein,and->Is a constant;
step C5, incorporating the objective function of step C4Updating the lightweight face recognition convolutional network ++using the training data set as an embedded spatial distillation loss function according to a back propagation algorithm based on random gradient descent>The model parameters which can be learned in the process of updating are obtained, and the face light-weight network model based on embedded space distillation is obtained after updating.
2. The method for constructing a lightweight network model of a human face according to claim 1, wherein the step C5 further comprises:
when using a random gradient descent back propagation algorithm to optimize the update network, the face recognition convolution network is directly fixedCenter vector of +.>The model convergence difficulty and even the model convergence failure are easy to be caused, and the face light-weight network model based on the embedded space distillation adopts an approximate update center vector +.>Is a strategy of (1):
wherein,indicate->After iteration, the face recognition convolution network>Corresponding->A center vector of the class; />The training dataset class labels are +.>Is the number of samples of (a); />Indicate->The number of samples corresponds to the lightweight face recognition convolutional network +.>Is provided.
3. The method for constructing a lightweight network model of a face according to claim 1, wherein the method for determining the lightweight network model of a face based on embedded spatial distillation in step C5 is as follows:
acquiring a specified number of face digital images marked with face identity information labels, and constructing a positive sample pair and a negative sample pair to be tested according to the face identity information labels to serve as a test data set; extracting features from the face digital images in all the test data sets by using the face lightweight network model based on embedded space distillation in the step C5, and calculating cosine distances of each sample pair in the test data sets by using the extracted features; and evaluating the face light-weight network model based on embedded space distillation by using a receiver operation characteristic curve (Receiver Operating Characteristic curve) according to the cosine distance of the positive and negative sample pairs.
4. A face recognition method based on a face lightweight network model construction method according to any one of claims 1-3, comprising:
step 1, acquiring a face digital image to be recognized;
and 2, inputting the face digital image into the face light-weight network model based on the embedded space distillation, extracting face features through the face light-weight network model based on the embedded space distillation, and carrying out matching recognition with pre-stored face digital image features.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110947540.0A CN113723238B (en) | 2021-08-18 | 2021-08-18 | Face lightweight network model construction method and face recognition method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110947540.0A CN113723238B (en) | 2021-08-18 | 2021-08-18 | Face lightweight network model construction method and face recognition method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113723238A CN113723238A (en) | 2021-11-30 |
| CN113723238B true CN113723238B (en) | 2024-02-09 |
Family
ID=78676221
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110947540.0A Active CN113723238B (en) | 2021-08-18 | 2021-08-18 | Face lightweight network model construction method and face recognition method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113723238B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116453201B (en) * | 2023-06-19 | 2023-09-01 | 南昌大学 | Face recognition method and system based on adjacent edge loss |
| CN116959064B (en) * | 2023-06-25 | 2024-04-26 | 上海腾桥信息技术有限公司 | Certificate verification method and device, computer equipment and storage medium |
| CN117275076B (en) * | 2023-11-16 | 2024-02-27 | 厦门瑞为信息技术有限公司 | Method for constructing face quality assessment model based on characteristics and application |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108664893A (en) * | 2018-04-03 | 2018-10-16 | 福州海景科技开发有限公司 | A kind of method for detecting human face and storage medium |
| CN109190521A (en) * | 2018-08-17 | 2019-01-11 | 北京亮亮视野科技有限公司 | A kind of construction method of the human face recognition model of knowledge based purification and application |
| CN110674688A (en) * | 2019-08-19 | 2020-01-10 | 深圳力维智联技术有限公司 | Face recognition model acquisition method, system and medium for video monitoring scene |
| CN111259738A (en) * | 2020-01-08 | 2020-06-09 | 科大讯飞股份有限公司 | Face recognition model construction method, face recognition method and related device |
| CN112183492A (en) * | 2020-11-05 | 2021-01-05 | 厦门市美亚柏科信息股份有限公司 | Face model precision correction method, device and storage medium |
| WO2021023202A1 (en) * | 2019-08-07 | 2021-02-11 | 交叉信息核心技术研究院(西安)有限公司 | Self-distillation training method and device for convolutional neural network, and scalable dynamic prediction method |
| CN112766422A (en) * | 2021-03-15 | 2021-05-07 | 山东大学 | Privacy protection method based on lightweight face recognition model |
-
2021
- 2021-08-18 CN CN202110947540.0A patent/CN113723238B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108664893A (en) * | 2018-04-03 | 2018-10-16 | 福州海景科技开发有限公司 | A kind of method for detecting human face and storage medium |
| CN109190521A (en) * | 2018-08-17 | 2019-01-11 | 北京亮亮视野科技有限公司 | A kind of construction method of the human face recognition model of knowledge based purification and application |
| WO2021023202A1 (en) * | 2019-08-07 | 2021-02-11 | 交叉信息核心技术研究院(西安)有限公司 | Self-distillation training method and device for convolutional neural network, and scalable dynamic prediction method |
| CN110674688A (en) * | 2019-08-19 | 2020-01-10 | 深圳力维智联技术有限公司 | Face recognition model acquisition method, system and medium for video monitoring scene |
| CN111259738A (en) * | 2020-01-08 | 2020-06-09 | 科大讯飞股份有限公司 | Face recognition model construction method, face recognition method and related device |
| CN112183492A (en) * | 2020-11-05 | 2021-01-05 | 厦门市美亚柏科信息股份有限公司 | Face model precision correction method, device and storage medium |
| CN112766422A (en) * | 2021-03-15 | 2021-05-07 | 山东大学 | Privacy protection method based on lightweight face recognition model |
Non-Patent Citations (2)
| Title |
|---|
| Francesco Guzzi ; Luca De Bortoli等.Distillation of a CNN for a high accuracy mobile face recognition system.《2019 42nd International Convention on Information and Communication Technology, Electronics and Microelectronics (MIPRO)》.2019,989-994. * |
| 基于深度特征蒸馏的人脸识别;葛仕明;赵胜伟;刘文瑜;李晨钰;;北京交通大学学报(06);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113723238A (en) | 2021-11-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN113723238B (en) | Face lightweight network model construction method and face recognition method | |
| CN111489358B (en) | Three-dimensional point cloud semantic segmentation method based on deep learning | |
| CN111695415B (en) | Image recognition method and related equipment | |
| CN109344731B (en) | Lightweight face recognition method based on neural network | |
| WO2021022521A1 (en) | Method for processing data, and method and device for training neural network model | |
| Yu et al. | Auto-fas: Searching lightweight networks for face anti-spoofing | |
| CN112464865A (en) | Facial expression recognition method based on pixel and geometric mixed features | |
| CN110222718B (en) | Image processing method and device | |
| CN109190521B (en) | Construction method and application of face recognition model based on knowledge purification | |
| CN112801054B (en) | Face recognition model processing method, face recognition method and device | |
| CN113628059A (en) | Associated user identification method and device based on multilayer graph attention network | |
| CN113593661A (en) | Clinical term standardization method, device, electronic equipment and storage medium | |
| CN112084895B (en) | Pedestrian re-identification method based on deep learning | |
| CN113139499A (en) | Gait recognition method and system based on light-weight attention convolutional neural network | |
| CN116226785A (en) | Target object recognition method, multi-mode recognition model training method and device | |
| CN116248392A (en) | Network malicious traffic detection system and method based on multi-head attention mechanism | |
| CN112988970A (en) | Text matching algorithm serving intelligent question-answering system | |
| Wang et al. | A novel multiface recognition method with short training time and lightweight based on ABASNet and H-softmax | |
| WO2022063076A1 (en) | Adversarial example identification method and apparatus | |
| Barbhuiya et al. | Gesture recognition from RGB images using convolutional neural network‐attention based system | |
| CN116797895A (en) | Judicial scene-oriented multi-mode fusion identity authentication method, medium and device | |
| CN114429646A (en) | Gait recognition method based on deep self-attention transformation network | |
| CN114495163A (en) | Pedestrian re-identification generation learning method based on category activation mapping | |
| CN108427967B (en) | Real-time image clustering method | |
| Dhamija et al. | An approach to enhance performance of age invariant face recognition |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20230926 Address after: B1f-112, Zone C, Huaxun building, software park, torch hi tech Zone, Xiamen City, Fujian Province, 361000 Applicant after: XIAMEN RUIWEI INFORMATION TECHNOLOGY CO.,LTD. Applicant after: Beijing Shenshen Technology Co.,Ltd. Address before: 100000 1708, floor 17, No. 1, Zhichun Road, Haidian District, Beijing Applicant before: Beijing Shenshen Technology Co.,Ltd. |
|
| TA01 | Transfer of patent application right | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |