CN108664893A - A kind of method for detecting human face and storage medium - Google Patents
A kind of method for detecting human face and storage medium Download PDFInfo
- Publication number
- CN108664893A CN108664893A CN201810290187.1A CN201810290187A CN108664893A CN 108664893 A CN108664893 A CN 108664893A CN 201810290187 A CN201810290187 A CN 201810290187A CN 108664893 A CN108664893 A CN 108664893A
- Authority
- CN
- China
- Prior art keywords
- network
- loss function
- light weight
- complex network
- face datection
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/161—Detection; Localisation; Normalisation
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/172—Classification, e.g. identification
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Oral & Maxillofacial Surgery (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
A kind of method for detecting human face and storage medium, this approach includes the following steps：Step 102, the same training image of a batch is inputted respectively to light weight network and complex network；Step 104, the output for light weight network and the classification chart of complex network using difficult sample mining method as a result, be filtered；Step 106, comprehensive loss function is constructed, the comprehensive loss function includes knowledge distillation loss function or the Face datection loss function based on label, and the knowledge distillation loss function is obtained according to the output result of light weight network and the classification chart of complex network；Step 108, it is based on loss function, the parameter of light weight network is updated, does not update the parameter of complex network；Step 110, it repeats the above steps, until light weight network training is extremely restrained.A kind of neural network algorithm of fast velocity modulation ginseng is provided, solves the problems, such as that lightweight neural network carries out Face datection.
Description
Technical field
The invention belongs to image processing and pattern recognition fields, and in particular to a kind of method for detecting human face is arrived, it can
To be applied to security monitoring, the numerous areas such as human-computer interaction.
Background technology
Face datection is an important technology, it has a demand in the application of many computer visions, for example, face with
Track, face alignment, recognition of face etc..In recent years, due to the development of convolutional neural networks, the performance of Face datection has obtained obviously
Raising.However, the existing usual calculating speed of Face datection model is slow, because they need bigger neural network
To keep preferable Face datection performance.Although also there is the detection framework based on one-step method to be proposed for the speed of acceleration detection
(such as SSD, YOLO), but they are still not fast enough for practical application scene, especially in the environment based on CPU.Another party
Face, if reaching rate request by reducing the parameter of convolutional network, the performance of detector if, can be decreased obviously.Therefore, it
Obtain the task that a lightweight human-face detector of good performance is a great challenge.
Knowledge distillation (knowledge distillation) be it is a kind of can allow the small big network of network learning by imitation, to carry
Rise the technology of small network performance.The validity of knowledge distillation is verified in classification and metric learning task.For
Detection task has no idea directly to distill (knowledge distillation) technology using original knowledge, because of detection
There are the unbalanced problem of classification (background classes will be far more than other classes) for the output of device, if only imitating as classification task
All output is practised, then is unable to get a good performance.The detector of most of lightweight be based on one-step method, rather than
Two step method, because the former has speed advantage.Compared with two step method, one-step method has lacked the region nomination for eliminating negative sample
Network, thus the unbalanced problem of classification is more serious.
Invention content
For this reason, it may be necessary to which providing a kind of novel can be applicable in one-step method to promote lightweight detection model performance, fast velocity modulation
The neural network algorithm of ginseng solves the problems, such as that lightweight neural network carries out Face datection.In the present invention, it inventor provides
A kind of method for detecting human face, this approach includes the following steps：
Step 102, the same training image of a batch is inputted respectively to light weight network and complex network；
Step 104, the output for light weight network and the classification chart of complex network using difficult sample mining method as a result, carried out
Filtering；
Step 106, comprehensive loss function is constructed, the comprehensive loss function includes knowledge distillation loss function or is based on marking
The Face datection loss function of label, the knowledge distillation loss function is according to the output of light weight network and the classification chart of complex network
As a result it obtains；
Step 108, it is based on loss function, the parameter of light weight network is updated, does not update the parameter of complex network；
Step 110, it repeats the above steps, until light weight network training is extremely restrained.
Preferably, the difficult sample mining method is filtered specially：
A threshold value T is set, for judging whether some probability in classification chart has enough confidence levels；T is one
Hyper parameter, value range are 0 to 1, each index in classification chart are traversed, when probability of the index in light weight network is more than
T and when the probability of complex network is less than T, which is added set Sm；Alternatively, working as probability of the index in light weight network
Less than T when the probability of complex network is more than T, S also is added in the indexm。
Optionally, the knowledge distillation loss function is：
Wherein, p(i)For i-th of probability score in the classification chart of complex network, q(i)It is then the of light weight network class figure
I probability score.
Further,
The Face datection loss function based on label is：
LG=Lcls+Lreg
Wherein, LclsIt is the two class Softmax loss functions for classification, LregIt is the robust regression loss letter for recurrence
Number；
The comprehensive loss function is that knowledge distillation loss function is weighted with the Face datection loss function based on label：
L=LG+cLKD
C is coefficient of balance.
Specifically, further include step, structure light weight network, complex network；
The Face datection model based on convolutional neural networks is built, as complex network, its extremely convergence of training；
The Face datection model with the convolutional neural networks of frame with complex network is built, it is described light as light weight network
The quantity for measuring every layer of filter in the frame of network is respectively less than complex network.
A kind of Face datection storage medium of knowledge based distillation, is stored with computer program, the computer program exists
Following steps are executed when being run：
Step 102, the same training image of a batch is inputted respectively to light weight network and complex network；
Step 104, the output for light weight network and the classification chart of complex network using difficult sample mining method as a result, carried out
Filtering；
Step 106, comprehensive loss function is constructed, the comprehensive loss function includes knowledge distillation loss function or is based on marking
The Face datection loss function of label, the knowledge distillation loss function is according to the output of light weight network and the classification chart of complex network
As a result it obtains；
Step 108, it is based on loss function, the parameter of light weight network is updated, does not update the parameter of complex network；
Step 110, it repeats the above steps, until light weight network training is extremely restrained.
Specifically, the difficult sample mining method is filtered specially：
A threshold value T is set, for judging whether some probability in classification chart has enough confidence levels；T is one
Hyper parameter, value range are 0 to 1, each index in classification chart are traversed, when probability of the index in light weight network is more than
T and when the probability of complex network is less than T, which is added set Sm；Alternatively, working as probability of the index in light weight network
Less than T when the probability of complex network is more than T, S also is added in the indexm。
Optionally, the knowledge distillation loss function is：
Wherein, p(i)For i-th of probability score in the classification chart of complex network, q(i)It is then the of light weight network class figure
I probability score.
Further,
The Face datection loss function based on label is：
LG=Lcls+Lreg
Wherein, LclsIt is the two class Softmax loss functions for classification, LregIt is the robust regression loss letter for recurrence
Number；
The comprehensive loss function is that knowledge distillation loss function is weighted with the Face datection loss function based on label：
L=LG+cLKD
C is coefficient of balance.
Specifically, computer program also carries out step, structure light weight network, complex network when being run；
The Face datection model based on convolutional neural networks is built, as complex network, its extremely convergence of training；
The Face datection model with the convolutional neural networks of frame with complex network is built, it is described light as light weight network
The quantity for measuring every layer of filter in the frame of network is respectively less than complex network.
It is different from the prior art, above-mentioned technology uses standardized calculation means, in entire evaluation procedure, introduces original
The index not having in evaluation system, and unified parameters so that quantitative criteria is relatively uniform, and therefore, the present invention solves network carriage
Feelings dynamically real-time problem analysis.
Description of the drawings
Fig. 1 is the method for detecting human face flow chart described in specific implementation mode.
Specific implementation mode
For the technology contents of technical solution, construction feature, the objects and the effects are described in detail, below in conjunction with specific reality
It applies example and attached drawing is coordinated to be explained in detail.
In the embodiment shown in fig. 1, it may be seen that a kind of method for detecting human face, this approach includes the following steps：
Step 100, structure the Face datection model based on convolutional neural networks be used as teacher's network, training the model up to
Convergence.
The frame of teacher's network is usually as student network, if but the quantity of every layer of filter is student network
Dry times, therefore performance can be more preferable.We are in order that allow the complexity of conventional convolutional neural networks to simplify, therefore herein
Content in, teacher's network can be replaced mutually with complex network, and student network equally can also be replaced with light weight network, light weight
The characteristics of network, is that it is the Face datection model with the convolutional neural networks of frame with complex network, the light weight network
The quantity of every layer of filter is respectively less than complex network in frame.The training method of teacher's network and conventional detection model complete one
Sample, by taking the present invention as an example, the loss function of teacher's network is:LG=Lcls+Lreg.Wherein, LclsIt is two classes for classification
Softmax loss functions, LregIt is robust regression loss function (the smooth L for recurrence1).Build light weight network, complexity
Network；
Student network is exactly the detection model finally to be obtained, and is carried out now to the parameter of student network with Xavier methods
Random initializtion.
And in some other specific embodiment, the present invention can carry out above-mentioned preparation process in advance and directly start
In step 102, the same training image of a batch is inputted respectively to light weight network and complex network；
Training image can be not processed, and can also carry out data augmentation technology herein, specific as follows：
For every training image of input, using data augmentation technology, to increase the Generalization Capability of model.With this hair
For bright, data augmentation includes the following steps：
(1) colour dither operates：Respectively with 0.5 probability, the randomly brightness of adjusting training image, contrast, saturation
The parameters such as degree.
(2) random cropping operates：On this training image, 5 square subgraphs are randomly cut out.Wherein 1
It is its maximum square subgraph, in addition the length of side of 4 square subgraphs is 0.3~1.0 times of training image short side.
In this 5 square subgraphs, 1 is randomly chosen as final training sample.
(3) flip horizontal operates：For the training sample that this chooses, flip horizontal is randomly carried out with 0.5 probability
Operation.
(4) change of scale operates：The training sample obtained by aforesaid operations, 1024 × 1024 sizes are zoomed to, are sent
Enter network for training.
Step 104, the output for light weight network and the classification chart (classification map) of complex network as a result,
It is filtered using difficult sample mining method；To solve the problems, such as the unbalanced problem low with Fitting efficiency of classification.
Knowledge distillation (knowledge distillation) method be desirable to student network by imitate teacher's network, from
And the result as teacher's network is exported as much as possible.In neural network, the information of layer more rearward and final prediction knot
Fruit association is closer, can provide better supervision message for learning by imitation.Therefore, last layer, which compares, is adapted to allow student
E-learning is imitated.In the Face datection frame based on single -step method, there are two modules for last layer, are classification chart respectively
(classification map) and regression figure (regression map).Why effective knowledge distillation is, is that it provides old
The soft label information that teacher's e-learning arrives is to student network, and for these soft labels compared with original mark label, information is more accurate
It is really and smooth, thus more conducively e-learning.In Face datection, the mark label for returning frame is exactly originally real number, has been compared
It is more accurate；And the mark label of classification task only has 0 and 1, is not very accurate.Therefore, classification chart is particularly suited for knowledge
Distill (knowledge distillation) study.
One typically the classification chart based on single -step method (classification map) output size be 2N × H × W,
Wherein N is the number of anchor point frame, and 2 indicate that each anchor point frame is required for predicting that the probability of positive class and negative class, H are the height of classification chart, W
For its width.Because the probability of positive class and negative class is by standardization, addition is always 1, can only be closed when carrying out knowledge distillation
The probability of positive class is noted, therefore the output of classification chart can be reduced to N × H × W.In the training process, teacher's network and student
Network can export respectively classification chart as a result, for the two as a result, it is desirable to determine classification chart (classification
Map which index) should be filtered, which can be used for knowledge distillation (knowledge distillation)
's.
Step 106 is then also carried out, constructs comprehensive loss function, the comprehensive loss function includes knowledge distillation loss letter
Number or the Face datection loss function based on label, the knowledge distillation loss function is according to point of light weight network and complex network
The output result of class figure obtains.In some optional embodiments, (knowledge is distilled by constructing knowledge
Distillation) loss function so that result is as possible for the classification chart (classification map) of current student network
It is drawn close to the classification chart of teacher's network, as a specific embodiment, the knowledge distillation loss function is：
Wherein, p(i)For i-th of probability score in the classification chart of complex network, q(i)It is then the of light weight network class figure
I probability score.
Further, in the training process, in addition to knowledge distills (knowledge distillation) loss function, also
There are the conventional Face datection loss function based on label, the area in the loss function and classics detection framework Faster RCNN
Domain referral networks are consistent：
The Face datection loss function based on label is：
LG=Lcls+Lreg
Wherein, LclsIt is the two class Softmax loss functions for classification, LregIt is the robust regression loss letter for recurrence
Number；When training, knowledge based distills the loss function of (knowledge distillation) and the loss function based on label
It is added, forms final comprehensive loss function.
The comprehensive loss function be knowledge distillation loss function with：
L=LG+cLKD
C is the coefficient for balancing two loss functions, is fixed as 50 in the present invention, optimal value should be by concrete scene
It determines.
As shown in Figure 1, then also carrying out step 108, it is based on loss function, updates the parameter of light weight network, not more the old and new
The parameter of teacher's network.In this step, according to obtained comprehensive loss function, using back-propagation algorithm, to student
The parameter of network is updated, to complete primary training.The parameter of teacher's network need not update, therefore be needed in training
Freezed.
Step 110, repeat the above steps 102-108, until light weight network training is extremely restrained.
The present invention can effectively promote the precision of lightweight human-face detector by the design of application above-mentioned steps so that
Face datection can also obtain satisfactory detection result in the equipment that computing resource is limited.Due to detection model and classification with
And there are the differences in network structure for metric learning model, so can not (knowledge directly be distilled knowledge
Distillation) method is directly used in Detection task.The inventors found that the Face datection model based on single -step method
In regression figure and without enough effective informations for learning, and classification chart (classification map) can then carry
It is used as student network and teacher's network migration for effective soft label information, therefore by classification chart (classification map)
The medium of knowledge.In addition, the output result of classification chart (classification map) has largely negative class sample, lead to class
Not unbalanced problem.The present invention proposes that a kind of difficult sample method for digging is used to filter simple negative sample so that classification reaches equal
Weighing apparatus, while also having filtered simple positive sample so that knowledge distills the more efficient of (knowledge distillation).Instruction
When practicing, knowledge based distills the loss function of (knowledge distillation) and based on the loss function of label with appropriate
Ratio be added, constitute complete loss function.In specific implement, the method for the present invention also carries out following steps, test
Image inputs trained student network model, exports testing result frame.Since the quantity of the detection block of output is very more, need
They are screened with merging.In the present embodiment, most detection is screened out by confidence threshold value T=0.05 first
Frame selects preceding N then according to confidence levela=400 detection blocks.Then the detection block for using non-maxima suppression removal to repeat,
And preceding N is selected according to confidence levelb=200 detection blocks are to get to final testing result.
Finally, the training method of knowledge based distillation (knowledge distillation) proposed by the present invention can be effective
Improve the detectability of lightweight Face datection model in ground.
A kind of Face datection storage medium of knowledge based distillation, is stored with computer program, the computer program exists
Following steps are executed when being run：
Step 102, the same training image of a batch is inputted respectively to light weight network and complex network；
Step 104, the output for light weight network and the classification chart of complex network using difficult sample mining method as a result, carried out
Filtering；
Step 106, comprehensive loss function is constructed, the comprehensive loss function includes knowledge distillation loss function or is based on marking
The Face datection loss function of label, the knowledge distillation loss function is according to the output of light weight network and the classification chart of complex network
As a result it obtains；
Step 108, it is based on loss function, the parameter of light weight network is updated, does not update the parameter of complex network；
Step 110, it repeats the above steps, until light weight network training is extremely restrained.
Specifically, the difficult sample mining method is filtered specially：
A threshold value T is set, for judging whether some probability in classification chart has enough confidence levels；T is one
Hyper parameter, value range are 0 to 1, each index in classification chart are traversed, when probability of the index in light weight network is more than
T and when the probability of complex network is less than T, which is added set Sm；Alternatively, working as probability of the index in light weight network
Less than T when the probability of complex network is more than T, S also is added in the indexm。
Optionally, the knowledge distillation loss function is：
Wherein, p(i)For i-th of probability score in the classification chart of complex network, q(i)It is then the of light weight network class figure
I probability score.
Further,
The Face datection loss function based on label is：
LG=Lcls+Lreg
Wherein, LclsIt is the two class Softmax loss functions for classification, LregIt is the robust regression loss letter for recurrence
Number；
The comprehensive loss function be knowledge distillation loss function with：
L=LG+cLKD
C is coefficient of balance.
Specifically, computer program also carries out step, structure light weight network, complex network when being run；
The Face datection model based on convolutional neural networks is built, as complex network, its extremely convergence of training；
The Face datection model with the convolutional neural networks of frame with Complex Neural Network is built, as light weight network, institute
The quantity for stating every layer of filter in the frame of light weight network is respectively less than complex network.
It should be noted that herein, relational terms such as first and second and the like are used merely to a reality
Body or operation are distinguished with another entity or operation, are deposited without necessarily requiring or implying between these entities or operation
In any actual relationship or order or sequence.Moreover, the terms "include", "comprise" or its any other variant are intended to
Non-exclusive inclusion, so that process, method, article or terminal device including a series of elements include not only those
Element, but also include other elements that are not explicitly listed, or further include for this process, method, article or end
The intrinsic element of end equipment.In the absence of more restrictions, being limited by sentence " including ... " or " including ... "
Element, it is not excluded that there is also other elements in process, method, article or the terminal device including the element.This
Outside, herein, " being more than ", " being less than ", " being more than " etc. are interpreted as not including this number；" more than ", " following ", " within " etc. understandings
It includes this number to be.
It should be understood by those skilled in the art that, the various embodiments described above can be provided as method, apparatus or computer program production
Product.Complete hardware embodiment, complete software embodiment or embodiment combining software and hardware aspects can be used in these embodiments
Form.All or part of step in the method that the various embodiments described above are related to can be instructed by program relevant hardware come
It completes, the program can be stored in the storage medium that computer equipment can be read, for executing the various embodiments described above side
All or part of step described in method.The computer equipment, including but not limited to：Personal computer, server, general-purpose computations
It is machine, special purpose computer, the network equipment, embedded device, programmable device, intelligent mobile terminal, smart home device, wearable
Smart machine, vehicle intelligent equipment etc.；The storage medium, including but not limited to：RAM, ROM, magnetic disc, tape, CD, sudden strain of a muscle
It deposits, USB flash disk, mobile hard disk, storage card, memory stick, webserver storage, network cloud storage etc..
The various embodiments described above are with reference to method, equipment (system) and the computer program product according to embodiment
Flowchart and/or the block diagram describes.It should be understood that can be realized by computer program instructions every in flowchart and/or the block diagram
The combination of flow and/or box in one flow and/or box and flowchart and/or the block diagram.These computers can be provided
Program instruction is to the processor of computer equipment to generate a machine so that the finger executed by the processor of computer equipment
It enables and generates to specify in one flow of flow chart or multiple flows and/or one box of block diagram or multiple boxes
The device of function.
These computer program instructions, which may also be stored in, can guide computer equipment computer operate in a specific manner to set
In standby readable memory so that the instruction generation being stored in the computer equipment readable memory includes the manufacture of command device
Product, command device realization refer in one flow of flow chart or multiple flows and/or one box of block diagram or multiple boxes
Fixed function.
These computer program instructions can be also loaded on computer equipment so that be executed on a computing device a series of
To generate computer implemented processing, the instruction to execute on a computing device is provided for realizing in flow operating procedure
The step of function of being specified in one flow of figure or multiple flows and/or one box of block diagram or multiple boxes.
Although the various embodiments described above are described, once a person skilled in the art knows basic wounds
The property made concept, then additional changes and modifications can be made to these embodiments, so example the above is only the implementation of the present invention,
It is not intended to limit the scope of patent protection of the present invention, it is every to utilize equivalent structure made by description of the invention and accompanying drawing content
Or equivalent process transformation, it is applied directly or indirectly in other relevant technical fields, the patent for being similarly included in the present invention
Within protection domain.
Claims (10)
1. a kind of method for detecting human face, which is characterized in that this approach includes the following steps：
Step 102, the same training image of a batch is inputted respectively to light weight network and complex network；
Step 104, the output for light weight network and the classification chart of complex network using difficult sample mining method as a result, carried out
Filter；
Step 106, comprehensive loss function is constructed, the comprehensive loss function includes knowledge distillation loss function or based on label
Face datection loss function, the knowledge distillation loss function is according to the output result of light weight network and the classification chart of complex network
It obtains；
Step 108, it is based on loss function, the parameter of light weight network is updated, does not update the parameter of complex network；
Step 110, it repeats the above steps, until light weight network training is extremely restrained.
2. method for detecting human face according to claim 1, which is characterized in that the hardly possible sample mining method is filtered specifically
For：
A threshold value T is set, for judging whether some probability in classification chart has enough confidence levels；T is a super ginseng
Number, value range is 0 to 1, traverse classification chart in each index, when probability of the index in light weight network more than T and
When the probability of complex network is less than T, which is added set Sm；Alternatively, when probability of the index in light weight network is small
In T when the probability of complex network is more than T, S also is added in the indexm。
3. method for detecting human face according to claim 2, which is characterized in that
The knowledge distillation loss function is：
Wherein, p(i)For i-th of probability score in the classification chart of complex network, q(i)It is then i-th of light weight network class figure
Probability score.
4. method for detecting human face according to claim 2 or 3, which is characterized in that
The Face datection loss function based on label is：
LG=Lcls+Lreg
Wherein, LclsIt is the two class Softmax loss functions for classification, LregIt is the robust regression loss function for recurrence；
The comprehensive loss function is that knowledge distillation loss function is weighted with the Face datection loss function based on label：
L=LG+cLKD
C is coefficient of balance.
5. method for detecting human face according to claim 1, which is characterized in that further include step, structure light weight network, complexity
Network；
The Face datection model based on convolutional neural networks is built, as complex network, its extremely convergence of training；
Structure is with complex network with the Face datection model of the convolutional neural networks of frame, as light weight network, the light weight net
The quantity of every layer of filter is respectively less than complex network in the frame of network.
6. a kind of Face datection storage medium, which is characterized in that be stored with computer program, the computer program is being run
Shi Zhihang following steps：
Step 102, the same training image of a batch is inputted respectively to light weight network and complex network；
Step 104, the output for light weight network and the classification chart of complex network using difficult sample mining method as a result, carried out
Filter；
Step 106, comprehensive loss function is constructed, the comprehensive loss function includes knowledge distillation loss function or based on label
Face datection loss function, the knowledge distillation loss function is according to the output result of light weight network and the classification chart of complex network
It obtains；
Step 108, it is based on loss function, the parameter of light weight network is updated, does not update the parameter of complex network；
Step 110, it repeats the above steps, until light weight network training is extremely restrained.
7. Face datection storage medium according to claim 6, which is characterized in that the hardly possible sample mining method is filtered
Specially：
A threshold value T is set, for judging whether some probability in classification chart has enough confidence levels；T is a super ginseng
Number, value range is 0 to 1, traverse classification chart in each index, when probability of the index in light weight network more than T and
When the probability of complex network is less than T, which is added set Sm；Alternatively, when probability of the index in light weight network is small
In T when the probability of complex network is more than T, S also is added in the indexm。
8. Face datection storage medium according to claim 7, which is characterized in that
The knowledge distillation loss function is：
Wherein, p(i)For i-th of probability score in the classification chart of complex network, q(i)It is then i-th of light weight network class figure
Probability score.
9. Face datection storage medium according to claim 7 or 8, which is characterized in that
The Face datection loss function based on label is：
LG=Lcls+Lreg
Wherein, LclsIt is the two class Softmax loss functions for classification, LregIt is the robust regression loss function for recurrence；
The comprehensive loss function is that knowledge distillation loss function is weighted with the Face datection loss function based on label：
L=LG+cLKD
C is coefficient of balance.
10. Face datection storage medium according to claim 6, which is characterized in that computer program is gone back when being run
Carry out step, structure light weight network, complex network；
The Face datection model based on convolutional neural networks is built, as complex network, its extremely convergence of training；
Structure is with complex network with the Face datection model of the convolutional neural networks of frame, as light weight network, the light weight net
The quantity of every layer of filter is respectively less than complex network in the frame of network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810290187.1A CN108664893B (en) | 2018-04-03 | 2018-04-03 | Face detection method and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810290187.1A CN108664893B (en) | 2018-04-03 | 2018-04-03 | Face detection method and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108664893A true CN108664893A (en) | 2018-10-16 |
| CN108664893B CN108664893B (en) | 2022-04-29 |
Family
ID=63782947
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810290187.1A Active CN108664893B (en) | 2018-04-03 | 2018-04-03 | Face detection method and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108664893B (en) |
Cited By (29)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109300170A (en) * | 2018-10-18 | 2019-02-01 | 云南大学 | Portrait photo shadow transmission method |
| CN109472360A (en) * | 2018-10-30 | 2019-03-15 | 北京地平线机器人技术研发有限公司 | Update method, update device and electronic device of neural network |
| CN110059747A (en) * | 2019-04-18 | 2019-07-26 | 清华大学深圳研究生院 | A kind of net flow assorted method |
| CN110263731A (en) * | 2019-06-24 | 2019-09-20 | 电子科技大学 | A kind of single step face detection system |
| CN110598603A (en) * | 2019-09-02 | 2019-12-20 | 深圳力维智联技术有限公司 | Face recognition model acquisition method, device, equipment and medium |
| CN110674714A (en) * | 2019-09-13 | 2020-01-10 | 东南大学 | Joint detection method of face and face key points based on transfer learning |
| CN110956255A (en) * | 2019-11-26 | 2020-04-03 | 中国医学科学院肿瘤医院 | Difficult sample mining method, device, electronic device, and computer-readable storage medium |
| CN111027551A (en) * | 2019-12-17 | 2020-04-17 | 腾讯科技（深圳）有限公司 | Image processing method, apparatus and medium |
| CN111178370A (en) * | 2019-12-16 | 2020-05-19 | 深圳市华尊科技股份有限公司 | Vehicle retrieval method and related device |
| CN111368634A (en) * | 2020-02-05 | 2020-07-03 | 中国人民解放军国防科技大学 | Human head detection method, system and storage medium based on neural network |
| WO2020143225A1 (en) * | 2019-01-08 | 2020-07-16 | 南京人工智能高等研究院有限公司 | Neural network training method and apparatus, and electronic device |
| CN111507135A (en) * | 2019-01-31 | 2020-08-07 | 北京奇虎科技有限公司 | Face detection method and device, computer equipment and storage medium |
| CN111553227A (en) * | 2020-04-21 | 2020-08-18 | 东南大学 | Lightweight face detection method based on task guidance |
| CN111639744A (en) * | 2020-04-15 | 2020-09-08 | 北京迈格威科技有限公司 | Student model training method and device and electronic equipment |
| CN112232397A (en) * | 2020-09-30 | 2021-01-15 | 上海眼控科技股份有限公司 | Knowledge distillation method and device of image classification model and computer equipment |
| CN112270379A (en) * | 2020-11-13 | 2021-01-26 | 北京百度网讯科技有限公司 | Training method of classification model, sample classification method, apparatus and equipment |
| CN112348167A (en) * | 2020-10-20 | 2021-02-09 | 华东交通大学 | Knowledge distillation-based ore sorting method and computer-readable storage medium |
| CN112633406A (en) * | 2020-12-31 | 2021-04-09 | 天津大学 | Knowledge distillation-based few-sample target detection method |
| CN113408589A (en) * | 2021-05-26 | 2021-09-17 | 北京迈格威科技有限公司 | Method for training image recognition model, image recognition method and image recognition device |
| CN113614748A (en) * | 2019-03-21 | 2021-11-05 | 国际商业机器公司 | Systems and methods for incremental learning for object detection |
| CN113657411A (en) * | 2021-08-23 | 2021-11-16 | 北京达佳互联信息技术有限公司 | Neural network model training method, image feature extraction method and related device |
| CN113723238A (en) * | 2021-08-18 | 2021-11-30 | 北京深感科技有限公司 | Human face lightweight network model construction method and human face recognition method |
| CN113850103A (en) * | 2020-06-28 | 2021-12-28 | 京东科技控股股份有限公司 | Information processing method, apparatus, equipment and storage medium |
| CN114078268A (en) * | 2020-08-17 | 2022-02-22 | 珠海全志科技股份有限公司 | A training method and device for a lightweight face recognition model |
| CN114444558A (en) * | 2020-11-05 | 2022-05-06 | 佳能株式会社 | Training method and training device of neural network for object recognition |
| CN114663942A (en) * | 2022-03-17 | 2022-06-24 | 深圳数联天下智能科技有限公司 | Feature detection method, model training method, device, and medium |
| CN115661560A (en) * | 2022-06-30 | 2023-01-31 | 斑马网络技术有限公司 | Face detection method in cockpit, target detection model training method and device |
| CN117542085A (en) * | 2024-01-10 | 2024-02-09 | 湖南工商大学 | Park scene pedestrian detection method, device and equipment based on knowledge distillation |
| CN118799844A (en) * | 2024-09-13 | 2024-10-18 | 南京智莲森信息技术有限公司 | A railway pole number plate and kilometer mark identification method and related device |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20080126311A1 (en) * | 2006-11-29 | 2008-05-29 | Red Hat, Inc. | Automatic index creation based on unindexed search evaluation |
| EP2584381A1 (en) * | 2011-10-21 | 2013-04-24 | ENI S.p.A. | Method for predicting the properties of crude oils by the application of neural networks |
| US20130245844A1 (en) * | 2012-03-19 | 2013-09-19 | Saudi Arabian Oil Company | Methods For Simultaneous Process and Utility Systems Synthesis in Partially and Fully Decentralized Environments |
| CN107220618A (en) * | 2017-05-25 | 2017-09-29 | 中国科学院自动化研究所 | Method for detecting human face and device, computer-readable recording medium, equipment |
| CN107239736A (en) * | 2017-04-28 | 2017-10-10 | 北京智慧眼科技股份有限公司 | Method for detecting human face and detection means based on multitask concatenated convolutional neutral net |
| CN107247989A (en) * | 2017-06-15 | 2017-10-13 | 北京图森未来科技有限公司 | A kind of neural network training method and device |
| EP3276540A2 (en) * | 2016-07-28 | 2018-01-31 | Samsung Electronics Co., Ltd. | Neural network method and apparatus |
| CN107818314A (en) * | 2017-11-22 | 2018-03-20 | 北京达佳互联信息技术有限公司 | Face image processing method, device and server |
| CN110674880A (en) * | 2019-09-27 | 2020-01-10 | 北京迈格威科技有限公司 | Network training method, device, medium and electronic equipment for knowledge distillation |
-
2018
- 2018-04-03 CN CN201810290187.1A patent/CN108664893B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20080126311A1 (en) * | 2006-11-29 | 2008-05-29 | Red Hat, Inc. | Automatic index creation based on unindexed search evaluation |
| EP2584381A1 (en) * | 2011-10-21 | 2013-04-24 | ENI S.p.A. | Method for predicting the properties of crude oils by the application of neural networks |
| US20130245844A1 (en) * | 2012-03-19 | 2013-09-19 | Saudi Arabian Oil Company | Methods For Simultaneous Process and Utility Systems Synthesis in Partially and Fully Decentralized Environments |
| EP3276540A2 (en) * | 2016-07-28 | 2018-01-31 | Samsung Electronics Co., Ltd. | Neural network method and apparatus |
| CN107239736A (en) * | 2017-04-28 | 2017-10-10 | 北京智慧眼科技股份有限公司 | Method for detecting human face and detection means based on multitask concatenated convolutional neutral net |
| CN107220618A (en) * | 2017-05-25 | 2017-09-29 | 中国科学院自动化研究所 | Method for detecting human face and device, computer-readable recording medium, equipment |
| CN107247989A (en) * | 2017-06-15 | 2017-10-13 | 北京图森未来科技有限公司 | A kind of neural network training method and device |
| CN107818314A (en) * | 2017-11-22 | 2018-03-20 | 北京达佳互联信息技术有限公司 | Face image processing method, device and server |
| CN110674880A (en) * | 2019-09-27 | 2020-01-10 | 北京迈格威科技有限公司 | Network training method, device, medium and electronic equipment for knowledge distillation |
Non-Patent Citations (2)
| Title |
|---|
| HINTON G等: "Distilling the knowledge in a neural network", 《PROC. OF THE ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS》 * |
| 雷杰等: "深度网络模型压缩综述", 《软件学报》 * |
Cited By (42)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109300170A (en) * | 2018-10-18 | 2019-02-01 | 云南大学 | Portrait photo shadow transmission method |
| CN109472360A (en) * | 2018-10-30 | 2019-03-15 | 北京地平线机器人技术研发有限公司 | Update method, update device and electronic device of neural network |
| US11328180B2 (en) | 2018-10-30 | 2022-05-10 | Beijing Horizon Robotics Technology Research And Development Co., Ltd. | Method for updating neural network and electronic device |
| WO2020143225A1 (en) * | 2019-01-08 | 2020-07-16 | 南京人工智能高等研究院有限公司 | Neural network training method and apparatus, and electronic device |
| CN111507135A (en) * | 2019-01-31 | 2020-08-07 | 北京奇虎科技有限公司 | Face detection method and device, computer equipment and storage medium |
| CN113614748A (en) * | 2019-03-21 | 2021-11-05 | 国际商业机器公司 | Systems and methods for incremental learning for object detection |
| CN110059747A (en) * | 2019-04-18 | 2019-07-26 | 清华大学深圳研究生院 | A kind of net flow assorted method |
| CN110263731A (en) * | 2019-06-24 | 2019-09-20 | 电子科技大学 | A kind of single step face detection system |
| CN110598603A (en) * | 2019-09-02 | 2019-12-20 | 深圳力维智联技术有限公司 | Face recognition model acquisition method, device, equipment and medium |
| CN110674714A (en) * | 2019-09-13 | 2020-01-10 | 东南大学 | Joint detection method of face and face key points based on transfer learning |
| CN110674714B (en) * | 2019-09-13 | 2022-06-14 | 东南大学 | Joint detection method of face and face key points based on transfer learning |
| CN110956255A (en) * | 2019-11-26 | 2020-04-03 | 中国医学科学院肿瘤医院 | Difficult sample mining method, device, electronic device, and computer-readable storage medium |
| CN110956255B (en) * | 2019-11-26 | 2023-04-07 | 中国医学科学院肿瘤医院 | Difficult sample mining method and device, electronic equipment and computer readable storage medium |
| CN111178370A (en) * | 2019-12-16 | 2020-05-19 | 深圳市华尊科技股份有限公司 | Vehicle retrieval method and related device |
| CN111178370B (en) * | 2019-12-16 | 2023-10-17 | 深圳市华尊科技股份有限公司 | Vehicle searching method and related device |
| CN111027551B (en) * | 2019-12-17 | 2023-07-07 | 腾讯科技（深圳）有限公司 | Image processing method, apparatus and medium |
| CN111027551A (en) * | 2019-12-17 | 2020-04-17 | 腾讯科技（深圳）有限公司 | Image processing method, apparatus and medium |
| CN111368634A (en) * | 2020-02-05 | 2020-07-03 | 中国人民解放军国防科技大学 | Human head detection method, system and storage medium based on neural network |
| CN111368634B (en) * | 2020-02-05 | 2023-06-20 | 中国人民解放军国防科技大学 | Human head detection method, system and storage medium based on neural network |
| CN111639744B (en) * | 2020-04-15 | 2023-09-22 | 北京迈格威科技有限公司 | Training method and device for student model and electronic equipment |
| CN111639744A (en) * | 2020-04-15 | 2020-09-08 | 北京迈格威科技有限公司 | Student model training method and device and electronic equipment |
| CN111553227A (en) * | 2020-04-21 | 2020-08-18 | 东南大学 | Lightweight face detection method based on task guidance |
| CN113850103A (en) * | 2020-06-28 | 2021-12-28 | 京东科技控股股份有限公司 | Information processing method, apparatus, equipment and storage medium |
| CN114078268A (en) * | 2020-08-17 | 2022-02-22 | 珠海全志科技股份有限公司 | A training method and device for a lightweight face recognition model |
| CN112232397A (en) * | 2020-09-30 | 2021-01-15 | 上海眼控科技股份有限公司 | Knowledge distillation method and device of image classification model and computer equipment |
| CN112348167B (en) * | 2020-10-20 | 2022-10-11 | 华东交通大学 | A kind of ore sorting method and computer readable storage medium based on knowledge distillation |
| CN112348167A (en) * | 2020-10-20 | 2021-02-09 | 华东交通大学 | Knowledge distillation-based ore sorting method and computer-readable storage medium |
| CN114444558A (en) * | 2020-11-05 | 2022-05-06 | 佳能株式会社 | Training method and training device of neural network for object recognition |
| WO2022100045A1 (en) * | 2020-11-13 | 2022-05-19 | 北京百度网讯科技有限公司 | Training method for classification model, sample classification method and apparatus, and device |
| US12340288B2 (en) | 2020-11-13 | 2025-06-24 | Beijing Baidu Netcom Science Technology Co., Ltd. | Method of training classification model, method of classifying sample, and device |
| CN112270379A (en) * | 2020-11-13 | 2021-01-26 | 北京百度网讯科技有限公司 | Training method of classification model, sample classification method, apparatus and equipment |
| CN112270379B (en) * | 2020-11-13 | 2023-09-19 | 北京百度网讯科技有限公司 | Classification model training methods, sample classification methods, devices and equipment |
| CN112633406A (en) * | 2020-12-31 | 2021-04-09 | 天津大学 | Knowledge distillation-based few-sample target detection method |
| CN113408589A (en) * | 2021-05-26 | 2021-09-17 | 北京迈格威科技有限公司 | Method for training image recognition model, image recognition method and image recognition device |
| CN113723238B (en) * | 2021-08-18 | 2024-02-09 | 厦门瑞为信息技术有限公司 | A face lightweight network model construction method and face recognition method |
| CN113723238A (en) * | 2021-08-18 | 2021-11-30 | 北京深感科技有限公司 | Human face lightweight network model construction method and human face recognition method |
| CN113657411A (en) * | 2021-08-23 | 2021-11-16 | 北京达佳互联信息技术有限公司 | Neural network model training method, image feature extraction method and related device |
| CN114663942A (en) * | 2022-03-17 | 2022-06-24 | 深圳数联天下智能科技有限公司 | Feature detection method, model training method, device, and medium |
| CN115661560A (en) * | 2022-06-30 | 2023-01-31 | 斑马网络技术有限公司 | Face detection method in cockpit, target detection model training method and device |
| CN117542085A (en) * | 2024-01-10 | 2024-02-09 | 湖南工商大学 | Park scene pedestrian detection method, device and equipment based on knowledge distillation |
| CN117542085B (en) * | 2024-01-10 | 2024-05-03 | 湖南工商大学 | Park scene pedestrian detection method, device and equipment based on knowledge distillation |
| CN118799844A (en) * | 2024-09-13 | 2024-10-18 | 南京智莲森信息技术有限公司 | A railway pole number plate and kilometer mark identification method and related device |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108664893B (en) | 2022-04-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108664893A (en) | A kind of method for detecting human face and storage medium | |
| CN113901977B (en) | A method and system for identifying electricity theft by power users based on deep learning | |
| CN113642431B (en) | Training method and device, electronic equipment and storage medium for target detection model | |
| CN109359693A (en) | A Power Quality Disturbance Classification Method | |
| CN110929577A (en) | An improved target recognition method based on YOLOv3 lightweight framework | |
| CN109376615A (en) | For promoting the method, apparatus and storage medium of deep learning neural network forecast performance | |
| CN107688784A (en) | A kind of character identifying method and storage medium based on further feature and shallow-layer Fusion Features | |
| CN110516677A (en) | A kind of neural network recognization model, target identification method and system | |
| CN107945153A (en) | A kind of road surface crack detection method based on deep learning | |
| CN109241846A (en) | Change in time and space estimating and measuring method, device and the storage medium of remote sensing image | |
| CN110222728A (en) | The training method of article discrimination model, system and article discrimination method, equipment | |
| CN114170531B (en) | Infrared image target detection method and device based on difficult sample transfer learning | |
| CN112699832B (en) | Target detection method, device, equipment and storage medium | |
| CN109345553A (en) | A kind of palm and its critical point detection method, apparatus and terminal device | |
| CN112905997B (en) | Method, device and system for detecting poisoning attack facing deep learning model | |
| CN111353377A (en) | Elevator passenger number detection method based on deep learning | |
| CN106778910A (en) | Deep learning system and method based on local training | |
| CN111428419A (en) | Suspended sediment concentration prediction method and device, computer equipment and storage medium | |
| CN108027896A (en) | System and method for decoding the pulse reservoir with continuous synaptic plasticity | |
| CN109614990A (en) | A kind of object detecting device | |
| WO2022100607A1 (en) | Method for determining neural network structure and apparatus thereof | |
| CN108961358A (en) | A kind of method, apparatus and electronic equipment obtaining samples pictures | |
| CN107729947A (en) | A kind of Face datection model training method, device and medium | |
| CN110855474A (en) | Network feature extraction method, device, equipment and storage medium of KQI data | |
| CN116777814A (en) | Image processing method, apparatus, computer device, storage medium, and program product |
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 | ||
| CB02 | Change of applicant information | ||
| CB02 | Change of applicant information |
Address after: 350003 21 floors, No. 1 Building, G District, 89 Software Avenue, Gulou District, Fuzhou City, Fujian Province Applicant after: FUJIAN HAIJING TECHNOLOGY DEVELOPMENT Co.,Ltd. Address before: 350003 1-2 floors of Building C, Building 10, Fuzhou Software Park B, 89 Software Avenue, Gulou District, Fuzhou City, Fujian Province Applicant before: FUZHOU HAIJING SCIENCE & TECHNOLOGY DEVELOPMENT CO.,LTD. |
|
| CB03 | Change of inventor or designer information | ||
| CB03 | Change of inventor or designer information |
Inventor after: Huang Haiqing Inventor after: Wang Jinqiao Inventor after: Chen Yingying Inventor after: Liu Zhiyong Inventor after: Zheng Suiwu Inventor after: Yang Xu Inventor after: Huang Zhiming Inventor after: Xie Dekun Inventor after: Tian Jian Inventor before: Huang Haiqing Inventor before: Liu Zhiyong Inventor before: Zheng Suiwu Inventor before: Yang Xu Inventor before: Huang Zhiming Inventor before: Xie Dekun Inventor before: Tian Jian |
|
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| PP01 | Preservation of patent right | ||
| PP01 | Preservation of patent right |
Effective date of registration: 20231212 Granted publication date: 20220429 |