CN105760859B - Reticulate pattern facial image recognition method and device based on multitask convolutional neural networks - Google Patents
Reticulate pattern facial image recognition method and device based on multitask convolutional neural networks Download PDFInfo
- Publication number
- CN105760859B CN105760859B CN201610165633.7A CN201610165633A CN105760859B CN 105760859 B CN105760859 B CN 105760859B CN 201610165633 A CN201610165633 A CN 201610165633A CN 105760859 B CN105760859 B CN 105760859B
- Authority
- CN
- China
- Prior art keywords
- image
- reticulate pattern
- convolutional neural
- neural networks
- 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
- 238000013527 convolutional neural network Methods 0.000 title claims abstract description 71
- 230000001815 facial effect Effects 0.000 title claims abstract description 60
- 238000000034 method Methods 0.000 title claims abstract description 35
- 238000012549 training Methods 0.000 claims abstract description 67
- 230000008569 process Effects 0.000 claims description 9
- 238000013528 artificial neural network Methods 0.000 claims description 8
- 238000013480 data collection Methods 0.000 claims 1
- 230000006870 function Effects 0.000 abstract description 19
- 230000000694 effects Effects 0.000 abstract description 9
- 238000011084 recovery Methods 0.000 abstract description 2
- 238000005516 engineering process Methods 0.000 description 9
- 230000009466 transformation Effects 0.000 description 7
- 230000001537 neural effect Effects 0.000 description 3
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 238000005457 optimization Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 241000208340 Araliaceae Species 0.000 description 1
- 235000005035 Panax pseudoginseng ssp. pseudoginseng Nutrition 0.000 description 1
- 235000003140 Panax quinquefolius Nutrition 0.000 description 1
- 230000009471 action Effects 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 235000008434 ginseng Nutrition 0.000 description 1
- 238000003709 image segmentation Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 238000007689 inspection Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 238000003909 pattern recognition Methods 0.000 description 1
- 201000009032 substance abuse Diseases 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
- G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
- G06V40/16—Human faces, e.g. facial parts, sketches or expressions
- G06V40/172—Classification, e.g. identification
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
Abstract
The invention discloses a kind of reticulate pattern facial image recognition methods and device based on multitask convolutional neural networks.This method comprises: collecting reticulate pattern facial image and corresponding clear face image pair, then multitask convolutional neural networks are utilized, it separately designs based on the objective function for returning and classifying, one face descreening model of training, finally reticulate pattern facial image is input in trained descreening model, the facial image of descreening is obtained, to carry out subsequent recognition of face task.Present invention employs the frame of multi-task learning, the objective function that the Task expression for restoring clear image by reticulate pattern image is assisted each other at two, and converted using the complex nonlinear that convolutional neural networks study is directed to.Method used by inventing not only effectively increases convergence rate when model training, and can be obviously improved the effect and generalization ability of image recovery, and the recognition accuracy of reticulate pattern facial image is greatly improved.
Description
Technical field
The present invention relates to computer vision, pattern-recognition, the technical fields such as machine learning are especially a kind of to be based on multitask
Reticulate pattern facial image recognition method (Multi-task ConvNet for Face Recognition, the letter of convolutional neural networks
Claim MTCN).
Background technique
As one kind of biometrics identification technology, recognition of face due to its untouchable and accurate, convenient feature,
With good development and application prospect.Face recognition technology has all played highly important effect in many application scenarios,
Such as airport security, frontier inspection clearance etc..Traditional face recognition technology acquires in different periods mainly under Same Scene
Data.But with the horizontal raising of current face's identification technology, in order to more easily use face recognition technology.Identity-based
The face recognition technology that card-living photo compares gradually starts to obtain more concerns.This technology can be launched more easily
Airport clearance and bank remotely open an account in equal application scenarios, can not have to register can correctly identify in advance, increase substantially
Convenience when face recognition technology uses.
Identity card-living photo comparison technology or the difficult problem of a comparison at present.On the one hand, due to acquisition environment and
The difference of equipment is acquired, substantially, this is a heterogeneous recognition of face problem.In heterogeneous recognition of face problem, due to numerous originals
Difference is very big in class because caused by, thus is very difficult to solve.On the other hand, in order to protect the ID Card Image of user not by
Criminal abuses, and the identity card picture read out by card reader of ID card joined random reticulate pattern mostly and do watermark.This
A little moires make privacy of user be able to be protected, and but carry out recognition of face to machine and cause huge interference.Random net
Line greatly reduce Face datection, critical point detection and feature extraction and etc. accuracy rate, to seriously affect final
Recognition effect.
In recent years, deep learning all achieves the effect to attract people's attention in the various fields of machine vision.Wherein look steadily the most
Purpose model surely belongs to convolutional neural networks, which can extract hierarchical feature using multilayer convolutional layer and pond layer,
Realize stronger non-linear expression.Convolutional neural networks are in object classification, action recognition, the neck such as image segmentation and recognition of face
Domain achieves the effect for being significantly stronger than conventional method.In some Low Level Vision problems, for example, image denoising, image oversubscription
Distinguish, the problems such as image deblurring in, this model also all achieves good results.
In order to solve the problems, such as the identification of reticulate pattern facial image, we can be recovered from reticulate pattern facial image by algorithm
Carry out the clear face image without reticulate pattern.One can be trained to be used to directly from reticulate pattern using convolutional neural networks as described above
The model of clearly facial image is recovered in human face data.But simply use reticulate pattern image as input, it is clear to scheme
As it is more general to train the neural network effect come, and training process is veryer long as output, restrain slower.Moreover,
Since training process lacks certain priori knowledge guidance, trained model generalization performance is poor.
Summary of the invention
In order to improve accuracy rate of the reticulate pattern facial image for recognition of face when, the invention proposes one kind to be based on multitask
The reticulate pattern facial image recognition method of convolutional neural networks.In order to improve accuracy rate of the reticulate pattern image for recognition of face when, this
Invention recovers the clear face image without reticulate pattern with reticulate pattern image first.The present invention uses based on convolutional neural networks
Model framework devises the Optimized model of a multitask according to the priori knowledge of specific tasks, at the same optimize residual error return and
Reticulate pattern two tasks of prediction, while optimization is conducive to be promoted the convergence rate of network training, and the model obtained finally has comparatively fast
Convergence rate and recovery effects, and effectively increase the generalization ability of model.
The invention proposes a kind of reticulate pattern facial image recognition methods based on multitask convolutional neural networks, specifically according to
Following steps are implemented:
Step S1, reticulate pattern facial image and corresponding clear face image are collected to as sample image pair, forms training
Data set obtains the label figure of instruction reticulate pattern position to by threshold method for each sample image.
Step S2, training obtains the convolutional Neural net for going out the clear face image without reticulate pattern from reticulate pattern face image restoration
Network model, comprising:
The sample image pair concentrated using the training data, one multitask convolutional neural networks of training, training process
In, multiple convolutional layers of the multitask convolutional neural networks first half handle the reticulate pattern facial image of input, later half
Part is divided into two tasks, and being utilized respectively that treated, data are trained corresponding loss objective function；Wherein described two
Main task in a task is used to return the difference of clear face image and reticulate pattern facial image, obtains residual image；Auxiliary
Task is used to predict the reticulate pattern position of reticulate pattern facial image, the reticulate pattern image predicted；Final trained convolutional Neural net
The final output of network model is being added for the residual image and reticulate pattern position, i.e., without the clear face image of reticulate pattern；
Step S3, using trained convolutional neural networks model, clear face image to be identified is recovered, and use
Clear face image to be identified carries out recognition of face.
The invention also provides a kind of reticulate pattern facial image identification device based on multitask convolutional neural networks, comprising:
Training sample acquisition module, for collecting reticulate pattern facial image and corresponding clear face image to as sample graph
As right, formation training dataset obtains the label figure of instruction reticulate pattern position to by threshold method for each sample image.
Convolutional neural networks training module obtains going out the clear people without reticulate pattern from reticulate pattern face image restoration for training
The convolutional neural networks model of face image, comprising:
The sample image pair concentrated using the training data, one multitask convolutional neural networks of training, training process
In, multiple convolutional layers of the multitask convolutional neural networks first half handle the reticulate pattern facial image of input, later half
Part is divided into two tasks, and being utilized respectively that treated, data are trained corresponding loss objective function；Wherein described two
Main task in a task is used to return the difference of clear face image and reticulate pattern facial image, obtains residual image；Auxiliary
Task is used to predict the reticulate pattern position of reticulate pattern facial image, the reticulate pattern image predicted；Final trained image recognition mould
The final output of type is being added for the residual image and reticulate pattern position, i.e., without the clear face image of reticulate pattern；
Identification module, for recovering clear face image to be identified using trained convolutional neural networks model,
And recognition of face is carried out using clear face image to be identified.
Beneficial effects of the present invention: the present invention is for a particular problem in recognition of face, identity card photograph-living photo ratio
To problem, proposing one has the method that meaning is widely applied.This method can be from net by multitask convolutional neural networks
The clear face image without reticulate pattern is recovered in print image, and carries out recognition of face using the clear face image recovered.
Multitask convolutional neural networks model proposed by the present invention, has used the optimal way of multitask multiple target, so that model is restrained
Faster, effect is more preferable, and Generalization Capability is stronger.The first descreening proposed through the invention identifying schemes again, can be significantly
Improve the accuracy rate of the recognition of face of reticulate pattern image.
Detailed description of the invention
Fig. 1 is the example of the reticulate pattern facial image studied and clear face image of the invention；
Fig. 2 is the method flow of the reticulate pattern facial image recognition method in the present invention based on multitask convolutional neural networks
Figure；
Fig. 3 is the network diagram of multitask convolutional neural networks model in the present invention.
Specific embodiment
To make the objectives, technical solutions, and advantages of the present invention clearer, below in conjunction with specific embodiment, and reference
Attached drawing, the present invention is described in further detail.
The present invention learns the transformation of a nonlinearity by the convolutional neural networks of multitask, is used to from reticulate pattern image
In recover clear image without reticulate pattern, and carry out subsequent recognition of face using clear image.
Fig. 1 is the exemplary diagram of reticulate pattern facial image and clear face image used in the present invention.
Fig. 2 is a kind of reticulate pattern facial image recognition method process based on multitask convolutional neural networks proposed by the present invention
Figure, this method as shown in Figure 2 including the following steps:
Step S1 acquires reticulate pattern facial image and corresponding clear face image to as training dataset, described image
Pair in the same size, grayscale image or color image；To described image to making the difference, instruction reticulate pattern is obtained according to certain threshold value
The bianry image of distributing position is as label figure.Wherein, reticulate pattern facial image is expressed as x, and corresponding clear face chart is shown as
Y, corresponding label chart are shown as lr.
Step S2, the image pair concentrated using the training data, training multitask convolutional neural networks model, to be used to
Clearly facial image y is recovered from reticulate pattern facial image x.In order to guarantee that input and output image size is identical, convolutional Neural
Network only includes convolutional layer, and is all added in each convolutional layer and mends side operation, i.e., by expanded images edge pixel, so that volume
After product operation, the size of image is constant.In one embodiment, the convolutional neural networks structure is by 6 layers of convolutional layer and corresponding
It corrects linear unit to constitute, wherein every layer of convolutional layer has the filter of 64 3*3 sizes.The number of plies of convolutional layer and every layer of convolution
The number and size of filter can carry out selection setting according to the actual situation in layer.As shown in figure 3, original image is input to institute
State the input layer in convolutional neural networks, after multilayer convolutional layer and nonlinear transformation layer operation, obtain convolution results and
Nonlinear transformation result；In task one, the convolution sum nonlinear transformation result of abovementioned layers is input to only one convolution
In the filter of core, a size and the identical characteristic pattern of input picture are obtained, this is as convolutional neural networks to the pre- of residual error
It surveys, for the loss with true value together calculating task one.In task two, equally by the convolution sum nonlinear transformation of abovementioned layers
As a result it is input in the filter of other only one convolution kernel, and obtained characteristic pattern is passed through into a S type function
Carry out nonlinear transformation, the result of transformation can be used as pollution position prediction as a result, with the true value one that is provided in training set
It rises and calculates target loss.
In the step, the sample image pair concentrated using the training data, one multitask convolutional neural networks of training,
In training process, multiple convolutional layers of the multitask convolutional neural networks first half to the reticulate pattern facial image of input at
Reason, latter half is divided into two tasks, and being utilized respectively that treated, data are trained corresponding loss objective function；Wherein
Main task in described two tasks is used to return the difference of clear face image and reticulate pattern facial image, obtains residual plot
Picture；Nonproductive task is used to predict the reticulate pattern position of reticulate pattern facial image, the reticulate pattern image predicted；Final trained convolution
The final output of neural network model is being added for the residual image and reticulate pattern position, i.e., without the clear face figure of reticulate pattern
Picture.
The nonlinearity capability of fitting that the convolutional neural networks are utilized in the present invention restores clear for reticulate pattern image
This specific tasks of image, construction are to input with reticulate pattern image, and the error image of clear image and reticulate pattern image is as output
Convolutional neural networks model.Particularly, which can return out the difference of clear image and reticulate pattern image, rather than directly
It returns and comes out clear image.In this way, by network as shown in Figure 2, using two relevant tasks, can train one can be with
Restore the convolutional neural networks of clear image, so as to the clear face for polluting reticulate pattern face image restoration at no reticulate pattern
Image.Although the output of network is the difference of clear image Yu reticulate pattern image, when identification, it is only necessary to by the defeated of network
The clear image recovered can be obtained with reticulate pattern image addition out.
Specifically, two tasks of multiple target convolutional neural networks are respectively to return the difference of clear image and reticulate pattern image
And the position of prediction reticulate pattern.Particularly, the objective function for returning difference can be by minimizing using mean square error as representative
A series of objective functions are completed, such as:
Wherein, J1(w1) indicate that clear image returns the objective function of difference, w1It needs to train for what the objective function was related to
Parameter, r=x-y be reticulate pattern image x and prediction clear image y residual error,The clear image that neural network is acquired is represented to return
Device, i, j indicate that the pixel coordinate in image, N indicate the training sample sum that training data is concentrated.
The task of prediction reticulate pattern position can be expressed as whether a certain pixel of forecast image is this two classification problem of reticulate pattern.
The problem, which can be used, minimizes a series of Classification Loss functions by representative of logistical regression loss as optimization aim, than
Such as:
Wherein, J2(w2) indicate to predict the objective function of reticulate pattern position, w2For the ginseng for the needs training that the objective function is related to
Number, φ expression represent the reticulate pattern position forecaster that neural network is acquired, and l is the binary map for indicating reticulate pattern position in input picture
Picture, size is contaminated for characterizing which region of input picture as input picture, and the region that wherein pixel is zero is not
Polluted area, pixel are 1 and illustrate that the region is contaminated.
Above-mentioned convolutional neural networks, the task of regression residuals are and to predict the task of reticulate pattern position as main task
It is nonproductive task, the final goal function of the convolutional neural networks are as follows: J (w1, w2)=J1(w1)+αJ2(w2), α is that auxiliary is appointed
The weight parameter of business.Parameter in above-mentioned objective function in all parameters and abovementioned layers can be above-mentioned total by minimizing
Objective function carry out.The training of model can be carried out by back-propagation algorithm, and continuous iteration updates each layer parameter, be come minimum
Change the objective function.
The convolutional neural networks are trained as follows:
Step S21: identical weight is added to two tasks in the initial stage of network training, i.e. initialization α=1；
Step S22: the sample input that the training data is concentrated is trained as the convolutional neural networks, until
Total objective function J (w1, w2)=J1(w1)+αJ2(w2) tend towards stability；
Step S23: reducing the weight parameter of the nonproductive task, goes to step S22 and continues to train, until the nonproductive task
Weight parameter be reduced to 0；
Step S24: continuing to train, until training loss J (w1, w2)=J1(w1)+αJ2(w2) do not continue to reduce, at this moment
The current value that each layer parameter in network can be preserved, the parameter as final mask.
Step S3 is entered into trained convolutional neural networks new reticulate pattern facial image, obtains clear
The difference of clear image and reticulate pattern image, the value and reticulate pattern image addition can recover the clear face image without reticulate pattern.
Next traditional recognition of face step can be used, by Face datection, after critical point detection and feature extraction, carries out phase
The aspect ratio pair answered completes recognition of face task.
For the specific embodiment and verifying effectiveness of the invention that the present invention will be described in detail, we propose the present invention
Method apply a reticulate pattern image recognition of face task.Specifically, the multitask convolutional Neural net in order to train descreening
Network model, we have prepared 500,000 with the facial image of reticulate pattern and its is corresponding without reticulate pattern clear image, and calculate
Come indicate correspondence image pair reticulate pattern position label figure.The network structure and objective function designed using us, with reticulate pattern people
Face image is input, utilizes the gradient anti-pass training neural network.The weight of different task is constantly adjusted in training process, until
Last network convergence obtains the model for restoring clear face image.
In order to test the validity of the model, we have additionally prepared the clearly identity card picture of 300 people, and (everyone one
) and accordingly the living photo 300 of individual is opened, it is notable that this 300 people are not present in training set.The data set
The effect of algorithm when identity card-living photo compares can be used to test.Using clear identity card according to and living photo be compared
When, accuracy rate when being identified using a kind of depth characteristic of the heterogeneous recognition of face of special disposal is as shown in 1 first row of table.
Subsequent experiment, we use this feature to carry out recognition of face.When being used for recognition of face for calibration tape reticulate pattern image
Discrimination, we give this 300 clearly face random reticulate patterns are added, generate the identity card with reticulate pattern and shine.It is using this
Column image and corresponding clear living photo carry out recognition of face, it has been found that its discrimination (such as table 1) is greatly lowered.Later,
We use trained multitask convolutional neural networks model, recover clearly from the above-mentioned facial image with reticulate pattern
Facial image, then recognition of face is carried out with corresponding living photo, specific recognition result is as shown in table 1.Although with original clear figure
The discrimination of piece has a certain gap, but has had a big promotion compared with the recognition result with reticulate pattern facial image.
The validity that embodiment valid certificates method proposed by the invention identifies reticulate pattern facial image.
Table 1 is used by the present invention treated face recognition accuracy rate and untreated reticulate pattern image and normal clear image
It is as follows in the accuracy rate comparing result table that recognition of face is:
| TPR@FPR=1% | TPR@FPR=0.1% | TPR@FPR=0.01% | |
| Clear image | 91.20 | 75.20 | 50.20 |
| Reticulate pattern image | 49.00 | 30.80 | 16.40 |
| Descreening image | 81.00 | 57.80 | 29.20 |
Particular embodiments described above has carried out further in detail the purpose of the present invention, technical scheme and beneficial effects
Describe in detail bright, it should be understood that the above is only a specific embodiment of the present invention, is not intended to restrict the invention, it is all
Within the spirit and principles in the present invention, any modification, equivalent substitution, improvement and etc. done should be included in protection of the invention
Within the scope of.
Claims (10)
1. a kind of reticulate pattern facial image recognition method based on multitask convolutional neural networks, which is characterized in that specifically according to
Lower step is implemented:
Step S1, reticulate pattern facial image and corresponding clear face image are collected to as sample image pair, forms training data
Collection obtains the label figure of instruction reticulate pattern position to by threshold method for each sample image；
Step S2, training obtains the convolutional neural networks mould for going out the clear face image without reticulate pattern from reticulate pattern face image restoration
Type, comprising:
The sample image pair concentrated using the training data, one multitask convolutional neural networks of training should in training process
Multiple convolutional layers of multitask convolutional neural networks first half handle the reticulate pattern facial image of input, latter half point
For two tasks, it is utilized respectively multitask convolutional neural networks first half treated that data are appointed to by two of latter half
The loss objective function for being engaged in respectively constituting is trained；Wherein the main task in described two tasks is for returning clear face
The difference of image and reticulate pattern facial image, obtains residual image；Nonproductive task is used to predict the reticulate pattern position of reticulate pattern facial image,
The reticulate pattern image predicted；The final output of final trained convolutional neural networks model is the residual image and reticulate pattern
The addition of position, i.e., without the clear face image of reticulate pattern；
Step S3, using trained convolutional neural networks model, clear face image to be identified is recovered, and use is wait know
Other clear face image carries out recognition of face.
2. the reticulate pattern facial image recognition method according to claim 1 based on multitask convolutional neural networks, feature
Be, in the step S1, reticulate pattern facial image collected and corresponding clear face image pair it is in the same size and right
Make in bianry image of each sample image that training data is concentrated to the position for obtaining instruction reticulate pattern distribution according to certain threshold value
For label figure.
3. the reticulate pattern facial image recognition method according to claim 1 based on multitask convolutional neural networks, feature
It is, the step S2 includes:
Step S21: initialization is so that the weight parameter of two tasks is equal, wherein total objective function is J (w1,w2)=J1
(w1)+αJ2(w2), J1(w1) indicate main task objective function, J2(w2) be nonproductive task objective function；w1、w2Respectively
Training parameter in main task and nonproductive task objective function；α is the weight parameter of nonproductive task, initializes α in the step
=1；
Step S22: the sample input that the training data is concentrated is trained as the convolutional neural networks, until total
Objective function J (w1,w2)=J1(w1)+αJ2(w2) tend towards stability, wherein；
Step S23: reducing the weight parameter of the nonproductive task, goes to step S22 and continues to train, until the power of the nonproductive task
Weight parameter is reduced to 0；
Step S24: continuing to train, until training loss no longer reduces, to obtain final convolutional neural networks model.
4. as claimed in claim 3 based on the reticulate pattern facial image recognition method of multitask convolutional neural networks, feature exists
In the objective function of main task indicates as follows:
Wherein, r=x-y is the residual error of reticulate pattern image x and prediction clear image y,The clear image that neural network is acquired is represented to return
Return device, i, j indicate that the pixel coordinate in image, N indicate the training sample sum that training data is concentrated.
5. as claimed in claim 3 based on the reticulate pattern facial image recognition method of multitask convolutional neural networks, feature exists
In the objective function of nonproductive task indicates as follows:
Wherein, r=x-y is the residual error of reticulate pattern image x and prediction clear image y, and φ expression represents the reticulate pattern that neural network is acquired
Position forecaster, i, j indicate that the pixel coordinate in image, N indicate that the training sample sum that training data is concentrated, l are to indicate net
The bianry image of reticulate pattern position in print image.
6. a kind of reticulate pattern facial image identification device based on multitask convolutional neural networks characterized by comprising
Training sample acquisition module, for collecting reticulate pattern facial image and corresponding clear face image to as sample image
It is right, training dataset is formed, obtains the label figure of instruction reticulate pattern position to by threshold method for each sample image；
Convolutional neural networks training module obtains going out the clear face figure without reticulate pattern from reticulate pattern face image restoration for training
The convolutional neural networks model of picture, comprising:
The sample image pair concentrated using the training data, one multitask convolutional neural networks of training should in training process
Multiple convolutional layers of multitask convolutional neural networks first half handle the reticulate pattern facial image of input, latter half point
For two tasks, it is utilized respectively multitask convolutional neural networks first half treated that data are appointed to by two of latter half
The loss objective function for being engaged in respectively constituting is trained；Wherein the main task in described two tasks is for returning clear face
The difference of image and reticulate pattern facial image, obtains residual image；Nonproductive task is used to predict the reticulate pattern position of reticulate pattern facial image,
The reticulate pattern image predicted；The final output of final trained image recognition model is the residual image and reticulate pattern position
Addition, i.e., without the clear face image of reticulate pattern；
Identification module recovers clear face image to be identified, and make for using trained convolutional neural networks model
Recognition of face is carried out with clear face image to be identified.
7. the reticulate pattern facial image identification device according to claim 6 based on multitask convolutional neural networks, feature
Be, reticulate pattern facial image collected and corresponding clear face image pair it is in the same size, and training data is concentrated
Each sample image to according to certain threshold value obtain instruction reticulate pattern distribution position bianry image as label figure.
8. the reticulate pattern facial image identification device according to claim 6 based on multitask convolutional neural networks, feature
It is, the convolutional neural networks training module includes:
Initialization module, for initializing so that the weight parameter of two tasks is equal, wherein total objective function is J (w1,
w2)=J1(w1)+αJ2(w2), J1(w1) indicate main task objective function, J2(w2) be nonproductive task objective function；w1、w2
Training parameter respectively in main task and nonproductive task objective function；α is the weight parameter of nonproductive task, in the step just
Beginningization α=1；
Initial training module, the sample input for concentrating the training data are instructed as the convolutional neural networks
Practice, until total objective function J (w1,w2)=J1(w1)+αJ2(w2) tend towards stability, wherein；
Depth training module turns initial training module and continues to train for reducing the weight parameter of the nonproductive task, until institute
The weight parameter for stating nonproductive task is reduced to 0；
Training result output module is trained for continuing, until training loss no longer reduces, to obtain final image recognition
Model.
9. the reticulate pattern facial image identification device based on multitask convolutional neural networks, feature exist as claimed in claim 8
In the objective function of main task indicates as follows:
Wherein, r=x-y is the residual error of reticulate pattern image x and prediction clear image y,The clear image that neural network is acquired is represented to return
Return device, i, j indicate that the pixel coordinate in image, N indicate the training sample sum that training data is concentrated.
10. the reticulate pattern facial image identification device based on multitask convolutional neural networks, feature exist as claimed in claim 8
In the objective function of nonproductive task indicates as follows:
Wherein, r=x-y is the residual error of reticulate pattern image x and prediction clear image y, and φ expression represents the reticulate pattern that neural network is acquired
Position forecaster, i, j indicate that the pixel coordinate in image, N indicate that the training sample sum that training data is concentrated, l are to indicate net
The bianry image of reticulate pattern position in print image.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610165633.7A CN105760859B (en) | 2016-03-22 | 2016-03-22 | Reticulate pattern facial image recognition method and device based on multitask convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201610165633.7A CN105760859B (en) | 2016-03-22 | 2016-03-22 | Reticulate pattern facial image recognition method and device based on multitask convolutional neural networks |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105760859A CN105760859A (en) | 2016-07-13 |
| CN105760859B true CN105760859B (en) | 2018-12-21 |
Family
ID=56345590
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201610165633.7A Active CN105760859B (en) | 2016-03-22 | 2016-03-22 | Reticulate pattern facial image recognition method and device based on multitask convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105760859B (en) |
Families Citing this family (35)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106295678B (en) | 2016-07-27 | 2020-03-06 | 北京旷视科技有限公司 | Neural network training and constructing method and device and target detection method and device |
| CN106529402B (en) * | 2016-09-27 | 2019-05-28 | 中国科学院自动化研究所 | The face character analysis method of convolutional neural networks based on multi-task learning |
| CN107992783A (en) * | 2016-10-26 | 2018-05-04 | 上海银晨智能识别科技有限公司 | Face image processing process and device |
| CN106548159A (en) * | 2016-11-08 | 2017-03-29 | 中国科学院自动化研究所 | Reticulate pattern facial image recognition method and device based on full convolutional neural networks |
| US10394929B2 (en) * | 2016-12-20 | 2019-08-27 | Mediatek, Inc. | Adaptive execution engine for convolution computing systems |
| CN108205802A (en) * | 2016-12-23 | 2018-06-26 | 北京市商汤科技开发有限公司 | Deep neural network model training, image processing method and device and equipment |
| CN106815566B (en) * | 2016-12-29 | 2021-04-16 | 天津中科智能识别产业技术研究院有限公司 | Face retrieval method based on multitask convolutional neural network |
| CN107563372B (en) * | 2017-07-20 | 2021-01-29 | 济南中维世纪科技有限公司 | License plate positioning method based on deep learning SSD frame |
| US11341631B2 (en) | 2017-08-09 | 2022-05-24 | Shenzhen Keya Medical Technology Corporation | System and method for automatically detecting a physiological condition from a medical image of a patient |
| CN107545277B (en) * | 2017-08-11 | 2023-07-11 | 腾讯科技（上海）有限公司 | Model training, identity verification method and device, storage medium and computer equipment |
| CN107491767B (en) * | 2017-08-31 | 2020-06-12 | 云从科技集团股份有限公司 | End-to-end unconstrained face key point detection method |
| CN107766844A (en) * | 2017-11-13 | 2018-03-06 | 杭州有盾网络科技有限公司 | Method, apparatus, equipment of a kind of reticulate pattern according to recognition of face |
| CN107993190B (en) * | 2017-11-14 | 2020-05-19 | 中国科学院自动化研究所 | Image watermark removing device |
| CN107798667A (en) * | 2017-11-23 | 2018-03-13 | 中电科新型智慧城市研究院有限公司 | Face Enhancement Method based on residual error study |
| CN108229349B (en) * | 2017-12-21 | 2020-09-01 | 中国科学院自动化研究所 | Reticulate pattern human face image recognition device |
| CN108596818B (en) * | 2017-12-22 | 2021-10-26 | 华南理工大学 | Image steganalysis method based on multitask learning convolutional neural network |
| US10699358B2 (en) * | 2018-02-22 | 2020-06-30 | Mcafee, Llc | Image hidden information detector |
| CN108399649B (en) * | 2018-03-05 | 2021-07-20 | 中科视拓（北京）科技有限公司 | Single-picture three-dimensional face reconstruction method based on cascade regression network |
| CN108510454A (en) * | 2018-03-21 | 2018-09-07 | 百度在线网络技术（北京）有限公司 | Method and apparatus for generating depth image |
| CN108734673B (en) * | 2018-04-20 | 2019-11-15 | 平安科技（深圳）有限公司 | Descreening systematic training method, descreening method, apparatus, equipment and medium |
| CN109147940B (en) * | 2018-07-05 | 2021-05-25 | 科亚医疗科技股份有限公司 | Apparatus and system for automatically predicting physiological condition from medical image of patient |
| CN109002790A (en) * | 2018-07-11 | 2018-12-14 | 广州视源电子科技股份有限公司 | A kind of method, apparatus of recognition of face, equipment and storage medium |
| CN110795976B (en) | 2018-08-03 | 2023-05-05 | 华为云计算技术有限公司 | Method, device and equipment for training object detection model |
| CN109543763B (en) * | 2018-11-28 | 2022-10-21 | 重庆大学 | Raman spectrum analysis method based on convolutional neural network |
| CN111260548B (en) * | 2018-11-30 | 2023-07-21 | 浙江宇视科技有限公司 | Mapping method and device based on deep learning |
| CN109801225B (en) * | 2018-12-06 | 2022-12-27 | 重庆邮电大学 | Human face reticulate pattern stain removing method based on multitask full convolution neural network |
| CN110008817B (en) * | 2019-01-29 | 2021-12-28 | 北京奇艺世纪科技有限公司 | Model training method, image processing method, device, electronic equipment and computer readable storage medium |
| CN110084143B (en) * | 2019-04-04 | 2021-09-28 | 广州大学 | Emotion information protection method and system for face recognition |
| CN110084147B (en) * | 2019-04-04 | 2021-09-28 | 广州大学 | Gender privacy protection method and system for face recognition |
| CN110070027A (en) * | 2019-04-17 | 2019-07-30 | 南京邮电大学 | Pedestrian based on intelligent internet of things system recognition methods again |
| CN110232352B (en) * | 2019-06-12 | 2022-11-25 | 东北大学 | Improvement method of multi-task cascade convolution neural network model for face recognition |
| CN110263756A (en) * | 2019-06-28 | 2019-09-20 | 东北大学 | A kind of human face super-resolution reconstructing system based on joint multi-task learning |
| CN110647805B (en) * | 2019-08-09 | 2023-10-31 | 平安科技（深圳）有限公司 | Reticulate pattern image recognition method and device and terminal equipment |
| KR20210057401A (en) * | 2019-11-12 | 2021-05-21 | 한국전자통신연구원 | Apparatus and method for authenticating user based on multiple biometric infomation |
| CN111612133B (en) * | 2020-05-20 | 2021-10-19 | 广州华见智能科技有限公司 | Internal organ feature coding method based on face image multi-stage relation learning |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090116747A1 (en) * | 2006-08-01 | 2009-05-07 | Duong Tuan A | Artificial intelligence systems for identifying objects |
| CN105184303A (en) * | 2015-04-23 | 2015-12-23 | 南京邮电大学 | Image marking method based on multi-mode deep learning |
-
2016
- 2016-03-22 CN CN201610165633.7A patent/CN105760859B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20090116747A1 (en) * | 2006-08-01 | 2009-05-07 | Duong Tuan A | Artificial intelligence systems for identifying objects |
| CN105184303A (en) * | 2015-04-23 | 2015-12-23 | 南京邮电大学 | Image marking method based on multi-mode deep learning |
Non-Patent Citations (4)
| Title |
|---|
| Face occlusion detection based on multi-task convolution neural network;Yizhang Xia等;《2015 12th International Conference on Fuzzy Systems and Knowledge Discovery (FSKD)》;20160114;正文第375-379页 * |
| Joint space learning for video-based face recognition;Dong Cao等;《2015 3rd IAPR Asian Conference on Pattern Recognition (ACPR)》;20151006;正文第16-20页 * |
| Multi-Task CNN Model for Attribute Prediction;Abrar H. Abdulnabi等;《IEEE Transactions on Multimedia》;20150910;第17卷(第11期);正文第1949-1959页 * |
| When Face Recognition Meets with Deep Learning: An Evaluation of Convolutional Neural Networks for Face Recognition;Guosheng Hu等;《2015 IEEE International Conference on Computer Vision Workshop (ICCVW)》;20160215;正文第384-392页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105760859A (en) | 2016-07-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105760859B (en) | Reticulate pattern facial image recognition method and device based on multitask convolutional neural networks | |
| CN110348319B (en) | Face anti-counterfeiting method based on face depth information and edge image fusion | |
| CN109934200B (en) | RGB color remote sensing image cloud detection method and system based on improved M-Net | |
| CN105095856B (en) | Face identification method is blocked based on mask | |
| CN105512638B (en) | A kind of Face datection and alignment schemes based on fusion feature | |
| CN112950661B (en) | Attention-based generation method for generating network face cartoon | |
| WO2020007307A1 (en) | Sky filter method for panoramic images and portable terminal | |
| CN103198493B (en) | A kind ofly to merge and the method for tracking target of on-line study based on multiple features self-adaptation | |
| CN106548159A (en) | Reticulate pattern facial image recognition method and device based on full convolutional neural networks | |
| CN109509164B (en) | Multi-sensor image fusion method and system based on GDGF | |
| CN106156684B (en) | A kind of two-dimensional code identification method and device | |
| CN104424457A (en) | Method for identifying two-dimensional code under the condition of nonlinear distortion | |
| CN103208004A (en) | Automatic recognition and extraction method and device for bill information area | |
| CN104915676A (en) | Deep-level feature learning and watershed-based synthetic aperture radar (SAR) image classification method | |
| CN107766890A (en) | The improved method that identification segment learns in a kind of fine granularity identification | |
| Wazalwar et al. | A design flow for robust license plate localization and recognition in complex scenes | |
| CN107066955B (en) | Method for restoring whole human face from local human face area | |
| CN110503613A (en) | Based on the empty convolutional neural networks of cascade towards removing rain based on single image method | |
| CN107480585A (en) | Object detection method based on DPM algorithms | |
| CN112668483A (en) | Single-target person tracking method integrating pedestrian re-identification and face detection | |
| CN103679136A (en) | Hand back vein identity recognition method based on combination of local macroscopic features and microscopic features | |
| CN103927555B (en) | Static manual alphabet identifying system and method based on Kinect sensor | |
| CN109740578A (en) | It is a kind of suitable for illumination, the face identification method of posture, expression shape change | |
| CN104778670A (en) | Fractal-wavelet self-adaption image denoising method based on multivariate statistical model | |
| CN108537143A (en) | A kind of face identification method and system based on key area aspect ratio pair |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |