CN109034078A - Training method, age recognition methods and the relevant device of age identification model - Google Patents

Training method, age recognition methods and the relevant device of age identification model Download PDF

Info

Publication number
CN109034078A
CN109034078A CN201810864567.1A CN201810864567A CN109034078A CN 109034078 A CN109034078 A CN 109034078A CN 201810864567 A CN201810864567 A CN 201810864567A CN 109034078 A CN109034078 A CN 109034078A
Authority
CN
China
Prior art keywords
age
training image
value
identification model
face
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
Application number
CN201810864567.1A
Other languages
Chinese (zh)
Other versions
CN109034078B (en
Inventor
葛彦昊
李季檩
汪铖杰
黄飞跃
吴永坚
黄小明
梁小龙
晏轶超
贺珂珂
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Tencent Technology Shenzhen Co Ltd
Tencent Cloud Computing Beijing Co Ltd
Original Assignee
Tencent Technology Shenzhen Co Ltd
Tencent Cloud Computing Beijing Co Ltd
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Tencent Technology Shenzhen Co Ltd, Tencent Cloud Computing Beijing Co Ltd filed Critical Tencent Technology Shenzhen Co Ltd
Priority to CN201810864567.1A priority Critical patent/CN109034078B/en
Publication of CN109034078A publication Critical patent/CN109034078A/en
Application granted granted Critical
Publication of CN109034078B publication Critical patent/CN109034078B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00—Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10—Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16—Human faces, e.g. facial parts, sketches or expressions
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • 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
    • G06V40/178—Human faces, e.g. facial parts, sketches or expressions estimating age from face image; using age information for improving recognition

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Multimedia (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Human Computer Interaction (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

This application involves a kind of training methods of age identification model, this method comprises: acquisition includes the training image collection of face, the training image that training image is concentrated is as the input of age identification model, obtain the corresponding prediction age value of face in each training image of age identification model output, according to the corresponding prediction age value of face in each target training image for corresponding to same mark age value, the statistical forecast age corresponding with each mark age value is calculated, age statistic error value is calculated according to statistical forecast age and corresponding mark age value, the parameter in age identification model is adjusted according to age statistic error value, until meeting the condition of convergence, obtain target age identification model.The training method of the age identification model improves the accuracy of age identification.Furthermore, it is also proposed that a kind of training device of age identification model, age recognition methods, device, computer equipment and storage medium.

Description

Training method, age recognition methods and the relevant device of age identification model
Technical field
This application involves computer processing technical fields, training method, year more particularly to a kind of age identification model Age recognition methods and relevant device.
Background technique
Age identification refers to be identified by age of the facial image to people.Traditional age identification model often can only The scene smaller to age range carries out age identification, right (for example, monitoring scene) under the bigger scene of age range It is poor in the recognition effect of certain age brackets.
Summary of the invention
Based on this, it is necessary to know in view of the above-mentioned problems, proposing a kind of pair of all age group recognition accuracy high age Training method, device, computer equipment and the storage medium of other model, age recognition methods, device, computer equipment and storage Medium.
A kind of training method of age identification model, which comprises
Acquisition includes the training image collection of face, and the face in each training image that the training image is concentrated exists Corresponding mark age value；
The training image that the training image is concentrated obtains the age identification mould as the input of age identification model The corresponding prediction age value of face in each training image of type output；
According to the corresponding prediction age value of face in each target training image for corresponding to same mark age value, meter Calculation obtains the statistical forecast age corresponding with each mark age value；
Age statistic error value is calculated according to the statistical forecast age and corresponding mark age value；
The parameter in the age identification model is adjusted according to the age statistic error value, until meeting convergence Condition obtains target age identification model.
A kind of training device of age identification model, described device include:
Image set obtains module, for obtains include face training image collection, training image concentration it is each There are corresponding mark age values for face in training image；
Training input/output module, the training image for concentrating the training image is as the defeated of age identification model Enter, obtains the corresponding prediction age value of face in each training image of the age identification model output；
First computing module, for according to the face pair in each target training image for corresponding to same mark age value The statistical forecast age corresponding with each mark age value is calculated in the prediction age value answered；
Second computing module, for age system to be calculated according to the statistical forecast age and corresponding mark age value Count error amount；
Module is adjusted, for adjusting according to the age statistic error value to the parameter in the age identification model It is whole, until meeting the condition of convergence, obtain target age identification model.
A kind of computer equipment, including memory and processor, the memory are stored with computer program, the calculating When machine program is executed by the processor, so that the processor executes following steps:
Acquisition includes the training image collection of face, and the face in each training image that the training image is concentrated exists Corresponding mark age value；
The training image that the training image is concentrated obtains the age identification mould as the input of age identification model The corresponding prediction age value of face in each training image of type output；
According to the corresponding prediction age value of face in each target training image for corresponding to same mark age value, meter Calculation obtains the statistical forecast age corresponding with each mark age value；
Age statistic error value is calculated according to the statistical forecast age and corresponding mark age value；
The parameter in the age identification model is adjusted according to the age statistic error value, until meeting convergence Condition obtains target age identification model.
A kind of computer readable storage medium is stored with computer program, when the computer program is executed by processor, So that the processor executes following steps:
Acquisition includes the training image collection of face, and the face in each training image that the training image is concentrated exists Corresponding mark age value；
The training image that the training image is concentrated obtains the age identification mould as the input of age identification model The corresponding prediction age value of face in each training image of type output；
According to the corresponding prediction age value of face in each target training image for corresponding to same mark age value, meter Calculation obtains the statistical forecast age corresponding with each mark age value；
Age statistic error value is calculated according to the statistical forecast age and corresponding mark age value；
The parameter in the age identification model is adjusted according to the age statistic error value, until meeting convergence Condition obtains target age identification model.
Training method, device, computer equipment and the storage medium of above-mentioned age identification model, according to corresponding to same mark The corresponding prediction age value of face infused in each target training image of age value is calculated and each mark age value pair Then age statistical error is calculated using statistical forecast age and corresponding mark age value in the statistical forecast age answered Value, and then the parameter in age identification model is adjusted according to age statistic error value, until meeting the condition of convergence, obtain Target age identification model.Age identification model is trained by using innovative age statistic error value, is obtained Target age identification model, even if can be accessed under the uncontrollable scene of age bracket to the face in all age group The relatively high recognition effect of accuracy.
A kind of age recognition methods, which comprises
Acquisition includes the images to be recognized of face；
Using the images to be recognized as the input of age identification model, the age identification model acceptable age statistics is missed For difference as error metrics standard, the age statistic error value is according to mark age value and to correspond to same mark age value Multiple training images in face corresponding prediction age value be calculated；
Obtain the age value corresponding with the face in the images to be recognized of the age identification model output.
A kind of age identification device, described device include:
Images to be recognized obtain module, for obtain include face images to be recognized；
Input module, for using the images to be recognized as the input of age identification model, the age identification model For acceptable age statistic error value as error metrics standard, the age statistic error value is according to mark age value and to correspond to What the corresponding prediction age value of face in multiple training images of same mark age value be calculated；
Output module, for obtaining that the age identification model trained exports and the people in the images to be recognized The corresponding age value of face.
A kind of computer equipment, including memory and processor, the memory are stored with computer program, the calculating When machine program is executed by the processor, so that the processor executes following steps:
Acquisition includes the images to be recognized of face；
Using the images to be recognized as the input of age identification model, the age identification model acceptable age statistics is missed For difference as error metrics standard, the age statistic error value is according to mark age value and to correspond to same mark age value Multiple training images in face corresponding prediction age value be calculated；
Obtain the age value corresponding with the face in the images to be recognized of the age identification model output.
A kind of computer readable storage medium is stored with computer program, when the computer program is executed by processor, So that the processor executes following steps:
Acquisition includes the images to be recognized of face；
Using the images to be recognized as the input of age identification model, the age identification model acceptable age statistics is missed For difference as error metrics standard, the age statistic error value is according to mark age value and to correspond to same mark age value Multiple training images in face corresponding prediction age value be calculated；
Obtain the age value corresponding with the face in the images to be recognized of the age identification model output.
Above-mentioned age recognition methods, device, computer equipment and storage medium include the to be identified of face by obtaining Image is then obtaining the output of age identification model with figure to be identified using images to be recognized as the input of age identification model The face corresponding age as in.Wherein, age identification model be acceptable age statistic error value as error metrics standard into Row training obtains, and age statistic error value is according to mark age value and corresponding to multiple training figures of same mark age value What the corresponding prediction age value of face as in be calculated, by using innovative age statistic error value work Age identification model is trained for error metrics standard, it is quasi- to the prediction of all age group to can be improved age identification model Exactness can access the relatively high knowledge of accuracy to the face in all age group under the uncontrollable scene of age bracket Other effect.
Detailed description of the invention
Fig. 1 is the applied environment figure of age recognition methods in one embodiment；
Fig. 2 is the training method flow chart of age identification model in one embodiment；
Fig. 3 is the schematic diagram in one embodiment before and after selective erasing；
Fig. 4 is the schematic diagram in one embodiment before and after Random-fuzzy；
Fig. 5 is the schematic diagram in one embodiment before and after super-resolution processing；
Fig. 6 is the schematic diagram of the facial image of different angle in one embodiment；
Fig. 7 is the flow chart of age recognition methods in one embodiment；
Fig. 8 is the schematic diagram that the face age is identified in one embodiment；
Fig. 9 A is the schematic diagram of the recognition of face in one embodiment under monitoring scene；
Fig. 9 B is the schematic diagram of multiple recognition of face effects in one embodiment under monitoring scene；
Figure 10 is the structural schematic diagram of age identification model in one embodiment；
Figure 11 is the architecture diagram of age identification model application in one embodiment；
Figure 12 is the schematic diagram of the age distribution counted in one embodiment；
Figure 13 is the flow chart of age recognition methods in another embodiment；
Figure 14 is the structural block diagram of the training device of age identification model in one embodiment；
Figure 15 is the structural block diagram of the training device of age identification model in another embodiment；
Figure 16 is the structural block diagram of age identification device in one embodiment；
Figure 17 is the structural block diagram of age identification device in another embodiment；
Figure 18 is the structural block diagram of computer equipment in one embodiment.
Specific embodiment
It is with reference to the accompanying drawings and embodiments, right in order to which the objects, technical solutions and advantages of the application are more clearly understood The application is further elaborated.It should be appreciated that specific embodiment described herein is only used to explain the application, and It is not used in restriction the application.
Fig. 1 is the applied environment figure of the training method of age identification model in one embodiment.Referring to Fig.1, which knows Other method is applied to the training system of age identification model.The training system of the age identification model includes terminal 110 and service Device 120.By network connection, terminal 110 specifically can be terminal console or mobile terminal for terminal 110 and server 120, mobile Terminal specifically can be at least one of mobile phone, tablet computer, laptop etc..Server 120 can use independent clothes The server cluster of business device either multiple servers composition is realized.Terminal 110 be used for will include face training image Collection uploads onto the server 120, server 120 be used to obtain include face training image collection, training image concentration it is each There are corresponding mark age values for face in training image, and the training image that training image is concentrated is as age identification model Input, the corresponding prediction age value of face in each training image of age identification model output is obtained, according to corresponding to The corresponding prediction age value of face in each target training image of same mark age value, is calculated and each mark year Age is worth the corresponding statistical forecast age, and age statistical error is calculated according to statistical forecast age and corresponding mark age value Value, is adjusted the parameter in age identification model according to age statistic error value, until meeting the condition of convergence, obtains target Age identification model.
In another embodiment, above-mentioned age recognition methods may be directly applied to terminal 110, and terminal 110 is for obtaining Take include face training image collection, training image concentrate each training image in face there are the corresponding mark ages Value, the training image that training image is concentrated obtain each of age identification model output as the input of age identification model The corresponding prediction age value of face in training image, according in each target training image corresponding to same mark age value The corresponding prediction age value of face, the statistical forecast age corresponding with each mark age value is calculated, it is pre- according to statistics It surveys the age and age statistic error value is calculated in corresponding mark age value, mould is identified to the age according to age statistic error value Parameter in type is adjusted, until meeting the condition of convergence, obtains target age identification model.
As shown in Fig. 2, in one embodiment, providing a kind of training method of age identification model, this method both may be used To be applied to terminal, server also can be applied to, the present embodiment is illustrated to be applied to terminal.The age identification model Training method specifically comprises the following steps:
Step S202, acquisition include the training image collection of face, the people in each training image that training image is concentrated There are corresponding mark age values for face.
Wherein, training image collection refers to the set for being trained to model and needing the image used.In order to be identified to the age Model is trained, and needs to be trained age identification model using the training image for including face.Training image is concentrated Training image in include face be all corresponding with corresponding age mark, that is, train label.
Step S204, the training image that training image is concentrated obtain age identification as the input of age identification model The corresponding prediction age value of face in each training image of model output.
Wherein, prediction age value is the age value that age identification model is predicted.By using training image as the age Then the input of identification model obtains the prediction age value corresponding with the face in training image of age identification model output, Convenient for it is subsequent according to prediction age value and it is corresponding mark age value between gap come to the parameter in age identification model into Row adjustment, so that age identification model is optimized towards more accurate direction.
Step S206, according to the corresponding prediction of face in each target training image for corresponding to same mark age value The statistical forecast age corresponding with each mark age value is calculated in age value.
Wherein, mark age value refers to the age value of desired age identification model output.Face in each training image All it is corresponding with corresponding mark age value.The training image for corresponding to same mark age value is gathered as one.Then it obtains Take the corresponding prediction age value of each training image in the set.In order to distinguish, same mark age value will be corresponded to Training image is known as " target training image ".
After getting the corresponding prediction age value of face in corresponding each target training image of same mark age, root The statistical forecast age corresponding with the mark age is calculated according to each target training image corresponding prediction age.At one In embodiment, the statistical forecast age is obtained according to the mean value calculation of the corresponding prediction age value of each target training image 's.In another embodiment, the statistical forecast age is the intermediate value according to the corresponding prediction age value of each target training image It is calculated.For example, it is assumed that the target training image for corresponding to same mark age value (for example, 25 years old) has 5, this The corresponding prediction age value of 5 target training images is respectively as follows: 19 years old, 24 years old, 26 years old, 28 years old, 32 years old.If using average value It calculates, then the corresponding statistical forecast age are as follows: (19+24+26+28+32)/5=25.8 years old.If using median calculation, i.e., The age in an intermediate position is taken, then the corresponding statistics age is 26 years old, if there are two in an intermediate position, taking two Numerical value is averaged.
Age statistic error value is calculated according to statistical forecast age and corresponding mark age value in step S208.
Wherein, age statistic error value was calculated according to statistical forecast age and corresponding mark age value, was For the standard of the error between measure statistical prediction age and corresponding mark age value.
Step S210 is adjusted the parameter in age identification model according to age statistic error value, receives until meeting Condition is held back, target age identification model is obtained.
Wherein, after age statistic error value being calculated, according to age statistic error value to the ginseng in age identification model Number is adjusted.Then age identification model adjusted is continued to train by repeating the above steps, restrains item until meeting Part obtains target age identification model.Target age identification model is trained age identification model.In one embodiment In, the condition of convergence is pre-set for measuring the threshold value of age statistic error value, until the age statistics being calculated is missed When difference is less than the threshold value of the setting, that is, illustrate that the training of age identification model is completed.
In one embodiment, age statistical error function can indicate are as follows: Wherein, Gt indicates the mark age, and Val indicates that prediction obtains prediction age, Mean (Val) refer to the statistical forecast age for corresponding to the mark age, the statistical forecast age is according to corresponding to the same mark age What multiple prediction ages were calculated.N indicates that current iteration round enters the number at the mark age of model, such as, it is assumed that when There are 60 different mark ages in preceding iteration round, then N=60.By using innovative age statistic error value conduct Error metrics standard is able to solve the unbalanced problem for causing prediction accuracy low of training sample age distribution, to improve year The accuracy of age identification.
The training method of above-mentioned age identification model, according to each target training image for corresponding to same mark age value In the corresponding prediction age value of face the statistical forecast age corresponding with each mark age value is calculated, then using system Age statistic error value is calculated in meter prediction age and corresponding mark age value, and then according to age statistic error value to year Parameter in age identification model is adjusted, until meeting the condition of convergence, obtains target age identification model.By using innovation The age statistic error value of property is trained age identification model, obtained target age identification model, even if in age bracket Under uncontrollable scene, the relatively high recognition effect of accuracy can be accessed to the face in all age group.
In one embodiment, acquisition include face training image collection after, further includes: to training image concentrate Training image carries out enhancing processing, obtains enhancing treated training image, enhancing processing include: selective erasing, Random-fuzzy, At least one of super-resolution processing；The training image that training image is concentrated is as the input of age identification model, comprising: Will enhancing treated training image as the input of age identification model.
Wherein, in order to improve the generalization of age identification model, the instruction that data enhancement method concentrates training image is introduced Practice image and carries out enhancing processing.Increase processing includes: at least one of selective erasing, Random-fuzzy, super-resolution processing.With Machine erasing, which refers to, at random wipes a certain position in training image, and reality still is able in the case where blocking convenient for subsequent Now to comprising face carry out the age and accurately identify.Random-fuzzy, which refers to, carries out Fuzzy Processing for training image, convenient for subsequent Still be able in blurred image situation to comprising face carry out the age and accurately identify.Super-resolution processing, which refers to, to instruct Practice image and carry out super-resolution processing and obtain clearly image relatively, is then trained as training image, i.e., as auxiliary To improve the generalization of age identification model.
The diversity of training image can be increased by enhancing processing.Instruction during training, before enhancing is handled Practicing image and enhancing, treated that training image is all used as the training sample of training pattern, knows to improve the age that training obtains The accuracy of other model.Enhance processing mode by introducing, it can be respectively for solving human face light condition is uncontrollable, face Under far and near fog-level is uncontrollable and the uncontrollable situation of face circumstance of occlusion, the accuracy of face age identification is improved.
In one embodiment, the training image concentrated to training image carries out enhancing processing, and obtaining enhancing, treated Training image, comprising: when enhancing processing includes selective erasing, erasing region is randomly selected from training image, by scratching area The pixel in domain carries out random assignment, obtains selective erasing treated training image；And/or when enhancing processing includes Random-fuzzy When, it randomly selects a direction and convolution algorithm is carried out to training image, obtain Random-fuzzy treated training image；And/or When enhancing processing includes super-resolution processing, super-resolution processing is carried out to training image by image super-resolution model, Training image after obtaining super-resolution processing.
Wherein, enhancing processing includes at least one of selective erasing, Random-fuzzy, super-resolution processing.Selective erasing It is that the pixel for wiping region is then subjected to random assignment, can be obtained by randomly selecting erasing region from training image Selective erasing treated training image.As shown in figure 3, carrying out the schematic diagram before and after selective erasing in one embodiment.With It is to carry out convolution algorithm to training image by randomly selecting a direction to obtain Random-fuzzy treated that training is schemed that machine is fuzzy Picture.As shown in figure 4, carrying out the schematic diagram before and after Random-fuzzy in one embodiment.Super-resolution processing is by trained The super-resolution model arrived carries out super-resolution processing to training image, relatively clear after super-resolution processing can be obtained Image.If Fig. 5 is to carry out the schematic diagram before and after super-resolution processing in one embodiment.
In one embodiment, include face training image concentrate include multiple angles facial image, it is multiple The facial image of angle belongs to multiple and different age brackets.
Wherein, due in some scenarios, for example, monitoring scene, since facial angle is uncontrollable, in order to in any The face of angle can carry out age identification.In the facial image that training image concentration includes multiple angles, such as Fig. 6 institute Show, is in one embodiment, training image concentrates the schematic diagram of the facial image in different angle.In addition, in order to right The face of all age group can access accurate identification, need to include that training for all age group is schemed in training sample Picture.The face age is accurately identified in order to be realized in the uncontrollable situation of facial angle, needs to wrap in training image Face containing multiple angles.Different age group is accurately identified in order to realize, also needs to include each in training image The training image of a age bracket.
In one embodiment, it according to the corresponding prediction age value of face in each target training image, obtains and marks The quasi- age value corresponding statistical forecast age, comprising: by the face corresponding prediction year in obtained each target training image Age value is averagely obtained the consensus forecast age, using the consensus forecast age as statistical forecast year corresponding with rated age value Age.
Wherein, the statistical forecast age uses the consensus forecast age, and the consensus forecast age, which refers to, will correspond to same standard year Each target training image of age value corresponding prediction age is averagely obtained.It can be solved by using the consensus forecast age Certainly the age is unevenly distributed the not high problem of caused prediction accuracy that weighs in training sample, improves age identification model to each The prediction accuracy of age bracket.
In one embodiment, age statistical error is calculated according to statistical forecast age and corresponding rated age value Value, comprising: obtain error transfer factor coefficient；It is calculated according to error transfer factor coefficient, statistical forecast age and corresponding rated age value Obtain age statistic error value.
Wherein, since people is when young, complexion with the age increase variation it is more obvious, and it is older when, complexion increases with the age Long variation is more unobvious, so people, when young, the error of prediction can be bigger.By obtaining error transfer factor coefficient, utilize Error transfer factor coefficient is adjusted the corresponding age statistic error value of different age group, so as to obtain for all age group To more accurate prediction effect.I.e. error transfer factor coefficient be used for according to the different ages to corresponding age statistic error value into Row adjustment.In one embodiment, error transfer factor coefficient and rated age value are at inverse correlation, i.e. rated age value is bigger, accordingly Error transfer factor coefficient it is smaller.It is predicted according to error transfer factor coefficient, statistics age and corresponding mark is calculated most jointly Whole age statistic error value.
In one embodiment, error transfer factor coefficient is obtained, comprising: according to statistical forecast age and corresponding rated age Error transfer factor control parameter is calculated in value；Error transfer factor coefficient, error transfer factor is calculated according to error transfer factor control parameter Coefficient and error transfer factor control coefrficient are at inverse correlation.
Wherein, error transfer factor control parameter is used to control the size of error transfer factor coefficient, i.e. control age statistic error value Adjustment amplitude.Error transfer factor control parameter is related to statistical forecast age and corresponding rated age value.Implement at one In example, error transfer factor control parameter is the maximum value in statistical forecast age and corresponding rated age value, i.e. error transfer factor control Parameter processed can be expressed as max (Mean (Val), Gt), and Gt indicates rated age value, and Mean (Val) refers to corresponding to standard year The statistical forecast age of age value.
After error transfer factor control parameter is calculated, error transfer factor function is obtained, by by error transfer factor control parameter It substitutes into error transfer factor function and error transfer factor value is calculated.Error transfer factor coefficient and error transfer factor control coefrficient are at inverse correlation.? In one embodiment, the relationship of error transfer factor coefficient and error transfer factor control parameter can be indicated using following formula: f (x)= 2k/ (x+k), x are error transfer factor control parameter, and k is the constant greater than x.
In one embodiment, the corresponding prediction age range of age prediction model is obtained, is obtained in prediction age range Maximum value, according to prediction age range maximum value and error transfer factor control parameter error transfer factor coefficient is calculated.
Wherein, prediction age range refers to the range of age of prediction, for example, the age that the prediction age models can be predicted Range is 0-70 years old, then the maximum value in the prediction age range is 70 years old.According to the maximum value and mistake of prediction age range Error transfer factor coefficient is calculated in difference adjustment control parameter.It can be indicated using following formula: f (x)=2K/ (x+K), wherein K For the maximum value for predicting age range, x is error transfer factor control parameter.
In one embodiment, age statistic error value is calculated by using the following formula: Wherein, f (x)=2K/ (max (Mean (Val), Gt)+K).Wherein, f (x) Indicate that error transfer factor coefficient, K indicate the maximum value of default age range, max (Mean (Val), Gt)) it is error transfer factor control ginseng Number, Gt indicate rated age value.
As shown in fig. 7, in one embodiment it is proposed that a kind of age recognition methods.This method both can be applied to end End, also can be applied to server, and the present embodiment is illustrated with being applied to terminal.The age recognition methods specifically includes as follows Step:
Step S702, acquisition include the images to be recognized of face.
Wherein, images to be recognized refers to the image at face age to be identified.Images to be recognized can be terminal and pass through calling What camera captured in real-time obtained includes the image of face, be also possible to obtain it is stored include face image. Face in images to be recognized can be at any angle, for example, can be side face, can also positive face etc..In images to be recognized It may include a face, also can wrap containing multiple faces.It is subsequent to need to identify in image if including multiple faces The each face corresponding age.
Step S704, using images to be recognized as the input of age identification model, age identification model acceptable age statistics For error amount as error metrics standard, age statistic error value is according to mark age value and corresponding to same mark age value What the corresponding prediction age value of face in multiple training images be calculated.
Wherein, the corresponding age value of face that age identification model includes in images to be recognized for identification.Age statistics Error amount is calculated according to the age statistical error function of setting, age statistical error function be with mark age value and The corresponding statistical forecast age is variable.Year is substituted by the mark age value that will acquire and corresponding statistical forecast age Corresponding age statistic error value can be calculated in age statistical error function.
Mark age value refers to the age value being labeled to the face in training image, it can be understood as face is corresponding Actual age, the i.e. age value of desired output.In one embodiment, mark age value is using view-based access control model subjective judgement The range estimation age.In the case where the actual age of the face in training image cannot be got (such as under monitoring scene), phase The mark age value answered is estimated by professional person, is then labeled.The statistical forecast age is to pass through The corresponding prediction age value of multiple training images that statistics corresponds to the same mark age obtains, prediction age value be pass through by Training image input age identification model obtains.
In one embodiment, the statistical forecast age can be obtained by the way of calculating intermediate value, for example, prediction is obtained The multiple prediction ages for corresponding to same mark age value be ranked up according to size, then choose in intermediate prediction year Age value is as the statistical forecast age.In another embodiment, the statistical forecast age is calculated by the way of mean value, The multiple prediction ages that will be predicted are averaged, using obtained mean value as the statistical forecast age.
For example, it is assumed that the corresponding mark age value of face in multiple training images is all 25 years old, respectively should Input of multiple training images as training image obtains the corresponding prediction age value of face in each training image, then The statistical forecast age is calculated according to the corresponding prediction age value of multiple training image.
Age statistic error value is calculated using statistical forecast age and mark age value, is then counted and is missed according to the age Difference is adjusted training to the parameter in model, can still be able to obtain in the unbalanced situation of sample age distribution The age identification model high to all age group recognition accuracy.Since under special screne, the age distribution of training sample is not Controllably, if the age distribution unbalanced prediction age for being easy to cause model is tended to be distributed bigger age range, and originally Age statistic error value is calculated by using statistical forecast age and mark age value in embodiment can be with effective solution The above-mentioned problem low due to the unbalanced caused prediction accuracy of age distribution.
Step S706 obtains the age value corresponding with the face in images to be recognized of age identification model output.
Wherein, by that will include input of the images to be recognized of face as the age identification model trained Obtain the prediction age value corresponding with the face in images to be recognized of output.As shown in figure 8, in one embodiment, identification The corresponding age schematic diagram of face out.
Above-mentioned age recognition methods, by obtain include face images to be recognized, using images to be recognized as having instructed The input of experienced age identification model, then obtain trained age identification model output with the face in images to be recognized Corresponding age value.Above-mentioned age identification model is that acceptable age statistic error value is trained to obtain as error metrics standard , age statistic error value is calculated according to mark age value and corresponding statistical forecast age, statistical forecast Age is the prediction age according to the corresponding age identification model output of the multiple training images for corresponding to same mark age value Be calculated.Mould is identified to the age as error metrics standard by using innovative age statistic error value Type is trained, and can be improved age identification model to the prediction accuracy of all age group, uncontrollable special in age bracket Under scene, the relatively high recognition effect of accuracy can be accessed to the face in all age group.
It as shown in Figure 9 A, is the schematic diagram being applied under monitoring scene in one embodiment.As shown in Figure 9 A, it obtains first Then video image under monitoring scene detects the face in video image, extract include face target Picture, for the ease of identification, extraction includes in the Target Photo of face only including a face.If deposited in video image In multiple faces, then correspondingly from video image extract in multiple Target Photos.Then it is identified Target Photo as the age Then the input of model obtains the age of age identification model output.It as shown in Figure 9 B, is in one embodiment, in monitoring field There are when multiple faces under scape, to the schematic diagram for the recognition result that the age of multiple faces is identified.
In one embodiment, age identification model is trained to obtain using convolutional neural networks model, the age Identification model includes: multiple convolutional layers, includes the active coating and pond layer of preset number between adjacent convolutional layer.
Wherein, age identification model be using convolutional neural networks model (Convolutional Neural Network, CNN) training obtains.Age identification model includes multiple convolutional layers, includes the active coating of preset number between adjacent convolutional layer With pond layer.Convolutional layer is used to carry out convolution algorithm to image to extract characteristics of image.Active coating is used to carry out image non-linear Operation, for describing the nonlinear characteristic of image.Active coating is realized using activation primitive.For example, ReLu function.Chi Hua Layer is used to project each weight in image to obtain the image after dimensionality reduction.
As shown in Figure 10, in one embodiment, age identification model includes 17 layers of convolutional layer and one layer of output layer, In the first convolutional layer be 7X7 convolutional layer, the convolutional layer of 2-17 is all the convolutional layer of 3X3.Wherein, 7X7,3X3 refer to convolution The size of the convolution kernel of layer.Between adjacent convolutional layer further include: ReLu layers (i.e. active coatings) and Pooling (pond) layer.Wherein, The Chi Huahe of the last layer is 5X5, the pond core 7X7 of other layers.In the present embodiment, the innovative pond layer to the last layer Internal structure modification for 5X5, can to a certain extent lift scheme ability to express while do not change pre-training substantially The parameters weighting of model, so as to avoid duplicate pre-training process.Wherein, pre-training process refers to each in initialization model The process of a parameter.In order to avoid duplicate pre-training process, by the last layer in original convolutional neural networks model The pond core of pond layer is revised as 5X5, the ability to express of model not only can be improved, but also can be to avoid duplicate pre-training process.
In one embodiment, before using images to be recognized as the input for the age identification model trained, comprising: right Face in images to be recognized is identified, target human face region is obtained；Know images to be recognized as the age trained The input of other model, comprising: using the corresponding target human face region of images to be recognized as the defeated of the age identification model trained Enter.
Wherein, target human face region refers to the region where face.By carrying out Face datection identification to images to be recognized, It determines that face is corresponding and cuts out frame, then target person face region is obtained according to cutting out frame and be cut out, later by target face Input of the region as the age identification model trained.Since the images to be recognized got may further include many useless Information first detects the face in images to be recognized before input model, then extracts target human face region, By using target human face region as the input for the age identification model trained, being conducive to improve the prediction of face age in this way Accuracy.
As shown in figure 11, in the embodiment of a monitoring scene, including front-end module (being present in terminal) and backstage mould Block (is present in server), wherein front-end module includes: video acquisition module and face detection module.Background module includes: people Face analysis module, statistical analysis module and database module.Specifically, it is adopted in real time by video acquisition module by camera first Collect the image in reality scene, then acquired image is sent to face detection module, Face datection by video acquisition module Module is used to carry out Face datection to acquired image, when detecting in image includes face, by the area where face Domain is cut out to obtain target facial image, the human face analysis module that obtained target facial image is uploaded onto the server, people Include has age identification model in face analysis module, the age is carried out to the face in target facial image according to age identification model Identification, obtains the corresponding age value of face.Then the age that analysis obtains is passed into statistical analysis module, statistical analysis module It is a reality as shown in figure 12 for being counted to obtain corresponding age distribution according to the age of the multiple faces got It applies in example, the schematic diagram of the age distribution counted.Database module is for storing face information and corresponding age, just In subsequent lookup.
As shown in figure 13, in one embodiment it is proposed that a kind of age recognition methods, this method comprises:
Step S1301, acquisition include the training image collection of face, the people in each training image that training image is concentrated There are corresponding mark age values for face.
Step S1302, the training image that training image is concentrated obtain age identification as the input of age identification model The corresponding prediction age value of face in each training image of model output.
Step S1303, it is corresponding pre- according to the face in each target training image for corresponding to same mark age value Age value is surveyed, the statistical forecast age corresponding with each mark age value is calculated.
Obtained each target training image corresponding prediction age is averagely obtained consensus forecast by step S1304 Age, using the consensus forecast age as the statistical forecast age corresponding with mark age value.
Age statistical error is calculated according to the statistical forecast age and corresponding mark age value in step S1305 Value.
Step S1306 is adjusted the parameter in the age identification model according to the age statistic error value, directly To the condition of convergence is met, target age identification model is obtained.
Step S1307, acquisition include the images to be recognized of face.
Step S1308, using images to be recognized as the input of age identification model；
Step S1309 obtains the year corresponding with the face in the images to be recognized of the age identification model output Age value.
As shown in figure 14, in one embodiment it is proposed that a kind of training device of age identification model, the device packet It includes:
Image set obtains module 1402, for obtains include face training image collection, training image concentration There are corresponding mark age values for face in each training image；
Training input/output module 1404, the training image for concentrating the training image is as age identification model Input, obtain the corresponding prediction age value of face in each training image of age identification model output；
First computing module 1406, for according to the people in each target training image for corresponding to same mark age value The statistical forecast age corresponding with each mark age value is calculated in the corresponding prediction age value of face；
Second computing module 1408, for year to be calculated according to the statistical forecast age and corresponding mark age value Age statistic error value；
Module 1410 is adjusted, for carrying out according to the age statistic error value to the parameter in the age identification model Adjustment obtains target age identification model until meeting the condition of convergence.
As shown in figure 15, in one embodiment, it is described acquisition include face training image collection after, further includes:
Enhance processing module 1403, the training image for concentrating to the training image carries out enhancing processing, increased Strong treated training image, the enhancing processing include: selective erasing, Random-fuzzy, at least one in super-resolution processing Kind；
The trained input/output module is also used to using enhancing treated the training image as age identification model Input.
In one embodiment, the enhancing processing module is also used to when it includes selective erasing that the enhancing, which is handled, from Erasing region is randomly selected in the training image, and the pixel in the erasing region is subjected to random assignment, obtains selective erasing Treated training image；And/or when enhancing processing includes Random-fuzzy, a direction is randomly selected to the training Image carries out convolution algorithm, obtains Random-fuzzy treated training image；And/or when enhancing processing includes super-resolution When processing, super-resolution processing is carried out to the training image by image super-resolution model, after obtaining super-resolution processing Training image.
In one embodiment, it is described include face training image concentrate include multiple angles facial image, The facial image of the multiple angle belongs to multiple and different age brackets.
In one embodiment, in each target training image that first computing module 1406 is also used to obtain The corresponding prediction age value of face is averagely obtained the consensus forecast age, using the consensus forecast age as with the mark The age value corresponding statistical forecast age.
In one embodiment, the second computing module 1408 is also used to obtain error transfer factor coefficient；According to the error tune Age statistic error value is calculated in integral coefficient, the statistical forecast age and corresponding mark age value.
In one embodiment, the second computing module 1408 is also used to according to the statistical forecast age and corresponding mark Error transfer factor control parameter is calculated in age value；The error transfer factor system is calculated according to the error transfer factor control parameter Number, the error transfer factor coefficient and the error transfer factor control coefrficient are at inverse correlation.
As shown in figure 16, in one embodiment it is proposed that a kind of age identification device, described device include:
Images to be recognized obtain module 1602, for obtain include face images to be recognized；
Input module 1604, for using the images to be recognized as the input of age identification model, the age identification For model acceptable age statistic error value as error metrics standard, the age statistic error value is according to mark age value and right In the corresponding prediction age value of face in multiple training images of same mark age value should be calculated；
Output module 1606, for obtaining age identification model output and the face pair in the images to be recognized The age value answered.
In one embodiment, the age identification model is trained to obtain using convolutional neural networks model, The age identification model includes: multiple convolutional layers, active coating and pond between the adjacent convolutional layer including preset number Change layer.
As shown in figure 17, in one embodiment, above-mentioned age identification device further include:
Identification module 1603 obtains target face for identifying to the face by the images to be recognized Region；
The input module is also used to using the corresponding target human face region of the images to be recognized as the age trained The input of identification model.
Figure 18 shows the internal structure chart of computer equipment in one embodiment.The computer equipment specifically can be end End, is also possible to server.As shown in figure 18, which includes processor, the memory connected by system bus And network interface.Wherein, memory includes non-volatile memory medium and built-in storage.The non-volatile of the computer equipment is deposited Storage media is stored with operating system, can also be stored with computer program, when which is executed by processor, may make place Manage the training method that device realizes age identification model.Computer program can also be stored in the built-in storage, the computer program When being executed by processor, processor may make to execute the training method of age identification model.It will be understood by those skilled in the art that Structure shown in Figure 18, only the block diagram of part-structure relevant to application scheme, is not constituted to application scheme The restriction for the computer equipment being applied thereon, specific computer equipment may include more more or fewer than as shown in the figure Component perhaps combines certain components or with different component layouts.
In one embodiment, the training method of age identification model provided by the present application can be implemented as a kind of computer The form of program, computer program can be run in computer equipment as shown in figure 18.It can in the memory of computer equipment Storage forms each program module of the training device of the age identification model, for example, the image set acquisition module 1402 of Figure 14, Training input/output module 1404, the first computing module 1406, the second computing module 1408 and adjustment module 1410.Each program The age that the computer program of module composition makes processor execute each embodiment of the application described in this specification identifies Step in device.For example, computer equipment shown in Figure 18 can pass through the image of age identification device as shown in figure 14 Collection obtains the training image collection that the acquisition of module 1402 includes face, the people in each training image that the training image is concentrated There are corresponding mark age values for face；Made by the training image that training input/output module 1404 concentrates the training image For the input of age identification model, the corresponding prediction of face in each training image of the age identification model output is obtained Age value；By the first computing module 1406 according to the face in each target training image for corresponding to same mark age value The statistical forecast age corresponding with each mark age value is calculated in corresponding prediction age value；Pass through the second computing module 1408 are calculated age statistic error value according to the statistical forecast age and corresponding mark age value；By adjusting module 1410 are adjusted the parameter in the age identification model according to the age statistic error value, until meeting convergence item Part obtains target age identification model.
In one embodiment it is proposed that a kind of computer equipment, including memory and processor, the memory storage There is computer program, when the computer program is executed by the processor, so that the processor executes following steps: obtaining It include the training image collection of face, there are corresponding mark years for the face in each training image that the training image is concentrated Age value；The training image that the training image is concentrated obtains the age identification model as the input of age identification model The corresponding prediction age value of face in each training image of output；According to each target for corresponding to same mark age value The corresponding prediction age value of face in training image, is calculated the statistical forecast age corresponding with each mark age value； Age statistic error value is calculated according to the statistical forecast age and corresponding mark age value；It is counted according to the age Error amount is adjusted the parameter in the age identification model, until meeting the condition of convergence, obtains target age identification mould Type.
In one embodiment, it is described acquisition include face training image collection after, the computer program also makes Obtain the processor and execute following steps: the training image concentrated to the training image carries out enhancing processing, obtains at enhancing Training image after reason, the enhancing processing includes: at least one of selective erasing, Random-fuzzy, super-resolution processing；Institute The training image for concentrating the training image is stated as the input of age identification model, comprising: treated by the enhancing Input of the training image as age identification model.
In one embodiment, the training image concentrated to the training image carries out enhancing processing, is enhanced Treated training image, comprising: when enhancing processing includes selective erasing, wiping is randomly selected from the training image Except region, the pixel in the erasing region is subjected to random assignment, obtains selective erasing treated training image；And/or work as When enhancing processing includes Random-fuzzy, randomly select a direction and convolution algorithm carried out to the training image, obtain with Training image after machine Fuzzy Processing；And/or when enhancing processing includes super-resolution processing, pass through image super-resolution Model carries out super-resolution processing to the training image, the training image after obtaining super-resolution processing.
In one embodiment, it is described include face training image concentrate include multiple angles facial image, The facial image of the multiple angle belongs to multiple and different age brackets.
In one embodiment, the corresponding prediction age value of face according in each target training image, Obtain the statistical forecast age corresponding with the mark age value, comprising: by the face in obtained each target training image Corresponding prediction age value is averagely obtained the consensus forecast age, using the consensus forecast age as with the mark age It is worth the corresponding statistical forecast age.
In one embodiment, described to be calculated according to the statistical forecast age with the corresponding mark age value Age statistic error value, comprising: obtain error transfer factor coefficient；According to the error transfer factor coefficient, the statistical forecast age and Age statistic error value is calculated in corresponding mark age value.
In one embodiment, the acquisition error transfer factor coefficient, comprising: according to statistical forecast age and corresponding Error transfer factor control parameter is calculated in mark age value；The error tune is calculated according to the error transfer factor control parameter Integral coefficient, the error transfer factor coefficient and the error transfer factor control coefrficient are at inverse correlation.
In one embodiment it is proposed that a kind of computer equipment, including memory and processor, the memory storage There is computer program, when the computer program is executed by the processor, so that the processor executes following steps: obtaining It include the images to be recognized of face；Using the images to be recognized as the input of age identification model, the age identifies mould For type acceptable age statistic error value as error metrics standard, the age statistic error value is according to mark age value and correspondence What the corresponding prediction age value of face in multiple training images of same mark age value be calculated；It obtains The age value corresponding with the face in the images to be recognized of the age identification model output.
In one embodiment, the age identification model is trained to obtain using convolutional neural networks model, The age identification model includes: multiple convolutional layers, active coating and pond between the adjacent convolutional layer including preset number Change layer.
In one embodiment, it is described using the images to be recognized as the input of age identification model before, the meter Calculation machine program also makes the processor execute following steps: the face by the images to be recognized identified, Obtain target human face region；It is described using the images to be recognized as the input for the age identification model trained, comprising: by institute State input of the corresponding target human face region of images to be recognized as the age identification model trained.
In one embodiment it is proposed that a kind of computer readable storage medium, is stored with computer program, the calculating When machine program is executed by processor, so that the processor executes following steps: acquisition includes the training image collection of face, institute State training image concentration each training image in face there are corresponding mark age values；The training image is concentrated Input of the training image as age identification model obtains the face in each training image of the age identification model output Corresponding prediction age value；According to the corresponding prediction of face in each target training image for corresponding to same mark age value The statistical forecast age corresponding with each mark age value is calculated in age value；According to the statistical forecast age and correspondence Mark age value age statistic error value is calculated；According to the age statistic error value in the age identification model Parameter be adjusted, until meet the condition of convergence, obtain target age identification model.
In one embodiment, it is described acquisition include face training image collection after, the computer program also makes Obtain the processor and execute following steps: the training image concentrated to the training image carries out enhancing processing, obtains at enhancing Training image after reason, the enhancing processing includes: at least one of selective erasing, Random-fuzzy, super-resolution processing；Institute The training image for concentrating the training image is stated as the input of age identification model, comprising: treated by the enhancing Input of the training image as age identification model.
In one embodiment, the training image concentrated to the training image carries out enhancing processing, is enhanced Treated training image, comprising: when enhancing processing includes selective erasing, wiping is randomly selected from the training image Except region, the pixel in the erasing region is subjected to random assignment, obtains selective erasing treated training image；And/or work as When enhancing processing includes Random-fuzzy, randomly select a direction and convolution algorithm carried out to the training image, obtain with Training image after machine Fuzzy Processing；And/or when enhancing processing includes super-resolution processing, pass through image super-resolution Model carries out super-resolution processing to the training image, the training image after obtaining super-resolution processing.
In one embodiment, it is described include face training image concentrate include multiple angles facial image, The facial image of the multiple angle belongs to multiple and different age brackets.
In one embodiment, the corresponding prediction age value of face according in each target training image, Obtain the statistical forecast age corresponding with the mark age value, comprising: by the face in obtained each target training image Corresponding prediction age value is averagely obtained the consensus forecast age, using the consensus forecast age as with the mark age It is worth the corresponding statistical forecast age.
In one embodiment, described to be calculated according to the statistical forecast age with the corresponding mark age value Age statistic error value, comprising: obtain error transfer factor coefficient；According to the error transfer factor coefficient, the statistical forecast age and Age statistic error value is calculated in corresponding mark age value.
In one embodiment, the acquisition error transfer factor coefficient, comprising: according to statistical forecast age and corresponding Error transfer factor control parameter is calculated in mark age value；The error tune is calculated according to the error transfer factor control parameter Integral coefficient, the error transfer factor coefficient and the error transfer factor control coefrficient are at inverse correlation.
In one embodiment it is proposed that a kind of computer readable storage medium, is stored with computer program, the calculating When machine program is executed by processor, so that the processor executes following steps: acquisition includes the images to be recognized of face；It will Input of the images to be recognized as age identification model, the age identification model acceptable age statistic error value is as mistake Difference metric standard, the age statistic error value are according to mark age value and corresponding to multiple training of same mark age value What the corresponding prediction age value of face in image be calculated；Obtain age identification model output with institute State the corresponding age value of face in images to be recognized.
In one embodiment, the age identification model is trained to obtain using convolutional neural networks model, The age identification model includes: multiple convolutional layers, active coating and pond between the adjacent convolutional layer including preset number Change layer.
In one embodiment, it is described using the images to be recognized as the input of age identification model before, the meter Calculation machine program also makes the processor execute following steps: the face by the images to be recognized identified, Obtain target human face region；It is described using the images to be recognized as the input for the age identification model trained, comprising: by institute State input of the corresponding target human face region of images to be recognized as the age identification model trained.
Those of ordinary skill in the art will appreciate that realizing all or part of the process in above-described embodiment method, being can be with Relevant hardware is instructed to complete by computer program, the program can be stored in a non-volatile computer and can be read In storage medium, the program is when being executed, it may include such as the process of the embodiment of above-mentioned each method.Wherein, provided herein Each embodiment used in any reference to memory, storage, database or other media, may each comprise non-volatile And/or volatile memory.Nonvolatile memory may include that read-only memory (ROM), programming ROM (PROM), electricity can be compiled Journey ROM (EPROM), electrically erasable ROM (EEPROM) or flash memory.Volatile memory may include random access memory (RAM) or external cache.By way of illustration and not limitation, RAM is available in many forms, such as static state RAM (SRAM), dynamic ram (DRAM), synchronous dram (SDRAM), double data rate sdram (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronization link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) directly RAM (RDRAM), straight Connect memory bus dynamic ram (DRDRAM) and memory bus dynamic ram (RDRAM) etc..
Each technical characteristic of above embodiments can be combined arbitrarily, for simplicity of description, not to above-described embodiment In each technical characteristic it is all possible combination be all described, as long as however, the combination of these technical characteristics be not present lance Shield all should be considered as described in this specification.
The several embodiments of the application above described embodiment only expresses, the description thereof is more specific and detailed, but simultaneously The limitation to the application the scope of the patents therefore cannot be interpreted as.It should be pointed out that for those of ordinary skill in the art For, without departing from the concept of this application, various modifications and improvements can be made, these belong to the guarantor of the application Protect range.Therefore, the scope of protection shall be subject to the appended claims for the application patent.

Claims (15)

1. a kind of training method of age identification model, which comprises
Acquisition includes the training image collection of face, and the face in each training image that the training image is concentrated, which exists, to be corresponded to Mark age value；
It is defeated to obtain the age identification model as the input of age identification model for the training image that the training image is concentrated The corresponding prediction age value of the face in each training image out；
According to the corresponding prediction age value of face in each target training image for corresponding to same mark age value, calculate To the statistical forecast age corresponding with each mark age value；
Age statistic error value is calculated according to the statistical forecast age and corresponding mark age value；
The parameter in the age identification model is adjusted according to the age statistic error value, until meeting convergence item Part obtains target age identification model.
2. the method according to claim 1, wherein it is described acquisition include face training image collection after, Further include:
The training image concentrated to the training image carries out enhancing processing, obtains enhancing treated training image, the increasing Strength reason includes: at least one of selective erasing, Random-fuzzy, super-resolution processing；
The training image that the training image is concentrated is as the input of age identification model, comprising:
Using enhancing treated the training image as the input of age identification model.
3. according to the method described in claim 2, it is characterized in that, the training image concentrated to the training image carries out Enhancing processing obtains enhancing treated training image, comprising:
When enhancing processing includes selective erasing, erasing region is randomly selected from the training image, by the erasing The pixel in region carries out random assignment, obtains selective erasing treated training image；And/or
When enhancing processing includes Random-fuzzy, randomly selects a direction and convolution algorithm is carried out to the training image, Obtain Random-fuzzy treated training image；And/or
When enhancing processing includes super-resolution processing, the training image is surpassed by image super-resolution model Resolution processes, the training image after obtaining super-resolution processing.
4. the method according to claim 1, wherein it is described include face training image concentration include more The facial image of a angle, the facial image of the multiple angle belong to multiple and different age brackets.
5. the method according to claim 1, wherein the face according in each target training image Corresponding prediction age value obtains the statistical forecast age corresponding with the mark age value, comprising:
The corresponding prediction age value of face in obtained each target training image is averagely obtained into the consensus forecast age, Using the consensus forecast age as the statistical forecast age corresponding with the mark age value.
6. the method according to claim 1, wherein described according to the statistical forecast age and corresponding described Age statistic error value is calculated in mark age value, comprising:
Obtain error transfer factor coefficient；
Age statistics is calculated according to the error transfer factor coefficient, the statistical forecast age and corresponding mark age value to miss Difference.
7. according to the method described in claim 6, it is characterized in that, the acquisition error transfer factor coefficient, comprising:
Error transfer factor control parameter is calculated according to the statistical forecast age and corresponding mark age value；
The error transfer factor coefficient, the error transfer factor coefficient and the mistake is calculated according to the error transfer factor control parameter Difference adjustment control coefrficient is at inverse correlation.
8. a kind of age recognition methods, which comprises
Acquisition includes the images to be recognized of face；
Using the images to be recognized as the input of age identification model, the age identification model acceptable age statistic error value As error metrics standard, the age statistic error value is according to mark age value and corresponding to the more of same mark age value What the corresponding prediction age value of face in a training image be calculated；
Obtain the age value corresponding with the face in the images to be recognized of the age identification model output.
9. the method according to claim 1, wherein the age identification model is using convolutional neural networks mould What type was trained, the age identification model includes: multiple convolutional layers, includes default between the adjacent convolutional layer The active coating and pond layer of number.
10. the method according to claim 1, wherein described identify mould for the images to be recognized as the age Before the input of type, comprising:
The face by the images to be recognized is identified, target human face region is obtained；
It is described using the images to be recognized as the input for the age identification model trained, comprising:
Using the corresponding target human face region of the images to be recognized as the input for the age identification model trained.
11. a kind of training device of age identification model, described device include:
Image set obtains module, for obtains include face training image collection, each training of training image concentration There are corresponding mark age values for face in image；
Training input/output module, training image for concentrating the training image as the input of age identification model, Obtain the corresponding prediction age value of face in each training image of the age identification model output；
First computing module, for corresponding according to the face in each target training image for corresponding to same mark age value It predicts age value, the statistical forecast age corresponding with each mark age value is calculated；
Second computing module is missed for age statistics to be calculated according to the statistical forecast age and corresponding mark age value Difference；
Module is adjusted, for being adjusted according to the age statistic error value to the parameter in the age identification model, directly To the condition of convergence is met, target age identification model is obtained.
12. device according to claim 11, which is characterized in that described device further include:
Enhance processing module, the training image for concentrating to the training image carries out enhancing processing, after obtaining enhancing processing Training image, enhancing processing includes: at least one of selective erasing, Random-fuzzy, super-resolution processing；
The input/output module is also used to using enhancing treated the training image as the input of age identification model.
13. a kind of age identification device, described device include:
Images to be recognized obtain module, for obtain include face images to be recognized；
Input module, for using the images to be recognized as the input of age identification model, the age identification model to be used For age statistic error value as error metrics standard, the age statistic error value is according to mark age value and to correspond to same Mark what the corresponding prediction age value of the face in multiple training images of age value be calculated；
Output module, for obtaining that the age identification model trained exports and the face pair in the images to be recognized The age value answered.
14. a kind of computer readable storage medium is stored with computer program, when the computer program is executed by processor, So that the processor is executed such as the step of any one of claims 1 to 10 the method.
15. a kind of computer equipment, including memory and processor, the memory is stored with computer program, the calculating When machine program is executed by the processor, so that the processor is executed such as any one of claims 1 to 10 the method Step.
CN201810864567.1A 2018-08-01 2018-08-01 Training method of age identification model, age identification method and related equipment Active CN109034078B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810864567.1A CN109034078B (en) 2018-08-01 2018-08-01 Training method of age identification model, age identification method and related equipment

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810864567.1A CN109034078B (en) 2018-08-01 2018-08-01 Training method of age identification model, age identification method and related equipment

Publications (2)

Publication Number Publication Date
CN109034078A true CN109034078A (en) 2018-12-18
CN109034078B CN109034078B (en) 2023-07-14

Family

ID=64648377

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810864567.1A Active CN109034078B (en) 2018-08-01 2018-08-01 Training method of age identification model, age identification method and related equipment

Country Status (1)

Country Link
CN (1) CN109034078B (en)

Cited By (19)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110033332A (en) * 2019-04-23 2019-07-19 杭州智趣智能信息技术有限公司 A kind of face identification method, system and electronic equipment and storage medium
CN110046941A (en) * 2019-04-23 2019-07-23 杭州智趣智能信息技术有限公司 A kind of face identification method, system and electronic equipment and storage medium
CN110287942A (en) * 2019-07-03 2019-09-27 成都旷视金智科技有限公司 Training method, age estimation method and the corresponding device of age estimation model
CN110334663A (en) * 2019-07-09 2019-10-15 北京字节跳动网络技术有限公司 Age recognition methods and device, storage medium and terminal based on image
CN110378306A (en) * 2019-07-25 2019-10-25 厦门美图之家科技有限公司 Age prediction technique, device and image processing equipment
CN110390290A (en) * 2019-07-17 2019-10-29 北京天正聚合科技有限公司 A kind of method, apparatus, electronic equipment and storage medium identifying target object
CN110674397A (en) * 2019-08-30 2020-01-10 北京百度网讯科技有限公司 Method, device, equipment and readable medium for training age point prediction model
CN110689066A (en) * 2019-09-24 2020-01-14 成都考拉悠然科技有限公司 Training method combining face recognition data equalization and enhancement
CN110826469A (en) * 2019-11-01 2020-02-21 Oppo广东移动通信有限公司 Person detection method and device and computer readable storage medium
CN111079624A (en) * 2019-12-11 2020-04-28 北京金山云网络技术有限公司 Method, device, electronic equipment and medium for collecting sample information
CN111459587A (en) * 2020-03-27 2020-07-28 北京三快在线科技有限公司 Information display method, device, equipment and storage medium
CN111695415A (en) * 2020-04-28 2020-09-22 平安科技（深圳）有限公司 Construction method and identification method of image identification model and related equipment
CN111783623A (en) * 2020-06-29 2020-10-16 北京百度网讯科技有限公司 Algorithm adjustment method, apparatus, device, and medium for recognizing positioning element
CN111860333A (en) * 2020-07-22 2020-10-30 安徽兰臣信息科技有限公司 Child face recognition algorithm for cross-age recognition
CN112149601A (en) * 2020-09-30 2020-12-29 北京澎思科技有限公司 Occlusion-compatible face attribute identification method and device and electronic equipment
CN112329693A (en) * 2020-11-17 2021-02-05 汇纳科技股份有限公司 Training method, recognition method, medium and device for gender and age recognition model
CN112668455A (en) * 2020-12-24 2021-04-16 平安科技（深圳）有限公司 Face age identification method and device, terminal equipment and storage medium
CN112966653A (en) * 2021-03-29 2021-06-15 深圳市优必选科技股份有限公司 Line patrol model training method, line patrol method and line patrol system
CN113705535A (en) * 2021-09-18 2021-11-26 中国银行股份有限公司 ATM transaction method for helping old people, related device and storage medium

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2010125915A1 (en) * 2009-04-28 2010-11-04 Ｎｅｃソフト株式会社 Age estimation device, method, and program
CN105678253A (en) * 2016-01-04 2016-06-15 东南大学 Semi-supervised age estimation device based on faces and semi-supervised age estimation method based on faces
CN106228139A (en) * 2016-07-27 2016-12-14 东南大学 A kind of apparent age prediction algorithm based on convolutional network and system thereof
CN107545249A (en) * 2017-08-30 2018-01-05 国信优易数据有限公司 A kind of population ages' recognition methods and device
WO2018054283A1 (en) * 2016-09-23 2018-03-29 北京眼神科技有限公司 Face model training method and device, and face authentication method and device
CN108052862A (en) * 2017-11-09 2018-05-18 北京达佳互联信息技术有限公司 Age predictor method and device
CN108108668A (en) * 2017-12-01 2018-06-01 北京小米移动软件有限公司 Age Forecasting Methodology and device based on image

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2010125915A1 (en) * 2009-04-28 2010-11-04 Ｎｅｃソフト株式会社 Age estimation device, method, and program
CN105678253A (en) * 2016-01-04 2016-06-15 东南大学 Semi-supervised age estimation device based on faces and semi-supervised age estimation method based on faces
CN106228139A (en) * 2016-07-27 2016-12-14 东南大学 A kind of apparent age prediction algorithm based on convolutional network and system thereof
WO2018054283A1 (en) * 2016-09-23 2018-03-29 北京眼神科技有限公司 Face model training method and device, and face authentication method and device
CN107545249A (en) * 2017-08-30 2018-01-05 国信优易数据有限公司 A kind of population ages' recognition methods and device
CN108052862A (en) * 2017-11-09 2018-05-18 北京达佳互联信息技术有限公司 Age predictor method and device
CN108108668A (en) * 2017-12-01 2018-06-01 北京小米移动软件有限公司 Age Forecasting Methodology and device based on image

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
许梦婷;康长青;刘伟杰;张万顺;: "人脸图像年龄估计系统的设计与实现", 信息通信, no. 01 *
赵一丁 等: "基于分类与回归混合模型的人脸年龄估计方法", 《计算机应用》 *
赵一丁 等: "基于分类与回归混合模型的人脸年龄估计方法", 《计算机应用》, 31 July 2017 (2017-07-31), pages 1999 - 2002 *

Cited By (28)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110033332A (en) * 2019-04-23 2019-07-19 杭州智趣智能信息技术有限公司 A kind of face identification method, system and electronic equipment and storage medium
CN110046941A (en) * 2019-04-23 2019-07-23 杭州智趣智能信息技术有限公司 A kind of face identification method, system and electronic equipment and storage medium
CN110287942A (en) * 2019-07-03 2019-09-27 成都旷视金智科技有限公司 Training method, age estimation method and the corresponding device of age estimation model
CN110287942B (en) * 2019-07-03 2021-09-17 成都旷视金智科技有限公司 Training method of age estimation model, age estimation method and corresponding device
CN110334663A (en) * 2019-07-09 2019-10-15 北京字节跳动网络技术有限公司 Age recognition methods and device, storage medium and terminal based on image
CN110390290A (en) * 2019-07-17 2019-10-29 北京天正聚合科技有限公司 A kind of method, apparatus, electronic equipment and storage medium identifying target object
CN110378306A (en) * 2019-07-25 2019-10-25 厦门美图之家科技有限公司 Age prediction technique, device and image processing equipment
CN110378306B (en) * 2019-07-25 2021-11-02 厦门美图之家科技有限公司 Age prediction method and device and image processing equipment
CN110674397A (en) * 2019-08-30 2020-01-10 北京百度网讯科技有限公司 Method, device, equipment and readable medium for training age point prediction model
CN110689066A (en) * 2019-09-24 2020-01-14 成都考拉悠然科技有限公司 Training method combining face recognition data equalization and enhancement
CN110689066B (en) * 2019-09-24 2020-06-16 成都考拉悠然科技有限公司 Training method combining face recognition data equalization and enhancement
CN110826469A (en) * 2019-11-01 2020-02-21 Oppo广东移动通信有限公司 Person detection method and device and computer readable storage medium
CN111079624B (en) * 2019-12-11 2023-09-01 北京金山云网络技术有限公司 Sample information acquisition method and device, electronic equipment and medium
CN111079624A (en) * 2019-12-11 2020-04-28 北京金山云网络技术有限公司 Method, device, electronic equipment and medium for collecting sample information
CN111459587A (en) * 2020-03-27 2020-07-28 北京三快在线科技有限公司 Information display method, device, equipment and storage medium
CN111695415A (en) * 2020-04-28 2020-09-22 平安科技（深圳）有限公司 Construction method and identification method of image identification model and related equipment
CN111695415B (en) * 2020-04-28 2024-04-12 平安科技（深圳）有限公司 Image recognition method and related equipment
CN111783623A (en) * 2020-06-29 2020-10-16 北京百度网讯科技有限公司 Algorithm adjustment method, apparatus, device, and medium for recognizing positioning element
CN111783623B (en) * 2020-06-29 2024-04-12 北京百度网讯科技有限公司 Algorithm adjustment method, device, equipment and medium for identifying positioning element
CN111860333A (en) * 2020-07-22 2020-10-30 安徽兰臣信息科技有限公司 Child face recognition algorithm for cross-age recognition
CN112149601A (en) * 2020-09-30 2020-12-29 北京澎思科技有限公司 Occlusion-compatible face attribute identification method and device and electronic equipment
CN112329693B (en) * 2020-11-17 2024-01-19 汇纳科技股份有限公司 Training method, identification method, medium and equipment for gender and age identification model
CN112329693A (en) * 2020-11-17 2021-02-05 汇纳科技股份有限公司 Training method, recognition method, medium and device for gender and age recognition model
CN112668455A (en) * 2020-12-24 2021-04-16 平安科技（深圳）有限公司 Face age identification method and device, terminal equipment and storage medium
CN112668455B (en) * 2020-12-24 2023-11-28 平安科技（深圳）有限公司 Face age identification method and device, terminal equipment and storage medium
CN112966653A (en) * 2021-03-29 2021-06-15 深圳市优必选科技股份有限公司 Line patrol model training method, line patrol method and line patrol system
CN112966653B (en) * 2021-03-29 2023-12-19 深圳市优必选科技股份有限公司 Line inspection model training method, line inspection method and line inspection system
CN113705535A (en) * 2021-09-18 2021-11-26 中国银行股份有限公司 ATM transaction method for helping old people, related device and storage medium

Also Published As

Publication number Publication date
CN109034078B (en) 2023-07-14

Similar Documents

Publication Publication Date Title
CN109034078A (en) Training method, age recognition methods and the relevant device of age identification model
CN110517278B (en) Image segmentation and training method and device of image segmentation network and computer equipment
US9521391B2 (en) Settings of a digital camera for depth map refinement
CN106875422B (en) Face tracking method and device
KR102442844B1 (en) Method for Distinguishing a Real Three-Dimensional Object from a Two-Dimensional Spoof of the Real Object
CN107451965B (en) Distort face image correcting method, device, computer equipment and storage medium
Bare et al. An accurate deep convolutional neural networks model for no-reference image quality assessment
CN108960290A (en) Image processing method, device, computer readable storage medium and electronic equipment
CN111199523B (en) Power equipment identification method, device, computer equipment and storage medium
CN107832802A (en) Quality of human face image evaluation method and device based on face alignment
CN109299658B (en) Face detection method, face image rendering device and storage medium
CN109492577A (en) A kind of gesture identification method, device and electronic equipment
Ou et al. A novel rank learning based no-reference image quality assessment method
CN108764371A (en) Image processing method, device, computer readable storage medium and electronic equipment
CN110059700A (en) The recognition methods of image moire fringes, device, computer equipment and storage medium
CN114332744B (en) Transformer substation self-adaptive security method and system based on machine vision
CN110992401A (en) Target tracking method and device, computer equipment and storage medium
CN111915541B (en) Image enhancement processing method, device, equipment and medium based on artificial intelligence
CN114257738A (en) Automatic exposure method, device, equipment and storage medium
CN113179421A (en) Video cover selection method and device, computer equipment and storage medium
Wang et al. Efficient video deblurring guided by motion magnitude
CN110532837A (en) Image processing method and household appliance in a kind of article fetching process
CN114092515B (en) Target tracking detection method, device, equipment and medium for obstacle shielding
US11182475B2 (en) Detecting fraudulent facial recognition
CN110392211A (en) Image processing method and device, electronic equipment, computer readable storage medium

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant