WO2019000466A1 - 人脸识别方法、装置、存储介质及电子设备 - Google Patents

人脸识别方法、装置、存储介质及电子设备 Download PDF

Info

Publication number
WO2019000466A1
WO2019000466A1 PCT/CN2017/091378 CN2017091378W WO2019000466A1 WO 2019000466 A1 WO2019000466 A1 WO 2019000466A1 CN 2017091378 W CN2017091378 W CN 2017091378W WO 2019000466 A1 WO2019000466 A1 WO 2019000466A1
Authority
WO
WIPO (PCT)
Prior art keywords
age
face data
data
face
derived
Prior art date
Application number
PCT/CN2017/091378
Other languages
English (en)
French (fr)
Inventor
梁昆
Original Assignee
广东欧珀移动通信有限公司
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 广东欧珀移动通信有限公司 filed Critical 广东欧珀移动通信有限公司
Priority to PCT/CN2017/091378 priority Critical patent/WO2019000466A1/zh
Priority to CN201780092003.4A priority patent/CN110741387B/zh
Priority to EP17915349.9A priority patent/EP3648008A4/en
Publication of WO2019000466A1 publication Critical patent/WO2019000466A1/zh
Priority to US16/730,571 priority patent/US11380131B2/en

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16Human faces, e.g. facial parts, sketches or expressions
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168Feature extraction; Face representation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16Human faces, e.g. facial parts, sketches or expressions
    • G06V40/168Feature extraction; Face representation
    • G06V40/171Local features and components; Facial parts ; Occluding parts, e.g. glasses; Geometrical relationships
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/10Human or animal bodies, e.g. vehicle occupants or pedestrians; Body parts, e.g. hands
    • G06V40/16Human faces, e.g. facial parts, sketches or expressions
    • G06V40/178Human faces, e.g. facial parts, sketches or expressions estimating age from face image; using age information for improving recognition
    • YGENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02DCLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D10/00Energy efficient computing, e.g. low power processors, power management or thermal management

Definitions

  • the present invention relates to the field of data processing technologies, and in particular, to a face recognition method, apparatus, storage medium, and electronic device.
  • Today's electronic devices are very convenient for taking photos or image information, and people are exposed to a large number of photos or other image information during the use of electronic devices. Sometimes people may be interested in the age of some people in the photo when viewing photos or other image information. In a normal scene, people will think roughly according to the appearance of the target characters in the image, and rely on their own experience to judge the age of these people. However, it is obviously not easy to judge the actual age of the target person in the image by virtue of experience alone.
  • an embodiment of the present invention provides the following technical solution: a method for recognizing a face, including the following steps:
  • the age data corresponding to the face data to be tested is obtained according to the age distribution interval.
  • the embodiment of the present invention further provides the following technical solutions:
  • a face recognition device wherein the face recognition device comprises:
  • a first generating unit configured to generate, according to the measured face data, a first derived face data set related to the face data to be tested, where the first derived face data set includes multiple different Derivative face data;
  • a first discriminating unit configured to perform age discriminant on the derived face data in the first derivative face data set, and generate an age distribution interval corresponding to the first derived face data set
  • a first determining unit configured to determine whether the age distribution interval matches the first reference age range
  • a second obtaining unit configured to obtain age data corresponding to the face data to be tested according to the age distribution interval.
  • the embodiment of the present invention further provides the following technical solutions:
  • a storage medium wherein the storage medium stores a plurality of instructions adapted to be loaded by a processor to perform a face recognition method as described above.
  • the embodiment of the present invention further provides the following technical solutions:
  • An electronic device comprising a processor, a memory, the processor being electrically connected to the memory, the memory for storing instructions and data, the processor for performing face recognition as described above method.
  • the embodiment of the invention provides a face recognition method, device, storage medium and electronic device, which can improve the efficiency and accuracy of face recognition.
  • FIG. 1 is a schematic flowchart of an implementation process of a face recognition method according to an embodiment of the present invention.
  • FIG. 2 is a flowchart of another implementation of a face recognition method according to an embodiment of the present invention.
  • FIG. 4 is a schematic flowchart of an implementation process of a training face data generation model according to an embodiment of the present invention.
  • FIG. 5 is a schematic diagram of an implementation process of obtaining an age discrimination model according to an embodiment of the present invention.
  • FIG. 6 is a schematic structural diagram of a face recognition apparatus according to an embodiment of the present invention.
  • FIG. 7 is another schematic structural diagram of a face recognition apparatus according to an embodiment of the present invention.
  • FIG. 9 is a schematic structural diagram of a first discriminating unit according to an embodiment of the present invention.
  • FIG. 10 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.
  • references to "an embodiment” herein mean that a particular feature, structure, or characteristic described in connection with the embodiments can be included in at least one embodiment of the invention.
  • the appearances of the phrases in various places in the specification are not necessarily referring to the same embodiments, and are not exclusive or alternative embodiments that are mutually exclusive. Those skilled in the art will understand and implicitly understand that the embodiments described herein can be combined with other embodiments.
  • the embodiment of the invention provides a face recognition method, which comprises the following steps:
  • the age data corresponding to the face data to be tested is obtained according to the age distribution interval.
  • the method before the generating the face data to generate the first derived face data set related to the face data to be tested, the method further includes:
  • the derived face data in the first derived face data set is separately determined by age, and an age distribution interval corresponding to the first derived face data set is generated, including:
  • the age discriminant model is configured to discriminate age data corresponding to the feature data based on feature data extracted from the derived face data.
  • the age discriminant model includes a plurality of feature data discriminant models
  • the feature data discriminant model is configured to discriminate a feature data of a preset type, and obtain age data corresponding to the feature data of the preset category according to the feature data.
  • the derived face data in the first derived face data set is subjected to age determination by a preset age discriminant model, and an age corresponding to the first derived face data set is generated. Before the distribution interval, it also includes:
  • reference face data wherein the reference face data includes age data
  • the feature data is associated with age data corresponding to the feature data to generate the age discriminant model.
  • the feature data is obtained from facial data using a deep learning and convolutional neural network.
  • the age discriminant model is formed by a plurality of feature data discriminant models being associated by a preset weight coefficient.
  • obtaining age data corresponding to the face data to be tested according to the age distribution interval including:
  • the age data corresponding to the face data to be tested is obtained according to the age distribution interval.
  • the plurality of derived face data are continuously distributed according to the age corresponding to the age, to form the first derivative face data set.
  • obtaining age data corresponding to the face data to be tested according to the age distribution interval including:
  • the age median data is used as the age data corresponding to the face data to be tested.
  • FIG. 1 a flow of implementation of a face recognition method according to an embodiment of the present invention is shown.
  • the drawings only show contents related to the present invention.
  • the face recognition method is applied to an electronic device that can be used for performing data processing, and specifically includes the following steps:
  • step S101 the face data to be tested is acquired.
  • the face data may be in the form of a data set, image, etc. associated therewith.
  • the face data to be tested can be manually confirmed, manually selected, etc., and the face data to be tested can also be obtained by a preset program selection and extraction. .
  • an image carrying a face in a phone image set is selected so that the image is confirmed as the face data to be tested.
  • the face detection algorithm determines the position of each face, and then extracts the face data based on the face position.
  • the face data will be exemplified in the form of an image.
  • a first derived face data set related to the face data to be tested is generated according to the face data to be tested, wherein the first derived face data set includes a plurality of different derivatives. Face data.
  • the first derived face data set is generated based on the face data to be tested, and includes a plurality of different derived face data corresponding to the face data to be tested.
  • the first derivative face data set may be generated by using a preset face data by using a preset generator, wherein the generator may generate face data based on the feature data of the face. It can be understood that if the face data generated by the generator is used, the generated face features will be consistent with, or approximate to, the face data to be tested, or the feature changes may be performed as expected. See the following examples for specific implementations. It is assumed that the first derivative face data adopts the feature data-based generation method.
  • the system extracts the feature data of the contour, the facial features, the skin, the expression, the hairstyle and the like in the face data to be tested, and then according to some people summarized from the face database.
  • Face feature association rules such as association rules between faces of different ages or regions, generate derivative face data corresponding to the above feature data.
  • the system extracts the feature data about the facial features from the face data to be tested as “five features A”, and in the process of generating the first derivative face data set,
  • the characteristic data "Five A” is based on the characteristic correlation law between different ages, and changes according to a certain age value - assuming that the aging is 2 years old, the characteristic data about the facial features is generated as "five features A1", and the "five features A1" Features are added to the generated derived face data. It can be understood that if the model is generated, that is, the generator is trusted, when the newly generated derivative face data is judged based on the feature data, the result should be the same as the result of the original face data. The difference is about 2 years old.
  • the number and distribution of the derived face data in the generated first derivative face data set can be set according to actual needs.
  • the number of derived face data is set to 10, wherein the derived face data is based on the facial feature association rule of different ages in the feature data, and between each adjacent two derived face data It will be generated on the basis of aging or rejuvenation of 0.5 years old. For example, five derived face data of aged 0.5 years, 1 year old, 1.5 years old, 2 years old, and 2.5 years old, respectively, and 0.5 years old and 1 year old, respectively.
  • step S103 the derived face data in the first derived face data set is respectively subjected to age discrimination, and an age distribution interval corresponding to the first derived face data set is generated.
  • the plurality of derived face data in the first derived face data set is subjected to age determination by a preset algorithm. For example, if there are 10 derived face data in the first derived face data set, In the process of age discrimination, the ages of 10 derived face data are separately determined, and 10 age data are obtained.
  • the 10 age data are formed into an age distribution interval. For example, when a derived face data obtains the largest age data result in the first derived face data set, it is 30 years old, and another derived data obtains the smallest age data result in the first derived face data set, At the age of 25, based on the age data of the two ends, it can be known that the age distribution interval corresponding to the first derivative face data set is 25-30 years old.
  • this step can effectively expand the derivative face data for detection related to the face data to be tested, and improve the fault tolerance rate of the face age recognition based on the generated derivative face data, thereby improving the accuracy of the discrimination.
  • step S104 it is determined whether the age distribution interval matches the first reference age range.
  • the reference age interval may be an age distribution interval obtained by age-determining the face data to be tested on a generator obtained based on the preset face database.
  • the face database can be an AFAD database, an AFLW database, or some other commercial face database to provide sufficient reference face data so that the confidence of the reference information is sufficiently high.
  • the age distribution interval generated by the age determination of the first derivative face data set is matched with the first reference age interval, and the median, average, interval endpoint, etc. of the two intervals may be acquired.
  • the data is compared, the difference is obtained, and it is judged whether the difference is less than a preset threshold, and if so, the two match, and if not, the two do not match.
  • the specific judgment condition may be determined according to specific needs, which is not limited by the embodiment of the present invention.
  • the age range of the first derivative face data set generated by age is 25-29
  • the first reference age range obtained by the face to be tested based on the face database is 27-31, in order to judge them.
  • the degree of matching between the two can be calculated by using the median of the generated age distribution interval and the median value of the first reference age interval.
  • the allowable difference be ⁇ 3, then 29-27 ⁇ 3, it can be considered that the age distribution interval corresponding to the first derived face data set matches the first reference age range obtained based on the face database, and then, from the age distribution interval The median value is selected as the age data corresponding to the face data to be tested. If the allowable difference is ⁇ 1, it can be considered that the age distribution interval corresponding to the first derivative face data set does not match the first reference age interval obtained based on the face database, and at this time, the interval of the age distribution interval is adjusted. Position such that the age distribution interval eventually matches the first reference age range.
  • the value of the age distribution interval may be adjusted, such as increasing or decreasing a range of values, and matching is performed again, and the median of the age distribution interval is output when the matching condition is satisfied.
  • step S105 if yes, the age data corresponding to the face data to be tested is obtained according to the age distribution interval.
  • the age distribution interval is based on facial feature association rules of different ages in the feature data, and each adjacent two derived face data is aging or rejuvenated by a certain age value The method is generated. If the age distribution interval matches the first reference age interval, the median or average value between the two age endpoints (maximum and minimum values) of the generated age distribution interval may be taken as The age data of the face data is measured. Of course, if the age distribution interval is generated by other means, the age data can also be obtained at other locations in the age distribution interval.
  • the age distribution interval generated by the first derivative face data set is 25-29
  • the derived face data in the first derivative face data set is the face data to be tested based on different ages.
  • the facial features are related to each other and are generated after the age of the face data is aged or rejuvenated, for example, five derived face data of 0.5 years old, 1 year old, 1.5 years old, 2 years old, and 2.5 years old, respectively.
  • the five face data of 0.5 years old, 1 year old, 1.5 years old, 2 years old, and 2.5 years old are respectively rejuvenated.
  • the number of aging and juvenile derived face data is equivalent, a total of 10, understandable, the face data to be tested
  • the face data to be tested is acquired; then, according to the face data to be tested, a first derivative face data set related to the face data to be tested is generated, wherein the first derivative face
  • the data set includes a plurality of different derived face data to expand the face data sample related to the face data to be tested by using the first derived face data set; and then, the derivative of the first derivative face data set
  • the face data is respectively subjected to age discrimination, and an age distribution interval corresponding to the first derivative face data set is generated.
  • the age distribution interval is obtained by performing age determination based on the first derived face data, and the value of a certain point can be broadened.
  • the age data corresponding to the face data to be measured is described.
  • the above face recognition method can obtain an age distribution interval obtained according to the face data to be tested, and use the annual distribution interval to perform age discrimination, which can effectively solve the problem in the process of discriminating the age of the face data.
  • the problem of poor recognition accuracy caused by poor image angle, overexposure or too dark, improves the fault tolerance of the algorithm in the age discrimination process, so that the face age recognition algorithm can dynamically adapt to many different environments and greatly improve the face.
  • FIG. 2 is another implementation flow of a face recognition method according to an embodiment of the present invention. For convenience of description, only parts related to the content of the present invention are shown.
  • step S201 the face data to be tested is acquired.
  • the face data may be in the form of a data set, image, etc. associated therewith.
  • the face data to be tested can be manually confirmed, manually selected, etc., and the face data to be tested can also be obtained by a preset program selection and extraction. .
  • an image carrying a face in a phone image set is selected so that the image is confirmed as the face data to be tested.
  • the face detection algorithm determines the position of each face, and then extracts the face data based on the face position.
  • the face data to be tested generates a model by using preset face data, that is, a first derivative face data set related to the face data to be tested is generated by a preset generator, where The first derived face data set includes a plurality of different derived face data.
  • step S202 the face data to be tested is subjected to face aging processing and face rejuvenation processing according to preset different age groups, and a plurality of derivatives representing different ages related to the face data to be tested are generated.
  • Face data is subjected to face aging processing and face rejuvenation processing according to preset different age groups, and a plurality of derivatives representing different ages related to the face data to be tested are generated.
  • step S203 the plurality of derived face data are continuously distributed according to the age of the corresponding age to form the first derived face data set; wherein the first derived face data set includes Multiple different derived face data.
  • the first derivative face data adopts the feature data-based generation method. Firstly, the system extracts the feature data of the contour, the facial features, the skin, the expression, the hairstyle and the like in the face data to be tested, and then according to some people summarized from the face database. Face feature association rules, such as association rules between faces of different ages or regions, generate derivative face data corresponding to the above feature data.
  • the system extracts the feature data about the facial features from the face data to be tested as “five features A”, and in the process of generating the first derivative face data set,
  • the characteristic data "Five A” is based on the characteristic correlation law between different ages, and changes according to a certain age value - assuming that the aging is 2 years old, the characteristic data about the facial features is generated as "five features A1", and the "five features A1" Features are added to the generated derived face data. It can be understood that if the model is generated, that is, the generator is trusted, when the newly generated derivative face data is judged based on the feature data, the result should be the same as the result of the original face data. The difference is about 2 years old.
  • the number and distribution of the derived face data in the generated first derivative face data set can be set according to actual needs.
  • the number of derived face data is set to 10, wherein the derived face data is based on the facial feature association rule of different ages in the feature data, and between each adjacent two derived face data It will be generated on the basis of aging or rejuvenation of 0.5 years old. For example, five derived face data of aged 0.5 years, 1 year old, 1.5 years old, 2 years old, and 2.5 years old, respectively, and 0.5 years old and 1 year old, respectively.
  • the generator in order to improve the confidence level of the first derived face data set generated by the face data generation model, ie, the generator, the generator may be trained to introduce the confrontation network (GAN) to achieve the above purpose.
  • GAN confrontation network
  • step 204 the plurality of derived face data are respectively subjected to age discrimination by a preset age discriminant model, and an age distribution interval corresponding to the first derived face data set is generated; wherein the age discriminant model And for determining age data corresponding to the feature data based on the feature data extracted from the derived face data.
  • a plurality of feature data discriminant models may be included, wherein the feature data discriminant model is used to discriminate a feature data of a preset type, and obtain the preset according to the feature data.
  • the age data corresponding to the feature data of the category It can be understood that each feature data discriminant model will give an age discriminating result based on the corresponding feature data.
  • FIG. 3 is a block diagram showing the implementation of discriminating the derived face data.
  • the feature data about the facial features is extracted from the derived face data, and
  • the system will let “discriminate model A” determine the age by default, and obtain an age data “age A” corresponding to “five features”.
  • the feature data such as the contour, the skin, the expression, the hairstyle, and the like can be discriminated, and the flow is similar to the facial rule discrimination process, and the age data are respectively obtained.
  • “age B” "age C”
  • age D "age E”.
  • the age data discriminated by all the feature data discriminant models in the age discriminant model is obtained, the data is aggregated and calculated to obtain the overall age data on the face data.
  • step S205 it is determined whether the age distribution interval obtained by each of the derived face data is determined by age, and whether the correlation between the first reference age interval and the first reference age interval is higher than a preset threshold.
  • the degree of correlation is judged by the degree of correlation.
  • the degree of correlation is judged by the degree of correlation.
  • the ratio of the coincidence of the two intervals to the entire range is used as the correlation.
  • the ratio of the overlapping portion of the two intervals to the entire interval range is greater than a predetermined threshold, the two can be considered to match, and of course, the median value of the generated age distribution interval and the first reference age interval may also be utilized. The median value is calculated by combining the other factors to determine the correlation between the two. When the correlation is greater than a predetermined threshold, the two can be considered to match. It can be understood that the calculation of the correlation and the preset threshold size can be appropriately adjusted according to different algorithms, and the final purpose is to define the matching degree of the distribution intervals of the two by a numerical value.
  • the correlation is higher than a preset threshold, obtaining age data corresponding to the face data to be tested according to the age distribution interval;
  • step S206 as an embodiment, when the correlation is higher than the preset threshold, the median age data of the age distribution interval is obtained according to the age distribution interval.
  • step S207 the age median data is used as the age data corresponding to the face data to be tested.
  • the age distribution interval is based on facial feature association rules of different ages in the feature data, and each adjacent two derived face data is aging or rejuvenated by a certain age value The method is generated. If the age distribution interval matches the first reference age interval, the median value between the two age endpoints (maximum value and minimum value) of the generated age distribution interval may be taken as the face to be tested. Age data for the data.
  • the age distribution interval generated by the first derivative face data set is 25-29
  • the derived face data in the first derivative face data set is the face data to be tested based on different ages.
  • the facial feature association law is generated after the aging or rejuvenation of the face data to be tested, if the aging derivative face data is the same as the aging face data, and the two adjacent ages are derived. If the age difference of the face data is consistent, it can be considered that the median of the distribution interval should match the actual age.
  • step S208 if the correlation is lower than a preset threshold, the interval position of the age distribution interval is adjusted such that the age distribution interval finally matches the first reference age interval.
  • the interval position of the age distribution interval is adjusted so that the age distribution interval finally matches the first reference age interval.
  • the value of the age distribution interval may be adjusted, such as increasing or decreasing a range of values, and matching is performed again, and the median of the age distribution interval is output when the matching condition is satisfied.
  • the age distribution interval represents the age data corresponding to the plurality of derived face data
  • the representation of the age distribution interval can effectively improve the fault tolerance rate in the age discrimination process, and can obtain and wait from the age distribution interval.
  • the accurate age data corresponding to the face data is measured, and the accuracy of the age determination of the face data to be measured is improved.
  • FIG. 4 is a schematic flowchart of an implementation process of a training face data generation model according to an embodiment of the present invention. For convenience of description, only parts related to the content of the present invention are shown.
  • Figure 4 provides a method for training a face data generation model, ie, a generator, based on a generated confrontation network (GAN), wherein:
  • Step S301 obtaining reference face data.
  • Step S302 Generate, according to the reference face data, a second derived face data set related to the reference face data, wherein the second derived face data set includes a plurality of different derived face data.
  • Step S303 Perform age determination on the derived face data in the second derived face data set to generate an age distribution interval corresponding to the second derived face data set.
  • Step S304 determining whether the age distribution interval corresponding to the second derived face data set matches the second reference age range.
  • Step S305 if not, updating the model parameter corresponding to the face data generation model until the age distribution interval corresponding to the second derivative face data set is determined by the age and the second reference age interval match.
  • the reference face data may be obtained from an AFAD database, an AFLW database, or some other commercial face database, and the face database provides sufficient reference face data so that the reference information has sufficient confidence. high.
  • determining whether the age distribution interval corresponding to the second derived face data set matches the second reference age range the concept of the accuracy rate may be referenced, and the multiple loop matching is completed during the training process of the generator. After the success (including successful or unsuccessful matching), the accuracy will be counted.
  • the accuracy is high enough, the generator can be considered to be sufficiently reliable.
  • the matching accuracy rate of a certain unit time can be set to be more than 98%.
  • the generator may be inferred, that is, the generator is based on the person
  • the model parameter of the face data generation derivative face data process has insufficient confidence.
  • the parameter adjustment may be performed based on the second reference age interval, so that the regenerated generated face data can be matched with the second reference age range, and the pair is different.
  • the reference face data is continuously performed such as acquiring reference face data - generating multiple derived face data - not matching - updating parameters - generating and matching the generated confrontation network (GAN) training process to continuously generate a confrontation against the generator.
  • Network (GAN) generates training.
  • the number of training samples can be effectively improved, thereby improving the training efficiency of the generator in the training process, obtaining a better training effect, and enabling the generator, ie, the generator to reach Higher confidence.
  • the figure shows an implementation flow for obtaining an age discriminant model.
  • the following steps may be implemented:
  • Step S501 acquiring reference face data, wherein the reference face data includes age data
  • Step S502 extracting feature data corresponding to the reference face data according to the reference face data
  • Step S503 correlating the feature data with age data corresponding to the feature data to generate the age discriminant model.
  • the device when extracting feature data, may be acquired from the face data using a deep learning and convolutional neural network to improve the accuracy of the obtained feature data.
  • the age discriminant model is formed by a plurality of feature data discriminant models being associated by a preset weight coefficient.
  • the discriminative accuracy of each feature data by the age discriminant model and the influence of each feature data on the overall discriminant of the face type are different.
  • the weight adjustment can be performed according to the above differences to make the age discriminant.
  • the age data available to the model is more accurate.
  • a plurality of obtained age data corresponding to different derivative face data are formed into an age distribution interval. For example, when a derived face data obtains the largest age data result in the first derived face data set, it is 30 years old, and another derived data obtains the smallest age data result in the first derived face data set, At the age of 25, based on the age data of the two ends, it can be known that the age distribution interval corresponding to the first derivative face data set is 25-30 years old.
  • this step can effectively expand the derivative face data for detection related to the face data to be tested, and improve the fault tolerance rate of the face age recognition based on the generated derivative face data, thereby improving the accuracy of the discrimination.
  • a generated confrontation network to acquire a face data generation model, that is, a generator, specifically based on reference face data
  • a generator to acquire a second derived face data set
  • the accuracy of the second derived face data set is discriminated with reference to the age range to optimize the generator to make the generator passable.
  • the face data generation model using the face data generation model, the first derivative face data set corresponding to the face data to be tested is generated by the generator; and then the first derivative face data is concentrated by the trained age discriminant model.
  • a plurality of derived face data are used to determine the age of the feature data; finally, the age data is aggregated to generate an age distribution interval, and an age data with a high degree of matching with the face data to be tested is obtained.
  • This can effectively solve the problem of low recognition accuracy due to poor image angle, overexposure or over darkness in the process of discriminating the age of face data, and improve the fault tolerance rate of the algorithm in the process of age discrimination.
  • the face age recognition algorithm can dynamically adapt to a variety of different environments, greatly improving the accuracy of face age recognition.
  • the embodiment of the invention further provides a face recognition device, wherein the face recognition device comprises:
  • a first acquiring unit configured to acquire face data to be tested
  • a first generating unit configured to generate, according to the measured face data, a first derived face data set related to the face data to be tested, where the first derived face data set includes multiple different Derivative face data;
  • a first discriminating unit configured to perform age discriminant on the derived face data in the first derivative face data set, and generate an age distribution interval corresponding to the first derived face data set
  • a first determining unit configured to determine whether the age distribution interval matches the first reference age range
  • a second obtaining unit configured to obtain age data corresponding to the face data to be tested according to the age distribution interval.
  • the face recognition device comprises:
  • a first acquiring unit configured to acquire face data to be tested
  • a first generating unit configured to generate, according to the measured face data, a first derived face data set related to the face data to be tested, where the first derived face data set includes multiple different Derivative face data;
  • a first discriminating unit configured to perform age discriminant on the derived face data in the first derivative face data set, and generate an age distribution interval corresponding to the first derived face data set
  • a first determining unit configured to determine whether the age distribution interval matches the first reference age range
  • a second obtaining unit configured to obtain age data corresponding to the face data to be tested according to the age distribution interval.
  • the first generating unit comprises:
  • a first generating sub-unit configured to generate, according to the face data to be tested, a first derived face data set related to the face data to be tested by using a preset face data generation model.
  • the face recognition device further includes:
  • a third obtaining unit configured to acquire reference face data
  • a second generating unit configured to generate, according to the reference face data, a second derived face data set related to the reference face data, wherein the second derived face data set includes a plurality of different derivatives Face data
  • a second discriminating unit configured to perform age discriminant on the derived face data in the second derived face data set, and generate an age distribution interval corresponding to the second derived face data set
  • a second determining unit configured to determine whether an age distribution interval corresponding to the second derived face data set matches a second reference age range
  • an update unit configured to update the model parameter corresponding to the face data generation model until the age distribution interval corresponding to the second derivative face data set is determined by the age and the second reference Age range matching.
  • the first determining unit comprises:
  • a first discriminating sub-unit configured to perform age discriminating by using the preset age discriminant models, and generate an age distribution interval corresponding to the first derivative face data set;
  • the age discriminant model is configured to discriminate age data corresponding to the feature data based on feature data extracted from the derived face data.
  • the age discriminant model includes a plurality of feature data discriminant models
  • the feature data discriminant model is configured to discriminate a feature data of a preset type, and obtain age data corresponding to the feature data of the preset category according to the feature data.
  • the first determining unit further includes:
  • a first obtaining subunit configured to acquire reference face data, wherein the reference face data includes age data
  • Extracting a subunit configured to extract feature data corresponding to the reference face data according to the reference face data
  • the first generation subunit is configured to associate the feature data with age data corresponding to the feature data to generate the age discriminant model.
  • the feature data is obtained from facial data using a deep learning and convolutional neural network.
  • the age discriminant model is formed by a plurality of feature data discriminant models being associated by a preset weight coefficient.
  • the first determining unit includes:
  • a first determining subunit configured to determine whether an age distribution interval obtained by each of the derived face data is determined by age, and whether a correlation between the first reference age interval and the first reference age interval is higher than a preset threshold
  • the second obtaining unit includes:
  • a second obtaining sub-unit configured to obtain, according to the age distribution interval, age data corresponding to the face data to be tested, if the correlation is higher than a preset threshold
  • the first generating unit includes:
  • a second generating sub-unit configured to perform face aging processing and face rejuvenation processing on the face data to be tested according to a preset different age group, and generate different representatives related to the face data to be tested Derived face data of age;
  • a third generation sub-unit configured to continuously distribute the plurality of derivative face data according to an age corresponding to the age, to form the first derivative face data set.
  • the second obtaining unit includes:
  • a third obtaining sub-unit configured to acquire, according to the age-distributed interval, the age corresponding to the first derivative face data set of the continuous distribution process according to the result of the age determination Median data;
  • the confirmation subunit is configured to use the age median data as the age data corresponding to the face data to be tested.
  • the apparatus includes a first obtaining unit 601, a first generating unit 602, a first determining unit 603, a first determining unit 604, and a second.
  • the face recognition device specifically includes:
  • the first obtaining unit 601 is configured to acquire the face data to be tested.
  • the face data may be in the form of a data set, image, etc. associated therewith.
  • the face data to be tested can be manually confirmed, manually selected, etc., and the face data to be tested can also be obtained by a preset program selection and extraction. .
  • an image carrying a face in a phone image set is selected so that the image is confirmed as the face data to be tested.
  • the face detection algorithm determines the position of each face, and then extracts the face data based on the face position.
  • the face data will be exemplified in the form of an image.
  • a first generating unit 602 configured to generate, according to the face data to be tested, a first derived face data set related to the face data to be tested, where the first derived face data set includes multiple Different derived face data.
  • the first derived face data set is generated based on the face data to be tested, and includes a plurality of different derived face data corresponding to the face data to be tested.
  • the generation method based on the feature data may be adopted, or the generated face data may be generated by using the generation method based on other methods. It can be understood that if the derivative face data generated by the feature data generation method is used, the feature will be consistent with the face data to be tested, or similar, or the feature change may be performed as expected.
  • the first derivative face data adopts the feature data-based generation method. Firstly, the system extracts the feature data of the contour, the facial features, the skin, the expression, the hairstyle and the like in the face data to be tested, and then according to some people summarized from the face database. Face feature association rules, such as association rules between faces of different ages or regions, generate derivative face data corresponding to the above feature data.
  • the system extracts the feature data about the facial features from the face data to be tested as “five features A”, and in the process of generating the first derivative face data set,
  • the characteristic data "Five A” is based on the characteristic correlation law between different ages, and changes according to a certain age value - assuming that the aging is 2 years old, the characteristic data about the facial features is generated as "five features A1", and the "five features A1" Features are added to the generated derived face data. It can be understood that if the generated model is credible, the result obtained when the newly generated derivative face data is judged based on the feature data is also about 2 years old from the result of the original face data to be tested. .
  • the number and distribution of the derived face data in the generated first derivative face data set can be set according to actual needs.
  • the number of derived face data is set to 10, wherein the derived face data is based on the facial feature association rule of different ages in the feature data, and between each adjacent two derived face data It will be generated on the basis of aging or rejuvenation of 0.5 years old. For example, five derived face data of aged 0.5 years, 1 year old, 1.5 years old, 2 years old, and 2.5 years old, respectively, and 0.5 years old and 1 year old, respectively.
  • the first determining unit 603 is configured to perform age determination on the derived face data in the first derived face data set, and generate an age distribution interval corresponding to the first derived face data set.
  • the plurality of derived face data in the first derived face data set is subjected to age determination by a preset algorithm. For example, if there are 10 derived face data in the first derived face data set, In the process of age discrimination, the ages of 10 derived face data are separately determined, and 10 age data are obtained.
  • the 10 age data are formed into an age distribution interval. For example, when a derived face data obtains the largest age data result in the first derived face data set, it is 30 years old, and another derived data obtains the smallest age data result in the first derived face data set, At the age of 25, based on the age data of the two ends, it can be known that the age distribution interval corresponding to the first derivative face data set is 25-30 years old.
  • this step can effectively expand the derivative face data for detection related to the face data to be tested, and improve the fault tolerance rate of the face age recognition based on the generated derivative face data, thereby improving the accuracy of the discrimination.
  • the first determining unit 604 is configured to determine whether the age distribution interval matches the first reference age interval.
  • the first reference age interval may be obtained by using the face data to be tested based on a preset face database to obtain a corresponding age distribution interval.
  • the face database can be an AFAD database, an AFLW database, or some other commercial face database to provide sufficient reference face data so that the confidence of the reference information is sufficiently high.
  • the age distribution interval generated by the age of the first derivative face data set is matched with the above age distribution interval, and the median, average value, interval endpoint, etc. of the two age distribution intervals may be acquired.
  • the data is compared, the difference is obtained, and it is judged whether the difference is less than a preset threshold, and if so, the two match, and if not, the two do not match.
  • the specific judgment condition may be determined according to specific needs, which is not limited by the embodiment of the present invention.
  • the age range of the first derivative face data set generated by age is 25-29
  • the first reference age range obtained by the face to be tested based on the face database is 27-31, in order to judge them.
  • the degree of matching between the two can be calculated by using the median of the generated age distribution interval and the median value of the first reference age interval.
  • the allowable difference be ⁇ 3, then 29-27 ⁇ 3, it can be considered that the age distribution interval corresponding to the first derived face data set matches the first reference age range obtained based on the face database, and then, from the age distribution interval The median value is selected as the age data corresponding to the face data to be tested. If the allowable difference is ⁇ 1, it may be considered that the age distribution interval corresponding to the first derived face data set does not match the first reference age interval obtained based on the face database, and in some embodiments, The value of the age distribution interval is increased or decreased by a certain range, and matching is performed again, and the median of the age distribution interval is output when the matching condition is satisfied.
  • the situation of highlight or low light during the shooting process can be effectively avoided. Under the problem of large deviation in age judgment, the accuracy of face age recognition is improved.
  • the second obtaining unit 605 is configured to obtain, according to the age distribution interval, age data corresponding to the face data to be tested.
  • the age distribution interval is based on facial feature association rules of different ages in the feature data, and each adjacent two derived face data is aging or rejuvenated by a certain age value The method is generated. If the age distribution interval matches the first reference age interval, the median or average value between the two age endpoints (maximum and minimum values) of the generated age distribution interval may be taken as The age data of the face data is measured. Of course, if the age distribution interval is generated by other means, the age data can also be obtained at other locations in the age distribution interval.
  • the age distribution interval generated by the first derivative face data set is 25-29
  • the derived face data in the first derivative face data set is the face data to be tested based on different ages.
  • the facial features are related to each other and are generated after the age of the face data is aged or rejuvenated, for example, five derived face data of 0.5 years old, 1 year old, 1.5 years old, 2 years old, and 2.5 years old, respectively.
  • the five face data of 0.5 years old, 1 year old, 1.5 years old, 2 years old, and 2.5 years old are respectively rejuvenated.
  • the number of aging and juvenile derived face data is equivalent, a total of 10, understandable, the face data to be tested
  • the first acquiring unit 601 acquires the face data to be tested; then, the first generating unit 602 generates a first derivative face data set related to the face data to be tested according to the face data to be tested.
  • the first derivative face data set includes a plurality of different derived face data to expand a face data sample related to the face data to be tested by using the first derived face data set;
  • the determining unit 603 performs age discrimination on the derived face data in the first derived face data set to generate an age distribution interval corresponding to the first derived face data set, and the age distribution interval is based on the first derivative
  • the face data is obtained by age discrimination, and the value of a certain point can be broadened to a value of a certain distribution interval to improve the error rate of the age discrimination.
  • the first determining unit 604 determines whether the age distribution interval is the first reference. The age range is matched; finally, if the matching is successful, the second obtaining unit 605 obtains the age corresponding to the face data to be tested according to the age distribution interval. It is.
  • the above face recognition method can obtain an age distribution interval obtained according to the face data to be tested, and use the annual distribution interval to perform age discrimination, which can effectively solve the problem in the process of discriminating the age of the face data.
  • the problem of poor recognition accuracy caused by poor image angle, overexposure or too dark improves the fault tolerance of the algorithm in the age discrimination process, so that the face age recognition algorithm can dynamically adapt to many different environments and greatly improve the face.
  • FIG. 7 is another structure of a face recognition apparatus according to an embodiment of the present invention, which includes:
  • the first obtaining unit 601 is configured to acquire the face data to be tested.
  • the face data may be in the form of a data set, image, etc. associated therewith.
  • the face data to be tested can be manually confirmed, manually selected, etc., and the face data to be tested can also be obtained by a preset program selection and extraction. .
  • an image carrying a face in a phone image set is selected so that the image is confirmed as the face data to be tested.
  • the face detection algorithm determines the position of each face, and then extracts the face data based on the face position.
  • the face data to be tested generates a first derivative face data set related to the face data to be tested by using a preset face data generation model, wherein the first derivative face is The data set includes a number of different derived face data.
  • the first generation unit 602 includes a second generation subunit 6021 and a third generation subunit 6022, wherein:
  • a second generating sub-unit 6021 configured to perform face aging processing and face rejuvenation processing on the face data to be tested according to a preset different age group, and generate multiple representatives related to the face data to be tested. Derived face data of different ages;
  • a third generation sub-unit 6022 configured to continuously distribute the plurality of derivative face data according to an age corresponding to the age to form the first derivative face data set; wherein the first derivative person
  • the face data set includes a plurality of different derived face data.
  • the first derived face data set is generated based on the face data to be tested, and includes a plurality of different derived face data corresponding to the face data to be tested.
  • the generation method based on the feature data may be adopted, or the generated face data may be generated by using the generation method based on other methods. It can be understood that if the derivative face data generated by the feature data generation method is used, the feature will be consistent with the face data to be tested, or similar, or the feature change may be performed as expected.
  • the first derivative face data adopts the feature data-based generation method. Firstly, the system extracts the feature data of the contour, the facial features, the skin, the expression, the hairstyle and the like in the face data to be tested, and then according to some people summarized from the face database. Face feature association rules, such as association rules between faces of different ages or regions, generate derivative face data corresponding to the above feature data.
  • the system extracts the feature data about the facial features from the face data to be tested as “five features A”, and in the process of generating the first derivative face data set,
  • the characteristic data "Five A” is based on the characteristic correlation law between different ages, and changes according to a certain age value - assuming that the aging is 2 years old, the characteristic data about the facial features is generated as "five features A1", and the "five features A1" Features are added to the generated derived face data. It can be understood that if the generated model is credible, the result obtained when the newly generated derivative face data is judged based on the feature data is also about 2 years old from the result of the original face data to be tested. .
  • the number and distribution of the derived face data in the generated first derivative face data set can be set according to actual needs.
  • the number of derived face data is set to 10, wherein the derived face data is based on the facial feature association rule of different ages in the feature data, and between each adjacent two derived face data It will be generated on the basis of aging or rejuvenation of 0.5 years old. For example, five derived face data of aged 0.5 years, 1 year old, 1.5 years old, 2 years old, and 2.5 years old, respectively, and 0.5 years old and 1 year old, respectively.
  • the generator in order to improve the confidence level of the first derived face data set generated by the face data generation model, ie, the generator, the generator may be trained to introduce the confrontation network (GAN) to achieve the above purpose.
  • GAN confrontation network
  • the first discriminating unit 603 includes a first discriminating sub-unit 6031, and the first discriminating sub-unit 6031 is configured to separately perform age determination by using the preset age discriminant data, and generate the first And an age distribution interval corresponding to the derived face data set; wherein the age discrimination model is configured to determine age data corresponding to the feature data according to the feature data extracted from the derived face data.
  • a plurality of feature data discriminant models may be included, wherein the feature data discriminant model is used to discriminate a feature data of a preset type, and obtain the preset according to the feature data.
  • the age data corresponding to the feature data of the category It can be understood that each feature data discriminant model will give an age discriminating result based on the corresponding feature data.
  • the feature data about the facial features is extracted from a derivative face data, and can be defined as “five features”.
  • the feature data is subjected to the age-determined feature data discrimination model "Difference Model A".
  • the system will let “discriminate model A” determine the age by default, and obtain an age data “age A” corresponding to “five features”.
  • the age data obtained by discriminating the feature data types such as the outline, the skin, the expression, and the hairstyle can be obtained.
  • the age data discriminated by all the feature data discriminant models in the age discriminant model is obtained, the data is aggregated and calculated to obtain the overall age data on the face data.
  • the first determining unit 604 includes:
  • a first determining sub-unit 6041 configured to determine whether a correlation between an age distribution interval obtained by age determination of each of the derived face data and a first reference age interval is higher than a pre-determination Set the threshold.
  • the degree of correlation is judged by the degree of correlation.
  • the degree of correlation is judged by the degree of correlation.
  • the ratio of the coincidence of the two intervals to the entire range is used as the correlation.
  • the ratio of the overlapping portion of the two intervals to the entire interval range is greater than a predetermined threshold, the two can be considered to match, and of course, the median value of the generated age distribution interval and the first reference age interval may also be utilized. The median value is calculated by combining the other factors to determine the correlation between the two. When the correlation is greater than a predetermined threshold, the two can be considered to match. It can be understood that the calculation of the correlation and the preset threshold size can be appropriately adjusted according to different algorithms, and the final purpose is to define the matching degree of the distribution intervals of the two by a numerical value.
  • the second obtaining unit 605 includes a second obtaining subunit 6051 and an adjusting subunit 6052, wherein:
  • the second obtaining sub-unit 6051 is configured to obtain, according to the age distribution interval, age data corresponding to the face data to be tested, if the correlation is higher than a preset threshold.
  • the adjusting sub-unit 6052 is configured to adjust an interval position of the age distribution interval if the correlation degree is lower than a preset threshold, so that the age distribution interval finally matches the first reference age interval.
  • the second obtaining unit 605 includes a third obtaining subunit 6053 and a confirming unit 6054, wherein:
  • a third obtaining sub-unit 6053 configured to acquire, according to the result of the age determination according to the age distribution interval, the first derivative face data set corresponding to the continuous distribution process of the age distribution interval Age median data;
  • the confirmation subunit 6054 is configured to use the age median data as the age data corresponding to the face data to be tested.
  • the age distribution interval is based on facial feature association rules of different ages in the feature data, and each adjacent two derived face data is aging or rejuvenated by a certain age value The method is generated. If the age distribution interval matches the first reference age interval, the median value between the two age endpoints (maximum value and minimum value) of the generated age distribution interval may be taken as the face to be tested. Age data for the data.
  • the age distribution interval generated by the first derivative face data set is 25-29
  • the derived face data in the first derivative face data set is the face data to be tested based on different ages.
  • the facial feature association law is generated after the aging or rejuvenation of the face data to be tested, if the aging derivative face data is the same as the aging face data, and the two adjacent ages are derived. If the age difference of the face data is consistent, it can be considered that the median of the distribution interval should match the actual age.
  • the interval position of the age distribution interval is adjusted so that the age distribution interval finally matches the first reference age interval.
  • the value of the age distribution interval may be adjusted, such as increasing or decreasing a range of values, and matching is performed again, and the median of the age distribution interval is output when the matching condition is satisfied.
  • the age distribution interval represents the age data corresponding to the plurality of derived face data
  • the representation of the age distribution interval can effectively improve the fault tolerance rate in the age discrimination process, and can obtain and wait from the age distribution interval.
  • the accurate age data corresponding to the face data is measured, and the accuracy of the age determination of the face data to be measured is improved.
  • FIG. 8 provides a face recognition apparatus for training a face data generation model, that is, a generator based on a generated confrontation network (GAN), including a third acquisition unit 606, a second generation unit 607, a second determination unit 608, and a The second determining unit 609 and the updating unit 610, wherein:
  • GAN generated confrontation network
  • the third obtaining unit 606 is configured to obtain reference face data.
  • a second generating unit 607 configured to generate, according to the reference face data, a second derived face data set related to the reference face data, where the second derived face data set includes a plurality of different Derived face data.
  • the second determining unit 608 is configured to perform age determination on the derived face data in the second derived face data set, and generate an age distribution interval corresponding to the second derived face data set.
  • the second determining unit 609 is configured to determine whether the age distribution interval corresponding to the second derived face data set matches the second reference age range.
  • the updating unit 610 is configured to: if not, update the model parameter corresponding to the face data generation model until the age distribution interval corresponding to the second derivative face data set is determined by the age and the second Refer to the age range match.
  • the reference face data may be obtained from an AFAD database, an AFLW database, or some other commercial face database, and the face database provides sufficient reference face data so that the reference information has sufficient confidence. high.
  • step S102 in FIG. 1 the process of generating the second derivative face data set can refer to step S102 in FIG. 1 , and the generation principle and process are substantially the same. The biggest difference is merely replacing the face data to be tested with the reference in the adult face database. Face data, for the convenience of description, it will not be described in detail.
  • the concept of the accuracy rate may be referred to when the face data generation model, ie, the training of the generator After the completion of multiple loop matching in the process (including successful or unsuccessful matching), the accuracy will be counted.
  • the value of the accuracy is high enough, the generator can be considered to be sufficiently reliable.
  • the matching accuracy rate of a certain unit time can be set to be more than 98%.
  • the generator when the matching accuracy ratio of the generated age distribution interval and the second reference age interval of the second derivative face data set is lower than the preset matching accuracy value, it may be inferred that the generator generates the derivative based on the face data.
  • the model parameter confidence of the face data process is insufficient.
  • the parameter adjustment may be performed based on the second reference age interval, so that the regenerated generated face data can be matched with the second reference age range, and the different reference faces are adopted.
  • the data is continuously performed, such as obtaining reference face data—generating multiple derived face data—mismatching—update parameters—regenerating and matching the generated confrontation network (GAN) training process to continuously generate a confrontation network (GAN) for the generator.
  • GAN confrontation network
  • GAN confrontation network
  • the figure shows the structure of the first discriminating unit 603, which includes a first obtaining subunit 6032, an extracting subunit 6033, and a first generating subunit 6034, wherein:
  • a first obtaining sub-unit 6032 configured to acquire, by the third obtaining unit 606, reference face data, where the reference face data includes age data;
  • the extraction sub-unit 6033 is configured to, by the second generation unit 607, extract feature data corresponding to the reference face data according to the reference face data;
  • the first generation subunit 6034 is configured to associate the feature data with age data corresponding to the feature data to generate the age discriminant model.
  • the feature data and the age data corresponding to the feature data are associated, and a model of the feature data-age data relationship can be formed, so that the age discriminant model can determine the corresponding age according to the above relationship by acquiring different feature data. data.
  • a model of the feature data-age data relationship can be formed, so that the age discriminant model can determine the corresponding age according to the above relationship by acquiring different feature data. data.
  • how it relates can be determined on a case-by-case basis.
  • the device when extracting feature data, may be acquired from the face data using a deep learning and convolutional neural network to improve the accuracy of the obtained feature data.
  • the age discriminant model is formed by a plurality of feature data discriminant models being associated by a preset weight coefficient.
  • the discriminative accuracy of each feature data by the age discriminant model and the influence of each feature data on the overall discriminant of the face type are different.
  • the weight adjustment can be performed according to the above differences to make the age discriminant.
  • the age data available to the model is more accurate.
  • a plurality of obtained age data corresponding to different derivative face data are formed into an age distribution interval. For example, when a derived face data obtains the largest age data result in the first derived face data set, it is 30 years old, and another derived data obtains the smallest age data result in the first derived face data set, At the age of 25, based on the age data of the two ends, it can be known that the age distribution interval corresponding to the first derivative face data set is 25-30 years old.
  • this step can effectively expand the derivative face data for detection related to the face data to be tested, and improve the fault tolerance rate of the face age recognition based on the generated derivative face data, thereby improving the accuracy of the discrimination.
  • a generated confrontation network to acquire a face data generation model, that is, a generator, specifically based on reference face data, using a face data generation model to acquire a second derivative face data set, and utilizing The second reference age interval discriminates the accuracy of the second derived face data set to optimize the face data generation model, so that the face data generation model is convincable.
  • the generator using the generator, the first derived face data set corresponding to the face data to be tested is generated by the first generating unit; and then the first discriminating unit passes the trained derived age discriminant model to the first derived face data.
  • the plurality of derived face data are concentrated to determine the age of the feature data; finally, the first determining unit aggregates the age data to generate an age distribution interval, and obtains age data with high matching degree with the face data to be tested. .
  • This can effectively solve the problem of low recognition accuracy due to poor image angle, overexposure or over darkness in the process of discriminating the age of face data, and improve the fault tolerance rate of the algorithm in the process of age discrimination.
  • the face age recognition algorithm can dynamically adapt to a variety of different environments, greatly improving the accuracy of face age recognition.
  • the present invention also provides an electronic device, which may be a smart phone, a tablet, a smart watch, or the like.
  • electronic device 700 includes a processor 701, a memory 702.
  • the processor 701 is electrically connected to the memory 702 and can control reading and writing of the memory 702.
  • the processor 701 is a control center of the electronic device 700, and connects various parts of the entire electronic device 700 using various interfaces and lines, by executing or loading an application stored in the memory 702, and calling data stored in the memory 702, executing The various functions and processing data of the electronic device 700 enable overall monitoring of the electronic device 700.
  • the processor 701 in the electronic device 700 loads the instructions corresponding to the process of one or more applications into the memory 702 according to the following steps, and is stored in the memory 702 by the processor 701.
  • the application thus implementing various functions:
  • the age data corresponding to the face data to be tested is obtained according to the age distribution interval.
  • the processor 701 is further configured to perform the following steps:
  • a first derivative face data set related to the face data to be tested including:
  • the processor 701 is further configured to perform the following steps:
  • the method further includes:
  • the processor 701 is further configured to perform the following steps:
  • the age discriminant model is configured to discriminate age data corresponding to the feature data based on feature data extracted from the derived face data.
  • the age discriminant model includes a plurality of feature data discriminant models
  • the feature data discriminant model is configured to discriminate a feature data of a preset type, and obtain age data corresponding to the feature data of the preset category according to the feature data.
  • the derived face data in the first derived face data set is subjected to age determination by a preset age discriminant model, and an age corresponding to the first derived face data set is generated. Before the distribution interval, it also includes:
  • reference face data wherein the reference face data includes age data
  • the feature data is associated with age data corresponding to the feature data to generate the age discriminant model.
  • the feature data is obtained from facial data using a deep learning and convolutional neural network.
  • the age discriminant model is formed by a plurality of feature data discriminant models being associated by a preset weight coefficient.
  • the processor 701 is further configured to perform the following steps:
  • obtaining age data corresponding to the face data to be tested according to the age distribution interval including:
  • the age data corresponding to the face data to be tested is obtained according to the age distribution interval.
  • the processor 701 is further configured to perform the following steps:
  • the plurality of derived face data are continuously distributed according to the age corresponding to the age, to form the first derivative face data set.
  • the processor 701 is further configured to perform the following steps:
  • the age median data is used as the age data corresponding to the face data to be tested.
  • Memory 702 can be used to store applications and data.
  • the application stored in the memory 702 contains instructions executable in the processor 701.
  • Applications can be composed of various functional units.
  • the processor 701 executes various functional applications and data processing by running an application stored in the memory 702.
  • the electronic device 700 further includes a display screen 703, a control circuit 704, a radio frequency circuit 705, an input unit 706, an audio circuit 707, a sensor 708, and a power source 709.
  • the processor 701 is electrically connected to the radio frequency circuit 705, the input unit 706, the audio circuit 707, and the power source 709, respectively.
  • the display screen 703 can be used to display information entered by the user or information provided to the user as well as various graphical user interfaces of the electronic device, which can be composed of images, text, icons, video, and any combination thereof.
  • the control circuit 704 is electrically connected to the display screen 703 for controlling the display screen 703 to display information.
  • the radio frequency circuit 705 is configured to transmit and receive radio frequency signals to establish wireless communication with a network device or other electronic device through wireless communication, and to transmit and receive signals with the network device or other electronic devices.
  • the input unit 706 can be configured to receive input digits, character information, or user characteristic information (eg, fingerprints), and to generate keyboard, mouse, joystick, optical, or trackball signal inputs related to user settings and function controls.
  • the input unit 706 can include a fingerprint identification module.
  • the audio circuit 707 can provide an audio interface between the user and the electronic device through a speaker and a microphone.
  • Electronic device 700 may also include at least one type of sensor 708, such as a light sensor, motion sensor, and other sensors.
  • sensor 708 such as a light sensor, motion sensor, and other sensors.
  • Power source 709 is used to power various components of electronic device 700.
  • the power supply 709 can be logically coupled to the processor 701 through a power management system to enable functions such as managing charging, discharging, and power management through the power management system.
  • the electronic device 700 may further include a camera, a Bluetooth unit, and the like, and details are not described herein again.
  • Embodiments of the present invention also provide a storage medium in which a plurality of instructions are stored in the storage medium, the instructions being adapted to be loaded by a processor to perform the method as described above.
  • the electronic device of the present invention uses a generated confrontation network (GAN) to acquire a face data generation model, that is, a generator, specifically based on reference face data, and uses a face data generation model to acquire a second derivative face data set, and utilizes the above
  • GAN generated confrontation network
  • the second reference age interval discriminates the accuracy of the second derived face data set to optimize the face data generation model, so that the face data generation model is convincable.
  • using the face data generation model generating a first derivative face data set corresponding to the face data to be tested; and then using the trained age discriminant model, the plurality of derivatives in the first derivative face data set
  • the face data is used to determine the age of the feature data.
  • the age data is combined to generate an age distribution interval, and the age data with high matching degree with the face data to be tested is obtained.
  • This can effectively solve the problem of low recognition accuracy due to poor image angle, overexposure or over darkness in the process of discriminating the age of face data, and improve the fault tolerance rate of the algorithm in the process of age discrimination.
  • the face age recognition algorithm can dynamically adapt to a variety of different environments, greatly improving the accuracy of face age recognition.
  • the storage medium may include: a read only memory (ROM, Read) Only Memory), Random Access Memory (RAM), disk or CD.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Oral & Maxillofacial Surgery (AREA)
  • General Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Physics & Mathematics (AREA)
  • Human Computer Interaction (AREA)
  • Multimedia (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Image Analysis (AREA)
  • Measurement Of The Respiration, Hearing Ability, Form, And Blood Characteristics Of Living Organisms (AREA)
  • Image Processing (AREA)

Abstract

一种人脸识别方法、装置、存储介质以及电子设备,所述方法包括:获取待测人脸数据(S101);根据待测人脸数据,生成与待测人脸数据相关的第一衍生人脸数据集(S102);将第一衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间(S103);判断年龄分布区间是否与第一参考年龄区间匹配(S104);若是,则根据年龄分布区间,获得待测人脸数据对应的年龄数据(S105)。所述方法可提高人脸年龄识别的准确率。

Description

人脸识别方法、装置、存储介质及电子设备 技术领域
本发明涉及数据处理技术领域,尤其涉及人脸识别方法、装置、存储介质及电子设备。
背景技术
随着电子设备的快速发展,其已经成为人们日常生活不可或缺的一部分。特别是应用在移动终端中各式各样的应用程序,不同的功能提供给使用者多元的使用体验,人们也越来越离不开各种应用程序所带来的娱乐与便捷。
如今的电子设备获取照片或图像信息非常便捷,人们在使用电子设备的过程中会接触到大量的照片或其他图像信息。有时候,人们在查看照片或其他图像信息时,可能会对照片里的一些人的年龄感兴趣。常规的场景中,人们会根据图像中的目标人物的样貌进行大致的思考,并凭借自己的经验来判断这些人的年龄。可是,仅仅凭借经验显然不容易判断出图像中目标人物的实际年龄。
技术问题
本发明实施例提供一种人脸识别方法、装置、存储介质及电子设备,可以提高人脸识别的效率及准确率。
技术解决方案
第一方面,本发明实施例提供了以下技术方案:一种人脸识别方法,包括以下步骤:
获取待测人脸数据;
根据所述待测人脸数据,生成与所述待测人脸数据相关的第一衍生人脸数据集,其中,所述第一衍生人脸数据集包括多个不同的衍生人脸数据;
将所述第一衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间;
判断所述年龄分布区间是否与第一参考年龄区间匹配;
若是,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据。
第二方面,本发明实施例还提供了以下技术方案:
一种人脸识别装置,其中,所述人脸识别装置包括:
第一获取单元,用于获取待测人脸数据;
第一生成单元,用于根据所述待测人脸数据,生成与所述待测人脸数据相关的第一衍生人脸数据集,其中,所述第一衍生人脸数据集包括多个不同的衍生人脸数据;
第一判别单元,用于将所述第一衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间;
第一判断单元,用于判断所述年龄分布区间是否与第一参考年龄区间匹配;以及
第二获取单元,用于若是,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据。
第三方面,本发明实施例还提供了以下技术方案:
一种存储介质,其中,所述存储介质中存储有多条指令,所述指令适于由处理器加载以执行如上所述的人脸识别方法。
第四方面,本发明实施例还提供了以下技术方案:
一种电子设备,其中,包括处理器、存储器,所述处理器与所述存储器电性连接,所述存储器用于存储指令和数据,所述处理器用于执行如权上所述的人脸识别方法。
有益效果
本发明实施例提供一种人脸识别方法、装置、存储介质及电子设备,可以提高人脸识别的效率及准确率。
附图说明
图1是本发明实施例提供的人脸识别方法的实现流程示意图。
图2是本发明实施例提供的人脸识别方法的另一实现流程图。
图3是本发明实施例提供的对衍生人脸数据年龄判别的实现框图。
图4是本发明实施例提供的训练人脸数据生成模型的实现流程示意图。
图5是本发明实施例提供的获得年龄判别模型的实现流程示意图。
图6为本发明实施例提供的人脸识别装置的结构示意图。
图7为本发明实施例提供的人脸识别装置的另一结构示意图。
图8为本发明实施例提供的人脸识别装置的又一结构示意图。
图9为本发明实施例提供的第一判别单元的结构示意图。
图10为本发明实施例提供的电子设备的结构示意图。
图11为本发明实施例提供的电子设备的另一结构示意图。
本发明的最佳实施方式
请参照图式,其中相同的组件符号代表相同的组件,本发明的原理是以实施在一适当的运算环境中来举例说明。以下的说明是基于所例示的本发明具体实施例,其不应被视为限制本发明未在此详述的其它具体实施例。
在以下的说明中,本发明的具体实施例将参考由一部或多部计算机所执行的步骤及符号来说明,除非另有述明。因此,这些步骤及操作将有数次提到由计算机执行,本文所指的计算机执行包括了由代表了以一结构化型式中的数据的电子信号的计算机处理单元的操作。此操作转换该数据或将其维持在该计算机的内存系统中的位置处,其可重新配置或另外以本领域测试人员所熟知的方式来改变该计算机的运作。该数据所维持的数据结构为该内存的实体位置,其具有由该数据格式所定义的特定特性。但是,本发明原理以上述文字来说明,其并不代表为一种限制,本领域测试人员将可了解到以下所述的多种步骤及操作亦可实施在硬件当中。
本文所使用的术语“模块”可为在该运算系统上执行的软件对象。本文所述的不同组件、模块、引擎及服务可为在该运算系统上的实施对象。而本文所述的装置及方法可以以软件的方式进行实施,当然也可在硬件上进行实施,均在本发明保护范围之内。
本发明中的术语“第一”、“第二”和“第三”等是用于区别不同对象,而不是用于描述特定顺序。此外,术语“包括”和“具有”以及它们任何变形,意图在于覆盖不排他的包含。例如包含了一系列步骤或模块的过程、方法、系统、产品或设备没有限定于已列出的步骤或模块,而是某些实施例还包括没有列出的步骤或模块,或某些实施例还包括对于这些过程、方法、产品或设备固有的其它步骤或模块。
在本文中提及“实施例”意味着,结合实施例描述的特定特征、结构或特性可以包含在本发明的至少一个实施例中。在说明书中的各个位置出现该短语并不一定均是指相同的实施例,也不是与其它实施例互斥的独立的或备选的实施例。本领域技术人员显式地和隐式地理解的是,本文所描述的实施例可以与其它实施例相结合。
本发明实施例提供一种人脸识别方法,其中,包括以下步骤:
获取待测人脸数据;
根据所述待测人脸数据,生成与所述待测人脸数据相关的第一衍生人脸数据集,其中,所述第一衍生人脸数据集包括多个不同的衍生人脸数据;
将所述第一衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间;
判断所述年龄分布区间是否与第一参考年龄区间匹配;
若是,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据。
在一些实施例中,所述根据所述待测人脸数据,生成与所述待测人脸数据相关的第一衍生人脸数据集,包括:
根据所述待测人脸数据,通过预设的人脸数据生成模型生成与所述待测人脸数据相关的第一衍生人脸数据集。
在一些实施例中,在所述将所述测人脸数据生成与所述待测人脸数据相关的第一衍生人脸数据集之前,还包括:
获取参考人脸数据;
根据所述参考人脸数据,生成与所述参考人脸数据相关的第二衍生人脸数据集,其中,所述第二衍生人脸数据集包括多个不同的衍生人脸数据;
将所述第二衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第二衍生人脸数据集对应的年龄分布区间;
判断所述第二衍生人脸数据集对应的年龄分布区间是否与第二参考年龄区间匹配;
若否,更新所述人脸数据生成模型对应的模型参数,直至所述第二衍生人脸数据集对应的年龄分布区间,经所述年龄判别后均与所述第二参考年龄区间匹配。
在一些实施例中,所述将所述第一衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间,包括:
将所述多个衍生人脸数据通过预设的年龄判别模型分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间;
其中,所述年龄判别模型用于根据从所述衍生人脸数据提取的特征数据,判别出与所述特征数据对应的年龄数据。
在一些实施例中,所述年龄判别模型包括多个特征数据判别模型,
其中,一所述特征数据判别模型用于判别一预设种类的特征数据,并根据所述特征数据获得与所述预设种类的特征数据对应的年龄数据。
在一些实施例中,在所述将所述第一衍生人脸数据集中的衍生人脸数据通过预设的年龄判别模型分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间之前,还包括:
获取参考人脸数据,其中,所述参考人脸数据包括年龄数据;
根据所述参考人脸数据,提取与所述参考人脸数据对应的特征数据;
将所述特征数据和与所述特征数据对应的年龄数据进行关联,生成所述年龄判别模型。
在一些实施例中,所述特征数据利用深度学习和卷积神经网络从人脸数据中获取。
在一些实施例中,所述年龄判别模型由多个特征数据判别模型按预设的权重系数进行关联集合而成。
在一些实施例中,所述判断所述年龄分布区间是否与第一参考年龄区间匹配,包括:
判断每一所述衍生人脸数据经年龄判别后获得的年龄分布区间,与第一参考年龄区间之间的相关度是否高于预设阈值;
所述若是,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据,包括:
若所述相关度高于预设阈值,则根据所述年龄分布区间,获取所述待测人脸数据对应的年龄数据。
在一些实施例中,所述根据所述待测人脸数据,生成与所述待测人脸数据相关的第一衍生人脸数据集,包括:
将所述待测人脸数据按预设的不同年龄段进行人脸老化处理和人脸幼化处理,生成与所述待测人脸数据相关的多个代表不同年龄的衍生人脸数据;
将所述多个衍生人脸数据按与之对应的年龄的大小进行连续分布处理,形成所述第一衍生人脸数据集。
在一些实施例中,所述若是,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据,包括:
根据所述年龄分布区间,获取所述年龄分布区间的年龄中值数据;
将所述年龄中值数据作为所述待测人脸数据对应的年龄数据。
下面将结合附图和实施例来对本发明内容作进一步说明。
请参阅图1,所示为本发明实施例提供的人脸识别方法的实现流程,为了便于说明,附图仅示出了与本发明相关的内容。
所述人脸识别方法,应用于可用于执行数据处理的电子设备中,具体包括如下步骤:
在步骤S101中,获取待测人脸数据。
在一些实施例中,人脸数据可以是与之关联的数据集、图像等形式存在。而对于待测人脸数据的获取,可通过将待测人脸数据人工输入、人工选取等方式在电子设备中进行确认,也可以通过预设的程序选取、提取等方式获取待测人脸数据。
例如,在某一手机app上,通过选取手机图像集中的携带人脸的图像,以使图像被确认为待测人脸数据。又或者,在某一带多个人脸的图像中,通过人脸检测算法,确定每一人脸的位置,再基于人脸位置提取人脸数据。上述内容仅供说明用途,不作为本发明获取待测人脸数据方式的限定,实际应用时可通过多种方法对待测人脸数据进行获取。
在接下来的实施例中,为了便于说明,人脸数据将以图像的形式进行举例。
在步骤S102中,根据所述待测人脸数据,生成与所述待测人脸数据相关的第一衍生人脸数据集,其中,所述第一衍生人脸数据集包括多个不同的衍生人脸数据。
在一些实施例中,第一衍生人脸数据集基于待测人脸数据进行生成,其包括多个不同的与所述待测人脸数据对应的衍生人脸数据。在生成的过程中,第一衍生人脸数据集可以是利用待测人脸数据通过预设的生成器生成,其中,生成器可基于人脸的特征数据生成人脸数据。可以理解的,若以基于生成器生成的衍生人脸数据,其生成的人脸特征会与待测人脸数据一致,或者相近似,或者按预期方式进行特征改变。具体实现方式请参见下文实施例。假设第一衍生人脸数据采用基于特征数据的生成方法,首先系统会提取待测人脸数据中有关轮廓、五官、皮肤、表情、发型等特征数据,然后根据从人脸数据库中总结的一些人脸特征关联规律,例如不同年龄或不同地区的人脸之间的关联规律,生成与上述特征数据对应的衍生人脸数据。
以基于特征数据中五官的生成过程为例:系统从待测人脸数据中提取到其具有关于五官的特征数据为“五官A”,在生成第一衍生人脸数据集的过程中,将该特征数据“五官A”基于不同年龄之间的特征相关规律,按某特定年龄数值变化——假设为老化2岁,生成关于五官的特征数据为“五官A1”,并将该“五官A1”的特征添加入所生成的衍生人脸数据当中。可以理解的,若生成模型,即生成器可信,则当新生成的衍生人脸数据经基于该特征数据进行年龄判断后得出的结果,也应该与原待测人脸数据所得出的结果相差2岁左右。
上述例子仅以“五官A”作为说明,其他特征数据类型,至少还包括轮廓、皮肤、表情、发型等,其生成原理与五官相近,可参照上述例子来生成该衍生人脸数据中的上述特征数据,并将上述特征数据类型通过一些规则进行综合,实现对整个衍生人脸数据的生成。
可以理解的,所生成的第一衍生人脸数据集当中的衍生人脸数据的数量以及分布,可以根据实际需求进行设定。例如:将衍生人脸数据的数量设置为10个,其中,衍生人脸数据会基于特征数据中的不同年龄的人脸特征关联规律,并以每一相邻两个衍生人脸数据之间均会按老化或幼化0.5岁的基准进行生成,例如,分别获得老化0.5岁、1岁、1.5岁、2岁、2.5岁的5个衍生人脸数据,以及分别幼化0.5岁、1岁、1.5岁、2岁、2.5岁的5个衍生人脸数据,老化与幼化的衍生人脸数据数量相当,则最终会产生数量为10个、年龄分布区间为在待测人脸数据的年龄±2.5岁范围的第一衍生人脸数据集。
在步骤S103中,将所述第一衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间。
在一些实施例中,第一衍生人脸数据集中的多个衍生人脸数据,会通过预设的算法进行年龄判别,例如,第一衍生人脸数据集中有10个衍生人脸数据,则在年龄判别的过程中,会对者10个衍生人脸数据分别进行年龄判别,获得10个年龄数据。
接下来,将这10个年龄数据形成年龄分布区间。例如,当一衍生人脸数据获得在这第一衍生人脸数据集中最大的年龄数据结果,为30岁,而另一衍生数据获得在这第一衍生人脸数据集中最小的年龄数据结果,为25岁,则基于上述两端点的年龄数据结果,可以知道该第一衍生人脸数据集对应的年龄分布区间为25-30岁。
可以理解的,衍生人脸数据的数量以及区间的划分方式,可以根据实际应用情景进行调整。同时,该步骤可以有效拓展了与待测人脸数据相关的用于检测的衍生人脸数据,并基于上述生成的衍生人脸数据来提高人脸年龄识别的容错率,从而提高判别的准确度。
在步骤S104中,判断所述年龄分布区间是否与第一参考年龄区间匹配。
在一些实施例中,参考年龄区间,可以是通过将待测人脸数据在基于预设的人脸数据库获得的生成器进行年龄判别后得到的年龄分布区间。该人脸数据库可以是AFAD数据库、AFLW数据库或者是其他一些商用人脸数据库,以提供足够的参考人脸数据,使得参考信息的可置信度足够高。
在匹配过程中,将与第一衍生人脸数据集经年龄判别后所生成的年龄分布区间与上述第一参考年龄区间进行匹配,可以采用获取两个区间的中值、平均值、区间端点等数据进行比对,获得差值,并判断差值是否小于预设阈值,若是,则两者匹配,若否,则两者不匹配。当然,具体的判断条件可根据具体需要而定,本发明实施例对此不作限定。
例如,假设第一衍生人脸数据集经年龄判别后所生成的年龄分布区间为25-29,而通过待测人脸基于人脸数据库获得的第一参考年龄区间为27-31,为了判断他们两者的匹配程度,可采用生成的年龄分布区间的中值以及第一参考年龄区间的中值进行差值计算。基于上述方法,第一衍生人脸数据集对应的年龄分布区间的中值为(25+29)/2=27,而第一参考年龄区间的中值为(27+31)/2=29。设允许的差值为±3,则29-27<3,可以认为第一衍生人脸数据集对应的年龄分布区间与基于人脸数据库获得的第一参考年龄区间匹配,然后,从年龄分布区间选取其中值作为与待测人脸数据对应的年龄数据。若设允许的差值为±1,可以认为第一衍生人脸数据集对应的年龄分布区间与基于人脸数据库获得的第一参考年龄区间不匹配,此时,调整所述年龄分布区间的区间位置,以使所述年龄分布区间最终与第一参考年龄区间匹配。在一些实施例中,可以调整该年龄分布区间的数值,例如增减一定范围的值,并再次进行匹配,当满足匹配条件时即输出年龄分布区间的中值。通过根据待测人脸数据生成第一衍生人脸数据集,并对第一衍生人脸数据集中的多个衍生人脸数据进行年龄判断,可以有效避免因拍摄过程中存在高光或弱光的情况下,所造成的年龄判断偏差较大的问题,提高人脸年龄识别的准确度。
在步骤S105中,若是,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据。
在一些实施例中,若年龄分布区间为基于特征数据中的不同年龄的人脸特征关联规律,并以每一相邻两个衍生人脸数据之间均会按老化或幼化一定年龄数值的方式进行生成,此时如果该年龄分布区间与第一参考年龄区间匹配,则可以取该生成的年龄分布区间的两个年龄端点(最大值与最小值)之间的中值或平均值作为待测人脸数据的年龄数据。当然,如果该年龄分布区间通过其他方式进行生成,年龄数据还可以在年龄分布区间的其他位置进行获取。
例如,假设第一衍生人脸数据集经年龄判别后所生成的年龄分布区间为25-29,并且,该第一衍生人脸数据集中的衍生人脸数据是待测人脸数据基于不同年龄的人脸特征关联规律,并均为待测人脸数据老化或幼化一定年龄后生成,例如,分别老化0.5岁、1岁、1.5岁、2岁、2.5岁的5个衍生人脸数据,以及分别幼化0.5岁、1岁、1.5岁、2岁、2.5岁的5个衍生人脸数据,老化与幼化的衍生人脸数据数量相当,共10个,可以理解的,待测人脸数据的年龄数据可以为年龄分布区间25-29之间的中值,即(25+29)/2=27岁。这样,可以从该年龄分布区间中获得与待测人脸数据对应的较为准确的年龄数据。
基于上述内容,首先,获取待测人脸数据;然后,根据待测人脸数据,生成与所述待测人脸数据相关的第一衍生人脸数据集,其中,所述第一衍生人脸数据集包括多个不同的衍生人脸数据,以利用第一衍生人脸数据集拓展与待测人脸数据相关的人脸数据样本;接下来,将所述第一衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间,该年龄分布区间为基于第一衍生人脸数据进行年龄判别后获得,可以将某一点的数值拓宽到某一分布区间的数值,以提高年龄判别的容错率;进一步的,判断所述年龄分布区间是否与第一参考年龄区间匹配;最后,若匹配成功,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据。上述人脸识别方法,可以得出一个根据待测人脸数据获得的年龄分布区间,并利用该年分布区间进行年龄判别,可有效解决在判别人脸数据的年龄的过程中,因可能出现的图像角度不佳、过曝或过暗导致的识别准确率不高的问题,提高年龄判别过程中算法的容错率,以使人脸年龄识别算法可动态适配多种不同环境,大幅提高人脸年龄识别的准确度。
参见图2,图2为本发明实施例提供的人脸识别方法的另一实现流程,为了便于说明,仅示出了与本发明内容相关的部分。
在步骤S201中,获取待测人脸数据。
在一些实施例中,人脸数据可以是与之关联的数据集、图像等形式存在。而对于待测人脸数据的获取,可通过将待测人脸数据人工输入、人工选取等方式在电子设备中进行确认,也可以通过预设的程序选取、提取等方式获取待测人脸数据。
例如,在某一手机app上,通过选取手机图像集中的携带人脸的图像,以使图像被确认为待测人脸数据。又或者,在某一带多个人脸的图像中,通过人脸检测算法,确定每一人脸的位置,再基于人脸位置提取人脸数据。上述内容仅供说明用途,不作为本发明获取待测人脸数据方式的限定,实际应用时可通过多种方法对待测人脸数据进行获取。
在一些实施例中,待测人脸数据,会通过预设的人脸数据生成模型,即通过预设的生成器生成与所述待测人脸数据相关的第一衍生人脸数据集,其中,所述第一衍生人脸数据集包括多个不同的衍生人脸数据。
在步骤S202中,将所述待测人脸数据按预设的不同年龄段进行人脸老化处理和人脸幼化处理,生成与所述待测人脸数据相关的多个代表不同年龄的衍生人脸数据;
在步骤S203中,将所述多个衍生人脸数据按与之对应的年龄的大小进行连续分布处理,形成所述第一衍生人脸数据集;其中,所述第一衍生人脸数据集包括多个不同的衍生人脸数据。
假设第一衍生人脸数据采用基于特征数据的生成方法,首先系统会提取待测人脸数据中有关轮廓、五官、皮肤、表情、发型等特征数据,然后根据从人脸数据库中总结的一些人脸特征关联规律,例如不同年龄或不同地区的人脸之间的关联规律,生成与上述特征数据对应的衍生人脸数据。
以基于特征数据中五官的生成过程为例:系统从待测人脸数据中提取到其具有关于五官的特征数据为“五官A”,在生成第一衍生人脸数据集的过程中,将该特征数据“五官A”基于不同年龄之间的特征相关规律,按某特定年龄数值变化——假设为老化2岁,生成关于五官的特征数据为“五官A1”,并将该“五官A1”的特征添加入所生成的衍生人脸数据当中。可以理解的,若生成模型,即生成器可信,则当新生成的衍生人脸数据经基于该特征数据进行年龄判断后得出的结果,也应该与原待测人脸数据所得出的结果相差2岁左右。
上述例子仅以“五官A”作为说明,其他特征数据类型,至少还包括轮廓、皮肤、表情、发型等,其生成原理与五官相近,可参照上述例子来生成该衍生人脸数据中的上述特征数据,并将上述特征数据类型通过一些规则进行综合,实现对整个衍生人脸数据的生成。
可以理解的,所生成的第一衍生人脸数据集当中的衍生人脸数据的数量以及分布,可以根据实际需求进行设定。例如:将衍生人脸数据的数量设置为10个,其中,衍生人脸数据会基于特征数据中的不同年龄的人脸特征关联规律,并以每一相邻两个衍生人脸数据之间均会按老化或幼化0.5岁的基准进行生成,例如,分别获得老化0.5岁、1岁、1.5岁、2岁、2.5岁的5个衍生人脸数据,以及分别幼化0.5岁、1岁、1.5岁、2岁、2.5岁的5个衍生人脸数据,老化与幼化的衍生人脸数据数量相当,则最终会产生数量为10个、年龄分布区间为在待测人脸数据的年龄±2.5岁范围的第一衍生人脸数据集。
在一些实施例中,为了提高通过人脸数据生成模型,即生成器生成的第一衍生人脸数据集的可置信度,可通过引入生成对抗网络(GAN)来训练该生成器来达到上述目的。
在步骤204中,将所述多个衍生人脸数据通过预设的年龄判别模型分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间;其中,所述年龄判别模型用于根据从所述衍生人脸数据提取的特征数据,判别出与所述特征数据对应的年龄数据。
进一步的,在年龄判别模型中,可以包括多个特征数据判别模型,其中,一所述特征数据判别模型用于判别一预设种类的特征数据,并根据所述特征数据获得与所述预设种类的特征数据对应的年龄数据。可以理解的,每一个特征数据判别模型,都会基于与其对应的特征数据给出一个年龄判别结果。
图3示出了对衍生人脸数据的判别实现框图,如图3所示,在对衍生人脸数据进行年龄判别的过程中,从衍生人脸数据中提取出关于五官的特征数据,并可将其定义为“五官”,则在年龄判别模型中,会有一针对关于五官的特征数据进行年龄判别的特征数据判别模型“判别模型A”。当特征数据“五官”需要进行年龄判别时,系统会默认让“判别模型A”对其进行年龄判别,并获得一个与“五官”对应年龄数据“年龄A”。可以理解的是,除了上述针对五官进行判别后获得的年龄数据以外,还可以通过对上述轮廓、皮肤、表情、发型等特征数据进行判别,其流程与五官的判别流程相似,并分别获得年龄数据“年龄B”“年龄C”“年龄D”“年龄E”。当获得年龄判别模型中所有特征数据判别模型所判别得出的年龄数据后,会对这些数据进行汇总计算,以获得整体的关于人脸数据的年龄数据。
在步骤S205中,判断每一所述衍生人脸数据经年龄判别后获得的年龄分布区间,与第一参考年龄区间之间的相关度是否高于预设阈值。
此处通过相关度的高低来评判两者的匹配程度,当然,相关度的高低与否,可具体根据算法来定,例如采用两者区间的重合部分占整个区间范围的比值作为相关度的比较参数,则当两者区间的重合部分占整个区间范围的比值大于某预设阈值时,则可以认为两者匹配,当然,也可以利用生成的年龄分布区间的中值以及第一参考年龄区间的中值进行差值计算,结合其他因素来判断两者之间的相关度,当相关度大于某预设阈值时,则可以认为两者匹配。可以理解的,该相关度的计算以及设置的预设阈值大小,可以根据算法不同作出适当调整,其最终目的在于以一种数值大小来定义两者分布区间的匹配程度。
在一些实施例中,若所述相关度高于预设阈值,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据;
在步骤S206中,作为一个实施例,当相关度高于预设阈值时,会根据所述年龄分布区间,获取所述年龄分布区间的年龄中值数据。
在步骤S207中,将所述年龄中值数据作为所述待测人脸数据对应的年龄数据。
在一些实施例中,若年龄分布区间为基于特征数据中的不同年龄的人脸特征关联规律,并以每一相邻两个衍生人脸数据之间均会按老化或幼化一定年龄数值的方式进行生成,此时如果该年龄分布区间与第一参考年龄区间匹配,则可以取该生成的年龄分布区间的两个年龄端点(最大值与最小值)之间的中值作为待测人脸数据的年龄数据。
例如,假设第一衍生人脸数据集经年龄判别后所生成的年龄分布区间为25-29,并且,该第一衍生人脸数据集中的衍生人脸数据是待测人脸数据基于不同年龄的人脸特征关联规律,并均为待测人脸数据老化或幼化一定年龄后生成,若老化的衍生人脸数据与幼化的衍生人脸数据均相同,并且相邻两个年龄对应的衍生人脸数据的年龄差均一致,则可以认为其分布区间的中值与实际年龄的匹配度应该是最高的。
在步骤S208中,若所述相关度低于预设阈值,则调整所述年龄分布区间的区间位置,以使所述年龄分布区间最终与第一参考年龄区间匹配。
当相关度较低,使得年龄区间不匹配时,此时,调整所述年龄分布区间的区间位置,以使所述年龄分布区间最终与第一参考年龄区间匹配。在一些实施例中,可以调整该年龄分布区间的数值,例如增减一定范围的值,并再次进行匹配,当满足匹配条件时即输出年龄分布区间的中值。
可以理解的,年龄分布区间表示了多个衍生人脸数据对应的年龄数据,通过年龄分布区间的表示方式,可以有效提高年龄判别过程中的容错率,并可从该年龄分布区间中获得与待测人脸数据对应的较为准确的年龄数据,提高对待测人脸数据进行年龄判别的准确率。
参见图4,为本发明实施例提供的训练人脸数据生成模型的实现流程示意图,为了便于说明,仅示出了与本发明内容相关的部分。
图4提供了一种基于生成对抗网络(GAN)来训练人脸数据生成模型,即生成器的方法,其中:
步骤S301,获取参考人脸数据。
步骤S302,根据所述参考人脸数据,生成与所述参考人脸数据相关的第二衍生人脸数据集,其中,所述第二衍生人脸数据集包括多个不同的衍生人脸数据。
步骤S303,将所述第二衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第二衍生人脸数据集对应的年龄分布区间。
步骤S304,判断所述第二衍生人脸数据集对应的年龄分布区间是否与第二参考年龄区间匹配。
步骤S305,若否,更新所述人脸数据生成模型对应的模型参数,直至所述第二衍生人脸数据集对应的年龄分布区间,经所述年龄判别后均与所述第二参考年龄区间匹配。
在一些实施例中,该参考人脸数据可以是从AFAD数据库、AFLW数据库或者是其他一些商用人脸数据库中获取,上述人脸数据库提供足够的参考人脸数据,使得参考信息的可置信度足够高。
可以理解的,该第二衍生人脸数据集的生成过程可参照图1中的步骤S102,其生成原理及过程大致相同,最大的区别仅仅是将待测人脸数据替换成人脸数据库中的参考人脸数据,为了便于说明,故对此不再赘述。
在一些实施例中,判断所述第二衍生人脸数据集对应的年龄分布区间是否与第二参考年龄区间匹配,可以引用准确率的概念,当在生成器的训练过程中多次循环匹配完成后(包括匹配成功或不成功),均会对准确率进行统计,当该准确率的数值足够高,则可以认为该生成器足够可信。例如,可以设定其某一单位时间的匹配准确率达到98%以上,当生成器,即生成器在训练的过程中的匹配准确率达到上述数值,则可以暂停训练。
当然,当生成的第二衍生人脸数据集对应的年龄分布区间与第二参考年龄区间的匹配准确率低于预设的匹配准确率值时,可以推断出该生成器,即生成器基于人脸数据生成衍生人脸数据过程的模型参数置信度不足,此时可以基于第二参考年龄区间进行参数调整,以使再次生成的衍生人脸数据可与第二参考年龄区间匹配,并通过对不同的参考人脸数据不断进行如获取参照人脸数据—生成多个衍生人脸数据—不匹配—更新参数—再次生成、匹配的生成对抗网络(GAN)训练过程,以不断对生成器进行生成对抗网络(GAN)生成训练。并且,利用生成多个衍生人脸数据来进行匹配,可以有效提高训练样本数量,由此可提高生成器在训练过程中的训练效率,获得更好的训练效果,使生成器,即生成器达到较高的置信度。
参见图5,该图示出了获得年龄判别模型的实现流程,在一些实施例中,为了获得上述年龄判别模型,可以通过以下步骤实现:
步骤S501,获取参考人脸数据,其中,所述参考人脸数据包括年龄数据;
步骤S502,根据所述参考人脸数据,提取与所述参考人脸数据对应的特征数据;
步骤S503,将所述特征数据和与所述特征数据对应的年龄数据进行关联,生成所述年龄判别模型。
在上述步骤中,将特征数据和与特征数据对应的年龄数据进行关联,可以形成特征数据—年龄数据关系的模型,使得年龄判别模型通过获取不同的特征数据即可根据上述关系判别出对应的年龄数据。当然,具体实现时,其如何关联可以根据具体情况而定。
在一些实施例中,在提取特征数据的时候,可以让设备利用深度学习和卷积神经网络从人脸数据中获取,以提高所获得的特征数据的准确度。
在一些实施例中,所述年龄判别模型由多个特征数据判别模型按预设的权重系数进行关联集合而成。因年龄判别模型对每个特征数据的判别准确度以及每个特征数据对脸型整体判别的影响大小均有所不同,具体进行权重配比的时候,可以根据上述不同进行权重调整,以使年龄判别模型所能获得的年龄数据的准确度更高。
最后,将多个获得的对应不同衍生人脸数据的年龄数据,形成年龄分布区间。例如,当一衍生人脸数据获得在这第一衍生人脸数据集中最大的年龄数据结果,为30岁,而另一衍生数据获得在这第一衍生人脸数据集中最小的年龄数据结果,为25岁,则基于上述两端点的年龄数据结果,可以知道该第一衍生人脸数据集对应的年龄分布区间为25-30岁。
可以理解的,衍生人脸数据的数量以及区间的划分方式,可以根据实际应用情景进行调整。同时,该步骤可以有效拓展了与待测人脸数据相关的用于检测的衍生人脸数据,并基于上述生成的衍生人脸数据来提高人脸年龄识别的容错率,从而提高判别的准确度。
总而言之,基于上述实施例,利用生成对抗网络(GAN)获取人脸数据生成模型,即生成器,具体为基于参考人脸数据,利用生成器获取第二衍生人脸数据集,并利用上述第二参考年龄区间对第二衍生人脸数据集的准确度进行判别,以优化该生成器,使该生成器可置信。然后,利用该人脸数据生成模型,即利用生成器生成与待测人脸数据对应的第一衍生人脸数据集;然后通过经训练的年龄判别模型,对该第一衍生人脸数据集中的多个衍生人脸数据进行特征数据的年龄判别;最后通过年龄判别后将年龄数据汇总生成年龄分布区间,并在其中获得与待测人脸数据匹配度较高的年龄数据。如此可有效解决在判别人脸数据的年龄的过程中,因可能出现的图像角度不佳、过曝或过暗导致的识别准确率不高的问题,提高年龄判别过程中算法的容错率,以使人脸年龄识别算法可动态适配多种不同环境,大幅提高人脸年龄识别的准确度。
本发明实施例还提供一种人脸识别装置,其中,所述人脸识别装置包括:
第一获取单元,用于获取待测人脸数据;
第一生成单元,用于根据所述待测人脸数据,生成与所述待测人脸数据相关的第一衍生人脸数据集,其中,所述第一衍生人脸数据集包括多个不同的衍生人脸数据;
第一判别单元,用于将所述第一衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间;
第一判断单元,用于判断所述年龄分布区间是否与第一参考年龄区间匹配;以及
第二获取单元,用于若是,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据。
在一些实施例中,所述人脸识别装置包括:
第一获取单元,用于获取待测人脸数据;
第一生成单元,用于根据所述待测人脸数据,生成与所述待测人脸数据相关的第一衍生人脸数据集,其中,所述第一衍生人脸数据集包括多个不同的衍生人脸数据;
第一判别单元,用于将所述第一衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间;
第一判断单元,用于判断所述年龄分布区间是否与第一参考年龄区间匹配;以及
第二获取单元,用于若是,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据。
在一些实施例中,所述第一生成单元包括:
第一生成子单元,用于根据所述待测人脸数据,通过预设的人脸数据生成模型生成与所述待测人脸数据相关的第一衍生人脸数据集。
在一些实施例中,所述人脸识别装置还包括:
第三获取单元,用于获取参考人脸数据;
第二生成单元,用于根据所述参考人脸数据,生成与所述参考人脸数据相关的第二衍生人脸数据集,其中,所述第二衍生人脸数据集包括多个不同的衍生人脸数据;
第二判别单元,用于将所述第二衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第二衍生人脸数据集对应的年龄分布区间;
第二判断单元,用于判断所述第二衍生人脸数据集对应的年龄分布区间是否与第二参考年龄区间匹配;以及
更新单元,用于若否,更新所述人脸数据生成模型对应的模型参数,直至所述第二衍生人脸数据集对应的年龄分布区间,经所述年龄判别后均与所述第二参考年龄区间匹配。
在一些实施例中,所述第一判别单元,包括:
第一判别子单元,用于将所述多个衍生人脸数据通过预设的年龄判别模型分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间;
其中,所述年龄判别模型用于根据从所述衍生人脸数据提取的特征数据,判别出与所述特征数据对应的年龄数据。
在一些实施例中,所述年龄判别模型包括多个特征数据判别模型,
其中,一所述特征数据判别模型用于判别一预设种类的特征数据,并根据所述特征数据获得与所述预设种类的特征数据对应的年龄数据。
在一些实施例中,所述第一判别单元,还包括:
第一获取子单元,用于获取参考人脸数据,其中,所述参考人脸数据包括年龄数据;
提取子单元,用于根据所述参考人脸数据,提取与所述参考人脸数据对应的特征数据;以及
第一生成子单元,用于将所述特征数据和与所述特征数据对应的年龄数据进行关联,生成所述年龄判别模型。
在一些实施例中,所述特征数据利用深度学习和卷积神经网络从人脸数据中获取。
在一些实施例中,所述年龄判别模型由多个特征数据判别模型按预设的权重系数进行关联集合而成。
在一些实施例中,所述第一判断单元,包括:
第一判断子单元,用于判断每一所述衍生人脸数据经年龄判别后获得的年龄分布区间,与第一参考年龄区间之间的相关度是否高于预设阈值;
所述第二获取单元,包括:
第二获取子单元,用于若所述相关度高于预设阈值,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据;
调整子单元,用于若所述相关度低于预设阈值,则调整所述年龄分布区间的区间位置,以使所述年龄分布区间最终与第一参考年龄区间匹配。
在一些实施例中,所述第一生成单元,包括:
第二生成子单元,用于将所述待测人脸数据按预设的不同年龄段进行人脸老化处理和人脸幼化处理,生成与所述待测人脸数据相关的多个代表不同年龄的衍生人脸数据;以及
第三生成子单元,用于将所述多个衍生人脸数据按与之对应的年龄的大小进行连续分布处理,形成所述第一衍生人脸数据集。
在一些实施例中,所述第二获取单元,包括:
第三获取子单元,用于根据所述年龄分布区间根据所述年龄判别后的结果,获取所述年龄分布区间的经所述连续分布处理后的所述第一衍生人脸数据集对应的年龄中值数据;以及
确认子单元,用于将所述年龄中值数据作为所述待测人脸数据对应的年龄数据。
参见图6,所示为本发明实施例提供的人脸识别装置的结构,所述装置包括第一获取单元601、第一生成单元602、第一判别单元603、第一判断单元604以及第二获取单元605。
所述人脸识别装置,具体包括:
第一获取单元601,用于获取待测人脸数据。
在一些实施例中,人脸数据可以是与之关联的数据集、图像等形式存在。而对于待测人脸数据的获取,可通过将待测人脸数据人工输入、人工选取等方式在电子设备中进行确认,也可以通过预设的程序选取、提取等方式获取待测人脸数据。
例如,在某一手机app上,通过选取手机图像集中的携带人脸的图像,以使图像被确认为待测人脸数据。又或者,在某一带多个人脸的图像中,通过人脸检测算法,确定每一人脸的位置,再基于人脸位置提取人脸数据。上述内容仅供说明用途,不作为本发明获取待测人脸数据方式的限定,实际应用时可通过多种方法对待测人脸数据进行获取。
在接下来的实施例中,为了便于说明,人脸数据将以图像的形式进行举例。
第一生成单元602,用于根据所述待测人脸数据,生成与所述待测人脸数据相关的第一衍生人脸数据集,其中,所述第一衍生人脸数据集包括多个不同的衍生人脸数据。
在一些实施例中,第一衍生人脸数据集基于待测人脸数据进行生成,其包括多个不同的与所述待测人脸数据对应的衍生人脸数据。在生成的过程中,可以采用基于特征数据的生成方法,也可以采用基于其他方式的生成方法对衍生人脸数据进行生成。可以理解的,若以基于特征数据生成方法生成的衍生人脸数据,其特征会与待测人脸数据一致,或者相近似,或者按预期方式进行特征改变。
假设第一衍生人脸数据采用基于特征数据的生成方法,首先系统会提取待测人脸数据中有关轮廓、五官、皮肤、表情、发型等特征数据,然后根据从人脸数据库中总结的一些人脸特征关联规律,例如不同年龄或不同地区的人脸之间的关联规律,生成与上述特征数据对应的衍生人脸数据。
以基于特征数据中五官的生成过程为例:系统从待测人脸数据中提取到其具有关于五官的特征数据为“五官A”,在生成第一衍生人脸数据集的过程中,将该特征数据“五官A”基于不同年龄之间的特征相关规律,按某特定年龄数值变化——假设为老化2岁,生成关于五官的特征数据为“五官A1”,并将该“五官A1”的特征添加入所生成的衍生人脸数据当中。可以理解的,若生成模型可信,则当新生成的衍生人脸数据经基于该特征数据进行年龄判断后得出的结果,也应该与原待测人脸数据所得出的结果相差2岁左右。
上述例子仅以“五官A”作为说明,其他特征数据类型,至少还包括轮廓、皮肤、表情、发型等,其生成原理与五官相近,可参照上述例子来生成该衍生人脸数据中的上述特征数据,并将上述特征数据类型通过一些规则进行综合,实现对整个衍生人脸数据的生成。
可以理解的,所生成的第一衍生人脸数据集当中的衍生人脸数据的数量以及分布,可以根据实际需求进行设定。例如:将衍生人脸数据的数量设置为10个,其中,衍生人脸数据会基于特征数据中的不同年龄的人脸特征关联规律,并以每一相邻两个衍生人脸数据之间均会按老化或幼化0.5岁的基准进行生成,例如,分别获得老化0.5岁、1岁、1.5岁、2岁、2.5岁的5个衍生人脸数据,以及分别幼化0.5岁、1岁、1.5岁、2岁、2.5岁的5个衍生人脸数据,老化与幼化的衍生人脸数据数量相当,则最终会产生数量为10个、年龄分布区间为在待测人脸数据的年龄±2.5岁范围的第一衍生人脸数据集。
第一判别单元603,用于将所述第一衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间。
在一些实施例中,第一衍生人脸数据集中的多个衍生人脸数据,会通过预设的算法进行年龄判别,例如,第一衍生人脸数据集中有10个衍生人脸数据,则在年龄判别的过程中,会对者10个衍生人脸数据分别进行年龄判别,获得10个年龄数据。
接下来,将这10个年龄数据形成年龄分布区间。例如,当一衍生人脸数据获得在这第一衍生人脸数据集中最大的年龄数据结果,为30岁,而另一衍生数据获得在这第一衍生人脸数据集中最小的年龄数据结果,为25岁,则基于上述两端点的年龄数据结果,可以知道该第一衍生人脸数据集对应的年龄分布区间为25-30岁。
可以理解的,衍生人脸数据的数量以及区间的划分方式,可以根据实际应用情景进行调整。同时,该步骤可以有效拓展了与待测人脸数据相关的用于检测的衍生人脸数据,并基于上述生成的衍生人脸数据来提高人脸年龄识别的容错率,从而提高判别的准确度。
第一判断单元604,用于判断所述年龄分布区间是否与第一参考年龄区间匹配。
在一些实施例中,第一参考年龄区间,可以是通过将待测人脸数据基于预设的人脸数据库获得相应的年龄分布区间。该人脸数据库可以是AFAD数据库、AFLW数据库或者是其他一些商用人脸数据库,以提供足够的参考人脸数据,使得参考信息的可置信度足够高。
在匹配过程中,将与第一衍生人脸数据集经年龄判别后所生成的年龄分布区间与上述年龄分布区间进行匹配,可以采用获取两个年龄分布区间的中值、平均值、区间端点等数据进行比对,获得差值,并判断差值是否小于预设阈值,若是,则两者匹配,若否,则两者不匹配。当然,具体的判断条件可根据具体需要而定,本发明实施例对此不作限定。
例如,假设第一衍生人脸数据集经年龄判别后所生成的年龄分布区间为25-29,而通过待测人脸基于人脸数据库获得的第一参考年龄区间为27-31,为了判断他们两者的匹配程度,可采用生成的年龄分布区间的中值以及第一参考年龄区间的中值进行差值计算。基于上述方法,第一衍生人脸数据集对应的年龄分布区间的中值为(25+29)/2=27,而第一参考年龄区间的中值为(27+31)/2=29。设允许的差值为±3,则29-27<3,可以认为第一衍生人脸数据集对应的年龄分布区间与基于人脸数据库获得的第一参考年龄区间匹配,然后,从年龄分布区间选取其中值作为与待测人脸数据对应的年龄数据。若设允许的差值为±1,可以认为第一衍生人脸数据集对应的年龄分布区间与基于人脸数据库获得的第一参考年龄区间不匹配,此时,在一些实施例中,可以将该年龄分布区间的数值进行增减一定范围,并再次进行匹配,当满足匹配条件时即输出年龄分布区间的中值。通过根据待测人脸数据生成第一衍生人脸数据集,并对第一衍生人脸数据集中的多个衍生人脸数据进行年龄判断,可以有效避免因拍摄过程中存在高光或弱光的情况下,所造成的年龄判断偏差较大的问题,提高人脸年龄识别的准确度。
第二获取单元605,用于若是,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据。
在一些实施例中,若年龄分布区间为基于特征数据中的不同年龄的人脸特征关联规律,并以每一相邻两个衍生人脸数据之间均会按老化或幼化一定年龄数值的方式进行生成,此时如果该年龄分布区间与第一参考年龄区间匹配,则可以取该生成的年龄分布区间的两个年龄端点(最大值与最小值)之间的中值或平均值作为待测人脸数据的年龄数据。当然,如果该年龄分布区间通过其他方式进行生成,年龄数据还可以在年龄分布区间的其他位置进行获取。
例如,假设第一衍生人脸数据集经年龄判别后所生成的年龄分布区间为25-29,并且,该第一衍生人脸数据集中的衍生人脸数据是待测人脸数据基于不同年龄的人脸特征关联规律,并均为待测人脸数据老化或幼化一定年龄后生成,例如,分别老化0.5岁、1岁、1.5岁、2岁、2.5岁的5个衍生人脸数据,以及分别幼化0.5岁、1岁、1.5岁、2岁、2.5岁的5个衍生人脸数据,老化与幼化的衍生人脸数据数量相当,共10个,可以理解的,待测人脸数据的年龄数据可以为年龄分布区间25-29之间的中值,即(25+29)/2=27岁。这样,可以从该年龄分布区间中获得与待测人脸数据对应的较为准确的年龄数据。
基于上述内容,首先,第一获取单元601获取待测人脸数据;然后,第一生成单元602根据待测人脸数据,生成与所述待测人脸数据相关的第一衍生人脸数据集,其中,所述第一衍生人脸数据集包括多个不同的衍生人脸数据,以利用第一衍生人脸数据集拓展与待测人脸数据相关的人脸数据样本;接下来,第一判别单元603将所述第一衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间,该年龄分布区间为基于第一衍生人脸数据进行年龄判别后获得,可以将某一点的数值拓宽到某一分布区间的数值,以提高年龄判别的容错率;进一步的,第一判断单元604判断所述年龄分布区间是否与第一参考年龄区间匹配;最后,第二获取单元605若匹配成功,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据。上述人脸识别方法,可以得出一个根据待测人脸数据获得的年龄分布区间,并利用该年分布区间进行年龄判别,可有效解决在判别人脸数据的年龄的过程中,因可能出现的图像角度不佳、过曝或过暗导致的识别准确率不高的问题,提高年龄判别过程中算法的容错率,以使人脸年龄识别算法可动态适配多种不同环境,大幅提高人脸年龄识别的准确度。
参见图7,图7为本发明实施例提供的人脸识别装置的另一结构,其中包括:
第一获取单元601,用于获取待测人脸数据。
在一些实施例中,人脸数据可以是与之关联的数据集、图像等形式存在。而对于待测人脸数据的获取,可通过将待测人脸数据人工输入、人工选取等方式在电子设备中进行确认,也可以通过预设的程序选取、提取等方式获取待测人脸数据。
例如,在某一手机app上,通过选取手机图像集中的携带人脸的图像,以使图像被确认为待测人脸数据。又或者,在某一带多个人脸的图像中,通过人脸检测算法,确定每一人脸的位置,再基于人脸位置提取人脸数据。上述内容仅供说明用途,不作为本发明获取待测人脸数据方式的限定,实际应用时可通过多种方法对待测人脸数据进行获取。
在一些实施例中,待测人脸数据,会通过预设的人脸数据生成模型生成与所述待测人脸数据相关的第一衍生人脸数据集,其中,所述第一衍生人脸数据集包括多个不同的衍生人脸数据。
在一些实施例中,第一生成单元602包括第二生成子单元6021以及第三生成子单元6022,其中:
第二生成子单元6021,用于将所述待测人脸数据按预设的不同年龄段进行人脸老化处理和人脸幼化处理,生成与所述待测人脸数据相关的多个代表不同年龄的衍生人脸数据;
第三生成子单元6022,用于将所述多个衍生人脸数据按与之对应的年龄的大小进行连续分布处理,形成所述第一衍生人脸数据集;其中,所述第一衍生人脸数据集包括多个不同的衍生人脸数据。
在一些实施例中,第一衍生人脸数据集基于待测人脸数据进行生成,其包括多个不同的与所述待测人脸数据对应的衍生人脸数据。在通过人脸数据生成模型生成的过程中,可以采用基于特征数据的生成方法,也可以采用基于其他方式的生成方法对衍生人脸数据进行生成。可以理解的,若以基于特征数据生成方法生成的衍生人脸数据,其特征会与待测人脸数据一致,或者相近似,或者按预期方式进行特征改变。
假设第一衍生人脸数据采用基于特征数据的生成方法,首先系统会提取待测人脸数据中有关轮廓、五官、皮肤、表情、发型等特征数据,然后根据从人脸数据库中总结的一些人脸特征关联规律,例如不同年龄或不同地区的人脸之间的关联规律,生成与上述特征数据对应的衍生人脸数据。
以基于特征数据中五官的生成过程为例:系统从待测人脸数据中提取到其具有关于五官的特征数据为“五官A”,在生成第一衍生人脸数据集的过程中,将该特征数据“五官A”基于不同年龄之间的特征相关规律,按某特定年龄数值变化——假设为老化2岁,生成关于五官的特征数据为“五官A1”,并将该“五官A1”的特征添加入所生成的衍生人脸数据当中。可以理解的,若生成模型可信,则当新生成的衍生人脸数据经基于该特征数据进行年龄判断后得出的结果,也应该与原待测人脸数据所得出的结果相差2岁左右。
上述例子仅以“五官A”作为说明,其他特征数据类型,至少还包括轮廓、皮肤、表情、发型等,其生成原理与五官相近,可参照上述例子来生成该衍生人脸数据中的上述特征数据,并将上述特征数据类型通过一些规则进行综合,实现对整个衍生人脸数据的生成。
可以理解的,所生成的第一衍生人脸数据集当中的衍生人脸数据的数量以及分布,可以根据实际需求进行设定。例如:将衍生人脸数据的数量设置为10个,其中,衍生人脸数据会基于特征数据中的不同年龄的人脸特征关联规律,并以每一相邻两个衍生人脸数据之间均会按老化或幼化0.5岁的基准进行生成,例如,分别获得老化0.5岁、1岁、1.5岁、2岁、2.5岁的5个衍生人脸数据,以及分别幼化0.5岁、1岁、1.5岁、2岁、2.5岁的5个衍生人脸数据,老化与幼化的衍生人脸数据数量相当,则最终会产生数量为10个、年龄分布区间为在待测人脸数据的年龄±2.5岁范围的第一衍生人脸数据集。
在一些实施例中,为了提高通过人脸数据生成模型,即生成器生成的第一衍生人脸数据集的可置信度,可通过引入生成对抗网络(GAN)来训练该生成器来达到上述目的。
第一判别单元603包括第一判别子单元6031,所述第一判别子单元6031用于将所述多个衍生人脸数据通过预设的年龄判别模型分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间;其中,所述年龄判别模型用于根据从所述衍生人脸数据提取的特征数据,判别出与所述特征数据对应的年龄数据。
进一步的,在年龄判别模型中,可以包括多个特征数据判别模型,其中,一所述特征数据判别模型用于判别一预设种类的特征数据,并根据所述特征数据获得与所述预设种类的特征数据对应的年龄数据。可以理解的,每一个特征数据判别模型,都会基于与其对应的特征数据给出一个年龄判别结果。
在对衍生人脸数据进行年龄判别的过程中,从一衍生人脸数据中提取出关于五官的特征数据,并可将其定义为“五官”,则在年龄判别模型中,会有一针对关于五官的特征数据进行年龄判别的特征数据判别模型“判别模型A”。当特征数据“五官”需要进行年龄判别时,系统会默认让“判别模型A”对其进行年龄判别,并获得一个与“五官”对应年龄数据“年龄A”。通常,除了上述针对五官进行判别后获得的年龄数据以外,还可以获得例如对轮廓、皮肤、表情、发型等特征数据类型判别后得出的年龄数据。当获得年龄判别模型中所有特征数据判别模型所判别得出的年龄数据后,会对这些数据进行汇总计算,以获得整体的关于人脸数据的年龄数据。
第一判断单元604包括:
第一判断子单元6041,所述第一判断子单元6041用于判断每一所述衍生人脸数据经年龄判别后获得的年龄分布区间与第一参考年龄区间之间的相关度是否高于预设阈值。
此处通过相关度的高低来评判两者的匹配程度,当然,相关度的高低与否,可具体根据算法来定,例如采用两者区间的重合部分占整个区间范围的比值作为相关度的比较参数,则当两者区间的重合部分占整个区间范围的比值大于某预设阈值时,则可以认为两者匹配,当然,也可以利用生成的年龄分布区间的中值以及第一参考年龄区间的中值进行差值计算,结合其他因素来判断两者之间的相关度,当相关度大于某预设阈值时,则可以认为两者匹配。可以理解的,该相关度的计算以及设置的预设阈值大小,可以根据算法不同作出适当调整,其最终目的在于以一种数值大小来定义两者分布区间的匹配程度。
一些实施例中,所述第二获取单元605包括第二获取子单元6051以及调整子单元6052,其中:
所述第二获取子单元6051用于若所述相关度高于预设阈值,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据。
所述调整子单元6052,用于若所述相关度低于预设阈值,则调整所述年龄分布区间的区间位置,以使所述年龄分布区间最终与第一参考年龄区间匹配。
在一些实施例中,所述第二获取单元605包括第三获取子单元6053以及确认单元6054,其中:
第三获取子单元6053,用于根据所述年龄分布区间根据所述年龄判别后的结果,获取所述年龄分布区间的经所述连续分布处理后的所述第一衍生人脸数据集对应的年龄中值数据;以及
确认子单元6054,用于将所述年龄中值数据作为所述待测人脸数据对应的年龄数据.
在一些实施例中,若年龄分布区间为基于特征数据中的不同年龄的人脸特征关联规律,并以每一相邻两个衍生人脸数据之间均会按老化或幼化一定年龄数值的方式进行生成,此时如果该年龄分布区间与第一参考年龄区间匹配,则可以取该生成的年龄分布区间的两个年龄端点(最大值与最小值)之间的中值作为待测人脸数据的年龄数据。
例如,假设第一衍生人脸数据集经年龄判别后所生成的年龄分布区间为25-29,并且,该第一衍生人脸数据集中的衍生人脸数据是待测人脸数据基于不同年龄的人脸特征关联规律,并均为待测人脸数据老化或幼化一定年龄后生成,若老化的衍生人脸数据与幼化的衍生人脸数据均相同,并且相邻两个年龄对应的衍生人脸数据的年龄差均一致,则可以认为其分布区间的中值与实际年龄的匹配度应该是最高的。
当相关度较低,使得年龄区间不匹配时,此时,调整所述年龄分布区间的区间位置,以使所述年龄分布区间最终与第一参考年龄区间匹配。在一些实施例中,可以调整该年龄分布区间的数值,例如增减一定范围的值,并再次进行匹配,当满足匹配条件时即输出年龄分布区间的中值。
可以理解的,年龄分布区间表示了多个衍生人脸数据对应的年龄数据,通过年龄分布区间的表示方式,可以有效提高年龄判别过程中的容错率,并可从该年龄分布区间中获得与待测人脸数据对应的较为准确的年龄数据,提高对待测人脸数据进行年龄判别的准确率。
图8提供了一种基于生成对抗网络(GAN)来训练人脸数据生成模型,即生成器的人脸识别装置,包括第三获取单元606、第二生成单元607、第二判别单元608、第二判断单元609以及更新单元610,其中:
第三获取单元606,用于获取参考人脸数据。
第二生成单元607,用于根据所述参考人脸数据,生成与所述参考人脸数据相关的第二衍生人脸数据集,其中,所述第二衍生人脸数据集包括多个不同的衍生人脸数据。
第二判别单元608,用于将所述第二衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第二衍生人脸数据集对应的年龄分布区间。
第二判断单元609,用于判断所述第二衍生人脸数据集对应的年龄分布区间是否与第二参考年龄区间匹配。
更新单元610,用于若否,更新所述人脸数据生成模型对应的模型参数,直至所述第二衍生人脸数据集对应的年龄分布区间,经所述年龄判别后均与所述第二参考年龄区间匹配。
在一些实施例中,该参考人脸数据可以是从AFAD数据库、AFLW数据库或者是其他一些商用人脸数据库中获取,上述人脸数据库提供足够的参考人脸数据,使得参考信息的可置信度足够高。
可以理解的,该第二衍生人脸数据集的生成过程可参照图1中的步骤S102,其生成原理及过程大致相同,最大的区别仅仅是将待测人脸数据替换成人脸数据库中的参考人脸数据,为了便于说明,故对此不再赘述。
在一些实施例中,判断所述第二衍生人脸数据集对应的年龄分布区间是否与第二参考年龄区间匹配,可以引用准确率的概念,当在人脸数据生成模型,即生成器的训练过程中多次循环匹配完成后(包括匹配成功或不成功),均会对准确率进行统计,当该准确率的数值足够高,则可以认为该生成器足够可信。例如,可以设定其某一单位时间的匹配准确率达到98%以上,当生成器在训练的过程中的匹配准确率达到上述数值,则可以暂停训练。
当然,当生成的第二衍生人脸数据集对应的年龄分布区间与第二参考年龄区间的匹配准确率低于预设的匹配准确率值时,可以推断出该生成器基于人脸数据生成衍生人脸数据过程的模型参数置信度不足,此时可以基于第二参考年龄区间进行参数调整,以使再次生成的衍生人脸数据可与第二参考年龄区间匹配,并通过对不同的参考人脸数据不断进行如获取参照人脸数据—生成多个衍生人脸数据—不匹配—更新参数—再次生成、匹配的生成对抗网络(GAN)训练过程,以不断对生成器进行生成对抗网络(GAN)生成训练。并且,利用生成多个衍生人脸数据来进行匹配,可以有效提高训练样本数量,由此可提高生成器在训练过程中的训练效率,获得更好的训练效果,使生成器达到较高的置信度。
参见图9,该图示出了第一判别单元603的结构,包括第一获取子单元6032、提取子单元6033以及第一生成子单元6034,其中:
第一获取子单元6032,用于第三获取单元606获取参考人脸数据,其中,所述参考人脸数据包括年龄数据;
提取子单元6033,用于第二生成单元607根据所述参考人脸数据,提取与所述参考人脸数据对应的特征数据;
第一生成子单元6034,用于将所述特征数据和与所述特征数据对应的年龄数据进行关联,生成所述年龄判别模型。
在上述步骤中,将特征数据和与特征数据对应的年龄数据进行关联,可以形成特征数据—年龄数据关系的模型,使得年龄判别模型通过获取不同的特征数据即可根据上述关系判别出对应的年龄数据。当然,具体实现时,其如何关联可以根据具体情况而定。
在一些实施例中,在提取特征数据的时候,可以让设备利用深度学习和卷积神经网络从人脸数据中获取,以提高所获得的特征数据的准确度。
在一些实施例中,所述年龄判别模型由多个特征数据判别模型按预设的权重系数进行关联集合而成。因年龄判别模型对每个特征数据的判别准确度以及每个特征数据对脸型整体判别的影响大小均有所不同,具体进行权重配比的时候,可以根据上述不同进行权重调整,以使年龄判别模型所能获得的年龄数据的准确度更高。
最后,将多个获得的对应不同衍生人脸数据的年龄数据,形成年龄分布区间。例如,当一衍生人脸数据获得在这第一衍生人脸数据集中最大的年龄数据结果,为30岁,而另一衍生数据获得在这第一衍生人脸数据集中最小的年龄数据结果,为25岁,则基于上述两端点的年龄数据结果,可以知道该第一衍生人脸数据集对应的年龄分布区间为25-30岁。
可以理解的,衍生人脸数据的数量以及区间的划分方式,可以根据实际应用情景进行调整。同时,该步骤可以有效拓展了与待测人脸数据相关的用于检测的衍生人脸数据,并基于上述生成的衍生人脸数据来提高人脸年龄识别的容错率,从而提高判别的准确度。
总而言之,基于上述实施例,利用生成对抗网络(GAN)获取人脸数据生成模型,即生成器,具体为基于参考人脸数据,利用人脸数据生成模型获取第二衍生人脸数据集,并利用上述第二参考年龄区间对第二衍生人脸数据集的准确度进行判别,以优化该人脸数据生成模型,使该人脸数据生成模型可置信。然后,利用该生成器,通过第一生成单元生成与待测人脸数据对应的第一衍生人脸数据集;然后第一判别单元通过经训练的年龄判别模型,对该第一衍生人脸数据集中的多个衍生人脸数据进行特征数据的年龄判别;最后第一判断单元通过年龄判别后将年龄数据汇总生成年龄分布区间,并在其中获得与待测人脸数据匹配度较高的年龄数据。如此可有效解决在判别人脸数据的年龄的过程中,因可能出现的图像角度不佳、过曝或过暗导致的识别准确率不高的问题,提高年龄判别过程中算法的容错率,以使人脸年龄识别算法可动态适配多种不同环境,大幅提高人脸年龄识别的准确度。
在一些实施例中,本发明还提供一种电子设备,该电子设备可以是智能手机、平板电脑、智能手表等设备。
图10所示,在一些实施例中,电子设备700包括处理器701、存储器702。其中,所述处理器701与所述存储器702电性连接,并可控制存储器702的读写。
处理器701是电子设备700的控制中心,利用各种接口和线路连接整个电子设备700的各个部分,通过运行或加载存储在存储器702内的应用程序,以及调用存储在存储器702内的数据,执行电子设备700的各种功能和处理数据,从而对电子设备700进行整体监控。
在本实施例中,电子设备700中的处理器701会按照如下的步骤,将一个或一个以上的应用程序的进程对应的指令加载到存储器702中,并由处理器701来运行存储在存储器702中的应用程序,从而实现各种功能:
获取待测人脸数据;
根据所述待测人脸数据,生成与所述待测人脸数据相关的第一衍生人脸数据集,其中,所述第一衍生人脸数据集包括多个不同的衍生人脸数据;
将所述第一衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间;
判断所述年龄分布区间是否与第一参考年龄区间匹配;
若是,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据。
在一些实施例中,所述处理器701还用于执行以下步骤:
所述根据所述待测人脸数据,生成与所述待测人脸数据相关的第一衍生人脸数据集,包括:
根据所述待测人脸数据,通过预设的人脸数据生成模型生成与所述待测人脸数据相关的第一衍生人脸数据集。
在一些实施例中,所述处理器701还用于执行以下步骤:
在所述将所述测人脸数据生成与所述待测人脸数据相关的第一衍生人脸数据集之前,还包括:
获取参考人脸数据;
根据所述参考人脸数据,生成与所述参考人脸数据相关的第二衍生人脸数据集,其中,所述第二衍生人脸数据集包括多个不同的衍生人脸数据;
将所述第二衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第二衍生人脸数据集对应的年龄分布区间;
判断所述第二衍生人脸数据集对应的年龄分布区间是否与第二参考年龄区间匹配;
若否,更新所述人脸数据生成模型对应的模型参数,直至所述第二衍生人脸数据集对应的年龄分布区间,经所述年龄判别后均与所述第二参考年龄区间匹配。
在一些实施例中,所述处理器701还用于执行以下步骤:
将所述多个衍生人脸数据通过预设的年龄判别模型分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间;
其中,所述年龄判别模型用于根据从所述衍生人脸数据提取的特征数据,判别出与所述特征数据对应的年龄数据。
在一些实施例中,所述年龄判别模型包括多个特征数据判别模型,
其中,一所述特征数据判别模型用于判别一预设种类的特征数据,并根据所述特征数据获得与所述预设种类的特征数据对应的年龄数据。
在一些实施例中,在所述将所述第一衍生人脸数据集中的衍生人脸数据通过预设的年龄判别模型分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间之前,还包括:
获取参考人脸数据,其中,所述参考人脸数据包括年龄数据;
根据所述参考人脸数据,提取与所述参考人脸数据对应的特征数据;
将所述特征数据和与所述特征数据对应的年龄数据进行关联,生成所述年龄判别模型。
在一些实施例中,所述特征数据利用深度学习和卷积神经网络从人脸数据中获取。
在一些实施例中,所述年龄判别模型由多个特征数据判别模型按预设的权重系数进行关联集合而成。
在一些实施例中,所述处理器701还用于执行以下步骤:
判断每一所述衍生人脸数据经年龄判别后获得的年龄分布区间,与第一参考年龄区间之间的相关度是否高于预设阈值;
所述若是,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据,包括:
若所述相关度高于预设阈值,则根据所述年龄分布区间,获取所述待测人脸数据对应的年龄数据。
在一些实施例中,所述处理器701还用于执行以下步骤:
将所述待测人脸数据按预设的不同年龄段进行人脸老化处理和人脸幼化处理,生成与所述待测人脸数据相关的多个代表不同年龄的衍生人脸数据;
将所述多个衍生人脸数据按与之对应的年龄的大小进行连续分布处理,形成所述第一衍生人脸数据集。
在一些实施例中,所述处理器701还用于执行以下步骤:
根据所述年龄分布区间,获取所述年龄分布区间的年龄中值数据;
将所述年龄中值数据作为所述待测人脸数据对应的年龄数据。
存储器702可用于存储应用程序和数据。存储器702存储的应用程序中包含有可在处理器701中执行的指令。应用程序可以组成各种功能单元。处理器701通过运行存储在存储器702的应用程序,从而执行各种功能应用以及数据处理。
如图11所示,在一些实施例中,电子设备700还包括:显示屏703、控制电路704、射频电路705、输入单元706、音频电路707、传感器708以及电源709。其中,处理器701分别与射频电路705、输入单元706、音频电路707、以及电源709电性连接。
显示屏703可用于显示由用户输入的信息或提供给用户的信息以及电子设备的各种图形用户接口,这些图形用户接口可以由图像、文本、图标、视频和其任意组合来构成。
控制电路704与显示屏703电性连接,用于控制显示屏703显示信息。
射频电路705用于收发射频信号,以通过无线通信与网络设备或其他电子设备建立无线通讯,与网络设备或其他电子设备之间收发信号。
输入单元706可用于接收输入的数字、字符信息或用户特征信息(例如指纹),以及产生与用户设置以及功能控制有关的键盘、鼠标、操作杆、光学或者轨迹球信号输入。其中,输入单元706可以包括指纹识别模组。
音频电路707可通过扬声器、传声器提供用户与电子设备之间的音频接口。
电子设备700还可以包括至少一种传感器708,比如光传感器、运动传感器以及其他传感器。
电源709用于给电子设备700的各个部件供电。在一些实施例中,电源709可以通过电源管理系统与处理器701逻辑相连,从而通过电源管理系统实现管理充电、放电、以及功耗管理等功能。
尽管图11中未示出,电子设备700还可以包括摄像头、蓝牙单元等,在此不再赘述。
本发明实施例还提供了一种存储介质,其中,所述存储介质中存储有多条指令,所述指令适于由处理器加载以执行如上所述的方法。
本发明的电子设备,利用生成对抗网络(GAN)获取人脸数据生成模型,即生成器,具体为基于参考人脸数据,利用人脸数据生成模型获取第二衍生人脸数据集,并利用上述第二参考年龄区间对第二衍生人脸数据集的准确度进行判别,以优化该人脸数据生成模型,使该人脸数据生成模型可置信。然后,利用该人脸数据生成模型,生成与待测人脸数据对应的第一衍生人脸数据集;然后通过经训练的年龄判别模型,对该第一衍生人脸数据集中的多个衍生人脸数据进行特征数据的年龄判别;最后通过年龄判别后将年龄数据汇总生成年龄分布区间,并在其中获得与待测人脸数据匹配度较高的年龄数据。如此可有效解决在判别人脸数据的年龄的过程中,因可能出现的图像角度不佳、过曝或过暗导致的识别准确率不高的问题,提高年龄判别过程中算法的容错率,以使人脸年龄识别算法可动态适配多种不同环境,大幅提高人脸年龄识别的准确度。
需要说明的是,本领域普通技术人员可以理解上述实施例的各种方法中的全部或部分步骤是可以通过程序来指令相关的硬件来完成,该程序可以存储于计算机可读的存储介质中,如存储在电子设备的存储器中,并被该电子设备内的至少一个处理器执行,在执行过程中可包括上述的实施例的流程。其中,存储介质可以包括:只读存储器(ROM,Read Only Memory)、随机存取记忆体(RAM,Random Access Memory)、磁盘或光盘等。
综上所述,虽然本发明已以优选实施例揭露如上,但上述优选实施例并非用以限制本发明,本领域的普通技术人员,在不脱离本发明的精神和范围内,均可作各种更动与润饰,因此本发明的保护范围以权利要求界定的范围为准。

Claims (22)

  1. 一种人脸识别方法,其中,包括以下步骤:
    获取待测人脸数据;
    根据所述待测人脸数据,生成与所述待测人脸数据相关的第一衍生人脸数据集,其中,所述第一衍生人脸数据集包括多个不同的衍生人脸数据;
    将所述第一衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间;
    判断所述年龄分布区间是否与第一参考年龄区间匹配;
    若是,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据。
  2. 如权利要求1所述的人脸识别方法,其中,所述待测人脸数据基于预设的人脸数据生成模型生成所述第一衍生人脸数据集,在所述将所述测人脸数据生成与所述待测人脸数据相关的第一衍生人脸数据集之前,还包括:
    获取参考人脸数据;
    根据所述参考人脸数据,生成与所述参考人脸数据相关的第二衍生人脸数据集,其中,所述第二衍生人脸数据集包括多个不同的衍生人脸数据;
    将所述第二衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第二衍生人脸数据集对应的年龄分布区间;
    判断所述第二衍生人脸数据集对应的年龄分布区间是否与第二参考年龄区间匹配;
    若否,更新所述人脸数据生成模型对应的模型参数,直至所述第二衍生人脸数据集对应的年龄分布区间,经所述年龄判别后均与所述第二参考年龄区间匹配。
  3. 如权利要求1所述的人脸识别方法,其中,所述将所述第一衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间,包括:
    将所述多个衍生人脸数据通过预设的年龄判别模型分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间;
    其中,所述年龄判别模型用于根据从所述衍生人脸数据提取的特征数据,判别出与所述特征数据对应的年龄数据。
  4. 如权利要求3所述的人脸识别方法,其中,所述年龄判别模型包括多个特征数据判别模型,
    其中,一所述特征数据判别模型用于判别一预设种类的特征数据,并根据所述特征数据获得与所述预设种类的特征数据对应的年龄数据。
  5. 如权利要求3所述的人脸识别方法,其中,在所述将所述第一衍生人脸数据集中的衍生人脸数据通过预设的年龄判别模型分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间之前,还包括:
    获取参考人脸数据,其中,所述参考人脸数据包括年龄数据;
    根据所述参考人脸数据,提取与所述参考人脸数据对应的特征数据;
    将所述特征数据和与所述特征数据对应的年龄数据进行关联,生成所述年龄判别模型。
  6. 如权利要求3-5所述的人脸识别方法,其中,所述特征数据利用深度学习和卷积神经网络从人脸数据中获取。
  7. 如权利要求3-5所述的人脸识别方法,其中,所述年龄判别模型由多个特征数据判别模型按预设的权重系数进行关联集合而成。
  8. 如权利要求1所述的人脸识别方法,其中,所述判断所述年龄分布区间是否与第一参考年龄区间匹配,包括:
    判断每一所述衍生人脸数据经年龄判别后获得的年龄分布区间,与第一参考年龄区间之间的相关度是否高于预设阈值;
    所述若是,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据,包括:
    若所述相关度高于预设阈值,则根据所述年龄分布区间,获取所述待测人脸数据对应的年龄数据;
    若所述相关度低于预设阈值,则调整所述年龄分布区间的区间位置,以使所述年龄分布区间最终与第一参考年龄区间匹配。
  9. 如权利要求1所述的人脸识别方法,其中,所述根据所述待测人脸数据,生成与所述待测人脸数据相关的第一衍生人脸数据集,包括:
    将所述待测人脸数据按预设的不同年龄段进行人脸老化处理和人脸幼化处理,生成与所述待测人脸数据相关的多个代表不同年龄的衍生人脸数据;
    将所述多个衍生人脸数据按与之对应的年龄的大小进行连续分布处理,形成所述第一衍生人脸数据集。
  10. 如权利要求9所述的人脸识别方法,其中,所述若是,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据,包括:
    根据所述年龄分布区间,获取所述年龄分布区间的年龄中值数据;
    将所述年龄中值数据作为所述待测人脸数据对应的年龄数据。
  11. 一种人脸识别装置,其中,所述人脸识别装置包括:
    第一获取单元,用于获取待测人脸数据;
    第一生成单元,用于根据所述待测人脸数据,生成与所述待测人脸数据相关的第一衍生人脸数据集,其中,所述第一衍生人脸数据集包括多个不同的衍生人脸数据;
    第一判别单元,用于将所述第一衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间;
    第一判断单元,用于判断所述年龄分布区间是否与第一参考年龄区间匹配;以及
    第二获取单元,用于若是,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据。
  12. 如权利要求11所述的人脸识别装置,其中,所述待测人脸数据基于预设的人脸数据生成模型生成所述第一衍生人脸数据集,所述人脸识别装置还包括:
    第三获取单元,用于获取参考人脸数据;
    第二生成单元,用于根据所述参考人脸数据,生成与所述参考人脸数据相关的第二衍生人脸数据集,其中,所述第二衍生人脸数据集包括多个不同的衍生人脸数据;
    第二判别单元,用于将所述第二衍生人脸数据集中的衍生人脸数据分别进行年龄判别,生成与所述第二衍生人脸数据集对应的年龄分布区间;
    第二判断单元,用于判断所述第二衍生人脸数据集对应的年龄分布区间是否与第二参考年龄区间匹配;以及
    更新单元,用于若否,更新所述人脸数据生成模型对应的模型参数,直至所述第二衍生人脸数据集对应的年龄分布区间,经所述年龄判别后均与所述第二参考年龄区间匹配。
  13. 如权利要求11所述的人脸识别装置,其中,所述第一判别单元,包括:
    第一判别子单元,用于将所述多个衍生人脸数据通过预设的年龄判别模型分别进行年龄判别,生成与所述第一衍生人脸数据集对应的年龄分布区间;
    其中,所述年龄判别模型用于根据从所述衍生人脸数据提取的特征数据,判别出与所述特征数据对应的年龄数据。
  14. 如权利要求13所述的人脸识别方法,其中,所述年龄判别模型包括多个特征数据判别模型,
    其中,一所述特征数据判别模型用于判别一预设种类的特征数据,并根据所述特征数据获得与所述预设种类的特征数据对应的年龄数据。
  15. 如权利要求13所述的人脸识别装置,其中,所述第一判别单元,还包括:
    第一获取子单元,用于获取参考人脸数据,其中,所述参考人脸数据包括年龄数据;
    提取子单元,用于根据所述参考人脸数据,提取与所述参考人脸数据对应的特征数据;以及
    第一生成子单元,用于将所述特征数据和与所述特征数据对应的年龄数据进行关联,生成所述年龄判别模型。
  16. 如权利要求13至15任意一项所述的人脸识别装置,其中,所述特征数据利用深度学习和卷积神经网络从人脸数据中获取。
  17. 如权利要求13至15任意一项所述的人脸识别装置,其中,所述年龄判别模型由多个特征数据判别模型按预设的权重系数进行关联集合而成。
  18. 如权利要求11所述的人脸识别装置,其中,所述第一判断单元,包括:
    第一判断子单元,用于判断每一所述衍生人脸数据经年龄判别后获得的年龄分布区间,与第一参考年龄区间之间的相关度是否高于预设阈值;
    所述第二获取单元,包括:
    第二获取子单元,用于若所述相关度高于预设阈值,则根据所述年龄分布区间,获得所述待测人脸数据对应的年龄数据;以及
    调整子单元,用于若所述相关度低于预设阈值,则调整所述年龄分布区间的区间位置,以使所述年龄分布区间最终与第一参考年龄区间匹配。
  19. 如权利要求11所述的人脸识别装置,其中,所述第一生成单元,包括:
    第二生成子单元,用于将所述待测人脸数据按预设的不同年龄段进行人脸老化处理和人脸幼化处理,生成与所述待测人脸数据相关的多个代表不同年龄的衍生人脸数据;以及
    第二生成子单元,用于将所述多个衍生人脸数据按与之对应的年龄的大小进行连续分布处理,形成所述第一衍生人脸数据集。
  20. 如权利要求19所述的人脸识别装置,其中,所述第二获取单元,包括:
    第三获取子单元,用于根据所述年龄分布区间根据所述年龄判别后的结果,获取所述年龄分布区间的经所述连续分布处理后的所述第一衍生人脸数据集对应的年龄中值数据;以及
    确认子单元,用于将所述年龄中值数据作为所述待测人脸数据对应的年龄数据。
  21. 一种存储介质,其中,所述存储介质中存储有多条指令,所述指令适于由处理器加载以执行如权利要求1-10所述的人脸识别方法。
  22. 一种电子设备,其中,包括处理器、存储器,所述处理器与所述存储器电性连接,所述存储器用于存储指令和数据,所述处理器用于执行如权利要求1-10所述的人脸识别方法。
PCT/CN2017/091378 2017-06-30 2017-06-30 人脸识别方法、装置、存储介质及电子设备 WO2019000466A1 (zh)

Priority Applications (4)

Application Number Priority Date Filing Date Title
PCT/CN2017/091378 WO2019000466A1 (zh) 2017-06-30 2017-06-30 人脸识别方法、装置、存储介质及电子设备
CN201780092003.4A CN110741387B (zh) 2017-06-30 2017-06-30 人脸识别方法、装置、存储介质及电子设备
EP17915349.9A EP3648008A4 (en) 2017-06-30 2017-06-30 FACIAL RECOGNITION METHOD AND APPARATUS, STORAGE MEDIUM, AND ELECTRONIC DEVICE
US16/730,571 US11380131B2 (en) 2017-06-30 2019-12-30 Method and device for face recognition, storage medium, and electronic device

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2017/091378 WO2019000466A1 (zh) 2017-06-30 2017-06-30 人脸识别方法、装置、存储介质及电子设备

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US16/730,571 Continuation US11380131B2 (en) 2017-06-30 2019-12-30 Method and device for face recognition, storage medium, and electronic device

Publications (1)

Publication Number Publication Date
WO2019000466A1 true WO2019000466A1 (zh) 2019-01-03

Family

ID=64740913

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/091378 WO2019000466A1 (zh) 2017-06-30 2017-06-30 人脸识别方法、装置、存储介质及电子设备

Country Status (4)

Country Link
US (1) US11380131B2 (zh)
EP (1) EP3648008A4 (zh)
CN (1) CN110741387B (zh)
WO (1) WO2019000466A1 (zh)

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110741387B (zh) * 2017-06-30 2023-07-25 Oppo广东移动通信有限公司 人脸识别方法、装置、存储介质及电子设备
CN109598176A (zh) * 2017-09-30 2019-04-09 佳能株式会社 识别装置和识别方法
WO2020213638A1 (ja) * 2019-04-18 2020-10-22 日本電気株式会社 人物特定装置、人物特定方法および記録媒体
US12054177B2 (en) * 2020-07-23 2024-08-06 Autobrains Technologies Ltd Child forward collision warning
US11366983B2 (en) * 2020-09-09 2022-06-21 International Business Machines Corporation Study-level multi-view processing system
CN113592696A (zh) * 2021-08-12 2021-11-02 支付宝(杭州)信息技术有限公司 加密模型训练、图像加密和加密人脸图像识别方法及装置

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140099029A1 (en) * 2012-10-05 2014-04-10 Carnegie Mellon University Face Age-Estimation and Methods, Systems, and Software Therefor
CN106295499A (zh) * 2016-07-21 2017-01-04 北京小米移动软件有限公司 年龄估计方法及装置
CN106469298A (zh) * 2016-08-31 2017-03-01 乐视控股(北京)有限公司 基于人脸图像的年龄识别方法及装置
CN106503623A (zh) * 2016-09-27 2017-03-15 中国科学院自动化研究所 基于卷积神经网络的人脸图像年龄估算方法

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2006085678A (ja) * 2004-08-16 2006-03-30 Fuji Photo Film Co Ltd 画像作成方法および装置並びにプログラム
US8861804B1 (en) * 2012-06-15 2014-10-14 Shutterfly, Inc. Assisted photo-tagging with facial recognition models
CN104143097B (zh) * 2013-05-09 2017-02-08 腾讯科技(深圳)有限公司 分类函数获取方法、人脸年龄识别方法、装置和设备
SG11201604981UA (en) * 2013-12-19 2016-07-28 Avigilon Fortress Corp System and method for identifying faces in unconstrained media
US11388541B2 (en) * 2016-01-07 2022-07-12 Noveto Systems Ltd. Audio communication system and method
CN110741387B (zh) * 2017-06-30 2023-07-25 Oppo广东移动通信有限公司 人脸识别方法、装置、存储介质及电子设备
US10956759B1 (en) * 2018-04-05 2021-03-23 Ambarella International Lp Age detection in vehicles using computer vision
CN109993125B (zh) * 2019-04-03 2022-12-23 腾讯科技(深圳)有限公司 模型训练方法、人脸识别方法、装置、设备及存储介质
KR20210077977A (ko) * 2019-12-18 2021-06-28 엘지전자 주식회사 지능형 전자 디바이스 및 이를 이용한 얼굴 인식 방법

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20140099029A1 (en) * 2012-10-05 2014-04-10 Carnegie Mellon University Face Age-Estimation and Methods, Systems, and Software Therefor
CN106295499A (zh) * 2016-07-21 2017-01-04 北京小米移动软件有限公司 年龄估计方法及装置
CN106469298A (zh) * 2016-08-31 2017-03-01 乐视控股(北京)有限公司 基于人脸图像的年龄识别方法及装置
CN106503623A (zh) * 2016-09-27 2017-03-15 中国科学院自动化研究所 基于卷积神经网络的人脸图像年龄估算方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP3648008A4 *

Also Published As

Publication number Publication date
EP3648008A1 (en) 2020-05-06
US11380131B2 (en) 2022-07-05
EP3648008A4 (en) 2020-07-08
CN110741387B (zh) 2023-07-25
CN110741387A (zh) 2020-01-31
US20200134294A1 (en) 2020-04-30

Similar Documents

Publication Publication Date Title
WO2019000466A1 (zh) 人脸识别方法、装置、存储介质及电子设备
WO2020045950A1 (en) Method, device, and system of selectively using multiple voice data receiving devices for intelligent service
WO2019231252A1 (en) Electronic device for authenticating user and operating method thereof
WO2021107506A1 (en) Electronic device for providing augmented reality service and operating method thereof
WO2021025482A1 (en) Electronic device and method for generating attestation certificate based on fused key
WO2016018004A1 (en) Method, apparatus, and system for providing translated content
WO2019045474A1 (ko) 비선형 특성을 갖는 오디오 필터를 이용하여 오디오 신호를 처리하는 방법 및 장치
WO2012005483A2 (ko) 정전 용량 변화 감지 방법, 장치 및 그 방법을 실행하는 프로그램이 기록된 기록매체, 및 그 방법을 사용한 터치 감지 방법, 장치 및 그 방법을 실행하는 프로그램이 기록된 기록매체
WO2020022578A1 (ko) 무선 충전을 이용하여 통신 채널을 제어하는 전자 장치 및 전자 장치의 동작 방법
WO2019083195A1 (en) ELECTRONIC APPARATUS AND METHOD FOR CONTROLLING TIME MEASUREMENT
WO2021162320A1 (en) Electronic device and method for operating high speed screen of electronic device
WO2018076875A1 (zh) 备份数据的同步方法、装置、存储介质、电子设备及服务器
WO2020209693A1 (ko) 인공지능 모델을 갱신하는 전자 장치, 서버 및 그 동작 방법
WO2021045463A1 (en) Electronic device and authentication method of electronic device
WO2019000467A1 (zh) 闹钟提醒方法、装置、存储介质及电子设备
WO2017034180A1 (ko) 애플리케이션 리스트를 제공하는 시스템 및 그 방법
WO2019050212A1 (ko) 지문을 인식하기 방법, 전자 장치 및 저장 매체
WO2019143081A1 (ko) 데이터 통신을 제어하는 방법 및 전자 장치
WO2018076453A1 (zh) 一种关联应用显示方法、装置及移动终端
WO2020209624A1 (en) Head mounted display device and operating method thereof
WO2022158686A1 (ko) 암호화된 정보에 기초하여, 인공지능 모델을 이용한 추론을 수행하는 전자 장치 및 그 동작 방법
WO2022075609A1 (ko) 멀티 챗봇을 이용하여 질의에 응답하는 전자 장치 및 그 제어 방법
EP3830821A1 (en) Method, device, and system of selectively using multiple voice data receiving devices for intelligent service
WO2021158058A1 (en) Method for providing filter and electronic device supporting the same
WO2019139373A1 (en) Method of providing notification and electronic device supporting same

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 17915349

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

WWE Wipo information: entry into national phase

Ref document number: 2017915349

Country of ref document: EP

ENP Entry into the national phase

Ref document number: 2017915349

Country of ref document: EP

Effective date: 20200130