WO2015078183A1

WO2015078183A1 - Image identity recognition method and related device, and identity recognition system

Info

Publication number: WO2015078183A1
Application number: PCT/CN2014/081636
Authority: WO
Inventors: 张维; 刘健庄; 李志锋
Original assignee: 华为技术有限公司
Priority date: 2013-11-29
Filing date: 2014-07-04
Publication date: 2015-06-04
Also published as: CN104680119A; CN104680119B

Abstract

An image identity recognition method, a model training method and an identity recognition system. The image identity recognition method may comprise: conducting feature extraction processing on a human face image to be recognized, so as to obtain an available comprehensive feature vector corresponding to the human face image to be recognized; based on the available comprehensive feature vector corresponding to the human face image to be recognized, calculating an identity feature vector corresponding to the human face image to be recognized; calculating the similarity between the identity feature vector corresponding to the human face image to be recognized and an identity feature vector corresponding to each sample human face image of Z sample human face images; and outputting identity information corresponding to Z1 sample human face images. The solution in the embodiments is beneficial to increasing the accuracy and generality of image identity recognition, thus meeting the requirements of more application scenarios, as far as possible.

Description

Image identification method and related device and identification system

This application claims priority to Chinese Patent Application No. 201310632582.0, entitled "Image Identification Method and Related Devices and Identification System", filed on November 29, 2013, the entire contents of which are incorporated by reference. In this application. Technical field

The present invention relates to the field of image processing technologies, and in particular, to an image identification method and related apparatus and an identification system. Background technique

Image-based face recognition has a wide range of applications in the fields of human identification, finance, security systems, and data identification. As people age, faces will change in contours, pigments and textures. For example, elderly people generally show wrinkles, pigmentation, and the like. These age-related changes make it very challenging to match faces of different ages. Figure 1 shows an example of the same person at different ages. It can be seen that the face of the same person is different in different age groups.

A popular technology is based on the generated model for face recognition. The core idea is to build a generated model to simulate the face aging process. The specific method is to compensate the face image of the test to offset it. The age difference of the reference face image in the database, and then the matching image is matched.

In the course of research and practice on the prior art, the inventors of the present invention have found that although the above-mentioned prior algorithms have achieved certain effects, they have the following limitations. First, it is very difficult to construct such a generation model because of the face. The aging process is very complicated, involving many factors such as human physiology, psychology, genetics, life and work habits, life experience, etc. It is difficult to accurately express it with a fixed aging model. Moreover, in many cases, the age compensation effect is not good, but a lot of noise is introduced, which has a negative effect on the later recognition. In addition, such algorithms require some extra information such as the age information of the face, but such information is often lacking in many applications. Therefore, the effectiveness and applicability of such algorithms are limited. Summary of the invention

Embodiments of the present invention provide an image identity recognition method, a related device, and an identity recognition system, with an expectation One step is to improve the accuracy and versatility of image identification, so as to meet the needs of more kinds of application scenarios.

A first aspect of the present invention provides an image identification method, which may include:

Performing a feature extraction process on the face image to be recognized to obtain an available integrated feature vector corresponding to the face image to be recognized, wherein the available integrated feature vector is used to describe the person corresponding to the face image to be recognized The identity factor of the identity is determined together with an age factor for describing the age of the person corresponding to the face image to be identified, wherein the identity factor and the age factor are not related to each other;

Calculating an identity feature vector corresponding to the to-be-recognized face image based on the available integrated feature vector corresponding to the to-be-identified face image;

Calculating a similarity between the identity feature vector corresponding to the face image to be identified and the identity feature vector corresponding to each sample face image in the Z sample face images, wherein the identity feature vector is determined by an identity factor, The Z is a positive integer;

And outputting the identity information corresponding to the Z1 sample face images, wherein the Z1 sample face images are a subset of the Z sample face images, and the Z1 sample face images correspond to the identity feature vectors and the The similarity of the identity feature vector corresponding to the recognized face image is greater than the identity feature vector of the sample face image other than the Z1 sample face image among the Z sample face images and The similarity of the identity feature vector corresponding to the face image to be identified, or the similarity between the identity feature vector corresponding to the Z1 sample face image and the identity feature vector corresponding to the face image to be recognized is greater than a fixed value, wherein the output identity information corresponding to the Z1 sample face images is the possible identity information corresponding to the face image to be identified.

In combination with the first aspect, in a first possible implementation manner,

Performing feature extraction processing on the face image to be recognized to obtain the available integrated feature vector corresponding to the face image to be recognized, including: preprocessing the face image to be recognized; The recognized face image performs feature extraction processing to obtain an available integrated feature vector corresponding to the face image to be recognized.

With reference to the first possible implementation manner of the first aspect, in a second possible implementation manner, the pre-processed facial image to be recognized is subjected to feature extraction processing to obtain the to-be-identified The available comprehensive feature vector corresponding to the face image includes: The original integrated feature vector is extracted from the recognized face image, and the extracted original integrated feature vector is subjected to dimensionality reduction processing to obtain an available integrated feature vector corresponding to the face image to be recognized.

With reference to the first aspect or the first possible implementation manner of the first aspect or the second possible implementation manner of the first aspect, in a third possible implementation manner, the original integrated feature vector or the available integrated feature vector Obtained based on the gradient direction histogram.

In combination with the first aspect or the first possible implementation of the first aspect or the second possible implementation of the first aspect or the third possible implementation of the first aspect, in a fourth possible implementation ,

The available comprehensive feature vector corresponding to the face image to be identified is described by an identity-age factor model, wherein the identity-age factor model is as follows:

Wherein, the ^ represents the available integrated feature vector, the ^ represents a sample feature average, the U represents an identity factor coefficient, and the V represents an age factor coefficient, and the representation is high

White noise, G 〜 Ν (0 σ ² ΐ), the X represents an identity factor, and y represents an age factor. With reference to the fourth possible implementation manner of the first aspect, in a fifth possible implementation, the calculating, by the available integrated feature vector corresponding to the to-be-identified face image, the face image to be recognized The corresponding identity feature vector includes: calculating, according to the available integrated feature vector corresponding to the face image to be recognized, an identity feature vector corresponding to the face image to be recognized:

Wherein

Representing the identity feature vector,

2 τ , ττττΤ τ

Wherein

With reference to the first aspect, or any one of the first to third possible implementation manners of the first aspect, in a sixth possible implementation, the available integrated feature vector corresponding to the to-be-recognized face image Described by the identity-age factor model,

Wherein, the identity-age factor model is as follows:

Wherein the T represents the available integrated feature vector, and the n represents the available synthesis

The total number of segments of the feature vector,

q represents the available comprehensive features corresponding to the segment q Collector vector, said

Indicates the stated

q corresponding sample feature average

a value, where the identifier represents an identity factor corresponding to the q, the representation

' q corresponds to an age factor, where ^q represents the Gaussian white noise corresponding to the t

With reference to the sixth possible implementation manner of the first aspect, in a seventh possible implementation, the calculating, by the available integrated feature vector corresponding to the to-be-identified face image, the face image to be recognized The corresponding identity feature vector includes: calculating, according to the available integrated feature vector corresponding to the face image to be recognized, an identity feature vector corresponding to the face image to be recognized:

TT

Two σ ² Ι + UU ¹ + VV

Wherein the q q q q q q

, said

Representing the identity feature vector, the identifier representing the identity feature vector corresponding to the segment q. With reference to the first aspect, or any one of the first to seventh possible implementation manners of the first aspect, in an eighth possible implementation manner, the identity feature vector corresponding to the face image to be recognized and the Z The similarity of the identity feature vector corresponding to each sample face image in the sample face image, the identity feature vector corresponding to the face image to be recognized and each sample face in the Z sample face images The cosine distance or Euclidean distance or Manhattan distance of the identity feature vector corresponding to the image is characterized.

A second aspect of the present invention provides a model training method, which may include:

Obtaining available comprehensive feature vectors corresponding to Z sample face images;

The identity-age factor model is trained by using the available integrated feature vectors corresponding to the Z sample face images to determine model parameters of the identity-age factor model;

Wherein the available comprehensive feature vector is described by an identity-age factor model,

The identity-age factor model corresponding to the available integrated feature subvector corresponding to the segment q of the T丄 is as follows:

q

Wherein

1, 2,

Wherein τ represents the available integrated feature vector, and the η represents the available synthesis

\ _η Τ

a total number of segments of the feature vector, the q representing an available integrated feature subvector corresponding to the segment q,

Indicates the stated

Corresponding sample feature average, wherein the identifier corresponding to the q corresponds to the identity factor

Wherein, the identity of the available comprehensive feature subvector corresponding to the segment q of the T丄 - age factor

Θ Two {β , U , V, σ _η ² } model parameters ^qqqqq . In conjunction with the second aspect, in a first possible implementation, the available integrated feature vector is obtained based on a gradient direction histogram.

A third aspect of the present invention provides a model training method, which may include:

The identity-age factor model is trained using the available integrated feature vectors corresponding to the Z sample face images to determine model parameters of the identity-age factor model.

Wherein, the identity-age factor model is as follows:

Wherein, the ^ represents the available integrated feature vector, the ^ represents a sample feature average, the U represents an identity factor coefficient, and the V represents an age factor coefficient, wherein the representation is high

White noise, G 〜 Ν (0 σ ² ΐ), the X represents an identity factor, and y represents an age factor, wherein the model parameters are υ, V, ^σ }.

In conjunction with the third aspect, in a first possible implementation, the available integrated feature vector is derived based on a gradient direction histogram.

A fourth aspect of the present invention provides an image identification device, which may include:

An extracting unit, configured to perform feature extraction processing on the face image to be recognized to obtain an available integrated feature vector corresponding to the face image to be recognized, wherein the available integrated feature vector is used to describe the person to be identified The identity factor of the identity of the person corresponding to the face image is determined together with an age factor for describing the age of the person corresponding to the face image to be recognized, wherein the identity factor and the age factor are not related to each other; The calculating unit calculates an identity feature vector corresponding to the to-be-recognized face image based on the available integrated feature vector corresponding to the to-be-recognized face image;

a matching unit, configured to calculate a similarity between an identity feature vector corresponding to the to-be-recognized face image and an identity feature vector corresponding to each sample face image in the Z sample face images, where the identity feature vector Determined by an identity factor, the Z is a positive integer;

An output unit, configured to output identity information corresponding to the Z1 sample face images, where the Z1 sample face images are a subset of the Z sample face images, and the matching unit calculates the Z1 The similarity between the identity feature vector corresponding to the sample face image and the identity feature vector corresponding to the face image to be recognized is greater than the Z sample face image among the Z sample face images. The similarity between the identity feature vector corresponding to the other face image of the sample and the identity feature vector corresponding to the face image to be recognized, or the identity feature vector corresponding to the Z1 sample face image and the face to be recognized The similarity of the identity feature vector corresponding to the image is greater than the set threshold, wherein the output identity information corresponding to the Z1 sample face image is the possible identity information corresponding to the face image to be identified.

In conjunction with the fourth aspect, in a first possible implementation manner,

The extracting unit is specifically configured to: perform pre-processing on the face image to be recognized; perform feature extraction processing on the pre-processed face image to obtain the available comprehensive image corresponding to the face image to be recognized Feature vector.

With reference to the first possible implementation manner of the fourth aspect, in a second possible implementation manner, performing feature extraction processing on the to-be-identified face image after the pre-processing is performed to obtain the to-be-identified The aspect of the face image corresponding to the available integrated feature vector, the extracting unit is specifically configured to: extract the original integrated feature vector from the pre-processed face image to be recognized, and extract the original integrated feature The feature vector performs a dimensionality reduction process to obtain an available integrated feature vector corresponding to the face image to be recognized.

With reference to the fourth aspect or the first possible implementation manner of the fourth aspect or the second possible implementation manner of the fourth aspect, in a third possible implementation manner, the original integrated feature vector or the available synthesis The feature vector is derived based on the gradient direction histogram.

In conjunction with the fourth aspect or the first possible implementation of the fourth aspect or the second possible implementation of the fourth aspect or the third possible implementation of the fourth aspect, in a fourth possible implementation In the mode, the available comprehensive feature vector corresponding to the face image to be identified is described by an identity-age factor model, wherein the identity-age factor model is as follows:

Wherein, the ^ represents the available integrated feature vector, the ^ represents a sample feature average, the U represents an identity factor coefficient, and the V represents an age factor coefficient, and the representation is high~

White noise, GN (VQ ' σ ² ΐ), the X represents an identity factor, and y represents an age factor.

With reference to the fourth possible implementation manner of the fourth aspect, in a fifth possible implementation, the calculating unit is specifically configured to: based on the available integrated feature vector corresponding to the to-be-recognized face image, Calculating an identity feature vector corresponding to the to-be-recognized face image:

Wherein the 1 represents the identity feature vector, where

With reference to the fourth aspect, or any one of the first to third possible implementation manners of the fourth aspect, in the sixth possible implementation manner, the available comprehensive feature corresponding to the to-be-recognized face image The vector is described by the identity-age factor model,

Wherein, the identity-age factor model is as follows:

Wherein

Wherein, the T 丄 represents the available integrated feature vector, and the n represents the available synthesis

T

The total number of segments of the feature vector,

Representing the available integrated feature sub-vector corresponding to the segment q,

Indicates the stated

q corresponding sample feature average, wherein the identifier corresponding to the q corresponds to the identity factor, the representation An age factor corresponding to q, wherein the Gaussian white noise corresponding to the q is

士 ~ Ν (0, σ ² Ι

The sound, g \ , q is combined with the sixth possible implementation of the fourth aspect. In a seventh possible implementation, the calculating unit is specifically configured to: based on the image of the face to be recognized Calculating an identity feature vector corresponding to the to-be-recognized face image by using the corresponding integrated feature vector:

its

In the q

The representation represents the identity feature vector, and the ^J q represents an identity feature vector corresponding to the segment q. With reference to the fourth aspect, or any one of the first to seventh possible implementation manners of the fourth aspect, in an eighth possible implementation manner, the identity feature vector corresponding to the to-be-recognized face image is The similarity of the identity feature vector corresponding to each sample face image in the Z sample face images, the identity feature vector corresponding to the face image to be recognized and each sample person in the Z sample face images The cosine distance or Euclidean distance or Manhattan distance of the identity feature vector corresponding to the face image is characterized. A fifth aspect of the present invention provides a model training apparatus, which may include:

An acquiring unit, configured to acquire an available integrated feature vector corresponding to the Z sample face images; a training unit, configured to use the available integrated feature vector corresponding to the Z sample face images to train the identity-age factor model to determine a model parameter of the identity-age factor model; wherein the available integrated feature vector is described by an identity-age factor model,

Wherein, the identity-age factor model is as follows:

Wherein the 1,2

Representing the available integrated feature vector, the n representing the available synthesis

The total number of segments of the feature vector,

The segmentation q corresponds to the available comprehensive

Collector vector, said

Indicates the stated

q corresponding age factor coefficient, the ^ represents the average of the sample characteristics corresponding to the q a value, where the representation represents an identity factor corresponding to the q, the representation

An age factor corresponding to q, wherein the Gaussian white noise corresponding to the q is

Shishi

Sound, ε g ~ Ν V 0,' σ q ² Ι

Wherein, the identity-age of the available comprehensive feature sub-vector corresponding to the segment q of the τ ^{丄 is} a model parameter ^{qqqqq of the} sub-model of e _Q = {β , U _q , V , a _Q ² }.

In conjunction with the fifth aspect, in a first possible implementation, the available integrated feature vector is derived based on a gradient direction histogram. A sixth aspect of the present invention provides a model training apparatus, including:

An acquiring unit, configured to obtain an available integrated feature vector corresponding to the z sample face images; a training unit, configured to use the available integrated feature vector corresponding to the z sample face images to train the identity-age factor model to determine a model parameter of the identity-age factor model, wherein the available integrated feature vector is described by an identity-age factor model,

Wherein, the identity-age factor model is as follows:

Wherein the ^ represents the available integrated feature vector, and the ^ represents the sample feature level p mean, the U represents an identity factor coefficient, the V represents an age factor coefficient, and the representation is high

White noise, G ~ N (VQ ' σ ² ΐ), the X represents an identity factor, and the y represents an age factor, wherein the model parameters are - υ, V, ^σ }. In conjunction with the sixth aspect, in a first possible implementation, the available integrated feature vector is obtained based on a gradient direction histogram.

A seventh aspect of the present invention provides an identity recognition system, including:

a client, configured to send a face image to be identified to the identity server;

The identity identification server is configured to receive the to-be-recognized face image from the client, and perform feature extraction processing on the to-be-recognized face image to obtain the face image to be recognized. An integrated feature vector, wherein the available integrated feature vector is an identity factor for describing an identity of a person corresponding to the face image to be recognized and an age for describing a person corresponding to the face image to be recognized The age factors are determined together, wherein the identity factor and the age factor are not related to each other; and the identity feature corresponding to the face image to be recognized is calculated based on the available integrated feature vector corresponding to the face image to be recognized Calculating a similarity between the identity feature vector corresponding to the face image to be identified and the identity feature vector corresponding to each sample face image in the sample face image, wherein the identity feature vector is determined by the identity factor Determining that the Ζ is a positive integer; the client is required to output the identity information corresponding to the Z1 sample face images, The Z1 sample face image is a subset of the sample face images, and the identity feature vector corresponding to the Z1 sample face image and the identity feature vector corresponding to the face image to be recognized a similarity, which is greater than an identity feature vector corresponding to the sample face image of the sample face image other than the Z1 sample face image and the identity feature corresponding to the face image to be recognized Similarity of vectors, or

The similarity between the identity feature vector corresponding to the Z1 sample face image and the identity feature vector corresponding to the face image to be recognized is greater than a set threshold, wherein the output Z1 sample face image corresponding to the identity information The possible identity information corresponding to the face image to be identified.

In conjunction with the seventh aspect, in a first possible implementation manner, The available comprehensive feature vector corresponding to the face image to be identified is described by an identity-age factor model, wherein the identity-age factor model is as follows:

With reference to the first possible implementation manner of the seventh aspect, in a second possible implementation, the to-be-identified face is calculated on the available integrated feature vector corresponding to the to-be-identified face image An aspect of the identity feature vector corresponding to the image, the identity recognition server is specifically configured to: calculate, according to the available integrated feature vector corresponding to the face image to be recognized, the image corresponding to the face image to be recognized Identity feature vector:

Wherein the representation of the identity feature vector, wherein

= σ ² / ₊ [/ +ν^ With reference to the seventh aspect, in a third possible implementation manner, the available comprehensive feature vector corresponding to the to-be-recognized face image is described by an identity-age factor model,

Wherein, the identity-age factor model is as follows:

Wherein

Wherein, the T 丄 represents the available integrated feature vector, and the n represents a total number of segments of the available integrated feature vector,

The available composite feature subvector corresponding to the segment q,

Indicates the stated

q corresponding sample feature average, wherein the identifier corresponding to the q corresponds to the identity factor, the representation The age factor corresponding to q, wherein the ⁴ represents the Gaussian white noise corresponding to the q

With reference to the third possible implementation manner of the seventh aspect, in a fourth possible implementation, the to-be-identified human face is calculated on the available integrated feature vector corresponding to the to-be-identified face image An aspect of the identity feature vector corresponding to the image, the identity recognition server is configured to calculate, according to the available integrated feature vector corresponding to the face image to be recognized, an identity feature corresponding to the image to be recognized vector:

The segmentation q corresponds to the identity feature vector.

An eighth aspect of the present invention provides an identification device, including:

Processor and memory;

The processor is configured to perform a feature extraction process on the face image to be recognized to obtain an available integrated feature vector corresponding to the face image to be recognized, where the available integrated feature vector is used to describe the to-be-used The identity factor of the identity of the person corresponding to the recognized face image is determined together with an age factor for describing the age of the person corresponding to the face image to be recognized, wherein the body And the age factor is not related to each other; the identity feature vector corresponding to the face image to be recognized is calculated based on the available integrated feature vector corresponding to the face image to be recognized; and the face to be recognized is calculated a similarity between the identity feature vector corresponding to the image and the identity feature vector corresponding to each sample face image in the Z sample face images, wherein the identity feature vector is determined by an identity factor, and the Z is a positive integer; The identity information corresponding to the Z1 sample face images, wherein the Z1 sample face images are a subset of the Z sample face images, and the Z1 sample face images correspond to the identity feature vector and the The similarity of the identity feature vector corresponding to the face image to be identified is greater than the identity feature vector corresponding to the sample face image other than the Z1 sample face image among the z sample face images. Determining the similarity of the identity feature vector corresponding to the recognized face image, or the identity feature vector corresponding to the Z1 sample face image and the face image pair to be recognized Similarity identity eigenvectors width greater than a set value, wherein Z1 sample identity information corresponding to the face image is output from the face to be recognized identity information corresponding to the image may be.

With reference to the eighth aspect, in a first possible implementation manner,

The processor is configured to: perform pre-processing on the face image to be recognized; perform feature extraction processing on the pre-processed face image to obtain an available comprehensive feature vector corresponding to the face image to be recognized .

In conjunction with the first possible implementation of the eighth aspect, in a second possible implementation, the processor is configured to extract an original integrated feature vector from the to-be-identified face image after performing pre-processing, Performing dimensionality reduction on the extracted original integrated feature vector to obtain an available integrated feature vector corresponding to the face image to be recognized.

With reference to the eighth aspect, or the first possible implementation manner of the eighth aspect, or the second possible implementation manner of the eighth aspect, in the third possible implementation manner, the original integrated feature vector or the available integrated feature vector Obtained based on the gradient direction histogram.

With reference to the eighth aspect or the first possible implementation manner of the eighth aspect or the second possible implementation manner of the eighth aspect or the third possible implementation manner of the eighth aspect, in a fourth possible implementation manner The available comprehensive feature vector corresponding to the face image to be identified is described by an identity-age factor model, wherein the identity-age factor model is as follows:

White noise, GN (VQ ' σ ² ΐ), the X represents an identity factor, and y represents an age factor. With reference to the fourth possible implementation manner of the eighth aspect, in a fifth possible implementation, the processor is configured to: calculate, according to the available integrated feature vector corresponding to the to-be-identified face image, by using: The identity feature vector corresponding to the recognized face image:

Wherein the 1 represents the identity feature vector, where

With reference to the eighth aspect, or any one of the first to third possible implementation manners of the eighth aspect, in the sixth possible implementation manner, the available integrated feature vector corresponding to the to-be-recognized face image Described by the identity-age factor model,

Wherein, the identity-age factor model is as follows:

Wherein

T

The total number of segments of the feature vector,

Indicates the stated

q corresponding to the average of the sample features, wherein the identity factor corresponding to the q, the age factor corresponding to the q, wherein the Gaussian white noise corresponding to the q Shi ~

Sound, ε g Ν \ 0 ,, σ q 2 Ι binding sixth possible embodiment of the eighth aspect, in the seventh possible embodiment, the processor is configured to: the following manner, based on the to-be Calculating an identity feature vector corresponding to the face image to be recognized by the available integrated feature vector corresponding to the recognized face image:

its

Medium, said

Wherein F represents the identity feature vector,

The segmentation q corresponds to the identity feature vector.

With reference to the eighth aspect, or any one of the first to the seventh possible implementation manners of the eighth aspect, in the eighth possible implementation, the identity feature vector corresponding to the to-be-recognized face image is The similarity of the identity feature vector corresponding to each sample face image in the Z sample face images, by the identity feature vector corresponding to the face image to be recognized and each sample person in the Z sample face images The cosine distance or Euclidean distance or Manhattan distance of the identity feature vector corresponding to the face image is characterized.

A ninth aspect of the present invention provides a model training device, which may include:

Processor and memory;

The processor is configured to obtain an available integrated feature vector corresponding to the Z sample face images; and use the available integrated feature vector corresponding to the Z sample face images to enter the identity-age factor model Performing line training to determine model parameters of the identity-age factor model; wherein the available integrated feature vector is described by an identity-type,

Wherein

The identity-age factor model corresponding to the available comprehensive feature subvector corresponding to the segmentation q is as follows:

its

, the l,

Wherein, the T 丄 represents the available integrated feature vector, and the n represents the available integrated T

The total number of segments of the feature vector,

Representing the available comprehensive features corresponding to the segment q

Collector vector, said

Indicates the stated

Shishi

Sound, ε g ~ Ν V 0,' σ q ² Ι

Wherein, the identity-age factor e _Q = {β , U _q , V , a _Q ² } corresponding to the available integrated feature subvector corresponding to the segment q of the τ丄

The model parameter ^{qqqqq of the submodel} .

In conjunction with the ninth aspect, in a first possible implementation of the ninth aspect, the available integrated feature vector is derived based on a gradient direction histogram. A tenth aspect of the present invention provides a model training device, including:

Processor and memory,

The processor is configured to acquire an available integrated feature vector corresponding to the z sample face images; and use the available integrated feature vector corresponding to the z sample face images to train the identity-age factor model to determine the Model parameters of the identity-age factor model,

Wherein, the identity-age factor model is as follows:

Ϋχ +Υγ + Wherein, the ^ represents the available integrated feature vector, the ^ represents a sample feature average, the U represents an identity factor coefficient, and the V represents an age factor coefficient, and the representation is high

White noise, G ~ N (VQ ' σ ² ΐ), the X represents an identity factor, and y represents an age factor, wherein the model parameters are - υ, V, ^σ }.

In conjunction with the tenth aspect, in a first possible implementation of the tenth aspect, the available integrated feature vector is obtained based on a gradient direction histogram.

The eleventh aspect of the present invention provides a computer storage medium, wherein the computer storage medium can store a program, and the program includes any one of the model training methods or the image identification method provided by the embodiment of the present invention. Part or all of the steps.

It can be seen that, in some embodiments of the present invention, the mutually unrelated identity factors and age factors are used to jointly determine the available integrated feature vectors of the face image to be recognized, and based on the available synthesis corresponding to the face image to be identified. The feature vector calculates an identity feature vector determined by the identity factor corresponding to the face image to be identified, and calculates an identity feature vector corresponding to the face image to be recognized and each sample face image in the sample face image. The similarity of the corresponding identity feature vector, and the similarity of the identity feature vector corresponding to the face image to be recognized among the sample face images satisfies the identity information corresponding to the required Z1 sample face image, The possible identity information corresponding to the face image to be identified is output. Since the mutually different identity factors and age factors are used to jointly determine the available integrated feature vectors of the face image to be recognized, it is advantageous to identify the identity-related features included in the available integrated feature vector of the face image to be recognized. Stripping out, which is beneficial to eliminate the influence of age-related features contained in the available comprehensive feature vector of the face image to be recognized on the identity recognition, thereby improving the accuracy and versatility of image identification, thereby facilitating It may meet the needs of more kinds of application scenarios. DRAWINGS

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following will be implemented The drawings used in the examples or the description of the prior art are briefly introduced. It is obvious that the drawings in the following description are only some embodiments of the present invention, and no one skilled in the art Further drawings can also be obtained from these drawings.

1 is a photograph of a face of the same person at different ages according to an embodiment of the present invention;

2 is a schematic flowchart of an image identification method according to an embodiment of the present invention; FIG. 3 is a schematic flowchart of a model training method according to an embodiment of the present invention;

4 is a schematic flow chart of another model training method according to an embodiment of the present invention;

FIG. 5 is a schematic diagram of an image identity recognition apparatus according to an embodiment of the present invention; FIG.

6 is a schematic diagram of a model training device according to an embodiment of the present invention;

FIG. 7 is a schematic diagram of another model training device according to an embodiment of the present invention; FIG.

FIG. 8 is a schematic diagram of an identity recognition system according to an embodiment of the present invention; FIG.

FIG. 9 is a schematic diagram of an identity recognition device according to an embodiment of the present invention;

FIG. 10 is a schematic diagram of a model training device according to an embodiment of the present invention; FIG.

FIG. 11 is a schematic diagram of another model training device according to an embodiment of the present invention.

detailed description

Embodiments of the present invention provide an image identification method, a related device, and an identification system, in order to further improve the accuracy and versatility of image identification, thereby satisfying the needs of more application scenarios as much as possible.

The technical solutions in the embodiments of the present invention will be clearly and completely described in the following with reference to the accompanying drawings in the embodiments of the present invention. It is an embodiment of the invention, but not all of the embodiments. All other embodiments obtained by a person of ordinary skill in the art based on the embodiments of the present invention without departing from the inventive scope should fall within the scope of the present invention.

The details are described below separately.

The terms "first", "second", "third", "fourth", etc. (if present) in the specification and claims of the present invention and the above figures are used to distinguish similar objects without being used for Describe a specific order or order. It is to be understood that the data so used may be interchanged as appropriate, such that the embodiments of the invention described herein can be implemented, for example, in a sequence other than those illustrated or described herein. In addition, the terms "including" and "having" and any variants thereof are intended to cover no Excluded, for example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include those that are not clearly listed or Other steps or units inherent to the method, product or device.

An embodiment of an image identification method according to the present invention, wherein an image identification method may include: performing feature extraction processing on a face image to be recognized to obtain an available integrated feature vector corresponding to the face image to be recognized, Wherein, the above-mentioned available integrated feature vector is jointly determined by an identity factor for describing the identity of the person corresponding to the face image to be identified, and an age factor for describing the age of the person corresponding to the face image to be recognized, wherein The identity factor and the age factor are not related to each other; the identity feature vector corresponding to the face image to be recognized is calculated based on the available integrated feature vector corresponding to the face image to be identified; and the face image to be recognized is calculated. The similarity between the identity feature vector and the identity feature vector corresponding to each sample face image in the Z sample face images, wherein the identity feature vector is determined by an identity factor, and the Z is a positive integer; and the Z1 sample faces are output. Identity information corresponding to the image, wherein, the above Z1 The sample face image is a subset of the Z sample face images, and the similarity between the identity feature vector corresponding to the Z1 sample face image and the identity feature vector corresponding to the face image to be recognized is greater than the Z samples. The similarity between the identity feature vector corresponding to the sample face image other than the above-mentioned Z1 sample face images and the identity feature vector corresponding to the face image to be recognized among the face images, or the Z1 sample face The similarity between the identity feature vector corresponding to the image and the identity feature vector corresponding to the face image to be identified is greater than a set threshold, wherein the output of the Z1 sample face image corresponding to the identity information is the face to be recognized Possible identity information corresponding to the image.

Referring to FIG. 2, FIG. 2 is a schematic flowchart diagram of an image identity recognition method according to an embodiment of the present invention. As shown in FIG. 2, an image identification method provided by an embodiment of the present invention may include the following contents:

201. Perform feature extraction processing on the face image to be recognized to obtain an available integrated feature vector corresponding to the face image to be identified, where the available integrated feature vector is used to describe the person corresponding to the face image to be recognized. The identity factor of the identity is determined together with an age factor for describing the age of the person corresponding to the face image to be identified, wherein the identity factor and the age factor are not related to each other.

Among them, the face image to be recognized can be specially processed through various possible feature extraction processing methods. The extraction process is performed to obtain an available integrated feature vector corresponding to the face image to be identified. In some possible application scenarios of the present invention, the above-mentioned available integrated feature vector can be obtained, for example, based on a gradient direction histogram or other means.

In some possible application scenarios of the present invention, performing feature extraction processing on the face image to be recognized to obtain the available integrated feature vector corresponding to the face image to be identified may include: preprocessing the face image to be recognized ( The pre-processing may include geometric correction, trimming, and/or normalization processing, etc.; performing feature extraction processing on the face image to be recognized after the pre-processing to obtain the available integrated feature vector corresponding to the face image to be recognized . Of course, if the obtained face image to be recognized has already met the relevant requirements for direct feature extraction, the step of pre-processing the face image to be recognized may also be omitted. In some embodiments of the invention, the original integrated feature vector may be derived, for example, based on a gradient direction histogram or based on other means.

In some possible application scenarios of the present invention, performing the feature extraction process on the pre-processed face image to obtain the available integrated feature vector corresponding to the face image to be recognized may include: The original integrated feature vector is extracted from the processed face image to be recognized, and the extracted original integrated feature vector is subjected to dimensionality reduction processing to obtain an available integrated feature vector corresponding to the face image to be recognized. The manner of the dimensionality reduction processing may be, for example, a dimensionality reduction processing method of the PCA+LDA. It can be understood that the main purpose of the dimensionality reduction processing is to reduce the computational complexity. If there is sufficient computing power to support, of course, the step of performing the dimensionality reduction processing on the extracted original integrated feature vector may not be performed, for example, the extraction may be directly performed. The original integrated feature vector is used as the available integrated feature vector corresponding to the face image to be identified.

202. Calculate an identity feature vector corresponding to the to-be-identified face image based on the available integrated feature vector corresponding to the face image to be identified.

203. Calculate a similarity between the identity feature vector corresponding to the face image to be identified and the identity feature vector corresponding to each sample face image in the Z sample face images, where the identity feature vector is determined by an identity factor. The above Z is a positive integer.

204. Output identity information corresponding to the Z1 sample face images (where the identity information corresponding to each sample face image in the Z1 sample face images is used to indicate the identity of the person corresponding to the sample face image. Any information (such as name, ID number, ID card picture, passport number and/or passport picture, etc.) may even be the sample face image itself (if the sample face image is used to indicate this The identity of the person corresponding to the sample face image)), wherein the Z1 sample face image is a subset of the Z sample face images, and the identity feature vector corresponding to the Z1 sample face image and the to-be-identified The similarity of the identity feature vector corresponding to the face image is greater than the identity feature vector corresponding to the sample face image other than the Z1 sample face image among the Z sample face images and the face to be recognized The similarity of the identity feature vector corresponding to the image, or the similarity between the identity feature vector corresponding to the Z1 sample face image and the identity feature vector corresponding to the face image to be recognized is greater than a set threshold. The identity information corresponding to the Z1 sample face images outputted above is the possible identity information corresponding to the face image to be identified.

In some possible application scenarios of the present invention, the available integrated feature vector corresponding to the face image to be identified is described by an identity-age factor model.

For reasons of simplicity and stability, the identity-age factor model can be expressed, for example, using a linear model. Moreover, the age feature vector and the identity feature vector can be considered to be obtained by linear transformation from the age factor and the identity factor, respectively.

Wherein, the above identity-age factor model can be as follows:

Wherein, the above ^ represents the above-mentioned available integrated feature vector, the above ^ represents the sample feature average value, the above U represents the identity factor coefficient, and the above V represents the age factor coefficient, and the above represents high

White noise, G ~ Ν ^I (0 σ ² ΐ), the above X^ represents the identity factor, and the above y represents the age factor. Based on the above identity-age factor model, it can be seen that any available features corresponding to any face image Vector

And noise. Among them, depending on the identity of the person corresponding to the face image, it can be considered not

V y

As the person's age changes, it can be used for character identification; depending on the age of the person corresponding to the face image, it can be used to estimate the person's age.

Wherein, the identity-age factor models describing the available integrated feature vectors T丄 corresponding to any face image have the same model parameters 9 -, P, u, v, °" }. Multiple sample face images can be used. The above-mentioned identity-age factor model can be trained by the comprehensive feature vector to determine the value of the model parameter - U, V, σ } of the identity-age factor model described above, wherein based on the above-mentioned possible application scenarios, the above is based on the above-mentioned to be identified. Calculating the identity feature vector corresponding to the face image to be identified by the available comprehensive feature vector corresponding to the face image, the method may include: calculating, according to the available integrated feature vector corresponding to the face image to be recognized, by using the following manner The identity feature vector corresponding to the recognized face image:

Among them, the above

Representing the above identity feature vector, σ 2Ί τ + , υ ττυττΤ ¹ + , Τ ΤΤ

∑ two

Wherein, the above W, wherein the predictive distribution of the identity factor *^ is as follows:

P(x

∑—丄U) Therefore,

In other possible application scenarios of the present invention, the available feature vector corresponding to the face image to be identified is described by an identity-age factor model.

Among them, the above identity-age factor model is as follows:

Among them, the above

Representing the above-mentioned available comprehensive feature vector, the above n represents the above available synthesis The total number of segments of the feature vector (ie, the available feature vector)

Divided into equal lengths or unequal lengths

η segments of degree, each segment corresponding to an available integrated feature sub-vector, including n segment pairs

The available composite feature subvectors, ie Τ include η available synthetic feature subvectors,

q

The identity factor coefficient corresponding to V t _a β. The above ^y represents the age factor coefficient corresponding to q, above ^

Represents the sample feature average corresponding to q. Where the above represents the identity corresponding to q

The above represents the corresponding age factor, wherein the above ⁴ represents

~ (

Corresponding Gaussian white noise, s. (1 0, o" ² /). where, the corresponding model

Θ two { , U , V, σ ²

The parameter ^qqqq , that is, is used to describe the available comprehensive features

The model parameter of the identity-age factor model of the available comprehensive feature subvector corresponding to the segmentation q o _q = β^ u _q , V _g , } . Among them, based on the above-identity-age factor model, it can be seen that any face image corresponding to the available comprehensive

The available integrated feature sub-vectors q corresponding to the segment q of the eigenvector T are composed of the following three parts u _a x _a vy ε: identity information ^qq , age information ^q and noise. Where ^q depends on the identity of the person corresponding to the face image, which can be considered ^qq

V y

It does not change with the age of the person and can be used for character identification. ^Qq depends on the age of the person corresponding to the face image and can be used to estimate the age of the corresponding person.

It can be understood that the above examples are described by taking the sequence of numbers from 1 to n as an example. Of course, in actual applications, each segment may be numbered out of order.

Wherein, the segment included in the available integrated feature vector T corresponding to any face image is described q

The identity-age factor models of the corresponding available feature eigenvectors q all have the same model parameters.

Θ two { , U , V, σ _η ² }

q, ^qqqq . The above-described identity-age factor model can be trained to determine the identity-age by using the available comprehensive feature vectors corresponding to the plurality of sample face images.

Θ 2, U , V, σ _η ² }

The value of the model parameter ^qqqqq of the factor model, for The values of the model parameters possessed by the identity-age factor model describing the available integrated feature vectors corresponding to each segment can be determined according to the above exemplary manner. The calculating the identity feature vector corresponding to the face image to be recognized based on the available integrated feature vector corresponding to the face image to be identified may include: based on the foregoing manner, based on the foregoing other possible application scenarios, The available feature vector corresponding to the recognized face image is used to calculate the identity feature vector corresponding to the face image to be identified:

its

Medium, above ^qq

Among them, the above

Representing the above identity feature vector,

Identity feature vector corresponding to n segments q

Among them, the predicted distribution of identity factors is as follows:

- _q ) -u _q ^T Σ ¹ U _q ) therefore,

In some possible application scenarios of the present invention, the similarity between the identity feature vector corresponding to the face image to be identified and the identity feature vector corresponding to each sample face image in the z sample face images, for example, by using the above The cosine distance or the Euclidean distance or the Manhattan distance (or the ability to characterize the similarity between the identity feature vector corresponding to the face image to be recognized and the identity feature vector corresponding to each sample face image in the Z sample face images Other parameters) to characterize.

For example, the cosine distance corresponding to two identity feature vectors can be obtained by the following formula:

― Γ—

among them,

The cosine distance of the F 丄 and the body traits to the vector quantity F 丄ⁿ . The manner in which the Euclidean distance or Manhattan distance of two identity vector vectors is obtained is not described in detail herein.

It can be seen that, in this embodiment, the mutually unrelated identity factors and age factors are used to jointly determine the available integrated feature vector of the face image to be recognized, and the above-mentioned comprehensive feature vector corresponding to the face image to be identified is used to calculate the above. Calculating an identity feature vector determined by an identity factor corresponding to the face image to be recognized, and calculating an identity feature corresponding to the face image to be recognized and an identity feature corresponding to each sample face image in the Z sample face images The similarity of the vector, and the similarity of the identity feature vector corresponding to the face image to be recognized among the Z sample face images satisfies the required Z1 The identity information corresponding to the sample face image is output as the possible identity information corresponding to the face image to be identified. Since the mutually different identity factors and age factors are used to jointly determine the available integrated feature vectors of the face image to be recognized, it is advantageous to identify the identity-related features included in the available integrated feature vector of the face image to be recognized. Stripping out, which is beneficial to eliminate the influence of age-related features contained in the available comprehensive feature vector of the face image to be recognized on the identity recognition, thereby improving the accuracy and versatility of image identification, thereby facilitating It may meet the needs of more kinds of application scenarios.

In order to better understand and implement the above image identification method, the following examples are exemplified by specific application scenarios. The above described scheme of embodiments of the present invention can be applied to a variety of applications where different face images across ages need to be verified or identified.

Among them, one possible application scenario is album management. In today's society, due to the rapid development and popularity of various digital devices (mobile phones, digital cameras, digital video cameras, tablets, etc.), users often store a large number of digital photos, most of which are used and used. People or their relatives and friends are closely related. Therefore, it is necessary to more effectively distinguish, store, manage, and retrieve these photos based on the identity of the faces appearing in the photos. This involves face recognition across ages. By identifying faces in different photos and finding photos with the same identity, photos can be stored and managed according to the recognition results.

Another possible application scenario may be for a government-related department to detect whether an applicant has multiple documents (such as a passport). In real life, there are times when a person applies for and owns more than one passport at different times with different names. So far, the relevant government departments have mainly judged whether there is such a phenomenon based on the applicant's name and ID number, but this method only prevents the gentleman from being a villain. For those who intend to use a false identity to declare multiple passports, traditional methods are difficult to solve. Therefore, a feasible idea is to use the solution proposed by the embodiment of the present invention to perform face recognition across ages. For example, the applicant for the passport uses the scheme proposed by the present invention to perform the similarity search in the standard face image database of the existing issued passport according to the passport standard photo provided by the applicant, and may find and input the most photographs in the database. Several (such as 50 or 100) photos of the image can then be further approved manually.

It can be seen that the solution of the embodiment of the present invention introduces a new stealth factor analysis model, which can significantly improve the face recognition performance across ages, and has wide application value. For example, the biggest person in the world The face age database (MORPH) has been tested on a large scale, and in some cases the recognition rate has reached the international leading level.

Several training methods for identity-age factor models are also provided below.

Referring to FIG. 3, FIG. 3 is a schematic flowchart of a model training method according to an embodiment of the present invention. As shown in FIG. 3, a model training method provided by an embodiment of the present invention may include the following contents:

301. Obtain an available integrated feature vector corresponding to the Z sample face images.

302. Train the identity-age factor model by using the available comprehensive feature vectors corresponding to the Z sample face images to determine model parameters of the identity-age factor model.

Wherein, the above available comprehensive feature vector is described by an identity-age factor model,

Among them, the above identity-age factor model is as follows:

Wherein, the above ^ represents the available integrated feature vector, the above ^ represents the sample feature average value, the U represents the identity factor coefficient, and the above V represents the age factor coefficient, and the above represents Gaussian white noise, ^G 〜 N ^V (Q σ ² ΐ

Representing the identity factor, the above y represents the age factor, where the above model parameters are - u, V, }.

Based on the above identity-age factor model, it can be seen that any available features corresponding to any face image The TU x V y vectors are composed of the following three parts: identity information, age information, and noise. Among them, depending on the identity of the person corresponding to the face image, it can be considered not

V y

In some possible implementations of the invention, the above-described available integrated feature vectors are derived based on a gradient direction histogram or based on other means.

It can be seen that the available integrated feature vector corresponding to the face image in the embodiment can be identified by identity.

- an age factor model description, wherein the above identity-age factor model is as follows:

Because of the model parameter ^- {^ U, , °" } of the identity-age factor model, the above-mentioned identity-age factor model can be trained to determine the model parameters by using the available comprehensive feature vectors corresponding to the two sample face images. , , υ, V, ^σ }, trained identity

The age factor model can lay a good foundation for the identification of any face image to be recognized. Wherein, since the available integrated feature vectors of the face image to be recognized are jointly determined by using mutually unrelated identity factors and age factors, it is advantageous to associate the identity-related features included in the available integrated feature vector of the face image to be recognized. The feature is stripped out, which is beneficial to eliminate the influence of the age-related features included in the available comprehensive feature vector of the face image to be recognized on the identity recognition, thereby facilitating the improvement of the accuracy and versatility of the image identification, thereby facilitating the image recognition. Meet as many application scenarios as possible. To facilitate a better understanding and implementation of the above model training method, the following is illustrated by some specific application scenarios.

Suppose the identity-age factor model to be trained is as follows:

Wherein, the above ^ represents the available comprehensive feature vector, the above ^ represents the average value of the sample feature, the U represents the identity factor coefficient, and the above V represents the age factor coefficient, and the above G represents high ~ Τ ί θ σ ² ΐ) X

White noise, ⁱ ' Λ above represents the identity factor, and Ύ above represents the age factor. The model parameter of the above identity-age factor model - { ' U, V, }. For example, the model parameters can be optimized by maximizing the joint probability distribution formula, wherein the maximum joint probability distribution formula can be, for example, as shown in the following formula 1:

Where, in formula 1, k represents the age of the person corresponding to the sample face image, and i represents the sample face

The identity of the person corresponding to the image,

A sample face map representing the identity i and age k Like the corresponding available comprehensive feature vector, ¹ represents the identity factor of the person corresponding to the sample face image whose identity is i, ^^ is the age factor of the person corresponding to the sample face image of age k, ^^ indicates in the given The model parameters are under the conditions of the Θ, and the joint probability distribution. Where L represents the joint probability distribution. Where, due to the two invisibility factors in Equation 1,

Can not be directly observed. For example, the factor sum can be analyzed by a coordinate ascending algorithm, that is, another invisible factor is optimized with one factor fixed. Where, for a given model parameter, the prior probability score ρ _β ( , y _k \ T) can be estimated

Cloth ⁰ maximizes the conditional expectation of joint probability distribution L to obtain prior probability distribution

, and then update the value of the model parameter Θ.

That is, given the initialization estimate, a new one is obtained by maximizing L _{C in} Equation 2 below:

Wherein, IJ and T in Equation 2 represent the available integrated feature vectors corresponding to each sample face image in the sample set (assuming there are Z sample face images), and L _C is the joint probability distribution L in the given initial model parameters. Conditional expectations. Among them, because the invisible factor and ^^ are unknown, Therefore L cannot be directly maximized. But by initializing the model parameter ^σ . To estimate the distribution of the invisibility factor 'and, and then to obtain the conditional expectation of the joint probability distribution L under the lower distribution, which is expected to be L _c .

The following is the maximum conditional expectation of the above-identification-age factor model adaptation method: τ二

Input as: Feature Vector Set of Sample Image with Identity and Age

The output is: Model parameters of the feature model 6 -, β, u, v, < } Specifically, the following parameters can be initialized first:

σ ² 0.1, brain - 0.1, 0.1)

The initialized ^σ , υ, ν are brought into the identity-age factor model formula, and the invisible factor sum is calculated based on the model parameters 9-, υ, ^ ² }.

Among them, 》 =

The N _d represents the number of sample face images whose identity is i in the training sample face image, and the N _sk represents the number of sample face images of the age of k in the training sample face image. Based on the calculated stealth factor and ^^ update model parameters - 3/: O 9ε9ϊϊ1£8/-oiAV

Wherein, N represents the total number of sample face images, and d is the length of the available integrated feature vector of the sample face image. Based on the above manner, the parameters ^σ , and V are obtained multiple times using a plurality of available integrated feature vectors in the available integrated feature vectors corresponding to the Z sample face images until convergence. Through the above algorithm, the model parameter θ = {β, 〃 σ ² } of the above identity-age factor model can be calculated more accurately. Of course, the model parameters can also be trained by other means.

The specific value of L σ }. It is not limited to the training method exemplified above.

Referring to FIG. 4, FIG. 4 is a schematic flowchart diagram of another model training method according to an embodiment of the present invention. As shown in FIG. 4, another model training method provided by an embodiment of the present invention may include the following contents: 401. Acquire an available integrated feature vector corresponding to the Z sample face images.

Of course, the identity factors and age factors in the available integrated feature vectors corresponding to the Z sample face images have been determined.

402. Train the identity-age factor model by using the available comprehensive feature vectors corresponding to the Z sample face images to determine model parameters of the identity-age factor model.

Among them, the above identity-age factor model is as follows:

Among them, the above 1,2

Representing the above-mentioned available comprehensive feature vector, the above n represents the above available synthesis

The total number of segments of the feature vector (ie,

Divided into equal lengths or unequal lengths

n segments of degree, each segment corresponding to an available integrated feature sub-vector,

Including n segment pairs

The available composite feature subvector, ie

Including n available synthetic feature subvectors), t _a T ut _a q represents the available integrated feature subvector corresponding to the segment q of 丄. The above represents q

The identity factor coefficient corresponding to V t _a β. The above represents the age factor coefficient corresponding to q, the above ^ q

Indicates the corresponding sample feature average. Where the above indicates the corresponding identity

Indicates the age factor corresponding to q, where the above represents s ~ (

Corresponding Gaussian white noise, q 0^, c^r ² /^). Where the corresponding model

6 _Q = {β , U _q , V , a _Q ² }

The parameter ^qqqq , that is, the model parameter for describing the identity-age factor model of the available integrated feature subvector corresponding to the segment q of the available integrated feature vector T

Θ q 二^^ q, U q, V q, σ _η q ² } J

Among them, based on the above-identity-age factor model, it can be seen that any face image corresponding to the available comprehensive

The available composite feature sub-vectors q corresponding to the segment q of the feature vector T are all composed of the following three parts. U _a X _a V y S

The composition is: identity information ^{4 q} , age information ^q and noise.

U X U X

Where ^q depends on the identity of the person corresponding to the face image, which can be considered ^qq

V y

It does not change with the age of the person and can be used for character identification. ^q depends on the age of the person corresponding to the face image and can be used to estimate the age of the corresponding person.

Due to

Θ 2 {β , U , V, σ _η ² } The model parameters ^q and qqq Q of the identity-age factor model. Therefore, the above-mentioned identity-age factor model can be used by using the available comprehensive feature vectors corresponding to the Z sample face images.

Θ two {β , U , V, σ _η ² } type training to determine the value of the model parameter ^qqqqq , the identity-age factor model corresponding to each segment trained can lay the identification of any face image to be identified Good foundation. Wherein, since the available integrated feature vectors of the face image to be recognized are jointly determined by using mutually unrelated identity factors and age factors, it is advantageous to associate the identity-related features included in the available integrated feature vector of the face image to be recognized. The feature is stripped out, which in turn helps to eliminate the identification The influence of the age-related features contained in the available feature vectors of the face image on the identity recognition, which in turn helps to improve the accuracy and versatility of the image identification, thereby facilitating the needs of more application scenarios as much as possible. .

In order to better understand and implement the above model training method, the following is illustrated by some specific application scenarios.

Hypothetical available feature vector

The corresponding identity to be trained - the age factor model is as follows:

its

, above l,

Representing the above-mentioned available integrated feature vector, where n represents the total number of segments of the available integrated feature vector (ie,

Divided into n segments of equal length or unequal length, each segment corresponding to one available integrated feature subvector, T丄 includes n segment pairs The available composite feature subvector, ie

Including η available synthetic feature subvectors),

q

The identity factor coefficient corresponding to V t _n β. The above represents the age factor coefficient corresponding to q, above ^

Represents the sample feature average corresponding to q. Where the above indicates the corresponding identity

Indicates the age factor corresponding to q, where the above represents s ~

Corresponding Gaussian white noise, q (0, σ ² /). Where t corresponds to the model

Θ two {β , U , V, σ _π ² }

The parameter ^qqqq , that is, the model parameter for describing the identity-age factor model of the available integrated feature subvector corresponding to the segmentation q of the available integrated feature vector T

Θ q two q, ⁷ U q, ⁷ V q, ⁷ σ _η q ² } J

For example, the model parameters can be optimized by maximizing the joint probability distribution formula, wherein the maximum joint probability distribution formula can be, for example, as shown in the following formula 3:

Wherein, in formula 2, k represents the age of the person corresponding to the sample face image, and i represents the identity of the person corresponding to the sample face image, wherein

Indicates that the identity is i and the age is k

Sample

Representation factor

Probability distribution.

Where ^ represents the joint probability distribution. Among them, the experiment can be performed on multiple sample face images corresponding to different ages and different identities. Among them, due to the two invisibility factors in the identity-age factor model and

Can not be directly observed. For example, the coordinate rising algorithm can be used to factor

He y is analyzed, that is, the other invisible factor is optimized when one factor is fixed. Chemical. Where, for a given model parameter

, can estimate the prior probability distribution

Pe _f q^,0) ^X q,i)

. Thus by maximizing the joint probability distribution L q

The condition expectation to get the prior probability distribution ^{Pe q} , ^{0) X(q )} , and then update the model parameters

That is, given the initial estimate ⁴ ', by maximizing L _{c in} Equation 4 below

Get a new q,

k

Where, in formula 4

Φ i)

The available synthetic feature subvectors corresponding to the segment q of the available synthetic feature vectors corresponding to each sample face image (assuming Z sample face images), L _C is the joint probability distribution ^ in the given initial model The conditional expectation of the parameter ( ^q , ⁰⁾ . Among them, because, the invisible factor and , ) are not β

Know, so L can't be directly maximized. But the invisible factor can be estimated by initializing the model parameters ( ^q ' ^Q )

Joint distribution, and then the joint probability distribution under the lower distribution

The condition of ^ is expected, and the condition is expected to be L _c .

The following is the maximum expected value (EM) algorithm for adapting the identity-age factor model described above: Input: Multiple segmentation of the face image with identity and age available. Segmentation in the integrated feature vector

The output is: Segment q Corresponding to the identity of the available integrated feature subvectors - Model parameters of the age factor model:

Θ q 二^^ q, U q, V q, σ _η q ² } J

Specifically, the following parameters can be initialized first:

2 0.1

U _q ^rand(-0AfiA) V ^rand(-0A,0A) Will be initialized

Brought into the identity-age factor model formula,

Model based parameters

Computational invisibility factor

among them,

~ ^{l T l}

3) 3)

T

among them,

T

C = ai + U _q U

Wherein, N _ci represents the number of sample face images whose identity is i in the training sample face image, and N _sk represents the number of sample face images of age k in the training sample face image.

Further updating the model parameters based on the calculated stealth factor

σ q ² U and q where,

U q =iC-DB ^l E)iA-FB ^l E

Wherein, the above N represents the total number of sample face images, and d is the length of the available integrated feature sub-vector corresponding to the index q of the available integrated feature vector of the sample face image.

Based on the above method, the available eigenvector vector corresponding to the segment q among the available integrated feature vectors corresponding to the Z sample face images is used to obtain the parameter ^σ多次 multiple times.

Until the convergence. The model parameter θ _η = {β , U _q , V , a _Q ² } of the identity-age factor model corresponding to the available integrated feature sub-vector corresponding to the segment q can be calculated more accurately by the above algorithm.

q ^qqqq . Of course, you can also train in other ways.

Θ ={Α, U _Q , V, σ. ² }

The specific value of the model parameter ^qqqqq . It is not limited to the training method exemplified above. Through the above-exemplified manner, the model parameters of the identity-age factor model corresponding to the available integrated feature sub-vectors corresponding to other segments may also be obtained. In order to facilitate a better understanding and implementation of the above described embodiments of the embodiments of the present invention, related apparatus for implementing the above aspects are also provided below. Referring to FIG. 5, an embodiment of the present invention provides an image identity recognition apparatus 500, which may include: The extracting unit 510, the calculating unit 520, the matching unit 530, and the output unit 540.

The extracting unit 510 is configured to perform feature extraction processing on the face image to be recognized to obtain an available integrated feature vector corresponding to the face image to be recognized, wherein the available integrated feature vector is used to describe the to-be-used The identity factor of the identity of the person corresponding to the recognized face image is determined together with an age factor for describing the age of the person corresponding to the face image to be recognized, wherein the identity factor and the age factor are not related to each other. ;

The calculating unit 520 is configured to calculate an identity feature vector corresponding to the to-be-recognized face image based on the available integrated feature vector corresponding to the to-be-identified face image;

The matching unit 530 is configured to calculate a similarity between the identity feature vector corresponding to the to-be-recognized face image and the identity feature vector corresponding to each sample face image in the Z sample face images, where the identity feature The vector is determined by an identity factor, and the Z is a positive integer;

The output unit 540 is configured to output identity information corresponding to the Z1 sample face images, where the Z1 sample face images are a subset of the Z sample face images, and the matching unit calculates the Z1 The similarity between the identity feature vector corresponding to the sample face image and the identity feature vector corresponding to the face image to be recognized is greater than the Z sample face image among the Z sample face images. The similarity between the identity feature vector corresponding to the other face image of the sample and the identity feature vector corresponding to the face image to be recognized, or the identity feature vector corresponding to the Z1 sample face image and the person to be identified The similarity of the identity feature vector corresponding to the face image is greater than the set threshold. The identity information corresponding to the output Z1 sample face image is the possible identity information corresponding to the face image to be recognized.

In some embodiments of the present invention, the extracting unit 510 may be specifically configured to: perform pre-processing on the face image to be recognized; perform feature extraction processing on the pre-processed face image to obtain the The available integrated feature vector corresponding to the recognized face image.

In some embodiments of the present invention, performing feature extraction processing on the to-be-identified face image after the pre-processing is performed to obtain an aspect of the available integrated feature vector corresponding to the face image to be recognized, and extracting The unit 510 may be specifically configured to: extract an original integrated feature vector from the to-be-identified face image after the pre-processing, perform a dimensionality reduction process on the extracted original integrated feature vector to obtain the to-be-identified person The available composite feature vector corresponding to the face image.

In some embodiments of the invention, the original integrated feature vector or the available integrated feature The amount is obtained based on the gradient direction histogram.

In some embodiments of the present invention, the available integrated feature vector corresponding to the face image to be identified is described by an identity-age factor model, wherein the identity-age factor model is as follows:

White noise, G Ν (0 σ ² ΐ), the X represents an identity factor, and y represents an age factor. In some embodiments of the present invention, the calculating unit 520 is specifically configured to calculate, according to the available integrated feature vector corresponding to the face image to be recognized, an identity feature vector corresponding to the face image to be recognized. :

Wherein the 1 represents the identity feature vector, where

σ ² / ₊ [/[/ +νν . In other embodiments of the present invention, the available comprehensive features corresponding to the face image to be identified The vector is described by the identity-age factor model,

Wherein, the identity-age factor model is as follows:

Wherein

T

The total number of segments of the feature vector,

Indicates the stated

In some embodiments of the present invention, the calculating unit 520 is specifically configured to calculate, according to the available integrated feature vector corresponding to the face image to be recognized, an identity feature vector corresponding to the face image to be recognized:

its

In the ^qq .,

The representation represents the identity feature vector, and the ^J q represents an identity feature vector corresponding to the segment q. In some embodiments of the present invention, the similarity between the identity feature vector corresponding to the face image to be recognized and the identity feature vector corresponding to each sample face image in the Z sample face images is determined by the The identity feature vector corresponding to the recognized face image is represented by a cosine distance or an Euclidean distance or a Manhattan distance of the identity feature vector corresponding to each sample face image in the Z sample face images.

It can be understood that the functions of the functional modules of the apparatus of this embodiment can be implemented according to the above method. The method in the example is specifically implemented, and the specific implementation process may refer to the related description of the foregoing method embodiment, and details are not described herein again.

It can be seen that, in this embodiment, the mutually unrelated identity factors and age factors are used to jointly determine the available integrated feature vector of the face image to be recognized, and the above-mentioned comprehensive feature vector corresponding to the face image to be identified is used to calculate the above. Calculating an identity feature vector determined by an identity factor corresponding to the face image to be recognized, and calculating an identity feature corresponding to the face image to be recognized and an identity feature corresponding to each sample face image in the Z sample face images The similarity of the vector, and the similarity of the identity feature vectors corresponding to the face image to be recognized among the Z sample face images satisfies the required identity information of the Z1 sample face images, as the above-mentioned to be identified The face image corresponds to the possible identity information for output. Since the mutually different identity factors and age factors are used to jointly determine the available integrated feature vectors of the face image to be recognized, it is advantageous to identify the identity-related features included in the available integrated feature vector of the face image to be recognized. Stripping out, which is beneficial to eliminate the influence of age-related features contained in the available comprehensive feature vector of the face image to be recognized on the identity recognition, thereby improving the accuracy and versatility of image identification, thereby facilitating It may meet the needs of more kinds of application scenarios. Referring to FIG. 6, an embodiment of the present invention further provides a model training apparatus 600, including:

The unit 610 and the training unit 620 are acquired.

The obtaining unit 610 is configured to obtain an available integrated feature vector corresponding to the Z sample face images. The training unit 620 is configured to train the identity-age factor model by using the available integrated feature vectors corresponding to the Z sample face images to determine model parameters of the identity-age factor model; wherein the available comprehensive features are The vector is described by the identity-age factor model,

Wherein, the identity-age factor model is as follows:

Wherein

1, 2,

\ _η Τ

a total number of segments of the feature vector, the ^ representing the available integrated feature subvector corresponding to the segment q,

Indicates the stated

q corresponding sample feature average, wherein the identifier corresponding to the q corresponds to the identity factor, the representation

' q corresponds to an age factor, where ^q represents the Gaussian white noise corresponding to the t Wherein

The identity-age factor corresponding to the available comprehensive feature subvector corresponding to the segment q

Θ two {β , U , V, σ _η ² }

The model parameter ^{qqqqq of the submodel} .

In some embodiments of the present invention, the available integrated feature vector is based on the gradient direction histogram. It can be understood that the functions of the functional modules of the device in this embodiment can be specifically implemented according to the method in the foregoing method embodiment. For a specific implementation process, reference may be made to the related description of the foregoing method embodiments, and details are not described herein again.

It can be seen that the available comprehensive feature vector corresponding to the face image in the embodiment can be described by the identity-age factor model, wherein the above identity-age factor model is as follows:

Due to

6 _Q = {β , U _q , V , a _Q ² } The model parameters ^q and ^qqq ^ of the identity-age factor model, so the identity-age factor model can be used for the above-mentioned identity-age factor model using the available comprehensive feature vectors corresponding to the z sample face images.

Θ two {β , U , V, σ _η ² }

The type is trained to determine the value of the model parameter ^qqqqq , and the identity-age factor model corresponding to each segment trained can lay a good foundation for the recognition of any face image to be recognized. Wherein, since the available integrated feature vectors of the face image to be recognized are jointly determined by using mutually unrelated identity factors and age factors, it is advantageous to associate the identity-related features included in the available integrated feature vector of the face image to be recognized. The feature is stripped out, which is beneficial to eliminate the influence of the age-related features contained in the available comprehensive feature vector of the face image to be recognized on the identity recognition, thereby facilitating the improvement of the accuracy and versatility of the image identification, thereby facilitating the image recognition. As much as possible More kinds of application scenarios.

Referring to FIG. 7, an embodiment of the present invention further provides another model training apparatus 700, which may include: an obtaining unit 710, configured to acquire an available integrated feature vector corresponding to a Z sample face image; and a training unit 720, configured to use the The identity-age factor model is trained by the available comprehensive feature vectors corresponding to the Z sample face images to determine model parameters of the identity-age factor model, wherein the available integrated feature vectors are described by an identity-age factor model.

Wherein, the identity-age factor model is as follows:

Because of the model parameter ^- {^ U, , °" } of the identity-age factor model, the above-mentioned identity-age factor model can be trained to determine the model parameters by using the available comprehensive feature vectors corresponding to the two sample face images. , , υ, V, ^σ }, the trained identity-age factor model can lay a good foundation for the recognition of any facial image to be recognized. Among them, because of the use of mutually unrelated identity factors and age factors Determining the available integrated feature vector of the face image to be recognized, thereby facilitating the stripping of the identity-related features included in the available integrated feature vector of the face image to be recognized, thereby facilitating the removal of the face image to be recognized. The influence of the age-related features included in the integrated feature vector on the identity recognition can be used to improve the accuracy and versatility of the image identification, and thus to meet the needs of more application scenarios as much as possible.

Referring to FIG. 8, an embodiment of the present invention further provides an identity recognition system, which may include:

Client 810 and identity server 820.

a client 810, configured to send a face image to be identified to the identity recognition server 820, where the identity recognition server 820 is configured to receive the image of the face to be recognized from the client 810, for the person to be identified The face image is subjected to feature extraction processing to obtain an available integrated feature vector corresponding to the face image to be recognized, wherein the available integrated feature vector is used to describe the identity of the person corresponding to the face image to be recognized. The factor is determined together with an age factor for describing an age of a person corresponding to the face image to be recognized, wherein the identity factor and the age factor are not related to each other; based on the face image to be recognized Calculating an identity feature vector corresponding to the face image to be recognized by using the integrated feature vector; calculating an identity feature vector corresponding to the face image to be recognized and each sample face image in the sample face image Corresponding to the similarity of the identity vector, wherein the identity feature vector is determined by an identity factor, and the Ζ is a positive integer; The client outputs Z1 sample face images corresponding to the identity information, where the Z1 sample face images are subsets of the Z sample face images, and the Z1 sample face images correspond to the identity features. The similarity between the vector and the identity feature vector corresponding to the face image to be identified is greater than the identity of the other sample face images other than the Z1 sample face images among the Z sample face images. a similarity between the feature vector and the identity feature vector corresponding to the face image to be recognized, or an identity feature vector corresponding to the Z1 sample face image and an identity feature vector corresponding to the face image to be recognized The degree is greater than the set threshold, and the identity information corresponding to the output Z1 sample face images is the possible identity information corresponding to the face image to be identified.

White noise, G 〜 Ν (0 σ ² ΐ), the X represents an identity factor, and y represents an age factor. In some embodiments of the present invention, the identity recognition server is configured to calculate an identity feature vector corresponding to the to-be-recognized face image based on the available integrated feature vector corresponding to the to-be-identified face image. Specifically, the identifier feature vector corresponding to the face image to be recognized is calculated based on the available integrated feature vector corresponding to the face image to be identified:

Wherein the representation of the identity feature vector,

Wherein

σ ² / + [/ ₊ νν

In some embodiments of the present invention, the available integrated feature vector corresponding to the face image to be identified is described by an identity-age factor model.

Wherein, the identity-age factor model is as follows:

The total number of segments of the feature vector,

Indicates the stated

q corresponding sample feature average

In some embodiments of the present invention, the identity recognition server is configured to calculate an identity feature vector corresponding to the to-be-recognized face image based on the available integrated feature vector corresponding to the to-be-identified face image. Specifically, the identifier feature vector corresponding to the to-be-recognized face image is calculated based on the available integrated feature vector corresponding to the to-be-recognized face image:

The qq

Wherein F represents the identity feature vector,

The segmentation q corresponds to the identity feature vector. It is to be understood that the functions of the functional modules of the device in this embodiment may be specifically implemented according to the method in the foregoing method embodiments. For the specific implementation process, reference may be made to the related description of the foregoing method embodiments, and details are not described herein again.

It can be seen that, in this embodiment, the identity identification server uses the mutually unrelated identity factors and age factors to jointly determine the available integrated feature vector of the face image to be recognized, and based on the available synthesis corresponding to the face image to be identified. The feature vector calculates an identity feature vector determined by the identity factor corresponding to the face image to be identified, and calculates an identity feature vector corresponding to the face image to be recognized and each sample face image in the Z sample face images The similarity of the corresponding identity feature vector, and the similarity of the identity feature vector corresponding to the face image to be recognized among the Z sample face images satisfies the identity information corresponding to the required Z 1 sample face image And outputting as the possible identity information corresponding to the face image to be identified. Since the mutually different identity factors and age factors are used to jointly determine the available integrated feature vectors of the face image to be recognized, it is advantageous to identify the identity-related features included in the available integrated feature vector of the face image to be recognized. Stripping out, which is beneficial to eliminate the influence of age-related features contained in the available comprehensive feature vector of the face image to be recognized on the identity recognition, thereby improving the accuracy and versatility of image identification, thereby facilitating It may meet the needs of more kinds of application scenarios. Referring to FIG. 9, FIG. 9 illustrates the structure of an identity recognition device 900 according to an embodiment of the present invention. The identity recognition device 900 includes: at least one processor 901, such as a CPU, at least one network interface 904 or other user interface 903, and a memory 905. At least one communication bus 902. Communication bus 902 is used to implement connection communication between these components. The identification device 900 optionally includes a user interface 903, including a display, a keyboard or a pointing device (eg, a mouse, a trackball, a touch Sensor board or touch screen display). The memory 905 may include a high speed RAM memory, and may of course also include a non-volatile memory such as at least one disk memory. The memory 905 can optionally include at least one storage device located remotely from the aforementioned processor 901.

In some embodiments, memory 905 stores the following elements, executable modules or data structures, or a subset thereof, or their extension set:

The operating system 9051, which contains various system programs for implementing various basic services and handling hardware-based tasks;

Application module 9052, which contains various applications for implementing various application services. The application module 9052 includes, but is not limited to, an extracting unit 510, a calculating unit 520, a matching unit 530, and an output unit 540.

For the specific implementation of each module in the application module 9052, refer to the corresponding modules in the embodiment shown in FIG. 5, and details are not described herein.

In some embodiments of the present invention, by calling a program or an instruction stored in the memory 905, the processor 901 may be configured to: perform feature extraction processing on the face image to be recognized to obtain an available integrated feature vector corresponding to the face image to be identified, Wherein, the above-mentioned available integrated feature vector is jointly determined by an identity factor for describing the identity of the person corresponding to the face image to be identified, and an age factor for describing the age of the person corresponding to the face image to be recognized, wherein The above identity factors are not related to the above age factors. Calculating an identity feature vector corresponding to the face image to be recognized based on the available integrated feature vector corresponding to the face image to be identified; calculating an identity feature vector corresponding to the face image to be recognized and a Z sample face image The similarity of the identity feature vector corresponding to each sample face image, wherein the identity feature vector is determined by an identity factor, and the above Z is a positive integer; and the identity information corresponding to the Z1 sample face image is output (where Z1 samples) The identity information corresponding to each sample face image in the face image is any information (such as name, ID number, ID card picture, passport number, and/or) that can be used to indicate the identity of the person corresponding to the sample face image. a passport picture, etc.), or even a sample face image itself (if the sample face image is used to indicate the identity of the person corresponding to the sample face image), wherein the Z1 sample face images are the above Z a subset of the sample face image, the identity feature vector corresponding to the Z1 sample face image and the face image to be recognized Similarity identity feature vectors is larger than in the Z-sample addition to the face images Z1 The similarity between the identity feature vector corresponding to the sample face image other than the sample face image and the identity feature vector corresponding to the face image to be recognized, or the identity feature vector corresponding to the Z1 sample face image and the above The similarity of the identity feature vector corresponding to the face image to be identified is greater than the set threshold. The identity information corresponding to the Z1 sample face images outputted above is the possible identity information corresponding to the face image to be identified.

The feature image may be extracted by the plurality of possible feature extraction processing modes to obtain the available integrated feature vector corresponding to the face image to be identified. In some possible application scenarios of the present invention, the above-described available integrated feature vector can be obtained, for example, based on a gradient direction histogram or other means.

For the sake of simplicity and stability, the identity-age factor model can use, for example, a linear model. Type is expressed. Moreover, the age feature vector and the identity feature vector can be considered to be obtained by linear transformation from the age factor and the identity factor, respectively.

Wherein, the above identity-age factor model can be as follows:

White noise, G ~ Ν ^I (0 σ ² ΐ), the above X^ represents the identity factor, and the above y represents the age factor. Based on the above identity-age factor model, it can be seen that any available features corresponding to any face image

T U X V y

They are composed of the following three parts: identity information, age information, and noise.

. Among them, depending on the identity of the person corresponding to the face image, it can be considered that it does not change with the age of the person.

Depending on the age of the person corresponding to the face image, it can be used to estimate the age of the person.

Wherein, the identity-age factor model describing the available integrated feature vector T丄 corresponding to any face image has the same model parameters 9-1, β, U, V, °" }. The above-mentioned identity-age factor model is trained by the available comprehensive feature vectors corresponding to the sample face images.

Θ2{β, U, σ ² } practice to determine the model parameters of the above identity-age factor model

The value. The calculating the identity feature vector corresponding to the to-be-recognized face image based on the available comprehensive feature vector corresponding to the face image to be identified, based on the foregoing possible application scenario, may specifically include: The available feature vector corresponding to the recognized face image is used to calculate the identity feature vector corresponding to the face image to be identified:

F II

Wherein, the above 1 indicates the above identity feature vector, wherein

Among them, the predicted distribution of the identity factor *^ is as follows:

Therefore,

In other possible application scenarios of the present invention, the available comprehensive feature vector corresponding to the face image to be identified is described by an identity-age factor model.

Among them, the above identity-age factor model is as follows:

Among them, the above

T represents the above-mentioned available integrated feature vector, and the above n represents the total number of segments of the available integrated feature vector (ie,

Divided into equal lengths or unequal lengths

n segments, each segment corresponding to an available integrated feature sub-vector, including the available integrated feature sub-vectors corresponding to n segments, ie

Including n available integrated feature sub-vectors, t _a T u \ _a q represents the available integrated feature sub-vectors corresponding to the segment q of the 丄. The above represents q

The identity factor coefficient corresponding to V t _a β. The above represents the age factor coefficient corresponding to q, the above ^ q Represents the average of the sample characteristics corresponding to 4. Wherein, the above indicates the identity corresponding to 4

Indicates the age factor corresponding to q, where

White noise, s q ~ (

Corresponding height 0, cr ² /). Where the corresponding model

Θ 2, U , V, σ _η ² }

The parameter ^qqqq , ie, is used to describe the available comprehensive features

the amount

The model parameter of the identity-age factor model of the available comprehensive feature subvector corresponding to the segmentation q

Θ, = β^ u _q , V _g , }. Among them, based on the above-identity-age factor model, it can be seen that any face image corresponding to the available comprehensive

Feature vector

All by the following three departments

U _a x _a vy ε is divided into: identity information ^qq , age information ^q and noise.

U X U X

V y

It does not change with the age of the person and can be used for character identification. ^q depends on the age of the person corresponding to the face image and can be used to estimate the age of the corresponding person. It can be understood that the above examples are described by taking the sequence of numbers from 1 to n as an example. Of course, in actual applications, each segment may be numbered out of order.

Θ two {β , U , V, σ _η ² }

q ^qqqq . The above-described identity-age factor model can be trained using the available comprehensive feature vectors corresponding to the plurality of sample face images to determine the above-identity-age e _Q = {β , U _q , V , a _Q ² }

The value of the model parameter ^qqqqq of the factor model, the value of the model parameter ^possessed by the identity-age factor model for describing the available integrated feature vector corresponding to each segment can be determined according to the above exemplary manner. The calculating the identity feature vector corresponding to the face image to be recognized based on the available integrated feature vector corresponding to the face image to be identified may include: based on the foregoing manner, based on the foregoing other possible application scenarios, The available feature vector corresponding to the recognized face image is used to calculate the identity feature vector corresponding to the face image to be identified:

Among them, the above

Representing the above identity feature vector, the above fq represents the above

The segmentation q corresponds to the identity feature vector. Where 1 consists of an identity feature vector corresponding to n segments q

X q

Among them, the predicted distribution of identity factors is as follows:

Therefore,

For example, the cosine distance corresponding to two identity feature vectors can be obtained by the following formula: ► TI ►

among them,

The cosine distance of Jn and the identity vector. The manner in which the Euclidean distance or Manhattan distance of two identity vector vectors is obtained is not described in detail herein. It can be seen that, in this embodiment, the mutually unrelated identity factors and age factors are used to jointly determine the available integrated feature vector of the face image to be recognized, and the above-mentioned comprehensive feature vector corresponding to the face image to be identified is used to calculate the above. Calculating an identity feature vector determined by an identity factor corresponding to the face image to be recognized, and calculating an identity feature corresponding to the face image to be recognized and an identity feature corresponding to each sample face image in the Z sample face images The similarity of the vector, and the similarity of the identity feature vectors corresponding to the face image to be recognized among the Z sample face images satisfies the required identity information of the Z1 sample face images, as the above-mentioned to be identified The face image corresponds to the possible identity information for output. Since the mutually different identity factors and age factors are used to jointly determine the available integrated feature vectors of the face image to be recognized, it is advantageous to identify the identity-related features included in the available integrated feature vector of the face image to be recognized. Stripping out, which is beneficial to eliminate the influence of age-related features contained in the available comprehensive feature vector of the face image to be recognized on the identity recognition, thereby improving the accuracy and versatility of image identification, thereby facilitating It may meet the needs of more kinds of application scenarios. Referring to FIG. 10, FIG. 10 illustrates a structure of a model training device 1000 according to an embodiment of the present invention. The model training device 1000 includes: at least one processor 1001, such as a CPU, at least one network interface 1004 or other user interface 1003, and a memory 1005. At least one communication bus 1002. Communication bus 1002 is used to implement connection communication between these components. Among them, model training equipment 1000 An optional user interface 1003 is included, including: a display, a keyboard or a pointing device (such as a mouse, trackball, touchpad or tactile display). The memory 1005 may include a high speed RAM memory, and may of course also include a non-volatile memory such as at least one disk memory. The memory 1005 can optionally include at least one storage device located remotely from the aforementioned processor 1001.

In some embodiments, memory 1005 stores the following elements, executable modules or data structures, or a subset thereof, or their extension set:

Operating system 10051, which contains various system programs for implementing various basic services and processing hardware-based tasks;

The application module 10052 includes various applications for implementing various application services.

The application module 10052 includes, but is not limited to, an acquisition unit 610 and a training unit 620.

For the specific implementation of each module in the application module 10052, refer to the corresponding module in the embodiment shown in FIG. 6, and details are not described herein.

In some embodiments of the present invention, by calling a program or instruction stored in the memory 1005, the processor 1001 may be configured to: obtain an available integrated feature vector corresponding to the Z sample face images (of course, the available comprehensive features corresponding to the Z sample face images) The identity factor and the age factor in the vector have been determined. The identity-age factor model is trained using the available comprehensive feature vectors corresponding to the Z sample face images to determine the model parameters of the identity-age factor model described above.

among them,

Among them, the above identity-age factor model is as follows:

q = i βq + ε

q Among them, the above

1, 2,

T

The total number of segments of the feature vector (ie, the available feature vector is divided into n segments of equal length or unequal length, each segment corresponding to one available integrated feature subvector, f丄 including n segments corresponding to the available Integrated feature subvector, ie

Including n available integrated feature sub-vectors, t _a T ut _a q represents the available integrated feature sub-vectors corresponding to the segment q of the 丄. The above represents q

The identity factor coefficient corresponding to V t _a β. The above q represents the age factor coefficient corresponding to q, the above ^ q

Indicates the age factor corresponding to q, where

s ~ (0 cr ² /)

Corresponding Gaussian white noise, q, . Where the corresponding model Θ 2, U , V, σ _η ² }

parameter

^Qqq , that is, used to describe the available comprehensive features

The segmentation of the quantity T corresponds to the identity of the available synthetic feature subvectors - the model parameters of the age factor model

Θ q 二^^ q, U q, V q, σ _η q ² } J where, based on the above identity-age factor model, it can be seen that any face image corresponds to the available comprehensive

Feature vector

The available composite feature subvectors corresponding to the segmentation q are composed of the following three parts.

U X U X

V y

It does not change with the age of the person and can be used for character identification. ^Qq depends on the age of the person corresponding to the face image and can be used to estimate the age of the corresponding person. It can be understood that the above examples are described by taking the sequence of numbers from 1 to n as an example. Of course, in actual applications, each segment may be numbered out of order.

due to

Θ a two {β U

, 'a, V

q' a, a, σ a } ) The model parameter of the identity-age factor model— ^q , so the above-mentioned identity-age factor model can be used for the available comprehensive feature vector corresponding to the sample face image

Θ two {β , U , V, σ _η ² } type training to determine the value of the model parameter ^qqqqq , the identity-age factor model corresponding to each segment trained can lay the identification of any face image to be identified Good foundation. Wherein, since the available integrated feature vectors of the face image to be recognized are jointly determined by using mutually unrelated identity factors and age factors, it is advantageous to associate the identity-related features included in the available integrated feature vector of the face image to be recognized. The feature is stripped out, which is beneficial to eliminate the influence of the age-related features contained in the available comprehensive feature vector of the face image to be recognized on the identity recognition, thereby facilitating the improvement of the accuracy and versatility of the image identification, thereby facilitating the image recognition. Meet as many application scenarios as possible.

T t assumes that the corresponding composite eigenvector vector corresponding to the available eigenvector vector ^ of the composite eigenvector i can be used as follows: The age factor model is as follows:

Among them, the above

1, 2,

Representing the above-mentioned available integrated feature vector, the above n represents the total number of segments of the available integrated feature vector (ie, the available integrated feature vector)

Divided into n segments of equal length or unequal length, each segment corresponding to an available integrated feature subvector, T丄 includes n available segments corresponding to the available integrated feature subvector, ie

Including n available synthetic feature subvectors), Represents the identity factor coefficient corresponding to q. The above ^ represents the corresponding age factor coefficient, the above ^ ^Γ

Indicates the age factor corresponding to q, where the above representation Where q corresponds to the model

, ie, used to describe the available composite feature vectors

Θ q 二 ^ ^ q, U q, V q, σ _η q ² } J The processor 1001 can optimize the model parameters, for example, by maximizing the joint probability distribution formula, wherein the maximum joint probability distribution formula can be, for example, the following formula 3 Shown as follows:

Where, in Equation 2, k represents the age of the person corresponding to the sample face image, and i represents the sample face

→k

The identity of the person corresponding to the image, where t q, {) indicates that the identity is i and the age is k

Sample representation

Child

Probability distribution.

Where ^ represents the joint probability distribution. Wherein, the experiment may be performed on a plurality of sample face images corresponding to different ages and different identities. Among them, due to the two invisibility factors in the identity-age factor model

^ q^ k) cannot be observed directly. For example, the coordinate rising algorithm can be used to optimize the factor ^x q, and the other invisible factor.

, can estimate the prior probability distribution

(q,0). Thus, the prior probability distribution ^{P 0)} ^^'^ ) \ ^T q can be obtained by maximizing the conditional expectation of the joint probability distribution L.

The new model parameter values _t ^

That is, given the initial estimate ( ^q ' ^Q ), by maximizing L _{c in} Equation 4 below β

Get a new one [ ,

Wherein, in the representation sample set in Equation 4

Each sample face image (assuming there are Z sample face images) corresponds to the available integrated feature subvector corresponding to the segment q of the available integrated feature vectors, L _c is the joint probability distribution ^ in the given initial model β → →

The conditional expectation of the parameter ( ^q , ⁰⁾ . Among them, because the stealth factor and y( , ) are not β

Know, so L can't be directly maximized. However, the distribution of the invisible factor can be estimated by initializing the model parameters ( ^q ' ^Q ), and then the joint probability distribution under the lower distribution is obtained.

The condition of ^ is expected, and the condition is expected to be L _c .

The following is the maximum expected value (EM) algorithm for adapting the identity-age factor model described above: Input: Multiple sample faces with identity and age can be segmented in the integrated feature vector

The output is: Model parameters of the identity-age factor model corresponding to the available integrated feature subvectors corresponding to segment q:

Θ q 二^^ q, U q, V q, σ _η q ² } J

Specifically, the following parameters can be initialized first:

2 0.1

U _q ^rand(- AfiA) V ^rand(-0A,0A)

2 jj y

Bring the initialized, q, into the identity-age factor model formula, and obtain

Calculate the stealth factor and based on the model parameters Θ ^q { {, β, U q, V q, σ _η q ² } ^

among them,

C = aq ² I + U q U q ^T ;

Φ

q

i+v _a ^T c~X.

Wherein, N _ci represents the number of sample face images whose identity is i in the training sample face image, and N _sk represents the number of sample face images of age k in the training sample face image. Further updating the model parameters u^ and vq based on the calculated stealth factors q, ) and k)

,

Wherein, N represents the total number of sample face images, and d is the length of the available integrated feature sub-vector corresponding to the index q of the available composite feature vector of the sample face image.

Based on the above manner, segmentation among the available integrated feature vectors corresponding to the Z sample face images q The corresponding synthetic feature subvector corresponding to q is used to find the parameters, u and v. multiple times until convergence. The model parameter of the identity-age factor model corresponding to the available comprehensive feature sub-vector corresponding to the segment q can be accurately calculated by the above algorithm.

Training method. Through the above exemplified manner, model parameters of the identity-age factor model corresponding to the available comprehensive feature sub-vectors corresponding to other segments may also be obtained. Referring to FIG. 11, FIG. 11 illustrates a structure of a model training device 1100 according to an embodiment of the present invention. The model training device 1100 includes: at least one processor 1101, such as a CPU, at least one network interface 1104 or other user interface 1103, and a memory 1105. At least one communication bus 1102. Communication bus 1102 is used to implement connection communication between these components. The model training device 1100 optionally includes a user interface 1103, including: a display, a keyboard or a pointing device (such as a mouse, a trackball, a touchpad or a touch sensitive display). Memory 1105 may contain high speed RAM memory and may of course also include non-volatile memory, such as at least one disk memory. The memory 1105 can optionally include at least one storage device located remotely from the processor 1101.

In some embodiments, the memory 1105 stores the following elements, executable modules or data structures, or a subset thereof, or their extension set:

Operating system 11051, which contains various system programs for implementing various basic services and processing hardware-based tasks;

The application module 11052 includes various applications for implementing various application services. The application module 11052 includes, but is not limited to, an acquisition unit 710 and a training unit 720.

For the specific implementation of each module in the application module 11052, refer to the corresponding module in the embodiment shown in FIG. 7, and details are not described herein. In some embodiments of the present invention, by calling a program or instruction stored in the memory 1105, the processor 1101 may be configured to: obtain an available integrated feature vector corresponding to the Z sample face images. The identity-age factor model is trained using the available comprehensive feature vectors corresponding to the Z sample face images to determine the model parameters of the identity-age factor model described above.

Among them, the above identity-age factor model is as follows:

Wherein, the above ^ represents the available integrated feature vector, the above ^ represents the average value of the sample feature, the U represents the identity factor coefficient, and the above V represents the age factor coefficient, and the above represents Gaussian white noise, Ν (V0 ' σ ² ΐ), The above X represents an identity factor, and the above y represents an age factor, wherein the above model parameter is u, V, }.

Based on the above-mentioned identity-age factor model, it can be seen that the available comprehensive features U X V y vectors corresponding to any face image are composed of the following three parts: identity information, age information and noise. Among them, depending on the identity of the person corresponding to the face image, it can be considered not

V y

As the age of the character changes, it can be used for character identification; depending on the age of the person corresponding to the face image, it can be used to estimate the age of the person. In some possible embodiments of the present invention, the above-mentioned available integrated feature vector is obtained based on a gradient direction histogram or based on other methods.

The age factor model can lay a good foundation for the identification of any face image to be recognized. Wherein, since the available integrated feature vectors of the face image to be recognized are jointly determined by using mutually unrelated identity factors and age factors, it is advantageous to associate the identity-related features included in the available integrated feature vector of the face image to be recognized. The feature is stripped out, which is beneficial to eliminate the influence of the age-related features included in the available comprehensive feature vector of the face image to be recognized on the identity recognition, thereby facilitating the improvement of the accuracy and versatility of the image identification, thereby facilitating the image recognition. Meet as many application scenarios as possible.

Assume that the identity-age factor model that processor 1101 is to train is as follows:

ε Wherein, the above ^ represents the available integrated feature vector, the above ^ represents the sample feature average value, the U represents the identity factor coefficient, and the above V represents the age factor coefficient, and the above represents Gaussian white noise, ^G 〜 N ^V (Q σ ² ΐ) The above X^ represents an identity factor, and the above y represents an age factor. Model parameters of the above identity-age factor model

υ, V, ^σ } . For example, the processor 1101 may optimize the model parameters by maximizing the joint probability distribution formula, wherein the maximum joint probability distribution formula may be, for example, as shown in the following formula 1:

, , ) where, in formula 1, k represents the age of the person corresponding to the sample face image, and i represents the sample face

The identity of the person corresponding to the image,

The available comprehensive feature vector corresponding to the sample face image whose identity is i and age k, represents the identity factor of the person corresponding to the sample face image whose identity is i, ^^ is the sample face image of age k Corresponding to the age factor of the character, ^^ represents the condition of the given model parameter Θ, and the joint probability distribution. Where L represents the joint probability distribution. Among them, the two invisible factors and ^^ in Equation 1 cannot be directly observed. For example, the factor sum can be analyzed by a coordinate ascending algorithm, that is, another invisible factor is optimized with one factor fixed. Where, for a given model parameter, a prior probability distribution can be estimated

. Thus, the prior probability distribution can be obtained by maximizing the conditional expectation of the joint probability distribution L.

, and then update the value of the model parameter Θ.

Among them, formula face

The available composite feature vector corresponding to the image (assuming there are Z sample face images), L _C is the conditional expectation of the joint probability distribution L given the initial model parameters. Among them, because the invisibility factor and ^^ are unknown, L cannot be directly maximized. But by initializing the model parameter ^σ . To estimate the invisibility factor

The distribution of sums, in turn, gives the conditional expectation of the joint probability distribution L under the lower distribution, which is expected to be L _C . The maximum conditional expectation (EM) algorithm for the above identity-age factor model adaptation is presented below: Input as: Feature Vector Set of Sample Image with Identity and Age

The output is: Model parameters of the feature model 6 -, β, U, V, σ } Specifically, the following parameters can be initialized first:

σ ² 0.1, biliary - 0·1, 0·1)

V ra^(— 0.1,0.1) .

2 β Bring the initialized ^σ , υ, V into the identity-age factor model formula and find ^. Based on the model parameters 9-^, υ, calculate the stealth factor 'and. among them,

τ

C two σΊ+UU

VC—V.

The N _d represents the number of sample face images whose identity is i in the training sample face image, and the N _sk represents the number of sample face images of the age of k in the training sample face image. The model parameters σ , and _{ν are} updated based on the calculated stealth factor and ^^.

among them,

-1

U = (C-DB ^→ E) (A-FB~ ^L E

Wherein, N represents the total number of sample face images, and d is the length of the available feature vector of the sample face image. Based on the above manner, the parameters ^σ and ^ are obtained multiple times using a plurality of available integrated feature vectors in the available integrated feature vectors corresponding to the Z sample face images until convergence. Through the above algorithm, the model parameters ^ = ^u , v, ^σ } of the above identity-age factor model can be calculated more accurately. Of course, the model parameters θ = {β, ί , ν, σ ² } can also be trained by other means.

Body value. It is not limited to the training method exemplified above. The embodiment of the present invention further provides a computer storage medium, wherein the computer storage medium can store a program, and the program includes some or all of the steps of the power feedforward control method described in the foregoing method embodiment.

It should be noted that, for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations, but those skilled in the art should understand that the present invention is not limited by the described action sequence. Because certain steps may be performed in other sequences or concurrently in accordance with the present invention. In addition, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily required by the present invention. In the above embodiments, the descriptions of the various embodiments are different, and the parts that are not detailed in an embodiment can be referred to the related descriptions of other embodiments.

In the several embodiments provided herein, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the unit is only a logical function division. In actual implementation, there may be another division manner, for example, multiple units or components may be combined or may be Integration into another system, or some features can be ignored, or not executed. Alternatively, the mutual coupling or direct connection or communication connection shown or discussed may be an indirect connection or communication connection through some interface, device or unit, and may be electrical or other form. The components displayed by the unit may or may not be physical units, that is, may be located in one place, or may be distributed to multiple network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solution of the embodiment.

In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

The integrated unit, if implemented in the form of a software functional unit and sold or used as a standalone product, may be stored in a computer readable storage medium. Based on such understanding, the technical solution of the present invention may contribute to the prior art or all or part of the technical solution may be embodied in the form of a software product stored in a storage medium. A number of instructions are included to cause a computer device (which may be a personal computer, server or network device, etc.) to perform all or part of the steps of the methods described in various embodiments of the present invention. The foregoing storage medium includes: a U disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a removable hard disk, a magnetic disk, or an optical disk, and the like, which can store program code. .

The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to be limiting; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art will understand that The technical solutions described in the embodiments are modified, or part of them The technical features are equivalent to those of the embodiments of the present invention.

Claims

Rights request

1. An image identity recognition method, characterized by including:

Perform feature extraction processing on the face image to be recognized to obtain an available comprehensive feature vector corresponding to the face image to be recognized, wherein the available comprehensive feature vector is used to describe the person corresponding to the face image to be recognized. The identity factor of the identity and the age factor used to describe the age of the person corresponding to the face image to be recognized are jointly determined, wherein, the identity factor and the age factor are not related to each other;

Calculate the identity feature vector corresponding to the face image to be recognized based on the available comprehensive feature vector corresponding to the face image to be recognized;

Calculate the similarity between the identity feature vector corresponding to the face image to be recognized and the identity feature vector corresponding to each of the Z sample face images, where, the identity feature vector is determined by the identity factor, The Z is a positive integer;

Output the identity information corresponding to the Z1 sample face images, where the Z1 sample face images are a subset of the Z sample face images, and the identity feature vectors corresponding to the Z1 sample face images are consistent with the Z1 sample face images. The similarity of the identity feature vector corresponding to the face image to be recognized is greater than the identity feature vector corresponding to the other sample face images among the Z sample face images except the Z1 sample face images. The similarity of the identity feature vectors corresponding to the face images to be recognized, or the similarity between the identity feature vectors corresponding to the Z1 sample face images and the identity feature vectors corresponding to the face images to be recognized is greater than the setting A fixed threshold value, wherein the identity information corresponding to the output Z1 sample face images is the possible identity information corresponding to the face image to be recognized.

2. The method according to claim 1, characterized in that,

The feature extraction processing of the face image to be recognized to obtain the available comprehensive feature vector corresponding to the face image to be recognized includes: preprocessing the face image to be recognized; preprocessing the face image to be recognized. The recognized face image is subjected to feature extraction processing to obtain an available comprehensive feature vector corresponding to the face image to be recognized.

3. The method according to claim 2, characterized in that: performing feature extraction processing on the pre-processed face image to be recognized to obtain available comprehensive features corresponding to the face image to be recognized. The vector includes: extracting an original comprehensive feature vector from the preprocessed face image to be recognized, and performing dimensionality reduction processing on the extracted original comprehensive feature vector to obtain the to-be-identified face image. Available comprehensive feature vectors corresponding to the recognized face images.

4. The method according to claims 1 to 3, characterized in that the original comprehensive feature vector or the available comprehensive feature vector is obtained based on a gradient direction histogram.

5. The method according to any one of claims 1 to 4, characterized in that,

The available comprehensive feature vector corresponding to the face image to be recognized is described by an identity-age factor model, where the identity-age factor model is as follows:

Wherein, the ^ represents the available comprehensive feature vector, the ^ represents the sample feature average, the U represents the identity factor coefficient, the V represents the age factor coefficient, and the represents high ~ N (0 σ ² ΐ) X y is white noise, V ' , represents the identity factor, and represents the age factor.

6. The method according to claim 5, characterized in that,

Calculating the identity feature vector corresponding to the face image to be recognized based on the available comprehensive feature vector corresponding to the face image to be recognized includes: in the following manner, based on the face image to be recognized corresponding to The comprehensive feature vector can be used to calculate the identity feature vector corresponding to the face image to be recognized:

Among them, the 1 represents the identity feature vector,

Among them, the

7. The method according to any one of claims 1 to 4, characterized in that the available comprehensive feature vector corresponding to the face image to be recognized is described by an identity-age factor model,

Among them, the identity-age factor model is as follows:

Among them, the T丄 represents the available comprehensive feature vector, and the n represents the available comprehensive feature vector.

The total number of segments of the feature vector, described

The available comprehensive characteristics corresponding to the segment q

eigenvector, described

express the stated

The average sample feature corresponding to q

value, where, said represents the identity factor corresponding to said q, said represents the

The age factor corresponding to the above q, where, the Gaussian white noise corresponding to the above q

8. The method according to claim 7, characterized in that,

Among them, the , represents the identity feature vector, and ^J q represents the identity feature vector corresponding to the segment q.

9. The method according to any one of claims 1 to 7, characterized in that the identity feature vector corresponding to the face image to be recognized corresponds to each of the Z sample face images. The similarity of the identity feature vector is determined by the cosine distance, Euclidean distance or Manhattan distance between the identity feature vector corresponding to the face image to be recognized and the identity feature vector corresponding to each of the Z sample face images. represented by distance.

10. A model training method, characterized by including:

Obtain the available comprehensive feature vectors corresponding to Z sample face images;

Utilize the available comprehensive feature vectors corresponding to the Z sample face images to train the identity-age factor model to determine the model parameters of the identity-age factor model;

Wherein, the available comprehensive feature vector is described by the identity-age factor model,

Among them, the

The identity-age factor model corresponding to the available comprehensive feature subvector corresponding to the segment q is as follows:

That

, the l, n Wherein, the τ represents the available comprehensive feature vector, and the η represents the available comprehensive feature vector.

T

The total number of segments of the feature vector, the q represents the available comprehensive features corresponding to the segment q of the -X.

eigenvector, described

express the stated

The average sample feature corresponding to q

value, where, the represents the identity factor corresponding to the q, and represents the

' The age factor corresponding to q, where, the ^q represents the Gaussian white noise corresponding to the t

Among them, the identity-age factor corresponding to the available comprehensive feature subvector corresponding to the segment q of T丄

Θ two {β, U, V, σ _η ² } model parameters ^qqqqq of the sub-model.

11. The method according to claim 10, wherein the available comprehensive feature vector is obtained based on a gradient direction histogram.

12. A model training method, characterized by including:

Use the available comprehensive feature vectors corresponding to the Z sample face images to train the identity-age factor model to determine the model parameters of the identity-age factor model,

Among them, the identity-age factor model is as follows:

Wherein, the ^ represents the available comprehensive feature vector, the ^ represents the sample feature average, the U represents the identity factor coefficient, the V represents the age factor coefficient, and the represents high ~ N (0 σ ² ΐ ⁾

13. The method according to claim 12, characterized in that the available comprehensive feature vector is obtained based on a gradient direction histogram.

14. An image identity recognition device, characterized by including:

An extraction unit, configured to perform feature extraction processing on the face image to be recognized to obtain an available comprehensive feature vector corresponding to the face image to be recognized, wherein the available comprehensive feature vector is used to describe the person to be recognized. The identity factor of the identity of the person corresponding to the face image and the age factor used to describe the age of the person corresponding to the face image to be recognized are jointly determined, wherein the identity factor and the age factor are independent of each other;

A computing unit calculates the comprehensive feature vector based on the available comprehensive feature vector corresponding to the face image to be recognized. The identity feature vector corresponding to the face image to be recognized;

A matching unit, used to calculate the similarity between the identity feature vector corresponding to the face image to be recognized and the identity feature vector corresponding to each of the Z sample face images, wherein, the identity feature vector Determined by the identity factor, Z is a positive integer;

The output unit is used to output the identity information corresponding to the Z1 sample face images, wherein the Z1 sample face images are a subset of the Z sample face images, and the matching unit calculates the Z1 sample face images. The similarity between the identity feature vector corresponding to the sample face image and the identity feature vector corresponding to the face image to be recognized is greater than that of the Z sample face images except for the Z1 sample face images. The similarity between the identity feature vectors corresponding to other sample face images and the identity feature vector corresponding to the face image to be recognized, or the identity feature vectors corresponding to the Z1 sample face images to the face to be recognized The similarity of the identity feature vectors corresponding to the images is greater than the set threshold, wherein the identity information corresponding to the output Z 1 sample face images is the possible identity information corresponding to the face image to be recognized.

15. The device according to claim 14, characterized in that,

The extraction unit is specifically configured to preprocess the face image to be recognized; perform feature extraction processing on the preprocessed face image to be recognized to obtain the available comprehensive information corresponding to the face image to be recognized. Feature vector.

16. The device according to claim 15, wherein: performing feature extraction processing on the pre-processed face image to be recognized to obtain an available synthesis corresponding to the face image to be recognized. In terms of feature vectors, the extraction unit is specifically configured to extract an original comprehensive feature vector from the preprocessed face image to be recognized, and perform dimensionality reduction processing on the extracted original comprehensive feature vector to obtain The available comprehensive feature vector corresponding to the face image to be recognized.

17. The device according to any one of claims 14 to 16, characterized in that the original comprehensive feature vector or the available comprehensive feature vector is obtained based on a gradient direction histogram.

18. The device according to any one of claims 14 to 17, characterized in that,

The available comprehensive feature vector corresponding to the face image to be recognized is described by an identity-age factor model, where the identity-age factor model is as follows: -no-

Among them, the ^ represents the available comprehensive feature vector, the ^ represents the average value of the sample characteristics, the U represents the identity factor coefficient, the V represents the age factor coefficient, and the represents high ~

White noise, GN (VQ ' σ ² ΐ), the X represents the identity factor, and the y represents the age factor.

19. The device according to claim 18, characterized in that,

The calculation unit is specifically configured to calculate the identity feature vector corresponding to the face image to be recognized based on the available comprehensive feature vector corresponding to the face image to be recognized in the following manner:

Among them, 1 represents the identity feature vector, where,

20. The device according to any one of claims 14 to 17, characterized in that, the available comprehensive feature vector corresponding to the face image to be recognized is described by an identity-age factor model,

Among them, the identity-age factor model is as follows:

Among them, the

T

The total number of segments of the feature vector, described

Represents the available comprehensive feature subvector corresponding to the segment q, the

express the stated

The average value of sample features corresponding to q, where, represents the identity factor corresponding to q, represents the age factor corresponding to q, wherein, represents the Gaussian white noise corresponding to q Shi ε ~

Sound, g Ν σ

q ² ϊ

21. The device according to claim 20, wherein the calculation unit is specifically configured to calculate the face to be recognized based on the available comprehensive feature vector corresponding to the face image to be recognized in the following manner: The identity feature vector corresponding to the image:

That

In, the qq

Among them, the

represents the identity feature vector, and fq represents the identity feature vector corresponding to the segment q.

22. The device according to any one of claims 14 to 20, characterized in that the identity feature vector corresponding to the face image to be recognized corresponds to each of the Z sample face images. The similarity of the identity feature vector is determined by the cosine distance, Euclidean distance or Manhattan distance between the identity feature vector corresponding to the face image to be recognized and the identity feature vector corresponding to each of the Z sample face images. represented by distance.

23. A model training device, characterized by including:

The acquisition unit is used to obtain the available comprehensive feature vectors corresponding to Z sample face images; A training unit, configured to train the identity-age factor model using the available comprehensive feature vectors corresponding to the Z sample face images to determine the model parameters of the identity-age factor model; wherein, the available comprehensive feature vectors Described by the identity-age factor model,

Among them, the identity-age factor model is as follows:

where, the 1,2 where,

represents the available comprehensive feature vector, and the n represents the available comprehensive T

The total number of segments of the feature vector, described

Indicates the available comprehensive characteristics corresponding to the segment q

eigenvector, described

express the stated

The age factor coefficient corresponding to q, the ^ represents the average sample characteristic corresponding to q value, where, said represents the identity factor corresponding to said q, said represents the

The age factor corresponding to q, where, represents the Gaussian white noise corresponding to q

scholar

Sound, ε g ~ Ν V 0,' σ q ² Ι

Among them, the available comprehensive feature subvector corresponding to the segment q of the τ丄 corresponds to the identity-age factor e _Q = {β , U _q , V , a _Q ² } and the model parameters ^qqqqq of the sub-model.

24. The device according to claim 23, wherein the available comprehensive feature vector is obtained based on a gradient direction histogram.

25. A model training device, characterized by including:

The acquisition unit is used to obtain the available comprehensive feature vectors corresponding to the Z sample face images; the training unit is used to train the identity-age factor model using the available comprehensive feature vectors corresponding to the Z sample face images to determine Model parameters of the identity-age factor model, wherein, the available comprehensive feature vector is described by the identity-age factor model,

Among them, the identity-age factor model is as follows:

Among them, the ^ represents the available comprehensive feature vector, and the ^ represents the sample feature average mean value, the U represents the identity factor coefficient, the V represents the age factor coefficient, and the U represents the high

White noise, G ~ N (VQ ' σ ² ΐ), the X represents the identity factor, the y represents the age factor, where, the model parameters — υ, V, ^σ }.

26. The device according to claim 25, wherein the available comprehensive feature vector is obtained based on a gradient direction histogram.

27. An identity recognition system, characterized by including:

Client, used to send face images to be recognized to the identity recognition server;

Wherein, the identity recognition server is configured to receive the face image to be recognized from the client, and perform feature extraction processing on the face image to be recognized to obtain the face image corresponding to the face image to be recognized. Available comprehensive feature vectors, wherein the available comprehensive feature vector consists of an identity factor used to describe the identity of the person corresponding to the face image to be recognized and an age factor used to describe the person corresponding to the face image to be recognized The age factor of is jointly determined, wherein the identity factor and the age factor are not related to each other; the identity feature corresponding to the face image to be recognized is calculated based on the available comprehensive feature vector corresponding to the face image to be recognized. Vector; calculate the similarity between the identity feature vector corresponding to the face image to be recognized and the identity feature vector corresponding to each of the Z sample face images, where the identity feature vector is represented by the identity factor It is determined that Z is a positive integer; I want the client to output identity information corresponding to Z1 sample face images, where the Z1 sample face images are a subset of the Z sample face images, so The similarity between the identity feature vectors corresponding to the Z1 sample face images and the identity feature vector corresponding to the face image to be recognized is greater than the Z1 sample face images among the Z1 sample face images. The similarity between the identity feature vector corresponding to other sample face images and the identity feature vector corresponding to the face image to be recognized, or the

The similarity between the identity feature vectors corresponding to the Z1 sample face images and the identity feature vector corresponding to the face image to be recognized is greater than the set threshold, wherein the identity information corresponding to the output Z1 sample face images Possible identity information corresponding to the face image to be recognized.

28. The identity recognition system according to claim 27, characterized in that,

White noise, G N (0 σ ² ΐ), the X represents the identity factor, and the y represents the age factor.

29. The identity recognition system according to claim 28, wherein the identity feature vector corresponding to the face image to be recognized is calculated based on the available comprehensive feature vector corresponding to the face image to be recognized. In aspect, the identity recognition server is specifically configured to calculate the identity feature vector corresponding to the face image to be recognized based on the available comprehensive feature vector corresponding to the face image to be recognized in the following manner:

where, the 1 represents the identity feature vector, where, the

= σ ² / ₊ [/ +ν^

30. The identity recognition system according to claim 27, characterized in that, the available comprehensive feature vector corresponding to the face image to be recognized is described by an identity-age factor model,

Among them, the identity-age factor model is as follows:

Among them, the

T

The total number of segments of the feature vector, described

express the stated

The average value of sample features corresponding to q, where, represents the identity factor corresponding to q, represents the The age factor corresponding to the above q, where, the ⁴ represents the Gaussian white noise corresponding to the above q

31. The identity recognition system according to claim 30, wherein the identity feature vector corresponding to the face image to be recognized is calculated based on the available comprehensive feature vector corresponding to the face image to be recognized. In aspect, the identity recognition server is specifically configured to calculate the identity feature vector corresponding to the face image to be recognized based on the available comprehensive feature vector corresponding to the face image to be recognized in the following manner:

U

where, the q— q _p q ' q where, the

quantity, the J ^J aq represents the identity feature vector corresponding to the segment q.

32. An identity recognition device, characterized by including:

processor and memory;

Wherein, the processor is configured to perform feature extraction processing on the face image to be recognized to obtain an available comprehensive feature vector corresponding to the face image to be recognized, wherein the available comprehensive feature vector is It is jointly determined by an identity factor used to describe the identity of the person corresponding to the face image to be recognized and an age factor used to describe the age of the person corresponding to the face image to be recognized, where the identity factor and The age factors are independent of each other; calculate the identity feature vector corresponding to the face image to be recognized based on the available comprehensive feature vector corresponding to the face image to be recognized; calculate the identity feature vector corresponding to the face image to be recognized The similarity between the identity feature vector and the identity feature vector corresponding to each of the Z sample face images, where the identity feature vector is determined by the identity factor, and Z is a positive integer; Z1 samples are output Identity information corresponding to the face image, wherein the Z1 sample face images are a subset of the Z sample face images, and the identity feature vectors corresponding to the Z1 sample face images are consistent with the to-be-identified The similarity of the identity feature vectors corresponding to the face images is greater than the identity feature vectors corresponding to the other sample face images among the z sample face images except the Z 1 sample face images and the identity feature vectors to be The similarity of the identity feature vector corresponding to the recognized face image, or the

33. The identity recognition device according to claim 32, characterized in that,

The processor is configured to: preprocess the face image to be recognized; perform feature extraction processing on the preprocessed face image to be recognized to obtain an available comprehensive feature vector corresponding to the face image to be recognized. .

34. The identity recognition device according to claim 33, characterized in that, the processor is configured to extract an original comprehensive feature vector from the preprocessed face image to be recognized, and perform the extraction of the original comprehensive feature vector. The comprehensive feature vector is subjected to dimensionality reduction processing to obtain an available comprehensive feature vector corresponding to the face image to be recognized.

35. The identity recognition device according to claims 32 to 34, characterized in that the original comprehensive feature vector or the available comprehensive feature vector is obtained based on a gradient direction histogram.

36. The identity recognition device according to any one of claims 32 to 35, characterized in that, the available comprehensive feature vector corresponding to the face image to be recognized is described by an identity-age factor model, wherein, the identity- The age factor model is as follows:

Among them, the

represents the available comprehensive feature vector, the mean value, the U represents the identity factor coefficient, the V represents the age factor coefficient, and represents the high

~

37. The identity recognition device according to claim 36, characterized in that,

The processor is configured to: calculate the identity feature vector corresponding to the face image to be recognized based on the available comprehensive feature vector corresponding to the face image to be recognized in the following manner:

F2 ΓΣ

Among them, the

represents the identity feature vector,

2 τ . ττττΤ Τ

∑ Two

Among them, the σΊ + UU ¹ +W

38. The identity recognition device according to any one of claims 32 to 35, characterized in that the available comprehensive feature vector corresponding to the face image to be recognized is described by an identity-age factor model,

Among them, the identity-age factor model is as follows:

Among them, the τ represents the available comprehensive feature vector, and the η represents the available comprehensive feature vector.

\ _η Τ

The total number of segments of the feature vector, the q represents the available comprehensive feature subvector corresponding to the segment q, the

express the stated

The average value of the sample features corresponding to q, where, represents the identity factor corresponding to the q, represents the age factor corresponding to the q, wherein, represents the Gaussian white noise corresponding to the q, ~ N[ 0 , _q ² I

39. The identity recognition device according to claim 38, characterized in that,

That

In, the qq

Wherein, the F represents the identity feature vector, and the

The identity feature vector corresponding to the segment q.

40. The identity recognition device according to any one of claims 31 to 38, characterized in that the identity feature vector corresponding to the face image to be recognized is identical to each of the Z sample face images. The similarity of the corresponding identity feature vectors is determined by the cosine distance or Euclidean distance between the identity feature vector corresponding to the face image to be recognized and the identity feature vector corresponding to each of the Z sample face images. Or represented by Manhattan distance.

41. A model training device, characterized by including:

processor and memory;

Wherein, the processor is configured to obtain available comprehensive feature vectors corresponding to Z sample face images; and use the available comprehensive feature vectors corresponding to the Z sample face images to train the identity-age factor model to determine the Model parameters of the identity-age factor model; wherein, the available comprehensive feature vector is described by the identity-age factor model,

Among them, the

That

in, stated

Among them, the 丄 represents the available comprehensive feature vector, and the n represents the available comprehensive T

The total number of segments of the feature vector, described

eigenvector, described

express the stated

The average sample feature corresponding to q

value, where, said represents the identity factor corresponding to said q, said represents the ' The age factor corresponding to q, where, the ^q represents the Gaussian white noise corresponding to the t

Θ 二 {β , U , V , σ _η ² }

Model parameters ^qqqqq of the submodel.

42. The model training device according to claim 41, wherein the available comprehensive feature vector is obtained based on a gradient direction histogram.

43. A model training device, characterized by including:

processor and memory,

Wherein, the processor is configured to obtain available comprehensive feature vectors corresponding to Z sample face images; and use the available comprehensive feature vectors corresponding to the Z sample face images to train the identity-age factor model to determine the Model parameters of the identity-age factor model,

Among them, the identity-age factor model is as follows:

Wherein, the ^ represents the available comprehensive feature vector, the ^ represents the sample feature average, the U represents the identity factor coefficient, the V represents the age factor coefficient, and the represents high White noise, G ~ ^I Νίθ σ ² Ι\, described

represents the age factor, where, the model parameters Θ2, β, U, V, σ ² }

44. The model training device according to claim 43, wherein the available comprehensive feature vector is obtained based on a gradient direction histogram.

45. A computer storage medium, characterized in that the computer storage medium can store a program, and when executed, the program includes the steps described in any one of claims 1 to 13.