WO2016145940A1

WO2016145940A1 - Face authentication method and device

Info

Publication number: WO2016145940A1
Application number: PCT/CN2016/070828
Authority: WO
Inventors: 郇淑雯; 朱和贵; 张祥德
Original assignee: 北京天诚盛业科技有限公司
Priority date: 2015-03-19
Filing date: 2016-01-13
Publication date: 2016-09-22
Also published as: CN105447441A; CN105447441B

Abstract

The present invention belongs to a field of image processing and mode recognition and discloses a face authentication method and device. The method comprises: acquiring a face image sample pair, dividing the acquired face image sample pair into blocks, and calculating based on the acquired blocks a local gray scale feature vector and a local histograms of oriented gradients (HOG) vector, calculating and determining whether the face image sample pair belong to a same person according to the calculated local gray scale feature vector and local HOG vector. The face authentication method and device of the present invention can effectively eliminate disturbing factors such as light, expressions and ages, and improve face recognition accuracy remarkably.

Description

Face authentication method and device

Technical field

The present invention relates to the field of image processing and pattern recognition, and in particular to a face authentication method and apparatus.

Background technique

As people pay more attention to the security and concealment of information, identity authentication technology has become more and more important. Traditional identity authentication methods, such as ID cards and passwords, are cumbersome to use. The biggest disadvantage is that it cannot distinguish between owners and imposters. Biometrics are inherent in human beings and are unique and difficult to replicate. Therefore, the study of biometrics is of great significance. Among them, the advantages of face recognition include: no need for user cooperation; non-contact type features; simple device implementation; and traceability. Because face recognition has the above advantages, this technology has a wide range of application scenarios. In social life, e-commerce, financial institutions, household registration management, etc. all have important application value.

Face authentication is a form of recognition. By effectively characterizing the face, the characteristics of the two face images are obtained, and the classification algorithm is used to determine whether the two photos are the same person. The face authentication process is mainly divided into three parts: face detection, feature extraction and authentication. Since the face is a three-dimensional deformation model, and the face authentication is based on the photo taken by the camera imaging model, the result of the authentication is easily affected by external factors such as illumination, posture, expression, and occlusion. At the same time, face recognition technology involves many interdisciplinary subjects such as pattern recognition, statistical learning, machine vision, applied mathematics and information science, and its wide application prospects have received more and more attention.

At present, the more mature face authentication technology methods in the prior art include: a face authentication method based on geometric features, a face authentication method based on overall features, and a method based on local features. These three methods are easily affected by factors such as light, posture, and expression, and the authentication accuracy is low.

Summary of the invention

The technical problem to be solved by the present invention is to provide a face authentication method and apparatus with strong anti-interference and high accuracy.

In order to solve the above technical problem, the present invention provides the following technical solutions:

A face authentication method includes:

Obtaining a pair of face image samples;

Segmenting the acquired face image sample pairs;

Calculating a local gray feature vector and a local direction gradient histogram HOG vector for the obtained block;

According to the calculated local gray feature vector and the local HOG vector, it is calculated whether the face image sample pair belongs to the same person.

A face authentication device includes:

Acquisition module: used to acquire a pair of face image samples;

Blocking module: used to block the acquired face image sample pairs;

a calculation module: configured to calculate a local gray feature vector and a local direction gradient histogram HOG vector for the obtained block;

The judging module is configured to calculate, according to the calculated local gray feature vector and the local HOG vector, whether the pair of face image samples belong to the same person.

The invention has the following beneficial effects:

Compared with the prior art, the face authentication method of the present invention first acquires a pair of face image samples; then, the acquired face image sample pairs are segmented, and the image segmentation process can be used to implement local information of the face image. Acquisition reduces the computational difficulty of image data processing; then calculates the local gray feature vector and the local direction gradient histogram HOG vector for the obtained block, since the gray feature of the image is the main parameter of image feature recognition, and the face The local features are stable, and the HOG operates on the local square cells of the image. It maintains good invariance to the geometric and optical deformation of the image, which makes the method effectively avoid illumination. Interference with factors such as expression and age, which effectively improves the accuracy of face authentication results; finally, the calculated face map is calculated based on the calculated local gray feature vector and local HOG vector. Like if the sample pair belongs to the same person. The face authentication method of the invention effectively avoids interference of factors such as illumination, expression and age, and also significantly improves the accuracy of face authentication.

DRAWINGS

1 is a schematic flowchart 1 of a face authentication method according to the present invention;

2 is a schematic flow chart of image grayscale difference feature extraction according to the present invention;

3 is a schematic flow chart of face authentication performed by the softmax regression model of the present invention;

4 is a schematic flowchart 2 of a face authentication method according to the present invention;

FIG. 5 is a schematic structural diagram 1 of a face authentication device according to the present invention; FIG.

FIG. 6 is a second schematic structural diagram of a face authentication device according to the present invention.

detailed description

The technical problems, the technical solutions, and the advantages of the present invention will be more clearly described in the following description.

In one aspect, the present invention provides a face authentication method, as shown in FIG. 1, comprising:

Step S101: acquiring a pair of face image samples;

Face image sample pair: In the process of face authentication, the face image detected in real time is compared with the face image in the existing database for 1:1 comparison. This process is called authentication, but in contrast. The two face images processed in the process (ie containing a real-time image and a library image) are called a face sample pair.

In this step, depending on the usage environment, the image sample pair may be composed of a face image collected in real time and a face image stored in a database, or may be composed of two face images stored in a database, or even two images. The face image composed in real time is composed.

Step S102: Blocking the acquired face image sample pairs;

In this step, the block processing method of the face image may adopt various methods known to those skilled in the art, such as a discrete non-coinciding blocking method, a window sliding traversal blocking method, and the like.

Step S103: Calculating a local gray level feature vector and a local direction gradient histogram HOG vector for the obtained block;

In this step, the calculation of the local gray feature vector and the direction gradient histogram HOG vector can effectively avoid the influence of illumination, expression, age and other factors on the result, and improve the accuracy of face authentication.

Step S104: Calculate, according to the calculated local gray feature vector and the local HOG vector, whether the pair of face image samples belong to the same person.

In the present invention, the method first acquires a pair of face image samples; then, the obtained face image sample pairs are segmented, and the image segmentation processing can be used to collect local information of the face image, thereby reducing image data processing. Computational difficulty; then calculate the local gray feature vector and the local direction gradient histogram HOG vector for the obtained block, since the gray feature of the image is the main parameter of the image feature recognition, and the local feature of the face has stability, and at the same time HOG operates on the local grid unit of the image, which can maintain the invariance of the geometric and optical deformation of the image, so that the method can effectively avoid the interference of illumination, expression and age. Furthermore, the accuracy of the face authentication result is effectively improved; finally, according to the calculated local gray feature vector and the local HOG vector, it is calculated whether the face image sample pair belongs to the same person. The face authentication method of the invention effectively avoids interference of factors such as illumination, expression and age, and also significantly improves the accuracy of face authentication.

As an improvement of the present invention, step S102 may further be: performing multi-scale, overlapping partitioning on the acquired pairs of face image samples.

In this step, since the local images at different scales contain different edges and other information, the multi-scale blocking strategy is used to extract local information at different scales, thereby supplementing the information loss caused by the single-scale features. At the same time, in order to extract sufficient image features, the method also uses an overlapping blocking strategy to segment the face image. This step further avoids the influence of lighting, expression and other factors on the authentication result, and also effectively enhances the accuracy of the system for face authentication.

The following is an example of a 4*4 image block divided into 4*4 images. In order to realize multi-scale and overlapping partitioning of images, if the step size is 2, 2*3=6 4*4 can be obtained. The small image below is the overall matrix of the image:

The following are the 6 small matrices after the image has been multi-scaled and overlapped:

The method can adopt different multi-scale and overlapping block strategies for different images. For example, the step size can be set to 5 at 10*10 scale; the step size can be set to 10 at 20*20 scale; 30*30 scale The step size can be set to 15.

As an improvement of the present invention, step S103 may further be: calculating the local gray vector cosine distance as the gray feature vector for the obtained block, and calculating the local HOG vector cosine distance as the HOG vector.

In this step, as shown in FIG. 2, an integrated vector is formed by combining the calculated local gray vector cosine distance and the local HOG vector cosine distance. In the present invention, the calculation process of the local gray vector cosine distance and the local HOG vector cosine distance adopted for the face images of different scales is the same. For the two face images, the same form of blocking strategy is performed on the same scale to obtain the same number of small images. For two small images at the same position, calculate the local gray vector cosine distance and the local HOG vector cosine distance:

1. The following is an example of the calculation of the local gray vector cosine distance, assuming that both images are 6*8 images (the value of each element in the matrix represents the gray value of the position of the image), each element The values have been determined. The image block is extracted for each image by using the 4*4 scale, and the block size is set to 2, the gray matrix of image 1 is A ₁ , and the gray matrix of image 2 is A ₂ :

For the image 1 and the image 2, the image sub-matrices a and b in the upper left dotted frame of the 4*4 scale of the same position are respectively extracted:

The image sub-matrices a', b' in the lower right dashed frame of the 4*4 scale of the same position are extracted for the image 1 and the image 2, respectively:

1) Pull the corresponding block images of image 1 and image 2 into vectors:

Image block in the upper left frame of image 1:

a=[11,45,47,97,58,58,87,11,56,47,45,13,25,35,1,74]

Image block in the upper left dotted line of image 2:

b=[54,11,1,11,39,13,1,13,36,74,1,47,23,87,48,47]

The image block inside the dotted line in the lower right of image 1:

a'=[45,13,87,47,1,74,15,54,1,87,33,39,1,49,55,36]

Image block in the dotted box at the bottom right of image 2:

b'=[1,47,18,62,78,47,97,7,55,87,11,83,47,18,62,26]

2) Calculate the cosine distance of the corresponding image block vector:

grayCosdis(a,b)=1-cos(a,b)=0.4809 (1)

grayCosdis(a,b)=1-cos(a',b')=0.3561 (2)

among them,

3) Since the 6*8 image can be divided into 2*3=6 image blocks, the corresponding 6 cosine distances will be obtained. These six values can form a vector [graycosdis1, graycosdis2, graycosdis3., graycosdis4, graycosdis5 , graycosdis6], indicating the grayscale difference feature of the face sample pair at the scale.

2. The HOG in the present invention is a feature defined for statistical information of gradient direction and intensity in a rectangular region of an image, and can well represent the edge or gradient structure of the target in the local region, thereby characterizing the shape of the target.

The HOG specific calculation process can be as follows:

A(x, y) represents the gray value of the image at the pixel point (x, y).

1) Calculate a step of the image. The derivation operation not only captures the outline and some texture information,

It also weakens the effects of lighting. The calculation method is as follows:

G _x (x,y)=A(x+1,y)-A(x-1,y) (3)

G _y (x,y)=A(x,y+1)-A(x,y-1) (4)

G _x (x, y) and G _y (x, y) represent the magnitudes of the vertical and horizontal gradients of the image at the (x, y) point.

The gradient size at the (x, y) point is defined as:

The gradient direction at the (x, y) point is defined as:

The gradient direction of [0, π] is divided into nine bins, and the amplitude of each pixel is defined as:

2) Accumulate the gradient magnitudes of all the pixels in the region of interest, such a region of interest

9 features can be extracted, and the accumulation method is as follows:

B _k =∑value _k (x,y)1≤k≤9 (8)

B _k represents the kth eigenvalue in the region of interest, and B is the gradient histogram vector in the region of interest, which can indicate the gradient structure in the region of interest.

For the extraction of local HOG features, it is the same as the local grayscale difference feature. Even in the case of changes in expression, age, illumination, etc., the local gradient distribution of the face is basically stable, so it is the same as the local grayscale difference feature. The partitioning method extracts the difference features of the local gradient histogram, that is, respectively calculates the gradient histogram vectors of the corresponding blocks of the pair of face samples to be authenticated, and then calculates the cosine distance of the two as the similarity measure of the image gradient distribution.

In this step, a method of combining local grayscale features and local HOG features is proposed to calculate the pair of face image samples, which can not only solve the influence of expression, age, illumination and other factors on the authentication result to some extent. The face authentication process can be accurately implemented. Compared with the prior art, the method is superior to the prior art in the complexity of the time and space of the algorithm.

As an improvement of the present invention, step S104 may further be: processing the calculated local gray feature vector and the local HOG vector, using an artificial neural network classification model of softmax regression, and calculating whether the face image sample pair is the same. personal. In the present invention, the calculated local gray feature vector and local HOG vector may also be implemented by SVM algorithm, naive Bayes algorithm, Gaussian process based classification algorithm, adaboost algorithm or k-nearest neighbor algorithm (KNN). The processing is performed to calculate whether the pair of face image samples belong to the same person.

In this step, the artificial neural network algorithm has better training and learning ability, and the extracted face feature vector is used as the input of the artificial neural network, and then the training of the network parameters is performed to obtain the face recognition classifier, thereby completing Certification of faces. For the preferred softmax regression model neural network algorithm in the present invention, the softmax model in the case of two classifications is used. The specific algorithm of the softmax regression model can be referred to as follows:

Assume that the training set consists of m labeled samples: X={(x ⁽¹⁾ , y ⁽¹⁾ )...(x ^(m) , y ^(m) )}, where the input sample x ⁽ⁱ⁾ ∈R ⁿ⁺¹ (our convention for symbols is as follows: the dimension of the feature vector x is n+1, where

For the intercept item). Softmax regression is a method for multi-classification problems. To avoid generality, we set the value of the label (that is, the expected output) y to {1...K} (K>2).

In the softmax model, for each given input x, we want to estimate the conditional probability value p(y=jx) for each tag class j using a hypothesis function h _θ (x). That is, it is assumed that the output of the function corresponding to each sample is a K-dimensional vector, and each component represents a probability estimate of each class under the condition of a given x. So, in this article, we will assume that the function is defined as:

among them

Where θ _j ∈R ⁿ⁺¹ is the parameter of the model, corresponding to the sample, θ _j,0 is the deviation term of the j category,

It is a normalized system. Figure 3 is a neural network representation of the softmax regression model, where 1 is the gray difference vector at the first scale, 2 is the gray difference vector at the second scale, and 3 is the HOG difference vector at the last scale. 4 is the probability of being the same person, 5 is the probability that the person is not the same person. When 4 is greater than 5, it is judged that the obtained face image sample pair belongs to the same person. When 4 is less than 5, the obtained face image sample pair is judged. Not belonging to the same person.

In the present invention, as shown in FIG. 4, step S102 may include step S105:

Aligning the face image sample pairs;

The aligned face image sample pairs are normalized and stretched.

Since the face changes in translation during the acquisition process, in order to avoid the influence of the posture and position information on the face authentication, the human image sample pairs are aligned in this step, wherein the face alignment method can be used. Various methods known to those skilled in the art may be used, such as a human eye positioning algorithm, an Active Shape Model (ASM) algorithm, and an Active Appearance Model (AAM) algorithm. At the same time, in order to strengthen the consistency of the face image sample pairs, thereby avoiding the influence of illumination, direction and noise, in this step, the aligned face image sample pairs are normalized and stretched, wherein the face is The method of normalizing and stretching the image may be performed by various methods known to those skilled in the art, such as image mean-variance normalization and gray-scale hyperbolic tangent stretching method, and maximum-minimum normalization. And the gray scale stretching method of the piecewise linear transformation function.

Preferably, step S105 is further:

The aligning the face image sample pairs further comprises: aligning the face image sample pairs by using a human face positioning based face alignment algorithm;

The normalizing and stretching processes on the aligned face image sample pairs further include: using the mean-variance normalization and the gray-scale hyperbolic tangent stretching process on the aligned face image sample pairs.

In this step, specifically:

1. Alignment algorithm based on human eye positioning

The core idea of the alignment algorithm based on human eye localization is to locate the two eye positions in the face region by using the positioning algorithm, and transform the image into a uniform size by panning, zooming, rotating, etc., so that the two images in the mutual comparison are in the image. The eyes are fixed in the same position, the specific steps are as follows:

1) Using the face feature point localization algorithm to perform eye positioning on the face image, the coordinates of the left eye are (x1, y1), and the coordinates of the right eye are (x2, y2);

2) Using MATLAB's own function cp2tform to change the two face images in the face sample pair respectively;

3) The two images obtained in step 2) constitute a new pair of face samples, and the subsequent operations of this patent are performed on pairs of new face samples.

2. Image Mean-variance Normalization and Grayscale Hyperbolic Tangent Stretching

The currently obtained face sample image is affected by illumination when it is acquired. This method reduces the influence of illumination to some extent by performing mean-variance normalization and gray-scale hyperbolic tangent stretching on the image.

The image mean-variance normalization formula is:

Where g(i,j) represents the gray value at (i,j) after normalization, and f(i,j) represents the pre-normalization (i,j) The gray value, mean(f) represents the average value of the image before normalization, and var(f) represents the variance of the gray value of the image before normalization.

Grayscale hyperbolic tangent stretching is performed after the graph mean-variance normalization operation.

among them

Indicates the gray value at (i, j) after stretching, and g(i, j) represents the gray value at (i, j) before stretching.

The face alignment algorithm based on human eye positioning, the mean-variance normalization algorithm and the gray-scale hyperbolic tangent stretching algorithm adopted in the present invention are all performed by using a basic algorithm in the field of mathematics, so that the present invention does not need to undergo complicated A lengthy mathematical process can be implemented.

In another aspect, the present invention also provides a face authentication device, as shown in FIG. 5, comprising:

The obtaining module 11 is configured to acquire a pair of face image samples;

Blocking module 12: configured to block the acquired face image sample pairs;

The calculating module 13 is configured to calculate a local gray level feature vector and a local direction gradient histogram HOG vector for the obtained block;

The determining module 14 is configured to calculate, according to the calculated local gray feature vector and the local HOG vector, whether the pair of face image samples belong to the same person.

In the face authentication device of the present invention, the acquisition module 1 first acquires a face image sample pair; the segmentation module 2 then blocks the acquired face image sample pairs, and the image segmentation process can implement local information on the face image. The acquisition is performed to reduce the computational difficulty of the image data processing; then the calculation module 13 calculates the local gray feature vector and the local direction gradient histogram HOG vector for the obtained block, since the gray feature of the image is the main parameter of the image feature recognition And the local features of the face have stability, while the HOG operates on the local grid unit of the image, which maintains good invariance to the geometric and optical deformation of the image, so that the invention can be effective Avoid interference from factors such as light, expression and age, which effectively improves people The accuracy of the face authentication result; the final judging module 14 calculates whether the pair of face image samples belong to the same person according to the calculated local gray feature vector and the local HOG vector. The face authentication device of the invention effectively avoids interference of factors such as illumination, expression and age, and also significantly improves the accuracy of face authentication.

As an improvement of the present invention, the blocking module 12 can be further configured to perform multi-scale, overlapping partitioning on the acquired pairs of face image samples.

In the present invention, when accepting external images, information of different scales is automatically processed, and partial images at different scales may contain different edges and the like. Therefore, multi-scale blocking strategies are used to extract local information at different scales, thereby supplementing The information caused by single-scale features is missing. At the same time, in order to extract sufficient image features, the present invention also uses an overlapping blocking strategy to segment the face image. The invention not only can solve the influence of illumination, expression and other factors on the authentication result to a certain extent, but also effectively enhances the accuracy of the system for face authentication.

Preferably, the calculation module 13 is further configured to calculate the local gray vector cosine distance as the gray feature vector for the obtained block, and calculate the local HOG vector cosine distance as the HOG vector.

The invention proposes a method for combining face image sample pairs by using a combination of local gray feature and local HOG feature, which can not only solve the influence of expression, age, illumination and other factors on the authentication result to some extent, and The face authentication process can be accurately implemented. Compared with the prior art, the present invention is superior to the prior art in the complexity of the time and space of the algorithm.

As an improvement of the present invention, the judging module 14 can be further configured to process the calculated local gray feature vector and the local HOG vector, and adopt an artificial neural network classification model of the softmax regression model to calculate and determine the face image sample pair. Whether it belongs to the same person.

As an improvement of the present invention, the judging module 14 can be further configured to process the calculated local gray feature vector and the local HOG vector, and adopt an artificial neural network classification model of softmax regression to calculate whether the face image sample pair is determined. Belong to the same person. The judging module 14 in the present invention can also use the SVM algorithm, the native bayes, the Gaussian process-based classification algorithm, the adaboost algorithm or the k-nearest neighbor for the calculated local gray feature vector and local HOG vector. Algorithm (KNN), etc., to calculate and judge the face image Whether the sample pair belongs to the same person.

The artificial neural network algorithm of the softmax regression model in the present invention is a mathematical model simulating the thinking mode of the human brain, which has the functions of learning, memory and error correction. In the present invention, it has the effect of improving the accuracy of face authentication. It takes the extracted face feature vector as the input of the artificial neural network, and then through the training of the network parameters to obtain the face recognition classifier, thereby completing the face authentication.

In order to avoid the influence of the posture and the position information and the illumination on the face authentication, the pre-processing module 15 is connected between the obtaining module 11 and the blocking module 12, as shown in FIG. 6, for performing the face image sample pair. Alignment processing; normalize and stretch the aligned face image sample pairs.

Since the face changes during the acquisition process such as translation and rotation, in order to avoid the influence of posture and position information on face authentication, the human image sample pairs are aligned in this step; at the same time, in order to strengthen the face image sample pair Consistency, so as to avoid the influence of factors such as illumination, direction and noise. In this step, the aligned face image sample pairs are normalized and stretched.

Preferably, the pre-processing module 15 is further configured to perform alignment processing on the face image sample pairs by using a human face alignment based face alignment algorithm; and using mean-variance normalization and grayscale on the aligned face image sample pairs. Hyperbolic tangent stretching treatment.

The invention utilizes the difference between the local gray scale difference and the local gradient histogram to obtain the difference representation of the human face, and uses the multi-scale feature to describe the face difference more comprehensively. The invention utilizes the local texture information to be robust to illumination, posture and expression, and the complexity of the time and space of the algorithm is also low. In the FERET database, the four sub-libraries Fb, Fc, DupI, and DupII achieved 97.91%, 73.20%, 63.71%, and 49.57% authentication rates respectively (error mismatch rate was 0.1%).

The above is a preferred embodiment of the present invention, it should be noted that Many modifications and refinements can be made by those skilled in the art without departing from the principles of the invention, and such modifications and refinements are also considered to be within the scope of the invention.

Claims

A face authentication method, comprising:

Obtaining a pair of face image samples;

Segmenting the acquired face image sample pairs;

Calculating a local gray feature vector and a local direction gradient histogram HOG vector for the obtained block;

According to the calculated local gray feature vector and the local HOG vector, it is calculated whether the face image sample pair belongs to the same person.
The face authentication method according to claim 1, wherein the segmenting the acquired face image sample pairs further comprises:

The acquired face image sample pairs are multi-scaled and overlapped.
The face authentication method according to claim 1, wherein the calculating the local gray feature vector and the local direction gradient histogram HOG vector for the obtained block further comprises:

For the obtained partition, the local gray vector cosine distance is calculated as the gray feature vector, and the local HOG vector cosine distance is calculated as the HOG vector.
The face authentication method according to claim 1, wherein the calculating whether the face image sample pair belongs to the same person according to the calculated local gray feature vector and the local HOG vector is further:

The calculated local gray eigenvectors and local HOG vectors are processed by the artificial neural network classification model of softmax regression, and it is calculated whether the pair of face image samples belong to the same person.
The face authentication method according to claim 1, wherein the step of segmenting the acquired face image sample pairs comprises:

Aligning the face image sample pairs;

The aligned face image sample pairs are normalized and stretched.
The face authentication method according to claim 5, wherein the face is facing The alignment of the image sample pairs is further performed by: aligning the face image sample pairs with a human face alignment algorithm based on human eye positioning;

The normalizing and stretching processes on the aligned face image sample pairs further include: using the mean-variance normalization and the gray-scale hyperbolic tangent stretching process on the aligned face image sample pairs.
A face authentication device, comprising:

Acquisition module: used to acquire a pair of face image samples;

Blocking module: used to block the acquired face image sample pairs;

a calculation module: configured to calculate a local gray feature vector and a local direction gradient histogram HOG vector for the obtained block;

The judging module is configured to calculate, according to the calculated local gray feature vector and the local HOG vector, whether the pair of face image samples belong to the same person.
The face authentication device according to claim 7, wherein the blocking module is further configured to perform multi-scale and overlapping segmentation on the acquired face image sample pairs.
The face authentication device according to claim 7, wherein the calculation module is further configured to calculate a local gray vector cosine distance as a gray feature vector for the obtained block, and calculate a local HOG vector cosine distance As a HOG vector.
The face authentication device according to claim 7, wherein the determining module is further configured to process the calculated local gray feature vector and the local HOG vector by using an artificial neural network classification model of softmax regression, It is calculated to determine whether the face image sample pair belongs to the same person.