CN103886235B - Face image biological key generating method - Google Patents

Abstract

The invention provides a face image biological key generating method. The face image biological key generating method includes steps of subjecting face images of users to characteristic space conversion, projecting the face images into a higher space, stabilizing facial feature information into an acceptable fluctuation range in the higher space, extracting digital sequences from stabilized feature vectors, and encoding in the digital sequences to generate a biological key. By adopting the face image biological key generating method, no face information of the users is required to be stored in mobile terminals and identification servers, and the face images of the users are also dispensed with transmission in the network. The users generate (user names and secret keys) pairs by acquiring the face images of their own and perform network identity authentication by various authentication methods derived from the (user names and secret keys) pairs. The private data of the users are subjected to direct encipherment protection by the method supporting the face biological secret key, and the method can be extended to application in the field of cloud storage safety. As long as key space of the face biological secret keys is larger enough, high safety can be guaranteed.

Description

A biometric key generation method for frontal face images

technical field

The invention belongs to the technical field of information security, and in particular relates to a method for extracting a human face biological key through a front face image. It can directly generate a biological key by ingesting a face image, which makes the face-based network identity authentication process more simplified and flexible, and provides a new authentication method for network identity authentication technology. Expand the application of face recognition technology in network security.

Background technique

In recent years, face recognition technology based on machine vision has received widespread attention because it can be widely used in daily life. One of the important applications is identity authentication. Face check-in systems and network-side identity authentication are all good examples of the use of face recognition technology. In the field of network identity authentication, face authentication adopts the following authentication mode: 1) Collect user's face, extract the face feature template, and store it in the remote network authentication server; The user's face generates face features and transmits them to the remote network authentication server; 3) The authentication server compares the user's face features with the stored face feature templates. If they are consistent, the authentication will pass, and if they are inconsistent, the authentication will fail. This authentication mode has certain disadvantages: it needs to store and transmit user face information in an unsafe network environment. If the user is not familiar with the authentication terminal, generally he will not cooperate with uploading his own face image. This limits the promotion of face authentication in network identity authentication. In addition, in an environment where cloud storage is booming, since face authentication does not support the encryption of face biometric information, users cannot use their own biometrics to protect their private data. This limits the development of face recognition technology in the field of information security, and face recognition technology fails to play its due role in cloud storage security.

Some researchers once proposed the concept of biological key, hoping to obtain a stable biological key sequence directly from biological characteristics. However, the current research does not give complete face feature information, and there is no practical face bio-key technology in actual production and life.

Contents of the invention

The invention proposes a method for generating a face biological key. The method transforms the user's frontal face image into the high-dimensional space after the feature space transformation, stabilizes the face feature information in the high-dimensional space to an acceptable fluctuation range, and then extracts the digital sequence from the stabilized feature vector, Generate a bio-key encoded from a sequence of numbers. The whole method does not need to store the user's face information on the mobile terminal or the authentication server side, and does not need to transmit the user's face image in the network. Users generate (username, key) pairs locally by collecting their own frontal face images, and perform network identity authentication through various authentication methods derived from (username, key) pairs. The method also supports directly encrypting and protecting user's private data with the face biological key, and can be extended to the application in the field of cloud storage security. As long as the key space of the face biometric key is large enough, high security can be guaranteed. The length of the human face biological key sequence extracted by the present invention is adjustable and can be greater than 256 bits.

The face biological key generation is divided into two parts, the first part is the face biological key training part, and the second part is the face biological key extraction part.

The specific steps of face biometric key training are as follows:

In the first step, the user collects the frontal face image through the camera, and the frontal face is adjusted by the user. The collection environment requires an indoor environment with sufficient light. Both natural light and artificial light source are acceptable, and the collection is repeated more than 8 times.

The second step is to grayscale the face image and adjust the size of the image to a uniform size of 128×128 or 64×64, which can be determined by the user based on experience.

The third step is to seek the self-quotient map of the face image, and obtain more than 8 self-quotient map images.

The fourth step is to organize more than 8 self-quotient images into a sample matrix and perform principal component analysis (PCA) processing to obtain the projection matrix of the face in the feature space, which is denoted as P1. More than 8 self-quotient images are projected through the projection matrix P1 to obtain the face feature vector. Organize the obtained eigenvectors into an M×D dimensional eigenvector matrix, denoted as S1, M is the number of self-quotient image images, D is the number of projected eigenvector elements, and generally D>M.

The fifth step is to expand the matrix S1 into two matrices, the L×L-dimensional random error square matrix EX, and the L×L-dimensional standard value square matrix EY, where L>D (the specific construction method is described in the specific implementation).

The sixth step is to solve the generalized inverse matrix of EX, denoted as IEX, multiply IEX by the matrix EY to the left to obtain the high-dimensional space projection matrix PEX=IEX×EY of the face feature vector, and store the projection matrices P1 and PEX on the user end.

Face biometric key training is complete.

The specific steps of face biometric key extraction are as follows:

In the first step, the user collects the frontal face image through the camera, and the frontal face is adjusted by the user. The collection environment requires an indoor environment with sufficient light, either natural light or artificial light source.

The second step is to grayscale the face image, and adjust the size of the image to be consistent with the size set during the training of the face biometric key.

The third step is to find the self-quotient map of the face image.

The fourth step is to convert the self-quotient image into a row vector, take the projection matrix P1 stored during the face biological key training, and multiply the projection matrix P1 to the left to get the feature vector of the face in the feature space, which is recorded as Z, and the length is for D.

The fifth step is to expand the vector Z into a 1×L-dimensional matrix EZ, and multiply the PEX matrix to the left to obtain a 1×L-dimensional vector ED; the expansion method is consistent with the face biometric key training.

The sixth step is to use the checkerboard method to further stabilize the values in the vector ED, and take the value of the previous DL values to be 1×DL-dimensional vector EE, where DL≤D. The value of each element in the vector EE is concatenated back and forth to generate a face biometric key. (The chessboard method is described in the specific implementation manner)

Beneficial effects of the present invention: the present invention proposes a method for generating a face biological key. It can change the traditional network biometric identity authentication mode. Users do not need to store personal face information on the mobile terminal or authentication server, and do not need to transmit face images in the network to complete face-based network identity authentication. At the same time, the method also supports directly encrypting and protecting the user's private data with the face biometric key, which can be extended to the application in the field of cloud storage security.

Description of drawings

Figure 1 is a flow chart of face biometric key extraction.

Figure 2 is a schematic diagram of the effect of the face self-quotation map.

Figure 3 is a schematic diagram of converting a face image into a row vector.

Figure 4 is a flowchart of face principal component analysis.

Fig. 5 is a schematic diagram of the process of generating a face biometric key.

detailed description

The present invention will be further described below in conjunction with accompanying drawing.

The current face recognition technology cannot correctly identify faces with multiple poses and complex lighting conditions. The preconditions for extracting the human face biological key in the present invention are a frontal face, sufficient illumination, and no multi-light source and multi-angle irradiation. Under this premise, face recognition has a high accuracy rate. After a series of stabilization processes, the biological key can be extracted from it. The extraction process of the face biometric key is shown in Figure 1.

Illumination conditions have a great influence on the existing face recognition effect, and the influence of illumination needs to be eliminated first. The present invention uses the self-quotient graph method to eliminate the influence of illumination, and the effect is shown in FIG. 2 .

The face biological key generation method proposed by the present invention is divided into two parts, the first part is the human face biological key training part, and the second part is the human face biological key extraction part.

The specific steps of face biometric key training are as follows:

In the first step, the user collects the frontal face image through the camera, and the frontal face is adjusted by the user. The collection environment requires an indoor environment with sufficient light. Both natural light and artificial light source are acceptable, and the collection is repeated more than 8 times.

The second step is to grayscale the face image. The grayscale processing formula is

f(i,j)=0.30R(i,j)+0.59G(i,j)+0.11B(i,j) (1)

R, G, and B are three components of red, green, and blue. Resize the image to a uniform size of 128×128, or 64×64 or the user can decide by experience.

The third step is to seek the self-quotient map of the face image, and obtain more than 8 self-quotient map images. The formula for calculating the self-quotient graph is

in Is the smoothed version of I, F is the filter kernel, and * is the convolution operator. Generally take the Gaussian filter kernel, the 3rd order and the 5th order can be used, such as F is taken

The fourth step is to organize more than 8 self-quotient images into a sample matrix. Note that there are M self-quotient graph images in total. A single image is spliced forward and backward in row units to form a one-dimensional row vector containing n elements, as shown in Figure 3; after all self-quotient images are converted into row vectors, they are arranged sequentially in row units from row 1 to row M , to get M×n-dimensional sample matrix S.

Perform principal component analysis (PCA) on the sample matrix. The analysis process is shown in Figure 4.

The sample matrix is centralized, and each column of the matrix is divided by the corresponding mean value. The formula is:

S _ij is any component of the sample matrix.

Calculate the sample matrix covariance matrix;

Diagonalize the covariance matrix C. First, the eigenvalue decomposition of C is performed, and the eigenvector matrix is obtained and orthogonalized to be P. P satisfies:

P ^T CP = Λ (5)

Λ is a diagonal matrix, and the values of the diagonal elements are the eigenvalues of matrix C.

Take the eigenvectors corresponding to the largest first D (D<n) eigenvalues to form a new eigenvector matrix P1 (n×D dimension), P1 is the projection matrix of the face in the feature space.

The M self-quotient images are respectively projected through the projection matrix P1, and the calculation formula is

S1=S×P1 (6)

S1 is an M×D dimensional feature vector matrix, D is the number of feature vector elements after projection, and generally D>M.

The fifth step is to expand the matrix S1 into two matrices, the L×L dimension random error square matrix EX, the L×L dimension EY, L>D, and the construction method is:

Take the M row vectors of the matrix S1, calculate the mean value, and obtain the mean value vector EB (1×D dimension);

Set the fluctuation range Er, such as Er=10; add random error disturbance to EB, the calculation formula is

S1 _j represents the jth row in the S1 matrix, and EX _j represents a row vector; the rand(0,1) function returns a random number between (0,1); assemble EX _j into an L×D dimensional matrix.

Construct LD nonlinear functions, the input variable is a one-dimensional row vector (x ₁ ,x ₂ ,…,x _D ), D elements, and the output is a one-dimensional row vector (x ₁ ,x ₂ ,…,x _D ,… , x _L ), L elements. The nonlinear function can be defined by the user. As an example, the following nonlinear function can be taken

Z(t)=(x ₁ -x ₂ )×sin(t)+(t ^{^} 2)×(x ₃ %10)(t is an integer, 0<t<LD)(8)

sin(t) trigonometric function, (t ^{^} 2) means the square of t, (x ₃ %10) means x ₃ modulo 10 operation.

Use the constructed Z(t) to operate on EX _j , and j traverses 1~L to get an L×L dimensional matrix, which is the random error square matrix EX.

The EY construction method is:

Repeat the mean value vector EB for L rows to obtain an L×D dimensional matrix, which is denoted as EYt. Use Z(t) to operate EYt _j , and j traverses 1-L to obtain an L×L dimensional matrix, which is the standard value square matrix EY.

The sixth step is to solve the generalized inverse matrix of EX, denoted as IEX, multiply IEX by the matrix EY to the left to obtain the high-dimensional space projection matrix PEX=IEX×EY of the face feature vector, and store the projection matrices P1 and PEX on the user end.

Face biometric key training is complete.

Figure 5 shows the change of digital sequence flow in the face biometric key extraction process, and the specific steps are as follows:

In the first step, the user collects the frontal face image through the camera, and the frontal face is adjusted by the user. The collection environment requires an indoor environment with sufficient light, and natural light or artificial light can be used.

The second step is to grayscale the face image, and adjust the size of the image to be consistent with the size set during the training of the face biometric key.

The third step is to find the self-quotient map of the face image. The method is the same as that of face biometric key training.

The fourth step is to convert the self-quotient map image into a row vector, and the method is the same as that of the face biological key training part. Take the projection matrix P1 stored during face biological key training, and multiply the projection matrix P1 to the left to get the feature vector of the face in the feature space, denoted as Z, and the length is D.

The fifth step is to expand the vector Z into a 1×L-dimensional matrix EZ. The extension method is consistent with the face biometric key training. For example, the extension function is taken as the extension function Z(t) described in the fifth step of the face biometric key training part.

The matrix EZ is multiplied by PEX to the left to obtain a 1×L-dimensional vector ED;

The sixth step is to use the checkerboard method to further stabilize the values in the vector ED. The chessboard method is described as follows:

Perform an operation on each element in ED (denoted as EDX _i ), the pseudocode is

mod() is a modulo function, and maxdis marks the grid size of the chessboard method, which is an odd number, and the specific value can be selected by the user based on experience.

Take the value of the previous DL number as 1×DL dimension vector EE, DL≤D. The value of each element in the vector EE is concatenated back and forth to generate a face biometric key. DL is selected by the user according to needs.

Those of ordinary skill in the art should recognize that the above embodiments are only used to illustrate the present invention, rather than as a limitation to the present invention, as long as they are within the essential scope of the present invention, changes and modifications to the above embodiments will be Fall within the protection scope of the present invention.

Claims (1)

1. a kind of front face image biological secret key generation method is it is characterised in that the method comprises the following steps：Face is biological Key training part and face biological secret key extract part；

Face biological secret key training part concretely comprises the following steps：

The first step, user gathers front face image by photographic head, and frontal faces are voluntarily adjusted by user, and collection environmental requirement is in room The abundant environment of interior light, repeated acquisition more than 8 times；

Second step, carries out gray processing process to facial image, and adjustment image size is uniform sizes 128 × 128 or 64 × 64；

3rd step, face image of asking for help from quotient graph, obtain more than 8 width from quotient graph image；

4th step, more than 8 width will be made into sample matrix from quotient graph image sets, carries out principal component analysiss process, obtain face in spy Levy the projection matrix in space, be designated as P1；More than 8 width will project through projection matrix P1 respectively from quotient graph image, obtain face Characteristic vector；The characteristic vector tried to achieve is organized as the eigenvectors matrix of M × D dimension, is designated as S1, M is from quotient graph image Number, D is characteristic vector element number after projection；

5th step, matrix S1 is expanded to 2 matrixes, random error square formation EX that L × L ties up, standard value square formation EY of L × L dimension, L>D, building method is：

Take the M row vector of matrix S1, average, obtain the mean vector EB of 1 × D dimension；

Set fluctuation range Er, such as Er=10；Increase random error disturbance for EB, computing formula is

S1_jRepresent the jth row in S1 matrix, EX_jRepresent a row vector；Rand (0,1) function returns random between (0,1) Number；By EX_jIt is assembled into the matrix of L × D dimension with behavior unit；

L-D nonlinear function of construction, input variable is one-dimensional row vector (x₁,x₂,…,x_D), D element, it is output as one-dimensional row Vectorial (x₁,x₂,…,x_D..., x_L), L element；Nonlinear function, by user's self-defining, takes following nonlinear function：

Z (t)=(x₁-x₂)×sin(t)+(t^2)×(x₃%10), t is integer, 0<t<L-D；

Sin (t) trigonometric function, (t^2) represent t square, (x₃%10) represent x₃Mould 10 computing；

With construction Z (t) to EX_jEnter row operation, j travels through 1～L, obtain L × L dimension matrix, i.e. random error square formation EX；

EY building method is：Mean vector EB is repeated L row, obtains L × D dimension matrix, be designated as EYt；With Z (t) to EYt_jTransported Calculate, j travels through 1-L, obtain L × L dimension matrix, i.e. standard value square formation EY；

6th step, solves the generalized inverse matrix of EX, is designated as IEX, IEX premultiplication matrix EY is obtained the higher-dimension sky of face feature vector Between projection matrix PEX=IEX × EY, user side store projection matrix P1 and PEX；

The training of face biological secret key completes；

Face biological secret key extraction part concretely comprises the following steps：

1st step, user gathers front face image by photographic head, and frontal faces are voluntarily adjusted by user, and collection environmental requirement is in room The abundant environment of interior light；

2nd step, carries out gray processing process to facial image, when adjustment image size is that uniform sizes are trained with face biological secret key The consistent size setting；

3rd step, face image of asking for help is from quotient graph；

4th step, will switch to row vector from quotient graph image, take the projection matrix P1 storing during the training of face biological secret key, and premultiplication is thrown Shadow matrix P1, obtains characteristic vector in feature space for the face, is designated as Z, and length is D；

5th step, vector Z is expanded to 1 × L dimension matrix EZ, extended method is consistent when being trained with face biological secret key, takes expansion Exhibition function trains spread function Z (t) of part the 5th step description for face biological secret key；Premultiplication PEX matrix, obtains 1 × L dimensional vector ED；

6th step, carries out further stabilized treatment with chessboard method to the numerical value in vectorial ED, take front DL number be worth 1 × DL tie up to Amount EE, DL≤D；To splice before and after the numerical value of each element in vectorial EE, that is, generate face biological secret key.