CN103778416B - A kind of self adaptation approaches facial image production method - Google Patents

Abstract

The invention discloses a kind of self adaptation and approach facial image production method, comprise the steps: that (1) initializes single sample image training set, the colorized face images of acquisition is converted to Gray Face image.(2) triangle decomposition is carried out to Gray Face image, it is thus achieved that the vector representation form of Gray Face image.(3) by selecting a number of basic image, produce Gray Face image approaches image.(4) by all of Gray Face image and approach image and form new training set.The present invention has taken into full account the individual information of Gray Face image, triangle decomposition by Gray Face image, produce Gray Face image approaches image, improve the quantity of the Gray Face image that can be used for training to greatest extent, thus more efficiently improve the performance of single sample face identification system.

Description

A kind of self adaptation approaches facial image production method

Technical field

The invention belongs to area of pattern recognition, relate to the self adaptation in single sample recognition of face problem and approach facial image product Generation method.

Background technology

Approaching facial image generation technology to be mainly used in single sample recognition of face problem, it is in information security, traffic Video monitoring, the field such as identity information identification has a wide range of applications demand, as in terms of traffic monitoring, current friendship Logical video monitoring system can contain video and the pictorial information of facial image by Real-time Collection in a large number, leads to these facial images Cross and approach facial image production method, multiple facial images of same person can be produced, thus improve recognition of face system The recognition accuracy of system, can implement arrest in time or position in real time to suspicious one's share of expenses for a joint undertaking.

Approach facial image production method and come from such a fact, i.e. under the restriction of the factor such as environment or condition, often Individual is merely able to collect a facial image.As at some traffic block ports, a photograph only taken to each passing vehicle Piece, collecting each driver's facial image is one；Identity card, passport, an only face on the certificate such as driving license Photo.Carry out recognition of face in these cases and broadly fall into single sample recognition of face problem.Approaching image generation technology is A kind of technology being produced multiple images by single sample image, it is intended to by removing some secondary information in single sample image, Retain the main human face image information of major part, produce multiple sample；Under the conditions of multisample, face identification system energy Enough accuracys rate significantly improving identification.

In recent years, approaching facial image generation technology becomes a hot research problem of single sample field of face identification, Illustrate the feasibility of method in theory, more propose and developed many sides having display meaning and using value Method.Two classes can be divided into by its producing method: based on method and the method based on single individual of sample of priori. It is difficult to ensure that based on the picture quality approaching image producing method generation of priori, and need to consume substantial amounts of storage Space and calculating time, unsatisfactory in actual applications.The method decomposed based on single individual of sample, has relatively Fast speed, produces picture quality reliable, but its efficiency producing image and quality can also improve further.

Content of the invention

Technical problem: it is an object of the invention to provide a kind of number of training that can increase single sample face identification system Amount, it is adaptable to the self adaptation of different face recognition algorithms approaches facial image production method.

Technical scheme: the self adaptation of the present invention approaches facial image production method, comprises the steps:

Step 1: gathering face coloured image, and being converted to Gray Face image, idiographic flow is as follows:

Step 1.1: gather the face coloured image of n different people, everyone one, every width face coloured image big Little for h_c×w_c× 3, wherein c is face coloured image numbering, c=1,2 ..., n, h_cRepresent c width facial image square The line number of battle array, w_cRepresent c width facial image matrix column number；

Step 1.2: the n in step 1.1 opens colorized face images being separately converted to size is h_c×w_cGray Face Image X_c；

Step 2: carry out triangle decomposition to processing the Gray Face image obtaining in step 1, then ask for Gray Face The vector representation form of the triangle decomposition of image, idiographic flow is as follows:

Step 2.1: by matrix L U triangle decomposition method, to processing the Gray Face image X obtaining in step 1_cEnter Row triangle decomposition, obtains triangle decomposition formula as follows:

$X_c=P_c^T{L}_c{U}_c$

Wherein P_cFor Gray Face image X_cTriangle decomposition permutation matrix, size is h_c×h_c, " T " represents to square Battle array seeks transposition,Represent P_cTransposed matrix；L_cFor Gray Face image X_cDecompose the lower triangular matrix obtaining, Its size is h_c×w_c, U_cFor Gray Face image X_cDecomposing the upper triangular matrix obtaining, size is w_c×w_c；

Step 2.2: the lower triangular matrix L that will obtain in step 2.1_cWith upper triangular matrix U_cIt is written respectively as vector representation Form, i.e.

$L_c=[l_1,l_2,...,l_i,...,l_{w_c}]$

$U_c=\left(\begin{array}{c}{u}_1\\ u_2\\ .\\ .\\ .\\ u_j\\ .\\ .\\ .\\ u_{w_c}\end{array}\right)$

Wherein l_iFor h_cDimensional vector, i is lower triangular matrix L_cColumn vector numbering, i=1,2 ..., w_c；u_jFor w_cDimension Row vector, j is upper triangular matrix U_cRow vector numbering, j=1,2 ..., w_c；

Step 2.3: the lower triangular matrix L that will obtain in step 2.2_cVector representation form and upper triangular matrix U_c's Vector representation form substitutes into X in step 2.1_cTriangle decomposition formula, obtain Gray Face image X_cTriangle decomposition Vector representation form is as follows:

$X_c=P_c^T(\underset{i=j=1}{\overset{w_c}{\mathrm{\Σ}}}{l}_i\·u_j);$

Step 3: ask for Gray Face image X_cApproach image, idiographic flow is as follows:

Step 3.1: obtain Gray Face image X according to following formula_cThe basic image B of triangle decomposition_i:

$B_i=P_c^T(l_i\·u_j),i=j;$

Step 3.2: determine and be used for producing Gray Face image X_cQuantity k of the basic image approaching image；

Step 3.3: choose Gray Face image X_cFront k the basic image of vector representation form of triangle decomposition, root Produce Gray Face image X according to following formula_cApproach image

$X_c^1=P_c^T(\underset{i=j=1}{\overset{k}{\mathrm{\Σ}}}{l}_i\·u_j);$

Step 4: the Gray Face image X that step 3 is obtained_cApproach imageWith former Gray Face image X_cGroup Become a new setTraining set as single sample face identification system.

In a kind of preferred version of the inventive method, step 3.2 is calculate for producing Gray Face figure according to following formula As X_cQuantity k of the basic image approaching image:

WhereinRepresent and round to zero, a=min{h_c,w_c, b=max{h_c,w_c}。

In a kind of preferred version of the inventive method, step 3 obtains Gray Face image X_cApproach imageAs First width approaches image, after step 3 completes, reuses facial image transposition method, asks for Gray Face image X_c The second width approach image, then by described Gray Face image X in described step 4_cThe first width approach imageApproach image with the second widthWith former Gray Face image X_cForm a new set Training set as single sample face identification system.

In the inventive method preferred version a kind of embodiment in, Gray Face image X_cThe second width approach image and be Obtain according to following flow process:

A: by Gray Face image X_cTransposition, obtains the transposition image of Gray Face imageIts size is w_c×h_c；

B: according to the transposition image to Gray Face image for the following formulaCarry out matrix L U triangle decomposition:

$X_c^T={\stackrel{\‾}{P}}_c^T{\stackrel{\‾}{L}}_c{\stackrel{\‾}{U}}_c$

WhereinTransposition image for Gray Face imageTriangle decomposition permutation matrix, its size is w_c×w_c, " T " represents to Matrix Calculating transposition,RepresentTransposed matrix；Transposition image for Gray Face image Decomposing the lower triangular matrix obtaining, its size is w_c×h_c,Transposition image for Gray Face imageDecompose The upper triangular matrix obtaining, size is h_c×h_c；

C: by the lower triangular matrix in step bAnd upper triangular matrixIt is written respectively as vector representation form, i.e.

${\stackrel{\‾}{L}}_c=[{\stackrel{\‾}{l}}_1,{\stackrel{\‾}{l}}_2,...,{\stackrel{\‾}{l}}_i,...,{\stackrel{\‾}{l}}_{w_c}]$

${\stackrel{\‾}{U}}_c=\left(\begin{array}{c}{\stackrel{\‾}{u}}_1\\ {\stackrel{\‾}{u}}_2\\ .\\ .\\ .\\ {\stackrel{\‾}{u}}_j\\ .\\ .\\ .\\ {\stackrel{\‾}{u}}_{h_c}\end{array}\right)$

WhereinFor w_cDimensional vector, i is lower triangular matrixColumn vector numbering, i=1,2 ..., h_c；For h_c Dimension row vector, j is upper triangular matrixRow vector numbering, j=1,2 ..., h_c；

D: by the lower triangular matrix in step cAnd upper triangular matrixVector representation form substitute in step bTriangle decomposition formula, obtain the transposition image of Gray Face imageTriangle decomposition vector representation form such as Under:

$X_c^T={\stackrel{\‾}{P}}_c^T(\underset{i=j=1}{\overset{h_c}{\mathrm{\Σ}}}{\stackrel{\‾}{l}}_i\·{\stackrel{\‾}{u}}_j);$

E: choose Gray Face imageFront k the basic image of vector representation form of triangle decomposition, according to following formula Produce Gray Face image X_cThe second width approach facial image

$X_c^2={\left({{\stackrel{\‾}{P}}_c}^T(\underset{i=j=1}{\overset{k}{\mathrm{\Σ}}}{\stackrel{\‾}{l}}_i\·{\stackrel{\‾}{u}}_j)\right)}^T.$

Beneficial effect: compared with prior art, the invention have the advantages that

The inventive method is better than traditional method in terms of producing picture quality and computational efficiency, single sample people for raising The accuracy rate of face identification system has important value.

In order to obtain produce that speed is fast, quality is high approaches facial image, meets the reality of single sample face identification system Application requires, the invention provides the adaptive image rebuilding method that approaches, the method utilizes former single sample image to produce Approach image, by removing some secondary information in single sample image, retain the main human face image information of major part, produce That gives birth to former single sample approaches image.Not only can ensure that generation approaches picture quality reliably, and recognition of face system can be made System significantly improves the accuracy rate identifying under single sample conditions.

1. the self adaptation that the present invention proposes approaches facial image production method, after being decomposed for difference list sample image, The quantity of basic image is different, uses adaptive base amount of images system of selection, with ensure to produce approach the quality of image can Lean on；

2. the self adaptation that the present invention proposes approaches facial image production method, in order to as far as possible many generations approach picture number Amount, it is proposed that single sample facial image and transposition image thereof are separately disassembled into two independent basic image collection, then base In the basic image collection of the two independence, use adaptive approach to produce two of single sample respectively and approach facial image；

3. the present invention is from actual application efficiency, proposes self adaptation and approaches facial image production method, and the method is adopted High by decomposition efficiency, computational methods simple matrix triangle decomposition.The method that the present invention proposes is higher than in computational efficiency The method such as singular value decomposition and ORTHOGONAL TRIANGULAR DECOMPOSITION；

Brief description

Fig. 1 is the inventive method overall flow figure.

Fig. 2 approaches the flow chart of image by original single sample for being produced.

Detailed description of the invention

Below in conjunction with Figure of description and embodiment, carry out clearly intactly illustrating to the detailed process of the inventive method.

The self adaptation of the present invention approaches facial image production method, and its flow chart is as it is shown in figure 1, specifically according to following step Suddenly carry out:

Step 1: gathering face coloured image, and being converted to Gray Face image, idiographic flow is as follows:

Step 1.1: gather the face coloured image of n different people, everyone one, every width face coloured image big Little for h_c×w_c× 3, wherein c is face coloured image numbering, c=1,2 ..., n, h_cRepresent c width facial image square The line number of battle array, w_cRepresent c width facial image matrix column number；

Step 1.2: the n in step 1.1 opens colorized face images being separately converted to size is h_c×w_cGray Face Image X_c；The common method that colorized face images is converted into Gray Face image has two kinds, i.e. seeks colored human face figure In Xiang, the mean value of each monochrome obtains Gray Face image and the weighted average seeking each monochrome in colorized face images is worth To Gray Face image, use method known to both can realize that colorized face images is converted into Gray Face figure Picture；

Step 2: carry out triangle decomposition to processing the Gray Face image obtaining in step 1, then ask for Gray Face The vector representation form of the triangle decomposition of image, idiographic flow is as follows:

Step 2.1: conventional triangle decomposition includes that LU decomposes and LDV decomposes two kinds；By matrix L U triangle decomposition Method, to processing the Gray Face image X obtaining in step 1_cCarry out triangle decomposition, obtain triangle decomposition formula as follows:

$X_c=P_c^T{L}_c{U}_c$

Wherein P_cFor Gray Face image X_cTriangle decomposition permutation matrix, size is h_c×h_c, " T " represents to square Battle array seeks transposition,Represent P_cTransposed matrix；L_cFor Gray Face image X_cDecompose the lower triangular matrix obtaining, Its size is h_c×w_c, U_cFor Gray Face image X_cDecomposing the upper triangular matrix obtaining, size is w_c×w_c；

Step 2.2: the lower triangular matrix L that will obtain in step 2.1_cWith upper triangular matrix U_cIt is written respectively as vector representation Form, i.e.

$L_c=[l_1,l_2,...,l_i,...,l_{w_c}]$

$U_c=\left(\begin{array}{c}{u}_1\\ u_2\\ .\\ .\\ .\\ u_j\\ .\\ .\\ .\\ u_{w_c}\end{array}\right)$

Wherein l_iFor h_cDimensional vector, i is lower triangular matrix L_cColumn vector numbering, i=1,2 ..., w_c；u_jFor w_cDimension Row vector, j is upper triangular matrix U_cRow vector numbering, j=1,2 ..., w_c；

Step 2.3: the lower triangular matrix L that will obtain in step 2.2_cVector representation form and upper triangular matrix U_c's Vector representation form substitutes into X in step 2.1_cTriangle decomposition formula, obtain Gray Face image X_cTriangle decomposition Vector representation form is as follows:

$X_c=P_c^T(\underset{i=j=1}{\overset{w_c}{\mathrm{\Σ}}}{l}_i\·u_j);$

Step 3: ask for Gray Face image X_cApproach image, its flow chart as in figure 2 it is shown, idiographic flow such as Under:

Step 3.1: obtain Gray Face image X according to following formula_cThe basic image B of triangle decomposition_i:

$B_i=P_c^T(l_i\·u_j),i=j;$

Step 3.2: determine and be used for producing Gray Face image X_cQuantity k of the basic image approaching image；

Step 3.3: choose Gray Face image X_cFront k the basic image of vector representation form of triangle decomposition, front K basic image is $B_1=P_c^T(l_1\·u_1),B_2=P_c^T(l_2\·u_2),...,B_k=P_c^T(l_k\·u_k);$ Produce gray scale according to following formula Facial image X_cApproach image

$X_c^1=P_c^T(\underset{i=j=1}{\overset{k}{\mathrm{\Σ}}}{l}_i\·u_j);$

Step 4: the Gray Face image X that step 3 is obtained_cApproach imageWith former Gray Face image Xc group Become a new setTraining set as single sample face identification system.

In a preferred embodiment of the present invention, step 3.2 is calculate for producing Gray Face image according to following formula X_cQuantity k of the basic image approaching image:

WhereinRepresent and round to zero, a=min{h_c,w_c, b=max{h_c,w_c}。

In a preferred embodiment of the present invention, step 3 obtains Gray Face image X_cApproach imageAs One width approaches image, after step 3 completes, reuses facial image transposition method, asks for Gray Face image X_c's Second width approaches image, then in step 4 by Gray Face image X_cThe first width approach imageWith the second width Approach imageWith former Gray Face image X_cForm a new setAs single sample people The training set of face identification system.

In above preferred embodiment of the present invention, Gray Face image X_cThe second width to approach image can be according to as follows Flow process obtains:

A: by Gray Face image X_cTransposition, obtains the transposition image of Gray Face imageIts size is w_c×h_c；

B: according to the transposition image to Gray Face image for the following formulaCarry out matrix L U triangle decomposition:

$X_c^T={\stackrel{\‾}{P}}_c^T{\stackrel{\‾}{L}}_c{\stackrel{\‾}{U}}_c$

WhereinTransposition image for Gray Face imageTriangle decomposition permutation matrix, its size is w_c×w_c, " T " represents to Matrix Calculating transposition,RepresentTransposed matrix；Transposition image for Gray Face image Decomposing the lower triangular matrix obtaining, its size is w_c×h_c,Transposition image for Gray Face imageDecompose The upper triangular matrix obtaining, size is h_c×h_c；

C: by the lower triangular matrix in step bAnd upper triangular matrixIt is written respectively as vector representation form, i.e.

${\stackrel{\‾}{L}}_c=[{\stackrel{\‾}{l}}_1,{\stackrel{\‾}{l}}_2,...,{\stackrel{\‾}{l}}_i,...,{\stackrel{\‾}{l}}_{h_c}]$

${\stackrel{\‾}{U}}_c=\left(\begin{array}{c}{\stackrel{\‾}{u}}_1\\ {\stackrel{\‾}{u}}_2\\ .\\ .\\ .\\ {\stackrel{\‾}{u}}_j\\ .\\ .\\ .\\ {\stackrel{\‾}{u}}_{h_c}\end{array}\right)$

WhereinFor w_cDimensional vector, i is lower triangular matrixColumn vector numbering, i=1,2 ..., h_c；For h_c Dimension row vector, j is upper triangular matrixRow vector numbering, j=1,2 ..., h_c；

D: by the lower triangular matrix in step cAnd upper triangular matrixVector representation form substitute in step bTriangle decomposition formula, obtain the transposition image of Gray Face imageTriangle decomposition vector representation form such as Under:

$X_c^T={\stackrel{\‾}{P}}_c^T(\underset{i=j=1}{\overset{h_c}{\mathrm{\Σ}}}{\stackrel{\‾}{l}}_i\·{\stackrel{\‾}{u}}_j);$

E: choose Gray Face imageFront k the basic image of vector representation form of triangle decomposition, according to following formula Produce Gray Face image X_cThe second width approach facial image

$X_c^2={\left({{\stackrel{\‾}{P}}_c}^T(\underset{i=j=1}{\overset{k}{\mathrm{\Σ}}}{\stackrel{\‾}{l}}_i\·{\stackrel{\‾}{u}}_j)\right)}^T.$

Claims (3)

1. a self adaptation approaches facial image production method, it is characterised in that the method comprises the steps:

Step 1: gathering face coloured image, and being converted to Gray Face image, idiographic flow is as follows:

Step 1.1: gather the face coloured image of n different people, everyone one, every width face coloured image big Little for h_c×w_c× 3, wherein c is face coloured image numbering, c=1,2 ..., n, h_cRepresent c width facial image square The line number of battle array, w_cRepresent c width facial image matrix column number；

Step 1.2: the n in described step 1.1 opens colorized face images being separately converted to size is h_c×w_cGray scale Facial image X_c；

Step 2: carry out triangle decomposition to processing the Gray Face image obtaining in described step 1, then ask for gray scale The vector representation form of the triangle decomposition of facial image, idiographic flow is as follows:

Step 2.1: by matrix L U triangle decomposition method, to processing the Gray Face image X obtaining in described step 1_c Carry out triangle decomposition, obtain triangle decomposition formula as follows:

$X_c=P_c^T{L}_c{U}_c$

Wherein P_cFor Gray Face image X_cTriangle decomposition permutation matrix, size is h_c×h_c, " T " represents to square Battle array seeks transposition,Represent P_cTransposed matrix；L_cFor Gray Face image X_cDecompose the lower triangular matrix obtaining, Its size is h_c×w_c, U_cFor Gray Face image X_cDecomposing the upper triangular matrix obtaining, size is w_c×w_c；

Step 2.2: the lower triangular matrix L that will obtain in described step 2.1_cWith upper triangular matrix U_cIt is written respectively as vector Representation, i.e.

$L_c=[l_1,l_2,...,l_i,...,l_{w_c}]$

$U_c=\left(\begin{array}{c}{u}_1\\ u_2\\ .\\ .\\ .\\ u_j\\ .\\ .\\ .\\ u_{w_c}\end{array}\right)$

Wherein l_iFor h_cDimensional vector, i is lower triangular matrix L_cColumn vector numbering, i=1,2 ..., w_c；u_jFor w_cDimension Row vector, j is upper triangular matrix U_cRow vector numbering, j=1,2 ..., w_c；

Step 2.3: the lower triangular matrix L that will obtain in described step 2.2_cVector representation form and upper triangular matrix U_cVector representation form substitute into X in step 2.1_cTriangle decomposition formula, obtain Gray Face image X_cTriangle divide The vector representation form solving is as follows:

$X_c=P_c^T(\underset{i=j=1}{\overset{w_c}{\Σ}}{l}_i\·u_j);$

Step 3: ask for Gray Face image X_cApproach image, idiographic flow is as follows:

Step 3.1: obtain Gray Face image X according to following formula_cThe basic image B of triangle decomposition_i:

$B_i=P_c^T(l_i\·u_j),i=j;$

Step 3.2: determine and be used for producing Gray Face image X_cQuantity k of the basic image approaching image；

Step 3.3: choose Gray Face image X_cFront k the basic image of vector representation form of triangle decomposition, root Produce Gray Face image X according to following formula_cApproach image

$X_c^1=P_c^T(\underset{i=j=1}{\overset{k}{\Σ}}{l}_i\·u_j);$

Step 4: the Gray Face image X that described step 3 is obtained_cApproach imageWith former Gray Face image X_cForm a new setTraining set as single sample face identification system.

2. self adaptation according to claim 1 approaches facial image production method, it is characterised in that described step It is to calculate for producing Gray Face image X according to following formula in rapid 3.2_cQuantity k of the basic image approaching image:

WhereinRepresent and round to zero, a=min{h_c,w_c, b=max{h_c,w_c}。

3. self adaptation according to claim 1 approaches facial image production method, it is characterised in that described step The Gray Face image X obtaining in rapid 3_cApproach imageApproach image as the first width, after step 3 completes, Reuse facial image transposition method, ask for Gray Face image X_cThe second width approach image, then in step 4 Middle by described Gray Face image X_cThe first width approach imageApproach image with the second widthWith former gray scale people Face image X_cForm a new setTraining set as single sample face identification system；

Described Gray Face image X_cThe second width approach image and obtain according to following flow process:

A: by Gray Face image X_cTransposition, obtains the transposition image of Gray Face imageIts size is w_c×h_c；

B: according to the transposition image to Gray Face image for the following formulaCarry out matrix L U triangle decomposition:

$X_c^T={\stackrel{\‾}{P}}_c^T{\stackrel{\‾}{L}}_c{\stackrel{\‾}{U}}_c$

WhereinTransposition image for Gray Face imageTriangle decomposition permutation matrix, its size is w_c×w_c, " T " represents to Matrix Calculating transposition,RepresentTransposed matrix；Transposition image for Gray Face image Decomposing the lower triangular matrix obtaining, its size is w_c×h_c,Transposition image for Gray Face imageDecompose The upper triangular matrix obtaining, size is h_c×h_c；

C: by the lower triangular matrix in step bAnd upper triangular matrixIt is written respectively as vector representation form, i.e.

${\stackrel{\‾}{L}}_c=\[{\stackrel{\‾}{l}}_1,{\stackrel{\‾}{l}}_2,\mathrm{..},{\stackrel{\‾}{l}}_i,\mathrm{...},{\stackrel{\‾}{l}}_{h_c}\]$

${\stackrel{\‾}{U}}_c=\left(\begin{array}{c}{\stackrel{\‾}{u}}_1\\ {\stackrel{\‾}{u}}_2\\ .\\ .\\ .\\ {\stackrel{\‾}{u}}_j\\ .\\ .\\ .\\ {\stackrel{\‾}{u}}_{h_c}\end{array}\right)$

WhereinFor w_cDimensional vector, i is lower triangular matrixColumn vector numbering, i=1,2 ..., h_c；For h_c Dimension row vector, j is upper triangular matrixRow vector numbering, j=1,2 ..., h_c；

D: by the lower triangular matrix in step cAnd upper triangular matrixVector representation form substitute in step bTriangle decomposition formula, obtain the transposition image of Gray Face imageTriangle decomposition vector representation form such as Under:

$X_c^T={\stackrel{\‾}{P}}_c^T(\underset{i=j=1}{\overset{h_c}{\Σ}}{\stackrel{\‾}{l}}_i\·{\stackrel{\‾}{u}}_j);$

E: choose Gray Face imageFront k the basic image of vector representation form of triangle decomposition, according to following formula Produce Gray Face image X_cThe second width approach facial image

$X_c^2={\left({{\stackrel{\‾}{P}}_c}^T\right(\underset{i=j=1}{\overset{k}{\Σ}}{\stackrel{\‾}{l}}_i\·{\stackrel{\‾}{u}}_j\left)\right)}^T.$