CN103971096A - Multi-pose face recognition method based on MB-LBP features and face energy diagram - Google Patents

Abstract

The invention provides a multi-pose face recognition method based on MB-LBP features and a face energy diagram. According to the method, a multi-pose face image training library is established, and after size normalization processing is carried out on face images, a face mean value energy diagram and a variance energy diagram of the training library are established; MB-LBP feature extracting is carried out on the obtained face mean value energy diagram and the variance energy diagram, and the MB-LBP features are stored as matching library information; when face detection is carried out, the face images are detected and face regions are extracted, and size normalization processing is carried out on the images of the face regions to obtain standard face images; MB-LBP feature extracting is carried out on the standard face images; finally, classification and recognition of multi-pose faces are completed by adopting a nearest neighbor classifier. According to the multi-pose face recognition method based on the MB-LBP features and the face energy diagram, the intrinsic macroscopic features of the multi-pose faces can be maintained well, the microstructure and the macrostructure of a face image mode are reserved, influences brought by single pixel noise can be removed, needed storage space is small, and the method has an excellent recognition rate and recognition speed.

Description

A kind of multi-pose Face recognition methods based on MB-LBP feature and face energygram

Technical field

The invention belongs to biological characteristics identity recognizing technology field, particularly relate to a kind of multi-pose Face recognition methods based on MB-LBP feature and face energygram.

Background technology

Automatic face recognition technology has the peculiar advantages such as the convenience of collection, non-invasion with respect to other living things feature recognition method such as fingerprint, iris, thereby has application prospect and economic worth very widely.If recognition of face is divided by attitude, can be divided into forward sight recognition of face and multi-pose Face identification.Wherein, the technology of forward sight recognition of face is comparatively ripe.And many technical matterss such as multi-pose Face recognition methods still has that memory space is large, calculation of complex, discrimination are low.The research of multi-pose Face identification lags behind, and becomes restriction face recognition technology and really obtains one of major obstacle of practical application.Therefore, carry out the research of multi-pose Face identification significant to the popularization of face recognition technology.

Existing multi-pose Face identification, typical document is wherein patent " a kind of multi-pose Face recognition methods based on face average and variance energygram " (Wang Kejun, Zou Guofeng etc. Chinese invention patent: 201310122161.3[P] .2013-07-24.), adopt the facial image structure face energygram of different attitudes to be used for realizing multi-pose Face identification.But face energygram does not possess the periodicity of gait energygram, cannot represent that the human face posture of different luffing angles and different swing angles changes.

The people such as Liao have proposed (Liao S C, et al.Learning multi-scale block local binary patterns for facerecognition.In Proceedings of the2007International Conference on Biometrics.Seoul, SouthKorea:Springer, 2007.828-837.) the face identification method based on multiple dimensioned local binary patterns (Multi-scale Block Local BinaryPattens, MB-LBP).Although MB-LBP feature extracting method has been obtained good effect in texture analysis and face recognition application experiment, but in the situation that affects of the complex factors such as, image-forming condition extreme variation violent in illumination variation, attitude, expression, age, sign ability and the classification capacity of MB-LBP feature are also restricted, and recognition performance sharply declines.

The present invention is by combining the method for MB-LBP feature and face energygram, adopt the feature extracting method of MB-LBP to carry out quadratic character extraction to face average energygram and variance energygram, when having reduced computation complexity, removed again redundant information, and then by extracted image texture characteristic the Classification and Identification for multi-pose Face.

Summary of the invention

The object of the present invention is to provide a kind of key message that can effectively extract pitching variation and the situation of change human face of vacillating now to the left, now to the right, simultaneously required storage space is little, computation complexity is low, a kind of multi-pose Face recognition methods based on MB-LBP feature and face energygram that discrimination and recognition speed are high.

The object of the present invention is achieved like this:

(1) set up Face Image with Pose Variations training storehouse, and train the face images in storehouse to carry out size normalization Face Image with Pose Variations;

(2), according to the pitching variation range of human face region image luffing angle different demarcation face, build narrow sense face average energygram and narrow sense face variance energygram in conjunction with pitching variation range, as the primary features of multi-pose Face identification;

Related narrow sense face average energygram F _kthe expression formula of (x, y) is:

$F_k(x,y)=\frac{1}{M_k}\underset{j=1}{\overset{M_k}{\mathrm{\Σ}}}{I}_j(x,y),k=\mathrm{1,2,3}$

In formula, M _krepresent the sum of image when same luffing angle scope, the angle of vacillating now to the left, now to the right change, I _j(x, y) is multi-pose Gray Face image, and k represents different luffing angle scopes, and k=1 represents to look up, and k=2 represents to look squarely, and k=3 represents to overlook, and j represents the image that j the angle of vacillating now to the left, now to the right changes, x, and y represents two dimensional image plane coordinate;

Related narrow sense face variance energygram D _kthe expression formula of (x, y) is:

$D_k(x,y)=\frac{1}{M_k}\underset{j}{\overset{M_k}{\mathrm{\Σ}}}{(I_j(x,y)-F_k(x,y))}^2,k=\mathrm{1,2,3}$

(3) adopt MB-LBP algorithm to carry out quadratic character extraction to the narrow sense face average energygram obtaining in step (2) and narrow sense face variance energygram, storage is for the MB-LBP characteristic information of Classification and Identification;

Related MB-LBP feature can be expressed as:

$\mathrm{MB}-\mathrm{LBP}=\underset{i=1}{\overset{8}{\mathrm{\Σ}}}s(B_i-B_c)\·2^i,$ Wherein

G _krepresent the gray-scale value of single pixel; B represents the average gray value of n block of pixels;

(4) read Face Image with Pose Variations to be detected, human face region is detected and extract face;

(5) extracted human face region is carried out to size normalized, obtain standard faces training image;

(6) standard faces training image is carried out to MB-LBP feature extraction, and store the MB-LBP characteristic information extracting;

(7) for the MB-LBP characteristic information of the training storehouse face energygram obtaining in the MB-LBP characteristic information of the standard faces image to be detected obtaining in step (6) and step (3), carry out Classification and Identification by the nearest neighbor classifier based on Euclidean distance, finally export face recognition result.

Beneficial effect of the present invention is:

Face energygram has merged the integrated information of several facial images, not only can well save storage space, reduce computation complexity, and can weaken the noise occurring in single-frame images, face energygram has contained the facial contour information under many attitude, is conducive to realize the recognition of face under the variation of wide-angle attitude.The present invention carries out feature extraction by MB-LBP operator to face energygram and the standard faces image detecting, for Classification and Identification.The MB-LBP feature of extracting can retain the intrinsic macroscopic features of multi-pose Face well, and facial image pattern micromechanism and macrostructure are comprised, and can remove the impact that single pixel noise brings, also make discrimination and recognition speed have remarkable lifting, improved the combination property of multi-pose Face identification.

Brief description of the drawings

Fig. 1 is the multi-pose Face identification process figure based on MB-LBP feature and face energygram;

Fig. 2 is that face is at three-dimensional variation diagram;

Fig. 3 is Face Image with Pose Variations and face average energygram;

Fig. 4 is Face Image with Pose Variations and face variance energygram;

Fig. 5 is the original image in face database;

Fig. 6 is the facial image after normalization;

Fig. 7 is part facial image in test facial image database;

Fig. 8 is the method for expressing figure of MB-LBP feature;

Fig. 9 is the texture image after face average energygram and variance energygram MB-LBP;

Figure 10 is the texture image after standard faces image M B-LBP;

Figure 11 is in conjunction with the method for MB-LBP feature and face energygram and the recognition effect contrast table of additive method.

Embodiment

For example the present invention is described in more detail below in conjunction with accompanying drawing:

Based on a multi-pose Face recognition methods for MB-LBP feature and face energygram, first the face images in Face Image with Pose Variations database is carried out to size normalization, be further built into face average energygram and variance energygram; , then face average energygram and variance energygram are carried out to MB-LBP feature extraction; Then, read Face Image with Pose Variations to be detected, and based on Adaboost algorithm, facial image is carried out to human face region detection, after extraction human face region, do size normalization, obtain standard faces image; Adopt again the standard faces image that the feature extracting method of MB-LBP obtains test library to carry out feature extraction; Finally, complete recognition of face by nearest neighbor classifier.

1. the size normalization of multi-pose Face training storehouse image

The present invention tests and has used the CAS-PEAL-R1 of institute of computing technology of the Chinese Academy of Sciences to share face database, based on the non-front face image subset in face database, comprising 1040 people's image, in experiment, choose wherein 50 people's image, everyone comprises 21 kinds of different attitudes and changes.First, in the face from everyone every kind pitching situation of change, choose 6 width, totally 18 width, build and obtain training face database, totally 50 × 18=900 width image.Then, all images are carried out to size normalization, in the present invention, the size unification of human face region is normalized to 230 × 270 pixels.

2. build face average energygram and variance energygram

Training storehouse image after normalization is respectively used to build three kinds of narrow sense face average energygram and variance energygrams under pitching situation of change.

2.1 narrow sense face average energygrams

Look squarely angle as zero degree using face, the maximum luffing angle that face can occur is-45 ° and 60 °, and the luffing angle of generalized case human face is distributed between [30 °, 30 °].The present invention by luffing angle variation range at [5 °, 5 °] between facial image be defined as and look squarely scope facial image, the facial image between [5 °, 30 °] is defined as and looks up scope facial image, facial image between [30 ° ,-5 °] is defined as and overlooks scope facial image.According to luffing angle scope difference, there is the concept of narrow sense face average energygram, specific as follows:

Narrow sense face average energygram (Narrow face mean energy image, NFMEI): refer to that the image stack summation that same people vacillates now to the left, now to the right under angle in same luffing angle scope, difference is averaging the average image obtaining again, according to face luffing angle scope, different every 1 people comprise 3 width average energygram pictures, are respectively and look up average energygram, look squarely average energygram and overlook average energygram.

Given multi-pose Gray Face image I _j(x, y), the computing formula of narrow sense face average energygram is as shown in (1):

Wherein, M _kthe sum that represents image when same luffing angle scope, the angle of vacillating now to the left, now to the right change, k represents different pitching

$F_k(x,y)=\frac{1}{M_k}\underset{j=1}{\overset{M_k}{\mathrm{\Σ}}}{I}_j(x,y),k=\mathrm{1,2,3}---\left(1\right)$

Angular range, k=1 represents to look up, and k=2 represents to look squarely, and k=3 represents to overlook, and j represents the image that j the angle of vacillating now to the left, now to the right changes, x, y represents two dimensional image plane coordinate.

2.2 narrow sense face variance energygrams

According to luffing angle scope difference, there is again the concept of narrow sense face variance energygram, specific as follows:

Narrow sense face variance energygram (Narrow face variance energy image, NFVEI): poor quadratic sum is averaging the image obtaining again to refer to image that same people vacillates now to the left, now to the right under angle in same luffing angle scope, difference and corresponding narrow sense face average energygram.

For multi-pose Face gray level image I _j(x, y), the definition of narrow sense face variance energygram is as shown in (2):

$D_k(x,y)=\frac{1}{M_k}\underset{j}{\overset{M_k}{\mathrm{\Σ}}}{(I_j(x,y)-F_k(x,y))}^2,k=\mathrm{1,2,3}---\left(2\right)$

Wherein, M _kthe sum that represents image when same luffing angle scope, the angle of vacillating now to the left, now to the right change, k represents different luffing angle scopes, and k=1 represents to look up, and k=2 represents to look squarely, and k=3 represents to overlook, F _k(x, y) represents the narrow sense average energygram of vacillating now to the left, now to the right and obtaining when angle changes within the scope of a certain luffing angle, and j represents the image that j the angle of vacillating now to the left, now to the right changes, x, and y represents two dimensional image plane coordinate.

In conjunction with Fig. 3 and Fig. 4, be depicted as a certain people and overlook, look squarely, look up seven kinds of image and corresponding narrow sense face average energygram and variance energygrams thereof of vacillating now to the left, now to the right while variation occur in three kinds of pitching situations.

3. pair face average energygram and variance energygram carry out MB-LBP feature extraction

Face energygram is the primary features that stack obtains to attitude facial image, can be directly used in the identification of assigning to.But owing to still existing data redundancy, recognition effect not good in face energygram, so adopting MB-LBP to do quadratic character to face energygram, the present invention extracts, for recognition of face.

The face identification method of multiple dimensioned local binary patterns, the method is calculated with the MB-LBP that relatively can realize between relatively replacing of average gray value between block of pixels (sub-block) traditional LBP operator pixel value.Each block of pixels is the square block that comprises neighbor.If adopt the length of side L of piece as parameter, 9 × L × L represents MB-LBP scale,

MB-LBP feature can be expressed as:

$\mathrm{MB}-\mathrm{LBP}=\underset{i=1}{\overset{8}{\mathrm{\Σ}}}s(B_i-B_c)\·2^i,$ Wherein

G _krepresent the gray-scale value of single pixel.B represents the average gray value of n block of pixels.

Fig. 8 is the method for expressing of MB-LBP feature.Fig. 9 is respectively by the texture image after part face average energygram and face variance energygram MB-LBP feature extraction.

4. reading Face Image with Pose Variations and human face region detects

4.1 human face postures change definition

In conjunction with Fig. 2, face is respectively translation and the rotation along axle in the variation of 3 dimension spaces, wherein along the left and right translation of axle, along the upper and lower translation of axle, along the front and back translation of axle and rotate the facial image that a certain angle causes and tilt can effectively be overcome by the method for geometrical normalization centered by axle.But rotating by central shaft the several normalization of variation of vacillating now to the left, now to the right that the pitching up and down causing changes, rotation causes centered by axle for facial image also cannot overcome.The present invention by facial image centered by axle and the variation causing become pitching change, can be divided into according to the difference of the anglec of rotation and look up, look squarely and overlook; Facial image rotates with axle the variation bringing becomes the variation of vacillating now to the left, now to the right.

4.2 human face regions detect

In conjunction with Fig. 5 and Fig. 6, the present invention needs first from multi-pose Face storehouse, to read the Face Image with Pose Variations with pitching variation and the variation of vacillating now to the left, now to the right.Then obtain effective human face region by AdaBoost algorithm.

Adaboost sorter is to be formed by the cascade of multilayer Weak Classifier, and the correct result being obtained by ground floor sorter triggers second layer sorter, and the correct result of being exported by the second layer triggers three-layer classification device, by that analogy.On the contrary, all can cause detection to stop immediately from the result being denied of any one layer of output.By the threshold value of every layer is set, most faces can be passed through, non-face can not passing through, has refused most non-face near the layer of cascade classifier rear end like this.Experiment shows, AdaBoost algorithm can detect human face region effectively.

5. human face region picture size normalization

Obtaining after human face region image, need to carry out size normalization to all images.In the present invention, the size of human face region is normalized to 230 × 270 pixels.Fig. 7 is the human face region image detecting.

6. pair standard faces image carries out MB-LBP feature extraction

The standard faces image obtaining is carried out to MB-LBP feature extraction, and the eigenwert that storage is extracted, for Classification and Identification.

Figure 10 is to the texture image after standard faces image M B-LBP.

7. Classification and Identification

Arest neighbors sorting technique based on Euclidean distance, the distance of calculating standard faces eigenmatrix and face energygram eigenmatrix, realizes classification, last Output rusults.

8. the Classification and Identification process of test face

Test process:

(1) first need to from test library, extract facial image, and based on Adaboost algorithm, image be carried out to human face region detection, obtain test face area image, it is carried out to size normalized, obtain 230 × 270 standard faces image T.

(2) then standard faces image T is carried out to MB-LBP feature extraction.

(3) last, the feature of standard faces image T and face average energygram and variance energygram is carried out to arest neighbors classification and draw classification results.

Claims (1)

1. the multi-pose Face recognition methods based on MB-LBP feature and face energygram, is characterized in that, comprises the steps:

(1) set up Face Image with Pose Variations training storehouse, and train the face images in storehouse to carry out size normalization Face Image with Pose Variations;

(2), according to the pitching variation range of human face region image luffing angle different demarcation face, build narrow sense face average energygram and narrow sense face variance energygram in conjunction with pitching variation range, as the primary features of multi-pose Face identification;

Related narrow sense face average energygram F _kthe expression formula of (x, y) is:

$F_k(x,y)=\frac{1}{M_k}\underset{j=1}{\overset{M_k}{\mathrm{\Σ}}}{I}_j(x,y),k=\mathrm{1,2,3}$

In formula, M _krepresent the sum of image when same luffing angle scope, the angle of vacillating now to the left, now to the right change, I _j(x, y) is multi-pose Gray Face image, and k represents different luffing angle scopes, and k=1 represents to look up, and k=2 represents to look squarely, and k=3 represents to overlook, and j represents the image that j the angle of vacillating now to the left, now to the right changes, x, and y represents two dimensional image plane coordinate;

Related narrow sense face variance energygram D _kthe expression formula of (x, y) is:

$D_k(x,y)=\frac{1}{M_k}\underset{j}{\overset{M_k}{\mathrm{\Σ}}}{(I_j(x,y)-F_k(x,y))}^2,k=\mathrm{1,2,3}$

(3) adopt MB-LBP algorithm to carry out quadratic character extraction to the narrow sense face average energygram obtaining in step (2) and narrow sense face variance energygram, storage is for the MB-LBP characteristic information of Classification and Identification;

Related MB-LBP feature can be expressed as:

$\mathrm{MB}-\mathrm{LBP}=\underset{i=1}{\overset{8}{\mathrm{\Σ}}}s(B_i-B_c)\·2^i,$ Wherein

G _krepresent the gray-scale value of single pixel; B represents the average gray value of n block of pixels;

(4) read Face Image with Pose Variations to be detected, human face region is detected and extract face;

(5) extracted human face region is carried out to size normalized, obtain standard faces training image;

(6) standard faces training image is carried out to MB-LBP feature extraction, and store the MB-LBP characteristic information extracting;

(7) for the MB-LBP characteristic information of the training storehouse face energygram obtaining in the MB-LBP characteristic information of the standard faces image to be detected obtaining in step (6) and step (3), carry out Classification and Identification by the nearest neighbor classifier based on Euclidean distance, finally export face recognition result.