CN106056076B

CN106056076B - A Method for Determining Illumination Invariants of Complex Illuminated Face Images

Info

Publication number: CN106056076B
Application number: CN201610371321.1A
Authority: CN
Inventors: 程勇; 韩袁琛; 曹雪虹; 焦良葆
Original assignee: Nanjing Institute of Technology
Current assignee: Nanjing Institute of Technology
Priority date: 2016-05-30
Filing date: 2016-05-30
Publication date: 2019-06-14
Anticipated expiration: 2036-05-30
Also published as: CN106056076A

Abstract

The invention discloses a kind of methods of the illumination invariant of determining complex illumination facial image, firstly, visual light imaging model --- the Lambert's model that research is classical, object analysis image-forming principle, the illumination estimation model to establish novel provide principle foundation；Then, when designing illumination estimation model, it is contemplated that under the conditions of complex illumination, the light conditions of a width facial image can be divided into it is unobstructed, block and these three regions of transition, thus be divided into two classes and discuss；Again, due to the correlation of illumination between adjacent pixel, the two class illumination estimation results defined before can be merged to obtain a final result；Finally, the illumination invariant of facial image can be derived from by classical simple Lambert's model.The method of the present invention can effectively eliminate the light differential of original image.And the numberical range of mentioned illumination invariant is between 0 and 1, and it is consistent with the numberical range of face intrinsic.

Description

A kind of method of the illumination invariant of determining complex illumination facial image

Technical field

The present invention relates to a kind of methods of the illumination invariant of determining complex illumination facial image, belong to face recognition technology Field.

Background technique

In recent years, in order to effectively eliminate influence of the complex illumination to recognition of face performance, domestic and foreign scholars have been proposed All multi-methods.Wherein, it is a kind of classical, effective method that illumination invariant is extracted from complex illumination facial image.Past, In order to isolate illumination invariant and imaging source from multiplying property model, assume initially that illumination invariant quickly changes, imaging Source is slowly varying, then implements illumination estimation using low-pass filtering and extracts illumination invariant indirectly.Such method can be divided into directly It connects and extracts illumination invariant with indirect both of which.Direct Model refer to from facial image extract high-frequency characteristic as illumination not Variable, effective high-frequency characteristic specifically include that Gradient Features, textural characteristics and transform domain high-frequency characteristic.Indirect pattern refer to first from Illumination is estimated in facial image, then implements illumination and the separation of face intrinsic, extracts illumination invariant, effective illumination estimation Method specifically includes that gaussian filtering, weighting Anisotropic fractals, logarithm total variation and is converted in smothing filtering.

Although these methods have been achieved for certain progress in complex illumination recognition of face, but still have limitation. On the one hand, it is assumed that the illumination invariant feature of face, which quickly changes, has certain narrow-mindedness.Because in face major part region Illumination invariant feature, such as eyebrow, pupil, mole and skin, be all it is slowly varying, only just there is illumination invariant spy between region Sign quickly variation.On the other hand, the angle of current low-pass filtering, smothing filtering and denoising model from acquisition image low-frequency information Estimate illumination (fuzzy image), contains excessive face intrinsic information, be only able to satisfy the slowly varying characteristic of illumination, ignore Image obtains the characteristic of model, is not associated with directly with image irradiation.

Summary of the invention

The technical problem to be solved by the present invention is to overcome the deficiencies of existing technologies, a kind of determining complex illumination face is provided The method of the illumination invariant of image no longer assumes that the frequency characteristic of face intrinsic on the basis of studying classical Lambert's model, But from the image-forming principle of image, illumination can be more accurately estimated from facial image, extracts more robust light According to invariant.

In order to solve the above technical problems, the present invention provides a kind of side of the illumination invariant of determining complex illumination facial image Method, comprising the following steps:

1) by analysis Lambert's model, complex illumination facial image model is determined；

2) illumination estimation model is designed, the image irradiation of facial image is solved；

3) image irradiation of the facial image solved according to the complex illumination facial image model and step 2) of step 1), meter Calculate human face light invariant.

In aforementioned step 1), complex illumination facial image model are as follows:

F (x, y)=I (x, y) R (x, y) (2)

Wherein, F (x, y) is facial image, and R (x, y) indicates that human face light invariant, I (x, y) indicate the figure of facial image As illumination.

Aforementioned step 2) designs illumination estimation model, and solving the image irradiation of facial image, detailed process is as follows:

Illumination estimation model I and illumination 2-1) are separately designed based on the slowly varying region of illumination and the quick region of variation of illumination Estimate modelⅱ:

Illumination estimation model I is defined as:

Illumination estimation modelⅱ is defined as:

F_a(x, y)=I_m(x, y)-F (x, y) (5)

Wherein, I_m(x, y) is the image irradiation under illumination estimation model I, I_s(x, y) is the figure under illumination estimation modelⅱ As illumination, o_{I, j}It is point (x, y) in Ω₁Consecutive points in neighborhood；Max () and min () are respectively indicated and are sought collective data Maximum value and minimum value；

2-2) calculate I_m(x, y) and I_s(x, y) will merge illumination estimation knot using illumination fusion in facial image F (x, y) Fruit I_ms(x, y) is defined as:

T=mean (F_g(x, y))+k × (max (F_g(x, y))-mean (F_g(x, y))) (7)

F_g(x, y)=F_a(x, y)/I_m(x, y) (8)

Wherein, mean () indicates to seek the average value of collective data；K is adjustable factors；

2-3) design the phase between image irradiation of the adaptive Anisotropic fractals of one kind to establish adjacent pixel Guan Xing, and by final image irradiation I (x, y) is defined as:

Wherein, G (x, y, Ω₂) be standard deviation be ρ, convolution kernel scale is Ω₂Gaussian kernel；P (x, y, Ω₂) it is I_ms(x, y) Corresponding anisotropy template；I_ms(i, j) is I_ms(x, y) is in Ω₂Pixel in neighborhood.

Adjustable factors k above-mentioned is taken as 0.6.

Standard deviation ρ above-mentioned is taken as 1.

Ω above-mentioned₁And Ω₂Neighborhood window is set as 3 × 3.

Human face light invariant above-mentioned indicates are as follows:

R (x, y)=F (x, y)/I (x, y) (11)

Advantageous effects of the invention: the method for the present invention can effectively eliminate the light differential of original image.And And the numberical range of mentioned illumination invariant is between 0 and 1, it is consistent with the numberical range of face intrinsic.

Detailed description of the invention

Fig. 1 is the Yale B in the embodiment of the present invention⁺The illumination invariant of face database.

Specific embodiment

The invention will be further described below in conjunction with the accompanying drawings.Following embodiment is only used for clearly illustrating the present invention Technical solution, and not intended to limit the protection scope of the present invention.

The invention mainly comprises extraction two parts of the foundation of illumination estimation model and illumination invariant.Firstly, research warp Visual light imaging model --- the Lambert's model of allusion quotation, object analysis image-forming principle, the illumination estimation model to establish novel provide Principle foundation；Then, when designing illumination estimation model, it is contemplated that under the conditions of complex illumination, the illumination feelings of a width facial image Condition can be divided into it is unobstructed, block and these three regions of transition, thus be divided into two classes and discuss；Again, due to adjacent The two class illumination estimation results defined before can be merged to obtain a final result by the correlation of illumination between pixel；Most Afterwards, by classical simple Lambert's model, the illumination invariant of facial image can be derived from.Specifically comprise the following steps:

1, Lambert's model is analyzed:

Image refers to that target object surface is reflected into the measurement of the light intensity formed on image acquisition sensor.Lambert's mould Type is widely used in complex illumination recognition of face as classical visible images imaging model.Formula (1) gives bright Primary model describes the image-forming principle of target object.

G (x, y)=ρ (x, y) n (x, y)^Ts (1)

Wherein, ρ (x, y) and n (x, y)^TThe reflectivity and normal vector of target object surface are respectively indicated, s indicates imaging Source, G (x, y) indicate the image of target object.

The reflectivity and normal vector of body surface are unrelated with imaging source, are the internal characteristics (illumination invariant) of object. Therefore, the image-forming principle of target object can be described with simple Lambert's model, i.e. a width facial image F (x, y) can be with It indicates are as follows:

F (x, y)=I (x, y) R (x, y) (2)

Wherein, R (x, y) indicates face intrinsic (illumination invariant), and numberical range belongs to [0,1], and I (x, y) indicates face The imaging source (image irradiation) of image.

From Lambert's model: facial image is the product that face intrinsic is multiplied with imaging source；The numerical value of face intrinsic Range belongs to [0,1]；The intensity of facial image is lower than the intensity of imaging source；The maximum value of facial image is than previous any light Method is closer to imaging source by estimate.

2, illumination estimation model is designed:

The light conditions of one width facial image can be divided into three parts: unobstructed region, occlusion area and transitional region (the unobstructed region between occlusion area).The illumination in these regions is presented below as feature respectively: the unobstructed area light of light It is slow according to brighter and variation；Light occlusion area illumination is more gloomy and variation is slow；Light transitional region illumination is by bright To dark and quick variation.Therefore, for the slowly varying region of illumination with quick region of variation separately design illumination estimation model I and II:

Illumination estimation model I is defined as:

Illumination estimation modelⅱ is defined as:

F_a(x, y)=I_m(x, y)-F (x, y) (5)

Wherein, o_{I, j}It is point (x, j) in Ω₁Consecutive points in neighborhood；Max (), min () and mean () difference table Show the maximum value, minimum value and average value for seeking collective data.

To the I in formula F (x, y)_m(x, y) and I_sAfter (x, y) is calculated, illumination estimation is improved using illumination fusion. In this process, we distinguish the shielding edge and other regions of light by image segmentation, and by facial image F (x, y) Middle fusion illumination estimation result I_ms(x, y) is defined as:

T=mean (F_g(x, y))+k × (max (F_g(x, y))-mean (F_g(x, y))) (7)

F_g(x, y)=F_a(x, y)/I_m(x, y) (8)

Wherein, mean () indicates to seek the average value of collective data；K ∈ [0,1] is an adjustable factors.

Since the illumination of neighborhood pixels there should be very big relationship, the adaptive Anisotropic fractals of one kind are designed to build Correlation between the illumination of vertical adjacent pixel, and by final image irradiation estimated result I (x, y) is defined as:

Adjustable factors k and standard deviation ρ are respectively set to 0.6 and 1, Ω in the present invention₁And Ω₂Neighborhood window is set as 3 × 3.

3. deriving illumination invariant:

After estimating illumination in facial image, the Lambert's model that can be described according to formula (2) derives facial image Illumination invariant.The illumination invariant of facial image F (x, y) may be expressed as:

R (x, y)=F (x, y)/I (x, y) (11)

Proved by experimental verification: the method for the present invention can effectively eliminate the light differential of original image, and described The numberical range of illumination invariant R is consistent with the numberical range of face intrinsic between 0 and 1.

Embodiment:

In order to verify the validity of the method for the present invention, Yale B and extension Yale B are combined into Yale B⁺Face database into Row experiment.The library complexity light illumination mode is still a challenging problem for robust illumination face recognition algorithms.Identification Stage, principal component analysis are used for feature extraction, and the nearest neighbor classifier based on Euclidean distance is classified for identification.Inventive algorithm With current advanced algorithm: MSR, Gradientfaces and Guo have carried out comparative experiments, provide corresponding recognition effect.

Yale B⁺Face database includes 38 people, and 64 kinds of illumination modes amount to 2432 width images.All graphical rules are adjusted Whole is 100*100.According to the difference of light source and center of face axis angle, face database is divided into 5 set altogether.Fig. 1 gives The illumination invariant that one people, 5 width images of each set and the present invention extract, it can be seen that the present invention can effectively eliminate difference Influence of the illumination to face intrinsic.

Firstly, selecting a collection to be combined into training set respectively from 5 set, other four set are used as test set, table 1-5 Give the experimental result of algorithms of different.It can be seen that the discrimination of the mentioned algorithm of the present invention is higher than other algorithms, especially collect When closing 5 as training set, hence it is evident that be better than other algorithms.Then, in order to verify the high efficiency of inventive algorithm, everyone is arbitrarily selected Piece image is selected as training set (total 38 width facial images), other images are as test set (total 2394 width face figures Picture), experiment 60 times is repeated, the average recognition rate and standard deviation of algorithms of different are as shown in table 6, it can be seen that inventive algorithm is put down Equal discrimination is apparently higher than other algorithms, and discrimination standard deviation is minimum.

Table 1: discrimination (%) of the set 1 as training set algorithms of different.

Method	Set 2	Set 3	Set 4	Set 5	Entire set
						MSR	99.78	95.49	94.52	94.04	95.71
Gradientfaces	100.00	98.87	87.28	94.74	95.29
						S&L	100.00	97.56	95.83	93.21	96.26
The method of the present invention	100.00	99.81	98.90	98.06	99.08

Table 2: discrimination (%) of the set 2 as training set algorithms of different.

Method	Set 1	Set 3	Set 4	Set 5	Entire set
						MSR	97.74	94.17	93.64	90.31	93.12
Gradientfaces	99.25	95.30	92.54	93.91	94.69
						S&L	98.12	96.62	96.05	90.58	94.48
The method of the present invention	100.00	98.12	99.56	98.02	98.74

Table 3: discrimination (%) of the set 3 as training set algorithms of different.

Method	Set 1	Set 2	Set 4	Set 5	Entire set
						MSR	99.62	98.25	96.27	97.65	97.74
Gradientfaces	100.00	100.00	98.03	99.03	99.16
						S&L	99.25	98.90	95.18	97.65	97.58
The method of the present invention	99.25	98.90	99.34	99.31	99.21

Table 4: discrimination (%) of the set 4 as training set algorithms of different.

Method	Set 1	Set 2	Set 3	Set 5	Entire set
						MSR	95.87	96.71	94.17	99.31	96.86
Gradientfaces	100.00	99.56	97.37	99.72	99..9
						S&L	99.25	98.68	94.93	99.45	98.03
The method of the present invention	98.50	100.00	99.25	99.31	99.34

Table 5: discrimination (%) of the set 5 as training set algorithms of different.

Method	Set 1	Set 2	Set 3	Set 4	Entire set
						MSR	96.62	91.45	92.67	99.34	94.74
Gradientfaces	96.24	91.67	90.23	99.56	94.04
						S&L	98.50	88.38	89.29	98.90	93.04
The method of the present invention	100.00	99.78	100.00	100.00	99.94

Table 6: everyone average recognition rate (%) of the image as training set algorithms of different is randomly selected.

Method	Set 1	Set 2	Set 3	Set 4	Set 5	Entire set
							MSR	84.19	81.36	74.37	74.21	80.20	78.45
Gradientfaces	87.65	80.16	73.94	82.18	90.19	82.95
							S&L	85.63	79.09	70.25	71.76	79.81	76.73
The method of the present invention	94.44	92.35	91.48	93.01	95.06	93.32

The above is only a preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art For member, without departing from the technical principles of the invention, several improvement and deformations can also be made, these improvement and deformations Also it should be regarded as protection scope of the present invention.

Claims

1. a method for determining the illumination invariant of complex illumination face image, is characterized in that, comprises the following steps:

1) Determine the complex illumination face image model by analyzing the Lambertian model;

2) Design an illumination estimation model to solve the image illumination of the face image; the specific process is as follows:

2-1) Design illumination estimation model I and illumination estimation model II based on the slow illumination change area and the illumination rapid illumination change area respectively:

The illumination estimation model I is defined as:

Illumination estimation model II is defined as:

F _a (x,y)= _Im (x,y)-F(x,y)(5)

Among them, _Im (x, y) is the image illumination under illumination estimation model I, Is ( _x , y) is the image illumination under illumination estimation model II, o _{i, j} is the point (x, y) in Ω ₁ Adjacent points in the neighborhood; max( ) and min( ) represent the maximum and minimum values of the set data, respectively;

2-2) Calculate _Im ( _x , y) and Is (x, y), and use illumination fusion to define the fusion illumination estimation result _Im (x, y) in the face image F (x, y) as:

t=mean(Fg( _x ,y))+k×(max(Fg( _x ,y))-mean(Fg( _x ,y))) (7)

F _g (x, y)=F _a (x, y)/I _m (x, y) (8)

Among them, mean( ) means to obtain the average value of the set data; k is an adjustable factor;

2-3) Design an adaptive anisotropic Gaussian filter to establish the correlation between the image illumination of adjacent pixels, and define the final image illumination I(x, y) as:

Among them, G(x, y, Ω ₂ ) is a Gaussian kernel with standard deviation ρ and convolution kernel scale Ω ₂ ; P(x, y, Ω ₂ ) is the anisotropy corresponding to I _ms (x, y) Template; I _ms (i, j) is the pixel point of I _ms (x, y) in the Ω ₂ neighborhood;

3) Calculate the face illumination invariant according to the complex illumination face image model in step 1) and the image illumination of the face image obtained in step 2).

2. a kind of method that determines the illumination invariant of complex illumination face image according to claim 1, is characterized in that, in described step 1), complex illumination face image model is:

F(x,y)＝I(x,y)·R(x,y) (2)

Among them, F(x, y) is the face image, R(x, y) represents the face illumination invariant, and I(x, y) represents the image illumination of the face image.

3 . The method for determining the illumination invariant of a complex illuminated face image according to claim 1 , wherein the adjustable factor k is set to be 0.6. 4 .

4 . The method for determining the illumination invariant of a complex illuminated face image according to claim 1 , wherein the standard deviation ρ is taken as 1. 5 .

5 . The method for determining the illumination invariant of a complex illuminated face image according to claim 1 , wherein the Ω ₁ and Ω ₂ neighborhood windows are set to 3×3. 6 .

6. the method for determining the illumination invariant of complex illumination face image according to claim 1, is characterized in that, described face illumination invariant is expressed as:

R(x,y)=F(x,y)/I(x,y) (11)