Real-time Human Age Estimation based on Facial Images using Uniform

Local Binary Patterns

Mohamed Selim

, Shekhar Raheja

and Didier Stricker

German Research Center for Artiﬁcial Intelligence, DFKI, Kaiserslautern, Germany

Technical University Kaiserslautern, Kaiserslautern, Germany

Keywords:

Age-group Estimation, Local Binary Patterns, Extended Local Binary Patterns, Real-time, K-Nearest

Neighbours.

Abstract:

This paper summarizes work done on real-time human age-group estimation based on frontal facial images.

Our approach relies on detecting visible ageing effects, such as facial skin texture. This information is de-

scribed using uniform Local Binary Patterns (LBP) and the estimation is done using the K-Nearest Neighbour

classiﬁer. In the current work, the system is trained using the FERET dataset. The training data is divided

into ﬁve main age groups. Facial images captured in real-time using the Microsoft Kinect RGB data are used

to classify the subjects age into one of the ﬁve different age groups. An accuracy of 81% was achieved on

the live testing data. In the proposed approach, only facial regions affected by the ageing process are used

in the face description. Moreover, the use of uniform Local Binary Patterns is evaluated in the context of

facial description and age-group estimation. Results show that the uniform LBP depicts most of the facial

texture information. That led to speeding up the entire process as the feature vector’s length has been reduced

signiﬁcantly, which optimises the process for real-time applications.

1 INTRODUCTION

In recent years, interest toward age-group estima-

tion using facial images has increased. Age informa-

tion can be very useful in various applications such

as Human-Computer Interaction, digital signage dis-

plays in shopping stores, and security surveillance.

This can be done using complex approaches that re-

quire long processing time. In this paper, we present

work on age-group estimation from facial images us-

ing features that are not computationally expensive,

and can be done in real-time.

In the current work, we aim at developing a real-

time age-group estimation system where we track and

capture faces in real-time. Estimation is carried out

using tracked facial images and extracting skin texture

information from them using Local binary patterns

(LBP) operator. This is followed by classiﬁcation of

a person’s age into ﬁve groups: 13-20, 21-30, 31-40,

41-50 and 51-above. We use images from FERET

(Phillips et al., 2000) database for training the system.

Due to insufﬁcient number of samples for persons in

age-group 1-12, this group is not involved in the sys-

tem.

Classiﬁcation is carried out using the k-Nearest

Neighbours algorithm. Finally, we look into reduc-

ing the dimensionality of the feature vectors used for

classiﬁcation to optimise the system for real-time ap-

plications. This is achieved by considering only spe-

ciﬁc regions of face which are affected the most due

to ageing. This helps in reducing the processing time

and speeding up the age-estimation process signiﬁ-

cantly and is discussed more in detail in section 3.3.

The system uses frontal facial images from the

FERET dataset in the training process. Various train-

ing criterion has been performed throughout the work.

First, the data was divided into classes based on

the subject age. Later, more in depth separation

of the data was performed where each age group

was divided into sub-classes based on different crite-

rion. For example, the senior age-group was divided

into two classes based on the gender (males-seniors,

females-seniors). This will be discussed later in fur-

ther details in section 4.

Preliminery results show that the system trained

using FERET dataset, and tested on live data(RGB

data) captured by the Microsoft Kinect camera

reached accuracy of roughly 81% of correct age-

group classiﬁcation. The processing time including

the face and feature extraction was on average 15 mil-

408

Selim M., Raheja S. and Stricker D..

Real-time Human Age Estimation based on Facial Images using Uniform Local Binary Patterns.

DOI: 10.5220/0005311604080415

In Proceedings of the 10th International Conference on Computer Vision Theory and Applications (VISAPP-2015), pages 408-415

ISBN: 978-989-758-090-1

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: General Age-Group Estimation Pipeline.

liseconds, which is suitable for real-time applications.

This paper is organized as follows, in section 2,

we present related work and background information

related to this work. In section 3, we present our

approach in details for human age-group estimation.

Following that we discuss the approach in section 5.

In section 4 the results of various training and testing

are presented. Finally, in section 6 conclusions are

presented with ideas for future work.

Our contribution is estimating the age group of a

subject from its facial image based on simple LBP

features applied on speciﬁc regions on the face that

are affected by the aging process. The time needed to

process a frame is around 15 milliseconds, thus our

approach is also suitable for real time applications.

2 RELATED WORK

The basic approach of almost all work in the area of

age estimation till now is more or less the same as

shown in the age-estimation pipeline in Fig. 1.

As shown in Fig. 1 an image containing face is

fed into the system. Some pre-processing operations

like cropping, smoothing and resizing are carried out

on the input image. The pre-processed image is then

passed for feature extraction. Subsequently, machine

learning techniques are then used to classify these im-

ages into age-groups.

(Lanitis et al., 2004) carried out a survey to com-

pare various age-estimation techniques like, quadratic

functions, shortest distance classiﬁer and neural net-

works. Their results showed that classiﬁers based on

quadratic functions and neural networks obtained the

best results. However, quadratic functions do not al-

ways map the ageing process perfectly since each in-

dividual ages in a different way and hence we decided

to work with neural networks.

Work of (Hewahi et al., 2010) relies on use of

Multilayer Perceptrons (MLPs) and facial landmarks

along with gender information. They focus on classi-

ﬁcation of ages into four age-groups and sub-dividing

each group into two for more ﬁne classiﬁcation. (Ko-

hail, 2012) uses the same approach with a little mod-

iﬁcation. This modiﬁcation basically dealt with com-

bining individual x,y-coordinates into one meaningful

unit using Singular Value Decomposition (SVD), be-

fore feeding them to neural network. This approach

relies on facial landmark detection on the subject face

which might not be always detected in live camera

feed due to image quality or face size.

Another approach which doesn’t rely on speciﬁc

landmark points on the face, but takes into consid-

eration the facial skin texture which varies in accor-

dance with age. If this variation information can be

described, it can serve as a very good feature for age-

estimation. We use Local Binary Patterns (Zhao and

Pietik

ainen, 2007) as a descriptor for the facial texture

information.

Figure 2: Local Binary Patterns.

LBP (see Fig. 2) is an operator that works on pix-

els. It considers the neighbourhood of each pixel and

assigns a label to this pixel after thresholding with the

neighbouring pixel values. It is based on the follow-

ing equation:

LBP

P,R

) =

p−1

∑

p=0

u(x

−x

, u(y) =

(

1 if y ≥ 0

0 if y < 0

(1)

where, x

is pixel under consideration, x

repre-

sents neighbouring P pixels, R is a radius and u(.) is

the step function. Since we consider 8 neighbours,

this results in an 8-bit binary pattern. In total there

can be 2

or 256 different binary patterns. Finally, the

pixel under consideration is set to the decimal equiv-

alent value of its corresponding binary pattern.

Real-timeHumanAgeEstimationbasedonFacialImagesusingUniformLocalBinaryPatterns

409

Figure 3: 256-bin histogram(not exact count).

Figure 4: 59-bin histogram(not exact count).

(Gunay and Nabiyev, 2008) used this approach to

classify ages into 10-year intervals. However, they

also used some weighting mechanism on the image

according to importance of the information in a spe-

ciﬁc region of the face. The weights on different re-

gions of image are decided approximately. (Karthi-

gayani and Sridhar, 2011) also focus on implement-

ing Local binary patterns (LBP) with artiﬁcial neural

networks. (Fard et al., 2013) implemented Local bi-

nary patterns (LBP) on 3 regions of face that are more

affected by ageing and showed considerable improve-

ment. Also, (Ylioinas et al., 2012) applied LBP and

LBP with contrast measure on 9 regions on the face.

However, in our approach, we consider more regions

with ﬁner granularity that are also affected by ageing

effects.

As before, the face is divided into 8 × 8 patches

and histograms are generated for each patch and con-

catenated together. The ﬁnal histogram is the input

feature vector. However, this time we focus on re-

ducing dimensionality of this feature vector and we

implement the concept of uniform and non-uniform

local binary patterns.

A local binary pattern is called uniform if the

binary pattern contains at most two bitwise transi-

tions from 0 to 1 or vice versa when the bit pat-

tern is traversed circularly. For example, the patterns

00000000 (0 transitions), 01110000 (2 transitions)

and 11001111 (2 transitions) are uniform whereas the

patterns 11001001 (4 transitions) and 01010010 (6

transitions) are not. In the computation of the LBP

labels, there is a separate label for each uniform pat-

tern and all the non-uniform patterns are labelled with

a single label. For example, when using (8,R) neigh-

bourhood, there are a total of 256 patterns, 58 of

which are uniform, which yields in 59 different la-

bels.

3 PROPOSED APPROACH

Our approach relies on detecting features that can de-

pict age-variations (e.g. skin texture, wrinkles, age-

spots, etc.). These features are described using Local

Binary Patterns (Zhao and Pietik

ainen, 2007). The

face is divided into 8x8 regions.

3.1 Image Preprocessing and Face

Extraction

Preprocessing the input image is a vital step. Every

image that is fetched from the database undergoes var-

ious preprocessing steps as shown in Fig. 5.

After the image is input to the system, the face is

detected using (Viola and Jones, 2001) haar-cascades.

Following the face detection, the two eyes are de-

tected using also haar-cascades. The face is then ro-

tated so that the two eyes lie on the same horizontal

line.

Cropping is an important step in the pipeline as the

face is the input to the LBP deescriptor. The face de-

tected by the haar-cascade face detector is not always

consistent (same cropping), thus, the face is cropped

using the distance between the eyes. The width and

height of the face is deﬁned as in ﬁgure 6. Those pa-

rameters were chosen to crop the face excluding the

ears and any background part behind the face. The

cropped out image is then smoothed by adding little

Gaussian blur and re-sized ﬁnally to a predetermined

size of 150x150.

3.2 LBP Features

In order to optimize the approach for real-time age-

group estimation, the size of the computed feature

vector was reduced. First, the size of the LBP his-

togram is reduced from 256 bins to 59 bins. More-

over, the local binary pattern operator is applied to

selected facial regions. Details on how this is done

are explained in the following paragraphs.

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

410

Figure 5: Image Preprocessing and Face Extraction Pipeline.

Figure 6: Face cropping using distance between eyes as

metric.

Instead of using normal Local Binary Patterns,

where the histogram consists of 256 bins, uniform Lo-

cal Binary Patterns are used. As mentioned earlier,

the uniform LBP histogram consists of only 59 bins,

where 58 bins are for the 58 uniform patterns, and 1

bin for all non uniform ones. Using the uniform LBP

as descriptor for the facial images must be evaluated

in the domain of facial images and for the application

of age-group estimation. A comparison is carried out

in the results section (see section 4) to evaluate the ac-

curacy of the system using normal LBPs vs. uniform

LBPs.

The facial image is divided into 8 × 8 patches and

LBP operator is applied on each patch one by one. For

every patch, a 59-bin histogram is generated. These

histograms are concatenated and yields one large his-

togram of size (64 × 59) 3776. This resulting his-

togram contains entire texture information for one fa-

cial image. Histograms are generated for all the im-

ages in training and test data set.

As described by (Fard et al., 2013), certain re-

gions of face are much better indicators of age as

Figure 7: Histogram generation from 8x8 patches.

compared to others. Regions like, forehead, cheeks,

eye-corners, etc. reﬂect ageing effects the most. On

the other hand, regions like eyes, eyebrows and lips

hardly give any information about age. Also as the

face is not in a rectangular shape, there is always some

area besides the chin which is also part of detected

face. This is not useful information and this region is

eliminated as well. Considering this fact, we decide

to work with only selected patches as shown in Fig.

Figure 8: Selective Facial Regions on a LBP image.

The regions which are covered by a square are the

ones considered for age-estimation. This approach

eliminated 24 patches, leaving behind only 40 patches

as compared to 64 patches used earlier. So, now the

histograms are of (40x59) 2360 bins. Earlier this size

was (64x59) 3776 bins. This results in histograms or

Real-timeHumanAgeEstimationbasedonFacialImagesusingUniformLocalBinaryPatterns

411

the input feature vectors which are 62.5% of previ-

ous size. This also made sure that the processing time

came down which is a crucial factor for real-time per-

formance.

In the classiﬁcation phase, minimum distance cri-

teria is used for matching and classiﬁcation between

histogram of test image and training images.

3.3 K-Nearest Neighbour Classiﬁer

Using the nearest neighbour classiﬁer might results in

misclassiﬁcation. However, if we can make a decision

by considering more number of neighbours, then clas-

siﬁcation accuracy is expected to go higher. Hence,

we work with k-nearest neighbours. The system fol-

lows a voting strategy and assigns the class with max-

imum votes to the current sample. More details about

choosing the appropriate K value is discussed in sec-

tion 4.

Figure 9: k-Nearest Neighbour classiﬁer with different val-

ues of k.

3.4 Training with FERET Dataset

FERET (Phillips et al., 2000) database consists of fa-

cial images from people of different ages and different

ethnicities. It has two sets of frontal facial images, fa

set and fb. For each subject the fb image was taken

seconds after the fa one.The fa images are of neutral

expression whereas, in the fb the subjects have slight

facial expression.

The FERET dataset is used for training and test-

ing. Prior to age-estimation, some images are dis-

carded in preprocessing stage for which faces are not

detected properly. Remaining images in the database

are used for training and testing purpose. For uni-

form distribution of images into training and test sets,

we carry out 10-fold cross-validation. With each it-

eration, 90% of images are taken as training set and

remaining 10% for test set. A point to note here is

source: www.scholarpedia.org/article/

Local Binary Patterns

Figure 10: fa image. Figure 11: fb image.

that, we apply 10-fold cross-validation to all the four

age-groups. Hence, it is ensured that, with every iter-

ation, training set contains 90% samples from all the

groups and test set contains 10% samples from each

age-group. It is also ensured that no single image is si-

multaneously present in both the sets. The age groups

are divided as shown in table 1.

Table 1: Age groups.

Age-group 1 13-20

Age-group 2 21-30

Age-group 3 31-40

Age-group 4 41-50

Age-group 5 51-above

3.5 Live Feed from Microsoft Kinect

We tried our new system for age-estimation in real-

time. We used Microsoft Kinect (Microsoft, 2012) to

grab images in real-time. We chose to work on higher

resolution of Kinect, 1280x960 so that ﬁne details of

facial texture can be captured. Some rotation is re-

quired to align the images with the horizontal and for

this we need to ﬁrst calculate the tilt. This tilt or angle

of rotation is determined with the help of eye coordi-

nates. Haar cascades are used to ﬁnd the eye locations

in these images. The grabbed images are then rotated

by the angle found earlier to make sure it is aligned

with the horizontal. After this, the distance between

eyes is used as a metric for cropping the face. Fi-

nally, this extracted face image is then passed on for

Local Binary Patterns (LBP) and corresponding his-

togram computation. Once these are computed, the

histogram is passed to the kNN classiﬁer to predict its

age-group.

4 RESULTS

Various tests were carried out in this work. First,

the results of testing the simple LBP operator -that

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

412

Figure 12: Sample images from FERET dataset with age-

group classiﬁcation and age ground truth.

yeilds a histogram of 256 bins- over the data with-

out selective regions on the face are shown. During

those tests, the data was divided into four groups only

(Kids, Young, Adults, Seniors), as shown in table 2.

Table 2: Age-Groups in the ﬁrst tests.

Age-Group Kids Young Adults Seniors

Age Range 1-10 11-25 26-50 50-above

Count 18 525 726 52

We tested the system on mixed dataset i.e, datasets

containing both male and female subjects and also on

datasets containing only males or only females. Some

sample images of the tests are shown in ﬁgure 12. It

was noticed, that the system was not able to classify

samples from Kids and Seniors age-group properly.

This is mainly because the FERET database contains

few number of images for these two age groups com-

pared to Young and Adults group. Following are re-

sults obtained (see Table 3) for training and testing

using Local Binary Patterns (LBPs).

Table 3: Age-group estimation results for mixed and gender

speciﬁc datasets.

Accuracy Mixed Males Females

Average 61.7453% 59.6787% 71.0657%

Kids - - -

Young 63.6538% 60.3846% 76.1538%

Adults 65.4167% 64.8% 71.4286%

Seniors 24.0% 22.5% -

In table 4 the error-distribution for above results

are shown. Here, Error of 1 means the error by nearest

age-group or error by 1-class gap. Similarly, Error

of 2 means the error by second-nearest age-group or

error by 2-class gap. And, Error of 3 means the error

by farthest age-group or error by 3-class gap.

Table 4: Error Distribution for Approach 2.

Error % Mixed Males Females

Error of 1 96.0396% 96.0843% 97.2028%

Error of 2 3.9604% 3.9156% 2.0979%

Error of 3 0% 0% 0.6993%

One thing that we noticed from the results above

was that, considering only misclassiﬁed samples,

more than 96% are misclassiﬁed by nearest class.

Only around 3% are misclassiﬁed by second nearest

age-group and negligible amount for the farthest age-

group. We strongly believe that, the system is not able

to differentiate properly for boundary values.

We also tried, to test our system on images free

from eye-glasses and facial hair like beard and mous-

tache. Since, Local Binary Patterns operate at the

pixel level, presence of eye-glasses and facial hair can

affect the performance as it changes the texture. Re-

sults obtained for this test are shown in Table 5 and

Table 6.

Table 5: Age-group estimation results for images without

eye-glasses and facial hair.

Accuracy Mixed Males Females

Average 63.108% 59.5125% 72.7419%

Kids 0% 0% -

Young 68.4444% 64.2857% 82.0833%

Adults 65.625% 65.0% 71.0526%

Seniors 10.0% 5.0% -

Table 6: Error Distribution.

Error % Mixed Males Females

Error of 1 96.4674% 95.3917% 96.8%

Error of 2 3.5326% 4.6082% 2.4%

Error of 3 0% 0% 0.8%

The results of the constrained images (no glasses

and no facial hair) are slightly higher than the uncon-

strained ones. However, in order to improve the re-

sults of the age-group classiﬁcation, more divisions

were carried on the Young and Adults age group as

shown in table 1. Also, uniform patterns were used.

Finally, selected face regions were used in the training

and testing of the following results. The results have

the average accuracy of the whole data with different

K-Value.

Using normal LBP Histograms, the results are

shown in table 7

Real-timeHumanAgeEstimationbasedonFacialImagesusingUniformLocalBinaryPatterns

413

Table 7: Accuracy achieved with LBP with different param-

eter values.

Results Unconstrained Data

K=1 K=5 K=10 K=20 K=30

All LBP 63.2% 69.8% 70.0% 69.5% 67.6%

Uniform 62.1% 69.3% 70.6% 69.3% 68.2%

Constrained Data

K=1 K=5 K=10 K=20 K=30

All LBP 63.2% 68.9% 69.3% 68.1% 65.2%

Uniform 61.0% 68.0% 68.5% 67.1% 66.8%

Clearly, the results are higher than those of the 4-

groups data division that was carried out before. For

more in-depth analysis, more test were carried out us-

ing the Extended Local Binary Patterns with different

values for the radius (1 and 2 pixels). The results are

shown in the following tables.

Table 8: Accuracy achieved with ELBP (1,8) Radius 1, 8

neighbouring pixels.

Results Unconstrained Data

K=1 K=5 K=10 K=20 K=30

All LBP 62.9% 69.4% 70.3% 69.0% 67.8%

Uniform 54.7% 66.9% 66.4% 67.5% 65.5%

Constrained Data

K=1 K=5 K=10 K=20 K=30

All LBP 62.4% 68.1% 68.9% 67.3% 65.9%

Uniform 57.3% 66.2% 65.9% 65.9% 64.0%

Table 9: Accuracy achieved with ELBP (2,8) Radius 2, 8

neighbouring pixels.

Results Unconstrained Data

K=1 K=5 K=10 K=20 K=30

All LBP

64.4% 69.3% 71.1% 68.5% 67.6%

Uniform 65.4% 69.3% 70.9% 68.3% 68.5%

Constrained Data

K=1 K=5 K=10 K=20 K=30

All LBP 64.1% 68.6% 68.2% 66.1% 65.0%

Uniform 64.9% 68.7% 70.4% 67.6% 66.2%

Looking at the results, The maximum accuracy

reached was at K of value 10. Consequently, the K

value was ﬁxed to 10. Also, using the 59 uniform

patterns yields nearly the same accuracy as using the

full Local Binary Patterns. Thus, the uniform patterns

contribute the most towards the correct classiﬁcation,

and at the same time, reduces the size of the feature

vector dramatically. Consequently, uniform patterns

were chosen for the real time system.

Real-time Testing. Real-time test was carried out

at the DFKI. We tested with 20 subjects (3 fe-

males and 17 males). The subjects were facing the

Kinect(Microsoft, 2012) camera at a distance between

1 meter to 1.5 meters. The subjects were all between

or nearby to age-group 21 − 30. The system per-

formed quite well on those subjects, with correct age-

group estimation for nearly 81% cases.

5 DISCUSSION

Considering the results in Table 7, Table 8 and Table

9, the following discussion points can be drawn:

• Accuracy is equally good even when we work

with only 59 uniform patterns instead of all the

256 local binary patterns. This means, that its the

58 unique uniform patterns that contribute max-

imum towards classiﬁcation. All the remaining

198 non-uniform patterns are put into the same

bin in histogram and still it does not affect the re-

sults.

• The accuracy of the system improved with in-

creased number of neighbours in the classiﬁer that

are contributing towards classiﬁcation.

• We achieved equally good or better accuracy with

signiﬁcantly reduced dimensionality.

• Smaller histograms means processing time is less

and hence the system is fast.

Based on the results, the proposed approach of us-

ing uniform Local Binary Pattern is suitable for hu-

man age-group estimation in real time. Since the time

taken on average for a single prediction with is 15

milliseconds, which is appropriate in the context of

real-time performance.

6 CONCLUSION

As it is clear from experimental results, using uniform

Local Binary Patterns (LBP) as a descriptor is suitable

for the age-group estimation based on facial images.

That is due to the face that skin texture is directly af-

fected by ageing, and the texture is described using

the LBP. Since, time taken by for single prediction

using Local binary patterns is in the order of few mil-

liseconds, this was found suitable for age-group esti-

mation in real-time.

Also, the reduced dimensionality of input feature

vector by using uniform patterns and selected facial

regions has signiﬁcantly reduced the processing time

which is a major source of concern for real-time per-

formance.

Regarding future work, it can be summarized in

the following points:.

VISAPP2015-InternationalConferenceonComputerVisionTheoryandApplications

414

• Include more datasets that have more samples in

Kids and Seniors age groups.

• Use higher resolution camera that can enable

the functionality of the system at further dis-

tances(>2m).

• Use regression to deal with the values of the age

at the boundaries of the age-groups

• Test with different poses of the face.

• Normalize the lighting of the faces using Self-

Quotient Images (Wang et al., 2004).

ACKNOWLEDGEMENTS

This work has been partially funded by the European

project Eyes of Things under contract number GA

643924.

REFERENCES

Fard, H. M., Khanmohammadi, S., Ghaemi, S., and Samadi,

F. (2013). Human age-group estimation based on anﬁs

using the hog and lbp features.

Gunay, A. and Nabiyev, V. (2008). Automatic age classi-

ﬁcation with lbp. In Computer and Information Sci-

ences, 2008. ISCIS ’08. 23rd International Symposium

on, pages 1–4.

Hewahi, N., Olwan, C., Tubeel, N., EL-Asar, S., and Abu-

Sultan, Z. (2010). Age estimation based on neu-

ral networks using face features. Journal of Emerg-

ing Trends in Computing and Information Science,

1(2):61 – 67.

Karthigayani, P. and Sridhar, S. (2011). A novel approach

for face recognition and age estimation using local bi-

nary pattern, discriminative approach using two lay-

ered back propagation network. In Trendz in Informa-

tion Sciences and Computing (TISC), 2011 3rd Inter-

national Conference on, pages 11–16.

Kohail, S. (2012). Using artiﬁcial neural network for human

age estimation based on facial images. In Innovations

in Information Technology (IIT), 2012 International

Conference on, pages 215–219.

Lanitis, A., Draganova, C., and Christodoulou, C. (2004).

Comparing different classiﬁers for automatic age esti-

mation. Systems, Man, and Cybernetics, Part B: Cy-

bernetics, IEEE Transactions on, 34(1):621–628.

Microsoft (2012). Kinect. http://www.microsoft.com/en-

us/kinectforwindows/.

Phillips, P. J., Moon, H., Rizvi, S. A., and Rauss, P. J.

(2000). The feret evaluation methodology for face-

recognition algorithms. IEEE Trans. Pattern Anal.

Mach. Intell., 22(10):1090–1104.

Viola, P. and Jones, M. (2001). Robust real-time object de-

tection. In International Journal of Computer Vision.

Wang, H., Li, S., and Wang, Y. (2004). Face recognition

under varying lighting conditions using self quotient

image. In Automatic Face and Gesture Recognition,

2004. Proceedings. Sixth IEEE International Confer-

ence on, pages 819–824.

Ylioinas, J., Hadid, A., and Pietikainen, M. (2012). Age

classiﬁcation in unconstrained conditions using lbp

variants. In Pattern Recognition (ICPR), 2012 21st

International Conference on, pages 1257–1260.

Zhao, G. and Pietik

ainen, M. (2007). Dynamic texture

recognition using local binary patterns with an appli-

cation to facial expressions. IEEE Trans. Pattern Anal.

Mach. Intell., 29(6):915–928.

Real-timeHumanAgeEstimationbasedonFacialImagesusingUniformLocalBinaryPatterns

415