Facial Age Simulation using Age-specific 3D Models and Recursive

PCA

Anastasios Maronidis and Andreas Lanitis

Department of Multimedia and Graphic Arts, Cyprus University of Technology, P.O. Box 50329, 3036 Lemessos, Cyprus

Keywords: Age Simulation, Age Specific Statistical Models, Recursive PCA, Anthropometric Measurements.

Abstract: Facial age simulation is a topic that has been gaining increasing interest in computer vision. In this paper, a

novel age simulation method that utilizes age-specific shape and texture models is proposed. During the

process of generating age-specific shape models, 3D face measurements acquired from real human faces are

used in order to tune a generic 3D face shape model to represent face shapes belonging to certain age

groups. A number of diagnostic studies have been conducted in order to validate the compatibility of the

tuned shape models with the corresponding age groups. The shape age-simulation process utilizes age-

specific shape models that incorporate age-related constraints during a 3D shape reconstruction phase. Age

simulation is completed by predicting the texture at the target age based on a recursive PCA method that

aims to superimpose age-related texture modifications in a way that preserves identity-related characteristics

of the subject in the source image. Preliminary results indicate the potential of the proposed method.

1 INTRODUCTION

Age simulation can prove an important task in a

broad range of applications such as the prediction of

facial appearance of missing persons, automated

update of records, face recognition robust to aging

variation, entertainment applications and

cosmetology. Simulating facial aging is a tough

problem due to the diversity of aging variation for

different subjects and the dependence of the aging

process on external factors related to a person’s

lifestyle.

Many studies related to face age progression

have shown that different mechanisms are

responsible for the aging of shape and texture.

Particularly, in formative years a face is

predominantly subjected to shape transformations,

while in adulthood the main changes are expressed

as texture transformations (Fu et al., 2010).

The contribution of this work is an integrated

system, which given a person’s face that belongs to

a certain age group, it predicts its shape and texture

in a target age group. Despite the fact that in the past

age progression techniques that combine shape and

texture manipulations were reported (Du et al.,

2012), (Lanitis et al., 2002), (Park et al., 2008) in

our method, shape and texture information is treated

separately through the use of age specific 3D shape

and texture models. Using 3D instead of 2D models

enables the application of the proposed method on

faces with different orientations.

During the training phase a set of anthropometric

face measurements (Farkas, 1994) in conjunction

with a generic 3D face Point Distribution Model

(PDM) (Cootes et al., 1995) are used in order to

generate 3D PDM’s specific for each age group. The

method for generating age-specific shape models is

based on our previous work (Lanitis et al., 2012)

where a set of stylistic rules are used for tuning a

generic 3D shape model to a certain design style. In

the case of texture, age-specific models are trained

using texture samples extracted from faces

belonging to different age groups.

During the age simulation phase the age-

transformed shape of the target face is obtained by

3D reconstructing the raw face using a 3D shape

model specific to the target age group. The use of an

age-group specific PDM ensures that during the 3D

reconstruction phase age-specific constraints are

enforced so that the resulting shape possesses

characteristics compatible with the target age group.

The texture of the resulting 3D face is segmented

into two equal parts with respect to the vertical axis

of face symmetry. Keeping the one out of the two

halves intact and considering the second half

occluded, the 3D target face is predicted by utilizing

663

Maronidis A. and Lanitis A..

Facial Age Simulation using Age-speciﬁc 3D Models and Recursive PCA.

DOI: 10.5220/0004290606630668

In Proceedings of the International Conference on Computer Vision Theory and Applications (VISAPP-2013), pages 663-668

ISBN: 978-989-8565-47-1

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

the recursive PCA technique (Wang et al., 2007),

(Park et al., 2003) in conjunction with the target age

group texture model. Finally, the resulting 3D

instance is back projected to 2D obtaining the target

face image. By keeping intact half of a face ensures

that the id information is maintained while the use of

an age group specific texture model ensures that age

information about the target age group is embedded

to the raw face. A block diagram of the proposed

methodology is shown in Fig. 1.

Figure 1: Diagram of the proposed methodology.

The remainder of the paper is organised as

follows: In Section 2, a brief literature review is

presented. In Section 3, the generation of age

specific shape models is described. The proposed

methodology along with some preliminary results is

presented in Section 4. Conclusions are drawn in

Section 5.

2 LITERATURE REVIEW

The process of generating age-specific shape models

is based on facial anthropometric measurements. In

(Zhuang et al., 2005) a database of anthropometric

information collected from a sample of the

American population that could be used for

respirator design is presented. Anthropometric

instruments like spreading callipers and steel

measuring tape have been utilized during the data

acquisition process. The data are categorized

according to the two genders and four ethnic groups.

A similar work is reported in (Goldstein, 1936),

where an attempt to observe the facial growth over a

sample of the American population has also been

made. The main difference between the two above

works is that the former investigates the face

progress over the age groups of 18-30, 30-45, 45-66

years old, while the latter over the age groups of 2.5-

3.5, 4.5-5.5, 6.5-7.5, 8.5-9.5, 10.5-11.5, 12.5-13.5,

14.5-15.5, 16.5-17.5, 18.5-19.5, 20.5-21.5 and 60-

106 years old.

The last few years a considerable number of

researchers described work related to age

progression. Ramanathan et al (Ramanathan et al.,

2009), and Fu et al (Fu et al., 2010) provide

comprehensive survey papers related to facial aging.

Ramanathan et al. (Ramanathan et al., 2006)

propose a computational model that characterizes

facial aging effects observed during formative years

(0–18 years). They highlight the importance of

incorporating age-based anthropometric face

measurements in developing an age progression

model. Subsequently, they propose a facial growth

model that adopts the ‘revised’ cardioidal strain

transformation model that fuses age-based

anthropometric measurements (e.g., landmarks and

proportions) extracted across different facial

features. Our approach is related to (Ramanathan et

al., 2006) since we also utilize anthropometric

measurements for the shape age simulation process.

However, we use a 3D approach, while

(Ramanathan et al., 2006) work in 2D. Another main

difference is that (Ramanathan et al., 2006) deals

only with shape, while we also deal with texture.

Park et al. (Park et al., 2008) propose a 3D facial

aging simulation technique for age invariant face

recognition. Τhe aging patterns of the shape and the

texture are learnt based on PCA coefficients. A

simplified version of Blanz and Vetter (Blanz et al.,

1999) 3D morphable model is used to model the

aging variations from a set of 2D face images. (Park

et al., 2008) use a generic 3D shape model for

reconstructing the faces of a 2D face aging database.

In contrast, we develop and use age-specific 3D

shape models by using anthropometric

measurements.

3 AGE SPECIFIC 3D PDM’S

In this section we describe the process of generating

age-specific 3D shape models.

3.1 Methodology

Based on studies on face anthropometry (Zhuang et

al., 2005), (Goldstein, 1936), we have quantized the

age range into the following 7 age groups: 2-6, 6-10,

10-14, 14-18, 18-30, 30-45 and 45-66 years old. The

age sampling adopted is dense in formative years

and sparse in adulthood entailing an efficient

representation of shape changes during age

progression (Fu et al., 2010). A set of six 3D (see

Fig. 2) measurements along with the corresponding

values that have been measured on real-world

human faces of diverse ages have been acquired

from (Zhuang et al., 2005) and (Goldstein, 1936).

Based on the measurements, each of the seven age

groups is represented by a 6-dimensional feature

VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications

664

vector, where each element contains the mean value

of the corresponding measurement among the

overall population of that age group. Given an

arbitrary synthetic 3D face shape instance it is

possible to use the vertex coordinates for measuring

the values of the six measurements in question and

subsequently estimate the distance of the face

instance from each of the seven age groups.

Figure 2: The 6 anthropometric head measurements. From

left to right: Bigonial Width, Bizygomatic Breadth, Head

Breadth, Menton-Sellion, Minimum Frontal and Nose

Width.

With the help of a generic 3D PDM trained with

laser scanned 3D faces (Blanz et al., 1999), using an

exhaustive grid search over the PDM parameters, we

generate synthetic face instances. For each of these

instances, we estimate the values of the six

anthropometric measurements so that a 6-

dimensional feature vector representing the face

instance is created. The Euclidean distance of this

feature vector from each one of the seven age group

mean vectors is then calculated and used as the

discrepancy measure. The face instance is finally

assigned to the nearest age group since it mostly

complies with this group.

Having obtained a set of 60 compliant samples

for each age group, we train a corresponding 3D

PDM. These PDM’s convey statistical information

relevant to the faces of that age, with respect to their

shape. Therefore, reconstructing a 2D face using an

age specific PDM has the advantage that the

resulting face should contain shape information of

both the id and the age group.

3.2 Diagnostic Tests

A series of diagnostic tests has been conducted in

order to verify the compatibility of the generated 3D

PDM’s to the corresponding age groups. Tests have

been carried out using synthetic and real samples.

3.2.1 Age Group Discrepancies

Experiments using Synthetic Faces: Using the 3D

PDM that corresponds to the 2-6 yrs age group

trained above, we generate 1000 random 3D face

shape instances. For each synthetic face instance we

estimate the values of the six face measurements

(see Fig. 2) and calculate the discrepancy of the

synthetic sample from each age group. As stated in

section 3.1 the discrepancy is defined as the

Euclidean distance of the six measurements and the

mean values for each age group. The distributions of

the discrepancies of instances generated using the

3D PDM that corresponds to the 2-6 yrs age group

from each of the seven age groups are plotted in Fig.

3. Discrepancy is depicted in the x-axis. It is clear

that face instances that have been generated by the

2-6 yrs PDM have the lowest discrepancy from the

2-6 yrs age group.

Figure 3: Distribution of discrepancies of instances

generated by 2-6 yrs 3D PDM from the seven age groups

(seven curves).

Similar but less pronounced results were

obtained when the same test was repeated using the

3D PDM’s for different age groups. This finding

corroborates the claim that shape changes in human

face mainly take place during formative years. The

overall results show in a clear way that instances

generated by an age specific PDM mostly comply

with the corresponding age group.

Experiments using Real Faces: Using the FG-NET

(http://fgnet.rsunit.com) aging database, the faces of

a certain age group are 3D reconstructed by using

each of the seven age group specific 3D PDM’s.

Thus, seven types of 3D reconstruction of the faces

from the age group in question are obtained. 3D

reconstruction has been accomplished by utilizing a

method similar to the one proposed in (Blanz et al.,

2004). For each reconstruction type, the mean

discrepancy of the corresponding 3D reconstructed

faces from the age group in question has been

calculated. This process has been applied to all age

groups. The results for the age groups 2-6, 6-10, 10-

14 and 14-18 yrs old are illustrated in Fig. 4. The

corresponding bar diagrams for the remaining age

groups have been omitted because they are almost

identical to that of 14-18.

In almost all the cases, samples that belong to the

same age group as the age group specific PDM

return the lowest total discrepancy. This is another

FacialAgeSimulationusingAge-specific3DModelsandRecursivePCA

665

way to ensure that the PDM’s that correspond to an

age group are indeed compatible with that age

group. The difference of the mean discrepancy

between the age group in question and the remaining

age groups is clearly more pronounced in the first

two diagrams. As the age progresses, the mean

discrepancies become more uniformly distributed.

This result emphasizes the increased intensity of

shape changes during the formative years.

Figure 4: Bar diagrams of mean discrepancies from the

age groups 2-6, 6-10, 10-14 and 14-18 of faces belonging

to these age groups reconstructed using different 3D

PDM’s. The 3D PDM that corresponds to the age group in

question is depicted with red.

3.2.2 Age Classification

Experiments using Synthetic Faces: Using the 6-

measurement representation of 100 random

instances generated by each of the seven age-group

specific PDM’s, we have conducted an age

classification experiment using five-fold cross

validation (CV). For classifying the samples, a

minimum Mahalanobis distance classifier has been

used. The average correct classification train and test

rates through the 5 CV steps are depicted in the first

row of Table 1. The 100% classification rates

obtained indicate that the six measurements offer a

reasonable representation for discriminating face

instances produced by different age-group specific

shape models.

Experiments using Real Samples: We have 3D

reconstructed the faces depicted in FG-NET dataset

employing a method similar to the one proposed in

(Blanz et al., 2004). For this purpose, we have used

both a generic 3D PDM and each of the age group

specific 3D PDM’s. Using again the six

measurements as feature representation we have

performed classification with respect to age. The

classification settings are the same as in the

experiment using synthetic faces. The results using

the generic and the Age Group Specific (AGS)

PDM’s are shown in the second and third rows of

Table 1, respectively.

Table 1: Results of age classification experiments.

Train Rate (%) Test Rate (%)

Synthetic faces 100.0 100.0

FGNET using generic PDM 69.2 52.7

FGNET using AGS PDM 100.0 99.7

Comparing these two rows it can be easily

observed that using age specific PDM’s returns

approximately double accuracy rate than using

generic PDM implying that unlike the generic 3D

PDM, age specific 3D PDM’s retain age-related

information during the 3D reconstruction process.

Despite the quite high accuracy rate obtained by

using age specific PDM’s, it must be stated that this

is just a diagnostic test assessing the specificity of

the generated models to each age group. The

aforementioned techniques cannot be treated as age

estimation results, since during the 3D

reconstruction process of FG-NET images we use

the a priori age information in selecting the

appropriate age group specific 3D PDM for each

face instance. Nevertheless, this test accentuates the

significance of age group specific 3D PDM’s in

representing real faces of different ages.

4 AGE SIMULATION

Given a face image that shows a subject at a certain

age, we attempt to predict the appearance of the

same person in a target age. Using the Target Age

Group Specific (TAGS) 3D PDM trained in the way

described in the previous section, we estimate firstly

the shape of the target face. We also train a TAGS

texture PCA model using samples belonging to the

target age group from the FG-NET aging database.

Using this model, we estimate the texture of the

target face using the method described in 4.2.

4.1 Shape

Given a raw face, we locate 68 landmarks, which

correspond to some predefined salient facial

features, (Lanitis et al., 2002). Using these

landmarks along with the trained TAGS 3D shape

model, we reconstruct the 3D version of the initial

2D face image. The resulted 3D face shape

approximates the corresponding face shape of the

VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications

666

given subject at the target age group, because the

age specific shape model contain statistical

information from that age group. This claim is

substantiated by the experimental results reported in

section 3.2.

4.2 Texture

For predicting the texture of a raw face at the target

age, we use a modified version of the recursive PCA

method (Park et al., 2003), (Wang et al., 2007).

Recursive PCA is a technique used extensively for

restoring partially damaged or occluded face images.

During the training phase, each sample from the FG-

NET aging database is 3D reconstructed using the

3D PDM that corresponds to the age group in which

the sample belongs to. Using the 3D reconstructed

samples, a Principal Component Analysis (PCA)

texture model for each age group is trained.

During age simulation, a raw face is 3D

reconstructed and divided into two equal segments

with respect to the vertical axis of face symmetry.

The one out of the two halves is used as the

reference part of the face, while the second half is

used as the control part. At each iteration, the whole

face is coded into TAGS texture model parameters

and back reconstructed in the face space. This code-

reconstruction scheme produces a new face. The

distance between the two faces is defined as the

summation of the absolute differences of the

intensities of the vertices that belong to the control

part of the two faces. If this value is minimized, the

process is completed and the simulated face is the

reconstructed one. Otherwise, the reference part of

the reconstructed face is replaced by the

corresponding part of the initial raw face and the

process is repeated.

The rationale behind this approach is to exploit

the statistical texture information about the target

age group contained in the texture model, as well as

to maintain information about the id through the use

of the original half. Therefore, it is expected that the

resulting face comprises a prediction of the actual

appearance of the id at the target age.

Indicative examples of both forward and reverse

age texture simulation are provided in Fig. 5. In Fig.

5 the leftmost picture is the raw face, the middle

picture comprises a mixture of the raw and the age-

transformed face and the rightmost picture consists

of the whole predicted face texture in the target age

group. In Fig. 5, by comparing the initial with the

reconstructed part it can be stated that the identity is

maintained in a high degree. Moreover, texture

characteristics pertaining to the target age group are

also embedded in the reconstructed part. Examples

of the application of the combined shape and texture

age simulation methodology in forward and reverse

age progression are provided in Fig. 6.

Figure 5: Facial texture simulation 5 to 45-66 yrs old (top)

and 50 to 2-6 yrs old (bottom).

5 CONCLUSIONS

In this work, an integrated technique for simulating

facial age progression has been proposed. The

major contribution of the proposed method is the

separate use of both shape and texture age specific

statistical models. The problem of incomplete

existing datasets inhibits the training of statistical

shape models that represent the several age groups.

For this purpose, anthropometric measurements have

been utilized for tuning a generic statistical shape

model into several age-specific ones. A

comprehensive diagnostic study of the proposed

technique has also been performed giving promising

results. Preliminary visual results obtained by using

the proposed method have been provided for visual

inspection in order to extract some intuitive findings.

The work is still in process. As part of our plans

for future work we plan to conduct further

diagnostic studies with respect to the use of the

recursive PCA method for age simulation. We also

plan to stage systematic performance evaluation

experiments for assessing the age progression results

using dedicated performance evaluation methods

(Lanitis, 2008). In addition an extended aging

dataset (Ricanek et al., 2006) will be employed so

that more representative age specific texture models

are utilized during the texture age simulation

process.

FacialAgeSimulationusingAge-specific3DModelsandRecursivePCA

667

Figure 6: Forward age progression from 3 to 10 yrs old (top row) and reverse age progression from 12 to 4 yrs old (bottom

row). From left to right the raw image, initial (red) and reconstructed (blue) landmarks overlaid on raw image, half

predicted image warped on raw image, total predicted image warped on raw image, total predicted image overlaid on raw

image (no warping), real face image at the target age, predicted image warped on target image, initial (red) and

reconstructed (blue) target image landmarks overlaid on target image are shown.

ACKNOWLEDGEMENTS

Part of the work presented was supported by the

Cyprus Research Promotion Foundation and the

European Union Structural Funds (project

ΤΠΕ/ΠΛΗΡΟ/0609(ΒΙΕ)/05)

REFERENCES

Blanz, V., Vetter, T., A Morphable Model for the

Synthesis of 3D Faces, In Computer Graphics, Annual

Conference Series (SIGGRAPH), pp. 187-194, 1999.

Blanz, V., Mehl, A., Vetter T. and Seidel H., A Statistical

Method for Robust 3d Surface Reconstruction From

Sparse Data, In Proc. Intl. Symposium on 3D Data

Processing, Visualization and Transmission, 2004.

Cootes, T. F., Taylor, C. J., Cooper, D. H., Graham, J.,

Active Shape Models - Their Training

and Application, Computer Vision and Image

Understanding, Vol. 61, No. 1, pp. 38-59, Jan. 1995.

Du, J.-X., Zhai, C.-M., Ye, Y.-Q., Face aging simulation

based on NMF algorithm with sparseness constraints,

Proc. of the 7th Intl. Conf. on Advanced Intelligent

Computing Theories and Applications: with aspects of

artificial intelligence, Springer Berlin / Heidelberg,

vol. 6839, pp. 516-522, 2012.

Farkas, L. G., Anthropometry of the Head and Face,

Raven Press, New York, 1994.

Fu, Y., Guo, G., Huang, T.S., Age Synthesis and

Estimation via Faces: A Survey, IEEE Trans. Pattern

Anal. Mach. Intell., vol. 32, no. 11, pp. 1955-1976,

Nov. 2010.

Goldstein, M. S., Changes in dimensions and form of the

face and head with age, Am J Phys Anthropol, 22:37-

89, 1936.

Lanitis, A., Taylor, C. J. and Cootes, T. F., Toward

Automatic Simulation Of Aging Effects on Face

Images. IEEE Trans. of Pattern Analysis and Machine

Intelligence, Vol 24, no 4, pp 442-455, 2002.

Lanitis, A., Evaluating the Performance of Face-Aging

Algorithms. Procs. of the IEEE Intl. Conf. on Face

and Gesture Recognition, 2008.

Lanitis, A., Stylianou, G., Voutounos C., Virtual

Restoration of Faces Appearing In Byzantine Icons,

International Journal of Cultural Heritage, Elsevier,

(doi:10.1016/j.culher.2012.01.001), 2012.

Park, J.-S., Oh, Y.-H., Ahn, S.-C., Lee, S.-W., Glasses

Removal from Facial Image Using Recursive PCA

Reconstruction, Intl Conf. Audio and Video-based

Biometric Person Authentication, pp. 369-376, 2003.

Park, U., Tong, Y., Jain, A. K., Face recognition with

temporal invariance: A 3d aging model, IEEE Intl.

Conf. on Face and Gesture Recognition, 2008.

Ramanathan, N., Chellappa, R., Biswas, S., Computational

Methods for Modeling Facial Aging: A Survey, J.

Visual Languages and Computing, vol. 20, no. 3, pp.

131-144, 2009.

Ramanathan, N., Chellappa, R., Modeling Age

Progression in Young Faces, Proc. IEEE Conf.

Computer Vision and Pattern Recognition, pp. 387-

394, 2006.

Ricanek, K., Tesafaye, T., MORPH A Longitudinal Image

Database of Normal Adult Age-Progression. Procs. of

the 7th IEEE Intl. Conf. on Face and Gesture

Recognition, pp. 341-345, 2006.

Wang, Z. M., Tao, J.H., Reconstruction of Partially

Occluded Face by Fast Recursive PCA, Int. Conf. on

Computational Intelligence and Security Workshops,

Harbin, December 15-19, 2007.

Zhuang, Z., Bradtmiller, B., Head and face anthropometric

survey of U.S. respirator users, J. Occup. Environ.

Hyg., 2, 567-576, 2005.

VISAPP2013-InternationalConferenceonComputerVisionTheoryandApplications

668