You are on page 1of 8

A Hierarchical Approach to Facial Aging

Amrutha Sethuram Karl Ricanek


Eric Patterson
Face Aging Group, Computer Science Department, UNCW, USA
{sethurama, ricanekk, pattersone}@uncw.edu
http://www.faceaginggroup.com

Abstract In this work, we define a hierarchical model for id-


iosyncratic facial aging and develop a synthetic facial ag-
Active Appearance Models (AAMs) have been used as ing system based upon hierarchical models. Adult aging
a promising tool in the field of synthetic age progression. is a distinct process from childhood growth and develop-
However, they are yet to be demonstrated on a large hu- ment. While in the formative years, pre-adult process leads
man population with wide variation. This paper presents to shape-based changes in the mid and lower face, adult ag-
a novel AAM-based hierarchical approach to facial aging. ing consists of shape deformation from weight change, mi-
This work is motivated from studies in medical and anthro- nor bone remodeling, and tissue degeneration. Significant
pological literature on classification of human faces based textural changes due to a variety of complex phenomenon
on gender, ethnic and age groups. The proposed hierar- occur due to adult aging [1]. Further, facial aging can be
chical model approach is a ethnicity and gender specific divided into intrinsic aging and extrinsic aging.Intrinsic ag-
aging paradigm. Specifically, the Caucasian (European de- ing, also known as the natural aging process is character-
scent) and African American ethnic groups are considered. ized by skin lines, furrows,dryness and laxity of the skin.
This work will further show that using individual hierarchi- Extrinsic aging caused due to external factors such as sun
cal models generate better age-progressed synthetic images exposure, repetitive facial expressions, gravity, sleeping po-
when compared to a general model approach. The results sitions, and smoking is marked by fine wrinkles and dyspig-
are evaluated by visual perception of the intended age group mentation. There is some evidence in the anthropological
and preservation of identity. Also, a quantitative evaluation literature that patterns of facial skin aging varies in differ-
was performed using FaceVACS, a commercial face recog- ent ethnic groups. The most obvious ethnic skin difference
nition system, as a surrogate measure. Higher match scores is related to skin colour which is dominant by the presence
for synthetic images generated by hierarchical models when of melanin. The different characteristics of facial skin ag-
compared to those generated by a general model suggests ing in Caucasian, African American, Indian, Hispanic and
the efficiency of the proposed hierarchical model approach. East Asian women is discussed in [2]. Findings from [25]
suggests that skin wrinkling and sagging skin appears a
decade sooner on Caucasian skin. Further from [25], indi-
cations are that skin wrinkling and sagging caused by pho-
1. Introduction toaging is a predominant problem of lighter skin type than
darker skin. Age related changes in facial wrinkles and sag-
The only thing that is constant is change. The changes
ging in the skin of Japanese, Chinese and Thai women are
that a human face undergoes with age is inevitable. The
compared in [3]. Comparison of the occurrence of age-
effects of aging lead to significant variations in both shape
associated changes in the facial skin between French and
and texture in a human face. The degree of variations is
Japanese women is discussed in [24]. Recently Tsuka-
influenced by factors such as gender, ethnicity, exposure to
hara et. al [26] compared the photoaging in Caucasian and
sunlight and effects due to smoking, drinking and weight
Japanese females living in Cincinnati and Tokyo and con-
loss or gain. Recently, sufficient emphasis has been laid on
cluded that wrinkle and facial sagging scores were higher
this “real-world” problem of aging which can considerably
in the Caucasian women. However, the more comprehen-
degrade the performance of facial recognition algorithms.
sive data is that of Hillebrand et al. [27] comparing dif-
Robustness of these algorithms to any such temporal vari-
ferent ethnic groups living in Los Angeles. Findings from
ations is highly desirable for applications in law enforce-
[27] concluded that increased skin wrinkling was observed
ment, forensics, medicine and other fields.

978-1-4244-7030-3/10/$26.00 ©2010 IEEE 100


in the order: Caucasian > Hispanic > African American demonstrated thus far for synthesizing changes in the face
> East Asian. Similarly, although the aging processes of is the use of AAM for a feature space that takes into account
the male and female face share many common features, dis- aging variation. General aging models based on AAMs
crete anatomic differences exist between the male and fe- [13], [14] have been used for growth and development (pre-
male face. In the literature found in plastic surgery, some adult aging) in [15], [16] while [17] [18] [19] [23] have
suggest the existence of differences in aging due to gender focused on adult aging. A thorough survey of some of the
differences. Aging changes in the male face is discussed in techniques is presented in [20], although the authors do not
[4]. go in much depth on the results generated via AAM-based
This suggests that in general, the aging pattern of an in- methods.
dividual belonging to a specific ethnic group and gender
can be formulated as an intra-class problem. Developing 3. Hierarchical Models for synthetic age pro-
separate models grouped by ethnicity and gender to fetch
gression
better aging patterns is thus justified. In this work, we
consider a novel hierarchical approach to synthetic facial Development of accurate, individualized, idiosyncratic
aging. Hierarchical models for male and female belong- models require a large set of longitudinal data for the same
ing to Caucasian American and African American ethnic individual across adult lifespan. However, it has still been a
groups are developed based on Active Appearance Models challenge to gather such data. The data used for developing
(AAM). Our goal is to create synthetically age progressed the hierarchical models were chosen mainly from Morph
images that portray actual aging effects documented in med- [21] [29] and the PAL Database [30]. A few images from
ical and anthropological literature, considering the intrin- FG-NET [22] [28] were included in the models for pur-
sic aging factors based on the individual ethnic and gender poses of testing
groups. Also, a general AAM model is developed using all
the images from the hierarchical groups. The age synthe- 3.1. Methodology
sized images from the hierarchical models and the general
model are then visually compared. In addition, FaceVACS For our hierarchical model approach, four models were
SDK,
c a state of the art face recognition SDK from Cog- considered - Caucasian American Male (CM), Caucasian
nitec, is used to generate match scores for the two models American Female (CF), African American Male (AM) and
to demonstrate the effectiveness of the proposed hierarchi- African American Female (AF). For each of these models,
cal model approach when compared to the general model. a total of ninety-five quality greyscale images in the (18-65
The remainder of the paper is organized as follows. In )year age range, that represent the ethnic and gender group
section 2, a brief review of literature concerning aging of the were selected to train each model. Five images were chosen
human face is provided. The hierarchical model approach is in the (18-20) year age group. Ten images were selected
discussed in section 3 with the results presented in section for every succeeding five year age range. Further to model
4. Lastly, we present some concluding remarks with a brief the shape and texture variation using the AAM model more
discussion of ongoing research. efficiently, separate AAM models were built for the younger
(18-45 years) group and the older group(46-65 years) within
each of these hierarchical groups.
2. Related work in face aging
The general methodology to develop the models and
The computational approaches that have been considered generate synthetic aged images is as shown in the flowchart
for simulation of aging can be broadly divided into geomet- in Figure 1. The representative images were first hand
ric or physical based [5] [6]and analysis-synthesis model- annotated using a detailed 161 dense annotation scheme
based methods [5] [7] [8] In general, computational meth- similar to the one in [21]. For our hierarchical model
ods take a two-step approach towards facial aging, involv- approach, individual AAM models were then developed,
ing estimating the current age of the face in an image and based on gender, ethnicity and age-group to generate pa-
then generating parameters to modify the image to appear rameters for the faces within the group. For example, an
age-progressed by a certain amount of years. Tiddeman et. individual African American Female model was built for
al [9] used wavelets to model wrinkles on age-based face the younger age group using only the images in that cat-
composites. Ramanathan et. al [5] characterized age-based egory. Another individual aged-AAM model for African
appearance variations using subspace methods and in [6], American Female in the older age group was developed and
they proposed a two-fold approach towards modeling facial so on and so forth. For the general model, all the images
aging in adults. Park et. al [10] [12] propose an automatic across gender, ethnicity and age-group were pooled together
aging simulation technique by using a 3D morphable model. to develop a single generalized AAM model as shown in the
Recently, a compositional and a dynamic model for face ag- flowchart. For each of these models, support vector regres-
ing has been presented in [11]. One of the best approaches sion (SVR) was then used to learn age-related properties of

101
Figure 1. Methodology to generate synthetic aged images using Hierarchical/General AAM model approach

Figure 2. Our model (last row) compared with the work of Patterson et. al [19] (top three rows). The top two rows are at original age 30
then (40,50,60,70), the third row is a synthetic version aged from 20 to 80. The last row shows aging the same individual at 20, 30, 40, 50,
60 and 65 years using our latest hierarchical model

each of the AAM parameters within the group training set. progress or regress the parameters in the feature space for
Feasible random faces were generated using Monte-Carlo synthesizing a new image.
simulation, classifying each using age-estimation learned Since individual models were built for the younger and
by SVR for each of the groups. Finally, for each of the older subgroups within the hierarchical models, for any
groups, a table of representative age parameters, where each young image that had to be age-progressed to the older
year is a bin represented by the average at that year cre- age group, the synthetic image of the source image age-
ated by the simulation process was generated. This table progressed to the 45th year using the younger model was in-
was used to index and difference AAM parameters to age- cluded in the older model before the older model was built.

102
Figure 3. Comparison of the first principal mode of the Hierarchical and General Models. Top left shows variation in the first mode for
each of the individual hierarchical models in the younger age group. Top right shows variation in the first mode for each of the individual
hierarchical models in the older age group. The bottom row shows variation in the first mode for the general model

In addition, to aid in the visual comparison of the synthetic 6 were generated using the Caucasian American Female
images generated by the hierarchical and the general model, Model. The synthetic images shown in Figures 5 and 7
few images from the FG-Net database and a few celebrity were generated using the Caucasian American Male Model.
images viz. Sean Connery, Bill Clinton, Oprah Winfrey etc Figure 8 shows the age progressed images of Bill Clinton
were included in the models. Some of the results are as and Daniel Craig that were generated by the same model.
shown in the next section. The African American Female model was used to generate
Figure 2 compares the synthetic images generated with synthetic age progressed images of Michelle Obama Figure
our hierarchical model with that of work done by Patterson 10 and Oprah Winfrey Figure 11. Figure 12 which shows
et. al [19]. It is evident from the figure that our hierarchi- the age progressed images of Sean Connery and figure 9
cal model can generate better synthetic images for the older were both aged using Caucasian American Male model. Fi-
decades, in terms of texture and age-representation. nally, age progressed images of Sidney Poitier were gener-
ated using the African American Male model as shown in
3.2. Results Figure 13.
It should be noted that only the youngest images were
Synthetic images generated using both the hierarchical
included in the models to generate age-progressed synthetic
models and the general model are presented in this section.
images. The older images are included in the figures only
Figure 3 shows the variation in the first principal mode for
to compare the images generated by the models.
both the hierarchical and the general models. It can be in-
ferred from the figure that individual AAM models devel- It is evident from Figures 4 - 13 that the hierarchical
oped based on gender, ethnicity and age-group can capture models can better model the shape and textural changes of
the intra-class variation better when compared to the gen- a specific group when compared to the general model. For
eral model. Such a specialized individual AAM model can the synthetic aged images generated from the hierarchical
generate better synthetic images that are representative of models, identity and intended age group is better preserved
gender, ethnicity and age-group as is evident from Figures than the general model. Further, in an attempt to quantify
4 - 13. The synthetic images shown in Figures 4 and the performance of the hierarchical model and the general

103
Figure 6. Row 1: Original images at 18, 30, 40, 48 and 61 years R
ow 2: Synthetic images at the same ages generated by CF Hierar-
chical Model Row 3 : Synthetic Images at the same ages generated
by General Model

Figure 4. Row 1: Original images at 20, 35, 49 and 60 years Row


2: Synthetic images at the same ages generated by CF hierarchical
model Row 3 : Synthetic Images at the same ages generated by
general model

Figure 7. Row 1: Original images at 18, 30, 46 and 54 years Row


2: Synthetic images at the same ages generated by CM Hierarchi-
cal Model Row 3 : Synthetic Images at the same ages generated
by General Model
Figure 5. Row 1: Original images at 21, 37, 53 and 63 years Row
2: Synthetic images at the same ages generated by CM hierarchical
model Row 3 : Synthetic Images at the same ages generated by
younger group and the general model were used to construct
general Model
the probe. The probe images were then matched against the
parent image in the gallery. A sample of the results are as
shown in Table 1 for various images. It is clear from the
model in generating synthetic images of a specific group, table that synthetic images generated from the hierarchical
match scores were obtained using the FaceVACS SDK ver- models consistently have a better match score than those
sion 7.1.0. A higher score generated by the SDK is an in- generated by the general model. This suggests that indi-
dication of a better match. The original young test images vidual models built specifically for a ethnic, gender and age
that were used in the model for synthetic age progression group - are better capable of generating synthetic images for
were used as gallery images. The corresponding images aging a person in accordance with the intra-class variation
that were reconstructed by the hierarchical models of the of the group.

104
SourceImage Hierarchical Model Synthetic Image Score General Model Synthetic Image Score
003A20.jpg 0.997 0.57
004A21.jpg 0.792 0.473
005A18.jpg 0.539 0.363
048A18.jpg 0.999 0.451
Charlie34.jpg 0.999 0.363
MichelleObama28.jpg 0.999 0.998
OprahWinfrey31.jpg 0.999 0.994

Table 1. Match Scores for the synthetic images against the original images using FaceVACS.

Figure 8. Row 1: Original images of Bill Clinton @ 46 years and Figure 10. Row 1: Original images of Michelle Obama at 28 and
Daniel Craig @39 years Row 2: Synthetic images of Clinton at 45 years Row 2: Synthetic images at 28, 35, 45, 55 and 65 years
46, 50, 55, 60 and 65 years generated by CM Hierarchical Model generated by AF Hierarchical Model Row 3 : Synthetic images at
Row 3 : Synthetic Images of Craig at at 46, 50, 55, 60 and 65 years 28, 35, 45, 55 and 65 years generated by General Model
generated by Hierarchical Model

4. Conclusion and Future Work


In this paper, we have presented a hierarchical model ap-
proach to facial aging. AAM based synthetic models were
developed for the Caucasian American Male, Caucasian
American Female, African American Male and African
American Female populations. In addition, each of these
groups were further divided into younger and older sub-
groups to capture the effects of aging better within the
models. The synthetic age-progressed images generated by
the hierarchical models were then compared with a general
model which was built using all the training images. Vi-
sual comparison of these images suggest that the hierarchi-
cal models could better model the process of aging for a
Figure 9. Row 1: Original images of an individual at 34 years Row specific ethnic and gender population. Such a specialized
2: Synthetic images at 34, 40,50, 60 and 65 years generated by CM model can then be used to age progress an individual within
Hierarchical Model Row 3 : Synthetic images at 34, 40,50, 60 and the same group in a more efficient manner accounting for
65 years generated by General Model the intra-class variation within the group. FaceVACS SDK
was used to generate match scores to evaluate the perfor-
mance of the two models. It was shown that match scores
for the original and reconstructed images from the models

105
Figure 11. Row 1: Original images of Oprah Winfrey at 31 and 55 Figure 13. Row 1: Original images of Sidney Poitier at 38 and 65
years Row 2: Synthetic images at 31, 45, 55 and 65 years gener- years Row 2: Synthetic images at 20, 38, 50 and 60 years gener-
ated by AF Hierarchical Model Row 3 : Synthetic images at 31, ated by AM Hierarchical Model Row 3 : Synthetic images at 20,
45, 55 and 65 years generated by General Model 38, 50 and 60 years generated by General Model

development. Collection of higher quality longitudinal data


for each of the groups will yield improved results. We are
currently working on sensitivity studies to come up with an
optimum annotation map for the images. We are also focus-
ing on expanding this experiment by creating larger crafted
training data for each of the idiosyncratic hierarchical mod-
els.

References
[1] A. M. Albert, K. Ricanek, and E. Patterson. A review of the
literature on the aging adult skull and face: Implications for
forensic science research and applications. Forensic Science
International, In Press.
[2] A. Perkins, A Kimball, G. Hillebrand. K. Miyamot0. The
different characteristics of facial skin aging in Caucasian,
African American, Indian, Hispanic, and East Asian Women.
In Journal of the American Academy of Dermatology, Volume
Figure 12. Row 1: Original images of Sean Connery at 29 and 62 56, Issue 2, Supplement 2, February 2007, Page AB33
years Row 2: Synthetic images at 29, 40, 50 and 62 years gener-
[3] K. Tsukahara, K. Sugata, O. Osanai,A. Ohuchi,Y. Miyauchi,
ated by CM Hierarchical Model Row 3 : Synthetic images at 29,
M. Takizawa, M. Hotta, T. Kitahara. Comparison of age-
40, 50 and 62 years generated by General Model
related changes in facial wrinkles and sagging in the skin of
Japanese, Chinese and Thai women. In Journal of Dermato-
logical Science, Volume 47, Issue 1, July 2007, Pages 19-28
were consistently higher for the hierarchical models when [4] Paul L. Leong. Aging changes in the male face Facial Plastic
compared to the general model. Our work thus suggests that Surgery Clinics of North America, Volume 16, Issue 3, August
specialized hierarchical models trained on a specific ethnic, 2008, Pages 277-279
gender and age group could generate more accurate syn- [5] N. Ramanathan and R. Chellapa. Face verification across age
thetic images. progression. In IEEE Conference Computer Vision and Pat-
As this is a new area of study, there is much room for tern Recognition, 2005.

106
[6] N. Ramanathan and R. Chellapa. Modeling Shape and Textu- [21] K. Ricanek and T. Tesafaye. Morph: A longitudinal image
ral Variations in Aging Faces. In 8th IEEE Int. Conf on Auto- database of normal adult age-progression. In 7th Int. Conf.
matic Face and Gesture Recognition, 2008. on Automatic Face and Gesture Recognition, pages 341-345,
[7] A. Lanitis and C. J. Taylor. Towards automatic face identifica- April 2006
tion robust to ageing variation. In Proceedings of the Fourth [22] Minear M,Park DC. A lifespan database of adult facial stim-
IEEE Int Conf on Automatic Face and Gesture Recognition, uli.Behavior research methods, instruments and computers : a
pages 391-396, 2000 journal of the Psychonomic Society, Inc Vol: 36 Pages: 630-3
[8] K. Ricanek, E. Boone, and E. Patterson. Craniofacial aging [23] E. Patterson, A. Sethuram, M. Albert, K. Ricanek “Com-
impacts on the eigenface face biometric. In Proc. of the Sixth parison of synthetic face aging to age progression by foren-
IASTED Int. Conf. on Visualization, Imaging, and Image Pro- sic sketch artist” Proceedings of the seventh IASTED Interna-
cessing, pages 249-253, Palma de Mallorca, Spain, Aug 2006 tional Conference on Visualization, Imaging, and Image Pro-
[9] B. Tiddeman, D.M Burt, and D.Perret. Prototyping and trans- cessing, 2007.
forming facial texture for perception research. Computer [24] E Tschachler, F. Morizot “Ethnic Differences in Skin Aging”
Graphics and Applications, IEEE, 21(5): 42-50, July-August Pages 23-31 in the book ‘ ‘Skin aging” B. Gilchrest and J.
2001 Krutmann, Springer 2006.
[10] Unsang Park, Yiying Tong, Anil K Jain.Face Recognition [25] A.V. Rawlings. Ethnic skin types: are there differences in
with Temporal Invariance: A 3D Aging Model. In 8th IEEE skin structure and function? International Journal of Cosmetic
Int. Conf. on Automatic Face and Gesture Recognition, 2008. Science, 2006, 28, 79-93
[11] J. Suo, S Zhu, S Shan.A compositional and dynamic model [26] Tsukahara T, Fujimara T, Yoshida Y et al. Comparison of age
for face aging. In IEEE Transactions on Pattern Analysis and related changes in wrinkling and sagging of the skin in Cau-
Machine Intelligence, Vol 32, No 3, March 2010. casian females and in Japanese females. International Journal
[12] Unsang Park, Yiying Tong, Anil K Jain.Age Invariant Face of Cosmetic Science, 2004, 55, 373-388
Recognition. To appear In IEEE Transactions on Pattern Anal- [27] Hillebrand G.G, Levine M.J, and Miyamoto K. The age
ysis and Machine Intelligence, 2010. dependent changes in skin condition in African Americans,
[13] A. Lanitis, C. J. Taylor, and T. F. Cootes. Automatic iden- Asian Indians, caucasians, East Asians and Latinos. IFSCC
tification and coding of human faces using flexible models. Magazine. 2001, 4, 259-266
IEEE Transactions on Pattern Analysis and Machine Intelli- [28] FG-NET Aging Database, http://www.fgnet.rsunit.com
gence, 19(7), July 1997.
[29] Face Aging Group, http://www.faceaginggroup.com
[14] G. J. Edwards, C. J. Taylor, and T. F. Cootes. Interpreting
face images using active appearance models. In Proceedings [30] PAL Face Database, https://pal.utdallas.edu/facedb/
of the Fifth European Conference on Computer Vision, pages
581-595, 1998
[15] A. Lanitis, C. J. Taylor, and T. F. Cootes. Simulating aging on
face images. In Proceedings of the Second Int. Conf. on Audio
and Video-based Biometric Person Authentication, 1999.
[16] A. Lanitis, C. J. Taylor, and T. F. Cootes. Toward automatic
simulation of aging effects on face images. IEEE Transactions
on Pattern Analysis and Machine Intelligence, 24(4), April
2002.
[17] E. Patterson, K. Ricanek, A. M. Albert, and E. Boone. Auto-
matic representation of adult aging in facial images. In Proc.
of the Sixth IASTED Int. Con. on Visualization, Imaging, and
Image Processing, Palma de Mallorca, Spain, Aug 2006.
[18] E. Patterson, A. Sethuram, Albert. M, K. Ricanek, M. King.
Aspects of Age Variation in Facial Morphology Affecting
Biometrics.1st IEEE Int. Conf. on Biometrics: Theory, Ap-
plications and Systems, 2007. pages 1-6, 2007
[19] E. Patterson, A. Sethuram, K. Ricanek, F. Bingham. Im-
provements in Active Appearance Model Based Synthetic
Age Progression for Adult Aging.3rd IEEE Int. Conf. on Bio-
metrics: Theory, Applications and Systems, 2009.
[20] N. Ramanathan, R. Chellappa and S. Biswas. Computational
methods for modeling facial aging: A survey.Journal of Visual
Languages and Computing, 2009.

107

You might also like