A Survey on Methods for Face Recognition across Aging
Shruthi M
Ranganatha K
Dept. of CSE
Canara Engineering College,
Mangalore
Dept. of CSE
Canara Engineering College,
Mangalore
mshruthi22@gmail.com
ranganatha.kulkarni82@g
mail.com
ABSTRACT
Face recognition is one of the most successful
applications of decades of research on image analysis. Face
recognition across aging is an important problem which poses
theoretical and practical challenges to the research community
since facial aging also has high impact on recognition
performance. The research in this field has gained lots of
importance recently due to the challenging problems of
human face aging processes and strong demand of robust face
recognition system across ages. It has many applications, such
as passport photo verification, homeland security, missing
person identification, surveillance, etc. This paper discusses
about many interesting studies that have been performed on
this topic. We also present the performance analysis of
various methods for face recognition across age progression.
Keywords
Face Recognition, Facial Aging, Age Progression
1. INTRODUCTION
Face recognition is one of the most popular
biometric methods used to verify the identity an individual.
Face recognition is defined as the biometric identification by
scanning a person’s face and matching it against a library of
known faces. It is typically used in security systems and can
be compared to other biometrics such as fingerprint or iris
recognition systems.
A. General Face Recognition Structure
The input of a face recognition system is always an
image or video stream. The output is an identification or
verification of the subject or subjects that appear in the image
or video. The figure below shows the generic structure of a
face recognition system. The face recognition procedure has 3
steps: Face Detection, Feature Extraction and Face
recognition [16].
Face detection is defined as the process of
extracting faces from scenes. So, the system positively
identifies a certain image region as a face.
Face
Detection
Feature
Extraction
Face
Recognition
regions, variations, angles or measures, which can be human
relevant (e.g. eye spacing) or not.
In face recognition step system recognizes the face. In an
identification task, the system would report an identity from
the database. This phase involves a comparison method, a
classification algorithm and accuracy measure.
B. Face Recognition across Age Progression
Human faces undergo a lot of change in appearance
as they age. The change produced by aging varies across
different ages as well as subjects and it is affected by different
factors. Specifically, during formative years, the variations in
the shape of a face are more prominent while in the later
stages of adulthood, texture variations such as wrinkles and
pigmentation are more visible.
Several face recognition algorithms and systems
have been proposed in the literature and constant efforts are
being made to improve the performance of these systems in
terms of both accuracy and time. Over the years, researchers
have studied the role of illumination variations, pose
variations, facial expressions etc. in affecting the performance
of face recognition systems. Other than these external factors,
aging is another phenomenon that affects performance of face
recognition systems. So it is also important to consider the
role played by aging when designing an efficient face
recognition system. Automatic matching of faces as people
age is useful for tasks like passport renewal, identifying
missing persons where authorities need to verify if the old and
new photographs belong to the same person [1].
The rest of the paper is organized as follows. In
Section 2, we present a brief survey on various methods
proposed for recognizing faces across age progression.
Section 3 presents performance analysis of the methods
described in section 2. In Section 4 we present some of the
facial aging databases that are used in different studies on this
problem. The final Section draws the conclusion.
2. LITERATURE SURVEY
At present, quite a number of scholars have devoted
to studying how to recognize faces across age progression.
Unlike problems such as pose variations, illumination
variations, facial expressions etc., the facial aging problem
poses numerous challenges to face recognition.
a) Bacteria Foraging Fusion [1]:
Fig 1: General Face Recognition Structure
Feature Extraction step involves obtaining relevant facial
features from the data. These features could be certain face
Yadav et al. proposed an algorithm that improves
the performance of face recognition by applying the bacteria
foraging fusion algorithm. The proposed algorithm mitigates
the effect of facial changes caused due to aging by combining
the Local binary pattern (LBP) features of global and local
facial regions at match score level, by means of the bacteria
foraging fusion algorithm.
b) NMF algorithm with sparseness constraints [2]:
Du et al. proposed a method of facial aging
simulation based on Non-negative matrix factorization (NMF)
with sparseness constraints. In this paper, first, an improved
prototyping method was adopted to perform the task of ageing
a human face, which aims to incorporate sparseness
constrained NMF to extract texture features of facial image
and find out which part of the factorized matrix should be
kept sparse, the coefficients H or the basis vectors W, so that
more satisfied result could be obtained. Next, the proposed
aging method was applied in the face recognition to improve
accuracy through the way of simulating virtual aged facial
images.
c) Local descriptors in application to the aging
problem [3]:
Bereta et al. evaluated local descriptors for face
recognition in the context of age progression. The main
objective of their study was to assess and compare the
accuracy of the most recent local descriptors along with their
application to age-invariant face recognition. They have tested
the following commonly encountered descriptors: local binary
pattern (LBP), improved LBP (ILBP), centre-symmetric LBP
(CSLBP), differential local ternary pattern (DLTP), three
patch LBP (TPLBP), multi-scale block LBP (MBLBP),
simplified form of Weber local descriptor (WLD), local XOR
pattern (LXP), local Gabor phase binary pattern (LGPBP),
local Gabor XOR pattern (LGXP). Moreover, the first seven
of above descriptors were tested in application to Gabor
magnitude images. They discussed these descriptors as
standalone entities not combined with Gabor wavelets, where
the descriptors are applied to Gabor magnitude and Gabor
phase images.
d) Face verification of age separated images
under the influence of internal and external factors
[4]:
Mahalingam and Kambhamettu studied the problem
of face verification with age variations using discriminative
methods. Face image is holistically represented using the
hierarchical local binary pattern feature (LBP) descriptor. The
LBP provides an effective representation across minimal age
variations, illumination, and minimal pose variations, which
makes it a suitable descriptor for description of images across
age. The spatial information is incorporated by combining the
LBP at each level of the Gaussian pyramid constructed for
each face image. They presented an AdaBoost classifier that
identifies the intrapersonal and extra-personal image pairs
across age gaps.
e) Periocular Biometrics [5]:
Juefei-Xu et al. presented a novel framework of
utilizing periocular region for age invariant face recognition.
To obtain age invariant features, they first performed preprocessing schemes, such as pose correction, illumination and
periocular region normalization. And then they applied robust
Walsh-Hadamard transform encoded local binary patterns
(WLBP) on pre-processed periocular region only. Finally,
they used unsupervised discriminant projection (UDP) to
build subspaces on WLBP featured periocular images.
f) Discriminative model [6]:
Li et al. proposed a Discriminative Model for Age
Invariant Face Recognition. They first represented each face
by designing a densely sampled local feature description
scheme, in which scale invariant feature transform (SIFT) and
multi-scale local binary patterns (MLBP) serve as the local
descriptors. Since both SIFT-based local features and MLBPbased local features span a high-dimensional feature space, to
avoid the over fitting problem, they developed an algorithm,
called multi-feature discriminant analysis (MFDA) to process
these two local feature spaces in a unified framework. By
random sampling the training set as well as the feature space,
multiple LDA-based classifiers are constructed and then
combined to generate a robust decision via a fusion rule.
g) 3D facial Aging Model [7] :
Park et al. proposed a 3D facial aging model and
simulation method for age-invariant face recognition and
showed how it can be used to compensate for the age
variations to improve the face recognition performance. The
aging modelling technique adapts view-invariant 3D face
models to the given 2D face aging database. Their approach is
based on the fact that exact craniofacial aging can be
developed only in 3D. The advantage of using a 3D model is
that it reduces the variations due to lighting and pose.
h) Graph matching [8]:
Mahalingam and Kambhamettu presented a graph
based face representation for efficient age invariant face
recognition. The graph contained information on the
appearance and geometry of facial feature points. An aging
model was constructed for each individual to model their age
variations mainly in shape and texture. A modified Local
Feature Analysis that uses Fisher score to extract the feature
points was been used effectively to extract feature points.
Uniform Local binary pattern (LBP) operator was applied to
each feature point to compute a feature descriptor for each
feature point, and was used in the graph representation. A two
stage approach for recognition was proposed in which a
simple deterministic algorithm that exploits the topology of
the graphs was proposed for efficient graph matching between
the probe image and the gallery image.
i) Facial Aging in Adults [9]:
Ramanathan and Chellappa proposed a two-step
approach towards modelling facial aging in adults. The model
comprises of a shape variation model and a texture variation
model. Attributing facial shape variations during adulthood to
the changing elastic properties of the underlying facial
muscles, the shape variation model was formulated by means
of physical models that characterize the functionalities of
different facial muscles. Identifying facial muscles into one of
three types namely, (i) linear muscles (ii) sheet muscles (iii)
sphincter muscles, they proposed transformation models for
each and modelled facial feature drifts as linear combinations
of the drifts observed on the individual facial muscles. The
texture variation model was designed specifically to
characterize facial wrinkles in predesignated facial regions
such as the forehead, nasolabial region etc.
In the absence of age-based anthropometric measurements
extracted on adults across ages, they collected facial growth
data by extracting facial features on the passport database.
Such growth data was collected on five different age groups
21-30 yrs, 31-40 yrs, 41-50 yrs, 51-60 yrs and 61-70 yrs. The
facial growth data collected in this manner was found to be
effective in characterizing facial growth based on age, gender,
ethnicity etc. and further during instances when individuals
gain or lose weight.
j) Aging Pattern Subspace [10]:
Geng et al. proposed the ‘Aging pattern subspace:
AGES’ approach to perform age estimation from face images.
Defining the ‘aging pattern’ as a sequence of personal face
images ordered in time, AGES is designed to characterize the
temporal nature of the underlying data and thereby capture the
appearance variations commonly observed in aging faces.
Since obtaining a ‘complete aging pattern’ for each individual
is difficult (the case when an individual’s face images are
available for all the ages of interest), they developed the
‘aging pattern subspace’ drawing inspiration from methods
that develop an eigenspace using incomplete data (data with
missing features). Upon developing the subspace, the aging
pattern and the age of a previously unseen face is determined
by the projection in the subspace that best reconstructs the
face.
k) Bayesian Framework [11]:
Ramanathan and Chellappa proposed a method to
estimate the age separation between a pair of face images of
an individual. Considering four age separation categories
namely, 1-2 years, 3-4 years, 5-7 years and 8-9 years, they
created intra-personal subspaces for each category using the
difference images obtained from pairs of age-separated face
images from the respective categories. Their approach was
based on the premise that facial appearance variations induced
as a result of facial aging would be more pronounced with
increase in age separation and as a consequence, the
‘difference image’ feature would get progressively more
exaggerated with increase in age separation. They used a
Bayesian framework to classify pairs of face images based on
age separation and further extended the approach to perform
face verification across age progression.
3. PERFORMANCE ANALYSIS
a) Bacteria Foraging Fusion [1]:
Experimental results were presented using the FGNet and IIITDelhi face aging databases. The IIITDelhi
database, which was collected by the authors, consisted of
over 2600 age-separated labeled face images of 102
individuals. To account for real life and natural conditions,
images included changes in the face due to illumination, pose,
and presence of accessories such as eyeglasses.
Table 1. Rank-1 accuracy obtained by using the proposed
algorithm on the IIITDelhi and FG-Net databases.
Database
Recognitio
n Accuracy
Oldest Image as
Probe
IIITDelh
FGi
Net
54.3%
64.5
%
Youngest Image
as Probe
IIITDelh
FGi
Net
33.3%
31.2
%
The results showed that it is more difficult to recognize a
person if the youngest image is used as probe compared to
using the oldest image as probe. In case of the “oldest image"
as probe, a rank-1 accuracy of 54.3% was achieved on the
IIITDelhi database and 64.5% on the FG-Net database.
b) NMF algorithm with sparseness constraints [2]:
Experiment results showed that apart from lighting,
gesture and expression, age span is also one of the factors that
affect face recognition, which causes a greater effect on
teenagers’ facial growth and recognition ratio is apparently
improved after adding additional virtual samples by aging
simulation. Experiment shows that Non-negative matrix
factorization (NMF) with coefficient H sparse is more capable
of feature extraction compared to the PCA method in the
course of texture aging, which significantly improves the
accuracy of face recognition across the age progression.
Table 2. Rank-one identification accuracies for FG-NET
Database
FG-NET
Recognition Accuracy
age groups
age groups
< =18
> 18
39.2%
58.7%
Local descriptors in application to the aging
problem [3]:
c)
They showed that the highest accuracies are
obtained by using local descriptors combined with Gabor
magnitudes, and multi-scale block LBP (MBLBP), in
particular. This descriptor is the most stable when looking at
different age groups and ranks. Recognition accuracies
produced by descriptors combined with Gabor phase are
significantly lower.
They showed that local descriptors and Gabor
wavelets can be potentially a useful tool for facial recognition
in the context of aging, however they need to be
complemented by the methods further increasing recognition
accuracy or reducing the dimensionality of feature
description.
The results showed that the recognition accuracies
may be substantially different when using various descriptors.
The MBLBP comes as the best descriptor among the
descriptors not combined with Gabor images, particularly if
the age differences between the training and test images are
short, i.e., between 0 and 5 years, or relatively high, i.e.,
between 21 and 30 years. Slightly worse results are obtained
when using the LBP and its modifications such as TPLBP,
CSLBP, and ILBP. The TPLBP is especially highly robust
with regard to short and middle age differences.
d) Face verification of age separated images
under the influence of internal and external factors
[4]:
Experiments on the FG-NET and MORPH database
provided an insight on several factors that affect the
performance of a face verification system. Experimental
results on the FG-NET and MORPH aging datasets indicated
that the performance of the proposed framework is robust
with respect to images of both adults and children. A detailed
empirical analysis on the effects of internal (age gap, gender,
and ethnicity) and external (pose, expressions, facial hair, and
glasses) factors in the face verification performance was also
studied. The results indicated that the verification accuracy
reduces as the age gap between the image pair increases.
e) Periocular Biometrics [5]:
The WLBP featured periocular images with
subspace modeling using UDP was able to obtain 100% rank1 identification rate and 98% VR at 0.1% FAR on FG-NET
database.
Table 3. Rank-one identification accuracy for
FG-NET
Database(#subjects,
#images)
in probe and gallery
FG-NET (82,1002)
Recognition
Accuracy
reports the rank 1 recognition score under the three settings.
The experimental results highlight the significance of
transforming shape and texture, when performing face
recognition across ages.
Table 6. Face recognition across ages
100%
f) Discriminative model [6]:
Experimental results show that this approach
outperforms a state-of-the-art commercial face recognition
engine on two public domain face aging data sets: MORPH
and FG-NET. The proposed approach addresses the face
aging problem in a more direct way without relying on a
generative aging model. This avoids the need of a training set
of subjects that differ only in their age with minimal
variations in illumination and pose, which is often a
requirement to build a generative aging model. Multi-feature
discriminant analysis (MFDA) method was adopted to refine
the feature space for enhanced recognition performance.
Experimental Setting
No transformations
Shape transformations
Shape and Texture
transformations
Recognition Accuracy
38%
41%
51%
j) Aging Pattern Subspace [10]:
Table 7. Rank-one identification accuracy
Database(#subjects,
#images)
in probe and gallery
FG-NET (10,10)
Recognition
Accuracy
38.1%
Table 4. Rank-one identification accuracy
Database(#subjects,
#images)
in probe and gallery
FG-NET (82,82)
MORPH Album 2 (10000,
20000)
Recognition
Accuracy
47.50%
83.9%
g) 3D facial Aging Model [7]:
They have evaluated their approach using a state-ofthe-art commercial face recognition engine (FaceVACS), and
showed improvements in face recognition performance on
three different publicly available aging databases (i.g., FGNET, MORPH, and BROWNS).Their method is capable of
handling both growth (developmental) and adult face aging
effects.
Table 5. Rank-one identification accuracy
Database(#subjects,
#images)
in probe and gallery
FG-NET (82,1002)
MORPH Album 1 (612,
612)
MORPH Album 2 (10000,
20000)
Recognition
Accuracy
37.4%
66.4%
79.8%
h) Graph matching [8]:
They evaluated the performance of the proposed
algorithms on FG-NET database. The experimental results
indicated that the combination of aging model and the graph
representation perform well in age invariant face recognition.
i) Facial Aging in Adults [9]:
On a database that comprises of 260 age separated
image pairs of adults, they performed face recognition across
age progression. The image pairs were compiled from both
the Passport database and the FG-NET database. They
adopted a very simple recognition algorithm namely, Principal
Component Analysis, to perform recognition across ages
under the following three settings: (i) No transformation in
shape and texture (ii) Performing shape transformation (iii)
Performing shape and textural transformation. Table below
k) Bayesian Framework [11]:
The method proposed in this paper, is very relevant
to applications such as renewal of passports. During renewal
of passports, the age difference between the image pairs in
known apriori. Given a pair of age-separated face images of
an individual, the age difference classifier establishes the
identity between the image pairs and further classifies them to
their corresponding age difference category.
Table 7. Rank-one identification accuracy
Database(#subjects,
Recognition
#images)
Accuracy
in probe and gallery
Public domain
15.0%
Private database (109,109)
The above study on facial aging provides a good
understanding of the problem and also brings out the
challenges linked with the problem.
4. AGING DATABASES
Over the years, many face databases [14] have been
collected to help study the different problems related to face
recognition. There are three publicly available databases that
comprise of age separated face image samples namely, the
MORPH Database, the FGNET Aging Database and the
FERET Database [12].
A. The MORPH Database
The MORPH Database (Craniofacial Longitudinal
Morphological Face Database) comprises face images of
adults taken during different ages. The database has been
organized into two albums: ‘MORPH Album 1’ and ‘MORPH
Album 2’. ‘MORPH Album 1’ comprises of 1690 digitized
images of 515 individuals under the age range 15 - 68 years.
‘MORPH Album 2’ comprises of 15204 images of nearly
4000 individuals. Apart from the face images, the database
also provides meta-information that is critical for the task of
studying age progression such as age, sex, ethnicity, height
and weight [12].
B. FG-NET Aging Database
The Face and Gesture Recognition Research
Network (FG-NET) aging database contains on average 12
pictures of varying ages between 0 and 69, for each of its 82
subjects. Altogether there are a mixture of 1002 color and
greyscale images, which were taken in totally uncontrolled
environments. Each was manually annotated with 68
landmark points. In addition there is a data file for every
image, containing type, quality, size of the image and
information about the subject such as age, gender, spectacles,
hat, moustache, beard and pose. One particular problem with
this dataset is the fact, that images are not equally distributed
over age and thus only few images of persons older than 40
are available [13].
C. FERET Database
The Face Recognition Technology (FERET)
program [15] database is a large database of facial images,
divided into development and sequestered portions. The
development portion is made available to researchers, and the
sequestered portion is reserved for testing face recognition
algorithms. The FERET Database, a comprehensive database
that addresses multiple problems related to face recognition
such as illumination variations, pose variations, facial
expressions etc., also comprises of a few hundred age
separated face images of subjects (the age separation
amounting to 18 months or more). The FERET dataset,
pertaining to facial aging, can be described as below [12]:

Gallery-set: Comprises of 1196 images

Duplicate I Probe-set: Comprises of 722 images of
subjects whose gallery match was taken 0 - 1031
days beforehand.

Duplicate II Probe-set: Comprises of 234 images of
subjects whose gallery match was taken 540 - 1031
days beforehand.
5. CONCLUSION
In this paper we have analysed some of the
interesting approaches to the solution of face recognition
across aging problem. This analysis is essential in developing
new robust algorithms for face recognition across aging
problem. This paper discusses how facial aging impacts the
recognition performance. This paper also discusses about the
performance analysis of various methods for face recognition
across age progression. There is still a strong demand for the
system which can improve the performance of face
recognition across aging both in terms of accuracy and time.
We need a robust methodology to solve this problem. Face
recognition across aging has many applications such as
forensic science, border control, driver's license and passport
verification, access control, localization of missing people,
etc.
REFERENCES
[1]. D. Yadav, M. Vatsa, R. Singh and M. Tistarelli, “Bacteria
Foraging Fusion For Face Recognition Across Age
Progression,” in IEEE Conference on CVPRW, Portland,
Oregon, USA 2013.
[2]. J. Du, C. Zhai, Y. Ye. Face aging simulation and
recognition based on NMF algorithm with sparseness
constraints. In Neurocomputing 116 (2013) 250–259.
[3]. M. Bereta, P. Karczmarek, W. Pedrycz, M. Reformat.
Local descriptors in application to the aging problem in face
recognition. In PatternRecognition 46 (2013)2634–2646.
[4]. G. Mahalingam and C. Kambhamettu. Face verification
of age separated images under the influence of internal and
external factors. In Image and Vision Computing 30 (2012)
1052–1061.
[5]. F. Juefei-Xu, K. Luu, M. Savvides, T. D. Bui, and C. Y.
Suen. Investigating Age Invariant Face Recognition Based
on Periocular Biometrics. IEEE, 2011.
[6]. Z. Li, U. Park, and A. Jain, A discriminative model for
age invariant face recognition. IEEE TIFS, 6(3):1028 –1037,
2011.
[7]. U. Park, Y. Tong, and A. K. Jain. Age-invariant face
recognition. IEEE TPAMI, 32(5): 947–954, 2010.
[8]. G. Mahalingam and C. Kambhamettu. Age invariant face
recognition using graph matching, In IEEE BTAS, 2010.
[9]. N. Ramanathan and R. Chellappa, “Modeling shape and
textural variations in aging adult faces,” in IEEE
International Conference on Automatic Face and Gesture,
Amsterdam, Netherlands, 2008.
[10]. X. Geng, Z. H. Zhou, and K. Smith-Miles, “Automatic
age estimation based on facial aging patterns,” IEEE Pattern
Analysis and Machine Intelligence, vol. 29 (12), pp. 2234–
2240, December 2007.
[11]. N. Ramanathan and R. Chellappa, “Face verification
across age progression,” in IEEE Conference on Computer
Vision and Pattern Recognition, vol. 2, San Diego, U.S.A,
June 2005, pp. 462–469.
[12]. N. Ramanathan, R. Chellappa and S. Biswas “Age
Progression in Human Faces: A Survey”.
[13]. Fg-net aging database, http://fipa.cs.kit.edu/429.php
[14]. http://www.face-rec.org/databases/
[15]. P. J. Phillips, H. Moon, and P. J. R. S. A. Rizvi, “The
FERET database and evaluation procedure for face
recognition algorithms,” IEEE Transactions on Pattern
Analysis and Machine Intelligence, vol. 22(10), pp. 1090–
1104,2000.
[16]. PF de Carrera, I. Marqu´es, Manuel Gra˜na, “Face
Recognition Algorithms”, 2010.