Conference PaperPDF Available

Age Estimation from Face Images: Human vs. Machine Performance


Abstract and Figures

There has been a growing interest in automatic age estimation from facial images due to a variety of potential applications in law enforcement, security control, and human-computer interaction. However, despite advances in automatic age estimation, it remains a challenging problem. This is because the face aging process is determined not only by intrinsic factors, e.g. genetic factors, but also by extrinsic factors, e.g. lifestyle, expression, and environment. As a result, different people with the same age can have quite different appearances due to different rates of facial aging. We propose a hierarchical approach for automatic age estimation, and provide an analysis of how aging influences individual facial components. Experimental results on the FG-NET, MORPH Album2, and PCSO databases show that eyes and nose are more informative than the other facial components in automatic age estimation. We also study the ability of humans to estimate age using data collected via crowdsourcing, and show that the cumulative score (CS) within 5-year mean absolute error (MAE) of our method is better than the age estimates provided by humans.
Content may be subject to copyright.
Age Estimation from Face Images: Human vs. Machine Performance
Hu Han, Charles Otto, and Anil K. Jain
Department of Computer Science and Engineering
Michigan State University, East Lansing, MI, U.S.A.
There has been a growing interest in automatic age esti-
mation from facial images due to a variety of potential ap-
plications in law enforcement, security control, and human-
computer interaction. However, despite advances in au-
tomatic age estimation, it remains a challenging problem.
This is because the face aging process is determined not
only by intrinsic factors, e.g. genetic factors, but also by ex-
trinsic factors, e.g. lifestyle, expression, and environment.
As a result, different people with the same age can have
quite different appearances due to different rates of facial
aging. We propose a hierarchical approach for automatic
age estimation, and provide an analysis of how aging influ-
ences individual facial components. Experimental results
on the FG-NET, MORPH Album2, and PCSO databases
show that eyes and nose are more informative than the
other facial components in automatic age estimation. We
also study the ability of humans to estimate age using data
collected via crowdsourcing, and show that the cumulative
score (CS) within 5-year mean absolute error (MAE) of our
method is better than the age estimates provided by humans.
1. Introduction
Humans can glean a wide variety of information from
a face image, including identity, age, gender, and ethnicity
(See Fig. 1). The identification characteristic of face im-
ages has been well explored in real-world applications [34],
including passports and driver licenses. Face mugshot re-
trieval is also a powerful way for law enforcement agen-
cies to identify potential suspects in criminal investigations.
Despite the broad exploration of person identification from
face images, there is only a limited amount of research [25]
on how to accurately estimate and use the demographic in-
formation contained in face images such as age, gender, and
For many practical applications, relying on humans to
supply demographic information from face images is not
Identity: ABC
Age: 42
Ethnic: Caucasian
Moustache: Yes
Beard: Yes
Mole: Yes
Scar: Yes
Gender: Male
Hair: Short, brown
Figure 1. A wide variety of information that can be extracted from
a face image, such as identity, age, gender, ethnicity, and scars,
marks and tattoos (SMT).
feasible. Hence, there has been a growing interest in au-
tomatic extraction of demographic information from face
images. Here we focus on age estimation, whose objective
is to determine the specific age or age range of a subject
based on a facial image. Some of the potential applications
of automatic age estimation are: (i) Law enforcement: Au-
tomatic age estimation systems can help to determine the
potential suspects more efficiently and accurately by filter-
ing the gallery database using the estimated age of the input
mugshot. (ii) Security control: An automatic age estima-
tion system can be used to prevent minors from purchasing
alcohol or cigarette from vending machines or accessing in-
appropriate web pages. (iii) Human-computer interaction
(HCI): The system can adjust the contents presented to a
user based on her age. For example, a smart shopping chart
can be designed to provide recommendations according to
the age of the customer.
Unlike other sources of variation in facial appearance
(lighting, pose, and expression) which can be controlled
during face image acquisition, face aging is an unavoidable
natural process. Moreover, face aging is affected not only
by internal factors, but external factors as well [1].
1.1. Background
A number of studies in the biological, psychological, and
cognitive sciences areas, have reported on how the human
The 6th IAPR International Conference on Biometrics (ICB), June 4 - 7, 2013, Madrid, Spain
Table 1. A Comparison of Published Methods for Automatic Age Estimation.
Publication Face representation Face aging database Performance measure
(#subjects, #images) Human perception of age and accuracy
Lanitis et al. [26] Holistic 2D shape and
texture Private (60, 500) Studied on a subset
with 32 images MAE1: 4.3 (Case2)
Hayashi et al. [18] Texture and Wrinkle Private (300, 300) Studied Hitting ratio2of age group
classification: 27%
Iga et al. [22] Gabor, color, texture
and intensity Private (101, 101) Studied Hitting ratio of age group
classification: 58.4%
Geng et al. [11] Holistic appearance
FG-NET (82, 1002)
MORPH (NA, 433)
Studied on a subset
of FG-NET with 51
MAE: 6.8 / 8.8
CS3: 65% / 46%
Fu and Huang [9] Holistic appearance,
Manifold Private YGA (1600, 8000) Not studied MAE: 56
CS: F: 55%, M: 50%
Suo et al. [36]
Holistic and local
topology, 2D shape,
color, and gradient
FG-NET (82, 1002)
Private (NA, 8000)
Studied with 500
images from the
two databases
MAE: 6.0 / 4.7
CS: 55% / 66%
Guo et al. [13] Holistic BIF FG-NET (82, 1002)
Private YGA (1600, 8000) Not studied
MAE: 4.8 / F: 3.9, M: 3.5
CS: 47% / F: 75%, M: 80%
Li et al. [27] Local patch, code book Private YGA (1600, 8000) Not studied MAE: 8.6
CS: 30.0%
Guo and Wang [12] Holistic BIF,
Kernel PLS MORPH II (NA, 55000) Not studied MAE: 4.2, CS: NA
Choi et al. [4] Holistic appearance,
Gabor, LBP
FG-NET (82, 1002)
PAL (NA, 430)
Private BERC (NA, 390)
Not studied
MAE: 4.7 / 4.3 / 4.7
CS: 73% / 70% / 65%
Luu et al. [28] Holistic contourlet
appearance model
FG-NET (82, 1002)
PAL (NA, 443) Not studied
MAE: 4.1 / 6.0
CS: 73%/NA
Chang et al. [2] Ordinal hyperplane
FG-NET (82, 1002)
MORPH II (NA, 5492) Not studied
MAE: 4.5 / 6.1
CS: 74.7% / 56.5%
Guo and Wang [14] Holistic BIF, PLS PAL (590, 844)
FACES (171, 1026) Not studied PAL/FACES
MAE: 6.1 / 8.1
Wu et al. [40] Grassmann manifold
of Facial shape
FG-NET (82, 1002)
Passport (109, 233) Not studied
MAE: 5.9 / 8.8
CS: 62% / 40%
Thukral et al. [37] Landmark based
hierarchical approach FG-NET (82, 1002) Not studied MAE: 6.2
Chao et al. [3] Label-sensitive relevant
component analysis FG-NET (82, 1002) Not studied MAE: 4.4
Proposed Component and holistic
FG-NET (82, 1002)
MORPH II (20569, 78207)
PCOS4(1802, 10036)
Studied with FG-NET
and 2200 images from
MAE: 4.6 / 4.2 / 5.1
CS: 74.8% / 72.4% / 64.0%
1MAE (mean absolute error) [26] is the average of the absolute difference between estimated and real ages. 2Hitting ratio [18] is the rank-1 age group classification accuracy. 3CS (cumulative score) [11] reflects the percentage of correct age
estimations with different absolute errors. In the table, we only give the CS within 5-year absolute error. 4Interested researchers may contact the Pinellas County Sheriff’s Office (PCSO) to access this database.
brain perceives, represents, and remembers faces. In par-
ticular, various aspects of human age estimation have been
studied in the field of psychology (Rhodes provides a review
[31]). Psychological studies often have the goal of examin-
ing the effects of a subject’s age, gender, and race on the
accuracy of age estimation that the subject provides [39].
These studies provide some context for the performance of
automatic age estimation methods reported in the literature;
however, the accuracy of age estimation by human subjects
on a large scale has not been reported for most databases
used in automatic age estimation research.
Computational models based on the above research have
also been proposed to provide insight into the problem of
automatic demographic information estimation. Fu et al.
provided a review of many of the existing methods until
2010 [8]; however, several novel approaches have recently
been proposed. In this paper, we provide a brief compari-
son of major approaches for automatic age estimation. Ta-
ble 1 shows that existing approaches either represent faces
with a holistic representation (e.g. [11, 13]) or use local fea-
tures, e.g. hair, wrinkles, and mustache for age estimation.
However, there is only a limited amount of analysis on how
aging influences individual facial components [30, 36]; the
study in [30] was limited to 0-10 year age gaps. Further,
to our knowledge, no large scale studies on the human abil-
ity to estimate age have been conducted on public-domain
face aging databases (e.g. FG-NET). Human performance
can provide an interesting baseline for the age estimation
task, since although we don’t expect humans to predict ages
with perfect accuracy, we do expect human estimates to at
least fall within some broad age range of the true age (e.g.
a human is unlikely to say a child is an adult).
1.2. Motivation
Face aging is determined by both intrinsic factors (e,g.
human genes) and extrinsic factors (e.g. work environment,
lifestyle, and health). The characteristics of facial aging in-
clude a gradual change in appearance which is not visibly
apparent over a short age gap. Based on this observation, a
hierarchical age estimator [4, 37] is proposed for automatic
age estimation. Unlike [4, 37], where the whole age range is
directly partitioned into multiple age groups, in this paper,
we use a binary decision tree based on SVM (SVM-BDT)
[38] to perform age group classification. Within each age
group, a separate SVM age regressor is trained to predict the
final age. In order to mitigate the age estimation errors due
to incorrect age group classifications, we use overlapping
age ranges while training the regression functions. With the
hierarchical age estimator, we perform automatic age esti-
mation using a component based representation (forehead,
eyebrows, eyes, nose, mouth, shape, and holistic face), and
analyze the influence of aging on individual facial compo-
The main contributions of this paper are as follows. (i)
A hierarchical age estimation approach is proposed for au-
tomatic age estimation, and a component based represen-
tation is used to analyze the aging process of each facial
component separately. (ii) Human perception ability in age
estimation is studied using crowdsourcing which allows a
comparison of the ability of machines and humans.
2. Face Aging Databases
Table 1 indicates that most face aging studies have
evaluated their approaches on the public domain FG-NET
database. However, the age distribution of FG-NET is
biased significantly to children. Another public domain
database that has been used for age estimation is the
MORPH database [32]. A few studied have reported results
on the MORPH Album2 data set, which contains 55,000
mugshot images. We performed experiments on the FG-
NET and MORPH Album2 databases, and additionally on
a 10,036 face image subset from a 1.5 million mugshot
database available from the Pinellas County Sheriff’s Of-
fice (PCSO), constructed so that faces in it have a uniform
age distribution in the range [17, 68]. FG-NET is from per-
sonal photo collection, but MORPH Album2 and PCSO are
databases from law enforcement agencies. Details of the
three databases that we have used are provided below.
2.1. FG-NET
FG-NET1consists of 1,002 images of 82 individuals.
The average number of images per individual is 12. Our
experiments on FG-NET used the entire dataset, using a
Leave-One-Person-Out (LOPO) protocol, similar to prior
work on age estimation. Although the age of subjects in
FG-NET ranges from 0-69 years, over 50% of the subjects
in FG-NET are between the ages 0 and 13.
2.2. MORPH Album2
MORPH2is a database of mugshot images, with asso-
ciated metadata giving the age, ethnicity, and gender of
each subject in the database. We performed experiments on
MORPH Album2, with a version of the database containing
78,207 images of 20,569 subjects. For our experiments, we
randomly selected subjects (using all images of each sub-
ject selected) until we had a total of 10,001 images. We
used these 10,001 images as a training set, and evaluated
our method’s performance on the remaining 68,206 images
in the dataset. Subjects in MORPH Album2 range from 15-
77 years old, although the number of images per age drops
off above 50 years old. MORPH Album2 is a much larger
database than FG-NET, and has a different age distribution
since it is comprised primarily of adults, and contains no
images of young children.
2.3. PCSO
This mugshot dataset was acquired from the Pinellas
County Sherrif’s Office (PCSO). Similar to MORPH Al-
bum2, this database contains mugshot images with associ-
ated metadata including a capture date for each image, and a
date of birth for each subject. The complete PCSO data set
contains some 1.5 million images, out of which we sampled
a subset of 10,036 images covering an age range of 17-68,
with exactly 193 images per age. Although both MORPH
Album2 and FG-NET cover wide age ranges, their age dis-
tributions are far from uniform. Experiments on this subset
of the PCSO dataset allow us to examine the effect of a uni-
form age distribution (a more challenging scenario, since in
this case we cannot sacrifice performance on the upper end
of the age range to improve overall performance) on our age
Face preprocessing Facial component
Component feature
age estimation
Age: 14
Figure 2. Overview of the proposed component based representation and hierarchical estimator for automatic age estimation.
estimation method while still maintaining a relatively large
dataset. Results are reported for the entire PCSO subset us-
ing a leave-one-fold-out (totally 5 folds) protocol.
3. Proposed Age Estimation Method
We propose a component-based representation for age
estimation from face images. As illustrated in Fig. 2, the
proposed approach consists of four main steps: preprocess-
ing, facial component localization, feature extraction, and
hierarchical age estimation.
3.1. Face Preprocessing
Images in some of the aging face databases, e.g. FG-
NET, may have been captured using different methods, in-
cluding scanned photographs, digitized film, or digital cam-
eras. As shown in Fig. 6, face images in FG-NET can be
either in gray-scale or color, and some of the color images
even have a color cast. To mitigate the influence of inconsis-
tent colors, we first convert all color face images into gray-
scale3. There are also in-plane and out-of-plane face rota-
tions. In order to improve the accuracy of facial component
localization, a nonreflective similarity transformation is ap-
plied to normalize each face image based on two eyes. In
the FG-NET database, 68 landmarks (including the eye cen-
ters) are provided for each face image. For the PCSO and
MORPH Album2 face databases which do not have labeled
landmarks, we automatically detect the eye centers using
the FaceVACS SDK [5]. The normalized face images are
then cropped to the same size and interpupillary distance
(IPD) (see Figs. 2 and 3).
3.2. Facial Component Localization
Component-based face recognition methods have been
reported in [15, 19, 20, 23]. However, these algorithms were
mainly proposed to resolve the misalignment problem in
face matching, and did not address the age estimation prob-
lem. We propose an efficient method to localize individual
facial components which is based on ASM [6] to automat-
3Additional methods for handling illumination variations, e.g. [16, 17],
will be investigated in our future work.
(a) Original (b) Preprocessed (c) Manual (d) Components
Figure 3. Facial component localization for one image from the
FG-NET database. (a) Facial landmarks detected in the original
face image; (b) Facial landmarks detected in the preprocessed face
image, which are closer to the true landmark locations; (c) Manual
facial landmarks; and (d) Localized facial components based on
the landmarks in (c).
ically detect a set of facial landmarks4, followed by com-
ponent localization based on subsets of landmarks corre-
sponding to the forehead, eyebrows, eyes, nose, and mouth.
We also utilize the vectorized face shape coordinates5and
the whole face image as two additional components in our
experiments. Fig. 3 shows that face preprocessing improves
not only the accuracy of ASM in detecting facial landmarks,
but also the localization accuracy of facial components. Per-
component scaling is applied to compensate for differences
in absolute component sizes.
3.3. Feature Extraction
Following the theory of a feedforward path in the cor-
tex [33], Serre et al. [35] proposed a biologically inspired
model (BIM) for robust object recognition, which has also
been found to be effective in automatic age estimation [13].
Instead of extracting biologically inspired features (BIF)
from a holistic face like in [13], we extract BIF features
from individual facial components.
In its simplest form, the BIM consists of two layers of
computational units, where simple S1units are followed by
complex C1units. The S1units correspond to the classical
4The open source software Stasm [29] with 76 predefined facial land-
marks is utilized as the landmark detector. For the FG-NET database, we
directly use the manual facial landmarks provided with each face image.
5The coordinates of facial keypoints have already been normalized to
the same scale in the face preprocessing step.
simple cells of Hubel and Wiesel found in the primary vi-
sual cortex (V1) [21]. They are usually implemented with
the real components of Gabor filters [10] which have been
shown to provide a good model of cortical simple cell re-
ceptive fields [24]. We determine the parameters, e.g. as-
pect ratio γ, the effective width σ, and the wavelength λ, the
orientation θ, and filter scale sas in [35], but we use 8 ori-
entations and 12 scales for the Gabor filters in our automatic
age estimation task.
The C1units correspond to cortical complex cells which
are robust to shift and scale variations. The C1units can be
calculated by pooling over the preceding S1units with the
same orientation but at two neighboring scales. We apply
“MAX” pooling operator and “STD” normalization (MAX-
STD) to extract C1features from the S1layer.
3.4. Hierarchical Age Estimation
We propose a hierarchical age estimation method as
shown in Fig. 2. Each facial component is first classified
into one of four disjoint age groups using a binary deci-
sion tree based on SVM (SVM-BDT) [38]. Within each age
group, a separate SVM age regressor is trained to predict
the final age. However, in contrast to the age group classi-
fication stage with SVM-BDT, we use all available training
samples with ages up to 5 years outside of the age range
for each group. In this way we mitigate error due to in-
correct age group classification of face images with ages
that are close to the group boundaries, and since we only
use overlapped age ranges in the regression stage this does
not introduce class label ambiguity during the classification
stage. We use the RBF kernel for all SVM classifiers and
regressors in this framework. For each data set, we select
the parameters Cand γof the RBF kernel using a 5-fold
cross-validation on the training set.
3.5. Age Estimate Fusion
We train separate hierarchical age estimators for the dif-
ferent feature types we extract from each face image. In
practice we have found that an average of the ages predicted
by several individual estimators can improve the overall per-
formance, as long as the estimators have comparable perfor-
mance. Further, we have found that since the best perform-
ing feature may depend on the age range, age estimation
performance can be improved by fusing different features
for different age groups.
4. Age Estimation by Humans
As a baseline for our automatic age estimation results,
we gathered age estimates made by human workers using
the Amazon Mechanical Turk service (AMT) crowdsourc-
ing service6. We collected human age estimates for the en-
How many years old
i s the person in the
image? Enter your
answer using digits
0-9 only.
(b) (c)
True age: 0
Human est.:
True age: 27
Human est.:
Figure 4. Age estimation using crowdsourcing experiment using
Amazon Mechanical Turk. (a) Human age estimation process; (b)
Overview of one HIT task. (c) Examples of consistent and incon-
sistent estimates by 10 workers.
tire FG-NET database, and a subset of the PCSO database
consisting of 2,200 face images (with a similar age distri-
bution to the larger subset used in our experiments). Me-
chanical Turk workers are anonymous, and we did not col-
lect any personally identifying or demographic information
from workers in these experiments. In order to keep the in-
dividual tasks assigned to workers (referred to as HITs or
human intelligence tasks) as fine-grain as possible, a single
task consisted of displaying a face image with the prompt
string “How many years old is the person in the image? En-
ter your answer using digits 0-9 only.”, along with a text
input box (See Fig. 4). For each face image, we asked ten
workers to provide the age estimates. We posted a total of
32,020 such HITs on AMT. Out of the 32,020 HITs, 10,020
HITs were for the FG-NET database (1,002 face images ×
10 workers) and 22,000 HITs were for the PCSO database
(2,200 face images ×10 workers). So, for each face image
10 workers provided age estimates. The payment for each
HIT was 3 cents, for a total cost of $1,120.70 (including
service fees).
As is typical of crowdsourcing experiments, some of the
response data was noisy and unusable. For example, some
workers submitted an empty text box, some entered a 3 digit
long age, some entered an age range (e.g. 52-60), and at
least one worker typed out the age in words e.g. “forty”.
When workers entered an age range, we took the middle
of the range as their estimate, and we manually converted
the few word entries to integers. We rejected the remaining
problematic cases (empty text boxes, age estimates larger
than 100), and in these cases obtained replacement age es-
timates from additional workers. Out of the 10,020 initial
FG-NET age estimates we found 16 of them to be unusable,
(a) FG-NET (b) MORPH (c) PCSO
0 5 10 15
Absolute Erro r (years)
0 5 10 15
Absolute Erro r (years)
0 5 10 15
Abs olute Erro r (years)
Accuracy (%)
Figure 5. Per-component age estimation performance on (a) the FG-NET, (b) the MORPH Album2, and (c) the PCOS databases.
and out of the initial 22,000 PCSO age estimates, we found
604 of them to be unusable7.
Given 10 age estimates of each face provided by human
workers, one needs to select a method for coming up with
a single age estimate. Given the noisy nature of the crowd-
sourced data, a simple strategy is to discard the highest and
lowest age estimates for each face image, take the mean
of the remaining 8 estimates and finally compute the MAE
as we would for automatic age estimates. However, taking
the mean of all age estimates for individual face images al-
lows too-high and too-low age estimates to counteract each
other [7, 39]. Hence, the MAE based on the mean age does
not represent the performance of an individual human. We,
therefore, discard the highest and lowest estimates for each
image, then calculate the mean absolute error of all remain-
ing individual age estimates.
5. Experimental Results
5.1. Aging of Individual Facial Components
The per-component age estimation performance of the
proposed method on FG-NET, MORPH Album2, and
PCSO are shown in Figs. 5 (a-c)8. On the FG-NET
database, the holistic BIF features perform best overall, fol-
lowed closely by BIF features extracted from the eye re-
gion, and the shape based features. On the other hand,
performance on the MORPH Album2 and PCSO databases
shows that the holistic BIF features significantly outperform
all others, followed by the eyes, with the remaining feature
types clustered relatively close together. A major considera-
tion here is that the FG-NET database provides manually se-
lected keypoint locations, while on the other two databases
the keypoint locations were detected automatically. Since
the shape features are based directly on the keypoint loca-
tions, their performance necessarily suffers when keypoint
7The human age estimates for the FG-NET database is
available to interested researchers through our lab’s website:
8There is no variation bars for the CS curves on the FG-NET and PCSO
databases, because the CS curves are calculated on the whole dataset, not
the average of individual folds.
localization is not very accurate. Our component localiza-
tion method also relies on the keypoint locations, so the
per component features do not perform as well as on FG-
NET on automatically detected keypoints in the MORPH
Album2 and PCSO databases.
5.2. Overall Performance
On the FG-NET database, we found that the best perfor-
mance was attained by a fusion of the three best performing
features, namely holistic BIF, shape, and eye region BIF.
We attained a MAE of 4.6 by combining these three fea-
tures for the lowest age group (0, 7), just combining holistic
and shape based features for the 2nd age group (8, 17) and
3rd age group (18, 25), and just combining eye and holistic
estimates for the oldest age group (26, 69).
On the MORPH Album2 and PCSO datasets, we found
that the holistic BIF features outperformed the shape and
per-component features to the extent that simple fusion
methods showed no improvement in age estimation over the
holistic BIF based estimate alone. For these two datasets,
we therefore report only the performance of the holistic BIF
based age estimator, which is as follows: an MAE of 4.2
on MORPH Album2, and an MAE of 5.1 on PCSO. While
the MORPH Album2 and PCSO datasets both consist of
mugshots of adults, we attribute this performance difference
to the different age distributions of these two databases. The
uniform age distribution of PCSO means that there are no
cases where performance on one part of the age range can
be sacrificed to improve performance on the overall dataset.
Examples of face images where our algorithm gave good
and poor age estimates are shown in Fig. 6.
Age estimation results from state-of-the-art methods on
FG-NET and MORPH Album2 are included in Table 1.
Better results than ours have been reported on FG-NET in
[2], [3], and [28] with MAE of 4.5, 4.5, and 4.1 respec-
tively. The best known performance on MORPH Album2
is 4.18 years MAE reported by Guo and Mu [12], which is
slightly better than our MAE of 4.2. Compared to [12], we
are using a larger version of MORPH Album2 (78,207 vs
55,000 images), but more significantly we handle ethnicity
True: 1
Est.: 1
True: 22
Est.: 22
True: 8
Est.: 8
True: 35
Est.: 35
True: 18
Est.: 18
True: 35
Est.: 35
True: 20
Est.: 20
True: 47
Est.: 47
(a) Good estimation
(b) Poor estimation
True: 5
Est.: 15
True: 31
Est.: 21
True: 24
Est.: 34
True: 13
Est.: 23
True: 19
Est.: 29
True: 39
Est.: 49
True: 27
Est.: 37
True: 53
Est.: 63
Figure 6. Examples of age estimation using the proposed approach. (a) Good age estimates; (b) Poor age estimates.
and gender variations differently. Guo and Mu constructed
10,000 image training sets with relatively balanced gender
and ethnic groups, and trained a model that simultaneously
predicts age, gender, and ethnicity. We randomly sample a
10,000 image training set which follows the race and eth-
nicity distribution of the complete dataset, and construct a
model that does not use either the race or ethnicity ground
truth during training. Although our method cannot predict
gender or ethnicity, we do achieve comparable age estima-
tion accuracy to [12] without directly modeling the effects
of gender and ethnicity variation in the dataset.
5.3. Comparisons with Human Age Estimation
On FG-NET the MAE of human age estimates over the
entire database is 4.7 with a variance of 24.8, while on the
PCSO data the MAE of human age estimates is 7.2 with
a variance of 32.0. On FG-NET (Fig. 7 (a)), the cumula-
tive scores of human age estimation and the proposed au-
tomatic method are fairly close to each other. However, on
the PCSO data ((Fig. 7 (b))) the automatic age estimation
scores are consistently better than the human age estima-
tion scores. To understand these results, we consider the
per age range errors. On the FG-NET data, the human age
estimates are significantly more accurate on the lower age
ranges (0-15) than higher age ranges. This explains the dif-
ference in the overall performance on FG-NET and PCSO,
since PCSO data does not include any face images of 0-15
year olds. In fact, the human MAE on FG-NET excluding
0-15 year olds is 7.4 with a variance of 32.8, which is fairly
consistent with the human MAE on PCSO database.
6. Conclusions and Future Work
We have proposed a hierarchical approach for automatic
age estimation, and analyzed the influence of aging on in-
dividual facial components using a component based rep-
resentation. Human perception ability to estimate age is
evaluated using crowdsourced data obtained via the Ama-
(a) FG-NET (b) PCSO
0 5 10 15
Absolute Erro r (years)
Accuracy (%)
Human Age Estimatio n
Automatic Age Estimation
0 5 10 15
Absolute Error (years)
Accuracy (%)
Human Age Estimation
Automatic Age Estimation
Figure 7. Human vs. automatic age estimation on (a) the FG-NET
database, and (b) a subset of the PCSO database with 2,200 im-
zon Mechanical Turk service, and compared with the per-
formance of the proposed automatic age estimation. Ex-
perimental results on the FG-NET, MORPH Album2, and
PCSO databases show that eyes and nose are more infor-
mative in age estimation than the other facial components
(forehead, eyebrows, mouth, and shape). We also show that
the performance of the proposed age estimation method is
better than or comparable to the age estimates provided by
humans on FG-NET and a small subset of PCSO database.
Although the fusion of per-component age estimates
provided no benefit on the MORPH Album2 and PCSO
databases, we nevertheless report comparable performance
to the best known result on MORPH Album2, and do so
without taking advantage of the ground truth demographic
information provided with the database. For future work,
we plan to investigate methods to improve our automatic
keypoint detection accuracy, additional strategies for age fu-
sion, as well as the possibility of further improving our re-
sults by incorporating demographic information in our age
estimation method.
[1] Aging Skin Net. Causes of aging skin.
[2] K.-Y. Chang, C.-S. Chen, and Y.-P. Hung. Ordinal hyper-
planes ranker with cost sensitivities for age estimation. In
Proc. IEEE CVPR, pages 585–592, 2011.
[3] W.-L. Chao, J.-Z. Liu, and J.-J. Ding. Facial age estima-
tion based on label-sensitive learning and age-oriented re-
gression. Pattern Recogn., 46(3):628 – 641, 2013.
[4] S. E. Choi, Y. J. Lee, S. J. Lee, K. R. Park, and J. Kim. Age
estimation using a hierarchical classifier based on global and
local facial features. Pattern Recogn., 44(6):1262–1281, Jun.
[5] Cognitec Systems GmbH. Facevacs software developer kit., 2010.
[6] T. Cootes, C. Taylor, D. Cooper, and J. Graham. Active
Shape Models - Their Training and Application. Comp. Vis.
Img. Und., 61(1):38 – 59, 1995.
[7] N. Ebner, M. Riediger, and U. Lindenberger. FACES - a
database of facial expressions in young, middle-aged, and
older women and men: Development and validation. Behav-
ior Research Methods, 42(1):351–362, 2010.
[8] Y. Fu, G. Guo, and T. Huang. Age synthesis and estimation
via faces: A survey. IEEE Trans. PAMI, 32(11):1955 –1976,
Nov. 2010.
[9] Y. Fu and T. Huang. Human age estimation with regression
on discriminative aging manifold. IEEE Trans. Multimedia,
10(4):578 –584, Jun. 2008.
[10] D. Gabor. Theory of communication. J. of the Institution of
Electrical Engineers, 93(26):429 –441, Nov. 1946.
[11] X. Geng, Z.-H. Zhou, and K. Smith-Miles. Automatic age
estimation based on facial aging patterns. IEEE Trans. PAMI,
29(12):2234–2240, Dec. 2007.
[12] G. Guo and G. Mu. Simultaneous dimensionality reduction
and human age estimation via kernel partial least squares re-
gression. In Proc. IEEE CVPR, pages 657 –664, 2011.
[13] G. Guo, G. Mu, Y. Fu, and T. S. Huang. Human age estima-
tion using bio-inspired features. In Proc. IEEE CVPR, pages
112–119, 2009.
[14] G. Guo and X. Wang. A study on human age estimation
under facial expression changes. In Proc. IEEE CVPR, pages
2547–2553, 2012.
[15] H. Han, B. Klare, K. Bonnen, and A. K. Jain. Matching
composite sketches to face photos: A component-based ap-
proach. IEEE Trans. IFS, 8(1):191–204, 2013.
[16] H. Han, S. Shan, X. Chen, and W. Gao. A comparative study
on illumination preprocessing in face recognition. Pattern
Recognition, 46(6):1691–1699, 2013.
[17] H. Han, S. Shan, X. Chen, S. Lao, and W. Gao. Separabil-
ity oriented preprocessing for illumination-insensitive face
recognition. In Proc. ECCV, pages 307–320, 2012.
[18] J. Hayashi, M. Yasumoto, H. Ito, and H. Koshimizu. Age
and gender estimation based on wrinkle texture and color of
facial images. In Proc. ICPR, pages 405–408, 2002.
[19] B. Heisele, P. Ho, J. Wu, and T. Poggio. Face recognition:
component-based versus global approaches. Comp. Vis. Img.
Und., 91(1-2):6–21, Jul. 2003.
[20] J. Huang, V. Blanz, and B. Heisele. Face recognition using
component-based svm classification and morphable models.
In Proc. Int’l Workshop Pattern Recogn. with SVM, pages
334–341, 2002.
[21] D. H. Hubel and T. N. Wiesel. Receptive fields, binocular
interaction and functional architecture in the cat’s visual cor-
tex. J. Physiology, 160(1):106–154, Jan. 1962.
[22] R. Iga, K. Izumi, H. Hayashi, G. Fukano, and T. Ohtani. A
gender and age estimation system from face images. In Proc.
SICE Annual Conference, pages 756–761, 2003.
[23] Y. Ivanov, B. Heisele, and T. Serre. Using component fea-
tures for face recognition. In Proc. FGR, pages 421 – 426,
May 2004.
[24] J. P. Jones and L. A. Palmer. An evaluation of the two-
dimensional gabor filter model of simple receptive fields in
cat striate cortex. J. Neurophysiology, 58(6):1233–1258,
Nov. 1987.
[25] N. Kumar, A. Berg, P. Belhumeur, and S. Nayar. Describable
visual attributes for face verification and image search. IEEE
Trans. PAMI, 33(10):1962–1977, 2011.
[26] A. Lanitis, C. Taylor, and T. Cootes. Toward automatic sim-
ulation of aging effects on face images. IEEE Trans. PAMI,
24(4):442–455, Apr. 2002.
[27] Z. Li, Y. Fu, and T. Huang. A robust framework for multi-
view age estimation. In Proc. IEEE CVPR Workshops, pages
9 –16, 2010.
[28] K. Luu, K. Seshadri, M. Savvides, T. Bui, and C. Suen. Con-
tourlet appearance model for facial age estimation. In Proc.
IJCB, pages 1 –8, 2011.
[29] S. Milborrow and F. Nicolls. Locating facial features with
an extended active shape model. In Proc. ECCV, pages 504–
513, 2008.
[30] C. Otto, H. Han, and A. Jain. How does aging affect facial
components? In Proc. ECCV, pages 189–198, 2012.
[31] M. Rhodes. Age estimation of faces: A review. Applied
Cognitive Psychology, 23(1):38 – 59, 1995.
[32] K. Ricanek and T. Tesafaye. MORPH: a longitudinal image
database of normal adult age-progression. In Proc. FGR,
pages 341–345, 2006.
[33] M. Riesenhuber and T. Poggio. Hierarchical models of ob-
ject recognition in cortex. Nature Neuroscience, 2(11):1019–
1025, Nov. 1999.
[34] S. Z. Li and A. K. Jain (eds.). Handbook of face recognition,
2nd edition. Springer-Verlag, London, 2011.
[35] T. Serre, L. Wolf, and T. Poggio. Object recognition with fea-
tures inspired by visual cortex. In Proc. IEEE CVPR, pages
994 – 1000, 2005.
[36] J. Suo, S.-C. Zhu, S. Shan, and X. Chen. A compositional
and dynamic model for face aging. IEEE Trans. PAMI,
32(3):385 –401, Mar. 2010.
[37] P. Thukral, K. Mitra, and R. Chellappa. A hierarchical ap-
proach for human age estimation. In Proc. IEEE ICASSP,
pages 1529 –1532, 2012.
[38] V. N. Vapnik. Statistical Learning Theory. John Wiley, New
York, 1998.
[39] M. C. Voelkle, U. L. N. C. Ebner, and M. Riediger. Let me
guess how old you are: Effects of age, gender, and facial
expression on perceptions of age. Psychology and Aging,
27(2):265–277, 2012.
[40] T. Wu, P. Turaga, and R. Chellappa. Age estimation and face
verification across aging using landmarks. IEEE Trans. IFS,
7(6):1780 –1788, Dec. 2012.
... In addition to the MAE, many researchers also adopt the Cumulative Score (CS) as a performance measure [15,39,89]. The CS (3) indicates the percentage of images for which the error is less than a certain number of years [39]. ...
... In addition to the MAE, many researchers also adopt the Cumulative Score (CS) as a performance measure [15,39,89]. The CS (3) indicates the percentage of images for which the error is less than a certain number of years [39]. The higher the CS, the more accurate is the age estimation system. ...
Full-text available
The precise age estimation of child sexual abuse and exploitation (CSAE) victims is one of the most significant digital forensic challenges. Investigators often need to determine the age of victims by looking at images and interpreting the sexual development stages and other human characteristics. The main priority - safeguarding children -- is often negatively impacted by a huge forensic backlog, cognitive bias and the immense psychological stress that this work can entail. This paper evaluates existing facial image datasets and proposes a new dataset tailored to the needs of similar digital forensic research contributions. This small, diverse dataset of 0 to 20-year-old individuals contains 245 images and is merged with 82 unique images from the FG-NET dataset, thus achieving a total of 327 images with high image diversity and low age range density. The new dataset is tested on the Deep EXpectation (DEX) algorithm pre-trained on the IMDB-WIKI dataset. The overall results for young adolescents aged 10 to 15 and older adolescents/adults aged 16 to 20 are very encouraging -- achieving MAEs as low as 1.79, but also suggest that the accuracy for children aged 0 to 10 needs further work. In order to determine the efficacy of the prototype, valuable input of four digital forensic experts, including two forensic investigators, has been taken into account to improve age estimation results. Further research is required to extend datasets both concerning image density and the equal distribution of factors such as gender and racial diversity.
... In the same vein, Han et al. (2013) employed a hierarchical model for age prediction and also examined the effect of aging on features extracted from a 3-D face image. Despite the fact that sparse and imbalanced datasets are prevalent restrictions in age estimation research, Chen et al. (2013) developed a method for crowd density and age prediction. ...
Full-text available
A computer-based age estimation is a technique that predicts an individual's age based on visual traits derived by analyzing a 2D picture of the individual's face. Age estimation is critical for access control, e-government, and effective human–computer interaction. The other-race effect has the potential to cause techniques designed for white faces to underperform when used in a region with black faces. The outcome is the consequence of intermittent training with faces of the same race and the encoding structure of the trained face images, which is based on the feature extraction technique used. This study contributes to a constructive comparison of three feature-extraction techniques, namely, local binary pattern (LBP), Gabor Wavelet (GW), and wavelet transformation, used in the development of a genetic algorithm (GA)-artificial neural network (ANN)-based age estimation system. The feature extraction techniques used are proven to produce a wealth of shape and textural information. The GA-ANN constitutes the age classifier module. The correct classification rate was chosen as the performance metrics in this study. The results demonstrated that the LBP is a more robust representation of the black face than the GW and Wavelet transformations, as evidenced by its accuracy rate of 91.76 compared to 89.41 and 84.71 achieved with the GW and Wavelet transformation age estimation systems, respectively.
... The majority of the research work in the field of soft biometric focuses on age [29,30,31], gender [32,33], and race [34] estimation from face images to date. Earlier, techniques like infrared iris images for finding out the age of people were used, which was possible by using the already existing iris datasets available [20,21]. ...
Full-text available
Soft Biometrics is a growing field that has been known to improve the recognition system as witnessed in the past decade. When combined with hard biometrics like iris, gait, fingerprint recognition etc. it has been seen that the efficiency of the system increases many folds. With the Pandemic came the need to recognise faces covered with mask in an efficient way- soft biometrics proved to be an aid in this. While recent advances in computer vision have helped in the estimation of age and gender - the system could be improved by extending the scope and detecting quite a few other soft biometric attributes that helps us in identifying a person, including but not limited to - eyeglasses, hair type and color, mustache, eyebrows etc. In this paper we propose a system of identification that uses the ocular and forehead part of the face as modalities to train our models that uses transfer learning techniques to help in the detection of 12 soft biometric attributes (FFHQ dataset) and 25 soft biometric attributes (CelebA dataset) for masked faces. We compare the results with the unmasked faces in order to see the variation of efficiency using these data-sets Throughout the paper we have implemented 4 enhanced models namely - enhanced Alexnet ,enhanced Resnet50, enhanced MobilenetV2 and enhanced SqueezeNet. The enhanced models apply transfer learning to the normal models and aids in improving accuracy. In the end we compare the results and see how the accuracy varies according to the model used and whether the images are masked or unmasked. We conclude that for images containing facial masks - using enhanced MobileNet would give a splendid accuracy of 92.5% (for FFHQ dataset) and 87% (for CelebA dataset).
... As a first illustrative application example, we analyze a popular regression task from computer vision: age prediction from facial images [34], [7], [1]. This problem represents a typical regression task but is often approached by classifying images into different age bins and therefore seems particularly suited to highlight the difference between XAIR and standard XAI. ...
Full-text available
In addition to the impressive predictive power of machine learning (ML) models, more recently, explanation methods have emerged that enable an interpretation of complex non-linear learning models such as deep neural networks. Gaining a better understanding is especially important e.g. for safety-critical ML applications or medical diagnostics etc. While such Explainable AI (XAI) techniques have reached significant popularity for classifiers, so far little attention has been devoted to XAI for regression models (XAIR). In this review, we clarify the fundamental conceptual differences of XAI for regression and classification tasks, establish novel theoretical insights and analysis for XAIR, provide demonstrations of XAIR on genuine practical regression problems, and finally discuss the challenges remaining for the field.
... In the same vein, Han and Otto [13] [14] employed a hierarchical model for age prediction and also examined the effect of aging on features extracted from a 3-D face image. ...
Full-text available
Age and gender estimation using face images by working on the facial features of the human, which are unique for each person can serve several applications such as person’s identification, access control, human–machine interaction, to avoid any type of fraud or misuse of someone’s identity, forensic, and in organizations. However, previous work for age estimation was based on handcraft features to encode the ageing patterns. With the advancement of deep architectures in CNN algorithms, CNN shows better performance than handcraft features. There exist several methods and relatively substantial literature on the field. However, biological variations and uncertainty will always be associated with age estimates because of the large variety in facial appearance and several other extrinsic and intrinsic factors. Thus, the proposed work is based on age and gender estimation using an improved convolutional neural network (CNN). The proposed model improves the computational performance, and it is generalized to another dataset also. The convolutional layer is the fundamental building block of a CNN. Like other neural networks, CNNs consist of neurons that are capable of learnable weights and biases. The proposed model is evaluated using UTKFace, IMDB-WIKI, FG-NET, CACD datasets. The efficacy of the proposed algorithm in comparison to existing algorithms achieves better accuracy in age estimation through extensive simulations. The proposed model is trained on UTKFace (aligned and cropped faces) dataset has shown 94.01% accuracy for age and 99.86% accuracy for gender estimation.
Full-text available
Introduction of large training datasets was essential for the recent advancement and success of deep learning methods. Due to the difficulties related to biometric data collection, facial image datasets with biometric trait labels are scarce and usually limited in terms of size and sample diversity. Web-scraping approaches for automatic data collection can produce large amounts of weakly labeled and noisy data. This work is focused on picking out the bad apples from web-scraped facial datasets by automatically removing erroneous samples that impair their usability. The unsupervised facial biometric data filtering method presented in this work greatly reduces label noise levels in web-scraped facial biometric data. Experiments on two large state-of-the-art web-scraped datasets demonstrate the effectiveness of the proposed method with respect to real and apparent age estimation based on five different age estimation methods. Furthermore, we apply the proposed method, together with a newly devised strategy for merging multiple datasets, to data collected from three major web-based data sources (i.e., IMDb, Wikipedia, Google) and derive the new Biometrically Filtered Famous Figure Dataset or B3FD. The proposed dataset, which is made publicly available, enables considerable performance gains for all tested age estimation methods and age estimation tasks. This work highlights the importance of training data quality compared to data quantity and selection of the estimation method.
Conference Paper
Full-text available
We consider the problem of automatic age estimation from face images. Age estimation is usually formulated as a regression problem relating the facial features and the age variable, and a single regression model is learnt for all ages. We propose a hierarchical approach, where we first divide the face images into various age groups and then learn a separate regression model for each group. Given a test image, we first classify the image into one of the age groups and then use the regression model for that particular group. To improve our classification result, we use many different classifiers and fuse them using the majority rule. Experiments show that our approach outperforms many state of the art regression methods for age estimation.
Full-text available
In this paper we propose a novel Contourlet Appearance Model (CAM) that is more accurate and faster at localizing facial landmarks than Active Appearance Models (AAMs). Our CAM also has the ability to not only extract holistic texture information, as AAMs do, but can also extract local texture information using the Nonsubsampled Contourlet Transform (NSCT). We demonstrate the efficiency of our method by applying it to the problem of facial age estimation. Compared to previously published age estimation techniques, our approach yields more accurate results when tested on various face aging databases.
In this paper, we study human age estimation in face images under significant expression changes. We will address two issues: (1) Is age estimation affected by facial expression changes and how significant is the influence? (2) How to develop a robust method to perform age estimation undergoing various facial expression changes? This systematic study will not only discover the relation between age estimation and expression changes, but also contribute a robust solution to solve the problem of cross-expression age estimation. This study is an important step towards developing a practical and robust age estimation system that allows users to present their faces naturally (with various expressions) rather than constrained to the neutral expression only. Two databases originally captured in the Psychology community are introduced to Computer Vision, to quantitatively demonstrate the influence of expression changes on age estimation, and evaluate the proposed framework and corresponding methods for cross-expression age estimation.
Age estimation and face verification across aging are important problems with a wide range of applications. It is well known that age and identity information are encoded in both texture and shape of the face. Building on recent advances in landmark extraction and statistical techniques for landmark-based shape analysis, we consider these problems using facial shapes. We show that by using well-defined shape spaces and their associated geometry, one can obtain significant performance improvements in both age estimation and face verification. Toward this end, we propose to model the facial shapes as points on a Grassmann manifold. Age estimation and face verification are then considered as regression and classification problems on this manifold. Algorithms for regression and classification are designed to take into account the geometry of the underlying space. The proposed method is flexible and can be used as a standalone age estimator or classifier, and we also present methods for fusion with texture-based algorithms.
Model-based vision is firmly established as a robust approach to recognizing and locating known rigid objects in the presence of noise, clutter, and occlusion. It is more problematic to apply model-based methods to images of objects whose appearance can vary, though a number of approaches based on the use of flexible templates have been proposed. The problem with existing methods is that they sacrifice model specificity in order to accommodate variability, thereby compromising robustness during image interpretation. We argue that a model should only be able to deform in ways characteristic of the class of objects it represents. We describe a method for building models by learning patterns of variability from a training set of correctly annotated images. These models can be used for image search in an iterative refinement algorithm analogous to that employed by Active Contour Models (Snakes). The key difference is that our Active Shape Models can only deform to fit the data in ways consistent with the training set. We show several practical examples where we have built such models and used them to locate partially occluded objects in noisy, cluttered images.
This paper provides a new age estimation approach, which distinguishes itself with the following three contributions. First, we combine distance metric learning and dimensionality reduction to better explore the connections between facial features and age labels. Second, to exploit the intrinsic ordinal relationship among human ages and overcome the potential data imbalance problem, a label-sensitive concept and several imbalance treatments are introduced in the system training phase. Finally, an age-oriented local regression is presented to capture the complicated facial aging process for age determination. The simulation results show that our approach achieves the lowest estimation error against existing methods.