# DeepUAge: Improving Underage Age Estimation Accuracy to Aid CSEM Investigation

## Full Text

This Markdown is an automated text extraction for search and reading tools. The [hosted PDF](UnderageFacialAgeEstimation.pdf) is the authoritative publication source.

Source PDF: [UnderageFacialAgeEstimation.pdf](UnderageFacialAgeEstimation.pdf)
Source SHA-256: `d2b3219108476861adf38bc4fcdfbf371af0a733829958f88a9f2905756a318e`
Generated with `pdftotext -nopgbrk`; formatting and reading order may differ from the PDF.

---

Forensic Science International: Digital Investigation 32 (2020) 300921

Contents lists available at ScienceDirect

Forensic Science International: Digital Investigation
journal homepage: www.elsevier.com/locate/fsidi

DFRWS 2020 EU e Proceedings of the Seventh Annual DFRWS Europe

DeepUAge: Improving Underage Age Estimation Accuracy to Aid CSEM
Investigation
Felix Anda*, Nhien-An Le-Khac, Mark Scanlon
Forensics and Security Research Group, University College Dublin, Ireland

a r t i c l e i n f o

a b s t r a c t

Article history:

Age is a soft biometric trait that can aid law enforcement in the identification of victims of Child Sexual
Exploitation Material (CSEM) creation/distribution. Accurate age estimation of subjects can classify
explicit content possession as illegal during an investigation. Automation of this age classification has the
potential to expedite content discovery and focus the investigation of digital evidence through the
prioritisation of evidence containing CSEM. In recent years, artificial intelligence based approaches for
automated age estimation have been created, and many public cloud service providers offer this service
on their platforms. The accuracy of these algorithms have been improving over recent years. These
existing approaches perform satisfactorily for adult subjects, but perform wholly inadequately for underage subjects.
To this end, the largest underage facial age dataset, VisAGe, has been used in this work to train a
ResNet50 based deep learning model, DeepUAge, that achieved state-of-the-art beating performance for
age estimation of minors. This paper describes the design and implementation of this model. An evaluation, validation and comparison of the proposed model is performed against existing facial age classifiers resulting in the best overall performance for underage subjects.
© 2020 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).

Keywords:
Child Sexual Exploitation Material (CSEM)
Age estimation
Underage facial age dataset
Child sexual abuse investigations
Deep learning

1. Introduction
The technological advancement of modern society encompasses a
myriad of interconnected digital devices. Petabytes of information
about our lives are consciously uploaded daily to various social
media platforms and communications via mobile messaging applications. Nevertheless, personal data is also collected surreptitiously
from location based services, facial detection by private or municipal
CCTV cameras and omnipresent internet-connected sensors.
The prevalence of multimedia content has necessitated the
storing of increasing volumes of information e resulting in an
increased local storage capacity on mobile device and the cloud
servers where the data is backed-up. Also, consumers continuously
demand higher quality photographs that require an increase on the
number of pixels in digital cameras. Moore's law has negatively
affected digital forensic investigations due to the amount of devices
that are seized in modern crime scenes coupled with the increasing
storage per device (Lillis et al., 2016). After the evidence acquisition

* Corresponding author.
E-mail addresses: felix.andabasabe@ucdconnect.ie (F. Anda), an.lekhac@ucd.ie
(N.-A. Le-Khac), mark.scanlon@ucd.ie (M. Scanlon).

phase, the devices are stored for further analysis. An exponential
accumulation of devices in digital forensic laboratories is an
ongoing issue that has contributed to backlogs in Law Enforcement
Agencies (LEAs) over the past years throughout the globe (Scanlon,
2016). Intelligent automation is needed to expedite digital investigations that are hampered by lack of resources, such as time
and skilled expertise. Moreover, Sanchez et al. (2019) verified that
digital forensic practitioners demand automated tools to detect
CSEM, age estimation and skin tone detection and intelligent
artefact prioritisation can expedite digital investigation (Du and
Scanlon, 2019). Our research tackles the age estimation problem
for minors. Our age estimation model discards any unnecessary
details, i.e., background and other noise. Based on the No-FreeLunch theorem, i.e., there is no single model that works best for
every problem, our focus was to find a model that suits best for the
underage age estimation. The approach may fail in other situations,
such as adult age ranges.
The application area for this work is to execute facial age estimation over images that have already been flagged with nudity and
be further used in conjunction with an ensemble of models that
endeavour to solve the bigger problem, i.e., the automated categorisation of this content as CSEM. The element of study is depicted

https://doi.org/10.1016/j.fsidi.2020.300921
2666-2817/© 2020 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/
).

S2

F. Anda et al. / Forensic Science International: Digital Investigation 32 (2020) 300921


 Creation of DCA, an encapsulation of an offline pre-processing
method suitable for law enforcement agencies (LEA) used for
the automatic age prediction of unlawful images. Available at
https://github.com/4ND4/DeepUAge.

2. Related work

Fig. 1. Pipeline of designed ensembles to detect CSEM.

as the Underage Facial Age Detector in Fig. 1. Although the pipeline
is comprised of several components, our goal is to achieve the
highest accuracy possible for underage subjects between the age of
1 to 18 years old. In the aforementioned pipeline, an initial image
filter will decrease the input of images based on nudity/pornography detection, such as the technique developed by Sae-Bae et al.
(2014). A nudity/pornography detector would serve as a primary
filter to aid the ensemble and facilitate the encapsulation of our age
estimation methodology.
Challenges arise due to the lack of underage facial images for
training models, the validation of age labels, and the use of estimated
or simply guessed values as ground truths in existing datasets.
Moreover, performance within this age group has been previously
evaluated with state-of-the-art age estimators and the inadequate
results are a motivation for our research. Anda et al. (2019) assessed
several online services such as Amazon Rekognition, Microsoft Azure
Cognitive Services, How-Old.net, and Deep EXpectation (DEX), an
offline Caffe1 model and winner of the ChaLearn LAP 2015 challenge.
To address the existing challenges, this work created a balanced and
validated underage dataset (VisAGe), a non-binary age estimation
model was trained, and the accuracy of underage age estimation was
improved.
The paper is organised as follows. In Section 2, an overview of the
related work is presented. Section 3 provides an overview of the
design and methodology of the developed model and its derivation
from the VisAGe dataset, and the development of the DCA contouring pre-processing technique. Section 4 describes the performance of
the DeepUAge model in comparison with other state-of-the-art facial
age estimators as well as the evaluation of the pre-processing
techniques used. Section 4.1 explains the evaluation for the age
estimation framework. Finally, the last section outlines the conclusions and discusses future work.
Contribution of this work:

 Creation of a dataset consisting of an underage facial age dataset
with validated, accurate age and gender labels.

 Creation of DeepUAge an age estimation state-of-the-art model
with a mean absolute error (MAE) rate of 2.73 years for underage subjects.

 Evaluation of several facial image pre-processing techniques
and the impact of facial landmark points for age estimations on
subjects from the age range 1 to 18 years old.

1

https://caffe.berkeleyvision.org/.

Many facial age estimation methods are built on datasets with
estimated or guessed age label, low sample counts, highly noisy
data, unbalanced/biased data, or age bins (e.g., binary groups that
separate underage subjects from adults, and/or small groups such
as child, teen, adult, etc.). Usually, the usage of groups hinder the
filtering of illegal content due to its non-dynamic structure. Some of
the aforementioned methods affect the reliability of the machine
learning assistance and hence decrease the credibility of the
approach. Furthermore, any introduction of reasonable doubt may
dismiss a case. To attempt to alleviate this issue, the Daubert
standard is suggested. It was introduced in 1993 and has been used
by most state courts in the USA as a rule of evidence to assess the
reliability of scientific evidence through the following factors: (1)
the method can be and has been tested, (2) subject to peer review,
(3) error rates are acceptable, (4) general acceptance in the scientific community (Nawara, 2010). In regards to these factors, Nutter
(2018) states that “machine learning easily satisfies three of the
four Daubert factors without extensive discussion”.
2.1. Facial age estimation
Automated facial age estimation requires three main properties:
validated facial age labelled datasets, robust face detection, and
machine learning. Many existing approached reuse the same
dataset(s) either for bench-marking, validation, training or testing
purposes. In the early stages, facial anthropometric2 models were
suggested for age prediction. In 2006, Ramanathan and Chellappa
(2006) proposed a cranio-facial growth model that classifies
growth related shape variations observed in underage faces. Their
model was capable of face recognition across age progression. The
dataset for the aforementioned model was the FG-Net ageing
dataset (82 subjects with ages ranging from 0 to 69 years old and
over 50% juvenile subjects3) and a separate dataset containing 233
images (Ramanathan and Chellappa, 2006).
In 2009, Guo et al. (2009) found that age estimation performance was able to improve when manifold learning uses biologically inspired features (BIF). The accuracy achieved is positively
influenced by a known gender; therefore, their approach consisted
in two different MAEs for each gender. Furthermore, the model was
a combination of “BIF locality sensitive discriminant analysis” and
“BIF marginal fisher analysis”, reaching MAE rates of 2.58 years for
males and 2.61 years for females. The data used was the large
Yamaha gender and age (YGA) database that contains 8,000 outdoor facial images of Asian ethnicity subjects. The dataset was
distributed equally by gender and divided in age ranges from 0 to
93 with intervals of 9 classes per age group until the age of 70 and a
group containing the rest due to the lack of images available for
subjects over 70.
Eidinger et al. (2014) presented an approach on age and gender
estimation using standard linear support vector machines (SVM)
with their own dropout-SVM scheme. The dataset employed was
Adience, which is a collection of Creative Commons (CC) images

2
The science of measuring sizes and proportions on human faces (Ramanathan
and Chellappa, 2006).
3
https://yanweifu.github.io/FG_NET_data/index.html.

F. Anda et al. / Forensic Science International: Digital Investigation 32 (2020) 300921

sourced from Flickr. Furthermore, the ages were labelled in heterogeneous categories in an unknown manner that contained in
average 2,205 images per age group, but were unbalanced.
More recently in 2018, Rothe et al. (2018) proposed a deep
learning solution based on a VGG16 convolutional neural network
architecture pre-trained on ImageNet.4 This achieved a MAE of
3.252 years. The data was trained on a dataset named IMDB-WIKI,
consisting in over half a million facial images of celebrities
(currently the largest public face image dataset annotated with age
and gender labels) that had been crawled from IMDB and cross
referenced with the age denoted in Wikipedia.
2.2. Underage facial age estimation
Work specialised in underage facial age estimation has been
limited due to the challenges of collecting data which is understandably subject to ethical implications, lack of underage datasets,
and scarcity of reliably annotated images. Nevertheless, apparent
age estimation on children was studied by Antipov et al. (2016) in
2016. Antipov et al. used a fine-tuned VGG-16 (a very deep convolutional network of 16 weight layers used for large-scale image
classification) to train a model of minors from the age range 0 to 12
years old. HeadHunter, a detector based on rigid templates, was
their choice for face detection and the alignment technique was
based on a multi-view facial landmark detection tool. The error rate
achieved for validation was 0.2609, a metric ε defined as the size of
the tail of the normal distribution with the mean m and the standard
deviation s with respect to the predicted value x^.
In the same year, Ferguson and Wilkinson (2017) determined
that manual human visual age estimation of childrens’ faces reveals
poor accuracy, confirming the difficultly to precisely predict age.
They also suggested that black and white images were classified
with less accuracy. The latter suggestion was confirmed with
several experiments that are mentioned in Section 4.1. In 2019,
Anda et al. improved facial age estimation with an ensemble
technique approach that was fine-tuned on DEX for the age group
16 to 17 years old, which falls in borderline adulthood age in many
jurisdictions. Their work also evaluated the state-of-the-art cloud
based facial age estimators, such as Amazon AWS Rekognition, and
Microsoft Azure Cognitive Services.
2.3. Facial pre-processing for age estimation
Facial image pre-processing may not be necessary if the source
is akin to a standard passport photograph. However, facial images
in-the-wild may have characteristics such as various pitch/roll/yaw
angles, multiple subjects per image, background noise, varying
image size and quality, etc. Such photos require image preprocessing and normalisation to align and remove unnecessary
features. Reisfeld and Yeshurun (1998) suggest that the knowledge
of the location and the scale of a face impacts positively on the
speed and reliability of face recognition systems.
In 2015, Han et al. (2015) designed a face pre-processing procedure to overcome image variations due to external factors. Their
approach entails: (1) converting a colour facial image into greyscale, (2) rectifying the face based on the two eyes and cropping to
60x60 pixels with a 32-pixel interpupillary distance (IPD), (3)
detecting the face and the eyes using Cognitec's commercial FaceVACS SDK, and (4) applying the Difference of Gaussians filtering. In
the same year, Liu et al. (2015) proposed a deeply learned regressor
and classifier for robust apparent age estimation. A three step
preprocessing procedure was implemented: face detection, facial

4

http://www.image-net.org/.

S3

landmark localisation, and facial normalisation. For the first, a face
detection toolkit developed by VIPL lab of CAS was used; for the
second, 5 facial landmarks were detected with a Coarse-to-Fine
Auto-Encoder Network. For the last step, external and internal
normalisation approaches were considered.
2.4. Facial age dataset
Large datasets for underage subjects with accurate labels are
rare. Accurately labelled age and gender datasets are preferred over
apparent age estimation and guessed metadata. In 2013, Dalrymple
et al. created a set of images with the following variations: 8 facial
expressions, 5 angles and 2 lightning conditions. The collection
consisted of combinations of these variations for 40 male and 40
female Caucasian children between 6 and 16 years-old. The real age
was documented and also estimated by external raters with a 79.7%
accuracy. Later in 2014, a 50 image dataset of female subjects aged
10 to 19 years from Germany, Italy, and Lithuania was created. In
2015, the In-The-Wild Child Celebrity (ITWCC) dataset was created
by Ricanek et al. and the set was composed of 304 subjects with a
total of 1,715 images (876 female and 839 male) from the age range
5 months to 32 years. Next in 2016, the Boys2Men collection was
released as a private dataset mainly focused on male images from
n-Santana et al., 2016). In
the age range 12 to 21 years-old (Castrillo
2018, Deb et al. released a dataset containing 3,682 face images of
919 subjects, in the age group 2 to 18. It is notable that in the past 5
years, the number of underage datasets has grown but still requires
validation, accurate age labels, and balance.
3. Design/methodology
This work presents an underage facial age estimation deep
convolutional neural network (DCNN) based on a residual neural
network of 50 layers (ResNet50). Our proposed method is pretrained on the ImageNet Dataset.5 The last Fully-Connected (FC)
Softmax layer with 1,000 outputs has been replaced by a FC Softmax activation function layer of 20 outputs that correspond to the
age classes studied (1 to 20 years old) to suit our needs. Subsequently, the parameters of the convolutional layers during the
training process have been frozen. The ResNet50 architecture
employed for facial age estimation and the replacement in the last
layer with 20 outputs can be seen in Fig. 2. The age estimation
problem was treated as a classification task and therefore, a categorical cross-entropy logarithmic loss function is used.
The batch size is a hyper-parameter that defines the number of
training samples used in one iteration and is directly proportional
to the memory space required. A batch size of 64 was chosen due to
RAM limitations with the development server. One forward pass
and one backward pass of all the training examples are referred to
as epochs. A reference of 100 epochs was chosen; however, the
training process was monitored and an early stopping implementation to prevent over-fitting was accomplished. The metric
used for accuracy was the MAE, i.e., the average of the absolute
mean error between the ground truth and the predicted values. The
formula used to evaluate the model is described in Equation (1).

MAE 1⁄4

n
X
jpredictedi  reali j
i1⁄41

n

(1)

The optimiser chosen was stochastic gradient descent (SGD)
that includes support for momentum, learning rate decay, Nesterov

5
ImageNet is a large scale ontology of images organised according to the
WordNet hierarchy (Deng et al., 2009).

S4

F. Anda et al. / Forensic Science International: Digital Investigation 32 (2020) 300921

Fig. 2. ResNet50 Pre-trained on ImageNet with 20 Outputs in the FC Softmax layer.
Grouping of Convolution Layers are Denoted by Colour.

momentum. SGD demonstrates excellent performance for largescale problems (Bottou et al., 2010). The learning rate is a hyperparameter that controls the number of changes affected by the
model in response to the estimated error each time the model
weights are updated. The selected value for the initialiser was 0.1
and the momentum was 0.9. The latter is a parameter that accelerates SGD in the relevant direction and reduces oscillations. The
input image size for width and height chosen was 224 with the
depth set to 3; values chosen due to the hardware limitations and
the decreased performance experienced with images of smaller
dimensions. Data augmentation techniques such as flip, rotation,
zoom, distortion, colour, contrast, brightness and random erasing6
were used. Moreover, augmentation in data-space provides a
greater benefit for improving performance and reducing overfitting
(Wong et al., 2016).
3.1. Proposed facial age and gender dataset
Age and gender estimation models require a large number of
images with real age and gender labels; moreover, the data must be
balanced within each class and this represents a challenge.
Nevertheless, it was possible to build a balanced set divided with
images from two sources; the VisAGe dataset supplemented by the
Anda et al. (2018) dataset generator.
The creation of VisAGe involved the annotating of the largest
set of underage images to date. These images are creative commons licensed with initially indicative age gathered from Flickr.
Each of these photos were processed by face and gender detection
algorithms, and other associated metadata were compiled such as
dimensions, title, tags, comments, dates, etc. Given the level of

6
“Random erasing randomly selects a rectangle region in an image and erases its
pixels with random values. In this process, training images with various levels of
occlusion are generated, which reduces the risk of over-fitting and makes the model
robust to occlusion” (Zhong et al., 2017).

error rates in automated facial age estimation and gender identification, each of these images were subjected to human age and
gender verification. Each photo was voted on by three human
assessors and if the decisions on age and gender were unanimous,
the photo was added to the dataset. The set used in this work
consists of 19,446 images from the age range 1 to 18. Further detail
regarding age and gender per class is depicted in Table 1. It is
notable that the age ranges contain an unbalanced amount of
images within each age and/or gender group. The average number
of male images per age is 521 versus 557 females. This unbalance
will be addressed in future work.
In order to present a sample of the dataset, an average image per
class has been calculated. Average faces from age 1 to 18 can be seen
in Fig. 3. The face cropping technique used in this paper is also
applied and visible on each face.
The age estimation model was trained with the majority of
images from the VisAGe dataset and was prepared in a balanced
fashion. Due to the size of the dataset available, 800 photos were
selected from each class. When enough images were available for
each class, a balanced amount of images were selected for both
male and females. In the age classed of 8 and higher, there were
insufficient images to fulfil the 800 training/testing images. As a
result, the remainder of the 800 images used were filled with the
Anda et al. (2018) facial age dataset generator with randomly obtained images from different datasets including FGNET (Lanitis and
Cootes, 2002), IMDB-WIKI (Rothe et al., 2018), FERET (Phillips et al.,
1998), MEDS (Founds et al., 2011).
Although the model application is for underage images, it was
necessary to consider 2 additional years over the maximum year
limit (18 years old). Since the best performance of existing approaches are approximately 2e3 MAE in years, the chosen limit was
20. Furthermore, the additional age classes were also completed
with the aforementioned dataset generator.
3.2. Facial image preprocessing
An important step for an image classification task is to filter
unnecessary features that would affect the learning process of a
machine learning algorithm. Initially, the approach implemented
used the Faceþþ7 API to obtain the face landmarks rather than the
open source machine learning library dlib (King, 2009). Once the
landmarks were collected, the left and right eye centre values were
processed to further compute the angle between the eye centroids.
Next, the median point between the two eyes in the input image
was computed and subsequently rotated. Finally an affine transformation was applied to the image with warping using the specified matrix, as per Equation (2):

dstðX; YÞ 1⁄4 srcðM11 x þ M12 y þ M13 ; M21 x þ M22 y þ M23 Þ

(2)

Another approach would be to solve the procrustes problem
(Gower, 1975) by subtracting centroids, scaling by the standard
deviation, and then using the singular value decomposition to
calculate the rotation. Once the face was aligned, the new facial
landmark positions were detected and a mask for the 273 contour
points provided by Faceþþ was created. The mask is overlaid and
the face is cropped. An example of a dlib cropped face vs a Faceþþ
cropped face can be seen in Fig. 4. The dlib landmark tool extracts
27 relevant contour points from the 68 points in total against 273/
1000 from Faceþþ.
The major drawback of using the Faceþþ API is that the images
must be sent to a remote cloud service and the accessibility is

7

https://www.faceplusplus.com/

F. Anda et al. / Forensic Science International: Digital Investigation 32 (2020) 300921
Table 1
VisAGe dataset - facial images per class per gender from 1 to 18 Years old.
Age

Combined

Male

Female

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
Total

4,236
2,722
2,280
2,434
1,227
984
974
686
453
401
371
211
354
217
337
589
285
660
19,446

2,292
1,485
1,071
1,110
515
462
418
315
256
217
154
103
171
142
91
184
204
193

1,944
1,237
1,209
1,324
712
522
556
371
197
184
217
108
183
75
246
405
81
467

limited to this cloud environment. In dealing with CSEM investigations, LEAs cannot transmit this sensitive information to a
third party service. To overcome this issue, a customised facial
cropping technique was implemented using dlib as a base and
extending it to predict the hairline with a facial proportion artistic
approach by Loomis (2017).

3.2.1. DCA facial proportion artistic approach
Initially, the Faceþþ and dlib pre-processing techniques were
analysed to decide which approach was most ideal to utilise for
the validation of the model. Exploration of the Faceþþ 1,000
landmark points detection tool has 273 contour points (refer to
the image on the middle of Fig. 4). Unfortunately, due to the
problem with remote cloud services stated in Section 3.2 likely
being insurmountable for CSEM investigation, the dlib tool was
selected. The novel pre-processing technique, DCA, was implemented instead based on dlib. The dlib library returns 68
landmark points from which 27 correspond to the face contour. In
the left of Fig. 4, the dlib jawline contour highlighted in green
from point number 1 to 27 can be seen. Portions of the head, such
as the forehead or wrinkles, are important features that impact
the age estimation of a subject; these features are not supported
with the 68 landmark detector. The DCA approach was addresses
this limitation, as can be seen on the right of Fig. 4). It uses facial
proportionality to reconstruct the face and obtain landmarks that
are close to the hairline.
Fig. 5 depicts the proportionality between the nose, eyes and
hairline contour. To predict the hairline landmarks, the following
steps that emulate the face drawing methodology was carried out
as follows:
1. Use the dlib landmark detector to obtain the coordinates x, y of
the lowest point of the nose which corresponds to the point 34.
2. Compute the average distance between the point 34 and the
intersecting points that lie close to a perpendicular drawn from
the nose point towards the contours. The square side is twice
this value.

d1 1⁄4

rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 
2

pleftx  p34x þ plefty  p34y

S5

rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi

2 
2
d2 1⁄4
prightx  p34x þ prighty  p34y
squareside 1⁄4

d1 þ d2
*2
2

squareside 1⁄4 d1 þ d2
3. Locate both vertexes v1 ; v2 of the square and use those values to
draw the shape within the circle as shown in Fig. 5.
4. From the centre points cx ; cy , draw a regular polygon of size
N 1⁄4 20 (Icosagon). Notice that the circle drawn in Fig. 5 corresponds to the half icosagon drawn in Fig. 6.
Our method uses the foundations of the Loomis face proportion
approach. In Fig. 6, the output of a digital sketch of our approach to
detect hairline landmark points is depicted. Last, a cropped mask
can be used to filter noise from facial images.

4. Results
As mentioned in Section 3.1, a balanced dataset of 16,000 images
was prepared. 80% of the images were used for training and 20%
were used for validation. The training stopped on 87 epochs,
maintained a loss under 1.799 with a favourable MAE of 1.57 years
and 2.73 years for validation and testing respectively. The model
was further tested with 1,000 additional images that were gathered
from the UTKFace dataset (Zhang et al., 2017) and the aforementioned dataset generator. These images amounted to 50 images per
class. The testing achieved a MAE of 2.73 years.
In the process of creating the DeepUAge model, the DCA facial
proportion artistic approach was developed. Its efficiency in preprocessing has been evaluated in comparison to other preprocessing techniques. It produced a MAE of 2.73 years, which
was the best performing of all the approaches evaluated. The results for this experiment can be found in Table 2. The accuracy of
the DeepUAge model in both validation and testing in comparison
with other facial age estimators for underage subjects is outlined in
Section 4.1.3.

4.1. Evaluation
4.1.1. DCA and other pre-processing techniques
By separating the age estimation problem into a smaller scope fo
age range, it was possible to validate the input data of juveniles
used to train the model. As the data used was not only frontal
“passport photo style” facial images, pre-processing procedures
have been adopted to minimise any negative impact of of noise on
the final results.
Several types of existing pre-processing techniques were evaluated; dlib contour aligned, dlib contour non-aligned, dlib
cropped, and Faceþþ contour. Variations of the dlib approaches
all produced a MAE greater than 4 years. In particular, the dlib
cropped technique was found to perform almost equally as not
using a pre-processing filter at all. Additionally, even the best performing dlib contour aligned, was still 43.72% greater in MAE
versus the Faceþþ contour. Unfortunately, due to the unsuitability
of Faceþþ for CSEM investigation, the technique was deemed unusable for the DeepUAge model. This motivated the development of
the DCA facial cropping techniques, which nonetheless achieved
better results than all other pre-processing techniques evaluated.

S6

F. Anda et al. / Forensic Science International: Digital Investigation 32 (2020) 300921

Fig. 3. VisAGe dataset - average face per age from 1 Year old (top-left) to 18 Years old (bottom-right).

Fig. 4. Image Taken from the FG-NET Aging Database (Wallhoff, 2006) with 64 dlib Landmarks (left), 1000 Faceþþ landmarks (middle), and the DCA approach (right).

correct catalogue of the images. Furthermore, for ages 2, 3, 9 and 11
the correct predicted age is between the minimum and the 1st
quartile of the classification distribution. There are several outliers
across all ages, predominately in ages 7, 13, 16, 17 and 19, where 4 or
more outliers can be observed.

Fig. 5. dlib landmark points (3, 34, 15) superimposed over a Loomis face proportion
approach sketch (Fussell, 2019) (image reproduced with permission).

4.1.2. DeepUAge performance
Fig. 7 illustrates a box plot of real age vs. DeepUAge predicted
age where accuracy and precision for the age range 1e20 can be
observed. Ages 18 and 20 have the largest range of predicted age.
Whilst the model manages to obtain correct classification for both
ages, the performance score for these classification age ranges is
low. Instead, the correct age prediction is found at the maximum
whiskers of the data for the age of 20 and at the 3rd quartile of the
data distribution for the age of 18.
Similar results can be seen across the age range 14 to 20 where
the correct age lies between the maximum and 3rd quartile of the
classification distribution. Conversely, the reverse can be observed
on the opposite end of the age spectrum where younger ages are
often overestimated. It can be observed that for ages 1, 5, 8 and 10
the minimum of the classification distribution contained the
Fig. 6. Contour reconstruction from Dlib.

F. Anda et al. / Forensic Science International: Digital Investigation 32 (2020) 300921
Table 2
Results of different preprocessing techniques.

S7

Table 3
Evaluation of facial age estimators for underage subjects.

Approach

MAE

Approach

MAE

DCA
Faceþþ contour
Dlib contour aligned
Dlib contour non-aligned
Dlib cropped
Non-processed

2.73
2.79
4.01
4.28
5.31
5.71

DeepUAge Test
Microsoft Azure Face API
Amazon Rekognition
Dummy estimator (All assumed 10 y/o)
Faceþþ
IMDB-WIKI WideResNet (Rothe et al., 2018)

2.73
3.60
3.74
5.00
18.21
20.43

4.1.3. Comparison of DeepUAge and other estimation techniques
State-of-the-art age estimators were compared against
DeepUAge, with the same testing set. The test performance achieved is inline with the state-of-the art age estimation classifiers.
The testing MAE of 2.73 for DeepUAge indicates that the average
magnitude of the model error was significantly lower than that of
its alternatives. The top 3 performers found in descending order
were DeepUAge, Microsoft Azure Face API, followed by Amazon
Rekognition, as can be seen in Table 3. Both Microsoft's and
Amazon's approaches were side by side in performance with only
being 0.14 MAE apart. DeepUAge stood at 0.87 MAE better than
Microsoft Azure Face API and exceeding Amazon Rekognition by a
MAE of 1.01.
Despite obtaining the best results of the age estimators evaluated, DeepUAge has a logarithmic loss (the variation of the actual
label from the machine learning model predicted value) of 1.799.
Our goal is to decrease this value further to as close to 0 as possible
and is addressed as our future work in Section 6. Although IMDBWIKI WideResNet was trained on a large celebrity dataset (over
half a million images), it had the lowest performance. Along with
Faceþþ, both estimators had the poorest performance overall. The
dummy estimator (an approach which classifies images to a fixed
predicted age of 10) managed to surpass the performance of two
intricate algorithms. This result can be due to the failure to validate
age labels and the lack of images in the underage age group.
As shown in Table 3, the overall performance of DeepUAge was
found to be best compared with the other models evaluated. This is
demonstrated at a better granularity in Fig. 8 where DeepUAge is
shown to catalogue age 9 to 18 with a higher accuracy than those of
the other age predicting services. In reference to Fig. 7, it can be
further concluded that the precision of DeepUAge is at its peak for
ages 7, 8, 10, 13, 16 and 17. In particular, it is most accurate at age
classification of 13 and 17 year olds; therefore providing better age
classification techniques for “early adolescent (age 10 to 14) and

Fig. 8. Average difference per age per facial age predictor.

late adolescent (age 15 to 19) years” (Signorielli, 1987). This age
range is important in CSEM investigations because it covers subjects within the borderline of adulthood.
5. Discussion
According to the No-Free-Lunch theorem, the age scope for our
age estimation problem was established. Although binary classification models are not used, numeric ages are used and the problem
was treated as a classification problem as opposed to a regression
problem. Several factors including environment, habits, makeup,
and ethnicity can have an impact on our perceived age and therefore make this problem more complex. Until now, the lack of underage datasets and scarce validation has had an influence in the
performance of age estimation models. This work has been able to
improve the performance of age estimation for underage subjects.
One weakness of our model is the data itself e demographic
balance in regards to ethnicity is lacking. Part one of the dataset
used relies on input labels from Flickr users. Naturally, it is not often
that a person classifies/tags images with these factors online or in
social media. To improve facial age estimation, the problem should
be further divided by gender and ethnicity and once a model is
generated for each type, create an ensemble that could work in a
hierarchical manner. The algorithm should be able to chose the
model based on the age range, gender, ethnicity, and other relevant
factors.
6. Conclusion and future work

Fig. 7. DeepUAge - Real Age vs Predicted Age.

It is true that the validation of data, the reduction of unnecessary
features, and the implementation of structured design reduced the
problem to a smaller one. Underage facial age estimation is a
challenging problem that can be assisted by the use of pre-trained
models. Such models hinge on accurate age labels. The performance of these models can be improved further by; 1) obtaining a

S8

F. Anda et al. / Forensic Science International: Digital Investigation 32 (2020) 300921

large dataset with balanced, validated class labelling, and 2)
ensuring optimal pre-processing of the data, e.g., subject alignment
and the elimination of unnecessary noise, such as photo backgrounds or hair.
Given that expert human age estimation of childrens’ faces often
achieves poor accuracy (Ferguson and Wilkinson, 2017), the model
proposed with a predictable error rate makes DeepUAge potentially
a valuable aid for law enforcement. Finally, it can be concluded that
ensemble pre-trained deep learning methods can have a positive
impact on CSEM investigation, e.g., as a triage technique or evidence analysis prioritisation tool.
6.1. Future work
As future work, the accuracy of our current age estimation algorithm can be improved, i.e., decreasing both the MAE and logarithmic loss. The main goal of this work is to aid law enforcement in the
detection and investigation of CSEM. From a victim identification
standpoint, we would like to analyse other components that are
present in a digital forensic CSEM crime scene including garments,
visual geolocation clues, object detection, etc. We also plan to make
the VisAGe dataset available and encourage others to contribute, e.g.,
to improve the demographic balance that is currently lacking. Lastly,
a safe framework will be created to interact with LEA and evaluate
the accuracy of our approach with real cases.
Acknowledgements
The authors wish to acknowledge the contribution of colleagues
in the UCD Forensics and Security Research Group and in the MITRE
Corporation in the curation of the VisAGe dataset. This work has
been supported in part by the Google Cloud Platform Research
Credits Program.
References
Anda, F., Lillis, D., Le-Khac, N.A., Scanlon, M., 2018. Evaluating automated facial age
estimation techniques for digital forensics. In: 2018 IEEE Security and Privacy
Workshops (SPW). IEEE, pp. 129e139.
Anda, F., Lillis, D., Kanta, A., Becker, B.A., Bou-Harb, E., Le-Khac, N.A., Scanlon, M.,
2019. Improving borderline adulthood facial age estimation through ensemble
learning. In: Proceedings of the 14th International Conference on Availability,
Reliability and Security. ARES ’19. ACM, New York, NY, USA, ISBN 978-1-45037164-3. https://doi.org/10.1145/3339252.3341491, 57:1e57:8. https://search.
crossref.org/?q1⁄4Improvingþborderlineþadulthoodþfacialþageþestimationþ
throughþensembleþlearning.
Antipov, G., Baccouche, M., Berrani, S., Dugelay, J., 2016. Apparent age estimation
from face images combining general and children-specialized deep learning
models. In: 2016 IEEE Conference on Computer Vision and Pattern Recognition
Workshops (CVPRW), pp. 801e809. https://doi.org/10.1109/CVPRW.2016.105.
Bottou, L., 2010. Large-scale machine learning with stochastic gradient descent. In:
Lechevallier, Y., Saporta, G. (Eds.), Proceedings of COMPSTAT’2010. PhysicaVerlag HD, Heidelberg, ISBN 978-3-7908-2604-3, pp. 177e186.
n-Santana, M., Lorenzo Navarro, J.J., Obrego
n, Freire, Boys2Men, C., 2016. An
Castrillo
age estimation dataset with applications to detect infants in pornography
content. In: First International Workshop on Biometrics and Image Forensics.
IEEE.
Dalrymple, K.A., Gomez, J., Duchaine, B., 2013. The dartmouth database of children's
faces: acquisition and validation of a new face stimulus set. PloS One 8 (11),
e79131.
Deb, D., Nain, N., Jain, A.K., 2018. Longitudinal study of child face recognition. In:
2018 International Conference on Biometrics (ICB). IEEE, pp. 225e232.
Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, Imagenet, L., 2009. A large-scale
hierarchical image database. In: 2009 IEEE Conference on Computer Vision and
Pattern Recognition. IEEE, pp. 248e255.
Du, X., Scanlon, M., 2019. Methodology for the automated metadata-based classification of incriminating digital forensic artefacts. In: Proceedings of the 14th
International Conference on Availability, Reliability and Security. ARES ’19. ACM,

New York, NY, USA, ISBN 978-1-4503-7164-3. https://doi.org/10.1145/
3339252.3340517,
43:1e43:8.
https://search.crossref.org/?
q1⁄4Methodologyþforþtheþautomatedþmetadatabasedþclassificationþofþincriminatingþdigitalþforensicþartefacts.
Eidinger, E., Enbar, R., Hassner, T., 2014. Age and gender estimation of unfiltered
faces. IEEE Trans. Inf. Forensics Secur. 9 (12), 2170e2179. https://doi.org/
10.1109/TIFS.2014.2359646.
Ferguson, E., Wilkinson, C., 2017. Juvenile age estimation from facial images. Sci.
Justice 57 (1), 58e62. https://doi.org/10.1016/j.scijus.2016.08.005. http://www.
sciencedirect.com/science/article/pii/S1355030616300739.
Founds, A.P., Orlans, N., Genevieve, W., Watson, C.I., 2011. .. In: NIST Special Database 32-Multiple Encounter Dataset II (MEDS-II). Tech. Rep. NIST.
Fussell, M., 2019. Facial Proportions - How to Draw a Face. https://
thevirtualinstructor.com/facialproportions.html. (Accessed 4 October 2019).
Gower, J.C., 1975. Generalized procrustes analysis. Psychometrika 40 (1), 33e51.
Guo, G., Mu, G., Fu, Y., Dyer, C., Huang, T., 2009. A study on automatic age estimation
using a large database. In: 2009 IEEE 12th International Conference on Computer Vision, pp. 1986e1991. https://doi.org/10.1109/ICCV.2009.5459438.
Han, H., Otto, C., Liu, X., Jain, A.K., 2015. Demographic estimation from face images:
human vs. Machine performance. IEEE Trans. Pattern Anal. Mach. Intell. 37 (6),
1148e1161. https://doi.org/10.1109/TPAMI.2014.2362759.
King, D.E., 2009. Dlib-ml: a machine learning toolkit. J. Mach. Learn. Res. 10,
1755e1758.
Lanitis, A., Cootes, T., 2002. FG-NET aging data base. Cyprus College 2 (3), 5.
Lillis, D., Becker, B., O'Sullivan, T., Scanlon, M., 2016. Current challenges and future
research areas for digital forensic investigation. In: ). The 11th ADFSL Conference on Digital Forensics, Security and Law (CDFSL 2016, vols. 9e20. ADFSL,
Daytona Beach, FL, USA.
Liu, X., Li, S., Kan, M., Zhang, J., Wu, S., Liu, W., Han, H., Shan, S., Chen, X., 2015.
AgeNet: deeply learned regressor and classifier for robust apparent age estimation. In: 2015 IEEE International Conference on Computer Vision Workshop
(ICCVW), pp. 258e266. https://doi.org/10.1109/ICCVW.2015.42.
Loomis, A., 2017. Figure Drawing for All It's Worth. Editoria Bibliomundi Serviçis
Digitais LTDA.
Nawara, J., 2010. Machine learning: face recognition technology evidence in criminal trials. U Louisville L Rev 49, 601.
Nutter, P.W., 2018. Machine learning evidence: admissibility and weight. U Pa J
Const L 21, 919.
Phillips, P.J., Wechsler, H., Huang, J., Rauss, P.J., 1998. The FERET database and
evaluation procedure for face-recognition algorithms. Image Vis Comput. 16 (5),
295e306.
Ramanathan, N., Chellappa, R., 2006. Modeling age progression in young faces. In:
2006 IEEE Computer Society Conference on Computer Vision and Pattern
Recognition (CVPR’06), vol 1, pp. 387e394. https://doi.org/10.1109/
CVPR.2006.187.
Reisfeld, D., Yeshurun, Y., 1998. Preprocessing of face images: detection of features
and pose normalization. Computer Cision and Image Understanding 71 (3),
413e430.
Ricanek, K., Bhardwaj, S., Sodomsky, M., 2015. A review of face recognition against
€ mme, A., Busch, C., Rathgeb, C., Uhl, A. (Eds.),
longitudinal child faces. In: Bro
Proceedings of the 14th International Conference of the Biometrics Special Interest Group (BIOSIG). Gesellschaft für Informatik e.V., Bonn, pp. 15e26.
Rothe, R., Timofte, R., Van Gool, L., 2018. Deep expectation of real and apparent age
from a single image without facial landmarks. Int. J. Comput. Vis. 126 (2),
144e157. https://doi.org/10.1007/s11263-016-0940-3.
Sae-Bae, N., Sun, X., Sencar, H.T., Memon, N.D., 2014. Towards automatic detection
of child pornography. In: 2014 IEEE International Conference on Image Processing (ICIP), pp. 5332e5336. https://doi.org/10.1109/ICIP.2014.7026079.
Sanchez, L., Grajeda, C., Baggili, I., Hall, C., 2019. A practitioner survey exploring the
value of forensic tools, AI, filtering, & safer presentation for investigating child
sexual abuse material (CSAM). Digit. Invest. 29, S124eS142. https://doi.org/
10.1016/j.diin.2019.04.005.
http://www.sciencedirect.com/science/article/pii/
S1742287619301549.
Scanlon, M., 2016. Battling the digital forensic backlog. In: Proceedings of the 2nd
International Workshop on Cloud Security and Forensics (WCSF 2016), vols
10e14. IEEE, Dublin, Ireland.
Signorielli, N., 1987. Children and adolescents on television: a consistent pattern of
devaluation. J. Early Adolesc. 7 (3), 255e268.
Wallhoff, F., 2006. Facial Expressions and Emotions Database. http://www-prima.
inrialpes.fr/FGnet/html/home.html.
Wong, S.C., Gatt, A., Stamatescu, V., McDonnell, M.D., 2016. Understanding data
augmentation for classification: when to warp?. In: 2016 International Conference on Digital Image Computing: Techniques and Applications (DICTA),
pp. 1e6. https://doi.org/10.1109/DICTA.2016.7797091.
Zhang, Z., Song, Y., Qi, H., 2017. Age progression/regression by conditional adversarial autoencoder. In: 2017 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), pp. 4352e4360. https://doi.org/10.1109/CVPR.2017.463.
Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.. Random erasing data augmentation.
CoRR 2017;abs/1708.04896. http://arxiv.org/abs/1708.04896.