A Data Augmentation Strategy for Improving Age Estimation to Support

CSEM Detection

Deisy Chaves

1 a

, Nancy Agarwal

2 b

, Eduardo Fidalgo

1 c

and Enrique Alegre

1 d

Department of Electrical, Systems and Automation, Universidad de Le

on, Le

on, Spain

GNIOT, Greater Noida, Uttar Pradesh, India

ﬁ

Keywords:

Age Estimation, Data Augmentation, Generative Adversarial Networks, Facial Occlusion, CSEM.

Abstract:

Leveraging image-based age estimation in preventing Child Sexual Exploitation Material (CSEM) content

over the internet is not investigated thoroughly in the research community. While deep learning methods

are considered state-of-the-art for general age estimation, they perform poorly in predicting the age group of

minors and older adults due to the few examples of these age groups in the existing datasets. In this work, we

present a data augmentation strategy to improve the performance of age estimators trained on imbalanced data

based on synthetic image generation and artiﬁcial facial occlusion. Facial occlusion is focused on modelling as

CSEM criminals tend to cover certain parts of the victim, such as the eyes, to hide their identity. The proposed

strategy is evaluated using the Soft Stagewise Regression Network (SSR-Net), a compact size age estimator

and three publicly available datasets composed mainly of non-occluded images. Therefore, we create the

Synthetic Augmented with Occluded Faces (SAOF-15K) dataset to assess the performance of eye and mouth-

occluded images. Results show that our strategy improves the performance of the evaluated age estimator.

1 INTRODUCTION

Facial age estimation is deﬁned as the problem of au-

tomatically predicting the real age of a person (i.e. the

number of years a person has been alive) or the age

group (i.e. pre-pubescent, pubescent, or adult) from

an image. In recent years, image-based age estima-

tion has become an emerging research area due to its

contributions to several real-world applications such

as human-computer interaction (Angulu et al., 2018),

(Osman and Yap, 2018), surveillance monitoring (Fu

et al., 2010), and age-invariant face recognition (An-

gulu et al., 2018). Moreover, some studies showed

the potential of age prediction to detect victims (pre-

pubescent and pubescents) aid the analysis of CSEM

(Anda et al., 2020), (Gangwar et al., 2021), (Grubl

and Lallie, 2022). In (Gangwar et al., 2021), authors

decompose the Child Sexual Abuse (CSA) detection

problem into two modules: pornographic detection

and age-group classiﬁcation as minor or adult.

Deep learning algorithms are state-of-the-art for

designing the age estimators from facial images

(Rothe et al., 2015), (Shen et al., 2018), (Yang et al.,

https://orcid.org/0000-0002-7745-8111

https://orcid.org/0000-0003-4392-0520

https://orcid.org/0000-0003-1202-5232

https://orcid.org/0000-0003-2081-774X

2018), (Agbo-Ajala and Viriri, 2021), (Wang et al.,

2022). However, their performance is hindered be-

cause the existing labelled datasets are limited in size.

Constructing accurate estimators requires many la-

belled face images for every individual age or age

group under consideration. Thus, the imbalance is-

sue in these datasets leads to an overﬁtted model and

prevents its generalisation capability from reaching

its full potential. There are several strategies to han-

dle small and imbalanced datasets, e.g. label distri-

bution learning, data re-sampling, redeﬁning class-

balance losses, and transfer learning (Geng et al.,

2014), (Kang et al., 2019), (Yan et al., 2022). These

approaches mainly focus on designing the model with

an equal probability of learning discriminative fea-

tures of minority and majority class samples.

Instead of working at the algorithm level, an-

other widely adopted strategy to deal with imbalanced

dataset issues is adding more samples of minority

classes to the training (Liu et al., 2020), (Zhong et al.,

2020), (Zhang and Bao, 2022). However, ﬁnding the

expected number of real images labelled with chrono-

logical ages is challenging and time-consuming for

the age estimation problem. Therefore, in this work,

we presented a data augmentation approach to gen-

erate target samples from the existing facial images

artiﬁcially. To the best of our knowledge, Style-

692

Chaves, D., Agarwal, N., Fidalgo, E. and Alegre, E.

A Data Augmentation Strategy for Improving Age Estimation to Support CSEM Detection.

DOI: 10.5220/0011719700003417

In Proceedings of the 18th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2023) - Volume 5: VISAPP, pages

692-699

ISBN: 978-989-758-634-7; ISSN: 2184-4321

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

GAN is a state-of-the-art method for synthesising

high-resolution images (Karras et al., 2019), (Karras

et al., 2020). In this work, StyleGANv2 and High-

Resolution Age Editing Face (HREAF) (Yao et al.,

2021) have been adopted.

Since we are focused on designing an age estima-

tion system that could assist in CSEM investigations,

occlusion plays an essential role as the criminals tend

to cover certain parts of the victim, such as eyes and

mouth, to hide their identity (Gangwar et al., 2017).

However, only a few works have studied the effect of

facial occlusion during age estimation (Yadav et al.,

2014),(Ye et al., 2018), (Cai and Liu, 2021), despite

regions such as eyes and mouth corners are impor-

tant features during the age prediction. In our pre-

vious work , we explored the use of eye occlusion

and improved the estimation of subjects between 0

and 25 years old. In this work, we use artiﬁcial eye

and mouth occluded images with different levels of

transparency for designing an age estimator focused

on minor and adult age groups that are robust to face

occlusion. Moreover, the Soft Stagewise Regression

Network (SSR-Net) (Yang et al., 2018) architecture is

applied to build a compact size age estimator. Fur-

thermore, for evaluation purposes, we generate two

artiﬁcial occluded versions of the APP-Real (Agusts-

son et al., 2017) and the FG-Net datasets (Fu et al.,

2014). Hence, the main contributions of this work are

summarised as follows:

• A data augmentation strategy for age estimation

using facial occlusion and synthetic images aid to

support CSEM detection.

• The Synthetic Augmented with Occluded Faces

(SAOF-15K)

, created for evaluating purposes,

which is composed of the eye and mouth-

occluded facial versions of the FG-Net and the

APP-Real dataset.

• Study of the effect of facial occlusion during age

estimation.

• The outcome of this study will be applied to the

European project Global Response Against Child

Exploitation (GRACE).

2 RELATED WORKS

2.1 Age Estimation Models

Age estimation is mainly addressed as a regression

problem to learn a non-linear function for mapping

https://gvis.unileon.es/dataset/synthetic-augmented-w

ith-occluded-faces-saof/

the facial features of the images to their chronologi-

cal age. However, learning a global ageing mapping

is challenging due to the inhomogeneous nature of

faces observed as intra-class facial appearance vari-

ations (i.e. signiﬁcant changes in the face attributes

of different persons at the same age) (Shen et al.,

2018). As a solution, some studies (Shen et al., 2018),

(Shen et al., 2019) presented modelling of multiple

local regressors to learn the variations in the ageing

pattern, e.g. random forest-based regression (Mon-

tillo and Ling, 2009) and CNN-based tree regression

(Shen et al., 2018). In some studies, the age estima-

tion is treated as a label distribution learning (LDL)

problem by modelling the correlation pattern between

the neighbouring ages based on the assumption that

facial images of a person at adjacent ages appear sim-

ilar (Geng et al., 2014), (Gao et al., 2018). Shen et

al. (Shen et al., 2019) combine the concept of mul-

tiple local functions and LDL to introduce the Deep

Label Distribution Learning Forests (DLDLFs) algo-

rithm for age estimation. The DLDLF model consists

of an ensemble of LDL trees employing the VGG-16

convolutional neural network architecture (Simonyan

and Zisserman, 2014).

Besides, age regression can also be formulated as

a classiﬁcation problem by partitioning the ages into

a set of discrete classes. The authors (Rothe et al.,

2015) leverage the softmax function at the outer layer

of the deep neural network to classify the predicted

age into labels between 0 to 100 years. The ﬁnal age

is predicted by combining the output probabilities of

the neurons via the expected value function. Inspired

by the above work, the authors (Yang et al., 2018)

perform multi-stage classiﬁcation, where each stage

corresponds to a set of age groups and reﬁnes the de-

cision of the previous step with a ﬁner granularity.

The advantage of the multi-stage framework is that

the number of output neurons is small at each stage,

leading to a more compact model than (Rothe et al.,

2015) without sacriﬁcing performance. Recently, in

(Shin et al., 2022) proposed a general regression algo-

rithm called Moving Window Regression (MWR) and

applied it to age estimation. MWR obtains an initial

rank estimate of an input instance based on the near-

est neighbour criterion. It is reﬁned by selecting two

reference instances to form a search window and esti-

mating the relative rank within the search window it-

eratively. The diverse characteristics in different rank

groups are managed using a local and a global regres-

sor.

In our work, we applied a similar stagewise

methodology (Yang et al., 2018), so we built a com-

pact size age estimator useful for machines with lim-

ited memory and computation resources.

A Data Augmentation Strategy for Improving Age Estimation to Support CSEM Detection

693

2.2 Age Estimation with Occluded

Faces

The majority of age estimation works have used fully-

visible facial images as input data for prediction and,

therefore, do not generalise well when certain parts of

the faces are hidden, or occluded (Chaves et al., 2020)

. In forensic applications such as CSEM detection,

age estimation from occluded images becomes crucial

as the criminals tend to cover the eyes of the victim to

hide their identity (Gangwar et al., 2017). Moreover,

the recent global COVID-19 pandemic, where people

should wear a face mask, has further urged the need

to consider occlusion in modelling age estimators.

Face occlusion is commonly studied in domains

like face recognition and face veriﬁcation (Min et al.,

2011), (Kortli et al., 2020), (Zhao et al., 2016). How-

ever, only a few studies have considered occluded

faces during the age estimation (Yadav et al., 2014),

(Cai and Liu, 2021), (Ye et al., 2018), (Chaves et al.,

2020). Yadav et al. (Yadav et al., 2014) conducted an

experiment where some regions of facial images, such

as T-region, binocular region, chin portion, faces with

masked eyes, and masked T-region, were shown to the

participants to understand which facial section con-

tains helpful information for age prediction. It is seen

that the chin area provides sufﬁcient clues for the age

group 0 − 5, and the faces with obfuscated T-region

are recognisable in the age group 6−10. In (Ye et al.,

2018), the authors mask eye regions to build a model

for age estimation with more discriminate features

around the mouth and nose. Furthermore, the work

(Cai and Liu, 2021) focused on occlusion caused by

wearing a mask and employed a self-supervised con-

trastive learning framework to model the relationship

between the fully-visible and masked face.

Our work is similar to (Chaves et al., 2020) where

artiﬁcially eye occluded images are included dur-

ing training to design an age model to support the

recognition of CSEM victims showed that the model

built using both occluded and non-occluded images is

more stable, robust, and efﬁcient. As an extension to

this approach, we considered several levels of trans-

parency during the occlusion and synthesised mouth

occluded images, which may assist the age estimation

of people wearing facial masks.

2.3 Data Augmentation

One of the solutions to deal with the imbalanced data

issue is to increase the number of minority samples in

the dataset. However, ﬁnding the expected number of

images labelled with chronological age is challenging

and time-consuming for the age estimation problem.

Data augmentation overcomes this problem by deﬁn-

ing a variety of simple operations such as rotation,

translation, scaling, ﬂipping, and cropping (Liu et al.,

2020), (Krizhevsky et al., 2012) to artiﬁcially gen-

erate different versions of images from the existing

ones. In addition to geometric image transformations,

synthetic instances can also be obtained by altering

the brightness of the image, infusing some noise to it,

or erasing its certain regions via masking technique

(Liu et al., 2020), (Zhong et al., 2020).

Besides these simple transformation principles,

data augmentation also consists of complex trans-

formations such as the work of Oliveira et al.

(de Pontes Oliveira et al., 2016) where a high-level

deformation function is proposed to induce variance

in speciﬁc facial features (e.g., chin, nose, and jaw)

by detecting ﬁducial points (de Pontes Oliveira et al.,

2016). Nowadays, the use of generative models such

as GAN (Generative Adversarial Network) is one

promising approach in data augmentation that relies

on neural networks to create artiﬁcial data (Wang

et al., 2018). The primary motivation for using these

models in age estimation is to generate a modiﬁed

version of an existing facial image with a younger

or older appearance (Golubovi

c and Risojevi

c, 2021).

In (Georgopoulos et al., 2020), the authors integrated

the style transfer optimising technique with GAN, al-

lowing them to modify the existing image based on

the features of another image. The framework assists

them in creating a variety of faces of extreme ages

(i.e. very old/young).

StyleGAN is nowadays the state-of-the-art

method for synthesising high-resolution images

(Karras et al., 2019), (Karras et al., 2020). Con-

ventional GAN modelling struggles with the feature

entanglement issue, where a small change in latent

input space affects multiple features and their ca-

pability to produce images with ﬁner details. In

contrast, styleGAN allows better control and offers

us to modify the details from high-level features

such as face shape to minute alterations (e.g., hair

colour, wrinkles) in the image by comprising the

methodology that learns progressively from low-level

to high-resolution details. Since GAN-based models

are complex and known for producing artefacts,

the study (Yao et al., 2021) presented the encoder-

decoder-based system to synthesise images for

performing ageing/de-ageing with fewer artefacts

and high resolution, i.e., High-Resolution Face Age

Editing (HRFAE). We employ HRFAE for creating

the older images and styleGANv2 (Karras et al.,

2020) for obtaining the images of the minor group.

styleGANv2 also solves the artefacts problem in

images caused by the styleGAN.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

694

3 METHODOLOGY

We proposed a two-step training strategy to improve

the performance of age estimation models built with

imbalanced data on facial images with and without

occlusion; see Figure 1. First, given a set of non-

occluded facial images, data is augmented by creating

a set of occluded eye and mouth images and adding

masks to simulate the observed conditions on CSEM.

A collection of synthetic images of minors and el-

ders are generated using GAN-based methods to in-

crease the samples with low frequency on the imbal-

ance datasets. Second, artiﬁcial occluded images and

synthetic generated ones are combined into one set

to build an age estimation model robust against facial

occlusion.

Figure 1: Strategy for training age estimation models with

imbalanced datasets.

3.1 Non-Occluded Dataset as Input

We use facial images of the Mivia Age Dataset pro-

vided last year for the Guess The Age contest (Greco,

2021) as input to build the age estimators. This

dataset includes 575.073 images of subjects aged be-

tween 1 and 81 years old. Apart from being one of

the largest publicly available datasets with age anno-

tations, the Mivia Age Dataset is highly imbalanced

on minors and the elderly making it ideal to evaluate

our proposed data augmentation strategy.

3.2 Augmentation of Facial Images

We increase the number of images in the training

dataset using three strategies: (i) creation of facial oc-

cluded images by adding a rectangular mask over the

eye and the mouth regions to simulate what offend-

ers do to difﬁcult victims identiﬁcation, (ii) genera-

tion of synthetic images of minors using transfer style

(StyleGANv2 (Karras et al., 2020)), and (iii) genera-

tion of elder faces using an ageing method (HREAF

(Yao et al., 2021)).

3.2.1 Creation of Facial Occluded Images

Given a non-occluded facial image, ﬁrst, the Multi-

Task Cascade CNN (MTCNN) (Zhang et al., 2016)

method is used to identify the location of the right and

the left eye, and the right and the left external points

of the mouth.

Second, the slope of the lines that connects these

points is obtained and used to determine the posi-

tion and the dimensions of the rectangular mask to

be drawn over the eyes and the mouth, respectively.

The dimensions of the rectangle covering the eyes are

the 25% of the height and the 95% of the width of the

bounding box containing the face. While the dimen-

sions of the rectangle to cover the mouth correspond

to the 25% of the height and the 55% of the width of

the bounding box containing the face.

Finally, several degrees of transparency were con-

sidered to draw the black rectangles, as illustrated in

Figure 2.

Figure 2: Eye and mouth occluded facial images created

artiﬁcially with different levels of transparency.

3.2.2 Generation of Synthetic Minor Images

We use StyleGANv2 to generate minor images be-

tween 1 and 17 years old. StyleGANv2 is an ex-

tension of StyleGAN to reduce the blob-shaped arte-

facts in the generated images. StyleGAN and Style-

GANv2 create images through the incremental ex-

pansion of discriminator and generator models from

small to large images during the training process.

A Data Augmentation Strategy for Improving Age Estimation to Support CSEM Detection

695

To generate a new minor image, ﬁrst, we select

two facial images of minors and generate latent vec-

tors using a StyleGANv2 trained from scratch using

2000 images per age group of the Mivia Age Dataset

– when available. We choose images of minors of the

same age and gender to keep the facial features rep-

resenting a particular age. Second, we combine the

latent vectors of two minor images using linear inter-

polation. Third, we generated the new minor image

from the latent vector using a truncation parameter set

experimentally to 0.7. We generate around 4000 im-

ages per age group; see Figure 3.

Figure 3: Minor facial images generated with StyleGANv2

per age group.

3.2.3 Generation of Synthetic Elder Images

We use HRFAE to generate elder images between 68

and 81 years old because the number of images on the

Mivia Age Dataset for this age group were not enough

to successfully train a StyleGANv2 model. HRFAE is

an encoder-decoder architecture for face age editing

that uses a latent space representation containing the

face identity, a feature modulation layer and its age.

We selected an adult image to generate a new elder

image and aged it using the HRFAE model trained

from scratch. As a result, we generate around 4000

images per age; see Figure 4.

3.3 Building of the Age Estimation

Model

We split the Mivia Age dataset into training and test

subsets, containing the 80% and 20% of all images,

respectively. We augmented the training dataset by

combining the non-occluded faces of the Mivia Age

dataset, their corresponding eye and mouth occluded

version created artiﬁcially, and the generated images

Figure 4: Generation of aged facial images using HRFAE

per age group.

of minors and the elderly using StyleGANv2 and

HRFAE (see section 3.2).

Images on the augmented training dataset are re-

sized to 64 × 64 pixels and used to build SSR-Net

models (Yang et al., 2018) from scratch. We select

a low image resolution (64 × 64 pixels) and the SSR-

Net model for age estimation because it allows us to

construct compact-size models which can be used in

any hardware including mobile devices regardless of

their memory capacity. SSR-Net primarily focuses on

reducing models’ size by classifying a small number

of classes within the age group and reﬁning them at

each stage.

We built the SSR-Net models on a 12GB Nvidia

TITAN Xp using a loss function focus on the Mean

Absolute Error (MAE) value and the Adam optimiser

with a decay learning rate varying between 30 and 60.

Models were trained until a maximum number of 180

epochs using an early stop strategy with a patience

parameter of 10 epochs.

4 EXPERIMENTAL EVALUATION

We evaluated the performance of the SSR-Net models

using the MAE computed on the entire test set, the

MAE values per age intervals obtained considering

eight range intervals between 1 and 81 years, and the

standard deviation (σ) of these MAE per age intervals.

First, we assess the effect of using eye and mouth-

occluded images during the training of SSR-Net mod-

els. For this, we compared the performance of models

built from scratch following the procedure described

in Section 3.3 with several training sets composed of

artiﬁcially occluded facial images using black masks

with several degrees of transparency: (a) original im-

ages (Org. Img.); (b) original images and eye oc-

cluded images (Org. Img & Eye Ocl. Img.) with a de-

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

696

Table 1: Effect of the data augmentation strategy on the performance of the SSR-Net models. The best values are shown in

bold.

Org. Img. & Org. Img. & Org. & Org. &

Evaluation Org. Eye Ocl. Img. Mouth Ocl. Img. Ocl. Ocl. &

Metric Img. 90% 75% 50% 25% 0% 90% 75% 0% Img. GANs Img.

MAE

, 1-10 yrs 18,11 16,28 17,99 17,5 17,26 20,54 17,77 19 17,61 16,28 13,00

MAE

, 11-20 yrs 5,01 4,17 5,04 4,70 4,40 6,73 4,62 5,78 4,75 4,17 4,40

MAE

, 21-30 yrs 2,77 2,74 3,37 2,90 2,56 4,04 2,86 3,44 2,76 2,74 2,90

MAE

, 31-40 yrs 3,47 3,17 3,44 3,24 3,75 4,10 4,05 3,55 3,11 3,17 3,23

MAE

, 41-50 yrs 3,84 3,49 3,64 3,42 5,18 3,98 5,11 3,54 3,36 3,49 3,43

MAE

, 51-60 yrs 3,61 3,26 3,91 3,27 5,55 3,83 5,62 3,55 3,30 3,26 3,13

MAE

, 61-70 yrs 5,36 3,99 5,15 3,68 6,81 3,89 7,68 4,78 4,28 3,99 3,68

MAE

, +70 yrs 9,90 6,92 8,84 5,71 10,65 4,92 12,75 8,26 7,18 6,92 7,12

MAE 3,64 3,30 3,78 3,35 4,30 4,20 4,51 3,77 3,31 3,30 3,31

σ 5,64 4,79 5,38 5,10 5,25 5,86 5,68 5,68 5,28 4,79 3,71

Table 2: Assessment of the age estimator with eye and mouth-occluded images on the Mivia Age, the APP-Real, and the

FG-Net datasets. The best values are shown in bold.

Evaluation org. Mivia Age Dataset APP-Real dataset FG-Net dataset

Metric Img. org. eye ocl. mouth ocl. org. eye ocl. mouth ocl. org. eye ocl. mouth ocl.

MAE

, 1-10 yrs 18,11 13,00 28,18 18,87 31,40 32,10 31,05 15,36 31,74 25,08

MAE

, 11-20 yrs 5,01 4,40 12,64 8,52 17,52 18,75 18,55 9,80 21,41 17,84

MAE

, 21-30 yrs 2,77 2,90 7,89 4,74 9,68 9,84 10,38 6,49 12,63 11,07

MAE

, 31-40 yrs 3,47 3,23 4,39 4,14 5,46 4,44 4,99 5,94 5,35 5,39

MAE

, 41-50 yrs 3,84 3,43 4,03 5,41 6,85 8,41 6,93 6,56 6,45 5,37

MAE

, 51-60 yrs 3,61 3,13 6,84 6,14 11,90 16,01 12,63 7,79 9,57 9,07

MAE

, 61-70 yrs 5,36 3,68 11,69 7,53 18,48 24,92 20,85 11,57 19,43 14,43

MAE

, +70 yrs 9,90 7,12 18,41 13,19 26,42 34,58 28,95 — — —-

MAE 3,64 3,31 6,46 5,34 12,70 13,58 13,06 11,01 22,08 18,03

σ 5,64 3,71 9,30 5,72 9,29 11,56 9,85 3,48 10,43 8,00

gree of transparency of 90%, 75%, 50%, 25% and 0%

(solid mask); (c) original images and mouth occluded

images (Org. Img & Mouth Ocl. Img.) with a degree

of transparency of 90%, 75%, and 0% (solid mask);

(d) original images with eye and mouth occluded im-

ages (Org. & Ocl. Img.); (e) original images with

eye and mouth occluded images and synthetic images

generated using GANs models (Org. & Ocl. & GANs

Img).

Table 1 condenses the general MAE, the MAE

per age interval, and the MAE standard deviation ob-

tained for the ﬁve evaluated conﬁgurations. The data

augmentation using eye and mouth-occluded images

effectively improves the performance during the age

estimation of minors and the elderly (MAE of 3, 30

and standard deviation, σ, of 4, 79). Moreover, the use

of synthetic images allows the building of more stable

models with the lowest MAE standard deviation (σ of

3, 71) in comparison to the model trained only with

original images (σ of 5, 64) and occluded ones (σ of

4, 79). Thus, the best SSR-Net model for estimating

the age is built using the combination of original, oc-

cluded and synthetic images; it achieves an MAE of

3, 31 and σ of 3, 71 on the Mivia Age Dataset.

Besides, we assess the performance of the best

age estimation model on occluded conditions using

the testing set of the Mivia Age dataset (20 % of the

data) and the reference datasets, APP-Real (Agusts-

son et al., 2017), and FG-Net (Fu et al., 2014). FG-

Net is composed of 1002 facial images of subjects be-

tween 0-66 years, and APP-Real comprises 7591 fa-

cial images of subjects between 0-95 years. Note that

after a manual inspection of the APP-Real dataset,

we use for evaluation 6884 images from this dataset.

Moreover, we created the SAOF-15K dataset, which

comprises 15772 images corresponding to two ver-

sions of the APP-Real and the FG-Net datasets (arti-

ﬁcially occluded eyes and mouth).

Table 2 shows the general MAE, the MAE per age

interval, and the MAE standard deviation. Although

eye and mouth occluded images were used to train

the age estimation models, results showed that the

performance decreased on this type of images, espe-

cially on eye occluded ones (MAE of 6, 46), in com-

parison to the prediction on non-occluded or origi-

nal images (MAE of 3, 31). This indicates that the

information in the eye region is more important for

predicting age than the information in the mouth re-

A Data Augmentation Strategy for Improving Age Estimation to Support CSEM Detection

697

gion. The same behaviour was observed on the eval-

uated datasets. Therefore, more robust strategies are

required to improve age prediction on occluded faces.

5 CONCLUSIONS

In this study, we presented a data augmentation strat-

egy to improve the estimation of age on imbalanced

datasets and support the detection of CSEM. The pro-

posal is based on the facial occlusion usually found

on CSEM and synthetic image generation to increase

the number of faces with few samples in the datasets.

We evaluated SSR-Net models built using the

Mivia Age and the SAOF-15K datasets. We create

the SAOF-15K dataset to assess the age estimation

with occluded faces from the APP-Real and the FG-

Net datasets by occluding the eye, and the mouth with

black masks since these datasets contain mainly non-

occluded faces. Results show that the best SSR-Net

model for estimating the image is built using the com-

bination of original, occluded and synthetic images.

Moreover, the data augmentation strategy allows for

improving the performance of non-occluded images.

However, it is not robust enough to accurately predict

age on the eye and mouth-occluded images. In future

work, we will evaluate more robust architectures as

the backbone for age estimation and an ensemble of

classiﬁers trained only with non-occluded or occluded

facial images. Also, we will assess in-painting tech-

niques to reconstruct the images with occluded faces.

ACKNOWLEDGEMENTS

This research has been funded with support from the

European Union’s Horizon 2020 Research and In-

novation Framework Programme, H2020 SU-FCT-

2019, under the GRACE project with Grant Agree-

ment 883341. This publication reﬂects the views

only of the authors, and the European Union’s Hori-

zon 2020 Research and Innovation Framework Pro-

gramme, H2020 SU-FCT-2019, cannot be held re-

sponsible for any use which may be made of the in-

formation contained therein.

REFERENCES

Agbo-Ajala, O. and Viriri, S. (2021). Deep learning

approach for facial age classiﬁcation: a survey of

the state-of-the-art. Artiﬁcial Intelligence Review,

54(1):179–213.

Agustsson, E., Timofte, R., Escalera, S., Bar

o, X., Guyon,

I., and Rothe, R. (2017). Apparent and real age esti-

mation in still images with deep residual regressors on

APPA-REAL database. In FG 2017 - 12th IEEE Inter-

national Conference on Automatic Face and Gesture

Recognition, pages 1–12.

Anda, F., Le-Khac, N.-A., and Scanlon, M. (2020). Deep-

uage: improving underage age estimation accuracy

to aid csem investigation. Forensic Science Interna-

tional: Digital Investigation, 32:300921.

Angulu, R., Tapamo, J. R., and Adewumi, A. O. (2018).

Age estimation via face images: a survey. EURASIP

Journal on Image and Video Processing, 2018(1):1–

35.

Cai, W. and Liu, H. (2021). Occlusion contrasts for self-

supervised facial age estimation. In Multimedia Un-

derstanding with Less Labeling on Multimedia Under-

standing with Less Labeling, pages 1–7. Association

for Computing Machinery.

Chaves, D., Fidalgo, E., Alegre, E., J

anez-Martino, F., and

Biswas, R. (2020). Improving age estimation in mi-

nors and young adults with occluded faces to ﬁght

against child sexual exploitation. In VISIGRAPP (5:

VISAPP), pages 721–729.

de Pontes Oliveira,

I., Medeiros, J. L. P., de Sousa, V. F.,

unior, A. G. T., Pereira, E. T., and Gomes, H. M.

(2016). A data augmentation methodology to improve

age estimation using convolutional neural networks.

In 2016 29th SIBGRAPI Conference on Graphics, Pat-

terns and Images (SIBGRAPI), pages 88–95. IEEE.

Fu, Y., Guo, G., and Huang, T. S. (2010). Age synthe-

sis and estimation via faces: A survey. IEEE trans-

actions on pattern analysis and machine intelligence,

32(11):1955–1976.

Fu, Y., Hospedales, T. M., Xiang, T., Yao, Y., and Gong, S.

(2014). Interestingness prediction by robust learning

to rank. In ECCV, pages 488–503.

Gangwar, A., Fidalgo, E., Alegre, E., and Gonz

alez-Castro,

V. (2017). Pornography and child sexual abuse detec-

tion in image and video: A comparative evaluation. In

8th International Conference on Imaging for Crime

Detection and Prevention (ICDP 2017), pages 37–42.

Gangwar, A., Gonz

alez-Castro, V., Alegre, E., and Fidalgo,

E. (2021). Attm-cnn: Attention and metric learning

based cnn for pornography, age and child sexual abuse

(csa) detection in images. Neurocomputing, 445:81–

104.

Gao, B.-B., Zhou, H.-Y., Wu, J., and Geng, X. (2018).

Age estimation using expectation of label distribution

learning. In IJCAI, pages 712–718.

Geng, X., Wang, Q., and Xia, Y. (2014). Facial age estima-

tion by adaptive label distribution learning. In 2014

22nd International Conference on Pattern Recogni-

tion, pages 4465–4470. IEEE.

Georgopoulos, M., Oldﬁeld, J., Nicolaou, M. A., Panagakis,

Y., and Pantic, M. (2020). Enhancing facial data di-

versity with style-based face aging. In Proceedings

of the IEEE/CVF Conference on Computer Vision and

Pattern Recognition Workshops, pages 14–15.

VISAPP 2023 - 18th International Conference on Computer Vision Theory and Applications

698

Golubovi

c, A. and Risojevi

c, V. (2021). Impact of data aug-

mentation on age estimation algorithms. In 2021 20th

International Symposium INFOTEH-JAHORINA (IN-

FOTEH), pages 1–6. IEEE.

Greco, A. (2021). Guess the age 2021: Age estima-

tion from facial images with deep convolutional neu-

ral networks. In Tsapatsoulis, N., Panayides, A.,

Theocharides, T., Lanitis, A., Pattichis, C., and Vento,

M., editors, Computer Analysis of Images and Pat-

terns, pages 265–274.

Grubl, T. and Lallie, H. S. (2022). Applying artiﬁcial in-

telligence for age estimation in digital forensic inves-

tigations. arXiv preprint arXiv:2201.03045.

Kang, B., Xie, S., Rohrbach, M., Yan, Z., Gordo, A., Feng,

J., and Kalantidis, Y. (2019). Decoupling representa-

tion and classiﬁer for long-tailed recognition. arXiv

preprint arXiv:1910.09217.

Karras, T., Laine, S., and Aila, T. (2019). A style-based

generator architecture for generative adversarial net-

works. In Proceedings of the IEEE/CVF conference

on computer vision and pattern recognition, pages

4401–4410.

Karras, T., Laine, S., Aittala, M., Hellsten, J., Lehtinen,

J., and Aila, T. (2020). Analyzing and improving

the image quality of stylegan. In Proceedings of the

IEEE/CVF conference on computer vision and pattern

recognition, pages 8110–8119.

Kortli, Y., Jridi, M., Al Falou, A., and Atri, M. (2020). Face

recognition systems: A survey. Sensors, 20(2):342.

Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Im-

agenet classiﬁcation with deep convolutional neural

networks. Advances in neural information processing

systems, 25.

Liu, X., Zou, Y., Kuang, H., and Ma, X. (2020). Face im-

age age estimation based on data augmentation and

lightweight convolutional neural network. Symmetry,

12(1):146.

Min, R., Hadid, A., and Dugelay, J.-L. (2011). Improv-

ing the recognition of faces occluded by facial acces-

sories. In 2011 IEEE International Conference on

Automatic Face & Gesture Recognition (FG), pages

442–447. IEEE.

Montillo, A. and Ling, H. (2009). Age regression from

faces using random forests. In 2009 16th IEEE In-

ternational Conference on Image Processing (ICIP),

pages 2465–2468. IEEE.

Osman, O. F. and Yap, M. H. (2018). Computational in-

telligence in automatic face age estimation: A survey.

IEEE Transactions on Emerging Topics in Computa-

tional Intelligence, 3(3):271–285.

Rothe, R., Timofte, R., and Van Gool, L. (2015). Dex: Deep

expectation of apparent age from a single image. In

Proceedings of the IEEE international conference on

computer vision workshops, pages 10–15.

Shen, W., Guo, Y., Wang, Y., Zhao, K., Wang, B., and

Yuille, A. (2019). Deep differentiable random forests

for age estimation. IEEE transactions on pattern anal-

ysis and machine intelligence, 43(2):404–419.

Shen, W., Guo, Y., Wang, Y., Zhao, K., Wang, B., and

Yuille, A. L. (2018). Deep regression forests for age

estimation. In Proceedings of the IEEE Conference

on Computer Vision and Pattern Recognition, pages

2304–2313.

Shin, N.-H., Lee, S.-H., and Kim, C.-S. (2022). Moving

window regression: A novel approach to ordinal re-

gression.

Simonyan, K. and Zisserman, A. (2014). Very deep con-

volutional networks for large-scale image recognition.

arXiv preprint arXiv:1409.1556.

Wang, H., Sanchez, V., and Li, C.-T. (2022). Improv-

ing face-based age estimation with attention-based dy-

namic patch fusion. IEEE Transactions on Image Pro-

cessing, 31:1084–1096.

Wang, Z., Tang, X., Luo, W., and Gao, S. (2018). Face ag-

ing with identity-preserved conditional generative ad-

versarial networks. In Proceedings of the IEEE con-

ference on computer vision and pattern recognition,

pages 7939–7947.

Yadav, D., Singh, R., Vatsa, M., and Noore, A. (2014). Rec-

ognizing age-separated face images: Humans and ma-

chines. PloS one, 9(12):e112234.

Yan, C., Meng, L., Li, L., Zhang, J., Wang, Z., Yin, J.,

Zhang, J., Sun, Y., and Zheng, B. (2022). Age-

invariant face recognition by multi-feature fusionand

decomposition with self-attention. ACM Transactions

on Multimedia Computing, Communications, and Ap-

plications (TOMM), 18(1s):1–18.

Yang, T.-Y., Huang, Y.-H., Lin, Y.-Y., Hsiu, P.-C., and

Chuang, Y.-Y. (2018). Ssr-net: A compact soft stage-

wise regression network for age estimation. In Pro-

ceedings of IJCAI-18, pages 1078–1084. IJCAI.

Yao, X., Puy, G., Newson, A., Gousseau, Y., and Hellier,

P. (2021). High resolution face age editing. In 2020

25th International Conference on Pattern Recognition

(ICPR), pages 8624–8631. IEEE.

Ye, L., Li, B., Mohammed, N., Wang, Y., and Liang, J.

(2018). Privacy-preserving age estimation for content

rating. In 2018 IEEE 20th International Workshop

on Multimedia Signal Processing (MMSP), pages 1–

6. IEEE.

Zhang, B. and Bao, Y. (2022). Cross-dataset learning for

age estimation. IEEE Access, 10:24048–24055.

Zhang, K., Zhang, Z., Li, Z., and Qiao, Y. (2016). Joint

face detection and alignment using multitask cascaded

convolutional networks. IEEE Signal Processing Let-

ters, 23(10):1499–1503.

Zhao, Z.-Q., Cheung, Y.-m., Hu, H., and Wu, X. (2016).

Corrupted and occluded face recognition via coop-

erative sparse representation. Pattern Recognition,

56:77–87.

Zhong, Z., Zheng, L., Kang, G., Li, S., and Yang, Y. (2020).

Random erasing data augmentation. In Proceedings

of the AAAI conference on artiﬁcial intelligence, vol-

ume 34, pages 13001–13008.

A Data Augmentation Strategy for Improving Age Estimation to Support CSEM Detection

699