(Available on SSRN.

SSRN is an open-access online preprint

community, owned by Elsevier.)

International Conference on Contents, Computing &

Communication (ICCCC-2022)

26th and 27th February,2022

Human Face Aging based on Deep Learning: A survey

Ms. T. S. Deshmukh 1, Dr. M. D. Kokate2
1Research scholar, Matoshri COE & Research Centre, Eklhare, Nashik, 422105 India
2Principal SNJB’s K. B. Jain College of Engineering, Chandwad, 423101 India.

Abstract

Face Age progression enables predicting the future appearance of a person. Facial aging affects human face in diverse forms f rom
early stages to old age. Over few decades, researchers are working on development of various face progression techniques to meet
the challenge of creating accurate aged face for smart systems applications. Here we present analysis of face age progression
methods for child and adult face. Deep Learning has found vast applications in the fields of Computer vision. With the advent of deep
generative networks success rate has increased in terms of aging accuracy and identity preservation. Here some databases regarding
face aging are described in terms of the number of persons along with their age ranges. The goal of this survey is to know ongoing
efforts used in age progression.

Keyword : Face age progression (FAP); Deep learning.

1. Introduction

Almost all countries are facing severe problem of human trafficking. Human trafficking victims are men,
women, as well as children. According to National Crime Records Bureau around 3 to 4 lakhs of persons go
missing every year [1]. Facial aging is a most important technology for finding out missing person, as parents or
relatives are can have at least one photograph of missing person. Other applications where we can use face aging
in human-computer interaction, entertainment, animation.

Every Indian must have a Aadhar card. No biometrics are captured of children below five years. On the basis of
age information and facial photograph Children UID is processed and are linked with the UID of their parents.
When they turn 5 and 15, they need to update their biometrics of fingers, iris and facial photograph. When they
grow up intimation to this effect will be mentioned in the Aadhaar letter. Later on, once the child is grown into
adult the photo in database is not update. With the help of the generated appearances the database can be
updated.

Various progressive changes are observed in a human face, which include skin texture, weight, facial hair, etc.
There is change in face and neck appearance with age. Face becomes flabby or we find dropping appearance due
to deprivation of muscle tone and derma thinning. Face has no desirable smooth surface as dermis dries out and
the underlying layer of fat reduce. Wrinkles cannot be avoided. They are likely to grow more quickly due to sun
exposure and cigarette smoking. Dark spots increase on the face in terms of numbers and size. Due to sun
exposure these pigment changes are largely.

Appearance of mouth may change due to missing teeth and lips may get smaller. Size of the lower face may
lessen out due to loss of bone mass in the jaw and makes forehead, nose, and mouth more noticeable. Nose may
also elongate to a slight extent.

Probably due to growth in cartilage the ears may look longer in some people. Greying of Eyebrows and
eyelashes. Wrinkles would be seen in skin around the eyes. Eyes may look sunken because of fat from the
eyelids settling into the cranial-orbit. The lower eyelids can slacken. The upper eyelid droop due to weakening
in muscle. A greyish-white ring may develop at the outer layer of the eye. The colouring of the iris is
reduced(pigmentation), hence very aged people eyes seem to be look like grey or light blue.

International Conference on Contents, Computing & Communication (ICCCC-2022)

Electronic copy available at: https://ssrn.com/abstract=4043509

 (Available on SSRN. SSRN is an open-access online preprint
community, owned by Elsevier.)

International Conference on Contents, Computing &

Communication (ICCCC-2022)

26th and 27th February,2022

2. Related Work

Human face ageing is due to different factors such as genetic, environment and behavioural factors. Hence while
predicting appearance, we have to take account on individual ageing rates instead of fixed once’s, which is a big
challenge.

In early days face age progression was divided into physical model based and prototype model based. Models
which are based on physical model are built on the biological model and human aging mechanism of aging such
as the muscle [2], wrinkles on face [3,4], facial shape [5,6]. Disadvantage of physical model based is that they
are computational expensive. Prototype based approach divides face into groups from which average faces
represent age pattern [7,8]. Disadvantage of the prototype approach is cannot preserve identity. Face age
progression methods based on deep learning has gained more attention as they perform well in terms of identity
retention and aging accuracy.

Conditional Adversarial Autoencoder (CAAE) [9] was delineated for face progression task to achieve face aging
and rejuvenation, by using single sample. In CAAE, the face is charted to a latent vector, and then the values are
imposed to the face manifold constraints on age. Personalized face features are preserved by the latent vector
preserves (i.e., personality) and progression is controlled by the age condition. Using CAAE only rough
wrinkles were generated. To overcome this problem, J. Zeng et.al. [10] created a generative model based on
CAAE with some development: 1) A component that seek to improve the convergence of very deep networks
called as an auxiliary classifier was additional inserted on the discriminator, and also categorizes them into the
target age group; and 2) a pre-trained deep face recognition model is used to preserve identity and pre-trained
age estimation model is used to preserve age similarity age. Wang et.al [11] introduced Identity Preservation
generative adversarial networks which collectively had both, an identity preservation and an age estimator.
IPCGANs can be put into different feature creation task, like hair colours, various facial expressions, no beard-
to-beard etc without making any changes in IPCGAN. If condition is removed from the framework of IPCGAN,
it can be used for image translation task.

An attention-based generative adversarial networks framework by Zhu et al. [12], who stated that without using
a pixel-wise loss, distorting or ghosting effects can be reduced. To avoid these artefacts generator is trained to
output an attention mask and a colour mask. The attention mask studies to spot the image zones related to the
age synthesis; the colour mask studies how to adjust those areas. Hence background area and the unique identity
of a person over time are well conserved. to produce enhanced vision of face image, Neha Sharma et.al. [13,14]
used a fusion-based generative adversarial network approach. Cycle-GAN achieved the face aging, then
Enhanced Super-resolution Generative Adversarial Network (ESRGAN) enhanced the aged look to improve the
clarity of image.

As face is the utmost affected part of aging of looks there is a requirement to extract robust face feature based on
this theory Boussaad et.al.[15] examined usefulness of deep-learning based methods as features extraction tool.
AlexNet, Inception V3, GoogleNet, SqueezeNet and ResNet50 these popular pre-trained deep-convolutional
neural network models’ evaluation was carried out. Experiment was conducted on a commonly used face-aging
database FG-NET. AlexNet seemed to be the finest to handle differences in age and the utmost capable for age
invariant face recognition. Author in forthcoming research wants to test these pre-trained models on additional
databases and combine them with other pre-processing approaches. A potent deep-featured encoding-based
discriminative model was introduced by Shakeel et. al. [16] for age-invariant face recognition. To learn
important deep features a pre-trained deep-CNN AlexNet model is being used, these values are encoded by
learning a code book, and each feature is being converted into S-dimensional codeword for picture
representation. Author has proposed feature-encoding structure based on locality constraint. Shakeel et. al. [16]
wants to design a new convolutional layer to achieve optimization of feature representation and feature fusion at

International Conference on Contents, Computing & Communication (ICCCC-2022)

Electronic copy available at: https://ssrn.com/abstract=4043509

 (Available on SSRN. SSRN is an open-access online preprint
community, owned by Elsevier.)

International Conference on Contents, Computing &

Communication (ICCCC-2022)

26th and 27th February,2022

same time. Huang et. al [17] presented a recursive generative adversarial network-based face age progression
architecture which is trained concurrently on the whole ageing span to reduce the generated errors as the authors
is of the opinion that training each of the face age progression structure units separately causes artefacts for age
progressions over long time spans since faults gather when being passed through the network chain.

With varying lifestyle human appearance will change a lot. As said face elderly look is due to diverse factors
such as genetic, environment and behavioural factors. Anwaar et.al [18] introduced image synthesis using age
and weight. For weight progressed face Constrained Local Model is used and then Conditional Adversarial Auto
Encoder is used to generate age-progressed face image. Face morphing of weight progressed facial image and
age-progressed facial images is carried out which creates future face image and keep natural looks.

Human face ageing takes place in two ways, from child to adult and then ageing process in adult is subjected to
change by texture. For face age progression and regression on children appearance, Chandaliya et.al
[19]. proposed an architecture based on generative adversarial network and Variational Autoencoder with
Perceptual loss i.e., like the same photo but shifted by one pixel called as Conditional Perceptual Adversarial
Variational Autoencoder. CPAVAE framework is robust against variation like illumination and pose. But
authors planned to take together age and gender loss to preserve identity. In [20], child face aging and
regression, gender is used as a constrain. Authors used Conditional Generative Adversarial Network (cGANs), a
convolutional neural network that is 19 layers deep (VGG19) based perceptual loss and LightCNN29 age
classifier to produce good results.

Table: Summary of FAP

Author Year Database Note Different loss & evaluation tool

Mapping of face to a latent vector.

Tool: Qualitative and
Zhifei Zhang Then vector is extrapolated to the
2017 MORPH, CACD quantitative (Survey form)
et.al [9] face manifold conditional on age.
comparisons.
No paired samples
A supplementary pretrained Losses: Image Adversarial loss,
classifier is added on the Age classification loss, Age
J. Zeng et.al discriminator. To preserve identity preserving loss, Identity
2018 UTKFace , FGNET
[10] and age similarity, deep face preservation loss.
recognition model and a pre-trained
age estimation model are used
Inception score metric used for
evaluating the quality of
Identity and age similarity is generated images.
Wang et.al [11] 2018 CACD conserved by using an age User Study evaluation.
estimation model. Losses included: adversarial
loss, perceptual loss, Age
classification loss.
Tool:To estimate the aging
accuracy Face++API.
to approximate the age
Pixel-wise loss is not used still distributions, i.e., mean value of
Zhu et al. [12] 2020 MORPH obscuring or ghosting effects can both real and generated faces in
be reduced. each age group,
Loss function used: Adversarial
loss, Attention loss, Age
classification loss.

International Conference on Contents, Computing & Communication (ICCCC-2022)

Electronic copy available at: https://ssrn.com/abstract=4043509

 (Available on SSRN. SSRN is an open-access online preprint
community, owned by Elsevier.)

International Conference on Contents, Computing &

Communication (ICCCC-2022)

26th and 27th February,2022

Tool: Face ++ toolkit,

parameters confidence score
IMDB-WIKI, number and age estimation
Neha Sharma A combination-based Generative
2020 CACD, UTKFace, value.
et.al. [13,14] Adversarial Network used.
FGNET,Celeb A Loss function: cycle consistency
loss, Identity loss, GAN loss,
perpetual loss ,content loss.
Tools: Pretrained CNN models
To extract robust face feature using
Boussaad et.al used Alexnet, GoogleNet,
2020 FGNET pretrained pre-trained deep-
[15] Inception V3, Squeeze Net,
convolutional neural network mode
ResNet
Tool: Face ++ is used for face
aging accuracy and identity
preservation. Inception score is
Face image is produced by
Huang et.al used to assess quality and variety
2020 MORPH, CACD maintaining image quality, accurate
[17] in created image.
aging, and identity preservation.
Loss function: adversarial
loss,age estimation loss, Identity
Consistency loss.
Anwaar et.al MORPH, CACD, Future face image produced using Tools: Structural Similarity
2020 Index (SSIM)
[18] UTK age and weight.
Tools: Inception Score , Frechet
Generative adversarial network and Inception score, Eucledian
The children Variational adversarial network Distance and Cosine similarity
Chandaliya
2019 longitudinal Face with Perceptual loss (CPAVAE), a Score
et.al.[19]
(private). model for face progression and Loss function:Perceptual loss ,
regression on little one face. adversarial loss ,Kullback-
Leibler divergence loss
The children Child face progression and VGG19 based perceptual loss
Chandaliya
2020 longitudinal Face regression. Using gender as
et.al [20]
(private). constrain.
3.Data Set

Performance of deep neural network depends on number of existing training information hence this segment
introduces datasets often used in face age progression works. Most frequently chosen data set by researchers is
(CACD) Cross Age celebrity dataset [21], and Academic MORPH [22] database. There are 160,000 plus images
of 2,000 celebrities in the CACD dataset with age varying from 16 to 62, only associated with estimated age
labels. The images in MORPH-II are taken in supervised scenario with neutral expression, uniform light, simple
backgrounds and correct age labels.

Other datasets used in FAP are FGNET, which is a publicly available dataset also used for age progression.
FGNET contains 1002 images of age ranging from 0 to 69. Images in FGNET are captured in uncontrolled
situations. UTKFace database contains more than 20,000 images from age varying from 0 to 116 with labels of
age, gender and ethnicity. It is subgroup of MORPH, CACD dataset and images taken from Google and Bing.
The 524,230 face images were selected from IMDB and Wikipedia websites to form new IMDB-WIKI dataset
[23]. CelebA dataset contains around 203 thousand [24] images with web crawled collection of celebrities with
high level of difference of poses, light, expression and accessories.

LFW [25] datasets hold 13233 Metadata images web crawled with low resolution of 250*250. Adience datasets
contains around 27,000 images with Age span from 0 to 60 plus and age and gender as labels. Images are web
crawled from Flickr. FFHQ datasets contains 70,000 images with high quality Flickr images and high variation
in ethnicity. Age span ranges from 0 to 70 plus. FFHQ Aging database is an extension of FFHQ datasets where
more labels are added to images like age, gender, pose, glasses, eye occlusion.
International Conference on Contents, Computing & Communication (ICCCC-2022)

Electronic copy available at: https://ssrn.com/abstract=4043509

 (Available on SSRN. SSRN is an open-access online preprint
community, owned by Elsevier.)

International Conference on Contents, Computing &

Communication (ICCCC-2022)

26th and 27th February,2022

ITWCC [26] datasets contain around 8000 images from age of 0 to 32 with age and gender as labels. These are
web crawled images of child celebrities with high differences of poses, light and facial expression. AGFW
dataset contains 19k of images with an age span of from 10 to 64 with age as labels. These images are web
crawled and gathered from “The Productive Aging Laboratory [28]”. AgeDB database [29] contains images of
famous persons with manually labelled ages. It contains 17k of images with age span of 9 – 95 years. The
Children Longitudinal Face [30] is a private dataset in which age ranges from 2 to 18 years. CLF dataset consist
of 3682 Images. A resolution of 354×472 pixels for the captured face images. Challenging acquisitions with
pose, light and expression, some images are also captured with occlusion like headscarves, cap, dressing on
face, whiskers, specs, birth marks such as moles, cuts, different eye colour, and blemish in CLF dataset images.

4. Future Work

There are many challenges to be resolved by upcoming works. Literature survey shows that most FAP
contributions is done on adult face ageing while less work is found on child face ageing. It is found that facial
changes in adults are mostly texture-based, the changes in the craniofacial are much more difficult to produce
while child is growing up, hence this is wide area for future work. It is seen that most research focus on
progression very less is mentioned about rejuvenation, which also offers good area to research.

One major challenge is database of young children, as there is private dataset available for child Face Age
progression. There is need to collect database for Indian ethnicity for all age ranges, as available datasets are of
mix ethnicity. Human face aging also depends upon external factors as lifestyle, working condition etc. Face
images with such labels should be collected.

While performing FAP this process also include some personalised facial characteristic for a sole person such as
mole, birthmark, scars etc… which do not change with time. The 3D face aging is open problem. Dataset
collection of 3D images, computing cost etc are some of the problems need to be addressed in 3D face aging.

5.Summary

In this paper, investigation of face age progression studies is been carried out. Face age progression is still an
upcoming field for several application. The number of contributions in FAP are based on condition-based
concepts. Challenges for deep face age progression are database. Elderly and children face images should be
increased. Additional information such as profession or nutrition type etc. of the face images helps FAP
methods to be tailored to the specific conditions of an individual.
References
1. National crime record bureau: https://ncrb.gov.in/
2. J. Suo, X. Chen, S. Shan, W. Gao, and Q. Dai. A concatenational graph evolution aging model. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 34(11):2083–2096, 2012.
3. N. Ramanathan and R. Chellappa. Modeling shape and textural variations in aging faces. In IEEE
International Conference on Automatic Face & Gesture Recognition, pages 1–8. IEEE, 2008.
4. J. Suo, S.-C. Zhu, S. Shan, and X. Chen. A compositional and dynamic model for face aging. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 32(3):385–401, 2010.
5. N. Ramanathan and R. Chellappa. Modeling age progression in young faces. In IEEE Computer
Society Conference on Computer Vision and Pattern Recognition, volume 1, pages 387–394. IEEE,
2006
6. A. Lanitis, C. J. Taylor, and T. F. Cootes. Toward automatic simulation of aging effects on face
images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(4):442–455, 2002.
7. D. M. Burt and D. I. Perrett. Perception of age in adult Caucasian male faces: Computer graphic
manipulation of shape and colour information. Proceedings of the Royal Society of London B:
Biological Sciences, 259(1355):137–143, 1995.

International Conference on Contents, Computing & Communication (ICCCC-2022)

Electronic copy available at: https://ssrn.com/abstract=4043509

 (Available on SSRN. SSRN is an open-access online preprint
community, owned by Elsevier.)

International Conference on Contents, Computing &

Communication (ICCCC-2022)

26th and 27th February,2022

8. I. Kemelmacher-Shlizerman, S. Suwajanakorn, and S. M. Seitz. Illumination-aware age progression. In

IEEE Conference on Computer Vision and Pattern Recognition, pages 3334–3341. IEEE, 2014
9. Z. Zhang, Y. Song and H. Qi, "Age Progression/Regression by Conditional Adversarial Autoencoder,"
2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 4352-4360,
doi: 10.1109/CVPR.2017.463.
10. J. Zeng, X. Ma and K. Zhou, "CAAE++: Improved CAAE for Age Progression/Regression," in IEEE
Access, vol. 6, pp. 66715-66722, 2018, doi: 10.1109/ACCESS.2018.2877706.
11. X. Tang, Z. Wang, W. Luo and S. Gao, "Face Aging with Identity-Preserved Conditional Generative
Adversarial Networks," 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition,
2018, pp. 7939-7947, doi: 10.1109/CVPR.2018.00828.
12. H. Zhu, Z. Huang, H. Shan and J. Zhang, "Look Globally, Age Locally: Face Aging With an Attention
Mechanism," ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), 2020, pp. 1963-1967, doi: 10.1109/ICASSP40776.2020.9054553.
13. Sharma, N., Sharma, R. & Jindal, N. “An Improved Technique for Face Age Progression and Enhanced
Super-Resolution with Generative Adversarial Networks”. Wireless Pers Commun 114, 2215–2233
(2020). https://doi.org/10.1007/s11277-020-07473-1.
14. Sharma, N., Sharma, R. & Jindal, N. Prediction of face age progression with generative adversarial
networks. Multimed Tools Appl 80, 33911–33935 (2021). https://doi.org/10.1007/s11042-021-11252-
w.
15. Leila Boussaad, Aldjia Boucetta, “Deep-learning based descriptors in application to aging problem in
face recognition”, Journal of King Saud University - Computer and Information Sciences,2020, ISSN
1319-1578, https://doi.org/10.1016/j.jksuci.2020.10.002.
16. M. Saad Shakeel, Kin-Man Lam, “Deep-feature encoding-based discriminative model for age-invariant
face recognition”, Pattern Recognition, Volume 93, 2019, Pages 442-457, ISSN 0031-3203,
https://doi.org/10.1016/j.patcog.2019.04.028.
17. Z. Huang, S. Chen, J. Zhang and H. Shan, "PFA-GAN: Progressive Face Aging With Generative
Adversarial Network," in IEEE Transactions on Information Forensics and Security, vol. 16, pp. 2031-
2045, 2021, doi: 10.1109/TIFS.2020.3047753.
18. Anwaar, M., Loo, C.K. & Seera, M., “Face image synthesis with weight and age progression using
conditional adversarial autoencoder”. Neural Comput & Applic 32, 3567–3579 (2020).
https://doi.org/10.1007/s00521-019-04217-6.
19. P. K. Chandaliya and N. Nain, "Conditional Perceptual Adversarial Variational Autoencoder for Age
Progression and Regression on Child Face," 2019 International Conference on Biometrics (ICB), 2019,
pp. 1-8, doi: 10.1109/ICB45273.2019.8987410.
20. P. K. Chandaliya, A. Sinha and N. Nain, "ChildFace: Gender Aware Child Face Aging," 2020
International Conference of the Biometrics Special Interest Group (BIOSIG), 2020, pp. 1-5.
21. Chen BC., Chen CS., Hsu W.H. (2014) Cross-Age Reference Coding for Age-Invariant Face
Recognition and Retrieval. In: Fleet D., Pajdla T., Schiele B., Tuytelaars T. (eds) Computer Vision –
ECCV 2014. ECCV 2014. Lecture Notes in Computer Science, vol 8694. Springer, Cham.
https://doi.org/10.1007/978-3-319-10599-4_49.
22. K. Ricanek and T. Tesafaye, "MORPH: a longitudinal image database of normal adult age-
progression," 7th International Conference on Automatic Face and Gesture Recognition (FGR06),
2006, pp. 341-345, doi: 10.1109/FGR.2006.78.
23. R. Rothe, R. Timofte and L. Van Gool, "DEX: Deep EXpectation of Apparent Age from a Single
Image," 2015 IEEE International Conference on Computer Vision Workshop (ICCVW), 2015, pp. 252-
257, doi: 10.1109/ICCVW.2015.41.
24. Z. Liu, P. Luo, X. Wang and X. Tang, "Deep Learning Face Attributes in the Wild," 2015 IEEE
International Conference on Computer Vision (ICCV), 2015, pp. 3730-3738, doi:
10.1109/ICCV.2015.425.
25. Gary B. Huang, Marwan Mattar, Tamara Berg, Eric Learned-Miller. Labeled Faces in the Wild: A
Database forStudying Face Recognition in Unconstrained Environments. Workshop on Faces in ’Real-
Life’ Images: Detection, Alignment, and Recognition, Erik Learned-Miller and Andras Ferencz and

International Conference on Contents, Computing & Communication (ICCCC-2022)

Electronic copy available at: https://ssrn.com/abstract=4043509

 (Available on SSRN. SSRN is an open-access online preprint
community, owned by Elsevier.)

International Conference on Contents, Computing &

Communication (ICCCC-2022)

26th and 27th February,2022

Frédéric Jurie, Oct 2008, Marseille, France. ffinria-00321923.

26. Ricanek, K., Bhardwaj, S. & Sodomsky, M., (2015). A review of face recognition against longitudinal
child faces. In: Brömme, A., Busch, C., Rathgeb, C. & Uhl, A. (Hrsg.), BIOSIG
2015. Bonn: Gesellschaft für Informatik e.V.. (S. 15-26).
27. M. Grimmer, R. Ramachandra and C. Busch, "Deep Face Age Progression: A Survey," in IEEE
Access, vol. 9, pp. 83376-83393, 2021, doi: 10.1109/ACCESS.2021.3085835.
28. Minear, M., Park, D.C. A lifespan database of adult facial stimuli. Behavior Research Methods,
Instruments, & Computers 36, 630–633 (2004). https://doi.org/10.3758/BF03206543
29. S. Moschoglou, A. Papaioannou, C. Sagonas, J. Deng, I. Kotsia and S. Zafeiriou, "AgeDB: The First
Manually Collected, In-the-Wild Age Database," 2017 IEEE Conference on Computer Vision and
Pattern Recognition Workshops (CVPRW), 2017, pp. 1997-2005, doi: 10.1109/CVPRW.2017.250.
30. D. Deb, N. Nain and A. K. Jain, "Longitudinal Study of Child Face Recognition," 2018 International
Conference on Biometrics (ICB), 2018, pp. 225-232, doi: 10.1109/ICB2018.2018.00042.

International Conference on Contents, Computing & Communication (ICCCC-2022)

Electronic copy available at: https://ssrn.com/abstract=4043509