FELLOWSHIP

PROJECT
APPARENT AGE
ESTIMATION

PREPARED BY:
SHIVANI JOSHI
AMEY JEDHE
MADHURA ARWADE
AMAN ARORA
RUDRAKSH BAHRI
 ACKNOWLEDGEMENT
We have taken efforts in this project. However, it would not have been
possible without the kind support and help of many individuals and
organizations. We would like to extend our sincere thanks to all of them.
We are highly indebted to our college authorities and respected teachers
for their guidance and constant supervision as well as for providing
necessary information regarding the project & also for their support in
completing the project.
We would like to express our special gratitude and thanks to university
persons for giving us such attention and time.
Our thanks and appreciations also go to our colleagues in developing the
project and people who have willingly helped us out with their abilities.
 TABLE OF CONTENTS

ABSTRACT
INTRODUCTION
WHAT IS AGE ESTIMATION?
DATASETS AVAILABKE AND USED
DIFFERENCE BETWEEN NEURAL NETWORKS AND
DEEP NEURAL NETWORKS
ARCHITECTURE OF AGE ESTIMATION
PREVIOUS WORKS
A Cascaded Convolutional Neural Network for Age
Estimation of Unconstrained Faces---Jun-Cheng
Chen1_, Amit Kumar1_
Facial Age Estimation with Age Difference---
Zhenzhen Hu, Yonggang Wen
CONCLUSION
REFFERENCES
FUTURE SCOPE
 ABSTRACT

Automatic age estimation has attracted much attention due to its

potential applications. Most of the proposed approaches have
mainly used low-level handcraft features to encode facial age
related visual information and train an age estimation model. We
focus on age classification task in which face image is assigned to
a label that represents an age range. We proposed a deep
learning based framework for age classification task. In our
proposed algorithm, Deep Convolutional Neural Networks (Deep
Conv Nets) are used to extract high-level complex age related
visual features and predict age range of input face image. Due to
lack of age labeled face images, we use the transfer learning
strategy to train the Deep Conv Nets. In addition, to describe the
relationships between labels that compose an ordered sequence,
we define a new loss function in the training process of age
classification task. The experiments are conducted on a widely
used age estimation dataset-Images of Groups of People. The
experimental results demonstrate the excellent performance of
our proposed algorithm against the state-of- the-art methods. Age
Estimation is basically automatic estimation of age, gender and
race of a person from his face image, which has many potential
applications ranging from forensics to social media. Given a face
image, we first extract informative features via an algorithm, and
then employ a hierarchical approach consisting of between-group
classification, and within group regression. Quality assessment is
also developed to identify low-quality face images that are difficult
to obtain reliable estimates. We propose a coarse-to-fine
approach for estimating the apparent age from unconstrained
face images using deep convolutional neural networks (DCNNs).
The proposed method consists of three modules. The first one is
a DCNN- based age group classifier which classifies a given face
image into age groups. The second module is a collection of
DCNN-based regressors which compute the fine-grained age
estimate corresponding in each age class. Finally, any erroneous
age prediction is corrected using an error-correcting mechanism.
Experimental evaluations on three publicly available datasets for
age estimation show that the proposed approach is able to
reliably estimate the age; in addition, the coarse-to-fine strategy
and the error correction module significantly improve the
performance. Age estimation from facial images is an important
problem in computer vision and pattern recognition. Typically the
goal is to predict the chronological age of a person given his or
her face picture. It is seldom to study a related problem, that is,
how old does a person look like from the face photo? It is called
apparent age estimation. A key difference between apparent age
estimation and the traditional age estimation is that the age labels
are annotated by human assessors rather than the real
chronological age. The challenge for apparent age estimation is
that there are not many face images available with annotated age
labels. Further, the annotated age labels for each face photo may
not be consistent among different assessors. We study the
problem of apparent age estimation by addressing the issues
from different aspects, such as how to utilize a large number of
face images without apparent age labels to learn a face
representation using the deep neural networks, how to tune the
deep networks using a limited number of examples with apparent
age labels, and how well the machine learning methods can
perform to estimate apparent ages. The apparent age data is from
the ChaLearn. Age estimation based on the human face remains
a significant problem in computer vision and pattern recognition.
In order to estimate an accurate age or age group of a facial
image, most of the existing algorithms require a huge face data
set attached with age labels. This imposes a constraint on the
utilization of the immensely unlabeled or weakly labelled training
data, e.g. the huge amount of human photos in the social
networks. These images may provide no age label, but it is easily
to derive the age difference for an image pair of the same person.
To improve the age estimation accuracy, we propose a novel
learning scheme to take advantage of these weakly labeled data
via the deep Convolutional Neural Networks (CNNs). Automatic
age and gender classification has become relevant to an
increasing amount of applications, particularly since the rise of
social platforms and social media. Nevertheless, performance of
existing methods on real-world images is still significantly lacking,
especially when compared to the tremendous leaps in
performance recently reported for the related task of face
recognition. In this paper we show that by learning
representations through the use of deep-convolutional neural
networks (CNN), a significant increase in performance can be
obtained on these tasks. To this end, we propose a simple
convolutional net architecture that can be used even when the
amount of learning data is limited.
 INTRODUCTION
Face analysis is an active research topic in computer vision with
applications in surveillance, human-computer interaction, access
control, and security. In this work, we focus on apparent age
estimation. Traditionally, the problem is tackled through pure
classification or regression approaches. In this paper, we present
a cascaded approach which incorporates the advantages of both
classification and regression approaches. Given an input image,
we first apply the age group classification algorithm to obtain a
rough estimate and then perform age group specific regression to
obtain an accurate age estimate.Like other facial analysis
techniques, age estimation is affected by many intrinsic and
extrinsic challenges, such as illumination variation, race,
attributes, etc. One may define the age estimation task as a
process of automatically labeling face images with the exact age,
or the age group (age range) for each individual. It was suggested
to differentiate the problem of age estimation along four concepts:

Actual age: real age of an individual.

Appearance age: age information shown on the visual
appearance.
Apparent age: suggested age by human subjects from the
visual appearance.
Estimated age: recognized age by an algorithm from the visual
appearance.
The proposed cascaded classification and regression approach
for apparent age estimation is based on a deep convolutional
neural network. Our method consists of three main stages:
(1) a single coarse age classifier
(2) multiple age regressors
(3) an error correcting stage to correct the mistakes made by the
age group classifer.

Since the number of samples for apparent age estimation is

limited, we exploit a DCNN model pre trained for large-scale face
identification task and fine tune the model for age group
classification and age regression tasks. The main contribution of
this work is to propose the age error correction module which
mitigates the common disadvantage of coarse-to-fine
approaches. Typically, the errors made at the initial classification
stage cannot be recovered by the regressors at the following
stage. In this work, we set up the baseline algorithm which is
based on the proposed regression algorithm and study how the
coarse-to-fine strategy and the error correction module improve
the prediction performance. Automatic age estimation task aims to
use machine learning algorithms to estimate a person's age
based on features extracted from face image. As an important
attribute of face, age estimation has attracted much attention due
to its potential applications in HumanComputer Interaction(HCI),
electronic customer relationship management, surveillance
monitoring and so on.

As a branch of face recognition, automatic age estimation task

also shares several similar processes such as face detection,
face alignment, feature extraction and estimation with other facial
related tasks. Features extracted from face images representing
age related visual information is a key factor to the performance
of age estimation algorithms. Low-level hand craft features are
usually extracted in these typical algorithms. Given features,
automatic age estimation tasks can usually be categorized to two
classes: age classification task and age regression task. In age
classification task, face image is usually assigned to a class label
that represents an age range and multi-class classification
algorithms are always used. In age regression tasks, algorithms
output an accurate predicted age number. In this paper, we focus
on the classification task that assigns a face image to an age
range label. The Deep Conv Nets in our proposed algorithm is
trained as a multi-class classifier. The public available age
estimation datasets usually do not contain enough labeled images
to train Deep Conv Nets that owns hundreds of millions of
parameters. We take the transfer learning strategy to overcome
this problem. Transfer learning has been widely used in some
image classification tasks and semi-supervised learning tasks. It
has been proved to be very efficient to train models on small
scale dataset. To improve performance in our proposed algorithm,
we define a new loss function in which a distance term is added to
describe the relationships of labels. In addition, in many face
related tasks, combination of results produced by different regions
in face image has been proved to perform much better than single
region or the whole face image. Thus, we train multiple
independent Deep Conv Nets models with patches cropped from
face images and combine the results as the final results to get a
better performance.

Below we describe different steps involved in training our models

in more detail.

Data collection: Data collection plays an important role in

training any deep neural network (DNN). In this paper, we
aim to label data for three separate tasks: age, gender and
emotion recognition. Collecting labeled data for some tasks,
such as real age estimation, is much more challenging
compared to popular classification or detection problems.
This disparity is due to the fact that human error in
 estimating real age is large (sometimes greater than the
computer vision estimations) and one cannot rely on human
annotators to label faces with their corresponding real age.
However, we have collected a large dataset of faces with
their corresponding age, gender and emotion labels. To our
knowledge, our datasets are the largest or among the largest
in either the academic or commercial world.

Below we provide some statistics on the data used for training our
models.

Face recognition: The base model for our facial recognition

is trained on over million images of more than thousands
individuals. The large variation in images of each identity
makes our deep model robust to common challenges in face
recognition.

Age estimation: Recently there have been some efforts in

collecting data with corresponding age labels. Additionally
the distribution of the data across different ages is highly
unbalanced. This led to the authors using only half of the
data for training in the original paper. To better address this
problem we collected a large dataset of large number of
images with corresponding age labels. In contrast to
previous works, our dataset has a more balanced distribution
across different ages.

Gender and emotion recognition: To better improve our

model, we added thousands of images of different ethnicities
as well as age groups. Additionally, we also annotated part
of our data with emotion labels for the task of emotion
recognition.
 Data pre-processing: We pre-process each image before
feeding them to our DNNs. These pre-processing steps
include face detection, facial landmark detection and
alignment. If more than one face is detected in an image, we
choose the most centered one. In the Chalearn dataset we
were able to detect all faces using a combination of
techniques. Given the face bounding boxes, we detect
accurate, mostly due to occlusion or low resolution of some
images. As shown, the performance of our method drops
mostly for gray-scale/old-style images where our network
tends to over estimate the age for those images.68 facial
landmarks and use those for alignment. Finally the aligned
faces are all cropped and resized to a fixed size. In contrast
to some previous works, which do not use any face
alignment, we found this to be important in our final accuracy
numbers.

Deep training: Our face recognition model is not only

computationally inexpensive, but also achieves outstanding
results on the LFW dataset. This model serves as the
backbone of our facial attribute recognition engine. We
designed a highly optimized deep network architecture for
accuracy and speed for each task. In some recent works,
researchers try to design a network which performs all tasks
at the same time, and they have shown marginal
improvements. However, having separate networks for each
task allowed us to design faster and more portable models
for each task. Additionally running all models combined
takes less time compared to the all-in-one model of and we
achieve better results. Age estimation is an important
problem in computer vision and pattern recognition.
Estimating the age from facial images has received great
interests in recent years.
Typically the goal of age estimation is to predict the chronological
age of a person given his or her face picture.

An age estimation system usually involves two components, i.e

aging pattern representation and aging function learning.
Representative approaches to age estimation can be found in
some survey papers. A related but different problem is apparent
age estimation, where the focus is on predicting how old does
the person look like? rather than what is the real age of this
person? It is relatively new to study the problem of apparent age
estimation. A key difference between apparent age estimation
and the traditional age estimation is that the age labels are
annotated by human assessors rather than the real chronological
age. In reality, some people may look younger than the real
chronological age, while some may look older. As a result, the
apparent age may be quite different from the real age for each
subject.

The challenge for apparent age estimation is that there are not
many face images available with annotated age labels.
Further, the annotated age labels for each face photo may not be
consistent among different assessors. The problem of apparent
age estimation is studied in this work. Particularly, we firstly utilize
a large number of face images without apparent age labels to
learn a face representation using the deep neural networks, then
we study how to fine-tune the deep networks using a limited
number of data with apparent age labels. We also explore how
well the machine learning methods can perform to estimate
apparent ages.
 WHAT IS AGE
ESTIMATION ?

Age estimation is an active topic today due to the growing

necessity of including this information in real-world systems. This
necessity comes from the fact that age is important to understand
requirements or preferences in different aspects of the daily life of
a person. Systems implementing age specific human computer
interaction can cope with these aspects. Some examples are
biometric systems that filter their database for the estimated age
range of a subject, vending machines capable of denying some
products such as alcohol or cigarettes to an underage customer,
or advertisements in different automated environments that can
be personalized according to the age of the individual interacting
with the system. Automatic facial age estimation is affected by the
traditional factors that make face analysis difficult in general.
When a person smiles, for instance, wrinkles are formed and
these can be misleading when only the appearance cues are
taken into account . Similarly, sagging of the face in a sad
expression can resemble the effects of aging.
Face analysis is an active research topic in computer vision with
applications in surveillance, human-computer interaction, access
control, and security. In this work, we focus on apparent age
estimation. Traditionally, the problem is tackled through pure
classification or regression approaches. Given an input image, we
first apply the age group classification algorithm to obtain a rough
estimate and then perform age group specific regression to
obtain an accurate age estimate.

Human facial age estimation aims to automatically label a facial

image with exact age or age group based on visual features
derived from a facial image. In addition, the proposed method
aggregates the linear prediction models for different age labels
into a matrix. Instead of learning all the prediction models
independently, we propose to learn all the prediction models
simultaneously by exploring the theory of matrix recovery, where
a trace norm regularization is introduced to capture the
dependence among different age labels and to control the model
complexity.
These images may provide no age label, but it is easily to derive
the age difference for an image pair of the same person. To
improve the age estimation accuracy, we propose a novel
learning scheme to take advantage of these weakly labeled data
via the deep Convolutional Neural Networks (CNNs).
The combination of these losses is designed to drive the neural
network to understand the age gradually from only the age
difference information. We also contribute a dataset including
more than one hundred thousand face images attached with their
taken dates. Each image is both labeled with the timestamp and
people identity. Experimental results on two aging face databases
show the advantages of the proposed age difference learning
system and the state-of-the-art performance is gained.
 Database Avaliable and Used

Adience Age Database

Adience Age Database is another large scale database, which

was collected to capture all the variations in appearance, noise,
pose, lighting and more, and can be expected to serve as images
taken without careful preparation or posing. There are 18,300
images in this dataset. However, different from the CACD, this
database doesnt have the accurate age annotation, i.e. age is
annotated in a range form like 0-2, 4-6, 8-13, 15-20, 25-32, 38-43,
48-53, 60+. As we use Euclidean loss for the deep networks, in
this work we transfer the range annotation into traditional single
label by just taking the mean value, e.g., 28-32 is converted to 30.

Morph Age Database

The Morph database was also used for our network training. It is
a large database containing two sections, I and II. Since Morph-I
is too small, we used Morph-II that contains about 55,000 face
images. The Morph is a multiethnic database. It has about 77%
Black faces and 19% White, while the remaining 4% includes
Hispanic, Asian,Indian, and Others. Although Morph has a great
number of faces, it is like mug-shot images, which is quite
different from the faces in the wild.

FGNET Age Database

FGNET consists of 1,002 images of 82 subjects, labeled with

accurate chronological age. Some of these photos were acquired
under controlled conditions. One characteristic of this database is
that it contains a lot of samples of young ages, i.e., from 1-13
years.

Lifespan Age Database

Lifespan is a database of 575 individual faces ranging from ages

18 to 93, and it was developed to be more representative of age
groups across the lifespan, with a special emphasis on recruiting
older adults. The database has faces of 218 adults age 18-29, 76
adults age 30-49, 123 adults age 50-69, and 158 adults age 70
and older.

Other Age Database

Besides the above public databases, in this work we alsoused a

private age database. This database is composed of controlled
face images which were captured in studio environment and in
the wild. Totally, we have about 8,941 faces with age labels in this
dataset.
Difference Between Neural Networks &
Deep Neural Networks

In machine learning, a convolutional neural network (CNN,

or ConvNet) is a class of deep,feed-forward artificial neural
network that have successfully been applied to analyzing visual
imagery. CNNs use a variation of multilayer perceptrons designed
to require minimal preprocessing.[1] They are also known as shift
invariant or space invariant artificial neural
networks (SIANN), based on their shared-weights architecture
and translation invariance characteristics.[2][3]
Convolutional networks were inspired by biological processes[4] in
which the connectivity pattern between neurons is inspired by the
organization. Individual neurons respond to stimuli only in a
restricted region of the visual field. CNNs use relatively little pre-
processing compared to other image classification algorithms.
This means that the network learns the filters that in traditional
algorithms were hand-engineered. This independence from prior
knowledge and human effort in feature design is a major
advantage.They have applications in image and video
recognition, recommender systems and natural language
processing.

Deep learning is a class of machine learning algorithms that use a

cascade of many layers of nonlinear processing units for feature
extraction and transformation. Each successive layer uses the
output from the previous layer as input. The algorithms may
be supervised or unsupervised and applications include pattern
analysis and classification are based on the learning of multiple
levels of features or representations of the data, are part of the
broader machine learning field of learning representations of data,
learn multiple levels of representations that correspond to
different levels of abstraction; the levels form a hierarchy of
concepts.
These definitions have in common multiple layers of nonlinear
processing units and the supervised or unsupervised learning of
feature representations in each layer, with the layers forming a
hierarchy from low-level to high-level features. The composition of
a layer of nonlinear processing units used in a deep learning
algorithm depends on the problem to be solved. Layers that have
been used in deep learning include hidden layers of an artificial
neural network and sets of complicated formulas.
At each layer, the signal is transformed by a processing unit, like
an artificial neuron, whose parameters are iteratively adjusted
through training.

Architecture of Age Estimation

Given the image data, we first applied face detection and

landmark localization. Images are rotated every 10 degrees
for further detection if no face is detected in the original image.
The rest of images that are still not detected by rotation are not
used in our approach. During the final submission, all the
undetected images are set to the average age from the overall
predictions as their age label. All the detected face images are
cropped and aligned by the eye locations. The face images from
the age estimation challenge are collected in the wild and the
number of training images is very limited. There are various
poses, illumination, and image quality issues in the dataset. To
handle this issue, we apply data augmentation to create new
training samples from the given training data before fine-tuning
the deep network. Recently, deep learning methods especially the
convolutional neural networks (CNN) have shown promising
performance in age estimation.

The first step is to pre-train the network using a large number of

face images e.g., from the dataset. In order to learn more robust
and representative features by deep neural networks, in our
method we expect to utilize a much deeper neural network rather
than previous work for real age estimation. Particularly, we chose
the FGNet, which is 22-layer deep neural network, to train our
deep models. We hope that through this pre-training step, the
deep network is able to capture general facial representations.
The second step for training our deep model is fine tuning the
network parameters using a large number of data with biological
age labels. We merged multiple age databases into one training
dataset. In this fine-tuning process,we consider the age
estimation as a regression problem, therefore we change the loss
function from softmax to Euclidean loss for this real-valued
regression task. Euclidean loss function E computes the sum of
squares of differences of its two inputs. Based on the pre-trained
deep face model, now we are aiming at fine-tuning the deep
model for age estimation. Since this is a different task and the
loss function is also changed to Euclidean loss, we initialize a
relatively large value for the learning rate. After the above steps,
the learned deep network is expected to have the capability to
predict real age given input facial image. However, such deep
model may not perform well for apparent age estimation since the
real age usually differs from the apparent age. Therefore, the third
step to train our deep age model is a further fine-tuning step,
which utilizes the apparent age data. In this fine-tuning, the
learning rate is set to a relatively small number for better
convergence. We boost the learning rate for the feature layers, so
that most of the age models change slowly but let the feature
layers learn fast. After this fine-tuning the deep model is
considered to have more capability to estimate the apparent age.

Previous Works

A Cascaded Convolutional Neural Network for Age

Estimation of Unconstrained Faces---Jun-Cheng Chen1_,
Amit Kumar1_

Most of the earlier age estimation methods have focused on using

shape or textural features. These features are then fed to a
regression method or a classifier to estimate the apparent age.
Holistic approaches usually adopt subspace-based methods,
while feature-based approaches typically extract different facial
regions and compute anthropometric distances. Geometry-based
methods are inspired by studies in neuroscience, which suggest
that facial geometry strongly influences age perception. As such,
these methods address the age estimation problem by capturing
the face geometry, which refers to the location of 2D facial
landmarks on images. Recently, Wu et al proposed an age
estimation method that presents the facial geometry.

To solve the regression problem on the Grassmann manifold,

then used the differential geometry of the manifold. However, the
Grassmannian manifold-based geometry method suffers from a
number of drawbacks. First, it heavily relies on the accuracy of
landmark detection step, which might be difficult to obtain in
practice. For instance, if an image is taken from a bearded
person, then detecting landmarks would become a very
challenging task. In addition, different ethnic-groups usually have
slightly different face geometry, and to appropriately learn the age
model, a large number of samples from different ethnic groups is
required.

Unlike the traditional methods discussed, the proposed method is

based on DCNN to encode the age information from a given
image. Recent advances in deep learning methods have shown
that compact and discriminative image representation can be
learned using DCNN from very large datasets. There are various
neural-network-based methods, which have been developed for
facial age estimation. However, as the number of samples for
estimating the apparent age task is limited, the traditional neural
network methods often fail to learn an appropriate model.
Thukral et. al. proposed a cascaded approach for apparent age
estimation based on classifiers using the naive-Bayes approach
and a support vector machine (SVM) and regressors using the
relevance vector machine (RVM).However, the difference
between and the proposed approach is that we leverage the rich
information contained in the DCNN model pretrained using a
large-scale face dataset for age estimation. Also, the proposed
error correction module mitigates the influences of the errors
made at initial classification stage.

Facial Age Estimation with Age

Difference--- Zhenzhen Hu, Yonggang Wen

In the past few years, much research has been conducted

in human facial age estimation. The earliest paper published
in the area of age classification from facial images was the
work by Kwon and Lobo. They proposed a human age
classification method based on the cranio facial development
theory and skin wrinkle analysis, where the human faces are
classified into three groups, namely, babies, young and senior
adults. Lanitis et al. adopted the statistical face model, Active
Appearance Models (AAMs) ,to extract the shape and texture
information of facial images. In their work, the aging pattern is
represented by a quadratic function called the aging function.
Later, Geng et al. proposed the AGING PATTERN SUBSPACE
(AGES) algorithm based on the subspace trained on a data
structure called aging pattern vector Yan et al regarded age
estimation as a regression problem with nonnegative label
intervals and solved the problem through semidefinite
programming. They also proposed an EM algorithm to solve the
regression problem and speed up the optimization process.
Instead of learning a specific
aging pattern for each individual, a common aging trend or
pattern can be learned from many individuals at different
ages. One possible way to learn the common aging pattern is
the age manifold , which utilizes a manifold embedding
technique to learn the low-dimensional aging trend from many
face images at each age. After that, various aging features
were developed for facial age estimation. Aging-related facial
feature extraction is more focused by the appearance model.
Hayashi et al. considered both texture (wrinkle) and skin color.
Another similar research topic in recent years is apparent age
estimation. In the competition organized by ChaLearn [35], age is
labeled by different volunteers given only the images containing
the single individuals. Compared with real age, the annotated
apparent age could be mutable, but the mean of labels from
different annotators are highly stable and thus can be defined as
the apparent age. Liu et al. and Yang et al. both adopted the
deep learning framework and label distribution learning for
apparent age estimation. In the work, they combined age
classifier and age regression models based on GoogLeNet. For
the age regressor, the Euclidean loss was used to measure the 1-
dimensional real value encoding. Although these two problems
may be similar with each other, there are still some differences.
First, there is no ground truth age of apparent age dataset and the
age annotation is very subjective. Second, in the challenge of
ChaLearn, the measurement of apparent age estimation is mean
normalized error, while for the real age estimation is mean
absolute error. The age estimator proposed by and is not suit for
standard aging dataset, such as FGNET and MORPH.

Conclusion

Face analysis is an active research topic in computer vision with

applications in surveillance, human-computer interaction, access
control, and security. In this work, we focus on apparent age
estimation. Traditionally, the problem is tackled through pure
classification or regression approaches. Given an input image, we
first apply the age group classification algorithm to obtain a rough
estimate and then perform age group specific regression to
obtain an accurate age estimate. Age and gender play
fundamental roles in social interactions. Despite the basic roles
these attributes play in our day-to-day lives, the ability to
automatically estimate them accurately and reliably from face
images is still far from meeting the needs of commercial
applications. This is particularly perplexing when considering
recent claims to super-human capabilities in the related task of
face recognition. Past approaches to estimating or classifying
these attributes from face images have relied on differences in
facial feature dimensions or tailored face descriptors. Most have
employed classification schemes designed particularly for age or
gender estimation tasks, including and others. Few of these past
methods were designed to handle the many challenges of
unconstrained imaging conditions. These images represent some
of the challenges of age and gender estimation from real-world,
unconstrained images.

REFFERENCES

https://google.co.in
https://git-hub.org
mathworks.com
www.wikipidea.com
https://www.tutorialspoint.com/matlab/
www.youtube.co.in
 FUTURE SCOPE

In this report recent works in the field of age estimation was

discussed. Many researchers had contributed and are still
working in this field. Though there is number of problems in
existing systems that need to be addressed such as occlusion
created by spectacles, cap or facial hairs, uneven illumination
However some of the problems like non frontal pose, collection of
images along with their age on large scale have already been
solved but still have room for improvement. Researchers are
working hard to use automatic age estimation in other research
fields as well one of its example is Age Invariant Face
Recognition. Automated systems can be installed in vehicles to
restrict under age drivers to drive a vehicle. In general, different
facial age estimation approaches and algorithms can be used to
get effective results in different real life scenarios.