See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/354152410

CNN Based Features Extraction for Age Estimation and Gender Classification

Article  in  Informatica · August 2021

DOI: 10.31449/inf.v45i5.3262

CITATIONS READS

8 1,723

1 author:

Mohammed kamel Benkaddour

Université Kasdi Merbah Ouargla
6 PUBLICATIONS   49 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Face recognition based on convolutional neural network and support vector machine View project

codage et transmission View project

All content following this page was uploaded by Mohammed kamel Benkaddour on 15 November 2021.

The user has requested enhancement of the downloaded file.

https://doi.org/10.31449/inf.v45i5.3262 Informatica 45 (2021) 697–703 697

CNN Based Features Extraction for Age Estimation and Gender

Classification
Mohammed Kamel Benkaddour
University Kasdi Marbah, Department of Computer Science and Information Technology
FNTIC Faculty, Ouargla, Algeria
E-mail: benkaddour.kamel@univ-ouargla.dz

Keywords: biometric, gender prediction, age estimation, deep neural network, convolutional neural networks (CNN)

Received: August 3, 2020

In recent years, age estimation and gender classification was one of the issues most frequently discussed
in the field of pattern recognition and computer vision. This paper proposes automated predictions of age
and gender based features extraction from human facials images. Contrary to the other conventional
approaches on the unfiltered face image, in this study, we show that a substantial improvement be obtained
for these tasks by learning representations with the use of deep convolutional neural networks (CNN). The
feedforward neural network method used in this research enhances robustness for highly variable
unconstrained recognition tasks to identify the gender and age group estimation. This research was
analyzed and validated for the gender prediction and age estimation on both the Essex face dataset and
the Adience benchmark. The results obtained show that the proposed approach offers a major
performance gain, our model achieve very interesting efficiency and the state-of-the-art performance in
both age and gender scoring.
Povzetek: V prispevku je opisana študija s konvolucijskimi nevronskimi mrežami (CNN) za prepoznavanje
starosti in spola iz obrazov na slikah.

1 Introduction
Human face image analysis is an important research area improvement in performance and enhance the recognition
in the field of pattern recognition and computer vision, in accuracy of age and gender classification.
which many researchers concentrate on creating new or The remainder of this paper is organized as follows:
improving existing algorithms to several face perception first briefly reviews the related work, then illustrates the
tasks, including face recognition, age classification, proposed method and network architecture in training and
gender recognition, etc. More closely, a face image is the testing of the data, after that the experimental results are
source of many kinds of important information’s, such as given and finally, conclusions are drawn in.
identity, expression, emotion, gender, race, age, etc.
Human facial image processing research is undergoing in 2 Related works
many directions and it still active and interesting, where
age estimation and gender distinction from face images The age and gender classification system is used to
play important roles in many computer vision based categorize human images into various groups determined
applications as parental controls of the websites, video by facial features. Accuracy of correctly classified image
services and shopping recommendation systems [1]. depending on the age, gender and their combination can
Over the last decades, many methods have been be calculated by number of correctly classified images
proposed to tackle the age and gender classification task, comparing to the stored values of the corresponding
the convolutional neural networks (CNN) is one of them images in the database. Estimating human age group and
more recently it has been employed in face image based gender automatically via facial image analysis has lots of
age and gender classification tasks [2][3]. However, as potential real-world applications, such as human computer
face images vary in a wide range under the unconstrained interaction and multimedia communication. Many
conditions (namely, in the wild), the performances of researchers give their contribution to the human age and
CNN still need to be improved, especially in age gender classification system which is a growing area in the
estimation tasks [4]. In this study, we concentrate on the research of the past decade [1][3].
issue of age group classification rather than that of exact The successful applications of CNN on many
age estimation. Though extensive approaches for age computer vision tasks have revealed that CNN is a
classification have been proposed, most of them focused powerful tool in image learning. If enough training data
on constrained images such as FG-NET [7] and MORPH are given, CNN is able to learn a compact and
[8]. After training our model, we provide a significant discriminative image feature representation. Therefore,
several studies have focused on the structuring of a new
architecture of the CNN networks to obtain the best
698 Informatica 45 (2021) 697–703 M. K. Benkaddour

recognition and prediction of age and gender classification

[5] [6] [7]. In this part, we describe the previous works on
age and gender classification.
Ari Ekmekji [9] Proposes a model for age and gender
classification, where he modified first some effective
architectures used for gender and age classification by
reducing the number of parameters, increasing the depth
of the network, and modifying the level of the used
dropout rate. In addition, the second face of his project
focuses on coupling the architectures for age and gender
recognition to take advantage of the gender-specific age
characteristics and age specific gender characteristics
inherent to images. Antipov et al [5], in this work they Figure 1: Convolutional network architecture.
design the state-of-the-art of gender recognition and age
shows a complete flow of CNN to process an input image
estimation models according to three popular datasets,
and classifies the objects based on values [10].
LFW, MORPH-II and FG-NET, where they analyzed four
The feature map of each pixel is calculated as follows:
important factors of the CNN for gender recognition and
age estimation: (1) the target age encoding and loss Cn= f (x * W + b) (1)
function, (2) the CNN depth, (3) the need for pretraining,
(4) the training strategy: mono-task or multi-task. Where “*” is the convolution computation, n the
Gil Levi and Tal Hassaner [3] suggested a structure of pixel in the feature map, x the pixel-value, W weights of
CNN constructed from five layers, two fully-connected the convolution kernel. f is a non-linear activation function
layers and three convolutional layers. The FERET dataset like sigmoid or ReLu (Rectified Linear unit), applied to
of images was used in the training and test samples. In this the result of the convolution.
structure, two methods were applied in the prediction After that some nonlinear function like the sigmoid or
process : the first was center crop for cropping the source ReLu is applied to the result of this convolution refered as
image to 227*227 in center image , and the second was a feature map. The different feature maps obtained for
over-sampling, segmenting the image into five regions different kernels are also called channels. Additionally,
each one making 256*256. Experimentally, they one also applies a pooling layer (or sub-sampling layer),
presented good results in gender and age prediction where the output of the convolution with some filter is
compared to the previous scholars. divided in tiled squared regions and each regions is
summarized by a single value. This value can be the
maximal response within the region in case of max-
3 Proposed method pooling or the average value of the region in case of mean-
pooling, but most of the data scientists uses ReLU since
3.1 Convolutional neural networks performance is better than other activation function. The
ReLU layer applies the function as follows:
In neural networks, convolutional neural network is one of
the main categories used for images recognition, images f(x) = max (0, x) (2)
classifications. In 1995, Yann LeCun and Yoshua Bengio We used the ReLU function in our work.After ReLU
introduced the concept of convolutional neural networks function, we apply a pooling layer to reduce the number
to classify the image successfully and for recognizing of parameters when the images are too large. Spatial
handwritten control numbers by Lenet-5 [2]. pooling also called subsampling or downsampling which
LeCun’s convolutional neural networks [11] are reduces the dimensionality of each map but retains the
organized in a sequence of alternating two types of layers important information. Spatial pooling can be of different
S-layers and C-layers, called convolution and sub- types : max-pooling, average-pooling or sum-pooling.
sampling with one or more fully connected layers (FC) in Max pooling take the largest element from the rectified
the ends. feature map. Taking the largest element could also take the
Convolutional Neural Networks are a special kind of average pooling (sum of all elements in the feature map).
multi-layer neural networks, they follow the path of its Convolutional layer and pooling layer compound the
predecessor neocognitron in its shape, structure, and feature extraction part. Afterwards, it comes the fully
learning philosophy [3]. Traditionally, neural networks connected layers which perform classification on the
convert input data into a one-dimensional vector [4], and extracted features by the convolutional layers and the
they have the ability to perform both feature extraction and pooling layers. These layers are similar to the layers in
classification. The input layer receives normalized images Multilayer Perceptron (MLP). Finally we flattened our
with the same sizes [2]. matrix into vector and feed it into a fully connected layer
Each input image will pass through a series of like neural network , we have an activation function such
convolution layers with filters (Kernels), Pooling, fully as softmax or sigmoid to classify the outputs. In the FC
connected layers (FC) and apply activation function to layer the neurons have complete connections with all the
classify an object with probabilistic values. Figure 1 activations of the previous layer, as it is in in regular neural
networks [11].
CNN Based Features Extraction for Age Estimation and... Informatica 45 (2021) 697–703 699

In the output layer, the activation function usually 3.3 Training and testing
depends on the task, for example in the case of binary or
Over the last decade, several research works have been
multiple binary classification the sigmoid activation is
published on facial gender and age estimation. The
used (which is guaranteed to have values in [0, 1]). For
algorithms usually take one of two approaches: age group
multi-way classification the softmax activation function is
or age-specific estimation. The former classifies a person
used to generate a value between 0–1 for each node.
as either child or adult, while the latter is more precise as
it attempts to estimate the age group of a person. Each of
3.2 Network architecture these approaches can be further decomposed into two key
The Architecture of our CNN as table1 shows ,is trained
on a database to age and gender classification .In this
section, we present our CNN model and explain the used
methods in the project .
Description of output
Name Type
size

Input layer Input image 3x92x112

Conv1 Convolution + ReLU 16 filters 3x3x3
S1 Max Pooling 16x2x2 stride 1
Conv2 Convolution + ReLU 32 filters 16x3x3
S3 Max Pooling 32x2x2 stride 1
Conv3 Convolution + ReLU 64 filters 32x3x3 Figure 2: Our Convolutional neural network
S3 Max Pooling 64x2x2 stride 1 architecture.
Con4 Convolution + ReLU 128 filters 64x3x3
S4 Max Pooling 128x2x2 stride 1 steps: feature extraction and pattern learning
FC1 Fully Connected + dropout 5376x1x1 classification.
FC2 Fully Connected + dropout 512 In our project, we divided the dataset into 2 folders
FC3 Fully Connected + dropou 128 train data and test data , each person in the dataset has 10
images , we put 8 of them in the training part and 2 in the
Table 1: Detailed architecture of our CNN used for testing part . In the age estimation there are 6 classes (0-
features extraction. 6), (8-20), (25-32), (38-43), (48-53) and (+ 60), the gender
"m : male" "f : female" for the label we took from the name
The Shape of the image is 92 x 112 x 3 where 92 on the images so that we renamed all the pictures for
represents the height, 112 the width, and 3 represents the example : "Adele-1-[f,(48-53)]", "dagran-2-[m,(38-43)]".
number of color channels. When we say 92 x 112 it means So each photo in training and test data has two labels,
we have 10,304 pixels in the data and every pixel has an gender label and age group label.
R-G-B value hence 3 color channels. CNN architectures
require that all input images have the same resolution 4 Datasets
(height and width). In our work we used the padding on
the convolutional layer with a default value to improves In this section, we introduce the datasets used for the
performance by keeping information at the borders. This evaluation of our proposed approach with a description of
means that the filter is applied only to valid paths to the their specifications for age group estimations and gender
entry. classifications accuracy.
The input images of our CNN pass through four
convolution layers that form the network, each layer is 4.2 Essex database
followed by ReLU function with conv1 consist 16 filters , The Essex face data is a dataset of the university Essex ,
conv2 consist 32 filters , conv3 consist 64 filters and the images are taken with a plain green background ,
conv4 consist 128 filters , then four pooling layers are used where the subjects sit at fixed distance from the camera
layer S1, S2 , S3 and S4 are subsampling layers , after the and are asked to speak, whilst a sequence of images is
preprocessing layers comes three fully connected layers taken. The speech is used to introduce facial expression
each one of them is followed by a and dropout layer.FC1 variation.This dataset contains images of male and female
with 5376 neurons , FC2 with 512 neurons and FC3 with subjects. The individuals are a students, so the majority of
128 neurons, yield to normalized class scores for either individuals are between (20- 32) years old but some older
gender or age, respectively. individuals are also present [15].
Finally there is softmax layer that sits on the top of We have made some changes to the images of this
FC3, which gives the loss probabilities for the age data by changing the image format from JPEG to BMP
estimation, the sigmoïd layer it also sits on the top of FC3 and its size from (180*200) to (92*112). In our work , we
and gives the loss class probabilities for the gender take from 20 photos 10 photos for every subject (2 for the
classification. test data, 8 for the train data) and manually labeled these
images. As an example, images corresponding to one
individual are shown in the Figure 3.
700 Informatica 45 (2021) 697–703 M. K. Benkaddour

Figure 5: Images from the Adience benchmark.

Figure 3: Accuracy rate of age estimation (Essex).

Figure 6: Sample images of one person from Essex

database.

4.3 Adience benchmark

We train and test our deep convolutional neural nework
model on the adience bechmark. Adience dataset is a
collection of face images from ideal real-life and
unconstrained environments contain approximately 26 K Figure 4: Accuracy of age estimation (Adience).
images of 2,284 with diffent age group with the parameter settings in our experiment listed in Table 1, we
corresponding gender label. The Adience benchmark plot the loss function and the accuracy function for age
measures both the accuracy in gender and age and display then gender separately and jointly.
a high-level of variations in noise, pose, and appearance,
among others [16]. 5.2 Results and discussion
In the first,we evaluate our method for classifying a person
5 Experiments, results and to the correct age. We train our network to classify face
discussions images into six age group classes and report the
In this section, we first introduce the parameter settings of performance of our classifier on essex dataset and adience
the experimental analysis, then results of the experiments benchmark.
evaluated on both Essex face dataset and Adience We also evaluate our method for classifying face
benchmark for gender prediction and age estimation are images to the correct labels gender. We test the
given, finally we draw on a performance comparison of performance on the same databse essex datasetand adience
our results with state-of-the-art works. benchmark. For this task, we train our network for
classification of two classes (female and male) and report
the result accuracy.
5.1 Evaluation of results We assess our solution for the age and gender
The evaluation of the proposed methodology can be classifications jointely on Essesex dataset and Adience
summarized using the classification accuracy rate and loss benchmark. The purpose is to predict whether a person's
rate of each set of data. gender is within a precise age range.
Correctly classified images in Subset The proposed work is based on gender and age
Accuracy rate = (3) classification using CNN. We tackled the classification of
Total Number of images in Subset
age group and gender of face images and posed the task as
Acc (total) = Accuracy(Age) + Accuracy(Gender) (4)
a multiclass classification problem as such train the model
Loss (total) = Loss(Age) + Loss(Gender) (5) with a classification-based loss function as training
targets.We perform data preprocessing on the training
Where Accuracy of correctly classified images
dataset image by the flatten on a one dimensional vector
depending on the age or gender is calculated by comparing
feature of input image for age and gender . After
the results of the algorithm of classification and the stored
successfully extracting features , these last will pass to the
age and gender values of the images in the images labels.
phase of classification by CNN , the result of the last layer
After training and testing our model according to
CNN Based Features Extraction for Age Estimation and... Informatica 45 (2021) 697–703 701

Parameters Value
Training iteration 40000 -80000
epochs
Dropout probability (keep dropout) 0.8
Dropout probability (ratsige dropout) 0.2
learning rate 0.00001
Activation function ReLU
Number of hidden units 512/128
Number of layers (Convolution-Pooling) 4
Classification function for age estimation Softmax Figure 7: Gender accuracy rate (Essex).
Classication function for gender Sigmoid

Table 2: The parameters settings used in our experiments.

Age 0-6 8-20 25-32 38-43 48-53 +60
classes
Acc (%) 89.5 93.1 95.3 97.1 99.3 98.2
(a) Essex dataset.
Age 0-6 8-20 25-32 38-43 48-53 +60
classes
Acc (%) 83.4 88.6 90.0 95.1 97.2 96.2 Figure 8: Gender accuracy rate (Adience benchmark).
(b) Adience benchmark.
Table 3: Gender accuracy per age for the Essex and
Adience benchmark datasets.

Method dataset Accuracy Accuracy

Gender % Age %
G. Levi and T. Essex N/A N/A
Hassncer [3]
Adience 77.8 45.1
benchmark
A. Ekmekji [8] Essex N/A N/A
Adience 86.8 50.7 Figure 9: Joint accuracy of age and gender
benchmark classification (Essex dataset).
A.Olatunbosun Essex N/A N/A
et al [18]
Adience 96.2 93.8
benchmark
S.M.Osman et Essex 93.93 92.3
al [20]
Adience N/A N/A
benchmark
Our Proposed Essex 99.10 95.41
model
Adience
benchmark 95.6 91.75
Figure 10: Joint loss of age and gender classification
Table 4: Age and gender classification: performance
(Adience benchmark).
comparison of our result with the state-of-the-art works.
According to parameter settings in our experiment
of convolutional network is passed to three fully
listed in Table 2, the analysis of Figure (5-6-7-8-9-10) and
connected layer , finally outputs are given to the fully
the result of Table 3 .The proposed network provides
connected layers to minimize the classification loss by the
significant accuracy improvements of age and gender
softmax function on six age classes ( groups 0-6, 8-20, 25-
classification, but requires considerable time for trainning
32, 38-43, 48-53, and + 60 years old ) , and by the sigmoid
our network to implement the correct prediction , where
function on two gender (female and male) classes .
702 Informatica 45 (2021) 697–703 M. K. Benkaddour

the training with 80000 epoch require a minimum of 36 References

hours. In Table 3, it is shown that the proposed approach
performs very well on adults group age, although he failed [1] P. Arumugam, S. Muthukumar , S. Selva Kumar and
to distinguish very young subjects. S. Gayathri,“ Human Age Group Prediction from
This performance is expected as even for humans Unknown Facial Image”, International Journal of
estimating age the young children is harder than the age of Advanced Research in Computer Science and
adults. Gender estimation mistakes are also frequently Software Engineering ,Volume 7,pp 722-726 ,May
occur for images of babies or very young children where 2017. https://doi.org/10.23956/ijarcsse/sv7i5/0103
obvious gender attributes are not yet visible. [2] M.K .Benkaddour and A.Bounoua," Feature
Some of the state-of-the-art methods for age and extraction and classification using deep convolutional
gender classification are compared with our model. The neural networks, PCA and SVC for face recognition",
results in Table 4 clearly show the improvement accuracy International Information and Engineering
of our method compared to previous work, The proposed Technology Association IIETA, 2017.
models achieve medium accuracy ratio, where their https://doi.org/10.3166/ts.34.77-91
systems make many misclassification , that are due to ex- [3] G. Levi and T. Hassncer, “Age and gender
tremely challenging viewing conditions of some of the classification using convolutional neural networks,”
Essesx and Adience benchmark images. As we note, the in 2015 IEEE Conference on Computer Vision and
best performance was obtained on both Essex face dataset Pattern Recognition Workshops (CVPRW), pp. 34–
and Adience benchmark. Our approach, therefore, 42, Boston, MA, USA, 2015.
achieves the best results not only on the age estimation but https://doi.org/10.1109/CVPRW.2015.7301352
also on gender classification, it outperforms the current [4] S. Chen, C. Zhang, M. Dong, J. Le, and M. Rao,
state-of-the-art methods. “Using ranking-cnn for age estimation,” in 2017
IEEE Conference on Computer Vision and Pattern
Recognition (CVPR), pp. 742–751, Honolulu, HI,
6 Conclusion USA, 2017. https://doi.org/10.1109/CVPR.2017.86
In this paper, we evaluated the application of deep neural [5] G. Antipov , M.Baccouche, J. Dugelay , Effective
network for human age and gender prediction using CNN. Training of Convolutional Neural Networks for Face-
During this study various design was developed for this Based Gender and Age Prediction , Pattern
task, age and gender classification is one of the key Recognition Volume 72 , 2017.
segments of research in the biometric as social https://doi.org/10.1016/j.patrec.2015.11.011
applications with the goal that the future forecast and the [6] G. Panis and A.Lanitis , An Overview of Research
information disclosure about the particular individual Activities in Facial Age Estimation Using the FG-
should be possible adequately. The procedure utilizes NET Aging Database ,Journal of American History
grouped approaches and calculations whereby the deep pp 455-462,2015. https://doi.org/10.1007/978-3-319-
learning is likewise the prime in usage designs in the CNN 16181-5_56
for age and gender classification. A convolutional neural [7] K. Ricanek and T. Tesafaye, “MORPH: A
networ pipeline for automatic age and gender recognition Longitudinal Image Database of Normal Adult Age-
has been proposed, we provide an extensive data set and Progression,” 7th International Conference on
benchmark for the study of age estimation and gender Automatic Face and Gesture Recognition (FGR06).
classification. https://doi.org/10.1109/fgr.2006.78
According to the results taken from the proposed age [8] A. Ekmekji , “Convolutional neural networks for age
and gender classification methodology from facial and gender classification,” Technical Report,
images has significantly higher accuracy. The proposed Stanford University, 2016.
method use facial images in the range of 0 – 68 as it is [9] E. Eidinger , R. Enbar and T.Hassner, “Age and
very difficult to recognize the gender and age of little gender estimation of unfiltered faces,” IEEE
babies and children by the geometric facial feature Transactions on Information Forensics and Security,
variations and also there are not enough facial images vol. 9, no. 12, pp. 2170–2179, 2014.
of people older than 60 years old to use in the https://doi.org/10.1109/TIFS.2014.2359646
experiment. [10] K. Fukushima, Neocognitron: Neural Network Model
In the proposed system we investigate the for a Mechanism of Pattern Recognition Unaffected
classification accuracy on Essex dataset and Adince by Shift in Position. Biological Cybernetics, 36, 193-
benchmark for age and gender classification, our proposed 202, 1980. https://doi.org/10.1007/BF00344251
method achieves the state-of-the-art performance, in both [11] Y. Lecun, L. Bottou, Y. Bengio, and P. Haffner,
age and gender estimation and significantly outperforming Gradient-Based Learning Applied to Document
the existing models. For future works, we will consider a Recognition. Proceedings of the IEEE, 86, 2278-
pretrained CNN architecture and focus the work to design 2324,1998. https://doi.org/10.1109/5.726791
such a framework for race prediction. [12] Essex face database, University of Essex, UK,
http://cswww.essex.ac.uk/mv/allfaces/index.html
[13] V. Carletti, A. Greco, G. Percannella, and M. Vento,
“Age from Faces in the Deep Learning Revolution,”
CNN Based Features Extraction for Age Estimation and... Informatica 45 (2021) 697–703 703

IEEE Transactions on Pattern Analysis and Machine

Intelligence, vol. 42, no. 9, pp. 2113–2132, Sep. 2020.
https://doi.org/10.1109/tpami.2019.2910522
[14] M. R. Dileep and A. Danti, “Human age and gender
prediction based on neural networks and three human
age and gender prediction based on neural networks
and three sigma control limits,” Applied Artificial
Intelligence, vol. 32, no. 3, pp. 281–292, 2018.
https://doi.org/10.1155/2020/1289408
[15] D. Yi, Z. Lei, and S. Z. Li, “Age Estimation by Multi-
scale Convolutional Network,” Lecture Notes in
Computer Science, pp. 144–158, 2015.
https://doi.org/10.1007/978-3-319-16811-1_10
[16] Adience Benchmark, for Gender and Age
Classification,
https://talhassner.github.io/home/projects/Adience/A
dience-data.html
[17] Z. Qawaqneh, A. A. Mallouh, and B. D. Barkana,
Deep neural network framework and transformed
MFCCs for speaker's age and gender classification,
Knowledge-Based Systems,Volume 115,Pages 5-14,
2017.
https://doi.org/10.1016/j.knosys.2016.10.008
[18] A. Anand, R. D. Labati, A. Genovese, E. Munoz, V.
Piuri, and F. Scotti, “Age estimation based on face
images and pre-trained convolutional neural
networks,” in Proceedings of the IEEE Symposium
Series on Computational Intelligence (SSCI), pp. 1–
7, Honolulu, HI, USA, November 2017.
https://doi.org/ 10.1109/SSCI.2017.8285381
[19] A.Olatunbosun and S. Viriri ,’’Deeply Learned
Classifiers for Age and Gender Predictions of
Unfiltered Faces’’ , the scientific world journal, 2020.
https://doi.org/10.1155/2020/1289408
[20] D.V.Sang, L.T. Cuong. Effective Deep Multi-source
Multi-task Learning Frameworks for Smile
Detection, Emotion Recognition and Gender
Classification. Informatica (Slovenia), 42, 2018.
https://doi.org/10.31449/inf.v42i3.2301
[21] S. M. Osman, N. Noor, and S. Viriri, “Component-
Based Gender Identification Using Local Binary
Patterns,” Lecture Notes in Computer Science, pp.
307–315, 2019. https://doi.org/10.1007/978-3-030-
28377-3_25
 704 Informatica 45 (2021) 697–703 M. K. Benkaddour

View publication stats