International Journal of Innovative Research in Computer Science & Technology (IJIRCST)

ISSN: 2347-5552, Volume-10, Issue-1, January 2022

https://doi.org/10.55524/ijircst.2022.10.1.4
Article ID IJIR-1148, Pages 15-21
www.ijircst.org

Age and Gender Prediction using Caffe Model and OpenCV

Sharik Shaban1, Ravinder Pal Singh2, and Dr. Monika Mehra3
1
M.Tech, Department of Electronics and Communication Engineering, RIMT University, Mandi Gobindgarh, Punjab, India
2
Associate Professor Department of Research, Innovation & Incubation, RIMT University, Mandi Gobindgarh, Punjab, India
3
Head of Department, Department of Electronics and Communication Engineering, RIMT University, Mandi Gobindgarh,
Punjab, India
Correspondence should be addressed to Sharik Shaban; [email protected]
Copyright © 2022 Sharik Shaban et al. This is an open access article distributed under the Creative Commons Attribution License, which permits
unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

ABSTRACT-Automatic classification of age and gender video, scanner, web and camera, whether a human face is
has become crucial for a rising number of applications, present or missing. Facial analysis in the recent past has
especially as social platforms and social media have risen. become well recognised in the computer-view
However, there is still substantial lack of performance of community[1–4]. The human face includes identity, age,
present approaches in real-world photos, especially when gender, emotions, and people's race characteristics [5, 6].
compared to the enormous performance jumps reported In my work, I haveused OpenCV and Haar Cascade
lately in the associated facial recognition job. In this research algorithm with pretrained Caffe Model for age and gender
we show that a considerable gain in performance may be detection.
achieved by the application of convolution neural networks
A. Open CV DNN
(CNN). This work is primarily designed to construct an
algorithm that accurately guesses a person's age and It is an OpenCV module. A pre-trained Tensorflow model
gender.Haar cascade is one of the most often utilised can be used. This is, however, a deep neural network that
approaches. In this research we provide a model that can help may be utilised with a pre-trained model to infer. Support for
Haar Cascade to determine a person's gender. The model several frames such as Cafe, Tensorflow, Darknet and
trained the classifier as positive and negative pictures using PyTorch is provided by OpenCV DNN. Various applications
diverse photos of men and women. Various face like face detection and object detection may be created using
characteristics are removed. With the help of Haar Cascade, this module.
the classifier determines if the picture input is men or B. Caffe Model
women. Even with insufficient data, it functions effectively.
Caffe is a profound learning context that takes into account
A deep education framework created with Caffe is used to do
expressiveness, speed and modularity. Berkeley AI Research
the age or sex approximation task. Our model is able to detect
(BAIR) is produced and supported by community
multiple faces in single image and predict age and gender of
contributors. During his PhD at UC Berkeley, Yangqing Jia
all faces present in the image.
designed the project. Under the BSD 2-Clause licence, Caffe
KEYWORDS- Haar Cascade, Caffe Model, OpenCV, is released.
Convolutional Neural Network. C. Caffe Data Storage
In 4-dimensional arrays called blobs coffee saves and
I. INTRODUCTION transmits data. Blobs give a single memory interface with
Usually interchangeable, the word "facial detection" and plenty of pictures, parameters or modifications to parameters
"facial recognition" has various significance. Face and face (or other information). By synchronising from the CPU host
detection are both complex computer vision tasks; yet, face to the GPU devices, the blobs cover the mental and
detection is normally the first step in several facial computational overhead of the mixed CPU/GPU process.
applications that identify the existence, location and size of Usually, data may be loaded from the disc to a CPU blob, the
human faces in digital images, whereas face detection CUDA kernel is called for GPU computing and the blob is
consists mainly of two phases.The process of recognising shut off to the next layer, disregarding the low level details
and locating the face is also viewed as a classifying task and retaining its high performance. Host memory and device
which should be properly classified by facial detection are assigned to efficient memory utilisation on request.
techniques (a small False Positive Rate) when an input is
presented as a digital picture of various sources that include

Innovative Research Publication 15

 International Journal of Innovative Research in Computer Science & Technology (IJIRCST)
D. Caffe Layers 2) Creating Integral Images
A caffe level is the core of a neural network layer: one or Functions that are important to the study must be
more blobs are used as input and one or more blobs are selected since this makes the process of detection faster
produced as output. Levels have two main duties in the and more accurate. Comprehensive pictures decrease
functioning of the whole network: a forward transmission,
the computer time on an image input by utilising just
which takes the inputs and creates the outputs and a return
transition, which takes the gradient from the output and
their four corner values to calculate the total of all
calculates the gradient from the parameters and inputs which, pixels in a particular rectangle. Therefore, not all the
in turn, are transmitted back to the previous layers. Including: functional values are computed, instead of certain basic
convolution, pooling, internal product, nonlinearity like calculations, on the collar pixel input picture
linear and logistic corrected, normalisation of local response values. It is worth noting that the pixel value (x, y) for an
and losses such as softmax and hinge. These are all kinds of integral picture is derived by summing the above and left
visionary jobs needed for the most advanced. Due to the pixels (x, y). In addition, the final total of the rectangle is
composition of networks, coding of custom layers needs computed by removing the total of the alternative pairings.
minimum effort. The integrated picture contributes to enabling these costly
computations to be performed rapidly in order to understand
E. Face Detection using Viola-Jones Algorithm if a feature with several attributes fits the requirements.
Viola Jones algorithm has its name from the work "Rapid
F. Ada Boost Training
Object Detection Using a Boosted Cascade of Simple
Features" by two computer vision researchers who presented In addition to selecting the best features, AdaBoost is also a
the approach in 2001, Paul Viola and Michael Jones. new idea which trains the classifiers. Although there are
Although Viola-Jones is an outdated frame, it is highly about 160.000 functionalities in the 24/24 detector window,
strong and is an amazing application for real-time face only certain of them are crucial to identify a face. So, in the
identification. This algorithm is quite slow to train, yet can 160,000 function, we apply the AdaBoost algorithm to find
recognise faces with amazing speed in real time. the best features. Each hair-like function represents a weak
Given a picture, the algorithm looks at numerous smaller learner in the Viola-Jones algorithm. AdaBoost examines the
sub-regions and tries to discover a face by looking at certain performance of all classifiers you deliver You assess the
attributes inside each sub-region. It must verify many distinct classification performance of all the photos used for training
locations and dimensions since a picture has multiple faces in all sub-regions. Some sub-regions have a significant
of varied sizes. For detecting faces in the system, Viola and reaction in the classification system. These are classed as
Jones exploited hairlike characteristics. positive, implying that the classifier believes it has a human
There are four primary phases in the Viola Jones algorithm, face. In the view of the classifiers, sub-regions that do not
which we address in the following sections: produce an effective response do not have a human face.
• Haar-like characteristics They are categorised as negative. The good performance
• Create a comprehensive picture classifiers are more important or more important. The
ultimate outcome is a strong classifier, which incorporates
• Training AdaBoost
the top performing weak classifiers, also termed an increased
• Creating cascades for the classifier classifier.
1) Selecting Haar-like features G. Cascading Classifiers
In the 19th century, Alfred Haar, a Hungarian Training in the Cascade classification calls for both positive
mathematician, presented a sequence of rescaled 'square' and negative imagery. The positive is the thing being
functions that formed the basis of a wavelet family. Voila and discovered, the negative is the thing not being found. It is
Jones have adapted and developed the so- called haarlike mostly a question of discarding non-faces and spending more
qualities to the notion of employing hair waves. Digital time on likely face areas to lower calculation costs. The
picture characteristics utilised for object detection are hair- Cascade classifier consists of phases with a powerful
like characteristics. The eyes area is darker than its classifier in each level. The objective of each level is to
neighbouring pixels and the nose area is brighter than the eye assess whether or not the supplied sub fenster is a face. Each
area. All human faces have some universal features of the step in this algorithm thus has a crucial function to play for
human face. the fast and reliable facial recognition process.
The pixel values of the two areas may be summed up and
compared simply to find out whether section is brighter or
darker. The pixel value in the darker area is lower than the
II. PROBLEM STATEMENT
pixel value in the brighter section. It might be an edge of the Since the 1960s, the field of facial recognition has been a
eyebrow on one side or it can be shiner in the central part subject for study. It was significant both because the issue
than the surrounding boxes that can be construed as a nose was practical and because of the cognitive scientists'
We may do this through the use of hair-like characteristics theoretical curiosity. Face recognition is intended to verify or
and interpret the various areas of a face with them. identify the identification of a person using a single
photograph or a video feed of his or her face. Face

Innovative Research Publication 16

 International Journal of Innovative Research in Computer Science & Technology (IJIRCST)
recognition systems, such as safety and health care, are used Python. The project has been motivated by problems like
to track patient consumption and help pain management lack of security, frauds, child molestation, robbery, criminal
operations properly. In this case Face recognition systems are identification.
employed.Researchers lately paid more attention to this
topic, conducted numerous experiments and continually V. METHODOLOGY
improved the existing models. In computer vision,
Convolution neural networks (CNNs) are frequently utilised A. Face Detection with Viola-Jones-Haar Cascades
to enhance state-of-the-art for numerous applications. The We start by deploying the OpenCV algorithm for Viola-
availability of vast volumes of training data is one of the most Jones and recognise faces in real time in the camera feed. All
significant components.Based on the fact that it is often the you need is to instal OpenCV and Python on your PC. This
case that a face recognizer is built up, especially if the dataset is really straightforward.
is restricted, it is difficult.One of the main problems with a We have many Haar Cascade models in OpenCV that are
restricted dataset is that if various lightnings, a person's face trained and stored in XML files. We utilise this file rather
may appear different, yet different people may have identical than create and train the model from scratch. Haar cascade
appearances. Assume you should design a mobile ID XML file which is a webcam classifier for the identification
unlocker. The person wouldn't be able to demand millions of of a certain item. Using OpenCV to detect the head face, you
photographs for the facial recognition system to be uploaded. have haarcascasde frontalface default.xml. OpenCV links to
In this circumstance, it would be an appropriate method to a camera that users may use to scan their faces for age, sex
use only one or a few samples. and emotional categorization.In this work, we will be using
the file "haarcascade frontalface alt2.xml." Let's begin
III. LITERATURE REVIEW coding now. One approach is to locate the route of the file
"haarcascade frontalface alt2.xml." We achieve it via the
Imane et al. [8] presented to the classifier and to the uniform Python language os module.
local binary pattern histogram (ULGBPHS) for pattern scan
a face detection system employing HOAR cascades, B. Working of Viola-Jones Classifier for Detecting Faces
standardisation and emotion detection by utilising CNN on For the detection of individual faces, we utilise haarcascade
FER 2013(KNN). The model employed 4 distinct machine frontalface default.xml. This file has a set of features with a
learning methods, i.e. 70 percent at 106 epochs with the use face (eyes, nose, beard,etc). If measurements are big, we will
of KNN and SVM algorithms, and the accuracy rates use the picture dimensions to improve the output:
(SVM, KnN, random forest, classification, regression trees) img=cv2.read
were good. This model can be improved with no algorithms print(‘original dimensions:’)
for machine languages. scale_percent=40% of original size
Sepidehsadat et al.[9] suggested that the network's attention Step 1: First, the path to HaarCascade FrontalFace
on the face should be made simpler by using a Gabor filter default.xml has to be found. We utilise a Python language os
since the output orientations are excellent for the facial module and utilise the OpenCV CascadeClassifier function
wrinkles, which will then be an input in the CNN. The to detect the xml file. The XML file path passes to the
network focuses on the useful characteristics with an age OpenCV function of CascadeClassifier().
accuracy of 7% and sexual accuracy of 2%. import os
In order to enhance the overall result, Ari Ekmekji[10] has face_cascade=
devised a model that combines interrelationships between cv2.CascadeClassifier('haarcascade_frontalface_alt.xml')
sex and age. The weaknesses include the complexity of
Step 2: After loading the classifier, let us open the webcam
separating the data into folds, training and cross-validating
using this simple OpenCV one-liner code video_capture =
the classifiers and merging the resultant classificators into a
cv2.VideoCapture(0)
test-ready classifier.
Next, we need to get the frames from the webcam stream, we
do this using the read() function. We use the infinite loop to
IV. OBJECTIVES get all the frames until the time we want to close the stream.
The goal of this project is to implement a real-time face while True:
recognition system using deep learning. This as it proceeds # Capture frame-by-frame
from the assumption that there is only limited images ret, frame = video_capture.read() The read() function
available to learn from. Algorithms should be evaluated returns:
based on accuracy. Humans are capable of determining an  The actual video frame read (one frame on each loop)
individual’s age and gender relatively easily using facial  A return code
attributes. Although it is challenging for machines to perform
the same task, in the past decade incredible strides have been If we have run out of frames, the return code notifies us that
made in automatically making prediction from face image. this will happen if we reading from a file. When we read from
The project identifies or detects the age and gender from the the camera, this does not important, because we may record
given face images. The tools used involve Convolutional it indefinitely, therefore we ignore that fact.
Neural Network along with programming language like

Innovative Research Publication 17

 International Journal of Innovative Research in Computer Science & Technology (IJIRCST)
Step 3: For this specific classifier to work, we need to instance and we set the green part to be nil.
convert the frame into grayscale. gray = Step 5: Then, the frame will be only displayed and a means
cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY) out of this endless loop and close the video stream will also
The object faceCascade contains a MultiScale() detecting be set. Pressing the 'q' button allows us to escape here.
function that gets an argument of a frame(image) and runs cv2.imshow('Video', frame)
the classification cascade over the picture. MultiScale means if cv2.waitKey(1) & 0xFF == ord('q'): break
the algorithm looks at multi-scale picture sub-regions to The next two lines are just to clean up and release the
recognise faces of different sizes. picture. video_capture.release()
faces = cv2.destroyAllWindows()
faceCascade.detectMultiScale(gray,scaleFactor=1.1,minNei
ghbors=5,minSize=(60, 60), Let us go through these 4) Classification using Convolutional Neural Network
arguments of this function: The coagulating layer is the basic building block of CNN.
Convolution is a mathematical technique which combines
1) Scale Factor
two data sets. In our scenario, the input data is converted to
Specifies how much each picture scale reduces the size of the
a feature map using a convolution filter.A convolutional
picture. You can scale the input image to a smaller one and
neural network consists of several layers. Implicit
detect it by means of an algorithm. 1.05 is a decent figure,
explanation about each of these layers is given below.
which suggests that you are resizing by a tiny step.
5) Convolution Layer (Conv Layer)
2) Min Neighbors
The Conv layer is the core building block of a Convolutional
parameter indicating how many neighbours should be Neural Network. The primary purpose of Conv layer is to
retained by each candidate rectangle. The identified extract features from the input image.
faces are affected by this parameter. Higher value
6) Pooling Layer (Sub-sampling or Down-sampling)
results in less but higher quality detections. For it, 3~6 Pooling layer decreases the sizes of maps by employing
is a positive value. functions such as the average or maximum value to
Flags: Operating mode summarise the sub-regions. The pooling process is done by
3) Min Size sliding over the input a window and passing the window
contents into a pooling function.The goal of pooling the
Minimum object size possible. Smaller than disregarded network is to minimise the number of parameters (so-called
objects. sampling) and increase the robustness of the learnt functions,
All detections for the target picture are now on the variable making them even more reliable on changes in scale and
faces. Pixel coordinates are stored for detections. The co- orientation.
ordinates in the top-left corner, the width and height of the
rectangle covering the identified face determine each 7) ReLU Layer
detection. ReLU is a non-linear operation, which represents the
Step 4: We will put a rectangle over it to show the identified Rectified Linear Unit. ReLU is a pixel-based procedure
face. which replaces all negative pixel values in the character map
The rectangle() of OpenCV creates pictures across with zero. The goal of ReLU is to add nonlinearity on our
rectangles, and the top-left and top-right corner pixels ConvNet, as we would like our ConvNet to learn most of the
coordinates need to be known. The co-ordinates display the real world data is nonlinear. Instead of ReLU other non-
pixel row and column. These coordinates may be simply linear functions such as tanh or sigmoid can be utilised,
obtained from the variable face. however in most circumstances, ReLU is superior.
We also know the position of the face, we define a new Output = Max(zero, Input)
region including a person's face and call it the face of the
8) Fully Connected Layer
ROI. We identify and round the eyeballs with the circular
function on the ROI face. The fully connected layer is just what its name implies: it is
for (x,y,w,h) in faces:cv2.rectangle(frame, (x, y), (x + w, y completely linked to the previous layer's output. All neurons
+ h),(0,255,0), 2) in the preceding layer (be it completely interconnected,
# Get Face pooling or convolution) are connected to each fully
face_img = image[y:y+h, h:h+w].copy() connected layer and are linked to each neuron. Adding a
The function rectangle() accepts the following arguments: completely linked layer is also a cost-effective technique to
The original picture learn non-linear combinations. Most characteristics learnt
the top-left detection point co-ordinates through convolution and pooling layers may be useful, but
the lower-right detection point co-ordinates these characteristics may be much better together.
The rectangle colour (a tuple that determines the quantity of 9) Softmax
the red, green, and blue (0.-255), the following parameters
The topmost level of the proposed architecture is a
are supported by rectangle().
The thickness of the rectangular lines is just 255 in our softmax layer that calculates the optimum loss term

Innovative Research Publication 18

 International Journal of Innovative Research in Computer Science & Technology (IJIRCST)
during training and also the probability for the category sensing: Age
during classification. While certain loss layers such as The output level of CNN in that CNN includes 8 values for
the multi-class SVM loss deal with the output of a fully 8 age ranges of - (0-2), (4-6), (8-12), (15-20), (25-32), (48-
connected layer as class scores, softmax treat these 53), (60-100), and is a class of 8 values.
OpenCV provides a method for deep learning classification
scores, also termed the multinomial logistic regression,
picture preprocessing:
because of the classes' unexpected log chances.
• blobFromImage ().
10) Flow of steps • Mean substraction
The flow are the steps for the proposed method • Scaling
Step 1: Load the dataset • And swaping of the canal optionally.
Step 2: Detection of Face using Haar Cascade Classifier This way blobFromImage builds an image blob in four
It is the main and crucial element of any image of the face. dimensions. Resize and crop the central picture, remove the
In certain cases, a picture could include rather than face the average value, the scaling scale values, switch Blue and Red
distinct items. Thus it is crucial to detect the face in this channels.
respect[43]. In this research, due of its quickest detection blob = cv2.dnn.blobFromImage(image, scalefactor=1.0,
property, the Viola-Jones face detection algorithm is size, mean, swapRB=True)
employed. The characteristics of the skin of the face picture Image: here is the pre-processed input image before it's sent
are necessary instead of other portions of the photos. through our deep neural classification network.
Step 3: Crop the Detected Face Image and Resize It Size factor: We may choose scale our pictures by a factor
First, the photos are rescaled to 256 x 256 before the input is whenever we conduct a mean subtraction. This default value
applied to CNN and the network will be supplied with the is 1.0 (i.e. no scaling), however another value can also be
crop 227 by 227. Before extracting the features, each picture supplied.
input will be scaled to meet the CNN input format. Size: Here we are providing the spatial size expected of the
CNN.
Step 4: Feature Extraction Mean: they are the mean value of our subtraction. They can
The resized input is sent for the function extraction phase be 3 times the amount of RGB, or they can be a single
during this step, which extracts the features using the number, in which case each channel of the image is removed.
CNN[11]. The network architecture provides knowledge If the mean subtraction is performed, ensure that 3fold is
about the layers, filters and FC layers employed. provided in (R, G, B) order, particularly when using
Step 5: Testing using Neural Network SwapRB=default True's behaviour.
Caffe is utilised for the classification of age and gender of blob = cv2.dnn.blobFromImage(face_img, 1, (227, 227),
humans in which the values of retrieved characteristics above MODEL_MEAN_VALUES, swapRB=False)
CNN layers are utilised for caffe and Keras. Here, though, #Predict Gender gender_net.setInput(blob)
coffee is only for categorisation. gender_preds = gender_net.forward()
gender = gender_list[gender_preds[0].argmax()] #Predict
VI. IMPLEMENTATION Age
age_net.setInput(blob) age_preds = age_net.forward()
A. Training a CNN using Caffe
age = age_list[age_preds[0].argmax()]
In training a CNN with coffee there are 4 steps:
Step 1: Preparation of data: This step allows us to clean and VII. RESULTS AND OBSERVATIONS
save the pictures in a manner that Caffe can utilise. We will
write a Python script to handle both pre- and storage images. A. Experimental Settings
Step 2: Definition of Model: In this phase, we select a CNN Because of a high number of parameters, CNN requires
architecture and describe its parameters in a.prototxt file. plenty of training data. Furthermore, training is extremely
Step 3: Definition of Solver: The model optimization solver time-consuming, optimization may need hours or months. In
is responsible. In a configuration file with extension .prototxt order to solve this obstacle, two stages are used to develop a
we define solver parameters. transfer learning strategy:
Step 4: Training Model: We train the template from the Prior to training: randomly initialised networks are initially
terminal with one Caffe command. We will obtain the trained trained by an accompanying task which has sufficient
model in a file with extension.caffemodel when the model pictures labelled.
has been trained. Fine-tune step: settings which have been learned during pre-
After the training phase, we will perform forecasts on new training are utilised to begin a new job.
unknown data using the learned model
.caffemodel. We are going to develop a script for Python. B. Results
The input can be either taken from the webcam or the images
B. Gender Detection with CNN can be given as the input. Here, the input image is taken from
The predicted gender may be one of ‘Male’ and ‘Female’ The webcam. The age and gender is found. This method achieves
forecast sex may be one of "man" and "female" CNN age high to medium accuracy. For testing we take sample images

Innovative Research Publication 19

 International Journal of Innovative Research in Computer Science & Technology (IJIRCST)
estimating age and gender. For classification we use Caffe Figure 2: Samples of Age misclassifications
model Figure 5.1 and Fig 5.2 represents the sample images
with correct age and gender classifications. Figure 5.3
represents the pictures of age misclassifications. Figure 5.4
represents the pictures of gender misclassifications.

Results of using Caffe Deep Learning Framework

If the detected face is a male, the output is M. If the detected
face is a female, the output is W.
For the age prediction, the CNN’s output layer(probability
layer) in this CNN consists of 8 values for 8
age classes of the following ranges- (0 – 2), (4 – 6), (8 – 12),
(15 – 20), (25 – 32), (38 – 43), (48 – 53),
(60 – 100).

Figure 3 : Multiple Human Age and Gender Prediction

Caffe Model Exact Match

Males Total: 50 38 76%
Females Total: 50 40 80%
Total : 100 78 78%
Table1: Caffe Model Accuracy

C. 7.2 Rate of Classification/Accuracy

Accuracy = No. of accurate prediction Total no. of
prediction
The overall accuracy of the system using caffe deep
learning framework is 78%.

VIII. CONCLUSION
Figure 1: Samples of face images with correct age and We assess the design of the CNN for good performance in
gender classifications this project. Age estimates and gender estimates via the
convolutive Neural Network are the suggested approach of
this research. The pre-trained CNN model was utilised to
extract the features from the picture. Accuracy was 76
percent correspondingly in the results analysis for age and
gender predictions using the caffe model. The model was
designed in python language. Real time and static detection
of the face have been performed. In a single picture, the
system can recognise numerous faces

REFERENCES
[1]. Transactions on Pattern Analysis and Machine
Intelligence, p. 1, 2019. E. Agustsson, R. Timofte, S.
Escalera, X. Baro, I. Guyon, and R. Rothe, “Apparent
and real age estimation in still images with deep
residual regressors on appa-real database,” in
Proceedings of the 2017 12th IEEE International
Conference on Automatic Face & Gesture Recognition
(FG 2017), pp. 87–94, Biometrics Wild, Bwild,
Washington, DC, USA, June 2017.
[2]. K. Zhang, C. Gao, L. Guo et al., “Age group and

Innovative Research Publication 20

 International Journal of Innovative Research in Computer Science & Technology (IJIRCST)
gender estimation in the wild with deep RoR [17]. Forensics and Security, 9(12), 2014
architecture,” IEEE Access, vol. 5, pp. 22492–22503,
2017.View at: Publisher Site | Google Scholar
[3]. A. Kuehlkamp, “Age estimation from face images,” in
Proceedings of the 6th IAPR International Conference
on Biometrics (ICB), pp. 1–10, Madrid, Spain, June
2013.
[4]. V. Carletti, A. S. Greco, G. Percannella, M. Vento, and
I. Fellow, “Age from faces in the deep learning
revolution,” IEEE
[5]. B. Bin Gao, H. Y. Zhou, J. Wu, and X. Geng, “Age
estimation using expectation of label distribution
learning,” in Proceedings of the Twenty-Seventh
International Joint Conference on Artificial
Intelligence, pp. 712–718, Stockholm, Sweden, July
2018.View at: Google Scholar
[6]. R. C. Malli, M. Aygun, and H. K. Ekenel, “Apparent
age estimation using ensemble of deep learning
models,” in Proceedings of the IEEE Computer Society
Conference on Computer Vision and Pattern
Recognition Workshops, pp. 714–721, Las Vegas, NV,
USA, June 2016.View at: Google Scholar
[7]. G. Antipov, M. Baccouche, S. A. Berrani, and J. L.
Dugelay, “Apparent age estimation from face images
combining general and children-specialized deep
learning models,” in Proceedings of the IEEE
Conference on Computer Vision and Pattern
Recognition Workshops (CVPRW), pp. 801–809, Las
Vegas,
[8]. NV, USA, June 2016.View at: Google Scholar
[9]. G. Antipov, M. Baccouche, S. A. Berrani, and J. L.
Dugelay, “Effective training of convolutional neural
networks for face-based gender and age prediction,”
Pattern Recognition, vol. 72, pp. 15–26, 2017.
[10]. R. Rothe, R. Timofte, and L. Van Gool, “Deep
expectation of real and apparent age from a single
image without facial landmarks,” International Journal
of Computer Vision, vol. 126, no. 2–4, pp. 144–157,
2018.View at: Publisher Site | Google Scholar
[11]. H. Han and A. K. Jain, “Age, gender and race
estimation from unconstrained face images,” Tech.
Rep., Michigan State University, East Lansing, MI,
USA, 2014, MSU Technical Report, MSU-CSE-14-
5.View at: Google Scholar
[12]. J. Huang, B. Li, J. Zhu, and J. Chen, “Age
classification with deep learning face
representation,” Multimedia Tools and Applications,
vol. 76, no. 19, pp. 20231–20247, 2017.\
[13]. E. Eidinger, R. Enbar, and T. Hassner, “Age and
gender estimation of unfiltered faces,” IEEE
Transactions on Information Forensics and Security,
vol. 9, no. 12, pp. 2170–2179, 2014.
[14]. Y. Sun, X. Wang, and X. Tang. Deep learning face
representation from predicting 10,000 classes. In Proc.
[15]. Conf. Comput. Vision Pattern Recognition, pages
1891–1898. IEEE, 2014
[16]. E. Eidinger, R. Enbar, and T. Hassner. Age and gender
estimation of unfiltered faces. Trans. on Inform.

Innovative Research Publication 21