I.J.

Image, Graphics and Signal Processing, 2023, 5, 60-76

Published Online on October 8, 2023 by MECS Press (http://www.mecs-press.org/)
DOI: 10.5815/ijigsp.2023.05.05

Face Recognition Using Modified Histogram of

Oriented Gradients and Convolutional Neural
Networks
Raveendra K*
Government Engineering College, K R Pet, India
E-mail: [email protected]
ORCID iD: https://orcid.org/0000-0003-3629-2407
*Corresponding Author

Ravi J
Global Academy of Technology, Bangalore, India
E-mail: [email protected]
ORCID iD: https://orcid.org/ 0000-0002-1488-9547

Khalid Nazim Abdul Sattar

Majmaah University, Al Majmaah 11952, Saudi Arabia
E-mail: [email protected]
ORCID iD: https://orcid.org/0000-0002-0759-0512

Received: 10 February 2023; Revised: 15 April 2023; Accepted: 11 May 2023; Published: 08 October 2023

Abstract: We are aiming in this work to develop an improved face recognition system for person-dependent and
person-independent variants. To extract relevant facial features, we are using the convolutional neural network. These
features allow comparing faces of different subjects in an optimized manner. The system training module firstly
recognizes different subjects of dataset, in another approach, the module processes a different set of new images. Use of
CNN alone for face recognition has achieved promising recognition rate, however many other works have showed
declined in recognition rate for many complex datasets. Further, use of CNN alone exhibits reduced recognition rate for
large scale databases. To overcome the above problem, we are proposing a modified spatial texture pattern extraction
technique namely modified Histogram oriented gradient (m-HOG) for extracting facial image features along three
gradient directions along with CNN algorithm to classify the face image based on the features. In the preprocessing
stage, the face region is captured by removing the background from the input face images and is resized to 100×100.
The m-HOG features are retrieved using histogram channels evenly distributed between 0 and 180 degrees. The
obtained features are resized as a matrix having dimension 66×198 and which are passed to the CNN to extract robust
and discriminative features and are classified using softmax classification layer. The recognition rates obtained for L-
Spacek, NIR, JAFFE and YALE database are 99.80%, 91.43%, 95.00% and 93.33% respectively and are found to be
better when compared to the existing methods.

Index Terms: Face recognition, Convolutional Neural Network, softmax, deep learning, HOG, L-Spacek, NIR, JAFFE
and YALE Feature extraction · Image database Preprocessing · Recognition accuracy.

1. Introduction

The issue of face recognition has been intensively researched in the fields of computer vision and pattern
recognition. Automatic face recognition is utilized in a broad variety of practical applications, such as security,
surveillance, identity verification, intelligent vision monitoring and immigration cleansing systems. Depending on the
application environment, it may be divided into two functions: facial examination and facial recognition. The first
objective is to assess whether or not a pair of facial images belongs to the same person, while the second objective is to
learn about the individual by examining a gallery of facial images and identifying identical ones. However, face
recognition in real-world applications is still a challenge [1]. The main reason for this is because the face is not stable

This work is open access and licensed under the Creative Commons CC BY 4.0 License. Volume 15 (2023), Issue 5
 Face Recognition Using Modified Histogram of Oriented Gradients and Convolutional Neural Networks

and has different facial features, ages, multiple angles, and, most importantly, changing light intensity. Furthermore,
several parameters, such as closure and posture, have an impact on face recognition performance.
In the last few years, computer-based learning has become more popular. Over the last decade, many perspectives
have been gained from computer vision issues and discriminating solutions, such as image classification, object
identification and face recognition [2, 3].
CNN can be trained using effective features derived from recurrent pooling and convolution operations on a large
database. Currently, CNN is emerging as a powerful algorithm and has become most sufficient for dense problem
predictions, especially for face recognition problems. From the past decade many algorithms have been proposed which
are based on CNNs [2, 3] shows accuracy improvements. The size of training database images for CNNs should be
increased for optimal accuracy.
Data addition (DA) is a proposed technique that uses domain-specific synthesis to deliberately infect a database in
order to add more unchanged cases [4]. Furthermore, it is a group of low-cost computer algorithms that were previously
utilized to reduce the usage of CNN [5]. However, large data sets contain high noise signals or variations, especially if
they are automatically collected from image or from movie sources.
The main objective of the work is to develop smart and efficient face biometric authentication system by
hybridizing texture feature extraction technique (HOG) with deep learning algorithm (CNN) and thereby increasing
recognition rate. Use of CNN alone for face recognition has achieved promising recognition rate, however many other
works have showed declined in recognition rate for many complex datasets. Further, use of CNN alone exhibits reduced
recognition rate for large scale databases. The authors of some of the articles have followed DA methods to improve
training data to create more examples of deep training demand.
To overcome the above problem, our method consists of hybrid approach which contains application of CNN for
spatial feature matrix obtained using m-HOG a feature extraction technique. The m-HoG features are extracted for input
face detected image. We present a collection of face images from different databases. In fact, we apply our model with
different datasets having color variations, fading, contrast and brightness change. Performance analysis is performed by
assessing parameter like recognition rate. Based on the results of the experimental analysis, our proposed model
performed well compared to the existing models.
In this work, an explanation of the system description and methodology consists of three parts: m-HOG feature
extraction, the use of CNN and the extraction of features and their comparisons have all been explored. The remaining
of the manuscript is organized as follows: Section 2 presents the literature review. Section 3 describes the suggested
technique in detail. The experimental results are summarized and discussed in Section 4. Section 5 concludes the
manuscript.

2. Literature Review

Gaili Yue and Lei Lu [6] developed a face recognition approach that makes use of histogram equalization and a
CNN. Preprocessing of the face image is carried out by histogram equalization. CNN is constructed with Google deep
learning framework TensorFlow1.3.0, whose structure is based on LeNet-5, for which the preprocessed face images are
fed for training. The efficacy of this method in terms of recognition rate is assessed using the ORL database.
Muhtahir O. Oloyede et al. [7] introduced a technique for recognizing face images which utilizes efficient image
enhancement approach for preprocessing face images, as well as a unique set of hybrid features and CNN. The
approach uses a metaheuristic optimization strategy for robust face image enhancement in an unconstrained
environment. When compared to the original image, recognition performance improves when more features are added
to the facial image. The derived hybrid feature is meant to enhance the classification efficiency of advanced CNN
architecture. Experimentation on typical face datasets have been performed to demonstrate an increase in the efficiency
of recognition which accounts for all the limitations in the face dataset. When compared to other facial conditions,
lower face occlusion has a significant influence on the system.
Wenqi Wu et al. [8] presented a Faster R-CNN-based face detection with variable scales. In this methodology, the
network employs a number of methodologies such as multitask learning, feature pyramiding, and feature fusion. An
efficient multitask region based network with improved facial recognition is built to obtain human face region of
interest. In order to build a human face proposal, the anchor is combined with facial landmarks. The proposal scale is
then used to propose a parallel-type Fast R-CNN network. The three networks differ in the weight of feature map fusion
and are characterized by their proposal scales. The experimental findings demonstrate that the proposed technique
outperforms common benchmarks such as FDDB, PASCAL and WIDER FACE when compared to UnitBox,
HyperFace, and FastCNN algorithms.
Xianzhang Pan [9] proposed a method for acquiring comprehensive characteristics for video-based facial
expression identification using CNNs and HOG. It retrieves deep features from video frames using a set of
convolutional kernels in CNNs with displacement, scaling, and deformed invariance. The HOG is then used to generate
discriminative features from CNN deep features that are highly associated with face emotions. Classification of
expressions is carried out using the Support Vector Machine (SVM) technique. The performance of this technique was
assessed using the RML, CK+, and AFEW5.0 datasets.

Volume 15 (2023), Issue 5 61

 Face Recognition Using Modified Histogram of Oriented Gradients and Convolutional Neural Networks

Yanhong Zhang et al. [10] used a patch method in CNN architectures to build a novel technique for learning
enough effective features for face recognition. The network has enabled the cropping of a face region into patches while
requiring no additional storage capacity for face patches. Furthermore, using the patch approach, a multi-branch CNN is
trained to learn the characteristics of every cropped patch and the patch features are then fused to generate a full face
representation. The performance of this method was evaluated by conducting experiments on the LFW and YouTube
platforms.
Feng Cen and Guanghui Wang [11] proposed a deep feature dictionary model for recognizing occluded faces. The
deep features are retrieved using a CNN and then linearly coded using a dictionary. In addition to the deep features from
training samples, the dictionary includes mapping vectors from people inside or outside the training set that are
correlated with occlusion patterns from testing face samples.. To normalize the coding coefficients, a squared Euclidean
norm is utilized. The performance of this approach was evaluated by conducting experiments on AR, FERET and
CelebA databases.
Pengfei Ke et al. [12] used LBP and CNN to examine and assess the impact of posture, lighting, mood, and other
variables on face recognition. A CNN is made up of 4 convolution layers, 2 max-pooling layers, an activation layer, a
fully linked layer, and an output layer. The LBP descriptor is used to generate LBP coded images, which are then fed
into a CNN to extract spatial information from images while minimizing feature dimensionality. Batch normalization is
performed to the convolution layer to improve the network structure. Experimenting with the CMU-PIE face database
demonstrated that the proposed approach considerably improves face recognition rates.
Fanzhi Kong [13] proposed a method for recognizing facial expressions using a CNN and Local Binary Pattern
(LBP) to extract face information. The fully connected layer interprets image abstract characteristics by feeding the
original image into a deep convolutional neural network, which eliminates the inherent inaccuracy of image
preprocessing and artificially selected features. A standard LBP feature extractor for facial expression images is then
constructed at the full connection layer. The enhanced LBP facial expression texture attributes are integrated with the
abstract face expression features learnt by a deep convolutional neural network. This will provide unique facial
expression features and enhance classification accuracy.
Erfan Zangeneh et al. [14] exploited deep convolutional neural networks (DCNN) to construct a novel linked
mapping strategy for recognizing low resolution facial images. The system is made up of two DCNN branches that use
nonlinear transformations to map high and low quality images into a common space. The branch dealing with high-
resolution image processing includes fourteen layers, while the branch dealing with low-resolution facial image
mapping has a five-layer super-resolution network coupled with a fourteen-layer network. Through back propagation,
the distance between two attributes that correspond to high- and low-resolution images is used to train networks. The
recognition rate of this approach was evaluated on different databases to measure its overall performance.
Raj Silwal et al. [15] propose an approach that combines deep learning, customized architecture, and an enhanced
loss function to recognize individuals in unconstrained environments. Multi-Block Local Binary Pattern modules
retrieve customized features, whereas CNN modules extract high-level unique features in the system. Individuals are
identified by fusing the characteristics from both modules and sending them through a fully connected layer using a
softmax classifier. The accuracy and processing time are measured to test its effectiveness.
Ashok Kumar Rai et al. [16] present a hyper spectral face recognition method that combines bands using the
Firefly algorithm and classifies using the CNN. The architecture for categorization and band fusion, consisting of four
major components the formulation of the hyper spectral imaging (HSI) problem as an Image-Set, the alignment of
several posture images inside the same person, the building of an HSI Face ConvNet to train and classify the images,
and finally, the application of an upgraded firefly approach to get fused images for classifying intra-person of hyper
spectral images. The effectiveness of this method was evaluated on CMU-HSFD & UWA-HSFD databases.
Leslie Ching Ow Tiong et al. [17] suggested the multimodal facial biometrics recognition for identifying faces in
surveillance applications. It employs multi-feature fusion layers in dual-stream CNNs, which results in considerable and
useful data learning. The network consists of two progressive parts with unique fusion algorithms for aggregating RGB
information and texture descriptions for multimodal face biometrics. The performance of this technique was evaluated
on four different datasets.
Tripti Goel and R Murugan [18] proposed a method for recognizing face images based on deep convolutional
networks and Kernel Extreme Learning Machine (KLEM) classifier. To get robust features from an image, the residual
network (ResNet) uses a set of convolution and pooling layers. These layers work together to get features that aren't
affected by lighting, posture, or expression. These collected features are then learnt and classified using the polynomial
function KELM, the parameters of which are also tuned using the particle swarm optimization process. To test the
efficiency of this technique, experiments are conducted on the AT&T, Yale, CMU PIE, and UMIST datasets.
Shuai Peng et al. [19] present a new technique using the Inception-ResNet network to reduce the complexity of
training such deep convolutional neural networks while simultaneously improving their performance. A fixed residual
scaling factor is employed in Inception-ResNet. Model training stability is improved by converting the value to a
trainable parameter and setting it to a modest value. The Inception-ResNet module uses Leaky ReLU and PReLU
instead of ReLU activation function, increasing the amount of training parameters while improving training stability
and performance. The performance is evaluated by conducting extensive tests on the datasets VGGFace2, MS1MV2,
IJBB, and LFW.

62 Volume 15 (2023), Issue 5

 Face Recognition Using Modified Histogram of Oriented Gradients and Convolutional Neural Networks

Xi Yin and Xiaoming Liu [20] demonstrated the Multitask CNN for recognizing pose invariant faces, with identity
classification as the primary task and pose, illumination, and expression estimations as subtasks. The dynamic-
weighting approach is used to dynamically allocate loss weights to each sub task, hence resolving the issue of task
balance during training, when a bigger loss weight is provided to a simpler sub task. During training, CNN learns
certain identification features, as well as a stochastic routing approach for feature fusion during testing. The approach
outperforms the current state of the art on the LFW, CFP, and IJB-A datasets when tested on wild datasets.
Yanan Wang et al. [21] suggested a bi-directional Collaborative representation-based classification system for
identifying facial images using CNN features. Deep convolutional neural networks were used to obtain the face
characteristics from the original gallery and query sets, resulting in an effective reverse representation strategy for
extracting more information between training and test samples. To prevent bi-directional optimization, the input sample
is characterized by a forward and a backward linear combination. For robust classification, the remainders from forward
representation and reverse modelling are combined at a low cost. Experiments using well-known face datasets
including AR, FERET and ORL demonstrate the algorithm's validity and resistance to small sample size issues.
Hana Ben Fredj et al. [22] developed a CNN framework to recognizing faces in unconstrained environment. In the
framework, aggressive data augmentation is used for learning. Further adaptive fusion of softmax loss and center loss as
supervised signals is added to improve the performance. The effectiveness of the model was tested on the LFW and
You Tube datasets.
M. Chandrakala and P. Durga Devi [23] proposed a two stage classifier using HOG features to recognize face
images. The desired features from the preprocessed images are extracted using HOG. At first, the k-NN classifier was
used. Then, unrecognized face images were tested with the SVM classifier, which led to more accurate face recognition.
Htwe Pa Pa Win et al. [24] proposed method for face recognition using CNN. The deep learning strategies of the
CNN are used for detecting face, extracting face features and for recognition. The effectiveness of this method was
tested by experimentation on FEI dataset resulting in better accuracy and reduced time complexity.
Peng Lu et al. [25] developed a method for face recognition that combines CNN with augmented dataset. The
small dataset is augmented to a larger one by transformations of the face images, such as flip, shift, scaling, and rotation.
Face features extracted from these augmented datasets resulted in higher recognition rate. This method was tested on
the ORL dataset.
Walid Hariri [26] introduced an efficient technique for recognizing masked faces. In this technique, first the
masked face region is removed, then three pre-trained deep CNN namelyVGG-16, Alex Net, and ResNet-50 is applied
to extract deep features. The classification is performed by multilayer perception. Experiment was conducted on Real-
World-Masked-Face-Dataset to evaluate the performance.
The survey identifies several methodologies and their associated experimental outcomes for various datasets. It is
discovered that despite the development of several algorithms, high-accuracy face recognition remains a difficult
problem owing to a variety of situations such as variations in intensity, lighting, direction, lightning, occlusion, and
changing position and facial expressions. Today's deep learning algorithms are quite successful and yield promising
results for applications such as pattern recognition and picture classification, among others. Convolutional Neural
Networks (CNNs) are feed forward networks comprised of several hidden layers that automatically learn features.
While using CNNs alone for face recognition has demonstrated a promising recognition rate, several other studies have
demonstrated a drop in recognition rates for a variety of complex datasets. Additionally, using CNNs alone results in a
lower recognition rate for big scale databases. We propose a hybrid approach in this study effort that combines modified
Histogram oriented gradient (m-HoG) for extracting facial image characteristics in three gradient directions with a deep
learning CNN algorithm for classifying the face image based on the features. The study indicates that the suggested
approach outperforms existing methods in terms of recognition rates.

3. Methodology
The proposed work consists of two different approaches, a spatial texture pattern extraction technique for
extracting facial image features and deep learning algorithm for further feature extraction and classification. Figure 1
shows the proposed model block diagram.
3.1 Database description
L-Spacek Database: This database is being explored to test the algorithm because its wide range of lightning
variants, different orientations and expressions. It contains 113 male subjects, each with 20 face images. Out of 113
subjects first 100 subjects are considered for training by choosing first 10 samples of each subject, with a total of 1000
images. 15th face image to 19th face image of the same 100 subjects are considered to compute validation accuracy.
The sample images from L-Space k dataset are shown in Figure 2(a).

Volume 15 (2023), Issue 5 63

 Face Recognition Using Modified Histogram of Oriented Gradients and Convolutional Neural Networks

NIR Database: The NIR database was chosen to test the algorithm due to its variability in terms of intensity, light
change and blurring effects, as well as diverse poses and face expressions. It consists of 115 subjects having 14 samples
per subject. Out of 115 subjects first 40 subjects are considered for training by choosing first 5 samples of each subject,
with a total of 200 images. The 6th face image of the same 40 subjects is considered for compute the recognition rate.
The sample images from NIR dataset are shown in Figure 2(b).

Training Pre- Prediction

m-HoG
Database Processing model

Conv Conv
1

Testing Pre- Classified

Database Processing m-HoG
Result

Conv Conv
1 2
Fig. 1. Proposed model block diagram

JAFFE Database: The JAFFE Database is taken into account since all of the individuals have identical/similar
facial images. The database contains
213 images of 10 Japanese female models posing in seven different face expressions. All 10 models face images
are considered for training by choosing first 10 samples of each model, with a total of 100 images. The 16th and 17th
face images of the same 10 models are considered to evaluate the performance. The sample images from JAFFE dataset
are shown in Figure 2(c).
YALE Database: This database is taken into consideration because of the changes in intensity, changes in
illumination, varied poses and face expressions. It has 15 subjects containing 11 samples per subject with a total of 165
images. All 15 subjects are considered for training by choosing first 9 samples of each subject, with a total of 135
images. The 10th face image of the same 15 subjects is considered to compute the recognition rate. The sample images
from Yale dataset are shown in figure 2(d).
3.2 Preprocessing
The suggested approach employs the Viola Jones algorithm [27] to locate the face in a complex environment. This
algorithm extracts characteristics from both face and non-facial areas by using Haar-like features. Haar features are tiny
kernels with varying forms and sizes that are used to identify the existence of a feature in a given image. The Adaboost
learning method was used to reduce any redundancies in the collected features. Adaboost is a machine learning
technique that determines the best among all characteristics, however these features are referred to as weak classifiers,
while strong classifiers are constructed by linearly combining weak classifiers. Finally, the cascade classifier
incorporates powerful classifiers at several levels that are utilized to recognize the desired face in the provided image.
For all databases, the region of interest is determined by cropping and resizing to 100×100.
3.3 Modified Histogram of Oriented Features
The gradient of the image in traditional HoG is computed using the intensity change only in x-y direction. This
results in ignoring the change in intensity along z direction, which leads to the lack of discriminative features from the
image. To overcome this drawback, a modified HoG (m-HoG) is proposed, in which the gradient is computed from the
change intensity levels along x-y, x-z and y-z directions which is as shown in Figure 3, making it more descriptive of
the image and hence contains discriminative characteristics. The m-HoG gradient is used to generate three distinct
HoG versions.

64 Volume 15 (2023), Issue 5

 Face Recognition Using Modified Histogram of Oriented Gradients and Convolutional Neural Networks

(a)

(b)

(c)

(d)
Fig. 2. Sample images of (a) L-Spacek (b) NIR (c) JAFFE and (d) YALE database

Modified HoG HoG1, HoG 2, HoG3

HoG1, HoG2, and HoG3 are the m-HoG characteristic features computed using the gradient in three directions.
The feature vector in this m-HoG version is three times as long as the standard HoG.
The face detected regions using Viola Jones Algorithm have been resized for fixed dimension of 100×100.
Resizing process is done to make sure that the number of m-HoG features generated for the subsequent images are
uniform. After resizing the image, gradient calculation is performed, which comprises computing the gradient values by
using a 1D centered point discrete derivative mask in x-y, y-z and x-z directions.

Fig. 3. Shows Gradient estimation along three directions

a. x-y direction
This method entails filtering the grayscale image with the following x and y filter kernels, which are represented
mathematically as

Volume 15 (2023), Issue 5 65

 Face Recognition Using Modified Histogram of Oriented Gradients and Convolutional Neural Networks

&º L >F s r s ? and & L >s F s ? (1)

So, given an image I , we can use a convolution technique to obtain the x and y derivatives as
+º L + &º =J@ + L + & (2)

Then the magnitude of the gradient is given by:

49
+) º ? +L k+º 6 E + 6 o (3)

And angular orientation of the gradient is given by:

x y tan 1
I y / Ix (4)

b. x-z direction
&º L >F s r s ? =J@ & L >rrs rrr F srr ? (5)

So, given an image I , we can use a convolution technique to obtain the x and z derivatives as

+º L + &º =J@ + L + & (6)

Then the magnitude of the gradient is given by:

49
)º? L k+º 6 E + 6 o (7)

And angular orientation of the gradient is given by:

1
x z tan Iz / Ix (8)

c. y-z direction
& L >s r F s ?=J@ & L >rrs rrr F srr ? (9)

So, given an image I , we can use a convolution technique to obtain the y and z derivatives as

+ L + & =J@ + L + & (10)

Then the magnitude of the gradient is given by:

0.5
Gy z I y2 I z2 (11)

And angular orientation of the gradient is given by:

1
y z tan I y / Iz (12)

The m-HOG returns the total features in a row vector and consists of 13068 features for each face image. Which is
three times larger than conventional HoG features. All such row vectors for the subsequent images are stacked one
below the other as shown in Figure 4 to create the trained dataset that would be used for feature matching. The obtained
13068 features are resized for size 66 × 198 matrix and which is given as input for CNN.

66 Volume 15 (2023), Issue 5

 Face Recognition Using Modified Histogram of Oriented Gradients and Convolutional Neural Networks

Fig. 4. Shows the m-HoG features values for the training dataset considering N number of face samples.

3.4 Convolution Neural Network

CNNs are the class of deep neural networks that are very effective for applications like pattern recognition and
classification. These CNNs are made up of many hidden layers and are of type feed forward network. CNNs contains
kernels that have trained weights and biases. Each kernel holds some input and performs convolution operation non-
linearly. A CNN architecture can be shown in Figure 5 and it consists of different layers like Convolution, pooling,
ReLU and fully connected layer.
Following the extraction of m-HoG features, CNN is employed in the study to extract further face features. The
proposed CNN architecture is shown in Figure 6 and it is made up of two convolutional layers followed by batch
normalization layers, two maximum pooling layers, and a final fully connected layer.
3.4.1 Convolutional layer
This layer consists of one or more convolutional feature extractors. Each feature extractor consists of an input
matrix x of size N×N, convolutional kernel w of size K×K and the bias b .The output of convolutional feature
extractor is a matrix h and it can be mathematically expressed as

h f conv x.w b

In which, f * ?is an activation function, conv * ?is a convolution.

In the proposed work, the output feature matrix of size 66×198 obtained by m-HoG technique for the input face
image is used as the input matrix ({for the CNN model. The model is composed of two convolutional layers, the first of
which contains four convolutional kernels of size 5×5 and the second of which contains sixteen convolutional kernels of
size 7×7.
3.4.2 Pooling layer
The pooling layer is also known as the subsampled layer. This layer is usually placed after the convolutional layer,
can collect features and minimize input image size. Averaging and Max Pooling are two common types of pooling
layers. The proposed CNN model employs Max Pooling to down sample the input image. The mathematical
expressions for the pooling layer are as follows:

l ,l
F Dmax P

l ,l
P Is the input vector, Dmax denotes the maximum pooling of the l and Fd enotes the output pooling face. The
proposed model implements 2×2 pooling filter and the size of feature extractor could be reduced to half of its original
size.
3.4.3 Rectified Linear Unit (ReLU)
A popular activation function, the rectified linear unit (ReLU), yields unsigned data type values. The ReLU layer
eliminates the negative values present in the feature maps by using a threshold value. Mathematically it can be
expressed as

Volume 15 (2023), Issue 5 67

 Face Recognition Using Modified Histogram of Oriented Gradients and Convolutional Neural Networks

r BKNQ O r
B: Q; L \
Q BKNQ R r

Where u represents the value of input matrix.

3.4.4 Fully-connected layer
This layer is similar to a classifier in that it is made up of neuron layers, with each neuron in one layer coupled to
every neuron in the layer before it. It can be stated mathematically as follows:

J f r a b

where r denotes input vector, a and b denotes the fully-connected layer's weight and bias, respectively, J is the output
of fully-connected layer, and f * is the activation function.

Fig. 5. Typical Convolutional Neural Network architecture

Input feature Matrix, 66 × 198

5 × 5 Conv, 4

Batch Norm, ReLU

Max Pool, 2 × 2

7 × 7 Conv, 16

Batch Norm, ReLU

Max Pool, 2 × 2

FC, 100

Fig. 6. Proposed Convolutional Neural Network architecture

4. Results and Discussions

After extracting the m-HoG features from the input face image, the size of the each image feature matrix is
changed to 66×198. We implemented different tests by making selecting different batch size. The proposed method is
implemented on MATLAB 2016a since it has built-in CNN library. The minimum batch size is set to 16, 32 and 64 and
the recognition rate is evaluated for all the batch sizes. The learning rate of the network is set to 0.001. The gradient
decent optimization algorithm is used. The performance of proposed face recognition system has been assessed on L
Spacek, JAFFE, NIR and YALE databases by computing recognition rate.

68 Volume 15 (2023), Issue 5

 Face Recognition Using Modified Histogram of Oriented Gradients and Convolutional Neural Networks

4.1 L-Spacek database

The Epoch is set to 50 to increase the reliability of the experimental findings, where one epoch is equal to training
all 1000 database images once. To enhance the accuracy of training samples of this database, all 1000 images are
trained 50 times in the experiment. The experiment is carried out using batch sizes of 16, 32, and 64. It is found that
better recognition accuracy is achieved for a batch size of 32.
Figure 7 (a) and (b) shows the Accuracy and loss (error) graphs obtained for both training and testing (validation)
features. Table 1 shows the values tabulated for validation accuracy and validation loss for different epochs. It is
noticed that the system accuracy is over 90% from epoch 1 and reaches 99.80% with minimal validation loss of 0.0098
for 23 epochs. After a specified number of training sessions, the training accuracy reaches 100% and the test
recognition rate remains consistent. This model's optimal test recognition rate for trained L-Spacek database is 99.80%.

(a) (b)
Fig. 7. (a) Shows Accuracy graphs for L-Spacek database (b) Shows Loss graphs for L-Spacek database

Table 1. shows the values tabulated for L-Spacek database for different epochs

Epoch Iteration Validation Accuracyin % Validation Loss Mini-Batch Accuracy in % Mini-Batch Loss
1 1 5.20 5.2135 3.13 5.1352
1 31 95.60 0.2272 96.88 0.3361
2 62 98.60 0.0556 100.00 0.1938
3 93 99.80 0.0155 96.88 0.0053
4 124 99.80 0.0146 100.00 0.0039
5 155 99.80 0.0137 100.00 0.0019
6 186 99.80 0.0133 100.00 0.0019
7 217 99.80 0.0127 100.00 0.0019
8 248 99.80 0.0123 100.00 0.0014
9 279 99.80 0.0120 100.00 0.0021
10 310 99.80 0.0118 100.00 0.0024
11 341 99.80 0.0115 100.00 0.0010
12 372 99.80 0.0113 100.00 0.0011
13 403 99.80 0.0111 100.00 0.0008
14 434 99.80 0.0110 100.00 0.0021
15 465 99.80 0.0108 100.00 0.0018
16 496 99.80 0.0106 100.00 0.0013
17 527 99.80 0.0104 100.00 0.0007
18 558 99.80 0.0103 100.00 0.0008
19 589 99.80 0.0102 100.00 0.0035
20 620 99.80 0.0101 100.00 0.0012
21 651 99.80 0.0100 100.00 0.0006
22 682 99.80 0.0099 100.00 0.0009
23 713 99.80 0.0098 100.00 0.0010

Volume 15 (2023), Issue 5 69

 Face Recognition Using Modified Histogram of Oriented Gradients and Convolutional Neural Networks

4.2 NIR dataset

The Epoch is set to 50 to increase the reliability of the experimental findings, where one epoch is equal to training
all 175 database images once. To enhance the accuracy of training samples of this database, all 175 images are trained
50 times in the experiment. The experiment is carried out using batch sizes of 16, 32, and 64. It is found that better
recognition accuracy is achieved for a batch size of 16.
Figure 8 (a) and (b) shows the Accuracy and loss (error) graphs obtained for both training and testing (validation)
features. Table 2 shows the values tabulated for validation accuracy and validation loss for different epochs. It is
noticed that the system accuracy is over 90% from epoch 12 and reaches 91.43% with minimal validation loss of 0.5321
for 13 epochs. After a specified number of training sessions, the training accuracy reaches 100% and the test
recognition rate remains consistent. This model's optimal test recognition rate for trained NIR database is 91.43%.

(a) (b)
Fig. 8. (a) Shows Accuracy graphs for NIR database (b) Shows Loss graph for NIR database

Table 2. Shows the values tabulated for NIR database for different epochs.

Epoch Iteration Validation Accuracy in % Validation Loss Mini-Batch Accuracy in % Mini-Batch Loss
1 1 5.71 5.7488 6.25 4.3789
1 11 25.71 4.1943 18.75 5.3290
2 22 48.57 2.2104 56.25 1.6501
3 33 65.71 1.4960 81.25 1.1888
4 44 85.71 0.5860 100.00 0.1583
5 55 85.71 0.5887 93.75 0.1311
6 66 88.57 0.4909 100.00 0.0196
7 77 91.43 0.4682 100.00 0.0211
8 88 85.71 0.3466 100.00 0.0166
9 99 88.71 0.4563 100.00 0.0204
10 110 88.57 0.4924 100.00 0.0312
11 121 88.57 0.6041 100.00 0.0063
12 132 91.43 0.5879 100.00 0.0064
13 143 91.43 0.5321 100.00 0.0117

4.3 JAFFE database

The Epoch is set to 50 to increase the reliability of the experimental findings, where one epoch is equal to training
all 100 database images once. To enhance the accuracy of training samples of this database, all 100 images are trained
50 times in the experiment. The experiment is carried out using batch sizes of 16, 32, and 64. It is found that better
recognition accuracy is achieved for a batch size of 32.
Figure 9 (a) and (b) shows the Accuracy and loss (error) graphs obtained for both training and testing (validation)
features. Table 3 shows the values tabulated for validation accuracy and validation loss for different epochs. It is
noticed that the system accuracy is over 90% from epoch 14 and reaches 95% with minimal validation loss of 0.0025
for 30 epochs. After a specified number of training sessions, the training accuracy reaches 100% and the test
recognition rate remains consistent. This model's optimal test recognition rate for trained JAFFE database is 95%.

70 Volume 15 (2023), Issue 5

 Face Recognition Using Modified Histogram of Oriented Gradients and Convolutional Neural Networks

Fig. 9. (a) Shows Accuracy graphs for JAFFE database (b) Shows Loss graphs for JAFFE database

Table 3. shows the values tabulated for JAFFE database for different epochs

Epoch Iteration Validation Accuracy in % Validation Loss Mini-Batch Accuracy in % Mini-Batch Loss
1 1 10.00 6.2005 3.13 3.3891
1 3 20.00 6.0339 37.50 6.1608
2 6 40.00 4.1914 40.63 4.5667
3 9 75.00 2.8158 75.00 2.5437
4 12 75.00 0.8419 87.50 0.4917
5 15 75.00 1.4347 87.50 0.9478
6 18 75.00 1.5400 87.50 0.8661
7 21 75.00 1.2882 96.88 0.4005
8 24 80.00 0.8498 96.88 0.1649
9 27 90.00 0.6335 100.00 0.0454
10 30 90.00 0.4934 100.00 0.0266
11 33 90.00 0.4013 100.00 0.0243
12 36 90.00 0.3455 100.00 0.0140
13 39 90.00 0.3114 100.00 0.0181
14 42 95.00 0.2904 100.00 0.0221
15 45 95.00 0.2775 100.00 0.0157
16 48 95.00 0.2700 100.00 0.0078
17 51 95.00 0.2652 100.00 0.0148
18 54 95.00 0.2623 100.00 0.0062
19 57 95.00 0.2601 100.00 0.0086
20 60 95.00 0.2589 100.00 0.0051
21 63 95.00 0.2574 100.00 0.0048
22 66 95.00 0.2556 100.00 0.0066
23 69 95.00 0.2542 100.00 0.0077
24 72 95.00 0.2533 100.00 0.0032
25 75 95.00 0.2528 100.00 0.0031
26 78 95.00 0.2525 100.00 0.0028
27 81 95.00 0.2525 100.00 0.0053
28 84 95.00 0.2529 100.00 0.0038
29 87 95.00 0.2532 100.00 0.0040
30 90 95.00 0.2534 100.00 0.0025
31 93 95.00 0.2537 100.00 0.0038

Volume 15 (2023), Issue 5 71

 Face Recognition Using Modified Histogram of Oriented Gradients and Convolutional Neural Networks

4.4 YALE database

The Epoch is set to 50 to increase the reliability of the experimental findings, where one epoch is equal to training
all 135 database images once. To enhance the accuracy of training samples of this database, all 135 images are trained
50 times in the experiment. The experiment is carried out using batch sizes of 16, 32, and 64. It is found that better
recognition accuracy is achieved for a batch size of 16.
Figure 10 (a) and (b) shows the Accuracy and loss (error) graphs obtained for both training and testing (validation)
features. Table 4 shows the values tabulated for validation accuracy and validation loss for different epochs. It is
noticed that the system accuracy is over 90% from epoch 6 and reaches 93.33% with minimal validation loss of 0.2614
for 14 epochs. After a specified number of training sessions, the training accuracy reaches 100% and the test
recognition rate remains consistent. This model's optimal test recognition rate for trained YALE database is 93.33%.

Fig. 10. (a) Shows Accuracy graphs for YALE database (b) Shows Loss graphs for YALE database

Table 4. Shows the values tabulated for YALE database for different epochs.

Validation Accuracy Mini-Batch

Epoch Iteration Validation Loss Mini-Batch Loss
in % Accuracy in %
1 1 6.67 7.2130 0.00 3.7148
1 8 20.00 6.8748 18.75 7.4419
2 16 53.33 1.4699 68.75 1.8690
3 24 53.33 1.3547 87.50 1.2084
4 32 73.33 0.9257 87.50 0.3331
5 40 86.67 0.4679 100.00 0.0134
6 48 93.33 0.3167 100.00 0.0298
7 56 86.67 0.2778 100.00 0.0041
8 64 86.67 0.2818 100.00 0.0011
9 72 86.67 0.2663 100.00 0.0052
10 80 86.67 0.2551 100.00 0.0013
11 88 86.67 0.2552 100.00 0.0208
12 96 86.67 0.2568 100.00 0.0114
13 104 93.33 0.2620 100.00 0.0006
14 112 93.33 0.2614 100.00 0.0013
15 120 93.33 0.2708 100.00 0.0192
16 128 93.33 0.2680 100.00 0.0009

From Table 5, it is noticed that the proposed face recognition system using CNN along with m-HoG feature
extractor achieves better recognition rate. It results 99.80% accuracy for L-Spacek, 93.33% Yale dataset, 95% for
JAFFE dataset and 91.43% for NIR dataset respectively.

72 Volume 15 (2023), Issue 5

 Face Recognition Using Modified Histogram of Oriented Gradients and Convolutional Neural Networks

Table 5. Recognition Rate evaluated for different datasets for different epochs.

Batch Size Training Accuracy Recognition Rate Learning Rate

Jaffe Dataset
16 100 90.00
32 100 95.00 0.001
64 91.15 70.00
NIR Dataset
16 100 91.43
32 97.25 88.57 0.001
64 90.11 82.46
L-Spacek Dataset
16 100 99.40
32 100 99.80 0.001
64 99.90 98.80
YALE Dataset
16 99.26 93.33
32 99.26 73.33 0.001
64 98.52 80.00

Table 6 shows the comparison of the recogntion rate achieved by the conventinal methods with that of the
proposed method for different datasets.It is noticed that the proposed work results in better recognition rate compare to
these conventional methods. The proposed method implements modified Histogram oriented Gradient (m-HoG) for
extracting facial image features and deep learning CNN algorithm to further extract features and classify the face image.
As a result, our suggested method performs better in rejecting unknown faces and accepting known faces with the
lowest possible error rate

Table 6 Comparison of recognition rates (in %) for different database with previous studies

Author-Year Database Technique Classifier Recognition Rate

Surbhi Gupta et al. [28] 2021 L-Spacek(Face94) SIFT+SURF 87.3%
SVM 98.03%
Deepali Virmani et al. [29] 2019 L-Spacek(Face94) EmbedNet
KNN 97.83%
Srinivas Halvi et al. 30] NIR 1DTD FFBP NN 72%
Shreyas Saxena et al. [31]
NIR CNN Softmax 85.9%
2016
Srijith Rajeev et al., [32] 2019 IIT Delhi NIR Circular with HOG SVM 80.41%
STWT 86.67%
NIR
Sateesh Kumar H C et al. [33] DTCWT 90.24%
2017 STWT ED 87.5%
YALE
DTCWT 87.5%
YALE 92.12%
Farid Ayeche et al. [34] 2021 HDGG SVM
JAFFE 91.43%
Rahul Ravi et al. [35] 2020 JAFFE CNN Softmax 77.27
Sujay H S et al.[36] 2019 YALE LBP SVM 86.66%
Shucheng Huang et al. [37]
YALE EDLPP - 88.56%
2016
V. Betcy Thanga Shoba et al.
YALE SURF+HOG+MSER - 91.2%
[38] 2020
L-Spacek 99.40%
NIR 91.43%
Proposed Work m-HOG+CNN Softmax
JAFFE 95.00%
YALE 93.33%

4.5 Work Limitations and Future scope

Face recognition rate will decline as the face database expands. More public datasets can be found in the face
databases. As a result, the key direction of future research will be to maintain recognition rate stability in large
databases. Compression techniques can be used to compress the feature dimensions for the datasets containing large set
of face images and thereby increasing the speed of recognition. Face biometric algorithms can be developed to
recognize the face images present in the video sequences. A real time face recognition system can be developed using
the hardware module for portability.

Volume 15 (2023), Issue 5 73

 Face Recognition Using Modified Histogram of Oriented Gradients and Convolutional Neural Networks

5. Conclusion
In this paper, we propose a model based on architectures of convolutional neural networks (CNN) to human face
recognition problems. This model has 2 convolutional neuron layers (CONV) and 1 fully connected neuron layers (FC).
Modified Histogram oriented gradient (m-HoG) for extracting facial image features and deep learning. The face images
are classified using the CNN algorithm based on their features. The proposed hybrid approach overcomes the
performance degradation problem of using CNN models alone. Further, m-HOG features are extracted to avoid loss of
features along three direction and the obtained histogram features are uniformly spread over 0 to 180 degrees. The
obtained features are resized as a matrix having dimension 66×198 and which are passed to the four layers CNN
structure made up of convolutional, max pooling, convolutional, max pooling layers respectively for training and face
image classification is carried out by softmax layer. We applied our model with different datasets having color
variations, fading, contrast and brightness change. Performance analysis is measured in terms of recognition accuracy. It
is found that L-Spacek and JAFFE datasets yields better recognition rate of 99.80% and 95.00% respectively. Whereas
the datasets like YALE and NIR having similarity, pose and brightness variations, yields recognition rate of 93.33% and
91.43% respectively.

Acknowledgment

The research was supported by Visvesvaraya Technological University, Jnana Sangama, Belagavi – 590018,
Karnataka, India.

Conflicts of Interest

The authors have no conflicts of interest to declare.

Author’s
Contribution Statement

Raveendra K: Conceptualization, Investigation, Data collection, Writing – original draft, Analysis and
Interpretation of results and and Draft manuscript preparation. Ravi J: Study Conception, Design, Supervision and
Investigation on challenges.

References

[1] Yong Xu, Qi Zhu, Zizhu Fan, David Zhang, Jianxun Mi and Zhihui Lai,“Usingtheideaofthesparserepresentationtop
coarse to-finefacerecognition,”InformationSciences.Vol. –148,pp.
82,3 381 July 2013
[2] Yi Sun, Xiaogang Wang, Xiaoou Tang, “Deep learning face representation by joint identification
-verification, ” Advances in
Neural Information Processing Systems, Vol. 27, pp. 1988–1996, 2014.
[3] Jae Young Choi, “Spatial pyramid face feature representation and weighted dissimilarity match
recognition, ”
The Visual Computer: International Journal of Computer Graphics, 34(11), 1535–1549 (2018)
[4] Larry S. Yaeger, Richard F. Lyon, Brandyn J. Webb, “Effective Training of a Neural Network Character Classifier fo
Recognition,”AdvancesinNeuralInformationProcessing –813 Systems,
(1996) pp.078
[5] Alex Krizhevsky, Ilya Sutskever and Geoffrey E. Hinton, “ImageNet classification with deep convolutional n
Advances in Neural Information Processing Systems, pp. 1097–1105 (2012)
[6] GailiYue and Lei Lu, “Face recognition based on histogram izationequal and convolution neural network,” Intern
Conference on Intelligent Human-Machine Systems and Cybernetics, pp. 336-339, 2018.
[7] Muhtahir O. Oloyede, Gerhard P. Hancke and Herman C. Myburgh, “Improving Face Recognition Systems Using a Ne
Image Enhancement Technique, Hybrid Features and the Convolutional Neural Network,” IEEE- Access, vol. 6,
75191, 2018.
[8] Wenqi Wu, Yingjie Yin, Xingang Wang and De Xu, “Face Detection With Different Scales Based on”Faster
-CNN, IEEE R
Transactions on Cybernetics, vol. 49, no. 11, pp. 4017-4028, Nov. 2019.
[9] Xianzhang Pan, “Fusing HOG and convolutional neural network -temporal features spatial for video-based facial expression
recognition,”IETImageProcessing,vol.1Issue 4, pp.
1, 617 – 182, January 2020.
[10] Yanhong Zhang, Kun Shang, Jun Wang, Nan Li and MonicaM.Y. Zhang. “Patch strategyfor deep face recognition,” IET
Image Processing, Vol.12, Issue 5, pp. 819- 825, May 2018.
[11] Feng Cen and Guanghui Wang, “Dictionary Representation of Deep Features lusion-Robust for Occ
Face Recognition,” IEEE
Access, vol. 7, pp. 26595-26605, 2019.
[12] Pengfei Ke, Maoguo Cai, Hanmo Wang and Jialong Chen, “A novel face recognition algorithm based on the combin
LBPandCNN,”InternationalConferenceonSignalp. Processing,
539-543, 2018. p
[13] Fanzhi Kong, “Facial expression recognition method based on deep convolutional neural network comb
LBPfeatures,”PersonalandUbiquitousComputing,vol. –539,,2pp.
3 2019.
3 15
[14] Erfan Zangeneh, Mohammad Rahmati and Yalda Mohsenzadeh, “Low resolution face recognition using
-branch deepa two
convolutionalneuralnetworkarchitecture,”ExpertSystemswith Applications,
-11, January 2020. vol.,3pp. 91 1

74 Volume 15 (2023), Issue 5

 Face Recognition Using Modified Histogram of Oriented Gradients and Convolutional Neural Networks

[15] Raj Silwal, Abeer Alsadoon, P. W. C. Prasad, Omar Hisham Alsadoon & Ammar Al-Qaraghuli,“Anoveldeeplearningsystem
for facial feature extraction by fusing CNN and MB-LBP and using enhanced loss function,” Multimedia Tools a
Applications, vol, 79, pp. 31027–31047, 2020.
[16] Ashok Kumar Rai , Radha Senthilkumar and Aswin KumarR, “Combiningpixel selection with covariancesimilaritya
in hyper spectral face recognition based on convolution neural network,” Microprocessors -8, and Micro
2020.
[17] Leslie Ching Ow Tiong, Song Tae Kim and Yong Man Ro, “Multimodal Facial Biometrics Recognition: -stream Dual
Convolutional Neural Networks with Multi-featureFusionLayers,”ImageandVisionComputing,vol.1pp. 2 , .2 0
[18] Tripti Goel and R Murugan, “Deep Convolutional - Optimized Kernel Extreme Learning Machine Based Classifier for Face
Recognition,”ComputersandElectricalEngineering, vol.,8pp.
-12, 2020. 5 1
[19] Shuai Peng, Hongbo Huang, Weijun Chen, Liang Zhang and Weiwei Fang, “More trainable inception
-ResNet for face
recognition,”Neurocomputing,vol.-19, pp.
14, 9 2020.
[20] Xi Yin and Xiaoming Liu “Multi -Task Convolutional Neural Network for Pose-Invariant Face Recognition,” IEEE
Transactions on Image Processing, vol. 27, no. 2, pp. 964-975, Feb. 2018
[21] Yanan Wang, Tian Na, Xiaoning Song and Guosheng Hu, “Bi -directional CRC algorithm using CNN based features for face
classification,”AsianConferenceonArtificialIntelligenceTechnology,–Vol. ,Issue
8012 November
1462, 61p.7451 2018
[22] Hana Ben Fredj, Safa BouguezziandChokriSouani,“FaceRecognitioninUnconstrainedEnvironmentwithCNN,”T
Computer, vol. 37, pp. 217–226, 2021.
[23] M. Chandrakala and P. Durga Devi, “Two -stage Classifier for Face Recognition using HOG Features,” Materials
Proceedings, vol. 47, pp. 5771-5775, 2021.
[24] Htwe Pa Pa Win, Phyo Thu Thu Khine and Khin Nwe Ni Tun, “ Face Recognition System based on Convolution Neura
Networks,”InternationalJournalonImage,GraphicsandSignalProcessing, -29, 2021. vol.,pp. 6 23
[25] Peng Lu, Baoye Song and Lin Xu, “Human face recognition based on convolutional neural network and augmen
Systems Science & Control Engineering, vol. 9, pp. 29-37, 2021.
[26] Walid Hariri, “Efficient masked face recognition method during the” COVID
-19 pandemic, Signal, Image and Video
Processing, vol. 16, pp. 605–612, 2021.
[27] Paul Violaand Michael J. Jones, “Robust -TimeReal
Face Detection,”International Journal ofComputerVision, Vol.,57N
2, pp. 137-154, 2004.
[28] Surbhi Gupta, Kutub Thakur and Munish Kumar, “2D -human face recognition using SIFT and SURF descriptors of face’
featureregions,”TheVisualComputer,vol. –456,
3pp.
7, 74 2021.
[29] Deepali Virmani, Palak Girdhar, Prateek Jain and Pakhi Bamdev, “FDREnet: Face Detection and Recogniti
Engineering, Technology & Applied Science Research, vol. 9, pp. 3933-3938, 2019.
[30] Srinivas Halvi, Nayina Ramapur , K B Raja and Shanti Prasad, “Fusion Based Face RecognitionSystemusing 1D Tr
Domains,” International Conference on Advances ng andin Computi
Communications, Procedia Computer Science, Vol. 115,
pp.383–390, 2017
[31] ShreyasSaxenaandJakobVerbeek,“HeterogeneousFaceRecognitionwithCNNs,”ECCVWorkshops, -491, 2016. pp.834
[32] Srijith Rajeev, Shreyas Kamath K.M, Qianwen Wan and Karen Panetta, “Illumination Invariant NIR Face Recognition usi
DirectionalVisibility,”ElectronicImaging, -1-273-7,pp.2019.
32 7
[33] Sateesh Kumar H C, C Chowda reddy, Raja K B and Venugopal K R “Face Recognition based on STWT and DTCWT using
two dimensional Q-shift Filters,”InternationalJournalofEngineeringResearchandApplication, -69, 2017. vol.pp.
7, 46
[34] Farid Ayeche and Adel Alti, “HDG and HDGG: an extensible feature extraction descriptor for effective
expressions,”PatternAnalysisions, andvol.Applicat
24, pp. 1095–1110, 2021.
[35] RahulRavi,YadhukrishnaS.VandRajalakshmiprithviraj,“AFaceExpressionRecognitionUsingCNN&LBP,”Inte
Conference on Computing Methodologies and Communication, pp. 684-689, 2020.
[36] Sujay S N and H S Manjunatha Reddy, “Extended Local Binary Pattern Features based Face Recognition usin
SVMClassifier,”vol.8,pp.2431 -4128, 2019.
[37] Shucheng Huang and Lu Zhuang, “Exponential Discriminant Locality Preserving Projection for Face
Neurocomputing, vol. 208, pp. 373-377, 2016.
[38] V. Betcy Thanga Shoba and I. Shatheesh Sam, “Hybrid Features Extraction on Face for Efficient Face Recog
Multimedia Tools and Applications, vol. 79, pp. 1-22, 2020.

Authors’ Profiles

Dr. Raveendra K is a Assistant Professor and HOD in the Department of Electronics and Communication
Engineering, Government Engineering College, K R Pet. He obtained his BE Degree in Electronics and
Communication Engineering from Kuvempu University, Shimoga. His specialization in Master degree was
Digital Communication Engineering from VTU, Belagavi. He was awarded PhD in Electrical and Electronics
Engineering Sciences from VTU Belagavi. His area of interest is in the field of Image Processing, Signal
Processing and Biometrics.

Volume 15 (2023), Issue 5 75

 Face Recognition Using Modified Histogram of Oriented Gradients and Convolutional Neural Networks

Dr. Ravi J is a Professor in the Department of ECE, GAT, Bengaluru. He received his BE Degree in
Instrumentation Technology from Bangalore University, Bengaluru. He obtained his Master degree in Digital
Electronics from VTU, Belagavi. He was awarded PhD in Electronics and Communication Engineering from
JNTU Anantapur. He has over 30 peer-reviewed research articles in international journals and conference
proceedings. His research interests include Image Processing, Biometrics VLSI and computer networks. He is life
member of IETE, ISTE, and ISAMP.

Dr. Khalid Nazim Sattar Abdul is a Assistant Professor in the Computer Science and Information Department,
College of Science, Al-Zulfi Campus, Majmaah University. He received his BE Degree in Computer Science &
Engineering from Ghousia College of Engineering, Bangalore University, Bangalore. He obtained his master’s
degree in computer science & Engineering from VTU, Belagavi. He was awarded PhD in Computer Science &
Engineering from School of Computer Science, Singhania University, Rajasthan. He has over 33 peer-reviewed
research articles in international journals and conference proceedings. His research interests include Image
processing, Software Engineering, Information Systems, Computer Networks.

How to cite this paper: Raveendra K, Ravi J, Khalid Nazim Abdul Sattar, "Face Recognition Using Modified Histogram of
Oriented Gradients and Convolutional Neural Networks", International Journal of Image, Graphics and Signal Processing(IJIGSP),
Vol.15, No.5, pp. 60-76, 2023. DOI:10.5815/ijigsp.2023.05.05

76 Volume 15 (2023), Issue 5