Volume 6, Issue 6, June – 2021 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Cross-Domain Face Recognition for Criminal

Identification
Adhert Johnson Adithyan Ranjan K C
Student, Computer science and Engineering Department Student, Computer science and Engineering Department
Sahrdaya College of Engineering Technology Sahrdaya College of Engineering Technology
Thrissur, India Thrissur, India

Christy Vargheese Akhil A S

Student, Computer science and Engineering Department Student, Computer science and Engineering Department
Sahrdaya College of Engineering Technology Sahrdaya College of Engineering Technology
Thrissur, India Thrissur, India

Princy T D
Asst. Professor, Computer science and Engineering Department
Sahrdaya College of Engineering Technology
Thrissur, India

Abstract:- The problem of cross-domain face recognition spectra images (VIS) with near-infrared images (NIR),
seeks to identify facial images obtained from different because near-infrared images are resistant to changing
domains, and it is gaining popularity due to its numerous lighting conditions. Given that the VIS-NIR faces are
applications in law enforcement identification and captured by different sensors, there is also a significant
camera surveillance Existing algorithms typically fail to domain difference between them. Cross-domain face
fully exploit semantic information for identifying cross- recognition, also referred to as heterogeneous face
domain faces, which could be a strong clue for recognition [4] seeks to match face images across domains.
recognition. In this paper, we present an efficient The training subset usually consists of a variety of face
algorithm for cross-domain face recognition that makes image pairs from two different domains, while the
use of semantic information in conjunction with deep remaining face image pairs are used for testing. Due to the
convolutional neural networks (CNN). We start with a various domain discrepancies, cross domain face recognition
soft face parsing algorithm that measures the boundaries may be a difficult task. In recent years, a number of cross-
of facial components as probabilistic values. For cross domain face recognition algorithms have been developed.
domain face recognition, we propose a hierarchical soft Early cross-domain face recognition methods typically used
semantic representation framework. CNN-derived deep hand-crafted architectures to either transform face images
features are computed and combined. Which could fully into the same domain using face synthesis techniques [5] or
exploit the same semantic clue across cross-domain directly match cross-domain faces using a standard subspace
faces. We present extensive experiments to show that the [6] or hand-crafted feature descriptors. Hand-crafted
proposed soft semantic representation algorithm approaches, on the other hand, are frequently
outperforms state-of-the-art methods. computationally expensive, hence their performance is
usually limited. With the rapid advancement of deep
I. INTRODUCTION convolutional neural networks (CNN) in classification tasks,
a growing number of deep learning-based cross-domain face
FACE images are frequently captured from various recognition approaches have recently been developed. To
domains in various tasks. In an enforcement application, one eliminate the domain discrepancy, an end-to-end deep
common scenario is to match face photos with sketches network was used for cross-domain face synthesis [8].
generated from supported eyewitness descriptions of the Cross-domain face recognition is also influenced by
suspect. Forensic artists are frequently invited to create a restricted Boltzmann Machines (RBMs) [9], CNN [10], and
hand-drawn sketch based on the description, or software is generative adversarial networks (GAN) [11].We offer an
frequently used to generate a composite sketch for the efficient soft semantic representation based framework
purpose of identifying the suspect. Face photos and sketches (SoftSR) in this research to utilise semantic information for
have a large domain difference because photos are captured crossdomain face recognition. We offer a soft face parsing
in a real-life environment, whereas both hand-drawn and technique that performs a soft classification on the pixels
software-generated sketches are created manually with around the facial component borders, taking into account the
artefact [1].Another cross-domain face recognition scenario disadvantages of existing approaches that use hard pixel
within the camera surveillance application is to match visual classification based semantic maps. We use the first face

IJISRT21JUN491 www.ijisrt.com 639

Volume 6, Issue 6, June – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
image as a guide to create a soft semantic map, in which analysis was incorporated into CNN by treating it as a
each pixel can belong to a semantic facial component in separate fully-connected layer capable of extracting domain-
part. For cross-domain face recognition, we also build a independent characteristics for recognition. To deal with
hierarchical soft semantic representation system..Both the thermal to visible face recognition and cross-resolution face
soft semantic component level and, as a result, the face recognition [49], Nasrabadi et al. developed attribute-guided
contour level are used to extract deep CNN features. algorithms. The variability between ID and spot
photographs, the bisample issue (where only two samples
They are then integrated to improve cross-domain face are given for each topic), and thus the vast scale classes data
identification performance. The equivalent semantic are all obstacles in recognising photographs from ID and
information across different domains is completely utilised spot photo domains. They proposed a classification-
in this manner. The following is a list of the article's verification-classification pipeline to transfer models in a
contributions : progressive manner. Deng et al. suggested a two-branch
Residual networks architecture for cross-domain face
1) For cross-domain face recognition, we offer a soft face recognition, featuring a residual compensation module and
parsing technique that takes advantage of semantic modality discrepancy loss, and they produced promising
information. The pixels in our soft semantic map, which can results on both sketch-photo matching and NIR-VIS
deal with erroneous segmentation around semantic face matching tasks. Yu et al. recently proposed using a
borders, are partially assigned to the component as generative adversarial network to turn a facial sketch into a
probabilistic values. photo with a loss at the feature level. By incorporating
2) To improve the performance of cross-domain face image generation for matching NIR and VIS faces, a variety
recognition, we develop a hierarchical soft semantic of ways were presented. However, The semantic clue was
representation framework that is coupled with deep ignored by the majority of existing cross-domain face
convolutional neural networks. recognition algorithms. According to our observations, the
3) We run extensive experiments on multiple cross-domain semantic information of facial components is generally
face datasets to test the proposed technique and show that it similar across domains (with the exception of caricature
outperforms state-of-the-art approaches. recognition, which involves shape exaggeration). We will
develop cross-domain face correlations and propose a soft
II. RELATED WORK semantic representation strategy that is robust across picture
domains based on this fact. Unlike the component-based
Here we look at some of the related works in the method (CBR), which relied on difficult component
field of Face Recognition. segmentation and hand-crafted characteristics, machine
learning methods provide the necessary tools. Model
A. Hand -crafted-method validation is carried out by 5-fold cross-validation, which
Face identification across domains began with an divides the dataset into ve subsets and uses one of the
eigen transformation-based method that converted face subsets as the testset and the other four as the training set.
pictures into sketches, and then a typical face recognition
algorithm was used to match synthesised sketches with III. PROPOSED SYSTEM
sketches in the target gallery. For cross-domain face
alteration, Wang et al. combined sparse feature selection We offer our proposed soft semantic representation for
with support vector regression. A number of Markov cross-domain face recognition in this part. We define X =
network-based approaches [5] have been presented in recent X1, X1, X M as a collection of M face photos in two
years to remodel cross-domain faces into a common domain domains, and Y = Y1, Y1, YM as a collection of M face
for recognition. Cross-domain face identification started photos in two domains. There are two phases to our process.
with an eigen transformation-based approach that converted To begin, the photos in both domains are subjected to a soft
face pictures into the sketch domain. Matching synthesised semantic face parsing algorithm, which provides a soft
sketches with sketches in the target gallery was done using a semantic parsing map for each face picture. Second, using
typical face recognition algorithm. For cross-domain face the soft face semantic map as a guide, we create a
morphing, Wang et al. [25] combined sparse feature hierarchical framework for extracting deep CNN features
selection with support vector regression. A variety of from both the soft semantic and contour levels. For cross-
Markov network-based strategies [5] have been proposed to domain face recognition difficulties, using a step-by-step
be revised in recent years. procedure is a common solution. Face photos, for example,
are processed first by a face completion method before
B. Methods based on deep learning being put into a CNN for cross-domain face matching. To
The intrinsic distribution of raw pixels is frequently obtain a new image pattern, first learn a discriminative face
investigated directly as deep learning progresses. To representation, and then project the filtered images into one
produce high-quality synthesised faces, an end-to-end cross- common subspace for identification. Similarly, to tackle
domain face synthesis approach [8] was developed. RBMs, cross-domain face recognition, we provide a two-phase
CNNs, and GANs were also used to influence cross-domain methodology. We'll go over each phase in detail, and for the
face recognition [9].Galea and Farrugia recommended that sake of brevity, we'll use face sketch-photo identification as
3D morphable models be fitted to face pictures and sketches an example.
so that fresh images may be synthesised for training data
augmentation. Later, it was developed to a super-deep CNN
with morphing faces and transfer learning [1] for matching
software-generated sketches to images. Mutual component

IJISRT21JUN491 www.ijisrt.com 640

Volume 6, Issue 6, June – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
A. Semantic Face Parsing with Soft Semantics boundaries become missing.
Semantic face parsing map as a cross-domain clue
for our recognition objective, inspired by earlier When the spatial parameter s is extremely big,
publications [12] on applying semantic face parsing to face- however, the parsing map will not be over smoothed. As a
related tasks. We use established ways to divide facial result, the default parameter setting for r must be a suitable
features into four categories: brows, eyes, nose, and mouth. value, whereas the default parameter setting for s can simply
We first recognise facial landmarks in a face picture (which be an outsized value. Because the proposed parsing
could be a photo or a sketch image) using convolutional method's computation time is exceedingly long,
experts restricted local models. It is possible to employ
dense facial landmarks to guide the creation of a semantic When the spatial parameter s is extremely big,
parsing map. The semantic face parsing map can be however, the parsing map will not be over smoothed. As a
generated in a straightforward method by linking result, the default parameter setting for r must be a suitable
surrounding landmarks around each facial component. Each value, whereas the default parameter setting for s can simply
polygon's pixels are given to the appropriate facial be an outsized value. Because the suggested parsing method
component category. The semantic region for the left has a very short processing time, which will be explained in
eyebrow, for example, is defined as the pixels with the complexity analysis chapter, tweaking the two
associated landmarks around the left eyebrow. However, parameters isn't very time demanding. For a variety of cross-
because there are some long-distance surrounding domain face recognition challenges,
landmarks, this direct connection operation may create a
piecewise linear effect. To address this problem, we use
cubic spline interpolation to smooth the component's
boundary. Finally, we might create an initial semantic
parsing map for each face image, with pixels within facial
component regions marked as 1 and the rest as 0.

However, the initial semantic parsing map suffers from

the same flaw as previous works in that it performs a hard
pixel classification around component borders. The pixels
on the inside of the boundary are labelled 1 while those on
the outside are labelled 0. However, the component's border
may be erroneous, which is overlooked in current
methodologies. For example, the nose bridge's boundary
may be ambiguous, and the above hard pixel classification
may provide misleading information for our next phase. As
a result, in this paper, we describe a soft semantic face
parsing system in which pixels near facial component
boundaries are classified as component pixels. We employ
each row/column in the initial semantic parsing map as the
input signal S, and every row/column in the input face
Fig. 1
image as the guidance signal P in our soft semantic face
parsing procedure. As a result, the sides of the input face
B. Cross-Domain Face Recognition Using Soft Semantic
image may be used to guide the semantic parsing map
Representation
smoothing. This vector shows the likelihood that a pixel
To tackle the domain discrepancy in cross-domain face
corresponds to at least one face component. To get the soft
recognition, we provide a hierarchical system providing soft
semantic level shape feature, we simply concatenate the R
semantic face parsing. Our technique integrates both soft
dimensional vectors in the least pixels together, which is
semantic level and contour level deep information, as
designated as follows: FS S = f (Xs m, R) (4), where FS S
illustrated in Fig. 1. The soft semantic level features are
denotes the soft shape feature that can be retrieved easily
independent of the domain discrepancy because they may
using the soft semantic parsing map Xs m as input.When
capture full information from both soft semantic component
there is a transparent border within the input face image,
forms and component contents. The former is frequently
such as an eye corner, the result is an outsized value and no
regarded as a high-level property of the face that is extracted
smoothing is performed. Xs m = SSFP(Xm , Xinim , σr, σs)
by extracting semantic characteristics at each pixel, whilst
(3)The soft semantic face parsing operation is denoted by
the latter can reflect the appearance of every semantic
SSFP. Here, the parameters s and r are used. The parameters
component. We also use contour level characteristics from
have an impact on the performance of our method. s is
the holistic face to provide a semantic description of the
frequently thought of as a spatial parameter introduced in
entire face, which are complementary, and the merger of
equation. It regulates the recursive edge-preserving filter's
these hierarchical deep features could enable efficient and
feedback coefficient. r is commonly thought of as a range
robust cross-domain face recognition performance. Below,
parameter that affects the smoothing operation's range area.
we go over the soft semantic representation in detail.
The smoothing phenomenon at the face component
1) Soft Semantic Level Feature: High-level semantic
boundaries becomes more noticeable as s and r grow. When
features have been shown to be successful in face labelling,
the range parameter r is set to a large value, the parsing map
face deblurring, and face synthesis, among other
becomes excessively smoothed, and useful face component
applications. We extract the soft semantic clue at each pixel

IJISRT21JUN491 www.ijisrt.com 641

Volume 6, Issue 6, June – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
position to describe whether the pixel partially belongs to a IV. EXPERIMENTS AND RESULTS
facial component category or not, because the semantic
information of facial components is usually consistent The experimental performance of our proposed
across faces from different domains. At each pixel, we method on numerous cross-domain face recognition
create an R-dimensional probabilistic vector, where R is the datasets, as well as comparisons to state-of-the-art
number of component categories. This vector shows the algorithms, are described in this section. Experiments are
likelihood that a pixel corresponds to at least one face performed using cross-domain face databases that are
component. To acquire the soft semantic level shape feature, publically available, such as the photo domain with sketch
we simply concatenate the R dimensional vectors in the least domain, UoM-SGFS [1] and a gathered forensic sketch
pixels together. database, the live photo domain with the ID photo domain,
2) Contour Level Feature: We will capture extensive and the VIS domain with the NIR domain. The e-PRIP
information from facial contour, which is robust to changes database has 123 photo-sketch pairs, with the photographs
in stance and expression, in addition to the soft semantic coming from the AR database and the sketches being
level feature, which concentrates on local components. The created using composite sketch software. The database is
key points in the initial semantic parsing map Xini m are divided into two subsets at random, each with 48 pairs. The
used to extract deep features with the help of the image remaining is for testing and classifier training. The PRIP-
patch level network described above. The contour level VSGC database has the same 123 pictures as the e-PRIP
feature, where FHC denotes the original hard semantic level database, but the sketches were made with a different
contour feature, concatenates the extracted deep features. composite sketch programme. For evaluation, the same
The contour level feature can supplement the soft semantic methodology used in the e-PRIP database is used. The
level feature by providing additional information. photographs for the 600 subjects in the UoM-SGFS database
3) Soft Semantic Representation Fusion: Each face were acquired from the FERET database. A composite artist
component category's soft semantic level features FS S and creates one sketch for each face pic.
FSC, as well as the contour level feature FHC, are retrieved
first. We combine the soft semantic level shape feature with One sketch is prepared for each face photo using the
the content feature retrieved at critical places inside this EFIT-V composite sketch software, and another sketch is
component region to create one soft semantic level created by further altering the formal sketch with the Corel
representation for each semantic facial component [FS S; PaintShop Pro-X7 software. As a result, EFIT-first V's
FSC]. We'll get six soft semantic level representations if composite sketches are collected as subset A, and the further
there are six component categories: left eyebrow, left eye, altered sketches are gathered as subset B after a short time.
right eyebrow, right eye, nose, and mouth. The contour level This article follows the standard evaluation process outlined
representation is formed by concatenating the deep features in. The forensic sketch database contains 168 forensic
extracted at the contour landmarks. The root normalises the sketch-photo pairs gathered from real-life law enforcement
elements of the six soft semantic level representations and, situations. Because the mug shot images are acquired in
as a result, the contour level representations, resulting in our varied environments and the forensic sketches are made
soft semantic representation. The set of emotive semantic according to eyewitness descriptions, this is a difficult face
representations is used for cross-domain face recognition. sketch-photo recognition database. We employ split
For starters, we create a classifier for each soft semantic protocols, which use 106 image pairings for training and the
level and contour level representation. Second, the rest for testing. .We follow the standard protocol to gauge
classifiers' scores are combined using a simple averaging our proposed soft semantic representation on matching faces
technique. Because the dimension of FSC and FHC is between live photo domain and ID photo domain. The
usually quite large, we use the principal component analysis CASIA NIR-VIS 2.0 database contain A total of 10,000
and Fisherface algorithms to map them into a discriminant additional photographs that have been added to the gallery
subspace. Further complex classifiers and fusion approaches to make it larger. There are 256 high-resolution live
could be used to improve the performance even more, and photographs in the NJU-ID database, along with
this is something that will be looked into in the future. In the accompanying low-resolution ID photographs. To test our
experimental section below, we'll go over the proposed soft semantic representation on matching faces
implementation details of our framework. between the live photo domain and the ID photo domain, we
use the standard protocol. There are 725 subjects in the
CASIA NIR-VIS 2.0 database, divided into two views.
View 1 is used for parameter adjustment, whereas view 2 is
utilised to evaluate performance. There are several facial
photos from both the VIS and NIR domains for each person.
In this study, we use a comparable experimental technique
given by databases, such as the UoM-SGFS database and
hence the CASIA NIR-VIS 2.0 database, if they are
accessible. Otherwise, we use a similar partitioning protocol
as described in the respective studies. We use 48 subjects for
training and 75 subjects for testing in the e-PRIP database
and the PRIP-VSGC database, according to the partition
protocol. For the forensic sketch database, we use an
experimental protocol in which 106 subjects are chosen at
random for training and the remaining 53 for testing. We use
Fig.2 a precise 10-fold cross validation technique for the NJU-ID

IJISRT21JUN491 www.ijisrt.com 642

Volume 6, Issue 6, June – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
database, dividing the database into 10 folds at random.

A. Implementation Specifications
On the e-PRIP database, we explain the
implementation details and parameter settings of our
proposed algorithm. All of the face photos used in this text
are first aligned and cropped to 200 by 250 pixels. The
parameters r and s are set to 0.4 and 40, respectively, during
the soft semantic face parsing phase. In the following
section, we'll look at the impact of those two variables. The
dimensions of picture patches utilised for extracting CNN
features in soft semantic representation extraction are
around 32 x 32. The important spots are sampled with a
density of 10 10 grid during the soft content feature
extraction step. The impact of those two parameters during Fig.3
the soft semantic feature extraction step will also be
examined further down. To examine facial character around
the keypoint, we merely require four image patches
surrounding each keypoint. The Brown dataset is used to
train the image patch level CNN. After that, we fine-tune the
cross-domain databases separately. The hard samples
approach is used, as well as data augmentation with random
rotation. We choose to utilise a conventional CNN rather
than a well-designed complicated module to showcase the
contribution of this text because the key contribution is to
employ the proposed soft semantic representation technique
for cross-domain face recognition. Fig.4

Our results show that the features extracted from the 2) The Soft Semantic Representation Extraction Parameters'
picture patch level CNN can perform well on cross-domain Influence:
face recognition tasks. In reality, within the task of image We need to extract CNN features of image patches
matching, the image patch level CNN must be present to sampled from cross-domain face images while using the
indicate significant potentiality and effectiveness. The suggested soft semantic face parsing algorithm for cross-
domain gap in our challenge can also be closed using picture domain face recognition. Because we use image patch level
patch level CNN as part of our network training technique, CNN for feature extraction, it's plausible to attribute the
which includes fine-tuning on cross-domain databases using performance to the image patch size choice. We test our
hard samples and data augmentation with random rotation. proposed method for feature extraction using different
picture patch sizes. It is difficult for CNN to extract
1) Soft Semantic Parsing Parameters: discriminative information from small picture patches when
There are two parameters r and s that need to be the image patch size is tiny. As a result, if we choose an
determined in our proposed soft semantic face parsing image patch size of 8/8 or 16/16, the performance will be
algorithm. In the graphic, we show a visual comparison of bad. With the increase in image patch size, more relevant
the impact of those two characteristics. If the borders inside and discriminative information, as well as semantic clues,
the input face image are obvious, our suggested soft are taken into account, which improves cross-domain face
semantic parsing algorithm can successfully smooth the recognition.
unclear facial component borders while keeping the strong
edge structure, as shown in the figure. Furthermore, we The computation cost for image patch level CNN, on
begin by setting s to 40. When r is small, no smoothing the other hand, increases as the image patch size grows. The
around the facial component boundaries is done. The soft picture patch size is set to 32 32 as the default setting in this
semantic parsing map gets highly smooth when r is large. article. The keypoint sampling density, which regulates the
Then, as seen in the second row of the picture, we set r to density of sampling key points within face components to
0.4. When s is small, there is no smoothing process. The extract picture patches for recognition,. We densely sample
findings are not over smoothed even when s is very large. key spots inside the region of face components in the soft
As a result, during this article, we set r and s to 0.4 and 40, content feature extraction phase and generate picture patches
respectively, to provide the simplest performance. around these key locations for deep feature extraction. The
keypoint sampling approach at various densities. When the
density is low, the image patches will be sampled using a
dense grid, which will increase the calculation cost. The
results are adequate. The effects of such factors are similar
for various cross-domain face recognition tasks. We
discovered that using the default parameters on different
datasets yields promising results, demonstrating the
method's generalisation capabilities.

IJISRT21JUN491 www.ijisrt.com 643

Volume 6, Issue 6, June – 2021 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
V. CONCLUSION [7]. HANAN ABDULLAH MENGASH, Using Data
Mining Techniques to Predict Student Performance to
Cross-domain face recognition may be a difficult task Support Decision Making in University Admission
due to the wide domain discrepancy. We provide a fully Systems March 30, 2020. Digital Object Identifier
unique soft semantic representation for cross domain face 10.1109/ACCESS.2020.2981905 Riyadh 11351, Saudi
recognition in this article. The first step is to create a soft Arabia
face parsing method that can produce a soft semantic [8]. AGORITSA POLYZOU AND GEORGE KARYPIS,
parsing map using the input face picture to avoid erroneous Feature Extraction for Next-Term Prediction of Poor
segmentation around component boundaries in the first hard Student Performance IEEE TRANSACTIONS ON
parsing map. After that, a hierarchical soft semantic LEARNING TECHNOLOGIES, VOL. 12, NO. 2,
representation extraction framework is described, with APRIL-JUNE 2019
sentimental semantic representation consisting of soft shape [9]. SANA BHUTTO,DR. QASIM ALI
features, soft content features, and hard semantic level ARAIN,MALEEHA ANWAR, Predicting Students'
contour features. The proposed strategy achieves the Academic Performance Through
simplest performance on numerous cross-domain face [10]. Supervised Machine Learning 2020 International
datasets by merging the collection of emotive semantic Conference on Information Science and
representations. The proposed soft semantic parsing method Communication Technology
will be applied to additional fields in the future. We devote a [11]. Y.-H. HU, C.-L. LO, AND S.-P. SHIH, ‘‘Developing
separate subsection to the suggested soft semantic face early warning systems to predict students’ online
parsing, which may alternatively be thought of as a pre- learning performance,’’ Comput. Hum. Behav., vol.
processing phase for face recognition. Uncertain visual 36, pp. 469–478, Jul. 2014.
boundaries around facial components are frequently [12]. M.ZIĘBA, S. K. TOMCZAK, AND J. M.
softened, resulting in incorrect segmentation. This is TOMCZAK, ‘‘Ensemble boosted trees with synthetic
beneficial for improved cross-domain face representation. features generation in application to bankruptcy
We believe that the features of smoothing incorrect prediction,’’ Expert Syst. Appl., vol. 58, pp. 93–101,
segmentation near ambiguous boundaries while retaining Oct. 2016.
edges at clear boundaries of a face picture are frequently [13]. C. PENG, X. GAO, N. WANG, and J. LI, “Graphical
integrated into other related tasks such as face deblurring, representation for heterogeneous face recognition,”
face denoising, and face synthesis. IEEE Trans. Pattern Anal. Mach. Intell., vol. 39, no. 2,
pp. 301–312, Feb. 2017
REFERENCES [14]. Y. SONG ET AL., “Joint face hallucination and
deblurring via structure generation and detail
[1]. C. GALEA and R. A. FARRUGIA, “Matching enhancement,” Int. J. Comput. Vis., vol. 127, nos. 6–7,
software-generated sketches to face photographs with pp. 785–800, Jun. 2019
a very deep CNN, morphed faces, and transfer
learning,” IEEE Trans. Inf. Forensics Security, vol. 13,
no. 6, pp. 1421–1431, Jun. 2018
[2]. Y. SHI and A. K. JAIN, “DocFace: Matching ID
document photos to selfies,” in Proc. IEEE 9th Int.
Conf. Biometrics Theory, Appl. Syst., Oct. 2018, pp.
56–67
[3]. H. ZHANG, H. HAN, J. CUI, S. SHAN, and X.
CHEN, “RGB-D face recognition via deep
complementary and common feature learning,” in
Proc. 13th IEEE Int. Conf. Autom. Face Gesture
Recognit., May 2018, pp. 1–8
[4]. X. WANG and X. TANG, “Face photo-sketch
synthesis and recognition,” IEEE Trans. Pattern Anal.
Mach. Intell., vol. 31, no. 11, pp. 1955–1967, Nov.
2009.
[5]. ABDULLAH ALSHANQITI AND ABDALLAH
NAMOUN, Predicting Student Performance and Its
Influential Factors Using Hybrid Regression and
Multi-Label Classification Digital Object Identifier
November19,2020.Digital Object Identifier
10.1109/ACCESS.2020.3036572 Madinah 42351,
Saudi Arabia
[6]. RAMIN GHORBANI AND ROUZBEH GHOUS,[1]
Comparing Different Resampling Methods in
Predicting Students’ Performance Using Machine
Learning Techniques Digital Object Identifier
10.1109/ACCESS.2020 April 22, 2020. Digital Object
Identifier 10.1109/ACCESS.2020.2986809 Tehran
16846-13114, Iran

IJISRT21JUN491 www.ijisrt.com 644