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

net/publication/334783128

Deep Learning For Face Recognition: A Critical Analysis

Preprint · July 2019

CITATIONS READS
0 3,609

1 author:

Andrew Shepley
University of New England (Australia)
8 PUBLICATIONS   2 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Andrew Shepley on 29 October 2019.

The user has requested enhancement of the downloaded file.

 DEEP LEARNING FOR FACE RECOGNITION: A
CRITICAL ANALYSIS

ANDREW JASON SHEPLEY

Charles Darwin University,

Ellengowan Drive, Casuarina, Darwin, Australia

Email: ashepley@myune.edu.au

ABSTRACT
Face recognition is a rapidly developing and widely applied aspect of biometric technologies. Its
applications are broad, ranging from law enforcement to consumer applications, and industry efficiency
and monitoring solutions. The recent advent of affordable, powerful GPUs and the creation of huge face
databases has drawn research focus primarily on the development of increasingly deep neural networks
designed for all aspects of face recognition tasks, ranging from detection and preprocessing to feature
representation and classification in verification and identification solutions. However, despite these
improvements, real-time, accurate face recognition is still a challenge, primarily due to the high
computational cost associated with the use of Deep Convolutions Neural Networks (DCNN), and the need
to balance accuracy requirements with time and resource constraints. Other significant issues affecting
face recognition relate to occlusion, illumination and pose invariance, which causes a notable decline in
accuracy in both traditional handcrafted solutions and deep neural networks. This survey will provide a
critical analysis and comparison of modern state of the art methodologies, their benefits, and their
limitations. It provides a comprehensive coverage of both deep and shallow solutions, as they stand today,
and highlight areas requiring future development and improvement. This review is aimed at facilitating
research into novel approaches, and further development of current methodologies by scientists and
engineers, whilst imparting an informative and analytical perspective on currently available solutions to
end users in industry, government and consumer contexts.

Keywords – facial recognition, face detection, feature extraction, face verification, fiducial point,
face alignment, convolutional neural networks, boosting, deep neural networks, video-based face
recognition, infrared face recognition

I. INTRODUCTION
Biometric recognition software plays an increasingly significant role in modern security,
administration and business systems. Biometrics include fingerprint, retinal scanning, voice
identification and facial recognition. Facial recognition has attracted particular interest, as it
provides a discreet, non-intrusive means of detection, identification and verification, without the
need for the subject’s knowledge or consent. It is now commonplace in applications such as airport
security, and has been widely embraced by law enforcement agencies, due to the improving
accuracy and deployability of systems, and the growing size of face databases. A recent example
of its success occurred in China; face recognition technologies were successfully used to identify
and track a wanted fugitive at a concert attended by over 60000 people, resulting in his arrest [2].
As a consequence of the proven ability of deep neural network based systems to outperform human
performance in face verification tasks, the international government sector is projected to be the
largest user of face biometrics systems, including face recognition [3], over the next 10 years. Face
recognition technologies are also at the forefront of a wide range of consumer applications and
devices, from user verification tasks enabling account access, to digital camera applications, and
social media tagging. Consumers and industry require rapid, affordable and efficient applications,
to meet demand in business, employment and education functions, which extend to employee
monitoring, roll call and security, reducing administrative costs, and procedural efficiency.
Despite significant progress in recent years, and widespread usage, many shortcomings still exist.
This survey will provide a comprehensive perspective on the current state of face detection,
verification and identification technologies, highlighting limitations which must be rectified in
order to progress to efficient, dynamic and versatile systems capable of meeting the needs of
modern day usage.

Since the 1990s, significant progress has been made in the realm of face detection and recognition
[4]. Current research in both face detection and recognition algorithms is focused on Deep
Convolutional Neural Networks (DCNN), which have demonstrated impressive accuracy on
highly challenging databases such as the WIDER FACE dataset [5] and the MegaFace Challenge
[6], as well as on older databases such as Labeled Faces in the Wild (LFW) [7]. Rapid
advancements have been triggered due to the increasing affordability of powerful GPUs [8], and
improvements in CNN architecture design which is focused on real world applications [9].
Furthermore, large annotated datasets, and a better understanding of the non-linear mapping
between input images and class labels has also contributed to the increase in research interest in
DCNNs. DCNNs are very effective due to their strong ability to learn non-linear features, however
they are inhibited by intensive convolution, and non-linear operations, which result in high
computational cost [10]. Nevertheless, DCCNs are predicted to encompass future research and
industry application, and are currently being deployed by large corporations such as Google,
Facebook, and Microsoft [11].

No comprehensive survey of face detection, recognition and verification methods has been
conducted since 2003, although several quality reviews have been conducted on specific aspects
of facial recognition solutions. However, they lack complete coverage of all existing research
areas and processes necessary to the development of effective solutions. To address this shortage,
this paper will review all relevant literature for the period from 2003-2018 focusing on the
contribution of deep neural networks in drastically improving accuracy. Furthermore, it will
provide a comprehensive analysis of the state-of-the-art approaches in face detection, verification
and identification, alongside a thorough coverage of the most recently developed databases and
benchmarks, and preprocessing methods. Specific applications, such as video-based face
recognition, and infrared recognition technologies will also be considered. Throughout the review,
 shortcomings and limitations are highlighted, and indicators of areas requiring future research and
improvement are emphasized. Causes of inaccuracy and computational bottlenecks are explored,
and recently proposed solutions evaluated in terms of real life applications. Furthermore, a
comparative analysis of state of the art methods will be provided, as well as a brief overview of
traditionally used methods which have recently been outperformed, but which provide an
alternative to computationally expensive deep methods. In particular, this survey will identify
areas of improvement in processing time, efficiency, cost and accuracy and analyse the means by
which effectiveness is measured. Our contributions are outlined as follows;
• Provide visually appealing comparative analysis of modern face recognition systems and
databases using tables, graphs and figures revealing performance and key features,
• Offer a critical analysis of the benefits and limitations of the state of the art methods, and a
broad range of solutions designed to address shortcomings,
• Summarise the role of current research in the context of traditional methodologies, outlining
the development of technologies over time,
• Highlight major shortcomings, and key areas requiring improvements in light of the latest
research undertaken in specific areas of facial recognition.

II. BRIEF CONTEXT

Since the seminal work of [12] introduced eigenfaces to the computer vision community, and [13]
proposed the use of Haar features for face detection, there has been a significant increase in interest
in facial recognition technologies. Technical progress in this field can be separated into four major
areas of research interest, as shown in Figure 1. Holistic approaches employ distributional
concepts such as manifold [[14], [15], [16]] and sparse representation [[17], [18]], and linear
subspace [[19], [20]] to create low-dimensional representations. These approaches were however
limited by the uncontrollable variations in facial appearance, which naturally deviate from
distribution assumptions. This limitation gave rise to the use of local feature-based techniques
such as Gabor wavelets [[21], [22] [23]], and LBP [24], which provided greater robustness due
greater invariance to transformations and environmental effects. These methods were limited by
lack of distinctiveness, a problem which was partially rectified by the use of shallow learning-
based local descriptors [[25], [26], [27]] which improved compactness and introduced automated
encoding. However, these shallow methods failed to achieve optimal performance due to their
inability to accurately represent the complex non-linear variations in facial appearance.
 Current state of the art on LFW and MegaFace
ArcFace Challenge. Although identification rate on the
2018 latter is still low (82.55% on SphereFace),
99.83% (LFW) verification rate is close to 100%. Significant
improvements in handling expression, pose,
SphereFace illumination and occlusion.
2017 High computational cost due to extensive use
99.42% (LFW) of GPUs and very deep network architectures.
Deep Learning

Issues of poor annotation and noise, together

FaceNet with image quality affect performance
2015
99.63% (LFW)
Surpassed human verification accuracy in
DeepFace unconstrained settings for the first time.
Commenced movement to focus research on
2014 deep learning methods such as CNNs

97.35% (LFW) High computational cost and suffers from loss

of accuracy in situations involving spoofing,
cross-pose, cross-age, low-resolution and
make-up
PCANet (2015) Improved distinctiveness and compactness of
86.28% (LFW) codebook.
Shallow

LE (Learning-based
Descriptor) (2010) Representation not robust to complex non-
84.45% (LFW) linear nature of face
Local Binary Patterns Robust to illumination and expression
2004 Removed the need for manual annotation
66-79% (FERET) Manually designing optimal encoding method
Local Handcraft

and codebook is very difficult

Gabor Wavelets Susceptible to surface issues such as
2002 blurring.Results in uneven distribution which
>70% (LFW) reduces informativity and compactness
Haar Features Provided method of detecting faces efficiently
2001 and effectively. Pioneered boosting based
93.9% (Detection on detection methods.
MIT-CMU test set) Sensitive to illumination, pose, image quality
Simple, efficient method of recognizing faces
EigenFaces in constrained environments.
Holistic

1991 Relatively ineffective in face recognition in

60.02% (LFW) unconstrained conditions due to lack of
robustness to lighting, pose, expression and
image quality changes

Figure 1: Timeline of developments in facial feature representations and face verification accuracy
 A major milestone in the development of facial recognition techniques was achieved by the
introduction of highly accurate deep learning methods such as DeepFace [28] and DeepID [29].
For the first time, face verification in unconstrained settings was achieved with accuracy
surpassing human ability. This development was only allowed for by the advent of significant
improvements in hardware, such as high capacity GPUs. Since then, the majority of research has
focused on the development of deep learning-based methods which attempt to model the human
brain, via high-level abstraction achieved using a concurrence of non-linear filters resulting in
feature invariance. The majority of these methods rely on increasingly deep CNNs, with an
emphasis on promoting sparsity and selectivity. Other deep learning methods.

III. RELATED WORKS

Facial recognition and verification is an area of high research interest due to its broad span of
applications, the available scope for improvement in accuracy and computational speed due to
innovation in hardware and increasingly large and accessible databases. Accordingly, literature
reviews have been conducted periodically to cover these changes. However, due to the vast range
of face recognition methods employed, most reviews focus on a particular issue or set of problems,
rather than addressing the entire range of dominant methods. For example, several recent surveys
have specifically addressed a range of methodologies which have attempted to achieve rotation
invariant face recognition (Oscos, Khoshgoftaar [30]), pose-invariant face recognition [31], skin-
based face detection in various conditions [32] and low-resolution face recognition [33]. Other
publications have reviewed face recognition techniques from various perspectives, for example,
only 2D methods were reviewed in [34] whilst several papers, including [35] and [36] examined
face recognition in terms of presentation attack detection and anti-spoofing technologies.
However, these surveys lack a comprehensive coverage of all currently relevant facial recognition
methodologies, and often do not refer to the most current databases and benchmarks, such as the
MegaFace Challenge benchmark.

Another disadvantage reflected in several recent surveys including [37], [38], [39] and [40] is the
excessive focus on local handcraft descriptors and shallow learning methods. In contrast, this
review presents a strong emphasis on the current research trends, particularly in the realm of deep
learning methods, such as the usage of DCNNs, which currently produce state of the art standards
in face detection, recognition and verification tasks. It must also be noted that various quality
reviews have been conducted on aspects of face recognition, such as face detection alone, [[41],
[42]], or identification and verification tasks [43]. In contrast, this survey will attempt to provide
a clear comparison and analysis of the advantages and disadvantages of a comprehensive range of
techniques in all areas of face recognition including detection, feature extraction and classification.
It will also offer an overview of which methodologies are most suited to a range of applications
whilst highlighting the areas within each area of research which could be further improved.
Attention is also given to the latest databases and benchmarks used to measure the accuracy and
scope of face recognition systems, noting whether they are publicly available or not.
IV. DATABASES
All facial recognition and
detection systems require the
use face datasets for training
and testing purposes. In
particular, the accuracy of
CNNs is highly dependent on
large training datasets [44].
For example, the
development of very large
Figure 2: Sample subset of the MegaFace Challenge dataset
datasets such as ImageNet [45],
which contains over 14 million images, has allowed the development of accurate deep learning
object detection systems [11]. More specifically, face detection and recognition datasets developed
alongside benchmarks such as the MegaFace Challenge [46], a subset of which is shown in Figure
2, the Face Detection Dataset and Benchmark (FDDB) dataset [47] and the Labeled Faces in the
Wild (LFW) dataset [48] provide a means to test and rank face detection, verification and
recognition systems using real-life, highly challenging images in unconstrained settings. Notable
and widely used datasets are listed in Table 1, along with information regarding their intended
usage, size and the number of identities they contain.

Upon analysis of the results attained by face verification and identification algorithms tested on
small datasets such as the LFW dataset, one may be led to believe there remains little scope for
improvement. This is far from true: when tested on millions of images, algorithms achieving
impressive results on smaller testing sets produce far from ideal accuracies [46]. The MegaFace
Challenge was created in response to the saturation of small datasets and benchmarks, providing
a large-scale public database and benchmark which requires all algorithms to be trained on the
same data and tested on millions of images, allowing fair comparison of algorithms without the
bias of private dataset usage. This addresses the problem of lack of reproducibility of results [49]
caused by the usage of private databases for training by state of the art CNN methods [50].
Although a shortage of cross-age identity sets is one limitation of the MegaFace dataset, results
thus far have indicated there is ample scope for algorithm improvement, with the highest
identification and verification accuracies attained by the state of the art method ArcFace [49]
reaching 82.55%, and 98.33% respectively. Similarly, the MS-Celeb-1M database was created to
provide both training and testing data, to enable the comparison of face recognition techniques by
use of a fixed benchmark. However, despite the benefits conferred by their size, both MegaFace
and MS-Celeb-1M are disadvantaged by annotation issues [51] and long tail distributions [52].
 Database Website
MegaFace [6] http://megaface.cs.washin
gton.edu/index.html
WIDER FACE [5] http://mmlab.ie.cuhk.edu.
hk/projects/WIDERFace/
Labelled Faces in the http://vis-
Wild (LFW) [7] www.cs.umass.edu/lfw/
VGG-Face [48] http://www.robots.ox.ac.u
k/~vgg/data/vgg_face/
VGG2-Face [53] http://www.robots.ox.ac.u
k/~vgg/data/vgg_face/
MS-Celeb-1M [54] https://www.msceleb.org/
FaceScrub [55] http://vintage.winklerbros.
net/facescrub.html
FERET [56] http://www.cs.cmu.edu/af
s/cs/project/PIE/MultiPie/
Multi-Pie/Home.html
Public
CASIA-WebFace [57] http://www.cbsr.ia.ac.cn/e
nglish/CASIA-WebFace-
Database.html
MAFA dataset [58] http://www.escience.cn/pe
ople/geshiming/mafa.html
CMU Multi-PIE [59] http://www.cs.cmu.edu/af
s/cs/project/PIE/MultiPie/
Multi-Pie/Home.html
MORPH Database http://www.faceaginggrou
[60] p.com/morph/
CACD [61] http://bcsiriuschen.github.
io/CARC/
YouTube Faces (YTF) https://www.cs.tau.ac.il/~
[62] wolf/ytfaces/
SCface [63] http://www.scface.org/
UMDFaces [64] http://www.umdfaces.io/
Features Application
4,700,000 images Large database and benchmark
672,000 identities suited for CNN comparison.
32,203 images Face detection with large
containing 393,703 illumination, expression, makeup,
faces occlusion, scale and pose
variations
13,233 images Benchmark for automatic still
5749 identities image face verification
2,600 000 images. Celebrity images in
2,622 identities. unconstrained conditions
3,310,000 images Downloaded from Google Image
9131 identities Search
10,000,000 images Celebrity images with associated
100,000 identities entity keys
107,818 images Celebrity images in
530 identities unconstrained conditions
14,126 images. Used for comparing effectiveness
1199 identities of recognition algorithms
494,414 images. Large database with many
10,575 identities identities
30, 811 images Face detection focusing on
containing 35,806 masked or occluded faces
faces
750,000 images. Varying illumination and
337 identities expression
1724 images. 515 Large longitudinal face database
identities (1962- showing effects of aging
1998) More recent
album: 55,134
images. 13,000
individuals
163,446 images. Focuses on age related issues
2000 identities
3425 videos. Benchmark for unconstrained
1595 identities. video-based face verification
4160 images. Video surveillance camera
130 identities footage
367,888 images Clean, highly annotated
8,277 identities & celebrity/public figure images
3,700,000 video obtained using GoogleScraper
 frames of 3100
identities
Face Detection http://vis- 2845 images Face detection for images
Dataset and www.cs.umass.edu/fddb/ 5171 faces containing 1 or more faces
Benchmark (FDDB)
[47]
PolyU near-infrared http://www4.comp.polyu. 35,000 samples. Used for NIR face recognition in
face database (PolyU- edu.hk/~biometrics/polyu 350 identities controlled and uncontrolled
NIRFD) [65] db_face.htm conditions

3D Mask Attack https://www.idiap.ch/data 76,500 images. 17 Recorded using Kinect. Aimed at

Database (3DMAD) set/3dmad identities. preventing 2D face spoofing
[66] attacks.

Social Face Unavailable 4,400,000 images. Facebook internal dataset

Classification (SFC) 4030 identities
[28]
FaceNet [67] Unavailable 100M-200M Google dataset of face thumbnails
images,
8,000,000 identities
Face Detection at Unavailable 741 images. 1856 Surveillance footage at nighttime
Night in faces. for detection
Surveillance (FDNS)
[68]
Private

CrowdFaceDB [69] Unavailable 385 videos. 257 Crowd video-based detection and
identities recognition

Bosphorus Database http://bosphorus.ee.boun.e 4666 images. 105 3D and 2D based – contains

[70] du.tr/Home.aspx identities occlusions, expressions and poses
conditionally available
KinectFaceDB [71] http://rgb-d.eurecom.fr/ 468 images. 52 Multimodal – 2D, 2.5D and 3D
conditionally available identities and video data using Kinect
sensor

Photoface Database http://www1.uwe.ac.uk/et 7356 images. 261 3D and 2D – contains natural

[72] /mvl/projects.aspx#PPDB identities real-life expressions and poses

Table 1: Summary and comparison of the main features and focuses of publicly available face detection and
recognition datasets, and notable private datasets

The LFW dataset and benchmark is often referred to as the de facto benchmark for automatic still
image face verification. It is a relatively small database which has been varied to include
additional, specific databases. These variations are very useful for developers when addressing
problematic aspects or issues associated with face recognition. Another small database is the
FERET benchmark and database, which was developed to provide a means to directly compare
different face recognition algorithms, identify state of the art methodologies, identify promising
approaches, and highlight areas requiring future research to further the development of face
recognition technologies. [73]. These benchmarks have thus far proven very useful in this purpose
but fail to account for the significant data needs of currently used CNN approaches. For this
reason, significantly larger databases and associated benchmarks including MegaFace, are being
 used for comparison of deep learning face recognition approaches.

V. FACE DETECTION
Face detection is a fundamental step in facial recognition and verification [74]. It also extends to
a broad range of other applications including facial expression recognition [75], face tracking for
surveillance purposes [76], digital tagging on social media platforms [77] and consumer
applications in digital technologies, such as auto-focusing ability in phone cameras [78]. This
survey will examine facial detection methods as applied to facial recognition and verification.
Historically, the greatest obstacle faced by face detection algorithms was the ability to achieve
high accuracy in uncontrolled conditions. Consequently, their usability in real life applications
was limited [41]. However, since the development of the Viola Jones boosting based face detection
method [13], face detection in real life settings has become commonplace. Significant progress
has since been made by researchers in this area [41] due to the development of powerful feature
extraction techniques including Scale Invariant Feature Transform (SIFT) [79], Histograms of
oriented Gradients (HoGs) [80], Local Binary Patterns (LBPs) [81] and methods such as Integral
Channel Features (ICF) [82]. For a recent and comprehensive review of these traditional face
detection methodologies, readers are referred to [83]. This review will alternatively focus on more
recently proposed deep learning methods, which were developed in response to the limitations of
HoG and Haar wavelet features in capturing salient facial information under unconstrained
conditions which include large variations in resolution, illumination, pose, expression, and color.
Essentially, it is the limitations of these feature representations which have thus far limited the
ability of classifiers to perform to the best of their ability [43]. Furthermore, due to the significant
increase in availability of large databases, DCNNs generally demonstrate higher performance in
object and face detection tasks, as demonstrated by [11], [84] and [85].

Deep Learning: Face Detection

Figure 3: From Haar features to tiny, and highly occluded faces. The use of CNNs in face
detection has significantly improved accuracy
Recently, the creation of large annotated databases such as the MegaFace Challenge, LFW and
WIDER FACE has encouraged the development of highly discriminative, state of the art deep
learning face detection. Consequently, DCNNs now perform significantly better in object and face
detection tasks, as demonstrated by [11], [84] and [85]. DCNN face detection methods can be
categorized as region based or sliding window approaches. The region-based approach uses an
object proposal generator such as Selective Search [86] to generate a pool of regions which may
include one or more faces. These proposals are then input into a DCNN which classifies them as
either including a face or not, returning the precise bounding box coordinates of faces in a given
image with minimal background inclusion [43]. Most current methods, including HyperFace [87]
and All-in-One-Face [88] use this approach. More efficient region-based CNNs (R-CNN) [89] use
a DCNN to generate proposals, perform bounding box regression and classification. [90] improved
upon R-CNN by using a combination of feature concatenation, multi-scale training improving
scale invariance, and hard negative mining, to reduce false positive rates, achieving state of the art
recall and accuracy. This method suffers from high computational cost, thus requires
improvements in efficiency and scalability to be deployable for real-time face detection. Other
methods which aim to improve upon state of the art region based methods include [91], which
developed a method to reduce redundant region proposals, and [92], who proposed a very
lightweight Single-State Headless face detector which achieved state of the art accuracy by
detecting faces directly from the early convolutional layers within the classification network.
However, despite impressive results, region-based deep face detectors are computationally
expensive due to the requirement of proposal generation.

An alternative and far more efficient approach to face detection is the sliding window-based
method, which computes accurate Face Detection Method WIDER FACE
bounding box coordinates at each (hard)
location in a feature map of a specific ScaleFaces [93] 76.4%
scale, using a convolution operation. HR [94] 81.9%
Scale invariance is achieved by SSH [92] 84.4%
generating an image pyramid containing S3FD [95] 85.8%
multiple scales. One such method is FAN [96] 88.5%
[97], which proposed a single-shot Table 2: Top 5 face detection methods
detector which uses the inbuilt
pyramidal DCNN cascade architecture to rapidly eliminate background regions at low resolutions,
allowing only challenging regions to be processed at high resolutions. A single forward pass is
sufficient to obtain detections, thus reducing computation time. [98] also achieves superior results
by use of a novel hard sample mining strategy together with a deep cascaded multitask framework
which leverages off the correlation between detection and alignment to improve performance.
Another sliding window-based method is [99] which developed DP2MFD, a deformable parts
model integrated with deep pyramid features, wherein the face is defined as a collection of parts
which are trained alongside the global face, to achieve scale invariant, state of the art detection.
Furthermore, [100] attempted to address the issue of multi-view face detection by proposing a
minimally complex Deep Dense Face Detector (DDFD) without the need for pose or landmark
annotation. It achieved similar results to highly complex methods but is limited due to inadequate
sampling strategies and the need to improve data augmentation. Faceness [101] claimed to achieve
effective face detection even in cases where over 50% of a given facial region was affected by
occlusion. It also claimed to overcome significant pose variation, with the added benefit of accept
arbitrary images of varying scale. This was achieved using a set of attribute aware deep networks
which were pre-trained with generic objects, followed by refinement using specific part-level
 binary attributes. However, the authors acknowledged that improvements in speed were possible,
noting the benefits of integrating model compression techniques and approximation of non-linear
filtering with low-rank expressions [102]. Another recently proposed method is ScaleFace: [93],
which designed a simplified multi-network CNN approach capable of detecting faces at a very
wide range of scales, by using a specialized set of DCNNs with varying structures, without the
use of a traditional image pyramid input. Finally, [94] evaluated the significant impact of
contextual information on the detection of very small faces, subsequently developed massively-
large receptive field based templates used to train separate detectors for different scales, improving
the state of the art results on WIDER FACE from 29-64% to 81% . Figure 3 shows illustrates the
significant progress achieved by face detection systems when detecting naturally occluded faces,
handling significant discrepancies in scale.

Despite the increasing accuracy and speed of face detection systems, the two greatest challenges
remain somewhat unresolved. Face detectors are required to cope with large and complex
variations in facial changes, and effectively distinguish between faces and non-faces in
unconstrained conditions. Furthermore, the large variation in face position and size within a large
search space presents challenges which reduce efficiency [103]. This calls for a trade-off between
high accuracy and computational efficiency. One benefit of less accurate Viola Jones inspired
cascade-based face detectors over CNN methods is their efficiency. Thus the greatest requirement
in the current field of research is the development of more efficient CNN face detection
techniques. [96] partly addressed this issue, achieving the current state of the art accuracy rate of
88.5% on the hard WIDER FACE test set by developing the Face Attention Network (FAN), a
novel face detector designed to improve recall in cases of occlusion without impacting on
computation speed. This was achieved by using an anchor-level attention to enhance facial
features within a face region, together with random crop data augmentation to tackle occlusion
and tiny faces. A comparison of the five highest accuracy face detectors as measured on the
WIDER FACE benchmark is provided in Table 2.

VI. FEATURE EXTRACTION

Feature extraction usually occurs immediately after face detection and can be considered as one
of the most important stages in face recognition systems, as their effectiveness is dependent upon
the quality of the extracted features. This is because facial landmarks and fiducial points identified
by a given network determine how accurately features are represented. Traditional fiducial point
locators are model-based, whilst many recent methods are cascaded regression based [43]. Lately,
key improvements have been made with the development of deep dual pathway methods [1], and
other confidence map based solutions, such as [104] and [105]. Traditional model-based fiducial
point methodologies include Active Shape Model (ASM), which suffers from low accuracy,
partially rectified by the work of [106], Active Appearance Model (AAM) [107], and Constrained
Local Models (CLM). CLMs are generally outperformed by cascaded regression [108], models
due to the latter’s inherent inability to mode the complex variation of local feature appearances. It
must be noted however, that highly effective methods based on CLMs have been developed. For
example [109] is based on CLMs but takes advantage of the neural network architecture, proposing
a Convolutional Experts Network (CEN) and Convolutional Experts Constrained Local Model
(CE-CLM) which uses CLM as local detector, achieving very competitive results particularly on
profile images.

Face recognition differs to object recognition in that it involves alignment before extraction [50].
This is reflected in the differences between CNNs used for face recognition and those used for
object recognition. An increase in data availability has resulted in development of learning-based
methods as opposed to engineered features due to their inherent ability to discover and optimize
features specific to a task. Consequently, learning methods have outperformed engineered
features [28]. In CNNs, a fiducial point detector is employed to localize important facial features
such as eye centers, mouth corners and nose tip. Once these landmarks have been identified, the
face is aligned and according to normalized canonical coordinates [110]. Subsequently, a feature
descriptor is extracted, encoding identity information. Similarity scores are then calculated using
these face representations. If this score is below a set threshold, the faces are classified as
belonging to the same identity. Model based DCNN methodologies employ training to learn a
shape model, which is used to fit new faces during testing. However, these learned models are
limited by their sensitivity to gradient descent optimization initialization and their lack the ability
to represent complex facial variations in pose, expression and illumination [1]. Deep cascade
regression based models, first proposed by [111], and improved by [112], rapidly outperformed
shallow methods such as [113] and [114]. These methods learn a model which directly maps the
appearance of the image to the target output. The effectiveness of these methods is highly
dependent on the robustness of the local descriptors used. However, most cascaded regression
methods involve independent regressor learning, which causes issues of cancelling regressor
descent direction, thus inhibiting learning. To address these shortcomings, [115] proposed a
combined convolutional recurrent network which allows training of an end-to-end system, in
which the recurrent module facilitates joint optimization of regressors by assuming cascades are
a non-linear, dynamic system. Thus, all information between cascade levels is used and shared
between layers.

Figure 4: Fiducial point detection by the state-of-the-art method proposed by [1]

Subsequently, more representative deformable 3D models used by [116] and [117] to estimate
facial poses and shape coefficients outperformed then state of the art shallow methods, including
[118] and [119]. These deep methods are however disadvantaged by the need to re-initialize
models when switching stages or networks, particularly in systems where local deep networks are
used to localize fiducial points based on facial patches, as seen in [111]. [120] used cascade
regression to predict a 3D to 2D projection matrix and base coefficients. The concept was further
developed in [116] which approached face alignment from a 3D model fitting perspective,
resulting in the development of a cascade of DCNN-based regressors which function to estimate
3D shape parameters, and the camera projection matrix. [121] addressed the problem of variations
in pose by using a simple, generic 3D surface to approximate the shape of all input faces. This
method is however limited by its reliance on the quality of detected landmarks which, if poor, can
cause the appearance of undesirable artifacts. Other works include [122] which used a DCNN to
fit a dense 3D face model to a given image, employing a Z-buffer to model depth data.
Alternatively [123] developed Local Deep Descriptor Regression (LDDR) which provides a
highly accurate means of localizing fiducial points using deep descriptors which are able to
accurately describe every pixel in a given image. [124] further presented an iterative method for
unconstrained fiducial point estimation and pose production by employing a novel CNN
architecture dubbed Heatmap-CNN (H-CNN) which captures both global and local features by
generating a probability value, indicating the presence of joint at a defined location. This allows
accurate, state of the art key point detection without the use of 3D mapping.

Multitask learning (MTL) approaches integrate face detection and fiducial point estimation within
the same process, allowing greater robustness, due to additional supervision. One effective
example is [91] which proposed Supervised Transformer Network, a cascade CNN which uses a
two stage process to predict face candidates and landmarks, followed by mapping landmarks to
canonical positions to normalize face patterns, followed by validation, achieving previous state of
the art results. [125] presents a semantic approach which uses a combination of a ConvNet with a
3D model to detect faces and their fiducial points in the wild, achieving competitive results. It
must be noted that network design significantly affects performance [1]. The abovementioned
systems face difficulty in pixel-level localization and classification tasks due to spatial-semantic
uncertainty [126] caused by the failure by to retain adequate spatial resolution after pooling and
convolutional layers are employed when using deep representation generated at the lower layers.
To rectify this issue, [87] concatenated shallow-level convolutional layers to the latest
convolutional layers prior to landmark regression, while [1] proposes a dual pathway model which
forces shallow and deep network layers to maximize the likelihood of highly specific candidate
region. Notably, [127] used aggregation of shallow and deep layers to generate more accurate
score map predictions, in the field of pose estimation.

Alternative methods include [128] [104] and [105], which generate a confidence map for each
landmark to indicate likelihood of landmark appearing at specific location in original image.
Prediction is made by selecting the location with maximum response as shown by the confidence
map. In comparison to cascade regression these methods are more effective, as they suppress false
predictions caused by noisy regions, thus improving robustness, with greater accuracy in
unconstrained conditions. [128] addresses the problem of reliance on high quality detection
bounding boxes coordinates by proposing a Convolutional Aggregation of Local Evidence
(CALE) which comprised of a CNN which performs facial part detection, mapping confidence
scores for the location of each landmark within the first few layers. The score maps and CNN
features are then aggregated by using joint regression in order to refine landmark location. CNN
regression guides contextual learning when predicating the position of occluded landmarks in
unconstrained conditions, thus increasing robustness. However, networks such as [105] and [129]
which rely on [130] to find the location of facial landmarks, are disadvantaged by the low quality
confidence map generated by DeconvNet. Other methods such as [104] minimize residual error in
score maps by use of a stacked cascaded architecture which refines key-point predictions, however
it is difficult to deploy on a small scale due to its heavy and largely redundant architecture.
Furthermore, [131] proposed an Lapalcian-pyramid architecture that provides effective refinement
 of 2D score maps generated by lower layers, by supervising the adding back of higher level
generated features using three softmax layers. [1] proposed a Globally Optimized Dual-Pathway
(GoDP) deep architecture to rectify these issues. This method aims to identify target pixels by
solving a cascaded pixel labeling problem without the use of high-level inference models or
complex stacked architectures. High quality 2D score maps are generated without the use of
stacked architecture, partially rectifying lack of spatial semantic information by discriminatively
extracting it from the deep network, as shown by Figure 4, and developed a novel loss function
which reduces false alarms. This method currently achieves state of the art results, outperforming
cascaded regression-based models on complex face alignment databases.

Currently, the greatest shortcoming present in the realm of unconstrained face alignment and
fiducial point detection is the lack of solution to the problem of aligning faces irrespective of pose
variation, and the general reliance of systems on accurate face detection. The 300 Faces in the
wild database [132] is generally used for comparison of fiducial point detection methods. This
face dataset is limited, and thus one area of improvement could include the creation of a large-
scale annotated dataset containing a broad range of unconstrained facial images specifically
designed for use in face alignment and fiducial point detection applications. This would improve
robustness across fiducial point detection generally, particularly with respect to pose and
expression variations, low illumination and poor quality. With respect to network structures,
deepening neural networks may capture more abstract information which may assist in detection,
however it is still unclear which network layers contribute most significantly to local features
relevant to fiducial point detection [43]. This is one area which may benefit from further research.
Furthermore, the high computational cost associated with localizing fiducial points still remains a
significant challenge in unconstrained conditions.

VII. FACE IDENTIFICATION AND VERIFICATION

Subsequent to feature extraction, facial recognition is performed. Recognition can be categorized
as either verification or identification. Modern face recognition systems using DCNNs involve
deep feature extraction, and lastly, similarity comparison. More specifically, verification involves
comparison of one-to-one similarity between a probe image and a gallery of a known identity,
whilst identification determines one to many similarities to determine the identity of the probe.
Both these processes require robust feature representation, and a discriminative classification
model or similarity measure. Traditional methods used for feature representation include LBP,
HoGs, and Fisher Vector. Relevant metric learning methods include cosine metric learning,
Mahalanobis metric learning, and one-shot similarity kernel. Others include large margin nearest
neighbor, Joint Bayesian and attribute-based classifiers. These methods are thoroughly reviewed
by [41]. Thus, for the sake of relevance and context, we have only included a brief overview of
the role these methods play in modern face recognition and have chosen to focus on the most
recently developed state of the art methodologies, which largely rely on DCNNs.

The modern CNN framework was designed in 1990 by [133] when they developed a system
known as LeNet-5 to classify handwritten digits by recognizing visual patterns from image pixels
without the need for preprocessing. [134] first presented a neural network used for upright, frontal,
grayscale face detection, which although primitive by today’s standards, compared in accuracy
with state-of-the-art methods at the time. Since then, research has accelerated significantly,
leading to the development of highly sophisticated DCCNs capable of detection, recognition and
verification with accuracy approaches that of humans. Although the development of CNNs was
impeded by lack of computing power [135], recent hardware advances have allowed rapid
improvement and a significant increase in CNN depth, and consequently, accuracy. One
outstanding feature is an increase in depth, and width to allow for improved feature representation
by improving non-linearity [135]. However, this leads to issues such as reduction in efficiency
and overfitting [9]. This section will explore the various methods which have aimed to address
these problems in the context of facial recognition, through an examination of general
improvements in DCCN architecture and loss functions. CCNs are generally more suitable to
object recognition than standard feedforward neural networks of similar size due to the use of
fewer connections and parameters which facilitates training and efficiency, with only slight
reduction in performance [11]. CNNs were designed specifically for classification of 2D images
[136] due to their invariance to translation, rotation and scaling [137]. A CNN is comprised of a
set of layers, including convolutional layers, which are a collection of filters with values known
as weights, non-linear scalar operator layers, and down sampling layers, such as pooling.
Activation values are the output of individual layers which are used as input in the next layer
[138]. For a thorough overview of basic CNN components, readers are referred to [135].

The use of CNNs in facial recognition tasks is comprised of two essential steps; namely, training
and inference. Training is a global optimization process [135] which involves learning of
parameters via observation of huge datasets. Inference essentially involves the deployment of a
trained CNN to classify observed data [138]. The training process involves minimization of the
loss function to establish the most appropriate parameters, and determination of the number of
layers required, the task performed by each layer, and networking between layers, where each
layer is defined by weights, which control computation. CNN face recognition systems can be
distinguished in three ways; the training data used to train the model, the network architecture and
settings, and the loss function design [49]. DCNN’s have the capacity to learn highly
discriminative and invariant feature representations, if trained with very large datasets. Training
is achieved using an activation function, loss function and optimization algorithm. The role of the
loss function is to determine the error in the prediction. Different loss functions will output
different error values for an identical prediction, and thus determine to a large extent the
performance of the network. Loss function type depends on the type of problem, e.g. regression
or classification. Minimization of the error is achieved using back propagation of the error to a
previous layer, whereby the weights and bias are modified. Weights are learned and modified
using an optimization function, such as stochastic gradient descent, which calculates the gradient
of the loss function with respect to weights, then modifies weights to reduce the gradient of the
loss function [138].
 Model/ Training Number Loss Face Face
Method Dataset of NNs function verification idenitification
DeepFace SFC 3 Cross- 97.35% -
[28] entropy
loss
DeepFR [48] VGG-Face 1 Triplet 98.95% -
CenterFace CASIA- 1 Center 99.28% 65.23% (MF1)
[139] WebFace, Loss 76.72% (MF1)
CACD,
Celebrity+
[140]
SphereFace CASIA- 1 Angular 99.47% 75.77% MF1
[141] WebFace softmax 89.14% MF1 (small protocol)
DeepID2+ CelebFaces+, 25 - 99.47% -
[142] WDRef
DCFL [143] CASIA- 1 Correlation 99.55% -
WebFace loss
FaceNet [67] FaceNet 1 Harmonic 99.63% 70.49% (large
triplet loss 86.47% (MF1) protocol)
CosFace CASIA - 1 Large 99.73% 79.54% (small
[144] WebFace Margin 97.96% (MF1) protocol)
Cosine 84.26% (large
protocol)
ArcFace Refined 1 Additive 99.83% 83.27% (MF1)
(LResNet10 MS-Celeb- Angular (LFW)
0E-IR) [49] 1M, VGG2 Margin 98.48 (MF1)
Loss
Table 3: State-of-the-art and competitive face verification and identification methods. All
verification results are recorded on the LFW dataset unless indicated. All identification results
are obtained on the MegaFace Challenge 1 (MF1) dataset.

Loss function modifications have been very popular as a means of improving accuracy of face
recognition systems, thus many variations have been proposed lately. Softmax loss function and
its variations [145] are commonly used as they promotes separation of features. However, it shows
ineffectiveness when intra-variations are greater than inter-variations. As such novel loss functions
such as the ArcFace [49] loss function have been proposed, and have shown greater effectiveness.
The additive angular margin (ArcFace) loss function produces more accurate geometrical
interpretation than previously used supervision signals, obtaining more discriminative deep
features by maximizing the decision boundary in angular space premised on L2 normalized
weights and features. It currently produces the start of the art face identification and verification
results on both LFW and the MegaFace Challenge, as shown by Table 3. This followed significant
progress in the development of multiplicative [141] and additive cosine margins [146] whch are
added into the Softmax loss to enhance its discriminative power. These angular and cosine margin-
based loss functions have shown improved performance over Euclidean-distance based loss
functions due to the use of angular similarity and separability between learned features.
Particularly, [141] proposed an angular loss function based on the Softmax loss which uses highly
discriminative feature representation optimized for cosine distance and similarly metric, achieving
prior state of the art results. [147] proposed a combination of a novel triplet loss function, and
feature fusion across layers which achieved state of the art performance in video-based face
recognition, while [139] proposed a loss which uses the centroid in each class as a regularization
constraint within the softmax function within a residual neural network. [145] used Softmax loss
regularized with a scaled L2 Norm constraint which was shown to optimize the angular margin
between classes. The last stage in face recognition is similarity comparison, which occurs after
training. This involves the conversion of test images to deep representations, similarity is
calculated by use of L2 distance or cosine distance, after which methods such as nearest neighbor
or threshold comparison are used to identify or verify faces. Other methods, including metric
learning and sparse representation classifiers are also used to post-process deep features to
improve accuracy and efficiency. It must however be noted that despite the high accuracy
produced using these novel loss functions, they suffer excessive GPU memory consumption
within the classification layer when handling large amounts of data. Additionally, the triplet and
contrastive loss functions are disadvnateged by the difficult task of selecting effective training
samples.

Notably, [28] proposed DeepFace, which uses a Siamese network architecture which employs the
same CNN to obtain descriptors for pairs of faces which are then compared via Euclidean distance.
This method uses metric learning during training to minimize difference between two images of
the same identity, and maximise the distance between those of differing identities. Although this
process achieved state of the art recognition, it was further improved in [148] by increasing the
size of the training data set. DeepFace was further enhanced by REF 24-27 in [48]. Generally, the
architecture of CNNs is determined on experience, on a trial and error basis. [136] proposed to
rectify this by developing a fully automated Adaptive Convolution Neural Network (ACNN) to
specifically address facial recognition. Its structure is created automatically based on performance
and accuracy requirements. Based on simple network initialization, convergence is then used to
determine whether or not expansion will occur depending on the allowable system average error
and desired recognition rate. It improves upon an Incremental Convolutional Neural Network
(ICNN) proposed by [149] as global expansion is controlled automatically, rather than artificially.
This study aimed to achieve a desirable balance between training time and recognition rate without
the need for performance comparison [136]. [28] also took an alternative approach as they did not
use standard convolution layers, instead relying heavily on an extensive database of over 4 million
faces to train a nine-layer deep feedforward neural network, and a 3D face model based alignment,
to generate a face representation. This network employed several locally connected layers without
weight sharing and over 120 million parameters. This method claimed to achieve close to human
level accuracy performance on the LFW dataset. However, the DeepID frameworks [142, 150]
were however the first to achieve state of the art verification results which outperformed human
performance, with the added benefit of using a smaller dataset. These approaches involved
learning of highly discriminative and informative features by using a collection of smaller, shallow
networks, and deep convolutional networks, specific to local and global face patches.
Improving performance by increasing depth and width has drawbacks such as overfitting, which
may lead to bottlenecks and needlessly increases computer resources, e.g. when a lot of weights
eventuate with 0 values [9]. This can be solved by modelling biological networks in transitioning
to thinly connected architectures rather than fully connected networks. [9] proposed a DCNN
named Inception, designed based on the Hebbian principle, i.e. neurons that fire together wire
together, and multi-scale processing to maintain a computational budget of 1.5 billion multiply-
adds at inference time to ensure cost effective real-world usage, on large databases. It claimed to
outperform state of the art object detection and image classification by focusing on improving the
structure of CNN. A 22-layer deep model was created, using 1 x 1 convolutions as dimension
reduction modules to remove computational bottlenecks allowing both depth and width of the
networks to be increased without impeding performance. With a similar goal of reducing
computational cost [138] highlighted the need for greater sparsity. Sparsity is defined by as the
proportion of zero values in a given layer’s activation and weight matrices. The goal of achieving
sparsity resulted in the creation of the Sparse Convolutional Neural Network (SCNN) architecture
which was designed to enhance computational efficiency and performance at inference by
manipulating zero valued activations and weights, to minimize unnecessary data processing and
storage [138]. This is achieved using the sparse planar-tiled input-stationary Cartesian product
(PT-IS-CP-sparse) dataflow. This approach accelerates the convolutional layers, but boasts added
benefit of utilizing both redundant weights and activations to improve performance. However both
approaches have been insufficiently tested on adequate databases, highlighting need for further
research in improvements in network sparsity. Furthermore, computational costs can be reduced
at deployment by using pruning [138]. Pruning is a means by which sparsity is created. It can
involve setting weights below a given threshold to zero, before retraining to regain accuracy. This
achieves a smaller, more efficient, yet accurate network. For example, [137] uses a three stage
approach to reduce network size and computational cost by feature map pruning in each
convolutional layer. Stage 1 involves training of a CNN using parameters that ensure feature map
size is only modified in max pooling layers. The second stage involves utilization of a screening
strategy that calculates discriminability values from feature maps – convolutional and feature
maps with low discriminability magnitudes are pruned. This is followed by a third stage which
involves piecewise pruning and retraining of each convolutional layer in the network. Often,
DCNNs can be pruned significantly without affecting accuracy [137], within the range of approx.
20-80% [138].

The importance of sparsity, selectiveness and robustness was also emphasized by [142], which
designed DeepID2+, improving upon DeepID2 by increasing the dimension of hidden
representations (128 feature maps were used) and adding supervision to early convolutional layers,
improving accuracy by 1.98% to achieve 98.70% accuracy on the LFW dataset. This study
achieved an accuracy of 99.47% on LFW by combining 25 DeepID2+ networks. This study also
reflected the Hebbian principle, as it noted that neural activations are moderately sparse – different
identities activate different subsets of neurons, while identical identities in different images
activate similar neurons. It was suggested that binary activation patterns are important in reducing
computational cost, further speculating that higher layers are sensitive to global features rather
than local variations which may result from occlusion. However, as shown by [11], the depth and
width of the DCNN is significant to allow adequate learning. The researchers successfully trained
one of the largest CNNs on ImageNet, achieving the best results ever recorded on the database at
the time. The network comprised of 8 learned layers: 5 convolutional layers and 3 fully connected
layers. It was noted that removing layers reduced performance. Overall, they constructed a highly
optimized GPU implementation which prevented overfitting by using data augmentation and
dropout and used minimal preprocessing – the only preprocessing involved was resizing of
images. The network was trained on original RGB images using Relu optimization, rather than
tanh, as it meant training time occurred six times faster and input normalization was not required.
Other notable face recognition systems include [8] which presented a CNNN system, comprised
of several DCNNs, designed to perform unconstrained face detection and preprocessing,
automated verification and recognition. The face detection module employed the deep pyramidal
deformable parts model proposed by [99] which has the ability to detect faces with varying sizes
and poses in unconstrained conditions, combined with the architecture proposed by [11] to extract
deep features. This system was quantitatively evaluated on a range of datasets included Labeled
Faces in the Wild (LFW) and the JANUS CS2, the latter containing highly challenging images
and videos. Thus it can be observed that increasing depth increases accuracy [136] but has
significant computational costs [9] and energy consumption [138]. Increasing photo resolution
also increases computational cost [9], which provides another area which may benefit from
additional research, despite the natural improvements in results consequent of greater access to
bigger datasets and faster GPUs [11].

In most DCNNs occlusion is naturally included in learned features alongside non-occluded

features, reducing accuracy by corrupting extracted image features. [151] proposed to rectify
issues caused by occlusion using MaskNet, a differentiable module, optimized by end to end
training, which adaptively generates various feature map masks for a range of occluded faces,
which can be incorporated into CNN architectures. It functions by assigning higher weights to
occluded features activated by non-occluded features, and lower weights to those activated by
occluded features, improving recognition in images affected by synthetic and natural occlusion. It
is claimed that the computational cost of adding the module is minimal. Furthermore, [44]
attempted to rectify blurring and issues arising from large standoffs and poor image quality, caused
by unconstrained illumination and environmental degradation, by incorporating a sensitivity term
into a DCNN cost function. The sensitivity term is comprised of the error divided by the activation
function derivative, followed by minimization of the total error using the gradient descent method.
This improves generalization and feature extraction by shifting the neural activations of the hidden
layers to the middle high gradient area of the activation function. This method demonstrated
effectiveness in day and night time images, and at varying standoff distances on the Long Distance
Heterogeneous Face dataset [152], however it was only tested on a small, augmented dataset of
images. One area of benefit is its effectiveness in reducing the need for intensive manual
preprocessing of highly blurred images. Another early methodology which achieved competitive
results without the benefit of large scale anotated datasets was proposed by [153] which used deep
convolutional belief networks based on local convolutional restricted Boltzmann machines.
Unsupervised representations were learned from unlabeled images, prior to being transferred to a
recognition task by using a classification model such as support vector machine (SVM) and metric
 learning methods. FaceNet [67] also trained a DCNN with a very large database 100-200 million,
using triplets of roughly aligned matching and non-matching face patches. [154] alternatively
proposed handling pose variations by using several models, tailored for each major pose variation,
resulting in greater accuracy. Other methods such as [48, 53] and [155] focused on establishing
databases or improving existing databases as a means to improve learning via improved
annotation, enlargement of databases, and handling of cross-pose factors using 3D morphable
models.

VIII. CONCLUSION
This survey presented a critical analysis of modern face recognition methodologies, developments
and challenges. It also provided a comparative analysis of the available databases, and related
benchmarks. It highlighted shortcomings of state of the art methods, and evaluated responses
designed to address these limitations, emphasizing outstanding issues yet to be addressed. Despite
drastic improvements in accuracy of representation due to the non-linearity of deep feature
representations, we can confidently conclude that there is no known ideal facial feature that is
sufficiently robust for face recognition in unconstrained environments. It must also be noted that
solutions achieving state of the art accuracy are largely inhibited by their dependence on
sophisticated GPUs and large databases, meaning there is still adequate need to focus research
attention on more traditional handcrafted feature representations. Thus, the focus of future
research must be on reducing the excessive computational cost of DCNNs, and their dependence
on large, accurately annotated databases. Refinement of pruning methods, and minimization of
training time is also an area requiring attention, as is network architecture, which would benefit
from increased sparsity, and selectiveness.

IX. CONFLICT OF INTEREST

This research is not affected by any conflicts of interest.

X. FUNDING STATEMENT
This research is supported by an Australian Government Research Training Program (RTP)
Scholarship.

XI. REFERENCES

1. Wu, Y., S.K. Shah, and I.A. Kakadiaris, GoDP: Globally Optimized Dual Pathway deep
network architecture for facial landmark localization in-the-wild. Image and Vision
Computing, 2018. 73: p. 1-16.
2. Team, T. Chinese authorities used Facial recognition AI to catch fugitive among 60,000
concert-goers in China. 2018; Available from: https://techstartups.com/2018/04/13/chinese-
authorities-used-facial-recognition-ai-catch-fugitive-among-60000-concert-goers-china/.
3. PRNewswire, Global Market Study on Face and Voice Biometrics: Government Sector
Projected to be the Most Attractive End Use Industry Segment During 2017 - 2025,
PRNewswire, Editor. 2017.
4. Xiaojun, L., et al., Feature Extraction and Fusion Using Deep Convolutional Neural
Networks for Face Detection. Mathematical Problems in Engineering, 2017. 2017.
5. Yang, S., et al., WIDER FACE: A Face Detection Benchmark. 2015.
6. Kemelmacher, I., et al., The MegaFace Benchmark: 1 Million Faces for Recognition at
Scale. 2016. 4873-4882.
7. Huang, G.B., et al., Labeled Faces in the Wild: A Database forStudying Face Recognition in
Unconstrained Environments, in Workshop on Faces in 'Real-Life' Images: Detection,
Alignment, and Recognition. 2008: Marseille, France.
8. Chen, J.-C., et al., Unconstrained Still/Video-Based Face Verification with Deep
Convolutional Neural Networks. International Journal of Computer Vision, 2018. 126(2): p.
272-291.
9. Szegedy, C., Liu,W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., et al, Going deeper with
convolutions. 2014.
10. Wu, S., et al., Funnel-structured cascade for multi-view face detection with alignment-
awareness. Neurocomputing, 2017. 221: p. 138-145.
11. Krizhevsky, A., I. Sutskever, and G. Hinton, ImageNet classification with deep
convolutional neural networks. Communications of the ACM, 2017. 60(6): p. 84-90.
12. Turk, M. and A. Pentland, Eigenfaces for recognition. J. Cognitive Neuroscience, 1991.
3(1): p. 71-86.
13. Viola, P. and M. Jones, Rapid object detection using a boosted cascade of simple features.
2001. p. I511-I518.
14. Xiaofei, H., et al., Face recognition using Laplacianfaces. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 2005. 27(3): p. 328-340.
15. Yan, S., et al., Graph Embedding and Extensions: A General Framework for Dimensionality
Reduction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2007. 29(1): p.
40-51.
16. Deng, W., et al., Comments on "Globally Maximizing, Locally Minimizing: Unsupervised
Discriminant Projection with Application to Face and Palm Biometrics". IEEE Transactions
on Pattern Analysis and Machine Intelligence, 2008. 30(8): p. 1503-1504.
17. Wright, J., et al., Robust Face Recognition via Sparse Representation. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 2009. 31(2): p. 210-227.
18. Zhang, L., M. Yang, and F. Xiangchu. Sparse representation or collaborative
representation: Which helps face recognition? in 2011 International Conference on
Computer Vision. 2011.
19. Moghaddam, B., W. Wahid, and A. Pentland. Beyond eigenfaces: probabilistic matching for
face recognition. in Proceedings Third IEEE International Conference on Automatic Face
and Gesture Recognition. 1998.
20. Belhumeur, P.N., J.P. Hespanha, and D.J. Kriegman, Eigenfaces vs. Fisherfaces:
recognition using class specific linear projection. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 1997. 19(7): p. 711-720.
21. Chengjun, L. and H. Wechsler, Gabor feature based classification using the enhanced fisher
linear discriminant model for face recognition. IEEE Transactions on Image Processing,
2002. 11(4): p. 467-476.
22. Shen, L., L. Bai, and M. Fairhurst, Gabor wavelets and General Discriminant Analysis for
face identification and verification. Vol. 25. 2007. 553-563.
23. Štruc, V., R. Gajšek, and N. Pavešić. Principal Gabor filters for face recognition. in 2009
IEEE 3rd International Conference on Biometrics: Theory, Applications, and Systems.
2009.
24. Ahonen, T., A. Hadid, and M. Pietikainen, Face Description with Local Binary Patterns:
Application to Face Recognition. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 2006. 28(12): p. 2037-2041.
25. Cao, Z., et al. Face recognition with learning-based descriptor. in 2010 IEEE Computer
Society Conference on Computer Vision and Pattern Recognition. 2010.
26. Chan, T.H., et al., PCANet: A Simple Deep Learning Baseline for Image Classification?
IEEE Transactions on Image Processing, 2015. 24(12): p. 5017-5032.
27. Lei, Z., M. Pietikäinen, and S.Z. Li, Learning Discriminant Face Descriptor. IEEE
Transactions on Pattern Analysis and Machine Intelligence, 2014. 36(2): p. 289-302.
28. Taigman, Y., et al., DeepFace: Closing the gap to human-level performance in face
verification. 2014. p. 1701-1708.
29. Sun, Y., Deep Learning Face Representation by Joint Identification-Verification, J. Huang,
Editor. 2015, ProQuest Dissertations Publishing.
30. Oscos, G.C., T.M. Khoshgoftaar, and R. Wald, Rotation invariant face recognition survey.
2015. p. 835-840.
31. Ding, C. and D. Tao, A Comprehensive Survey on Pose-Invariant Face Recognition. ACM
Transactions on Intelligent Systems and Technology (TIST), 2016. 7(3): p. 1-42.
32. Rajeshwari, J., K. Karibasappa, and M.T. Gopalkrishna, Survey on skin based face detection
on different illumination, poses and occlusion. 2015. p. 728-733.
33. Wang, Z., et al., Low-resolution face recognition: a review. International Journal of
Computer Graphics, 2014. 30(4): p. 359-386.
34. Mejda, C., et al., A Survey of 2D Face Recognition Techniques. Computers, 2016. 5(4): p.
21.
35. Ramachandra, R. and C. Busch, Presentation Attack Detection Methods for Face
Recognition Systems: A Comprehensive Survey. ACM Computing Surveys (CSUR), 2017.
50(1): p. 1-37.
36. Galbally, J., S. Marcel, and J. Fierrez, Biometric Antispoofing Methods: A Survey in Face
Recognition. Access, IEEE, 2014. 2: p. 1530-1552.
37. Hailing Zhou, A., et al., Recent Advances on Singlemodal and Multimodal Face
Recognition: A Survey. Human-Machine Systems, IEEE Transactions on, 2014. 44(6): p.
701-716.
38. Azeem, A., et al., A survey: Face recognition techniques under partial occlusion.
International Arab Journal of Information Technology, 2014. 11(1): p. 1-10.
39. Sharif, M., et al., Face recognition: A survey. Journal of Engineering Science and
Technology Review, 2017. 10(2): p. 166-177.
40. Ochoa-Villegas, M.A., et al., Addressing the illumination challenge in two-dimensional face
recognition: A survey. IET Computer Vision, 2015. 9(6): p. 978-992.
41. Zafeiriou, S., C. Zhang, and Z. Zhang, A survey on face detection in the wild: Past, present
and future. Computer Vision and Image Understanding, 2015. 138: p. 1-24.
42. Hjelmås, E. and B.K. Low, Face Detection: A Survey. Computer Vision and Image
Understanding, 2001. 83(3): p. 236-274.
43. Ranjan, R., et al., Deep Learning for Understanding Faces: Machines May Be Just as Good,
or Better, than Humans. Signal Processing Magazine, IEEE, 2018. 35(1): p. 66-83.
44. Jalali, A., R. Mallipeddi, and M. Lee, Sensitive deep convolutional neural network for face
recognition at large standoffs with small dataset. Expert Systems With Applications, 2017.
87: p. 304-315.
45. Deng, J., et al. ImageNet: A large-scale hierarchical image database. in 2009 IEEE
Conference on Computer Vision and Pattern Recognition. 2009.
46. Nech, A. and I. Kemelmacher, Level Playing Field for Million Scale Face Recognition.
2017. 3406-3415.
47. Vidit Jain, E.L.-m., FDDB: A benchmark for face detection in unconstrained settings. 2010.
48. M. Parkhi, O., A. Vedaldi, and A. Zisserman, Deep Face Recognition. Vol. 1. 2015. 41.1-
41.12.
49. Deng, J., J. Guo, and S. Zafeiriou, ArcFace: Additive Angular Margin Loss for Deep Face
Recognition. 2018.
50. Hu, G., et al., When Face Recognition Meets with Deep Learning: An Evaluation of
Convolutional Neural Networks for Face Recognition. 2016. p. 384-392.
51. Wu, X.H., Ran Sun, Zhenan Tan, Tieniu. A Light CNN for Deep Face Representation with
Noisy Labels. 2015.
52. Zhang, X., et al., Range Loss for Deep Face Recognition with Long-Tailed Training Data.
2017. p. 5419-5428.
53. Cao, Q., et al., VGGFace2: A dataset for recognising faces across pose and age. 2017.
54. Guo, Y., et al., MS-Celeb-1M: A Dataset and Benchmark for Large-Scale Face Recognition.
Vol. 9907. 2016. 87-102.
55. Ng, H.W. and S. Winkler, A data-driven approach to cleaning large face datasets. 2015.
343-347.
56. NIST, N.I.o.S.a.T. Face Recognition Technology (FERET). 2018 10/05/2018]; Available
from: https://www.nist.gov/programs-projects/face-recognition-technology-feret.
57. Yi, D., et al., Learning Face Representation from Scratch. 2014.
58. Ge, S., et al., Detecting Masked Faces in the Wild with LLE-CNNs. 2017. 426-434.
59. multipie.org. The CMU Multi-PIE Face Database. 2018; Available from:
http://www.cs.cmu.edu/afs/cs/project/PIE/MultiPie/Multi-Pie/Home.html
60. Ricanek Jr, K. and T. Tesafaye, MORPH: A longitudinal image database of normal adult
age-progression. 2006. p. 341-345.
61. Chen, B.-C., C.-S. Chen, and W.H. Hsu, Face Recognition and Retrieval Using Cross-Age
Reference Coding With Cross-Age Celebrity Dataset. Multimedia, IEEE Transactions on,
2015. 17(6): p. 804-815.
62. Wolf, L., T. Hassner, and I. Maoz. Face recognition in unconstrained videos with matched
background similarity. in CVPR 2011. 2011.
63. Grgic, M., K. Delac, and S. Grgic, SCface – surveillance cameras face database. An
International Journal, 2011. 51(3): p. 863-879.
64. Bansal, A., et al., UMDFaces: An Annotated Face Dataset for Training Deep Networks.
2016.
65. Zhang, B., et al., Directional binary code with application to PolyU near-infrared face
database. Pattern Recognition Letters, 2010. 31(14): p. 2337-2344.
66. Erdogmus, N. and S. Marcel, Spoofing in 2D face recognition with 3D masks and anti-
spoofing with Kinect. 2013. p. <xocs:firstpage xmlns:xocs=""/>.
67. Schroff, F., D. Kalenichenko, and J. Philbin, FaceNet: A unified embedding for face
recognition and clustering. 2015. p. 815-823.
68. Li, J., et al. Real-time face detection during the night. in 2017 4th International Conference
on Systems and Informatics (ICSAI). 2017.
69. Dhamecha, T.I., et al., CrowdFaceDB: Database and benchmarking for face verification in
crowd. Pattern Recognition Letters, 2018. 107: p. 17-24.
70. Savran, A., et al., Bosphorus Database for 3D Face Analysis. 2008. 47-56.
71. Min, R., N. Kose, and J.L. Dugelay, KinectFaceDB: A Kinect Database for Face
Recognition. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2014. 44(11):
p. 1534-1548.
72. Zafeiriou, S., et al., Face Recognition and Verification Using Photometric Stereo: The
Photoface Database and a Comprehensive Evaluation. IEEE Transactions on Information
Forensics and Security, 2013. 8(1): p. 121-135.
73. Phillips, P.J., et al., The FERET database and evaluation procedure for face-recognition
algorithms. Image and Vision Computing, 1998. 16(5): p. 295-306.
74. Chouchene, M., et al., Optimized parallel implementation of face detection based on GPU
component. Microprocessors and Microsystems, 2015. 39(6): p. 393-404.
75. Fan, W. and N. Bouguila, Face detection and facial expression recognition using
simultaneous clustering and feature selection via an expectation propagation statistical
learning framework. An International Journal, 2015. 74(12): p. 4303-4327.
76. Kumar, D., et al. Automated panning of video devices. in 2017 International Conference on
Signal Processing and Communication (ICSPC). 2017.
77. Hanumanthappa, M., S.R. LourdhuSuganthi, and S. Karthik. Tagging event image set using
face identification. in 2015 International Conference on Soft-Computing and Networks
Security (ICSNS). 2015.
78. Chen, W., et al., Automatic synthetic background defocus for a single portrait image. IEEE
Transactions on Consumer Electronics, 2017. 63(3): p. 234-242.
79. Lowe, D.G. Object recognition from local scale-invariant features. in Proceedings of the
Seventh IEEE International Conference on Computer Vision. 1999.
80. Dalal, N. and B. Triggs, Histograms of oriented gradients for human detection. 2005. p.
886-893.
81. Ojala, T., M. Pietikäinen, and D. Harwood, A comparative study of texture measures with
classification based on feature distributions. Pattern Recognition, 1996. 29(1): p. 51-59.
82. Dollar, P.a.T., Zhuowen and Perona, Pietro and Belongie, Serge, Integral Channel Features.
Proceedings of the British Machine Vision Conference, 2009: p. 91.1-91.11.
83. Zafeiriou, S., C. Zhang, and Z. Zhang, A survey on face detection in the wild: Past, present
and future. Computer Vision and Image Understanding, 2015. 138: p. 1.
84. Ioffe, S. and C. Szegedy, Batch Normalization: Accelerating Deep Network Training by
Reducing Internal Covariate Shift. 2015.
85. He, K., et al., Delving deep into rectifiers: Surpassing human-level performance on
imagenet classification. 2015. p. 1026-1034.
86. Uijlings, J., et al., Selective Search for Object Recognition. International Journal of
Computer Vision, 2013. 104(2): p. 154-171.
87. Ranjan, R., V.M. Patel, and R. Chellappa, HyperFace: A Deep Multi-task Learning
Framework for Face Detection, Landmark Localization, Pose Estimation, and Gender
Recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017: p. 1-1.
88. Ranjan, R., et al., An All-In-One Convolutional Neural Network for Face Analysis. 2017.
17-24.
89. Jiang, H. and E. Learned-Miller. Face Detection with the Faster R-CNN. in 2017 12th IEEE
International Conference on Automatic Face & Gesture Recognition (FG 2017). 2017.
90. Sun, X., P. Wu, and S.C.H. Hoi, Face detection using deep learning: An improved faster
RCNN approach. Neurocomputing, 2018. 299: p. 42-50.
91. Chen, D., et al., Supervised Transformer Network for Efficient Face Detection. 2016.
92. Najibi, M., et al., SSH: Single Stage Headless Face Detector. 2017.
93. Yang, S., et al., Face Detection through Scale-Friendly Deep Convolutional Networks.
2017.
94. Hu, P. and D. Ramanan, Finding Tiny Faces. 2016.
95. Zhang, S., et al., S³FD: Single Shot Scale-invariant Face Detector. 2017.
96. Wang, J., Y. Yuan, and G. Yu, Face Attention Network: An effective Face Detector for the
Occluded Faces. 2017.
97. Li, H., et al., A convolutional neural network cascade for face detection. 2015. p. 5325-
5334.
98. Zhang, K., et al., Joint Face Detection and Alignment Using Multitask Cascaded
Convolutional Networks. IEEE Signal Processing Letters, 2016. 23(10): p. 1499-1503.
99. Ranjan, R., V.M. Patel, and R. Chellappa, A deep pyramid Deformable Part Model for face
detection. 2015. p. <xocs:firstpage xmlns:xocs=""/>.
100. Farfade, S., M. J. Saberian, and L.-J. Li, Multi-view Face Detection Using Deep
Convolutional Neural Networks. 2015.
101. Yang, S., et al. From Facial Parts Responses to Face Detection: A Deep Learning
Approach. in 2015 IEEE International Conference on Computer Vision (ICCV). 2015.
102. Hinton, G., O. Vinyals, and J. Dean, Distilling the Knowledge in a Neural Network. 2015.
103. Zhang, S., et al., Detecting Face with Densely Connected Face Proposal Network. 2017. 3-
12.
104. Wei, S.-E., et al., Convolutional Pose Machines. 2016.
105. Peng, X., et al., A Recurrent Encoder-Decoder Network for Sequential Face Alignment. Vol.
9905. 2016.
106. Hoang Thai, L. and V. Nhat Truong, Face Alignment Using Active Shape Model And
Support Vector Machine. 2011.
107. J. Edwards, G., T. F. Cootes, and C. Taylor, Face Recognition Using Active Appearance
Models. 1998. 581-595.
108. Tzimiropoulos, G., Project-Out Cascaded Regression with an application to face alignment.
2015. 3659-3667.
109. Zadeh, A., T. Baltrusaitis, and L.-P. Morency, Convolutional Experts Constrained Local
Model for Facial Landmark Detection. 2017. 2051-2059.
110. Bansal, A., et al., The Do’s and Don’ts for CNN-Based Face Verification. 2017. 2545-2554.
111. Sun, Y., X. Wang, and X. Tang, Deep convolutional network cascade for facial point
detection. 2013. p. 3476-3483.
112. Zhang, Z., et al., Facial Landmark Detection by Deep Multi-task Learning. 2014.
113. Cao, X., et al., Face Alignment by Explicit Shape Regression. Vol. 107. 2012. 2887-2894.
114. Burgos-Artizzu, X., P. Perona, and P. Dollár, Robust Face Landmark Estimation under
Occlusion. 2013. 1513-1520.
115. Trigeorgis, G., et al., Mnemonic Descent Method: A Recurrent Process Applied for End-to-
End Face Alignment. 2016.
116. Jourabloo, A. and X. Liu. Large-Pose Face Alignment via CNN-Based Dense 3D Model
Fitting. in 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR).
2016.
117. Zhu, S., et al., Unconstrained Face Alignment via Cascaded Compositional Learning. 2016.
3409-3417.
118. Ren, S., et al., Face Alignment at 3000 FPS via Regressing Local Binary Features. Vol. 25.
2014. 1685-1692.
119. Zhu, S., et al., Face alignment by coarse-to-fine shape searching. 2015. 4998-5006.
120. Jourabloo, A. and X. Liu, Pose-Invariant 3D Face Alignment. 2015.
121. Hassner, T., et al. Effective face frontalization in unconstrained images. in 2015 IEEE
Conference on Computer Vision and Pattern Recognition (CVPR). 2015.
122. Zhu, X., et al., Face Alignment Across Large Poses: A 3D Solution. 2016. 146-155.
123. Kumar, A., et al., Face Alignment by Local Deep Descriptor Regression. 2016.
124. Kumar, A., A. Alavi, and R. Chellappa, KEPLER: Keypoint and Pose Estimation of
Unconstrained Faces by Learning Efficient H-CNN Regressors. 2017.
125. Li, Y., et al., Face Detection with End-to-End Integration of a ConvNet and a 3D Model.
2016.
126. Ghiasi, G. and C. Fowlkes, Laplacian Reconstruction and Refinement for Semantic
Segmentation. 2016.
127. Newell, A., K. Yang, and J. Deng, Stacked Hourglass Networks for Human Pose
Estimation. Vol. 9912. 2016. 483-499.
128. Bulat, A. and G. Tzimiropoulos, Convolutional aggregation of local evidence for large pose
face alignment. 2016.
129. Xiao, S., et al., Robust Facial Landmark Detection via Recurrent Attentive-Refinement
Networks. Vol. 9905. 2016. 57-72.
130. Noh, H., S. Hong, and B. Han, Learning Deconvolution Network for Semantic
Segmentation. 2015.
131. Ghiasi, G. and C. C. Fowlkes, Laplacian Pyramid Reconstruction and Refinement for
Semantic Segmentation. Vol. 9907. 2016. 519-534.
132. Sagonas, C., et al., 300 Faces In-The-Wild Challenge: database and results. Vol. 47. 2016.
133. LeCun, Y., et al., Handwritten digit recognition with a back-propagation network. 1990.
134. Rowley, H.A., S. Baluja, and T. Kanade, Neural network-based face detection. Pattern
Analysis and Machine Intelligence, IEEE Transactions on, 1998. 20(1): p. 23-38.
135. Gu, J., et al., Recent advances in convolutional neural networks. Pattern Recognition, 2018.
77: p. 354-377.
136. Zhang, Y., et al., Adaptive Convolutional Neural Network and Its Application in Face
Recognition. Neural Processing Letters, 2016. 43(2): p. 389-399.
137. Zou, J., et al., Convolutional neural network simplification via feature map pruning.
Computers and Electrical Engineering, 2018.
138. Parashar, A., et al., SCNN: An Accelerator for Compressed-sparse Convolutional Neural
Networks. ACM SIGARCH Computer Architecture News, 2017. 45(2): p. 27-40.
139. Wen, Y., K.Z.Z. Li, and Y. Qiao, A Discriminative Feature Learning Approach for Deep
Face Recognition. ECCV 2016. Lecture Notes in Computer Science, vol 9911. Springer,
Cham, 2016.
140. Liu, Z., et al., Deep Learning Face Attributes in the Wild. 2014.
141. Liu, W., et al., SphereFace: Deep Hypersphere Embedding for Face Recognition. 2017.
142. Sun, Y., X. Wang, and X. Tang. Deeply learned face representations are sparse, selective,
and robust. in 2015 IEEE Conference on Computer Vision and Pattern Recognition
(CVPR). 2015.
 143. Deng, W., et al., Deep Correlation Feature Learning for Face Verification in the Wild.
IEEE Signal Processing Letters, 2017. 24(12): p. 1877-1881.
144. Wang, H., et al., CosFace: Large Margin Cosine Loss for Deep Face Recognition. 2018.
145. Ranjan, R., C. D. Castillo, and R. Chellappa, L2-constrained Softmax Loss for
Discriminative Face Verification. 2017.
146. Wang, F., et al., Additive Margin Softmax for Face Verification. IEEE Signal Processing
Letters, 2018. 25(7): p. 926-930.
147. Ding, C. and D. Tao, Trunk-Branch Ensemble Convolutional Neural Networks for Video-
Based Face Recognition. Pattern Analysis and Machine Intelligence, IEEE Transactions on,
2018. 40(4): p. 1002-1014.
148. Taigman, Y., et al., Web-scale training for face identification. 2015. p. 2746-2754.
149. Gu, J.L. and H.J. Peng, Incremental convolution neural network and its application in face
detection. Vol. 21. 2009. 2441-2445.
150. Sun, Y., et al., DeepID3: Face Recognition with Very Deep Neural Networks. 2015.
151. Wan, W. and J. Chen. Occlusion robust face recognition based on mask learning. in 2017
IEEE International Conference on Image Processing (ICIP). 2017.
152. Kang, D., et al., Nighttime face recognition at large standoff: Cross-distance and cross-
spectral matching. Pattern Recognition, 2014. 47(12): p. 3750-3766.
153. Huang, G.B., H. Lee, and E. Learned-Miller, Learning hierarchical representations for face
verification with convolutional deep belief networks. 2012. 2518-2525.
154. AbdAlmageed, W., et al., Face Recognition Using Deep Multi-Pose Representations. 2016.
1-9.
155. Masi, I., et al., Do we really need to collect millions of faces for effective face recognition?
2016. p. 579-596.

View publication stats