Facial Proﬁle Biometrics: Domain Adaptation and Deep Learning

Approaches

Malak Alamri

1,2 a

and Sasan Mahmoodi

College of Computer and Information Science, Computer Science Department, Jouf University, Al jouf, K.S.A.

School of Electronic and Computer Science, University of Southampton, Southampton, U.K.

Keywords:

Facial Proﬁle, Biometrics, Bilateral Symmetry, Domain Adaptation, Deep Learning.

Abstract:

Previous studies indicate that human facial proﬁles are considered as a biometric modality and there is a

bilateral symmetry in facial proﬁle biometrics. This study examines the bilateral symmetry of the human

face proﬁles and presents the analysis of facial proﬁle images for recognition. A method from few-shot

learning framework is proposed here to extract facial proﬁle features. Based on domain adaptation and reverse

validation, we introduce a technique known as reverse learning (RL) in this paper for the same side proﬁles to

achieve a recognition rate of 85%. In addition, to investigate bilateral symmetry, our reverse learning model is

trained and validated on the left side face proﬁles to measure the cross recognition of 71% for right side face

proﬁles. Also in this paper, we assume that the right face proﬁles are unlabelled, and we therefore apply our

reverse learning method to include the right face proﬁles in the validation stage to improve the performance

of our algorithm for opposite side recognition. Our numerical experiments indicate an accuracy of 84.5% for

cross recognition which, to the best of our knowledge, demonstrates higher performance than the state-of-

the-art methods for datasets with similar number of subjects. Our algorithm based on few-shot learning can

achieve high accuracies for a dataset characterized with as low as four samples per group.

1 INTRODUCTION

The increased global use of biometrics technology has

resulted in an increase in security threats. Compa-

nies and government agencies have several difﬁculties

recognising authorized users. Tokens, such as identi-

ﬁcation cards and passwords, are the most common

security system authentication methods. However,

they are becoming increasingly unreliable due to du-

plication and the limited capacity of human memory

(Abdelwhab and Viriri, 2018). Thus, the use of bio-

metrics, such as facial recognition and ﬁngerprints, is

growing in popularity worldwide. Face recognition

is a very important security measure that works by

identifying possible suspects in videos or surveillance

frames. A video may only record a small section of

the face, making correct identiﬁcation impossible in

some cases (see Figure 1).

It should be highlighted that the majority of bio-

metrics literature focuses on the front face. Despite

its signiﬁcance, the facial proﬁle has received less at-

tention. It is common knowledge that facial recogni-

https://orcid.org/0000-0002-3484-0994

Figure 1: The Surveillance Camera Footage Released in

Link with a Stabbing at Block 1 Club in 2019 (Everett,

2019).

tion systems may rely heavily on the ability to com-

pare and match the features of two face images. How-

ever, a proﬁle view of a face may disclose certain as-

pects of the face’s structure that are not visible in a

frontal view (Alamri and Mahmoodi, 2022). Com-

bining the matching results derived from the frontal

and proﬁle views of a face can assist in reducing in-

stances of incorrect identiﬁcation. In the past decade,

a variety of algorithms for automatically recognising

individuals based on their facial proﬁles have been

introduced. Several facial proﬁle identiﬁcation tech-

niques are commonly used, including Scale-space ﬁl-

1046

Alamri, M. and Mahmoodi, S.

Facial Proﬁle Biometrics: Domain Adaptation and Deep Learning Approaches.

DOI: 10.5220/0013365700003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 1, pages 1046-1053

ISBN: 978-989-758-731-3; ISSN: 2184-4305

tering (Lipo

cak and Loncaric, 1999), Dynamic time

warping (Bhanu and Zhou, 2004), Attributed string

(Gao and Leung, 2002) and Hidden Markov model

(HMM) (Wallhoff and Rigoll, 2001; Wallhoff et al.,

2001). A facial proﬁle biometric system enhances the

effectiveness of multiple facial identiﬁcation systems

and makes recognition more realistic.

There are several deep learning-based methods

that have explicitly investigated facial proﬁle recog-

nition and proposed methods for extracting discrim-

inative features from facial proﬁles. Sengupta et al.

illustrated how several algorithms performed when

using a restricted protocol and how each one de-

graded from frontal–frontal to frontal–proﬁle (Sen-

gupta et al., 2016). In their study, the frontal-

frontal and frontal-proﬁle experiment achieved clas-

siﬁcation accuracy of 96.40% and 84.91%, respec-

tively, using a deep features-based method. More-

over, in recent years, generative models, such as

generative adversarial network (GAN)-based meth-

ods, have been widely used to synthesise the frontal

view from the proﬁle view in order to improve fa-

cial proﬁle recognition systems (Zhao et al., 2018;

Li et al., 2019; Yin et al., 2020). In addition, deep-

learning-based methods have also demonstrated high

levels of performance. Facial recognition has been

greatly improved by deep learning techniques, which

are trained on a large-scale dataset and demonstrate

high-level recognition rates under challenging condi-

tions (Parkhi et al., 2015; Deng et al., 2019; Meng

et al., 2021). However, it is important to note that

all deep learning methods require a large number of

samples per class during training to achieve accept-

able performance levels.

Our contributions in this paper are as follows: 1)

higher recognition rate than the literature is achieved

on large datasets containing more than 200 subjects,

2) a technique in the framework of few-shot learn-

ing is proposed here to extract features from facial

proﬁle images even with as low as four samples per

class, 3) our method is based on domain adaptation

(Wang et al., 2020), and reverse validation (Morvant

et al., 2011; Zhong et al., 2010) to guarantee a high

accuracy to improve recognition performance for bi-

lateral symmetry by minimizing the conditional dis-

tributions between the training and validation data

(Zhong et al., 2010), 4) the ability to identify mir-

ror symmetric images is used generally since mirrored

images were captured in different conditions than the

ﬁrst set of images, 5) facial proﬁle appears largely

to be bilaterally symmetric. We are the ﬁrst to use

few-shot learning (FSL) to examine the bilateral sym-

metry of a facial proﬁle. Additionally, we consider

challenging facial proﬁle images with various planar

poses, while measuring the performance by training

pre-processing facial proﬁle dataset and applying face

alignment to register all facial proﬁles for fair com-

parisons. The work given here represents the most

recent state-of-the-art methods in the recognition of

facial proﬁles and their bilateral symmetry.

The second section of this paper discusses the im-

age preprocessing phases, including face detection,

face alignment. In addition, the third section explains

the feature extraction stage using two few-shot learn-

ing techniques. Section four describes the experimen-

tal design, recognition performance, and subsequent

discussions. Moreover, Section ﬁve summarises the

conclusions drawn from the previous sections and

highlights the implications of the study.

2 FACIAL PROFILE AND

LANDMARKS DETECTION

To begin with, our focus was on detecting the bound-

ing box of a facial proﬁle in all of the images in the

dataset and then identifying the important landmarks

within the bounding box as described in (Alamri,

2024). To achieve this, we used a pre-trained his-

togram of oriented gradients (HOG) with a linear sup-

port vector machine (SVM) object detector, as intro-

duced in (Dalal and Triggs, 2005).

For landmarks detection, the ensemble of regres-

sion trees (ERT) (Kazemi and Sullivan, 2014) method

is used to directly estimate the positions of facial land-

marks by utilizing a sparse subset of pixel intensities.

The majority of face alignment algorithms rely on a

face detection bounding box to initialise the shape.

Thus, to detect facial landmarks, the face must ﬁrst

be extracted from the image. This extracted region of

interest (ROI) is then used to obtain the landmarks.

Shape predictors, also known as landmark predictors,

are used to predict key (x,y)-coordinates of a given

’shape’. Following the algorithm proposed by Kazemi

and Sullivan (Tzimiropoulos and Pantic, 2013), shape

predictors are used to locate individual facial struc-

tures such as the eyes, eyebrows, nose, lips/mouth,

and jawline. They require two inputs: the greyscale

version of the image and a rectangle object contain-

ing the coordinates of the face area.

The alignment and transformation framework was

based on the relevant literature (Zeng and Yi, 2011;

Walker et al., 1991). These transformations involved

estimating a combination of rotation, translation, and

scale that mapped the key points from one set to an-

other on a template.

Facial Proﬁle Biometrics: Domain Adaptation and Deep Learning Approaches

1047

Figure 2: Convolutional neural network architecture.

3 EXTRACTING VISUAL

FEATURES FROM FACIAL

PROFILE IMAGES

Deep neural networks are widely used in computer

vision for extracting features. In this section, we pro-

pose a structure inspired by previous research (Wang

et al., 2020; Parnami and Lee, 2022; Garcia and

Bruna, 2017) to leverage few-shot learning based on

the ResNet50 neural network (He et al., 2016) and

ArcFace (Deng et al., 2019). A typical CNN consists

of feature extraction and classiﬁcation components.

During training, the model learns the unique facial

features and generates feature embeddings through

the feature extraction process. In the FSL strategy,

after training is completed, it becomes possible to

skip the classiﬁcation step and generate feature em-

beddings for each face image. Figure 2 illustrates the

CNN architecture and the key stages used for feature

extraction.

• ResNet50 (He et al., 2016), short for residual net-

work, is a speciﬁc type of CNN known for its abil-

ity to maintain a lower error rate, even in deeper

networks. Our network accepts input images of

size 244 × 244 pixels and computes 7 × 7 grid

feature maps in the last layer before the fully con-

nected network. The pre-trained model was opti-

mised on the ImageNet dataset, which consisted

of multiple classes. Due to the difference between

the datasets used to train the networks and our tar-

get data, we chose to exclude the output of the

last fully connected layer. Our numerical experi-

ments demonstrated that the learned feature space

efﬁciently represented human faces using these 7

× 7 × 2048 feature maps, resulting in 49 2048-

dimensional vectors.

We then applied Sequential Floating Forward Se-

lection (SFFS) (Shirbani and Soltanian Zadeh,

2013) and principal component analysis (PCA)

(Jolliffe, 2002) algorithms as the next step to

reduce the dimensionality of the feature space

which resulted in achieving the best performance.

After reducing the dimensionality, we used the

feature vectors to train a KNearest Neighbour

(KNN) recognition method with k = 4 identiﬁed

as the optimal value to achieve the best perfor-

mance.

• ArcFace (Deng et al., 2019) is an innovative deep

face recognition algorithm proposed by Jiankang

Deng et al. It is considered a discriminative

model. The authors proposed the additive angu-

lar margin loss function, which has proven to be

highly effective in obtaining discriminative fea-

tures for facial recognition. It consistently outper-

forms other state-of-the-art loss functions. Arc-

Face addresses the challenge of effectively learn-

ing discriminative face features by incorporating

an angular margin that pushes the learned features

of different classes apart in the angular space. By

enhancing the separability between classes, Ar-

cFace signiﬁcantly improves the model’s ability

to distinguish between similar faces. This model

is trained using the LFW dataset (Huang et al.,

2008) to obtain pre-trained weights. The LFW

dataset consists of 13,233 facial photos collected

from the web. The ArcFace model expects inputs

of size 112 × 112 and returns 512-dimensional

vector representations. Importantly, ArcFace sim-

pliﬁes the process by compressing the extracted

features to only 512 components, thereby elimi-

nating the need for PCA.

We selected ResNet50 (Montero et al., 2022) as

the backbone for all of the network architectures

tested in the ArcFace repository, due to its optimal

balance between accuracy and parameter count.

In this context, the base model was ResNet50, and

we utilised ArcFace as the loss function. Gen-

erally, the term backbone refers to the feature-

extracting network responsible for processing in-

put data into a speciﬁc feature representation. The

backbone plays a crucial role in extracting and

encoding features from the input data, captur-

ing both low-level and high-level features. Af-

ter feature extraction, we calculated the ArcFace

loss and used it to update the network parame-

ters through backpropagation. We chose ArcFace

(Deng et al., 2019) as the baseline for two rea-

sons: it employs a SoftMax-loss-based methodol-

ogy, eliminating the need for an exhaustive train-

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

1048

ing data preparation stage, and it has been demon-

strated to yield the best results for the original face

recognition task.

4 FACIAL PROFILE

RECOGNITION USING

DOMAIN ADAPTATION AND

REVERSE VALIDATION

METHODOLOGY

Domain adaptation, a subﬁeld of machine learning,

aims to train a model on a source dataset and achieve

a high level of accuracy on a target dataset that differs

signiﬁcantly from the source (Farahani et al., 2021).

Conversely, if some samples in the target dataset lack

labels, reverse validation (Morvant et al., 2011; Wang

et al., 2020) is employed to identify the best model

that minimises the difference between the conditional

distributions of the source and target datasets. In

this study, we propose a method grounded in do-

main adaptation and reverse validation, assuming that

some facial proﬁle image samples lack labels and ex-

hibit distinct conditional distributions from the source

(training) dataset. Inspired by (Ganin et al., 2016),

we introduced an algorithm called reverse learning

(RL) based on domain adaptation and reverse valida-

tion. Unlike conventional prediction methods, our ap-

proach employed a two-step prediction process. We

use SFFS to select the most optimal features from the

distributions of the training and validation data. In our

RL algorithm, the two-step prediction helped identify

the best model with optimised hyper-parameters and

features between the training and validation datasets,

addressing issues associated with covariate shift as-

sumptions (David et al., 2010).

Figure 3: Training process of our reverse learning (RL) al-

gorithm: 1 training the model on the training set, 2

training the model on the validation set.

In the ﬁrst step, the process began with a simple

training procedure using SFFS, As shown in Figure 3.

Figure 4: Sample facial proﬁle images from the

XM2VTSDB dataset.

In the second step of the prediction process, after

making predictions for the validation set, we trained

on the validation set, made predictions for the training

set, and aimed to minimise errors using the SFFS al-

gorithm. In the second prediction step, the validation

set served as the training set. The second prediction

was ﬁnally used to quantify the error on the training

set, named the reverse learning metric. Finally, we

evaluated the performance of our system by present-

ing unseen test data from the target dataset.

4.1 Face Proﬁle Dataset

We assessed our method using the XM2VTSDB

database, a research resource established and main-

tained by the University of Surrey (see (Messer et al.,

1999) for details). This database is an extended

version of the M2VTS database, as it comprises

more video recordings of each subject during each

session compared to M2VTS. Participants from the

XM2VTSDB database attended four sessions and the

database was developed over a signiﬁcant period, en-

abling a wide range of appearance variations of the

individual subjects (see Figure 4).

4.2 Performance Analysis

In this section, we introduce a domain adaptation ap-

proach that is employed to assess the effect of this

strategy on the recognition accuracy of facial proﬁles

from opposite sides of the face. We detected a notice-

able shift in data within the same feature space when

evaluating the model ﬁtted on the training set. This

discrepancy arose from the limited number of sam-

ples per class and the variations in subjects’ appear-

ances across the four images, as well as between the

training and test sets. To address this challenge, we

employed a technique that aimed to identify the gap

between the training, validation, and test sets.

We conducted a series of experiments on side-

view face recognition to assess the effectiveness of

the RL method compared to traditional approaches.

We then employed a similar technique in two distinct

settings:

Facial Proﬁle Biometrics: Domain Adaptation and Deep Learning Approaches

1049

4.2.1 Experiment for Same View Facial Proﬁle

(Left Side)

In this setting, we employed the leave-one-out cross-

validation (LOOCV) method to measure the recog-

nition rate. In addition, we opted to use KNN as a

simple classiﬁer. The model was trained and evalu-

ated from the perspective of the left side, with the test

data also originating from the left side. We evalu-

ated the performance of this experiment by analysing

four samples from the left side view of each indi-

vidual in the dataset. For this experiment, 80% of

the data on the same side was allocated for training

and validation, while the remaining 20% served as

the unseen testing set. By utilising ResNet50 fea-

tures, we achieved an average recognition accuracy of

85% when training KNN with 80% of the samples in

the dataset (traditional strategy). Interestingly, we ob-

served an increase in accuracy to 91% when employ-

ing the RL strategy. However, when using ArcFace

features, the average recognition accuracy dropped to

71%. Notably, this accuracy remained consistent even

with the application of the RL approach.

Figure 5: Recognition via CMC performance for ResNet50

features.

4.2.2 Experiment for Mirror Bilateral

Symmetry View (Left and Right Side)

In this case, the facial proﬁle side images were ini-

tially mirrored from right to left, following the way in

which previous processes had been applied to detect

facial proﬁles. We evaluated the performance of this

experiment by analysing eight samples from the left

and right side view of each individual in the dataset.

In this experiment, we trained and validated on the

left side and measured the recognition rate on the right

side as unseen test data. Our numerical experiments

revealed that utilising RL can enhance recognition ac-

curacy to 75%, compared to 70% with the traditional

method, based on ResNet50 features.

Additionally, the model was trained on left-side

facial proﬁles and validated on unlabelled right-side

proﬁles. Testing was then conducted using unseen

data from the right side. By using right facial proﬁles

as unlabelled data in the validation process, a recog-

nition rate of 82% was achieved. This marked a sig-

niﬁcant improvement of 12% compared to traditional

methods. Interestingly, when ArcFace features were

utilised, the accuracy remained consistent across all

three techniques, at 56%. However, it is worth noting

that the RL approach did not lead to any enhancement

in accuracy.

One of the most important techniques for evaluat-

ing model performance is cumulative matching char-

acteristics (CMC) curves (DeCann and Ross, 2013).

The CMC curve shows the recognition at different

ranks, indicating the probability of ﬁnding an accu-

rate match at a particular rank. Figures 5 and 6 present

the recognition performance of facial proﬁles. In Fig-

ure 5, which used ResNet50 features, the accuracy for

same-side recognition improved from 85% at rank-1

to 96% at rank-10 when employing RL. Moreover, the

accuracy for BS increased from 71% at rank-1 to 91%

at rank-10 when using only right-side images as test

data. However, when right-side images were also in-

cluded in the validation stage, the accuracy rose from

82% at rank-1 to 96% at rank-10.

In Figure 6, which used ArcFace features, the ac-

curacy for RL on the same side improved from 71%

at rank-1 to 87% at rank-10. Additionally, the ac-

curacy for BS using only right-side images as test

data improved from 56% at rank-1 to 81% at rank-10.

Additionally, the accuracy for BS using only right-

side images as test data improved from 56% at rank-

1 to 81% at rank-10. Similarly, when right-side im-

ages were also used in the validation stage, the accu-

racy increased from 56% at rank-1 to 83% at rank-

10. The inferior performance of ArcFace compared

to ResNet50 can be attributed to the geometric nature

of its loss function, which does not effectively opti-

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

1050

Table 1: Recognition rates with and without domain adaptation based on RL algorithm; (BS) bilateral symmetry, (L) left side,

(R) right side, (SD) standard deviation.

View Method

Dataset

Model

ResNet50 ArcFace

Training Validation Test Accuracy SD Accuracy SD

Same side

Traditional L L L 85% 0.016 71% 0.012

RL L L L 91% 0.014 71% 0.011

Traditional L L R 70% 0.021 56% 0.009

RL L L R 75% 0.011 56% 0.008

RL L R R 82% 0.042 56% 0.041

Figure 6: Recognition via CMC performance for ArcFace

features.

mize for the best results. In particular, ArcFace’s loss

function can be limiting, especially in cases such as

facial proﬁle. Our numerical experiment results are

summarised in Table 1 and reveal the following two

ﬁndings:

1) Using ResNet50 features, the recognition rate for

the same side exceeded 85% with the traditional

method and improved by 6% with the RL method.

For the opposite side, the recognition rate using

the traditional method is 70%, which increased to

75% with the RL method when the left side was

included in both the validation and training sets.

Notably, the accuracy signiﬁcantly improved by

12% when right facial proﬁles were included in

the validation process.

2) When employing ArcFace features, the recog-

nition rates for both the same and opposite

sides remained unchanged, and the model perfor-

mance did not alter even after introducing the RL

method.

Our numerical experiments utilising ResNet50 in-

dicated that the RL method proposed here outper-

formed the state-of-the-art for datasets containing

more than 200 subjects. The dataset utilised in this

study differs from the one used in previous work. To

clarify, our dataset presented greater challenges due to

the following reasons: 1) it comprised 230 subjects,

which is substantially larger than all datasets consid-

ered in (Ding et al., 2013; Bhanu and Zhou, 2004; Xu

and Mu, 2007), 2) some of the facial proﬁles in our

dataset were occluded, yet our method still achieved

a high recognition rate of 91% using our RL method,

surpassing all methods presented in (Ding et al., 2013;

Bhanu and Zhou, 2004; Xu and Mu, 2007), and 3) in

the method proposed here, we assumed that the val-

idation set was unlabelled. Such an assumption ren-

dered our dataset more challenging than those in the

literature.

In our study of 230 subjects, we achieved a cross-

recognition rate of 82%, which is the highest among

our results. This performance can be compared to the

work presented in (Toygar et al., 2018), where only 46

subjects with left and right proﬁle images were used.

The best cross-recognition rate reported in (Toygar

et al., 2018) was 81.52%, when the model was trained

on the left side and tested on the right side. However,

in our study, facial proﬁles with a larger number of

subjects achieved an 82% cross-recognition rate. This

indicates that facial proﬁles contain sufﬁcient infor-

mation to be considered an independent and important

biometric modality. Table 2 presents the recognition

rates of facial proﬁles from various methods found in

the literature.

Facial Proﬁle Biometrics: Domain Adaptation and Deep Learning Approaches

1051

Table 2: Recognition rates of facial proﬁles in the literature

with various methods.

Publication Dataset Method Accuracy

Same side view

(Ding et al.,

2013)

44 DWT +

92.50%

(Bhanu and

Zhou, 2004)

30 DWT 90%

(Xu and Mu,

2007)

38 PCA 77.63%

Other side view

(Toygar et al.,

2018)

46 BSIF +

LPQ +

LBP

81.52%

Table 1 presents the results which demonstrate

that ResNet50 performed better than ArcFace in

both same-side and cross-recognition experiments.

While ArcFace demonstrated strong performance

when dealing with a slightly angled view of faces

and excelled in large-scale face identiﬁcation tasks,

its performance was lower when applied to facial pro-

ﬁles.

5 CONCLUSION

This study introduced a method for facial proﬁle

recognition that combines few-shot learning, domain

adaptation, and reverse validation techniques. We em-

ployed a similarity registration technique to ensure

the precise alignment of all facial proﬁle images. By

utilising two pre-trained CNN models, ResNet50 and

ArcFace, we implemented few-shot learning. Among

these models, ResNet50 demonstrated superior per-

formance compared to ArcFace. Speciﬁcally, our RL

algorithm, which utilised ResNet50, outperformed

the traditional training methods discussed in this

study. The results obtained from our RL method re-

veal signiﬁcant improvements in classiﬁcation rates.

Speciﬁcally, the recognition rate for same side (left

side) reached 91%, surpassing the state-of-the-art per-

formance achieved with datasets of similar sizes.

Additionally, promising outcomes were observed

in cross-recognition, with a rate of 82% achieved

when right-side images were used in the validation

stage. Furthermore, a recognition rate of 75% was

attained when left-side images were employed in val-

idation, with right-side images used for testing. Nu-

merical experiments indicated a notable 7% enhance-

ment in cross-recognition accuracy when right-side

faces were included in the validation process. There-

fore, we can conclude that depending on the applica-

tion, recognition is viable even when facial proﬁles

are the sole biometric modality available, including

scenarios involving bilateral symmetry. Finally, we

obtained promising results using only four samples

per subject. Alternatively, leveraging a transfer learn-

ing or training a neural network from scratch could

have been considered with more training samples.

REFERENCES

Abdelwhab, A. and Viriri, S. (2018). A survey on soft bio-

metrics for human identiﬁcation. Machine Learning

and Biometrics, 37.

Alamri, M. (2024). Facial proﬁle recognition using com-

parative soft biometrics. PhD thesis, University of

Southampton.

Alamri, M. and Mahmoodi, S. (2022). Face proﬁle bio-

metric enhanced by eyewitness testimonies. In 2022

26th International Conference on Pattern Recognition

(ICPR), pages 1127–1133. IEEE.

Bhanu, B. and Zhou, X. (2004). Face recognition from face

proﬁle using dynamic time warping. In Proceedings of

the 17th International Conference on Pattern Recog-

nition, 2004. ICPR 2004., volume 4, pages 499–502.

IEEE.

Dalal, N. and Triggs, B. (2005). Histograms of oriented

gradients for human detection. In 2005 IEEE com-

puter society conference on computer vision and pat-

tern recognition (CVPR’05), volume 1, pages 886–

893. Ieee.

David, S. B., Lu, T., Luu, T., and P

al, D. (2010). Impossi-

bility theorems for domain adaptation. In Proceedings

of the Thirteenth International Conference on Artiﬁ-

cial Intelligence and Statistics, pages 129–136. JMLR

Workshop and Conference Proceedings.

DeCann, B. and Ross, A. (2013). Relating roc and cmc

curves via the biometric menagerie. In 2013 IEEE

Sixth International Conference on Biometrics: The-

ory, Applications and Systems (BTAS), pages 1–8.

IEEE.

Deng, J., Guo, J., Xue, N., and Zafeiriou, S. (2019). Ar-

cface: Additive angular margin loss for deep face

recognition. In Proceedings of the IEEE/CVF con-

ference on computer vision and pattern recognition,

pages 4690–4699.

Ding, S., Zhai, Q., Zheng, Y. F., and Xuan, D. (2013). Side-

view face authentication based on wavelet and ran-

dom forest with subsets. In 2013 IEEE International

Conference on Intelligence and Security Informatics,

pages 76–81. IEEE.

Everett, A. (17-4-2019). Warrington guardian-new cctv im-

ages after man stabbed in face in block 1 club.

Farahani, A., Voghoei, S., Rasheed, K., and Arabnia, H. R.

(2021). A brief review of domain adaptation. Ad-

vances in data science and information engineering:

proceedings from ICDATA 2020 and IKE 2020, pages

877–894.

Ganin, Y., Ustinova, E., Ajakan, H., Germain, P.,

Larochelle, H., Laviolette, F., Marchand, M., and

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

1052

Lempitsky, V. (2016). Domain-adversarial training of

neural networks. The journal of machine learning re-

search, 17(1):2096–2030.

Gao, Y. and Leung, M. K. (2002). Human face proﬁle

recognition using attributed string. Pattern Recogni-

tion, 35(2):353–360.

Garcia, V. and Bruna, J. (2017). Few-shot learn-

ing with graph neural networks. arXiv preprint

arXiv:1711.04043.

He, K., Zhang, X., Ren, S., and Sun, J. (2016). Deep resid-

ual learning for image recognition. In Proceedings of

the IEEE conference on computer vision and pattern

recognition, pages 770–778.

Huang, G. B., Mattar, M., Berg, T., and Learned-Miller,

E. (2008). Labeled faces in the wild: A database

forstudying face recognition in unconstrained envi-

ronments. In Workshop on faces in’Real-Life’Images:

detection, alignment, and recognition.

Jolliffe, I. T. (2002). Principal component analysis for spe-

cial types of data. Springer.

Kazemi, V. and Sullivan, J. (2014). One millisecond face

alignment with an ensemble of regression trees. In

Proceedings of the IEEE conference on computer vi-

sion and pattern recognition, pages 1867–1874.

Li, P., Wu, X., Hu, Y., He, R., and Sun, Z. (2019). M2fpa: A

multi-yaw multi-pitch high-quality dataset and bench-

mark for facial pose analysis. In Proceedings of the

IEEE/CVF international conference on computer vi-

sion, pages 10043–10051.

Lipo

cak, Z. and Loncaric, S. (1999). A scale-space ap-

proach to face recognition from proﬁles. In Computer

Analysis of Images and Patterns: 8th International

Conference, CAIP’99 Ljubljana, Slovenia, September

1–3, 1999 Proceedings 8, pages 243–250. Springer.

Meng, Q., Zhao, S., Huang, Z., and Zhou, F. (2021).

Magface: A universal representation for face recog-

nition and quality assessment. In Proceedings of the

IEEE/CVF Conference on Computer Vision and Pat-

tern Recognition, pages 14225–14234.

Messer, K., Matas, J., Kittler, J., Luettin, J., and Maitre, G.

(1999). Xm2vtsdb: The extended m2vts database. In

Second international conference on audio and video-

based biometric person authentication, volume 964,

pages 965–966. Citeseer.

Montero, D., Nieto, M., Leskovsky, P., and Aginako, N.

(2022). Boosting masked face recognition with multi-

task arcface. In 2022 16th International Conference

on Signal-Image Technology & Internet-Based Sys-

tems (SITIS), pages 184–189. IEEE.

Morvant, E., Habrard, A., and Ayache, S. (2011). Sparse

domain adaptation in projection spaces based on good

similarity functions. In 2011 IEEE 11th International

Conference on Data Mining, pages 457–466. IEEE.

Parkhi, O. M., Vedaldi, A., and Zisserman, A. (2015). Deep

face recognition.

Parnami, A. and Lee, M. (2022). Learning from few exam-

ples: A summary of approaches to few-shot learning.

arXiv preprint arXiv:2203.04291.

Sengupta, S., Chen, J.-C., Castillo, C., Patel, V. M., Chel-

lappa, R., and Jacobs, D. W. (2016). Frontal to proﬁle

face veriﬁcation in the wild. In 2016 IEEE winter con-

ference on applications of computer vision (WACV),

pages 1–9. IEEE.

Shirbani, F. and Soltanian Zadeh, H. (2013). Fast sffs-based

algorithm for feature selection in biomedical datasets.

AUT Journal of Electrical Engineering, 45(2):43–56.

Toygar,

O., Alqaralleh, E., and Afaneh, A. (2018). Sym-

metric ear and proﬁle face fusion for identical twins

and non-twins recognition. Signal, Image and Video

Processing, 12:1157–1164.

Tzimiropoulos, G. and Pantic, M. (2013). Optimization

problems for fast aam ﬁtting in-the-wild. In Proceed-

ings of the IEEE international conference on com-

puter vision, pages 593–600.

Walker, M. W., Shao, L., and Volz, R. A. (1991). Estimat-

ing 3-d location parameters using dual number quater-

nions. CVGIP: image understanding, 54(3):358–367.

Wallhoff, F., Muller, S., and Rigoll, G. (2001). Recognition

of face proﬁles from the mugshot database using a hy-

brid connectionist/hmm approach. In 2001 IEEE In-

ternational Conference on Acoustics, Speech, and Sig-

nal Processing. Proceedings (Cat. No. 01CH37221),

volume 3, pages 1489–1492. IEEE.

Wallhoff, F. and Rigoll, G. (2001). A novel hybrid face pro-

ﬁle recognition system using the feret and mugshot

databases. In Proceedings 2001 International Con-

ference on Image Processing (Cat. No. 01CH37205),

volume 1, pages 1014–1017. IEEE.

Wang, Y., Yao, Q., Kwok, J. T., and Ni, L. M. (2020). Gen-

eralizing from a few examples: A survey on few-shot

learning. ACM computing surveys (csur), 53(3):1–34.

Xu, X. and Mu, Z. (2007). Feature fusion method based on

kcca for ear and proﬁle face based multimodal recog-

nition. In 2007 IEEE international conference on au-

tomation and logistics, pages 620–623. IEEE.

Yin, Y., Jiang, S., Robinson, J. P., and Fu, Y. (2020). Dual-

attention gan for large-pose face frontalization. In

2020 15th IEEE international conference on auto-

matic face and gesture recognition (FG 2020), pages

249–256. IEEE.

Zeng, H. and Yi, Q. (2011). Quaternion-based iterative so-

lution of three-dimensional coordinate transformation

problem. J. Comput., 6(7):1361–1368.

Zhao, J., Cheng, Y., Xu, Y., Xiong, L., Li, J., Zhao, F.,

Jayashree, K., Pranata, S., Shen, S., Xing, J., et al.

(2018). Towards pose invariant face recognition in

the wild. In Proceedings of the IEEE conference on

computer vision and pattern recognition, pages 2207–

2216.

Zhong, E., Fan, W., Yang, Q., Verscheure, O., and Ren,

J. (2010). Cross validation framework to choose

amongst models and datasets for transfer learning.

In Machine Learning and Knowledge Discovery in

Databases: European Conference, ECML PKDD

2010, Barcelona, Spain, September 20-24, 2010, Pro-

ceedings, Part III 21, pages 547–562. Springer.

Facial Proﬁle Biometrics: Domain Adaptation and Deep Learning Approaches

1053