Technical Progress of Face Recognition Technology in Occluded and

Unobstructed Environment

Yanqi Peng

School of Computer Science & Technology, Beijing Jiaotong University, Beijing, China

Keywords: Face Recognition, Unobstructed Recognition, Occluded Recognition, Deep Learning, Dataset.

Abstract: With the development of society, face recognition technology has become increasingly widespread across

various fields, and the accuracy of unobstructed face recognition has steadily improved. However, as real-

world face detection scenarios become more complex and variable, the challenge of occluded face recognition

has emerged as a significant research focus in recent years. This article analyzes and compares the two distinct

approaches to face recognition: unobstructed face recognition and occluded face recognition, discussing their

respective challenges and advancements. It also provides an overview of commonly used face recognition

datasets. Finally, the paper looks ahead to the future development and potential breakthroughs in both types

of face recognition technologies. By synthesizing and integrating research on these two approaches, this paper

helps researchers understand the current state of face recognition technology, highlights existing gaps in

current research, encourages innovation and further exploration, and lays the groundwork for the introduction

of new methodologies in future face recognition studies.

1 INTRODUCTION

As an important research direction in the field of

computer vision, face recognition technology has

been widely used in recent years, especially in

security monitoring, financial payment, intelligent

transportation and identity authentication and other

fields, becoming a key technology to improve

automation and security. With the rise of deep

learning technology, the accuracy and application

range of face recognition have been greatly improved,

especially based on convolutional neural network

(CNN) and generative adversarial network (GAN)

and other methods, which have promoted the rapid

development of face recognition. However, in

practical applications, face recognition technology

still faces many challenges, especially under the

influence of environmental factors and occlusion

conditions, the recognition performance may be

significantly reduced. Face occlusion is an important

factor affecting the accuracy of the recognition

system, occlusion from a variety of situations, such as

wearing masks, glasses, hair, hand occlusion and so

on. Especially in public places and specific scenes,

the occlusion phenomenon is more common. This

https://orcid.org/0009-0000-5515-0004

brings great technical problems to the existing

recognition system. Under the condition of no

occlusion, traditional face recognition methods can

usually provide high accuracy because the feature

area of the face is clear and can be fully utilized.

However, when the face is partially obstructed, the

existing system cannot fully capture the important

visual information, resulting in a significant decline

in recognition accuracy, or even cannot complete the

recognition task. This paper aims to discuss the

application progress of face recognition technology in

unobstructed and occluded environments, analyze the

advantages and disadvantages of current main

technologies, and discuss the possible development

direction in the future. By reviewing the existing

occlusion recognition methods and technical

innovations, this paper aims to provide references for

future improvements in the adaptation ability and

robustness of face recognition systems in complex

environments.

Peng, Y.

Technical Progress of Face Recognition Technology in Occluded and Unobstructed Environment.

DOI: 10.5220/0013677400004670

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

In Proceedings of the 2nd International Conference on Data Science and Engineering (ICDSE 2025), pages 21-25

ISBN: 978-989-758-765-8

2 FACE RECOGNITION IN AN

UNOBSTRUCTED ENVIRONMENT

Unobstructed face recognition refers to the feature

extraction (facial collection features and texture

features) of the face image detected from the static

image or video image under the premise of no

covering object, and then comparing it with the

database, and finally making a decision to determine

the identity of the recognized person. This technology

is the foundation of the field of face recognition and

has a wide range of research significance and

application value. Unobstructed face recognition can

provide performance judgment benchmarks for other

types of face recognition (including obstructed face

recognition and low-resolution face recognition),

which is also the basis for the implementation of

many optimization algorithms. At present,

unobstructed face recognition has long been widely

used in the daily lives, such as access control systems,

face-scanning mobile payment, monitoring systems

and real-name authentication. It can be seen that

unobstructed face recognition plays a crucial role in

both practical application and scientific research.

Next, I will introduce four methods of unobstructed

face recognition.

Lin proposed a super-resolution identification

method based on deep confidence networks (Lin,

2013). This method first judges the posture in the

image and then recognizes the face. Because high-

resolution and low-resolution images have certain

similarities in the feature domain, it is difficult to deal

with posture problems with linear features. Therefore,

it is proposed to use advanced deep neural networks

to obtain nonlinear connections between the two. She

implements the discrimination model through a group

of DBNs composed of multiple RBM structures and

adds a logical regression layer at the top. Considering

the uncertainty of facial posture, she proposed posture

mapping based on DBNs. Experiments show that it is

easy to lose personal characteristics during the

mapping process, which is also a shortcoming of this

method.

Eleyan uses 18 algorithms to extract local

statistical descriptors for face images and studies the

impact of a single feature descriptor and the fusion of

two or three feature descriptors on the accuracy of

face recognition (Eleyan, 2023). Experiments show

that the fusion of some two or three feature

descriptors can improve performance, but other

fusion techniques still need to be explored to get

better results.

Yang and others studied small sample problems in

the field of face recognition and established a

framework theory to solve singularity problems so

that F-S identification and J-Y identification can be

widely used as methods to directly deal with high-

dimensional and small sample problems (Yang et al.,

2003). They combined the advantages of these two

methods to propose a new combination linear

identification method, which was experimented on

the entire ORL face library and compared the method

with the classic feature faces and Fisherfaces

methods. Experiments show that the combination

identification method is obviously superior to the

other three methods. In addition, they also conducted

many other experiments, and without exception, the

results showed that the combination identification

method was the best method.

Huang optimized the method based on the Local

Binary Patterns (LBP) orator and creatively proposed

the LBP subpattern algorithm and the LBP pyramid

algorithm (Huang, 2009). Compared with the original

LBP mode, the LBP submodel is more flexible. In the

face of different types of images, this method can

obtain LBP sub-modes that can better represent the

characteristics of each type of image, and can

independently select the percentage of the total mode.

At the same time, through Principal Component

Analysis (PCA) transformation, the method can also

achieve dimension reduction and eliminate

interference factors in the image, which is conducive

to more accurate face recognition detection. The LBP

pyramid algorithm uses multi-scale space theory to

build a new type of multi-scale LBP based on

changing the resolution or scale of the image. This

method greatly reduces the computational

complexity, enhances the feasibility of this method,

and has higher accuracy in feature extraction than

other multi-scale LBPs. Then, on this basis, a multi-

scale LBP submodel is proposed, which has the

advantages of the above two methods and can cope

with changes in image types.

Unobstructed face recognition technology is

easier to implement than unobstructed face

recognition technology because it can extract more

facial features. There is no need to deal with the loss

of information caused by the loss of features caused

by face occlusion, and the algorithm is simpler. At

present, the development of unobstructed face

recognition is relatively mature, which can quickly

recognize faces, with high accuracy and good

performance. However, the method cannot continue

to maintain the accuracy of the model when it

encounters obstruction, which has certain limitations.

And because in real life, people cannot guarantee that

there is no cover every time face recognition is carried

out, which also leads to a decline in the detection

ICDSE 2025 - The International Conference on Data Science and Engineering

accuracy of unobstructed face recognition. Therefore,

the research on masked face recognition is

particularly urgent and important. In recent years,

masked face recognition has also been a hot research

direction in the field of face recognition.

3 FACE RECOGNITION IN A

BLOCKED ENVIRONMENT

Although face recognition technology has matured, it

still faces many challenges in real life. This is

because, in real life, it is impossible to ensure that the

face is not covered every time it is tested, such as

masks, hair, hats, etc., or light occlusion caused by

uneven external light. Especially with the outbreak of

COVID-19 in 2020, people's health awareness has

gradually increased. In many public places, people

will choose to wear masks, which greatly increases

the difficulty of face detection. Therefore, the study

of masked face recognition has become an inevitable

choice. In this chapter, I will introduce five methods

of masking face recognition.

Xu and others proposed a face image recognition

method based on the circular generation of

confrontation networks (Xu et al., 2022). The model

reconstructs the entire image and outputs the image

that restores the original image features, completes

the face repair, and trains through two pairs of

distinguishers and generators to ensure the accuracy

of the repair. After the repair is completed, the

residual network ResNet-50 is used to extract facial

features and the loss function RegularFace is

introduced to deal with the impact of different classes

of interclass distance on classification. Although the

repair effect will be affected by factors such as linear

nonlinear occlusion and the occlusion area of the

occlusion part, resulting in large differences in repair

effects, in general, when using the repaired pictures

for face detection, the accuracy of the detection will

be significantly improved.

Zhou et al. proposed a block-based obscured face

recognition algorithm in combination with

convolutional neural networks (Zhou et al., 2018).

The algorithm obtains the feature points of a face

through the self-coding network (CFAN) and divides

them into four areas: left and right eyes, mouth and

nose. After the blocking is completed, an occlusion

discrimination network is trained based on the

InceptionV3 network to perform occlusion

discrimination for each area, and feature fusion and

similarity detection are carried out according to the

discrimination results to obtain the final facial

features. The literature compares this method with the

classical Sparse Representation-based Classification

(SRC), Group Sparse Representation-based

Classification (GSRC) and Robust Sparse Coding

(RSC) algorithms in terms of the covering part and

the covering area (Wright et al., 2008; Yang & Zhang,

2010). When covering a large area, the algorithm has

maintained an extremely high accuracy. However, in

the case of sunglasses occlusion, the algorithm is not

as accurate as the RSC algorithm, which may be

because too much feature block occlusion leads to the

loss of eye features, so the algorithm still needs to be

improved.

Zhou and others optimized the scale invariant

feature transformation (SIFT) algorithm of human

faces (Zhou & Lai, 2011). The traditional SIFT

algorithm can find out the key points of most matches

in the image, but there are still some mismatches.

Therefore, they proposed a new matching idea, taking

a key point in an image and finding out the first two

key points in another image that are closest to that

point. In these two key points, if the nearest distance

divided by the sub-close distance is less than a certain

proportional threshold, the two key points are

considered to be matched. Experiments were

conducted on the AR face library and the Manchester

face library, and the results showed that the

recognition rate of the optimization method was about

10% higher than that of the traditional SIFT

algorithm.

Li and others proposed a masked face recognition

method based on the detection and elimination of face

heterogenous areas of the average face (Li et al.,

2015). This method obtains the error face image by

performing the difference between the test face and

the average face formed by the training picture and

segmenting the error image to obtain the information

description of the occlusion area. This is a very

critical step in the whole algorithm and determines

the division of the obscured area of the test set and the

training set later. The training set and the test set form

a new data set after removing the corresponding

blocking parts. The calculation difficulty of this

algorithm is relatively small, which reduces the

difficulty of implementation. However, experiments

have proved that the segmentation effect of the error

face image of the algorithm is poor in the presence of

light changes, and the impact of light intensity on the

algorithm needs to be further explored.

Li and others proposed an algorithm that

combines machine learning based on thinking

evolution with local features (Li G. & Li W., 2014).

Considering that face recognition in practical

applications will be affected by uncertainties such as

Technical Progress of Face Recognition Technology in Occluded and Unobstructed Environment

the environment, they proposed the LBP offset

feature group to avoid the problem of feature offset

instability. And the evolution of the convergence

process and alienation process is carried out in all

local areas. In order to avoid the replacement of the

score of the blocking area by the winner, they

proposed a new convergence rule, setting a threshold,

and using the threshold to divide the new competitive

group and the winner group. Based on the AR face

library, the experiment was compared with the

Robust Distance Weighted Local Binary Pattern

(RDW-LBP), Fuzzy Principal Component Analysis

(FPCA) and Orthogonal Matching Pursuit with

Cholesky Decomposition (OMP-Cholesky)

algorithms. The experimental results show that the

algorithm has maintained the highest recognition

accuracy.

With the continuous progress of science and

technology, the development of masked face

recognition is also getting faster and faster.

Obstructed face recognition makes up for the

shortcomings of unobstructed face recognition. Even

if the identification object wears a mask or cannot

capture all its facial features due to light, shooting

angle and other reasons, the covered face recognition

technology can still maintain a good recognition rate.

This technology provides strong technical support for

public safety, which can help quickly identify

dangerous personnel and criminals and improve

public safety. In terms of personal experience, for

example, when face unlocking is required, users may

wear hats, masks and other covering objects.

Applications with masking face recognition can also

maintain a good experience for users. However, as

mentioned above, it is still impossible to achieve

high-efficiency recognition in some application

scenarios with masked face recognition, such as

excessive area of the covered part, so the research on

masked face recognition needs to be further

strengthened.

4 THE ROLE AND IMPACT OF

DATA SET

The use of professional data sets is very important for

the research and application of face recognition. The

theoretical analysis in the research process needs to

be constantly modified and improved according to the

experimental results. This article will introduce

several commonly used unobstructed and covered

face recognition data sets. First of all, let's introduce

the unobstructed face recognition data set: Celebrity

Faces Attribute Dataset (CelebA) data set is a large-

scale face recognition data set established by the

Chinese University of Hong Kong. It contains more

than 200,000 face pictures, which is suitable for

people's Research judgment of face properties; the

Face Detection Data Set and Benchmark (FDDB)

data set is often used for test result evaluation; the

VGGFace2 data set is suitable for training deep

neural networks. Commonly used masked face

recognition data sets include: The maskedFace-Net

data set was born after the COVID-19 pandemic. All

pictures show faces wearing different types of masks,

which is very suitable for research on masked face

recognition; Multi-Attribute Fa Ce Annotations

(MAMA) also provides a large number of face data

sets with occlusion, including occlusion in all

directions and angles; the Wider Face test set is huge

and the results are highly reliable.

5 CONCLUSION

As an important part of the field of computer vision,

face recognition technology has made remarkable

progress in recent years and will continue to lead the

wave of scientific and technological development.

This technology not only plays a key role in security

monitoring but also gradually penetrates into many

daily applications such as smartphone unlocking and

financial payment verification. In the future, with the

continuous progress of artificial intelligence and deep

learning, face recognition is expected to maintain its

position at the forefront of scientific research and

promote innovation and change in more fields. This

article reviews the research results of face recognition

under unobstructed and covered conditions. Although

some effective y optimization algorithms have been

proposed, the existing methods still have limitations,

especially in complex environments, the performance

needs to be improved. Therefore, the future

improvement direction should focus on optimizing

existing algorithms to meet the needs of more diverse

real scenarios. In addition, this article briefly

introduces several data sets commonly used in face

recognition, such as CelebA. These data sets provide

valuable resources for researchers to help train and

evaluate model performance. Understanding and

rational use of these data sets is crucial to promoting

the development of face recognition technology. In a

word, the development of computer face recognition

technology is long and challenging. Although many

achievements have been made, in order to make this

technology more perfect, scientific researchers need

to work together to carry out in-depth exploration into

ICDSE 2025 - The International Conference on Data Science and Engineering

algorithm optimization, privacy protection and

ethical considerations. Only in this way can this

advanced technology better serve the society and

meet people's growing needs for safety and

convenience.

REFERENCES

M Eleyan, A. (2023). Face recognition using ensemble

statistical local descriptors. Journal of King Saud

University-Computer and Information Sciences, 35(9),

101729.

Huang Feifei. (2009). Research on face recognition based

on LBP (master's thesis, Chongqing University).

Li Dongmei, Xiong Chengyi, Gao Zhirong, Zhou Cheng &

Wang Hanxin. (2015). Occlusion face recognition

based on the elimination of heterovalued areas.

Computer Science (03), 289-295.

Li Gen & Li Wenhui. (2014). Blocking face recognition of

machine learning based on the evolution of thinking.

Journal of Jilin University (Engineering Edition) (05),

1410-1416.

Lin Miaozhen. (2013). Research on face recognition based

on deep learning (master's thesis, Dalian University of

Technology).

Wright, J., Yang, A. Y., Ganesh, A., Sastry, S. S., & Ma, Y.

(2008). Robust face recognition via sparse

representation. IEEE Transactions on Pattern Analysis

and Machine Intelligence, 31(2), 210-227.

Xu Runhao, Cheng Jixiang, Li Zhidan & Fu Xiaolong.

(2022). Obstructive face recognition based on the

circular generation of the opposition network.

Computer Engineering (05), 289-296+305.

Yang Jian, Yang Jingyu, Ye Hui. (2003). Theoretical

research and application of Fisher linear discrimination

analysis. Journal of Automation (04), 481-493.

Yang, M., & Zhang, L. (2010). Gabor feature based sparse

representation for face recognition with gabor

occlusion dictionary. In Computer Vision-ECCV 2010:

11th European Conference on Computer Vision (pp.

448-461). Springer Berlin Heidelberg.

Zhou Lingli & Lai Jianhuang. (2011). Obstructed face

recognition based on the optimization algorithm of

scale-invariant feature transformation. Computer

Application (S1), 64-66+75.

Zhou Xiaojia, Zhu Yunbin & Zhang Yue. (2018). Block-

based masked face recognition algorithm. Computer

Applications and Software (02), 183-187+193.

Technical Progress of Face Recognition Technology in Occluded and Unobstructed Environment