Аннотация:
The purpose of this thesis is to examine the impact of obfuscation on the performance
of facial recognition algorithms. The primary form of obfuscation is facial masking,
due to the immediate relevance of widespread adoption during the pandemic era,
and the potential impact of this practice on the reliability of facial recognition. The
work begins by replicating recent work [1] and comparing the results. The study then
examines a new dataset "SpeakingFaces", and generates the obfuscation set by using
the open-source "MaskTheFace" script. The set of facial recognition techniques are
applied to the obfuscated dataset, and the results evaluated.
The work consists of seven chapters.
The first chapter describes the motivation for using masked face recognition technology
to authenticate users to ensure security and confidentiality. The objectives of
this thesis are presented in detail, and the stages and procedures for the development
of this system are described. A brief historical overview of the emergence of this
subject area is also presented.
The second chapter describes the results of previous studies based on a review
of the relevant literature. A survey and comparative analysis of recently published
works on creating synthetic datasets and systems for recognizing masked faces, which
were used as the basis of this thesis, is being carried out.
The third chapter presents the architecture of the masked face recognition system,
which is a replication of previously conducted other research works. This work contains the basic architecture of the face recognition system, a model based on a
trained neural network and an assessment of the accuracy of face recognition both
with and without a mask. These segments have a detailed description with specific
templates and parameter settings.
The fourth chapter describes in detail the creation of a dataset using open-source
MTCNN scripts based on mxnet and TheMaskFace, and using these scripts, the processes
of creating datasets designed for the masked face recognition system. This
section describes the collection methods, data attributes, constraints, and preprocessing
required to use the data. We evaluate the system’s accuracy based on the
achievements of other researchers.
The fifth chapter describes the detailed implementation of the previously selected
research work. All the results of the previous chapters are considered, and the known
shortcomings of previous works are considered.
The sixth chapter presents all the achievements achieved within the masked face
recognition system framework. The analysis of the operation of the facial recognition
system in masks with an artificially created dataset and comparison with other
datasets is carried out. A comprehensive accuracy assessment and performance analysis
is reported. The goals and objectives of the thesis are analyzed with comments
on potential shortcomings of the work and completed.
The seventh chapter presents the final part of the research work. The results
obtained during the research work, the time and performance constraints, and the
work planned in the future are described.