Tesis y Tesistas 2020 - Postgrado - Fac. de Informática - UNLP
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
especialización
COMPUTACIÓN GRÁFICA,
IMÁGENES Y VISIÓN
POR COMPUTADORA
Esp. Cesar Armando Estrebou
cesarest@lidi.info.unlp.edu.ar
Advisor
Dr. Laura Lanzarini
Thesis defense date
June 26, 2020
SEDICI
http://sedici.unlp.edu.ar/handle/10915/100550
Human skin identification
algorithms and their relation
with colour systems
Keywords: image segmentation, skin detection, colour systems, skin dataset
Motivation
The systems of gesture detection have become a boom
during the recent years due to their technological progress
and the great interface potential for all kinds of applications.
A gesture detection system has four stages or sequences
where the output of a stage is the input of the following. In a
first stage, the movement and/or configuration of the hands
is captured by means of a sensor such as a webcam, a timeof-flight
camera, position sensor gloves or a combination
of any of them. Then, the segmentation or cutting of the
corresponding area of the hands is carried out. In a third stage,
from the image of the cut hand, the representative features
of the hand are removed in order to get a representation of
its configuration and movement that can be processed by a
computer. Finally, in the last stage, the gesture meaning is
identified after a combined classification of what the hand
gesture and its movement represent.
Together with my research team, we have been working on
the last two stages of removing features and classifying
gestures for many years. Recently, we have started working
on the first two stages, exploring different options using
a conventional video camera for the segmentation stage
without markers to cut hands. In the state-of-the-art
literature, there are several papers with a wide range of
options but they often choose as a partial or complete
solution the segmentation based on skin colour pixels to
remove the right areas of the hands and the face.
There are plenty of authors who publish solutions that
transform a wide range of colour systems and use different
algorithms to determine if a pixel in an image is skin or
not. Personally, I have started investigating the colour
systems used in different papers in order to find an answer
to the difficult question: what is the most appropriate
colour pattern to segment skin in colours? After reading
and studying many publications with specific solutions and
papers that revise the state-of-the-art, I could not find a
convincing answer. Furthermore, many authors in their
papers do not justify the use of different colour systems in
their skin segmentation applications, but if they do it, they
justify it vaguely and without enough reasons.
The main goal of this work was to determine the most
convenient representation colour pattern for the segmentation
based on human skin pixels in images. In order to carry this
out, a study was done on the influence of the colour patterns
in the different pixel-based segmentation algorithms that are
generally used in research papers of this field.
As a secondary goal, the idea was to document and describe
in detail the colour representation patterns, the pixel-based
segmentation algorithms and the skin dataset employed to
perform tests and obtain results. This is vital information to
copy and complement the results obtained by the authors
of the publications for future experiments.
Thesis Final Work contrinutions
This work described and studied in depth the colour
representation systems, an analysis of the skin distribution
was done for each of them, the algorithms that are
frequently used for skin segmentation were described and
more than 30 research publications were included for their
revision and study. The emphasis was on a survey as broad
as possible in order to analyse different works that include
as many colour representations models and segmentation
algorithms as possible.
This work has three interesting points to highlight after the
survey carried out, and they can be useful for future works
or publications. The first point is the collection of colour
representation models that are frequently used together
128