@PhDThesis{Souza:2017:AnCoPa,
author = "Souza, Vanessa Cristina Oliveira de",
title = "An{\'a}lise computacional de padr{\~o}es estruturais
n{\~a}o-lineares a partir de imagens digitais com estudos de caso
em ci{\^e}ncias ambientais e espaciais",
school = "Instituto Nacional de Pesquisas Espaciais (INPE)",
year = "2017",
address = "S{\~a}o Jos{\'e} dos Campos",
month = "2017-06-09",
keywords = "an{\'a}lise de flutua{\c{c}}{\~a}o destendenciada
bi-dimensional, textura, classifica{\c{c}}{\~a}o,
paraleliza{\c{c}}{\~a}o, GPGPU/CUDA, Bi-dimensional detrended
fluctuation analysis, texture, classification, parallelization,
GPGPU/CUDA.",
abstract = "A An{\'a}lise de Flutua{\c{c}}{\~a}o Destendenciada (DFA) tem
sido amplamente utilizada na verifica{\c{c}}{\~a}o de
propriedades de escala de s{\'e}ries temporais unidimensionais.
Al{\'e}m de revelar a presen{\c{c}}a ou n{\~a}o de mem{\'o}ria
na s{\'e}rie (persist{\^e}ncia), o m{\'e}todo DFA possibilita
compreender melhor o processo que originou o sinal analisado e a
for{\c{c}}a da correla{\c{c}}{\~a}o. Nesse ponto, o operador do
DFA (\$\alpha\$) infere tamb{\'e}m sobre a rugosidade do
sinal, isto porque quanto maior a persist{\^e}ncia, menor a
rugosidade. Quando tal caracter{\'{\i}}stica {\'e} expandida
para sinais bi-dimensionais, em especial imagens digitais, a
no{\c{c}}{\~a}o de persist{\^e}ncia infere tamb{\'e}m sobre a
textura desses sinais. O DFA foi generalizado para operar em
sinais bi-dimensionais em 2006 (DFA-2D) e, desde ent{\~a}o,
diversos estudos v{\^e}m sendo feitos, especialmente utilizando o
\$\alpha\$ como um operador textural. Diferente dos operadores
texturais comuns que atuam sobre a varia{\c{c}}{\~a}o de brilho
na imagem, o DFA-2D utiliza o arcabou{\c{c}}o te{\'o}rico da
teoria dos fractais e infere a textura a partir da
caracter{\'{\i}}stica de autossimilaridade do sinal. Neste
contexto, os objetivos desse trabalho foram ; i) explorar
quest{\~o}es controversas ou n{\~a}o tratadas ainda para o
DFA-2D na literatura e ii) avan{\c{c}}ar no estado da arte da
t{\'e}cnica bi-dimensional em suas vers{\~o}es mono e
multifractal, avaliando a resposta em um conjunto diverso de dados
e tamb{\'e}m melhorando sua performance por meio da
paraleliza{\c{c}}{\~a}o, utilizando GPGPU/CUDA. Sendo assim,
essa tese pretendeu contribuir com dois aspectos do DFA-2D
criticados na literatura : a dificuldade de
interpreta{\c{c}}{\~a}o do operador \$\alpha\$ e o alto custo
computacional. Foram estudados conjuntos de dados simulados cuja
resposta {\'e} conhecida (fBm e fGn), conjuntos cl{\'a}ssicos de
dados simulados n{\~a}o tratados na literatura com o DFA-2D (como
ru{\'{\i}}dos do tipo 1/f e redes de mapas acoplados). A
an{\'a}lise tamb{\'e}m incluiu conjuntos de dados reais, cujas
aplica{\c{c}}{\~o}es utilizando o DFA-2D s{\~a}o in{\'e}ditas,
como a classifica{\c{c}}{\~a}o morfol{\'o}gica de gal{\'a}xias
e a infer{\^e}ncia de rugosidade efetiva para estudos de energia
e{\'o}lica, por meio da classifica{\c{c}}{\~a}o de Modelos
Digitais de Eleva{\c{c}}{\~a}o topogr{\'a}fica (MDE). Para os
sinais simulados, o DFA-2D (mono e multifractal) caracterizou de
forma satisfat{\'o}ria os sinais. Para os sinais reais, os
resultados mostraram que, mesmo quando o operador \$\alpha\$
n{\~a}o atinge uma boa taxa de classifica{\c{c}}{\~a}o, seu uso
{\'e} muito informativo, caracterizando o sinal, e n{\~a}o
apenas classificando-o. Al{\'e}m disso, o DFA-2D aplicado a dados
reais apresentou dificuldades e desafios impercept{\'{\i}}veis
nos sinais simulados. Por fim, a paraleliza{\c{c}}{\~a}o
mostrou-se eficaz, diminuido consideravelmente o tempo de
processamento pelo DFA-2D. ABSTRACT: The Detrended Fluctuation
Analysis (DFA) has been widely used to verify the scaling
properties of unidimensional time series. Besides revealing the
presence or absence of memory in the series (persistence), the DFA
method allows understanding the process that originated the
analyzed signal, as well as the strength of the correlation. The
DFA operator () also infers about the signal roughness, because
the larger the persistence, the smaller the roughness. When such
feature is expanded to bi-dimensional signals, especially in
digital images, the notion of persistence also infers on the
texture of these signals. The DFA method was generalized to
operate on bi-dimensional signals in 2006 (DFA-2D) and, since
then, several studies have been performed using the () as a
textural operator. Differently from the common textural operators,
which focus on the brightness variation of an image, the DFA- 2D
exploits the theoretic framework from the fractal theory and
infers the texture using the auto similarity feature of the
signal. In this context, the goals of this work were: i) to
explore controversial or not yet treated issues for the DFA-2D in
the literature and ii) advance the state of the art in the
bi-dimensional technique in its two versions mono e multifractal
evaluating the response in a diverse set of data and also
improving its performance by means of parallelization, using
GPGPU/CUDA. Therefore, this thesis provides contributions in two
aspects of the DFA-2D which are criticized in the literature: the
interpretation difficulty related to the \$\alpha\$ operator
and high computational cost of the method. Simulated sets of data
that present a well-known response (fBm and fGn) has been studied,
as well as classical sets of simulated data not treated in the
literature using the DFA-2D (such as noise data of type 1/f and
coupled map networks). The analysis also included sets of real
data, providing umprecedent applications using the DFA-2D, such as
the morphological classification of galaxies and the inference of
effective roughness for eolic energy studies, by using the
classification of Models of Digital Elevation (MDE). The DFA-2D
(mono and multifractal) was able to categorize simulated signals
in a satisfactory manner. For real signals, the results show that,
even when the \$\alpha\$ operator does not reach a good
classification rate, its usage is very informative, characterizing
the signal and not only classifying it. Besides, the DFA-2D
applied to real data presented difficulties and challenges
imperceptible in simulated signals. Lastly, the parallelization
has proven to be effective in order to lower the processing time
required by the DFA-2D.",
committee = "Stephany, Stephan (presidente) and Rosa, Reinaldo Roberto
(orientador) and Assireu, Arcilan Trevenzoli (orientador) and
Guimar{\~a}es, Lamartine Nogueira Frutuoso and Zamith, Marcelo
Panaro de Moraes and Bolzan, Maur{\'{\i}}cio Jos{\'e} Alves",
englishtitle = "Computational analysis of non-linear structural patterns from
digital images with case studies in environmental and space
sciences",
language = "pt",
pages = "189",
ibi = "8JMKD3MGP3W34P/3P7RU3B",
url = "http://urlib.net/ibi/8JMKD3MGP3W34P/3P7RU3B",
targetfile = "publicacao.pdf",
urlaccessdate = "02 maio 2024"
}