Título de la tesis

Invariance and Reliability in Statistical Shape Models


Federico M. Sukno


Alejandro F. Frangi


The present thesis concentrates on the applicability of statistical modeling in fa-
cial analysis. Starting from the paradigm of shape and appearance models de-
veloped in the last decade, new algorithms are proposed to allow improving their
reliability and invariance to different types or rotations. The extensions are for-
mulated in a generic way, such that the models keep the wide applicability of the
original approach.

The proposed techniques were experimentally validated in facial analysis tasks.
This field became especially relevant in the last few years with an important growth
of its international market. A remarkable fact in this sense is the recent adoption
of facial biometrics as the standard technology for new biometric passports, taking
over other important biometric modalities such as iris or fingerprints. Although the
latter are able to achieve lower error rates, the face appearance is the natural way
for identification among humans and it is perceived less intrusive. Additionally, it
is among the very few biometric modalities that can work, in theory, without the
explicit collaboration of the person to be identified.

Chapter 1 provides a brief overview on biometrics and presents the components
of a generic biometric system for facial recognition, with especial focus on shape
and appearance models. Specifically, Active Shape Models (ASMs) constitute the
key methodological component of this thesis, and are briefly covered in Chapter 2.
ASMs allow for the automatic segmentation and analysis of images based on
generative models. Introduced in 1992 (by T. Cootes et al.), a considerable number
of works has been published on the application of ASMs to diverse types of im-
ages, among which medical and facial images are the most numerous. On the other
hand, ASMs proofed themselves too simple for modeling purposes in some appli-
cations. As a result, later publications focused on extensions and improvements
to the original formulation. One of the most important was the introduction of
the Active Appearance Models (AAMs) in 1998. The AAMs soon became popular
enough to be considered a separated methodology in its own right, independently
from ASMs.

This thesis introduces three novel extensions to ASM. They aim at improving
the behavior of these models with a special focus on invariance and reliability. The
hypothesis is that the extension and improvement of the segmentation algorithms
will lead to a more accurate delineation of facial features, allowing for a more
appropriate extraction of image information.

Our first extension addresses the problem of accurate segmentation of promi-
nent features of the face in frontal shots, and is covered in Chapter 3. We propose a
method that generalizes linear ASMs using a non-linear intensity model and incor-
porating a reduced set of differential invariant features as local image descriptors.
These features are invariant to rigid transformations, and a subset of them is chosen
by Sequential Feature Selection. The new approach overcomes the unimodality and
Gaussianity assumptions of classical ASMs regarding the distribution of the inten-
sity values across the training set. Our methodology has demonstrated a significant
improvement in segmentation accuracy when compared to the linear ASM, which
also derived in lower error rates on identity verification tasks.

The second extension (Chapter 4) concentrates on the invariance of the matching
algorithm in the presence of out-of-plane rotations when working with quasi-planar
objects. By constraining the analysis to certain parts of the face, the outlines can be
approximately considered coplanar. Then, based on projective geometry concepts,
ASMs are modified so that they can work independently from the viewpoint (within
the range limitations of feature visibility). As a consequence, an ASM constructed
with frontal view images can be directly applied to the segmentation of pictures
taken from other viewpoints. Validation of the method is presented in images sys-
tematically divided into three different rotations (to both sides), as well as upper
and lower views due to nodding. The presented tests are among the largest quan-
titative results reported to date in face segmentation under varying poses.

The third extension (Chapter 5) provides an automatic reliability measure of
the segmentation for each analyzed image. That is, the model is able to estimate
whether the segmentation obtained for certain image is trustworthy or not. This is
very important when ASMs are used into fully-automatic systems, since accurate
segmentation is crucial for the subsequent interpretation of the image. The auto-
matic estimation of reliability can be promising for a number of applications. We
demonstrate two of them: automatic model selection and reliable identity verifica-
tion. Results were highly satisfactory in both cases. The strength of the proposed
approach relies on its low false positive rate, which means that incorrect segmenta-
tions are very unlikely to be misclassified as reliable.

In this way, the first two extensions share the concept of invariance (to rotations
in and out of the image plane). On the other hand, it will be shown that the first
extension also increases the accuracy of the segmentation, while the third extension
is devoted to the estimation of how reliable is the segmentation of each image.
In all cases, intensive experiments have been performed to validate the proposed
algorithms, with encouraging results.


Defendida - 2008


Sobresaliente Cum Laude

Validación de código - XHTML 1.1
Validación de CSS
Validación de Accesibilidad Web