Título de la tesis

Techniques for statistical shape model building and fusion


Kostantyn Butakov


Alejandro F. Frangi


In the present thesis we address the problems of building and combining act
shape and appearance models. The models are one of the widespread tool
object modeling and segmentation with a shape and texture prior. When the
models are employed several problems can arise:

1. expensive training process (in terms of time and memory requirements)
2. a training set of images with the delineated object (usually manually) is required
3. a high degree of uncertainty of the delineations (due to the presence of noi
including unfeasibility of manual delineations in 3D.

To overcome these problems we propose:

1. A framework for weighted fusion of multiple active shape or active appear-
ance models based on eigenspace combination. Such combination strategy can
be treated as a linear interpolation of the models. The benefit of the fusion
is that the combined model can represent any object, which can be assumed
to be a linear combination of the objects corresponding to the fused models.
In other words, if an object has a number of typical appearances (different
face expressions or different face poses, or different cardiac pathologies), it is
possible to choose the most representative ones and assume any other to be
a linear combination of the representative set. Then the combined model can
be used to accurately segment the object in question and weights can be used
for classification to determine, which representative appearance is closer. The
possible applications of this framework are: batch model construction, object
classification based on combination weights, reduction of training sets to only
representative appearances.

2. A view-independent face segmentation algorithm based on the fusion of ac-
tive appearance models. This algorithm can be used to segment any facial
pose and also determine the pose angle using the estimated combination
weight. Only the views corresponding to the extreme head poses and the
frontal one are taken for training, all the other poses are assumed to be a lin-
ear combination of these. Estimation of combination coefficients through re-
construction error minimization allows finding the optimal combined model,
which is more specific to the pose under consideration than a single model
constructed for all poses.

3. Combination of computed tomography (CT) and synthetic ultrasound (US)
and single photon emission tomography (SPECT) images to automatically
learn shape variation and voxel intensity variation, where it is demonstrated
how intensity information can be learned for the two modalities where the
resolution or quality is too low to manually annotate the images, especially in
3D. In this case generation of synthetic images through realistic simulation of
the imaging process allows learning the appearance for a given set of shapes
(obtained from high quality CT scans).


Defendida - 2009


Sobresaliente Cum Laude

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