Spatial Temporal Analysis of the Sea Breeze Front
This thesis presents two systems. The purpose of the first system is
to locate and segment the sea breeze front in a sequence of satellite images. The purpose of the second system is to register and retexture video sequences. The theoretical foundation of both systems is laid in the field of computer vision and image processing.
Being able to automatically locate and segment the sea breeze front in a sequence of satellite images, would prove highly useful for researches working on understanding and predicting the sea breeze front.
The sea breeze is a cool breeze blowing inland from the sea.
This breeze often manifests itself in satellite imagery as a thin line
of clouds. The boundary of the sea breeze front clouds will be located by exploiting their distinctive texture. The movement and non-rigid
deformations of the sea breeze front is modeled in the spatiotemporal
space by an open active surface. The surface is regularized by certain
spatial and temporal smoothness priors. We put forward a method based
on cross-validation that automatically estimates sensible values for
the regularization parameters. We report that our system is able to
reliably segment the sea breeze front in each image. Our system is
compared with the only know current system for segmenting the sea
breeze front in satellite imagery. We show that our system surpasses
the old system in accuracy.
Retexturing videos of deformable surfaces is an important problem in computer vision as it has a wide variety of applications. A key step in producing visually pleasing retexturing results is registration. Traditional registration methods require a certain amount of texture on the surface in order to capture all the deformation details. However, in cases such as cartoon videos, there is a high number of smooth contours and only little or spurious texture. We propose a novel method for registering and retexturing cartoon-like videos by means of joint contour detection and point to point curve matching.
The main idea is to fit a parametric 3D active surface model in the spatiotemporal space, utilizing a regularization term which limits the change in curvature over time. We show that with cross-validation it is possible to automatically estimate a suitable value for the regularization parameter, controlling the tradeoff between the regularization and the data term. We report convincing registration and retexturing results on cartoon videos.
Censor er: Søren Overgaard, DMI
Vejleder er: Søren Olsen