Learning the dynamics of deformable objects and recursive boundary estimation using curve evolution techniques
Massachusetts Institute of Technology. Dept. of Electrical Engineering and Computer Science.
Alan S. Willsky.
MetadataShow full item record
The primary objective of this thesis is to develop robust algorithms for the incorporation of statistical information in the problem of estimating object boundaries in image data. We propose two primary algorithms, one which jointly estimates the underlying field and boundary in a static image and another which performs image segmentation across a temporal sequence. Some motivating applications come from the earth sciences and medical imaging. In particular, we examine the problems of oceanic front and sea surface temperature estimation in oceanography, soil boundary and moisture estimation in hydrology, and left ventricle boundary estimation across a cardiac cycle in medical imaging. To accomplish joint estimation in a static image, we introduce a variational technique that incorporates the spatial statistics of the underlying field to segment the boundary and estimate the field on either side of the boundary. For image segmentation across a sequence of frames, we propose a method for learning the dynamics of a deformable boundary that uses these learned dynamics to recursively estimate the boundary in each frame over time. In the recursive estimation algorithm, we extend the traditional particle filtering approach by applying sample-based methods to a complex shape space.(cont.) We find a low-dimensional representation for this shape-shape to make the learning of the dynamics tractable and then incorporate curve evolution into the state estimates to recursively estimate the boundaries. Experimental results are obtained on cardiac magnetic resonance images, sea surface temperature data, and soil moisture maps. Although we focus on these application areas, the underlying mathematical principles posed in the thesis are general enough that they can be applied to other applications as well. We analyze the algorithms on data of differing quality, with both high and low SNR data and also full and sparse observations.
Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science, 2005.This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections.Includes bibliographical references (p. 161-176).
DepartmentMassachusetts Institute of Technology. Dept. of Electrical Engineering and Computer Science.
Massachusetts Institute of Technology
Electrical Engineering and Computer Science.