|
Matthias Demant |
||
Chair of Pattern Recognition and Image Processing Institute for Computer Science Albert-Ludwigs-University Georges-Koehler-Allee 052, room 01-017 D-79110 Freiburg i.Br., Germany |
|||
Telefon: | +49-(0)761-203-8211 |
||
Fax: | +49-(0)761-203-8262 |
||
Email: | |||
Abstract
The soil nematode C.elegans has established as a powerful model organism in biological research. By changing the DNA, it is possible to visualize certain proteins in the animal with the "GFP" (Green Fluoreszent Protein) which are fluorescent markers. The read-out of the fluorescent signals in a manual imaging process is very time consuming. A possible alternative is the usage of a special high-throughput sorter COPAS. The sorter captures the 1-dimensional fluorescent profile of a worm.In this thesis, different methods have been analyzed which compile significant statistics of entire worm populations. Based on these statistics differences and similarities between different populations should be detectable. The "Dynamic Time Warping" (DTW) was applied as a registration technique as well as a similarity measure. The DTW was refined to ensure a smooth alignment.
Combined with the DTW, unsupervised learning methods were evaluated to describe the data. Hierarchical clustering methods were considered to allow a clearly stuctured comparison of the fluorescence profiles. To analyze and quantify the population self-organizing maps (SOM) were examined. The entries should provide a meaningful representation of the population and maintain the topology after the projection onto the SOM. The SOM serves as base for the comparison of populations. In order to allow a real time classification, Gabor and Haar wavelet features were considered. Slides
Abstract
Stereo vision has been an intensive subject of research over the past 25 years. Recent years have seen a high amount of progress in the calculation of accurate and dense depth maps. The increased concentration of interest in this area is based on its importance in measurement technology. In this technology branch the estimation of highly accurate depth maps take a major role.
A large number of algorithms for stereo correspondence have been developed.
Generally they are structured in a two stage process:
In a first step a rough depth estimation with pixel-level accuracy is intended. Using global
or semi-global approaches well known problems like occlusion and ambigu-
ous matches can be handled at this level. The results of current pixel-level
estimation algorithms lie within the range of 2 pixels. In a second pass, the
refinement step, highly accurate, dense, depth maps are pursuited.
Putting the emphasis on the subpixel refinement step, this work compares
two possible refinement methods. The focus lies on a fast adjustment of a
noisy rough depth estimation.
The first technique was chosen to ensure the scope of an depth estimation on sub-pixel level as well as the objective of a small computation time. Its concept is based on the correlation of patches. A Gabor filter is used as a windowing function to extract patches around each pixel in both images. Within the nature of a Gabor filter, as a combination of a gaussian and a complex exponential, lies the expansion of the signal in a space-frequency domain. This technique seems appropriate, since the correlation between corresponding patches can be achieved rapidly.
The second technique takes into account that the base of numerous recent refinement techniques consist of fitting a cost function to an adequate sam- pled cost volume near the estimated depth values. Both algorithms have been implemented and evaluated on real world data. In this evaluation process results on different image regions have been analyzed.
In a mathematical part the interaction between scene planes, texture and focus has been analyzed. The interest lies in the relation between the three dimensional shapes of fronto-parallel and slanted surfaces. Especially the applicability to our sub-pixel algorithm has been considered. Slides