My diploma thesis [1] presents a novel approach to discover objects in unlabeled image data using a combination of traditional methods including image segmentation, feature extraction, clustering, and dynamic programming. The key idea consists of using image segmentation to group features in an image, and use these feature groups to represent the individual segments in a way that is invariant to rotation, scale, and translation. Such feature segments can then be related to each other by an appropriate distance measure to identify segments that occur repeatedly in different contexts. Finally, neighborhood relations among segments can be learned in a similar fashion to discover stable feature segment constellations that indicate the presence of reoccuring structures, i.e., putative objects in the images. The thesis contains a number of interesting and novel approaches to common problems such as segmentation [2], measuring the similarity of high-dimensional features [3], or establishing rotational invariance.
References
1
,
Dynamisches Lernen von Nachbarschaften zwischen Merkmalsgruppen zum Zwecke der Objekterkennung,
In: diploma thesis, University of Dortmund, 2006,
[pdf|bibtex]
2
,
A Topology-Independent Similarity Measure for High-Dimensional Feature Spaces,
In: Artificial Neural Networks. 17th International Conference (ICANN 2007). Vol. 4669. LNCS Part 2. Porto, Portugal: Springer, pp. 331–340, 2007,
[pdf|doi|bibtex]
3
,
Image Segmentation Based on Height Maps,
In: Computer Analysis of Images and Patterns. Vol. 4673. Lecture Notes in Computer Science. Springer Berlin Heidelberg, pp. 612–619, 2007,
[pdf|doi|bibtex]