Spatio-chromatic image content descriptors and their analysis using extreme value theory

Vasileios Zografos, Reiner Lenz
Proceedings of the 17th Scandinavian conference on Image analysis (SCIA), page 579--591 - 2011
Download the publication : SCIA2011.pdf [3.4Mo]  
We use the theory of group representations to construct very fast image descriptors that split the vector space of local RGB distributions into small group-invariant subspaces. These descriptors are group theoretical generalizations of the Fourier Transform and can be computed with algorithms similar to the FFT. Because of their computational efficiency they are especially suitable for retrieval, recognition and classification in very large image datasets. We also show that the statistical properties of these descriptors are governed by the principles of the Extreme Value Theory (EVT). This enables us to work directly with parametric probability distribution models, which offer a much lower dimensionality and higher resolution and flexibility than histogram representations. We explore the connection to EVT and analyse the characteristics of these descriptors from a probabilistic viewpoint with the help of large image databases.

Images and movies

 

BibTex references

@InProceedings\{ZL11,
  author       = "Zografos, Vasileios and Lenz, Reiner",
  title        = "Spatio-chromatic image content descriptors and their analysis using extreme value theory",
  booktitle    = "Proceedings of the 17th Scandinavian conference on Image analysis (SCIA)",
  series       = "Lecture Notes in Computer Science, ISSN 0302-9743; 6688",
  pages        = "579--591",
  year         = "2011",
  editor       = "Anders Heyden and Fredrik Kahl",
  publisher    = "Springer-Verlag"
}

Author publication list