Seminar Details
| Date |
15-12-2005 |
| Time |
15:30 |
| Room/Location |
Sala conferenze 322, 3 piano |
| Title |
Approximate Nearest-Neighbor Searches: Application to Content-Based Image Retrieval |
| Speaker |
Sid-Ahmed Berrani, Ph.D. |
| Affiliation |
France Telecom R&D - TECH/IRIS , 4, rue du Clos Courtel - BP 91226. 35512 Cesson Sévigné Cedex. Fran |
| Link |
http://www.irisa.fr/texmex/Sid-Ahmed.Berrani/
|
| Abstract |
Content-based image retrieval systems rely on image processing
methods to automatically describe the images and on multidimensional
indexing techniques to reduce the cost of similarity searches.
Unfortunately, the existing multidimensional indexing techniques are
not well-adapted to the properties of the image descriptors
(dimension, distribution...). Their performance degrades drastically
when the dimension of image descriptors increases (dimensionality
curse problem).
In this talk, this problem will be analyzed and a new efficient
method of nearest-neighbor (NN) searches will be presented. The idea
of this method is to reduce the retrieval cost by searching for the
approximate neighbors of the query instead of searching for the exact
result, realizing hence a trade-off between result quality and query
execution time. This method allows, however, a fine and intuitive
probabilistic control of the precision of the search.
In the second part of the talk, an application of the proposed NN
search method within a real CBIR system for image copy identification
will then be presented.
Keywords: Multimedia Indexing, Content-Based Image Retrieval,
Similarity Searches, (Approximate) Nearest-Neighbor Searches,
Dimensionality Curse |
|
|
 |
