Universal compression of images: send the model?
Event details
| Date | 12.12.2018 |
| Hour | 10:00 › 12:00 |
| Speaker |
Aline ROUMY, INRIA-Rennes (France) Aline Roumy received the Engineering degree from Ecole Nationale Superieure de l'Electronique et de ses Applications (ENSEA), France in 1996, the Master degree in 1997 and the Ph.D. degree in 2000 from the University of Cergy-Pontoise, France. During 2000-2001, she was a research associate at Princeton University, Princeton, NJ. On November 2001, she joined INRIA, Rennes, France as a research scientist. She has held visiting positions at Eurecom and UC Berkeley. She serves as an Associate Editor for the Annals of telecommunications and for the IEEE Transactions on Image Processing. Her current research interests include the area of statistical signal and image processing, coding theory and information theory. |
| Location | |
| Category | Conferences - Seminars |
ABSTRACT:
Universal compression is the problem of compressing a data source without knowledge of the source probability distribution.
This occurs in particular when images and videos have to be compressed.
At the heart of the matter are the questions of modeling the data and coding the model.
In the first part of the presentation, the impact of the model will be discussed and I will present new, learned models of images and evaluate their impact on image compression performance. In particular, a transform and a predictor that both aim at capturing spatial correlation, will be learned (via a deep autoencoder and a set of deep neural networks).
In the second part of the presentation, I will introduce a novel source coding problem allowing massive random access to a large database of correlated sources. I will show that it is possible to extract arbitrary sources from an appropriately compressed database purely by bit extraction, and at the same rate as if the database was decoded and the requested sources were re-encoded.
Then the question of source modeling in view of compression will be discussed.
Universal compression is the problem of compressing a data source without knowledge of the source probability distribution.
This occurs in particular when images and videos have to be compressed.
At the heart of the matter are the questions of modeling the data and coding the model.
In the first part of the presentation, the impact of the model will be discussed and I will present new, learned models of images and evaluate their impact on image compression performance. In particular, a transform and a predictor that both aim at capturing spatial correlation, will be learned (via a deep autoencoder and a set of deep neural networks).
In the second part of the presentation, I will introduce a novel source coding problem allowing massive random access to a large database of correlated sources. I will show that it is possible to extract arbitrary sources from an appropriately compressed database purely by bit extraction, and at the same rate as if the database was decoded and the requested sources were re-encoded.
Then the question of source modeling in view of compression will be discussed.
Practical information
- Informed public
- Free
- This event is internal