Image semantic compression at extremely low bitrates

Thesis
Starting date
End date
Location
IRISA Rennes
Room
Petri turing
Speaker
Tom Bordin (COMPACT)
Main department
D6 - Signal, Image, Language
D6 - Signal, Image, Language
Theme
In this thesis, we present our work on semantic image compression. Our aim is focused on compression rates far beyond those achieved by standard approaches.
To achieve this, we introduce a new framework based on semantic representations of images, shifting the focus from pixel-level fidelity to the of meaningful content.
The objective becomes clear: retain what matters to human perception while reconstructing images that are realistic. We address mainly two research questions:
How can we construct a compact yet expressive semantic representation of an image? and How can we design a decoder capable of reconstructing realistic and
semantically faithful images from such a representation, even at extremely low bitrates? Throughout the thesis, we show that foundation models embed images into
rich semantic spaces that can be exploited for compression. We further adapt the decoder to these representations by harnessing the power of diffusion models,
proposing techniques to guide them effectively without retraining. We propose a state-of-the-art framework that achieves drastically low compression rates while
preserving semantic fidelity, confirmed through subjective evaluation.
 
Composition of the jury
1. (Ext) BAJIC Ivan, H, Professor, School of Engineering Science, Simon Fraser University,
Canada (Rapporteur)
2. (Ext) TARTAGLIONE Enzo, H, Maître de Conférence, Telecom Paris, France (Rapporteur),
3. (Ext) LARABI Chaker, H, Professeur, Université de Poitiers, France (Président),
4. (ED) HELLIER Pierre, H, Chair de Professeur Junior, Université de Rennes, France
(Examinateur)
5. (Ext) BIDOKHTI Shirin Saeedi, F, Associate Professor, University of Pennsylvania,
(Examinatrice),
6. (ED) MAUGEY Thomas, H, Directeur de Recherche, Inria, Rennes, France (Directeur)