Adaptive coding in JPEG AI
Description
The recently adopted image coding standard JPEG AI uses neural networks to encode images efficiently. In doing so, it achieves a comparatively low bitrate while maintaining high image quality. Both in classical and learning-based coding, it has been shown that techniques that adaptively respond to image content enable higher compression rates.
When adopting a standard, only the decoding side is ever defined, so that encoded images can be viewed on any end device. Since the encoding process is not part of the standard, this leaves many opportunities for further improvements to be developed.
As part of a thesis project, various content adaptive encoding optimizations can be implemented and tested.
Requirements
Requirements for a thesis in this topic are
- Solid foundation in Python
- Experience with image and video compression (e.g. lecture, lab course)
- Experience with deep learning (PyTorch)
Supervisor
Simon Deniffel
simon.deniffel@fau.de
Room 06.036