The Video Communication and Applications (VCA) department at the Fraunhofer Heinrich Hertz Institute (HHI) received an award last week at this year's Picture Coding Symposium (PCS 2025) in Aachen. The paper “UNNIP: Memory Optimized Universal Neural Network Intra-Prediction for Custom On-Chip Hardware Architectures” took second place in the Best Paper Awards.
Doctoral student Viktor Herrmann, a member of the Embedded Systems working group led by Prof. Dr. Benno Stabernack, is conducting research into special hardware architectures for ML-based video coding. In this context, Viktor Herrmann has developed an innovative approach that specifically combines the theoretical principles of neural networks with hardware-oriented optimization.
The work addresses a central problem of ML-based video coding: the high memory requirements of neural networks, which have so far severely limited their use in hardware architectures with limited energy and chip space. By using a universal feature extraction backbone and specifically optimized machine learning methods, the required memory requirements could be reduced to about one-tenth of those of conventional approaches.
Research focus: Memory-optimized ML for video encoding
A universal feature extraction backbone significantly reduces memory requirements—a crucial step for the practical application of ML-supported video encoding. An FPGA-based prototype, i.e., a preliminary version of the future chip that only shows some of its future functions, demonstrates that the architecture also works under real-time conditions for Full HD video at 30 FPS.
Impetus for future video encoding standards
The work shows how ML-supported processes can be efficiently integrated into video encoding frameworks. The team is thus making an important contribution to the further development of future standards and the practical implementation of powerful encoder and decoder hardware.
With its 2nd place in the Best Paper Award, the VCA department at Fraunhofer HHI, in collaboration with the University of Potsdam, underscores its leading role in the international research community and sends a strong signal about the institute's innovative strength.
Further information about the symposium can be found here.
