CTO Giovanni Ballocca and engineer Alessandra Mosca were part of a research team that used AI to boost EVC compression efficiency by 26%

A research team including representatives from Sisvel Tech has integrated AI-based tools into the MPEG-5 EVC standard, boosting the codec’s performance by 26%. A peer-reviewed paper detailing their work was published last month by Multimedia Tools and Applications, the premier journal in the field.

The research was conducted under the auspices of the MPAI standards organisation, a group founded in 2020 to develop AI-enabled data coding specifications. Sisvel Technology Chairman and CEO Massimo Marcarini serves on MPAI’s board of directors.

Sisvel CTO Giovanni Ballocca and Sisvel Tech patent engineer Alessandra Mosca are among the study’s eight authors; they contributed to preparing and analysing data, conducting the experimentation and writing the paper. Their co-authors included experts from the CYENS Centre of Excellence in Cyprus, the University of Turin, the Polytechnic Institute of Paris, RAI and MPAI.

“To the best of our knowledge, this work represents the first attempt to enhance the EVC standard using deep learning tools,” the authors write. The EVC codec, finalised by standards body MPEG in 2020, was chosen because it already achieves state-of-the-art compression efficiency using traditional, non-AI frameworks.

The paper makes two major contributions:

1. It proposes a method for replacing two coding blocks within EVC with AI-based tools; and

2. It demonstrates significant performance gains that result from integrating this AI-based encoding framework.

The authors identified two areas within the EVC specification where neural networks could make an impact. The first is ‘super-resolution’, a post-processing step in which video is restored to its native resolution. AI was deemed a good candidate for completing this task because deep neural networks have demonstrated the ability to learn complex mapping functions between low-resolution and high-resolution images.

AI tools were also deployed to handle a data-saving process called ‘intra-prediction’. This is where the codec predicts what a block of pixels looks like by referencing pixels nearby that have already been decoded.

Detailed information on the datasets, experiments and results can be found in the paper. The headline finding was a 26% increase in BD-rate, a compression efficiency metric that measures average bitrate savings between two codecs (or configurations) while keeping video quality constant.

The researchers also showed that the gains from the use of AI tools were additive rather than isolated improvements, and that they should be generalisable to real-world applications.

The team has identified several avenues for future work:

1. Applying advanced techniques such as in-loop filtering to achieve even greater performance gains;

2. Further optimising the proposed AI architecture to reduce its computational complexity while maintaining the efficiency gains; and

3. Exploring the application of these tools to other video codecs.

The full paper is available here.