Brain-inspired Intelligence for Semantic Video Compression

The BrainCode project proposes novel compression techniques for extended reality (XR) data which are energy efficient while ensuring a reconstruction quality that satisfies the human visual semantic perception. There are several challenges concerning the complexity and the power consumption of the latest video compression standards which have not yet taken into account by the signal processing community.

We propose that these challenges can be addressed by machine learning based architectures in order to avoid the exhaustive

comparisons between sequential frames. We aim at releasing a semantic video compression algorithm that uses Convolutional Neural

Networks (CNNs) and drives the bit allocation with respect to the content of the visual scene. Another goal of BrainCode is to mimic

the visual system as an intelligent mechanism that processes the visual stimulus. This can be claimed as it consumes low power, it

deals with high resolution dynamic signals and the dynamic way it transforms and encodes the visual stimulus is beyond the current

compression standards. During the last decades, a lot of effort has been made to understand how the visual system works, what is the

structure and role of each layer and individual cell that lies along the visual pathway, and how the huge visual information is

propagated and compacted through the nerve cells before it reaches the visual cortex. There are very interesting mathematical

models which approximate the neural behaviour and they have been widely used for image processing applications including

compression. The BrainCode project searches the latest neuroscience models for the design of a groundbreaking XR video compression

architecture. The efficiency of the above approaches is expected to improve several image processing applications like computer

vision, virtual reality, and video compression among other, where the real-time processing of the visual scene plays a substantial role.