Learning-based imagevideo compression is emerging as a promising technology in recent years. This submission includes three major outcomes of our recent study. First, ANFIC presents an end-to-end learned lossy image codec. It features Augmented Normalizing Flows as its backbone, achieving the state-of-the-art performance among the existing learned codecs. Second, our CANF-VC extends ANFIC to video compression based on conditional coding, a new video coding paradigm. It represents the first end-to-end video codec that achieves comparable performance to HEVC. Third, we showcase how reinforcement learning can be leveraged to train an agent that learns to control the encoder parameters without making any change to the codec.

ANFIC presents the first attempt to leverage variational autoencoder (VAE)-based compression in a flow-based framework. It advances by stackingextending hierarchically multiple VAE＇s. CANF-VC extends ANFIC to video compression. With recent research on conditional coding showing the sub-optimality of the traditional hybrid-based coding, CANF-VC represents a new attempt that leverages the conditional ANF to learn a video codec for conditional video coding. Lastly, in our reinforcement learning based encoder control framework, we improve upon DDPG algorithm for maximizing reconstruction quality under the rate requirement. All three methods show superior compression efficiency compared to other learned methodstraditional codecs.

Image/video compression is a commercially proven technology. ISO/IEC & ITU-T have standardized several image/video coding technologies, which now find wide applications in social media, cloud gaming, video conferencing, etc. The rising of deep learning is opening up new opportunities for disruptive innovations to revolutionize this seemingly mature sector. Our ANFICCANF-VC represent the state-of-the-art learning-based imagevideo coding technologies in this fast-growing area. In addition, our reinforcement learning-based coding framework is able to improve on existing codecs without making any change to them. It represents the forefront technology in AI-assisted compression.