Screen Content Encoding Network Based on Deep Contextual Information

This paper proposes a deep contextual video coding network that integrates transformational jumps. The network takes a sequence of original in-screen content video frames as input and extracts motion features through motion estimation and a motion information encoder, with motion compensation occurring during decoding. It also extracts contextual information in the feature domain through a context encoder. This contextual information assists in both context encoding and decoding, as well as entropy encoding. By using contextual information as a condition for conditional encoding, the network transitions from predictive coding to conditional coding, aiding in the high-quality reconstruction of high-frequency content. The motion information encoder introduces a transformational jump branch into the analysis and synthesis process. This branch has the ability to extract coarse features and reconstruct them, thereby enhancing the encoding and decoding of visual signals. Finally, experiments confirm that this algorithm effectively improves the encoding performance of screen content videos.