Hierarchical Random Access Coding for Deep Neural Video Compression

Abstract

Recent advancements in neural video compression networks have achieved notable results; however, most existing models focus primarily on low-delay configurations, where the order of display matches the coding order. In this paper, we propose a hierarchical random access coding approach that leverages bidirectional temporal redundancy to enhance the coding efficiency of current deep neural video compression models. Our framework integrates a video frame interpolation network to improve inter-frame prediction and introduces a hierarchical coding structure. Experimental results demonstrate that the proposed framework increases coding efficiency by 48.01% on the UVG dataset and 50.96% on the HEVC-class B dataset, significantly outperforming previous deep neural video compression networks.