GOP Structure-Independent Quantization Parameter Cascading in Video Coding

The popular lossy coding standards set unequal quantization parameter (Qp) values to its successive frames or slices aimed at minimizing its encoder distortion subject to a bitrate constraint. Nowadays, this Qp cascading (QPC) strategy is normally achieved by the derivation of an optimization problem with regard to the group-of-picture (GOP) structure of successive frames. In this paper, we propose to extract coding unit (CU)-level predicting information and project them to frame-level dependency with real-time neural networks. Thus, we successfully estimate and update the frame-level dependency in a GOP without prior knowledge to the GOP-level prediction structure. Finally, the dependency is utilized in adaptive QPC for diverse GOP structures. The experimental results on high efficiency video coding (H.265/HEVC) demonstrate the effectiveness of our proposed framework, which achieves adaptive QPC for both random access (RA) and low delay (LD) structures with superior rate-distortion (RD) performances.