基于失真传递的时域自适应量化算法

码率控制是视频编码器中的关键模块,其算法直接决定编码器率失真性能.视频编码帧间预测导致的编码失真会在时域产生传递效应,考虑该传递效应是优化码率控制算法性能的关键.宏块树码率控制是一种典型的时域量化控制算法,核心是根据编码单元失真传递量(相对传递代价ρ)自适应地调整量化参数(偏移量δ),合适的δ-ρ映射关系是宏块树量化控制算法的核心.宏块树算法采用基于经验的δ-ρ模型,对不同视频序列的普适性有待改进,模型准确度和精度也需进一步优化.针对上述问题,将竞争决策方法用于探索最优δ-ρ映射关系,提出了一种率失真性能优化的失真时域传递自适应量化δ-ρ模型,以改进时域自适应量化算法.实验结果表明,信噪比BD-PSNR较原模型提升了0.14dB以上,SSIM性能提升了0.29dB.算法能更好地控制码率时域分配,降低失真时域传递恶化.

A temporally adaptive quantization algorithm with constrained distortion propagation in video coding

Rate control is crucial to rate distortion performance optimization in video coding design. In video coder, temporal prediction bring about distortion propagation along adjacent frames, and it is an efficient way to further improve the video coding efficiency by taking the temporal distortion dependency into consideration. The MBTree rate control is a typical temporal quantization control algorithm, in which the quantization parameter offset δ is employed for quantization adjustment according to the distortion propagation amount, i.e. the relative propagation cost ρ. An appropriate δ-ρ model is therefore the key for the MBTree-like adaptive quantization algorithm. Nevertheless, the current δ-ρ model in MBTree algorithm is designed in an empirical way with rough accuracy. This model has unsatisfactory universality to different video sequences, thus there is still room left to be improved. This paper focuses on this problem and applies the competitive decision mechanism in exploring the optimized δ-ρ model, and then proposes an improved δ-ρ model with rate distortion optimization. The simulation results show that the improved MBTree algorithm based on the proposed model can achieve up to 0.14 dB BD-PSNR improvement and 0.29 dB SSIM improvement. The proposed algorithm can also implement better bit allocation in temporal domain and reduce the temporal distortion fluctuation, achieving adaptive quantization control.