基于深度残差网络的VP9超级块快速划分算法

Fast superblock partition decision based on deep residual network for VP9

VP9和HEVC同属新一代高效视频编码标准,因为代码开源且无需专利授权费的优势,在编码领域中受到了越来越多的关注与使用。在VP9的编码过程中,由于需要对超级块遍历所有划分的情况进行编码,耗费大量的时间在各种划分模式中进行编码,占用了大量的复杂度。本文提出一种基于深度残差网络的VP9超级块划分算法,针对VP9中超级块划分的过程进行快速预测,从而规避其递归计算最优划分结构的过程,实现了VP9超级块的快速划分。该算法通过建立一个级联的残差网络完成特征的提取,然后使用三层分类器来预测VP9的超级块划分。实验结果表明,与传统的基于率失真优化的超级块划分算法相比,文中提出的VP9超级块划分算法平均节省了65.16%的时间,同时只平均增加2.28%的码率。

In the open-source video codec VP9,superblocks with size 64×64 can be recursively split into multiple smaller blocks.In a brute-force recursive search,VP9 encoder performs the encoding process for all partitioning structures and selects the optimal one with the minimal rate-distortion cost,which consumes a large proportion of the encoding complexity.Therefore,in order to reduce the encoding complexity,this paper proposes a deep residual network based fast superblock partitioning decision.We first analyze the distribution of the partitioning modes in a statistical way and model the partition modes as a multi-class classification problem.Then we design a hierarchical CNN architecture to predict the three-level block partitioning modes.To reduce the complexity of neural network,the three-level classifiers share the same deep features.Finally,the experimental results present that our method can reduce VP9 encoding time by 65.16%with 2.28%BD-BR increment com-pared with the original VP9 encoder.