基于卷积神经网络的HEVC帧内预测算法优化

HEVC intra-frame prediction algorithm optimization based on CNN

作为HEVC标准中最基础、最重要的技术之一,帧内预测对实现视频编码的高速、高质量和高压缩率具有重要的作用。文中针对帧内预测复杂性问题进行研究,提出一种基于深度卷积神经网络(CNN)的方法,通过学习来预测CTU的划分,从而减少HEVC帧内编码的复杂性。通过建立一个大规模的CTU划分数据库,并利用CNN的能力学习各种CTU划分模式,能够准确地预测CTU的划分,从而避免了传统的穷举搜索,实现了HEVC编码复杂性的显著降低,提高了编码效率。实验结果表明,提出的方法在测试序列和图像上分别将帧内编码时间减少了62.25%和69.06%,与其他最先进的方法相比,比特率分别仅增加了2.12%和1.13%,达到了优化的目的。

As one of the most fundamental and crucial technologies in the HEVC(high efficiency video coding)standard,intra-frame prediction plays a crucial role in achieving high speed,high quality and high compression efficiency in video coding.This paper addresses the complexity issue of intra-frame prediction and proposes a method based on deep convolutional neural networks(CNNs)to predict CTU(coding tree unit)partition by learning,thereby reducing the complexity of HEVC intra-frame coding.By establishing a large-scale CTU partition database and using the learning capability of CNN to study various CTU partition patterns,the CTU partition is predicted accurately,so as to avoid the traditional exhaustive searches,reduce the complexity of HEVC encoding significantly and improve the coding efficiency.Experimental results demonstrate that the proposed method reduces intra-frame coding time by 62.25%and 69.06%for the test sequences and the images,respectively.In comparison with the other advanced methods,its bitrate increases by only 2.12%and 1.13%,which achieves the purpose of optimization.