一种基于深度可分离卷积的VVC帧内编码快速块划分算法

A fast block partitioning algorithm for VVC intra coding based on depthwise separable convolution

最近,联合视频探索工作组(JVET)将通用视频编码(VVC)作为新一代视频编码标准,它利用复杂的四叉树加多类型树(QTMTT)划分结构有效地提升了编码性能,但也导致编码复杂度急剧攀升,大幅地增加了编码时间。为解决上述问题,提出了一种基于深度可分离卷积的VVC帧内编码快速块划分算法,将编码单元(CU)的原始像素值作为输入,利用轻量化的深度可分离卷积神经网络提取CU纹理信息特征指导CU的划分模式选择,实现精准的划分模式预测。该方案通过跳过低概率的划分模式,减少CU划分模式的遍历,大幅地降低编码器的复杂度。实验结果表明,所提算法在VTM15.2平台上实现了18%~48%的编码时间节省,仅仅带来了平均0.15%的性能损失,并且轻量化的深度可分离卷积计算带来的额外复杂性也可以忽略不计。

The joint video exploration team(JVET)proposed versatile video coding(VVC)as a new video coding standard,and its quadtree plus multi-type tree(QTMTT)partition structure brings effective coding performance im-provements.However,it brings about a sharp increase in encoding complexity,which greatly increases the encoding time.In order to solve the above problems,a fast block partitioning algorithm for VVC intra coding based on depth-wise separable convolution was proposed.The pixel of coding unit(CU)was used as input,and the texture informa-tion feature of CU was extracted through depth-separable convolution.Therefore,accurate partition mode prediction was realized in the QTMT structure in VVC,and the complexity of the encoder was reduced by skipping low-probability partition modes.Experimental results show that the proposed algorithm saves 18%to 48%of encod-ing time on the VTM 15.2,and only brings an average performance loss of 0.15%.And the additional complexity brought by the lightweight depthwise separable convolution calculation is also negligible.