摘要
多视点彩色加深度(MVD)视频是三维(3D)视频的主流格式。在3D高效视频编码中,深度视频帧内编码具有较高的编码复杂度;深度估计软件获取的深度视频由于不够准确会使深度图平坦区域纹理增加,从而进一步增加帧内编码复杂度。针对以上问题,本文提出了一种联合深度处理的深度视频帧内低复杂度编码算法。首先,在编码前对深度视频进行预处理,减少由于深度图不准确而出现的纹理信息;其次,运用反向传播神经网络(BPNN,backpropagation neural network)预测最大编码单元(LCU,largest coding unit)的最大划分深度;最后联合深度视频的边缘信息及对应的彩色LCU最大划分深度进行CU提前终止划分和快速模式选取。实验结果表明,本文算法在保证虚拟视点质量的前提下,BDBR下降0.33%,深度视频编码时间平均节省50.63%。
Multi-view video plus depth （MVD） is the main format of three-dimension （3D） video. Under the framework 3D high efficiency video coding （HEVC） the computational complexity of depth video intra coding is huge. Moreover,the inaccuracy of estimated depth video will bring about the growth of texture in the flat region, which significantly increases the computational complexity of intra coding. To tackle these problems,a preprocessing based depth video fast intra coding algorithm is proposed in this paper. Firstly,in order to remove the texture caused by inaccurate depth video, a smoothing method is applied before encoding. Then, the largest partition depth （LPD） of largest coding unit （LCU） in depth video can be predicted by backpropagation neural network （BPNN）. Lastly,the process of LCU partitio- ning can be terminated in advance according to edge information of depth video and the LPD of LCU in corresponding color video. Experimental results demonstrate that the algorithm can save the time of depth video coding by 50.63 % on average with negligible quality degradation of virtual view.
出处
《光电子．激光》
EI
CAS
CSCD
北大核心
2017年第10期1131-1138,共8页
Journal of Optoelectronics·Laser
基金
国家"863"计划(2015AA015901)
国家自然科学基金(61771269
61620106012
61671258
U1301257)
浙江省自然科学基金(LY15F010005LY17F010005
LY16F010002)
宁波大学科研基金(理)/学科项目(xkxl1502)资助项目
