基于3D-HEVC的低复杂度编码模型与方法研究

The development of three-dimension video is a main task of the new generation information technology at this stage, and how to reduce the complexity of 3D-HEVC encoder is an unsolved core issue. Currently, most of researches focus on improving the coding performance of 3D-HEVC. However, the introduction of new prediction and transformation coding techniques dramatically increase computational complexity for 3D-HEVC, which obstructs the rapid extension and practical use of the new generation three-dimension video. In consideration of the characteristics of multi-view texture video in new generation three-dimension video, a structured fast model will be built to reveal the constructed redundancy relationship of multi-view texture video. Meanwhile, according to harmonic analysis theory, texture video coding will be designed for fast depth level, coding unit and adaptive prediction based on machine learning and JND (Just Noticeable Distortion) technologies. For distributives of depth map, an optimization model will be deduced based on foreground/background, edges and mode decision, and proposes low complexity depth map coding approach. In addition, a low complexity texture-depth joint coding model will be set by using the relevant and distinct feature between texture video and depth map coding, which can significantly reduce the computational complexity of 3D-HEVC while maintaining nearly the same rate distortion performance as the original encoder. Research of the project can provide new theoretical achievements and methodological foundation for accomplishing low complexity three-dimension video coding, and make a contribution for three-dimension video international standardization development.

发展三维视频是信息技术现阶段的主要任务，其中复杂度高的3D-HEVC编码是尚未解决的核心问题。目前研究方法大多集中在提高3D-HEVC编码性能，但新的预测和变换工具引入使得3D-HEVC计算复杂度急剧增加，阻碍了三维视频编码技术的发展与应用。本项目针对三维视频中多视点纹理视频的特性，构造结构化的快速编码模型，揭示多视点纹理视频结构化冗余关系，运用机器学习和JND理论设计基于深度分级、编码单元和自适应模式选择的快速纹理视频压缩；针对深度图自身的分布特性，推导基于前景背景、边缘和模式的判决优化模型，探讨计算复杂度低的深度编码方法；针对纹理视频和深度图编码之间既关联又区别的特点，建立二者联合的低复杂度纹理-深度编码模型，有望在整体不降低编码效率的前提下，大幅度提高3D-HEVC编码速度。本项目研究为探索低复杂度三维视频编码提供新的理论成果和方法基础，对促进三维视频编码的国际标准化进程有积极作用。

本项目主要探索新一代三维视频的低复杂度编码系统中若干未解决的关键科学问题。获得主要成果如下：1）针对目前3D-HEVC压缩三维纹理视频时编码结构复杂效率低的问题，构建新的3D-HEVC视频编码结构，减少不需要执行的压缩模式，提高3D-HEVC纹理视频的编码速度。2）针对目前3D-HEVC压缩深度图时深度模型预测计算复杂度高的问题，设计快速的三维视频深度图像编码方法，可对同质区域进行粗略编码而对边缘部分进行精细编码，从而加快3D-HEVC深度编码速度。3）针对目前3D-HEVC快速算法大多对纹理视频和深度编码两者进行单独优化的问题，通过构建纹理-深度联合的编码模型，从整体上保持三维视频质量同时获得编码效率高、复杂度低的3D-HEVC编码方法。4）针对目前3D-HEVC压缩三维视频时较少考虑视觉上的冗余特性且计算复杂度较高，通过研究人类视觉系统特性，开发基于纹理变化的空间-时间敏感度JND编码模型，从而对三维视频编码单元进行快速划分。本项研究开展以来，已在国内外重要期刊上发表论文30篇，其中SCI检索21篇；申请发明专利24项，授权发明专利6项，计算机软件著作权8项，出版学术专著1部，获河南省科技进步一等奖1项、二等奖2项、三等奖1项。上述研究成果，有力支撑了低复杂度三维视频编码系统的开发利用，为探索低复杂度三维视频编码提供新的理论成果和方法基础。