面向下一代视频编码标准的高效算法研究

This research project aims at the design of efficient compression algorithms for the next generation video coding, including intra-picture predictive coding, in-loop filtering and RGB video coding. To achieve this goal, we will first develop an efficient coding method for intra-predicted residuals through hiding the DC coefficient and a pixel-domain compensation algorithm. Secondly, we will design an advanced de-blocking filter based on partitioning image blocks' boundary region into two parts, i.e., local variation part and boundary variation part, according to the characteristics of the reconstructed pixels within the region. A more effective in-loop filtering will be achieved by applying a carefully-designed curve-fitting only on the boundary variation part. Thirdly, we will focus on efficient coding of RGB video signals to make the coded outputs fitting perfectly to displayers that require RGB inputs. To this end, we will extend the rate-distortion optimization (RDO) theory that was originally established in the YCbCr space to the RGB space and then develop effective solutions through an intelligent quantization technique. Putting together all efforts in these three attempts, we aim at achieving more than 5% coding gain over the existing H.265/HEVC standard.

本课题瞄准下一代视频编码标准，针对帧内预测编码、环路滤波和RGB视频编码优化三个问题展开新型高效编码算法的研究工作。内容包括：（1）开发基于直流系数隐藏和像素域信息补偿的帧内预测编码算法，通过对预测残差的高效编码提高帧内编码的效率；（2）设计基于图像块边界区域变化特征的去块效应滤波：将图像块边界区域内的信号分解为代表图像本身纹理变化和代表图像块边界变化的特征分量，对后者实施基于曲线拟合策略的滤波，通过提高去块效应滤波的效率而实现高效环路滤波；（3）针对显示设备需要输入RGB信号的特点，利用RGB和YCbCr信号的转换关系，将当前YCbCr空间的编码“率-失真”优化问题延伸定义到RGB空间，并开发新型的智能量化算法进一步提高RGB视频的编码效率，实现彩色视频信号高效编码输出和显示设备输入的有效统一。本课题的研究目标是：实现整体编码效率较H.265/HEVC提高5%以上。

本项目瞄准新一代视频编码标准，针对帧内预测编码、环路滤波和RGB视频编码优化三个问题展开新型高效编码算法的研究工作。主要创新如下：(1)提出了以直流系数隐藏和预测为基础的新型帧内编码方法，构建了基于局部能量变化特征的图像复原模型用于视频重建，新型帧内编码方法与HEVC帧内编码方法相比，性能提升最高可达2%；(2)研究了视频先验信息对编码性能的影响，提出了基于视频时空先验信息和深度学习的环路滤波算法，该方法在应用于HEVC编码后，性能提升最高可达11%；(3)针对RGB彩色图像编码率-失真优化问题展开研究，提出了以RGB-YCbCr颜色空间转换误差为指导的马赛克图像编码方法，该方法与当时性能最优的方法比较，性能提升5%左右。课题组成员发表学术论文19篇，其中SCI期刊论文12篇、EI会议7篇；申请发明专利18项，其中授权7项，转让1项。