PeterLee 2007-04-20
【会议时间地点】
32nd Meeting: San Jose , CA , USA , 20-21 April, 2007
【提案内容】
VCEG-AF05:Extension of Intra Prediction Using Multiple Reference Lines
帧内预测改进提案。原帧内预测只利用了块上方的一行和左边的一列作为参考,本文提出了多个水平参考行和多个垂直参考列的思想,多行预测提高了预测的准确性。
Gain: ΔPSNR of 0.05dB and ΔBitrate of 0.42%.
VCEG-AF06:Simulation results of Bidirectional Intra Prediction on KTA software version 1.3
双向帧内预测VCEG-AE14的实验结果:
The experimental results revealed that the bitrate reduction of up to 4.8 % was achieved compared with H.264 intra prediction.
双向帧内预测提案内容:
l 两种预测模式得到的预测块加权叠加,
l 新的子快编码顺序
VCEG-AF07:Internal bit depth increase except frame memory
提升codec处理过程中的像素位深(帧存储除外):
对于数字信号的处理,处理过程中的位数越深,精度则越高。视频源的采样位数一般为8bits,如果在codec处理过程中提高像素位数到12位,可以使预测更精确,从而提高压缩效率。然而,帧存储中的像素如果也提升到12位,则为明显增加缓冲区占用,所以,在提案中,帧存储依然采用8bits,12位和帧存储8位的转换,作者引入了adaptive rounding的方法。
实验结果:
As a result, average of over 7% and up to 15.90% of bitrate saving for 720p sequences was achieved and the bitrate loss by frame memory compression was less than 1% on average.
该提案为VCEG-AE13的改进版,系统框图如下:
VCEG-AF08:Simulation Results of AQMS on KTA Software version 1.3
Adaptive Quantization Matrix Selection (AQMS)
实验结果:
The experimental results show that the bitrate reduction of up to 5.7% with AQMS is achieved.
VCEG-AF09:Syntax of adaptive filter coefficients in the KTA reference model.
Based on VCEG-AD08。
2D自适应插值算法:
H.264原插值算法采用的是1D6点滤波器,VCEG-AD08提出使用2D维纳滤波插值的方法,且每个滤波系数都可以自定义设置,目的在于提高插值精确度,从而提升编码效率。
VCEG-AF10:Geometry-adaptive Block Partioning
背景:
一个宏块会包含多个几何区域,不同几何区域通常对应不同的对象,其纹理和运动情况都不相同。H.264四叉树分块方法,允许16x16至4x4的分块大小,这样的分块方法很容易发生一个块跨多个几何区域的情况。
基于几何形状自适应的分块方法:
根据几何形状,将宏块分成两个区域,分割线归正为一条直线。这两个区域分别作运动估计补偿。
两种分块方法的比较,如下图:
实验结果:
Results show that geometry-adaptive block partitioning consistently improves coding efficiency of Inter-P frames for common test QCIF and CIF sequences in a 5% of average.
主观质量比较:
PS:
基于几何形状自适应的分块方法也可用于I帧的编码。
VCEG-AF11:KTA 1.3 software manual
KTA tools implemented in version 1.3 are:
· Adaptive interpolation filters – Hannover University .
· Motion compensated prediction with 1/8-pel vector resolution – Hannover University .
· Motion vector competition – France Telecom R&D.
· Adaptive prediction error coding in spatial and frequential domain – Hannover University .
· Adaptive quantization matrix selection – Toshiba.
VCEG-AF12:On spatio-temporal adaptation of MC interpolation filter
基于时空自适应的插值滤波算法。
对VCEG-AF09算法的速度优化,VCEG-AF09提出的2D插值算法复杂度提高了3倍左右。本文采用了6-tap filter, 4-tap filter, 2d 6-tap filter, 2d 4-tap filter四种filter,通过帧级和宏块级的算法自适应地选择filter.
实验结果:
Experimental results show that on average, the complexity of interpolation filter could be reduced by 56% with around 0.08 dB penalty on coding efficiency.
VCEG-AF13:A new scheme for motion vector predictor encoding
H.264原来的PMV计算方法是从当前块相邻的几个MV中取中值。本文对几个候选MV的bits代价作了比较,选择代价最小的MV作为PMV。
实验结果:
The proposed method achieves 1.11% bitrate reduction on average in case three spatial predictors are used.
VCEG-AF15:Improved Intra Coding
在原帧内预测编码算法上的改进:
1、方向预测向量采用如下公式:
.
其中,w表示权重向量,S表示参考像素值。
2、transform的改进:
帧内预测之后的残差系数的分布特点和预测模式有很强的关联性,因此,根据不同的预测模式可以采用不同的transform变换矩阵。
实验结果:
The average improvements are 8.6% for 8 CIF sequences and over 10% for 4 4CIF sequences.
VCEG-AF16:A note on macroblock-level bitstream consistency
本文认为H.264的宏块级码流结构设计存在冗余。