在一个典型的视频编码系统中,其中运动估计(ME)的运算量大约占总运算量的60％～90％,尤其未来的数字电视将以高清和超高清标称,意味着在保持一定的信噪比条件下更大的搜索范围.本文提出了一种基于以上算法优势特点的运动估计算法及其硬件实现――分层全搜索运动估计(HFSME),该算法支持AVS标准的多参考帧技术(B,P帧均采用2幅图像作为参考帧)、率失真优化(RDO)和4种宏块分块模式(16×16,16×8,8×16,8×8),并满足AVS高清实时编码器的需求.从性能上该算法及其硬件实现支持AVS Baseline级高清1080P,帧率达到30 fps,搜索面积达到234×98像素,相较于同样性能下的全搜索算法,峰值信噪比(PSNR)相差不大,但运算量只有全搜索算法运算量的1/4.同时,相较于文献[12]中的结构,本文设计的基于HFSME算法的IME(整像素运动估计)结构设计,在处理单元(PE)规模上是文献[12]结构的2倍,但是搜索范围能力是其4倍,并且吞吐率是其10倍,具有最优的性能,带来了良好的性价比.%In a classic video encoder, Motion Estimation(ME) comes as a high computational complexity, about 60％-90％ in a whole system. An efficient algorithm named as Hierarchical Full Search ME(HFSME) is proposed and implemented on the circuit. Based on AVS HD encoder, it supports the AVS variable block size match in which both B and P frames adopt two pictures as the reference frames, Rate Distortion Optimization(RDO) and four kinds of macro-block partition modes concurrently. It totally produces 18 motion vectors for each macro-block. Experimental results show that the architecture can achieve 30 fps under the conditions of two reference frames, 234×98 search range and 1920×1080 picture size. The algorithm & architecture proposed in this paper can meet the requirements of AVS HD Real-Time Encoder. The computational complexity of HFSME proposed in the paper is only a quarter of that of full search. At the same time, HFSME can provide the highest computational capability which is about 10 times higher than that of reference [12].
展开▼
