An efficient fractional motion estimation architecture for avs real-time full Hd video encoder

Abstract

Fractional motion estimation (FME) as a complement to integer motion estimation (IME) conducts higher compression rate in video coding. Based on multiple reference frames and all modes of variable-block-size (VBS) motion search, the common FME algorithm and architecture generally causes high computational complexity. It consumes plenty of time and hardware resources to process the real-time (30fps) high definition (HD) video compression. In this paper, we propose an efficient FME algorithm and architecture that significantly reduce the number of the candidate sub-pixels by predicting quarter motion vector and cost less hardware by saltatory interpolation. Compared with the conventional FME architecture, the proposed method reduces resources significantly and accelerates the processing speed. 75% circuit area and 68% computational complexity are saved with only 0.1dB drop in performance and similar bit rate saving on average. Experiments of various sequences show that HD video (1920×1080) can be processed in real-time at frequency of 100MHz and it is verified in an AVS HD encoder on a Xilinx Virtex-6 FPGA prototype system.