Approximate interpolation filters for the fractional motion estimation in HEVC encoders and their VLSI design

Motion Estimation (ME) is one of the most complex HEVC steps, consuming more than 60% of the average encoding time, most of which is spent on its fractional part (Fractional Motion Estimation - FME), in which sub-pixel samples are interpolated and searched over to find motion vectors with higher precision. This paper presents hardware designs for the sub-pixel interpolation unit of the FME step. The designs employ approximate computing techniques by reducing the number of taps in each filter to reduce memory access and hardware cost. The approximate filters were implemented in the HEVC reference software to assess their impact on coding performance. A complete interpolation architecture was implemented in VHDL and synthesized with different filter precision and input sizes in order to assess the effect of these parameters on hardware area and performance. The approximate designs reduce the number of adders/subtractors by up to 67.65% and memory bandwidth by up to 75% with a tolerable loss in coding performance (less than 1% using the Bjontegaard Delta bitrate metric). When synthesized to an FPGA device, 52.9% less logic elements are required with a modest increase in frequency.