Separable adaptive interpolation filter for video coding

Motion-compensated prediction using fractional-pel motion vectors followed by transform coding of the resulting prediction error is used in hybrid video coding. In the case of fractional-pel motion, pixels at fractional-pel positions have to be determined by interpolation. For this purpose, fixed interpolation filters are applied in H.264/AVC. By using fixed interpolation filters, time varying effects such as aliasing, quantization errors, errors from inaccurate motion estimation, camera noise, etc cannot be considered accurately. Thus, the coding efficiency of the motion compensated prediction is limited. The concept of adaptive interpolation filtering addresses these effects resulting in an increased coding efficiency. Since a non-separable adaptive filter is used in prior art, it is associated with a significantly increased computational expense at encoder and decoder. In order to reduce the computational expense, a separable adaptive interpolation filtering is proposed in this paper that achieves the same coding efficiency than the non-separable adaptive filter. With this separable interpolation filter, the computational expense of the filtering is reduced by 24% in case of 4times4 motion-compensated blocks, 36% in case of 8times8 motion-compensated blocks, and 42% in case of 16times16 motion-compensated blocks compared to a non-separable filter, whereas the computational expense is measured by the number of calculation operations.