Fixed-point realization of affine motion compensation in HEVC
Chihiro Tsutake, Toshiyuki Yoshida · 2018 International Workshop on Advanced Image Technology (IWAIT) · 2018
Affine motion estimation/compensation (AME/AMC) is known as one of promising techniques for inter prediction coding, which can efficiently remove temporal redundancy in video sequences. In our previous work, we have proposed two 3-parameter affine motion models and their block-matching-based efficient search technique for HEVC, and have demonstrated its superiority over conventional techniques. Although the AME/AMC in the technique has been implemented in afloating-point arithmetic, fixed-point arithmetic is preferable in an actual hardware realization. This paper thus investigates the optimal fractional bit lengths for the AME and AMC to realize them in fixed-point arithmetic. They have been determined by considering both of the coding and computation efficiencies in the experiments given in this paper.