Directional Multi-hypothesis Prediction for Improving H.264/AVC Video Coding

In this letter, we propose efficient multi-hypothesis prediction (MHP) algorithms for high performance video codec. First, candidate MHPs are generated by combining the seed predictors around the initial predictor which is obtained by motion estimation. Among candidate MHPs, we select the optimal MHP predictor which have the lowest rate-distortion (RD) cost and signal the index as side information. We also propose directional MHP (DMH), which combines only two predictors along the predefined direction to minimize both encoding complexity and signaling overhead. In the experimental results, proposed MHP algorithms show better performance than H.264/AVC. Specifically, the average bit saving of DMH over H.264/AVC is up to 9.55%. The results also indicate that the performance is better at the high bit rate.