PU-type-dependent reference frame selection method for HEVC

Abstract

This paper proposes an efficient reference frame selection method for HEVC. The proposed method makes use of the correlation between the best reference frame (BRF) in the $2N\times 2N$ prediction unit (PU) and the BRFs in the other types of PUs. By efficiently using the BRF information of the $2N\times 2N$ PU, the proposed method significantly reduces the number of reference frames to be examined. To minimize the performance degradation, the proposed method changes the examination order of the inter-prediction modes and makes use of the best merge mode information. It also uses the rate-distortion cost information of the $2N\times 2N$ PU so as not to skip important reference frames. Because the proposed method deals with both uni-prediction and bi-prediction, it can be efficiently used, not only for LP but also for LB and RA coding structures. As a result, the proposed method reduces the encoding time by 12.7, 16.4 and 5.6 % with only 0.26, 0.29 and 0.04 % BD-BR increase for the LP, LB and RA coding structures, respectively.