Prediction unit depth selection based on statistic distribution for HEVC intra coding

The new video coding standard, HEVC, achieves better compression performance relative to H.264 at the expense of a significant increase in computational complexity. It adopts coding tree structure, and implements partition and mode decision for each coding unit (CU) and prediction unit (PU). For intra coding, five levels of PUs from 64×64 to 4×4 should be searched and each PU selects one best mode from 35 candidate modes, which is very computation intensive. Therefore, reducing the number of levels for PUs, i.e., PU depth selection, is an efficient way to decrease intra coding complexity. In this paper, we fully investigate the relationship between the variances of different sizes of pixel blocks and the best partition of PUs. It is shown that 90th percentile of the variances distribution can be used to distinguish large PUs and small PUs. Furthermore, 16×16 PU search can be disabled for smooth area when variances of 64×64 pixel blocks are smaller than the 90th percentile of variances distribution of 64×64 PUs and this 90th percentile is smaller than an absolute threshold. Meanwhile, further split can be disabled if the variance of the current PU is smaller than 70th percentile of the variances of its parent-level PUs when the current PU is 16×16 or 8×8. The experimental results show that the proposed algorithm can reduce 55% encoding time of reference encoder HM10.0 at all Intra setting on average with negligible coding performance loss less than 1.22% BD-Rate for HD or Ultra-HD videos.