A Wavelet-Guided Content-Adaptive Framework For Fast Intra-Mode Prediction In H.264/AVC

Abstract

Intra-mode decision is an expensive step in H.264/AVC encoding, as prediction modes need to be evaluated to find the mode with the least rate-distortion cost (RDC). In this paper, we propose a Wavelet-Guided Content-Adaptive Intra-Mode Prediction (WG-CAMP) approach to reduce such intra-mode evaluation in the coding process while still offering competitive coding performance. The proposed approach employs a two-level Daubechies-2 (db2) wavelet decomposition on each 16 × 16 luminance block and incorporates horizontal, vertical, and diagonal wavelet information in the space using a structure tensor to estimate the orientation of the dominant block. The estimated orientation is mapped to the three nearest directional H.264/AVC intra modes, while vertical, horizontal, and DC modes are kept as safety candidates and a small set of content-adaptive candidate information is used for rate-distortion optimisation (RDO). We compare our proposed framework with Full RD Search and SATD-based Fast Search algorithms on Coastguard, Container, Foreman, and News CIF sequences with QP of 20, 24, 28, and 32. Experiments showed that WG-CAMP reduces the average encoding time from 72.07 s for Full RD Search to 25.87 s, bringing 63.73% time savings and a 2.79× speedup and improves the average PSNR from 39.443 dB to 39.491 dB and SSIM from 0.95214 to 0.95273. In particular, the proposed method introduces only a 0.78% BDBR increase compared to Full RD Search (2.11% for SATD Fast Search) while keeping the Full-RD optimal mode in approximately 81.03% of blocks with an average candidate set of about 5.04 modes. These results prove that wavelet-domain orientation provides a good low-complexity descriptor of content-adaptive intra-mode prediction, and the coding efficiency and computational complexity trade-off can be made with WG-CAMP. Our WG-CAMP framework is a good candidate for low-complexity H.264/AVC encoding in applications where less encoding time is needed, and the RD performance is not significantly affected.