Simplified Level Estimation for Rate-Distortion Optimized Quantization of HEVC

Abstract

The rate-distortion optimized quantization (RDOQ) in High Efficiency Video Coding (HEVC) brings a significant coding gain but also leads to huge computation at encoder. Since the level estimation (LE) demands dominant computational complexity among the three major internal stages inside the RDOQ, this paper addresses speeding up of the RDOQ process by simplifying the LE stage. Under the observation that in particular cases, some quantization levels are very likely to be changed (or unchanged) after the RDOQ, and zero levels do not need to go through the RDOQ process, we propose an adaptive dead-zone scheme which can make more transform coefficients of high probability to change from one to zero have zero level by quantization itself before the LE stage. Moreover, for the levels expected to keep unchanged, we introduce two approaches to simplify their decision for optimal quantized levels. Our experiment results show 14.39%, 12.88%, and 12.91% of average quantization time saving under all intra (AI-Main), random access (RA-Main), and low delay B (LB-Main) configuration, respectively. Furthermore, the average saving time for the entire encoding is 4.00% for AI-Main, 2.27% for RA-Main, and 2.33% for LB-Main, respectively. In the Bjontegaard delta bit rate (BDBR) sense, our proposed simplified level estimation framework achieves the above time reduction with merely no performance loss.