Fast Mode and Partition Decision Using Machine Learning for Intra-Frame Coding in HEVC Screen Content Coding Extension

This paper presents a fast mode and partition decision framework for screen content coding (SCC) based on machine learning. Extensive statistical studies and complexity evaluations are conducted to explore the distribution of different coding modes and their complexities. The proposed encoder scheme is designed based on the results from these studies. Firstly, a CU is classified as either a natural image block (NIB) or a screen content block (SCB). An SCB only goes through SCC modes at the current CU level. An NIB is further classified as "partitioned" or "non-partitioned". The partitioned block will bypass current level intra modes. The non-partitioned block is classified as "directional" or "non-directional" and only goes through a subset of intra candidates. Decision tree classifiers are designed with chosen features that can distinguish different types of blocks. Furthermore, additional mode/partition checking is terminated once the current mode coding rate is lower than a statistics-based threshold. Compared with HEVC-SCC reference software, our proposed fast encoder can balance the encoding efficiency and complexity by adjusting decision confidence thresholds and rate thresholds. Under all-intra configurations, our "rate-distortion preserving" setting can achieve 40% complexity reduction with only 1.46% BD-loss. Our "complexity-reduction boosting" setting can achieve 52% complexity reduction with 3.65% BD-loss.