Content Classification Based Reference Frame Reduction and Machine Learning Based Non-square Block Partition Skipping for Inter Prediction of Screen Content Coding

Abstract

Screen Content Coding (SCC) is the extension of the latest video compression standard High Efficiency Video Coding (HEVC). SCC is mainly developed for reducing the bit-rate of videos generated from computers. However, under inter configuration, SCC has large complexity which brings heavy burden to encoding. This paper proposes a content classification based reference frame reduction method and a non-square prediction unit (PU) skipping method to accelerate SCC. In reference frame reduction method, according to number of colors, input coding tree unit (CTUs) will be divided into two classes: natural contents and screen contents. For each class, reference frame can be reduced based on different standard. In PU partition skipping method, five features are extracted from a CTU. The classic learning tool SVM is used to classify CTUs, then six non-square PU partition in depth 1, 2, 3 can be skipped. Finally, 40.83% encoding time saving on average is achieved with only 0.71% BD-rate degradation compared with SCC reference software (SCM6.0).