Signal-independent separable KLT by offline training for video coding

Abstract

After the works on High Efficiency Video Coding (HEVC) standard, the standard organizations continued to study the next generation of video coding standard, named Versatile Video Coding (VVC). The compression capacity of the VVC standard is expected to be substantially improved relative to the current HEVC standard by evolving the potential coding tools greatly. Transform is a key technique for compression efficiency, and core experiment 6 (CE6) in JVET is established to explore the transform-related coding tools. In this paper, we propose a novel signal-independent separable transform based on the Karhunen-Loève transform (KLT) to improve the efficiency of both intra and inter residual coding. In the proposed method, the drawbacks of the traditional KLT are addressed. A group of mode-independent intra transform matrices is calculated from abundant intra residual samples of all intra modes, while the inter separable KLT matrices are trained with the residuals that cannot be efficiently processed by the discrete cosine transform type II (DCT-II). The separable KLT is developed as an additional transform type apart from DCT-II. The experimental results show that the proposed method can achieve 2.7% and 1.5% bitrate saving averagely under All Intra and Random Access configurations on top of the reference software of VVC (VTM-1.1). In addition, the consistent performance improvement on test set also validates the property of signal independency and the strong generalization capacity of the proposed separable KLT.