Low Frequency Non-Separable Transform (LFNST)

Transform is an essential part to remove spatial redundancy in prediction residual blocks for all modern block-based hybrid video coding. In addition, the existing directional intra prediction causes directional pattern in prediction residual and it leads to predictable pattern for transform coefficients. The predictable patterns in transform coefficients are mostly observed in low frequency components. Therefore, in this paper, a non-separable transform, which is called low-frequency non-separable transform (LFNST), is introduced to further compress the redundancy between low-frequency primary transform coefficients, which are transform coefficients from the conventional directional intra prediction. The proposed LFNST is a direct matrix multiplication so that it doesn't require multiple iterations which causes undesirable delay in transform processing. The proposed LFNST is included in next generation video coding standard called versatile video coding (VVC).