A general framework for directional intra prediction with varying angle for video coding

Video coding standards such as H.264/AVC, HEVC, VVC, etc., all employ intra prediction where a directional prediction uses a single prediction direction inside a coding block. The underlying idea is that the object directionalities remain unchanged over the entire coding block. This is a simplistic model as the object directionalities over a block can change, especially if the block size is relatively large. In such a case, using a single prediction direction over a coding block will lead to smaller coding blocks after rate-distortion (R-D) optimization with side information for partitioning into smaller blocks. A better idea would be to allow the prediction direction to change gradually over the same coding block such that objects with slow changing directionalities can be modelled in larger blocks without splitting them. This would demand additional complexity and the side information required for model parameters. This paper proposes a general framework which supports this change of prediction direction inside a block and presents practical solutions for low complexity implementations. Simulation results with VVC test model (VTM 4) software show Luma BD-rate gain of 0.11% in All Intra (AI) configuration.