Novel intra prediction via position-dependent filtering

This paper presents a novel intra prediction algorithm, named position-dependent filtering (PDF), to improve the intra prediction accuracy. Different from the existing schemes where the samples along one prediction direction are predicted with the same set of filtering coefficients, in the proposed PDF, position-dependent filtering coefficients are employed, i.e., different sets of filtering coefficients are pre-defined for samples with different coordinates in one coding block. For each intra prediction mode, the set of linear filtering coefficients for each position within one block is obtained from off-line training using the least square method. Moreover, to further reduce the algorithm complexity, a simplified PDF (sPDF) is proposed. In sPDF, only a subset of reference samples are used for prediction and the others are discarded because of the minor contribution to intra prediction. The proposed algorithm has been implemented in the latest ITU-T VCEG KTA software. Experimental results demonstrate that, compared with the original KTA with new intra coding tool enabled, up to 0.53 dB of average coding gain is achieved by the proposed method, while applicable computational complexity is retained for practical video codecs.

Research highlights
► We propose a novel intra prediction method, named position-dependent filtering (PDF).
► In PDF, position- and mode-dependent filtering coefficients are employed.
► The coefficients are obtained from off-line training using the least square method.
► Simplified PDF (sPDF) is proposed using a subset of reference samples for prediction.
► They achieve higher coding gain, while keeping applicable computational complexity.