Context-adaptive recursive-filtering-based intra prediction in video coding

Abstract

Conventional intra prediction modes in image and video coding generate an estimation of a target block by copying or projecting its causal neighboring pixels along certain angles. Such simple directional model does not work well for complex image structures. A set of context-adaptive intra prediction modes based on recursive filtering is proposed in this paper. The prediction of a block is generated by applying linear filtering over certain previously reconstructed or predicted pixels in the causal neighborhood of each pixel recursively. The filter coefficients are estimated with least squares optimization using previously reconstructed pixels in the above and/or left regions of the current block. The configurations for the filters such as filter taps, position of reference pixels, as well as the location and shape of the training regions are all flexible, making the proposed prediction modes highly adaptive to local image texture contexts. A data-driven approach is used to select the optimal subset of all the possible filter configurations while retaining as much coding gains as possible. The proposed approach is tested on the state-of-the-art AV1 video coding standard. AV1 supports sophisticated intra prediction tools such as recursive filtering, quadratic interpolation filtering, intra block-copy, and the palette mode. Experimental results show that the context-adaptive recursive-filtering-based intra prediction modes can achieve significant improvement in compression efficiency.