Structured sparse representation of residue in screen content video coding

A novel transform method providing a regionally structured sparse model of residue for a video coding is proposed. Conventional learning-based video coding methods consider only the number of transformed coefficients in a dictionary construction but disregard any structural form of residue. A structural form of coefficients is used for another factor, specified by a set of local block patterns, in a transform design so as to yield the regionally compact distribution of the coefficients. A transform coefficient coding method is also developed because the distribution of the coefficients is different with that of conventional transforms. It is demonstrated with experiments that the proposed method outperforms the coding efficiency and visual quality over high efficiency video coding/screen content coding reference software.