Sparse representation based intraframe and semi‐intraframe video coding schemes for low bitrates

Abstract This paper proposes some extensions of the successful sparse coding of still images to intraframe and semi‐intraframe video coding. The presented frameworks apply the efficient K‐singular value decomposition and recursive least squares dictionary learning methods for sparse representation of videos to study their coding performances. In the proposed semi‐intraframe schemes, namely, SISC1 and SISC2, only frame‐blocks with more than a threshold deviation from the blocks of the previous frame are transmitted/coded. This reduces the required bitrate and prevents the sparse coding of similar blocks, leading to more efficient video coding methods. The results show that the dictionary learning‐based intraframe coding improves the rate‐distortion performance of the conventional Motion‐JPEG and Motion‐JPEG2000 at low bitrates for more than about 3 and 0.5 dB of PSNR on average (for 0.2–1 bpp compression), respectively. The proposed methods outperform the basic dictionary learning‐based coding, especially for slower changing videos, generally, with more than 3 dB superiority on average over the tested bitrates. These schemes even present superior performance than the HEVC in the intramode for the complex textured or cluttered scenes. The proposed SISC2 method also saves up to about 50% of the sparse coding computational cost by preventing the coding of more similar frame‐blocks.