Lossy image coding in the pixel domain using a sparse steering kernel synthesis approach

摘要: Kernel regression has been proven successful for image de-noising, deblocking and reconstruction. These techniques lay the foundation new coding opportunities. In this paper, we introduce a novel compression scheme: Sparse Steering Synthesis Coding (SSKSC). This pre- postprocessor JPEG performs non-uniform sampling based on smoothness of an image, reconstructs missing pixels using adaptive kernel regression. At same time, reduces blocking artifacts from coding. Crucial to technique is that performed while maintaining only small overhead signalization. Compared JPEG, SSKSC achieves gain low bits-per-pixel regions 50% or more PSNR SSIM. A typically in 0.0–0.5 bpp range, SSIM can mostly be achieved 0.0–1.0 range.