Enhancement of lossy compressed images by modeling with Bernstein polynomials

Abstract

A non-iterative post-processing enhancement technique is proposed for images degraded by either the JPEG-DCT or the JPEG-LS (LOCO) lossy coding algorithm. A degraded image is classified into active and smooth regions. A distance transform is applied to the resulting classification, and is used to determine the size and order of a Bezier surface patch. These Bezier blending surfaces, built with Bernstein polynomials, provide an interesting representation for the image. This approach mitigates the quantization noise while preserving strong edges and textures. Results illustrate the significant visual improvement achieved with a computational complexity of O(n).