Non local demosaicing

Demosaicing is the process by which from a single CCD sensor measuring only one color component per pixel, red, green or blue, one can infer a fully color information at each pixel. This inference involves a deep understanding of the interaction between colors and the involvement of local geometry. Although quite successful in making such inferences with very small relative error, state of the art demosaicing methods fail when the local geometry cannot been inferred from the neighboring pixels. In such a case, which occurs (e.g.) when thin structures of fine periodic patterns were present in the original, state of the art methods can create disturbing artifacts, known as zipper effect, blur, and color spots. The aim of this paper is to show that by taking into account the more non-local image geometry, these artifacts can be avoided and an almost satisfactory solution found even in the most critical cases. Since the evaluation of artifacts and their removal is problematic, this paper will introduce three new quality assessment principles which can be measured accurately and consistently. One, the “grey to grey” principle, leads to measure how much false color has been added by demosaicing, in particular the zipper effects. The other two criteria, “noise to noise” and ”method noise”, have been already introduced for denoising purposes. The first one measures how well demosaicing manages not to create structure from a non structured datum while the second one evaluates the loss of details, texture and geometry.