Dangers of Demosaicing : Confusion From Correlation

Abstract

Images from colour sensors using Bayer filter arrays require demosaicing before viewing or further analysis. Advanced demosaicing methods use empirical knowledge of inter-channel correlations to reduce interpolation artefacts in the resulting images. These inter-channel correlations are however different for standard RGB cameras and hyperspectral imagers using colour sensors with added narrow-band spectral filtering. We study the effects of conventional demosaicing methods on hyperspectral images with a dataset originally collected without a colour filter array. We find that using advanced methods instead of bilinear interpolation results in an overall increase of 9–14 % in absolute error and a decrease of 1–3 % in PSNR, but also observed a decrease in MSE of 11–13 %. For the corresponding RGB images, the advanced methods improved fidelity as expected. The results also demonstrate that the reconstruction methods that take advantage of correlation transport noise present in a single component to other reconstructed layers.