Predicting the Quality of Demosaiced Images Using the Sparsity of Chroma Gradients
Authors: Carlo Noel Ochotorena, Cecille Adrianne Ochotorena, and Yukihiko Yamashita
Abstract
The design of most modern cameras utilizes a color filter array that downsamples and interleaves the red, green, and blue pixels of an image into a single mosaiced image. Such a design makes it necessary to interpolate the missing pixels for each color channel using a process known as demosaicing. While it is possible to fill in these pixels, the resulting images are inexact estimates of the true image, with different algorithms offering various levels of success. However, this degree of success cannot be directly quantified in the absence of the true image, making it difficult to design adaptive algorithms for demosaicing. This paper explores a no-reference simple metric for inferring the quality of the estimated image by measuring the sparsity of chroma gradients along four directions (SCG4). The said measure is shown to be significantly correlated with respect to the PSNR in simulations using the Kodak image database.