Simulation of Digital Camera Images from Hyperspectral Input

Perceptual assessment of demosaicing algorithm performance

“…Before use in the experiment, each image was rendered for display on our monitor using standard methods for color balancing and gamma correction [2]. Images of ten different scenes were used.…”

“…It would provide a near-optimal solution if the R, G, and B responses were statistically independent, as then responses from one sensor class would provide no predictive information about responses in another. For natural im-ages and typical choices of R, G, and B sensor spectral sensitivities, however, such statistical independence of responses across sensor classes does not hold [2], [10]- [12]. The other demosaicing algorithms we considered attempt to take advantage of the correlations between responses in different sensor classes.…”

“…Specifying a lower correlation has the effect of increasing the prior probability of high spatial frequencies in the image. 2 For both the Bayesian 1 and Bayesian 2 algorithms, the estimates at each pixel were obtained from the mean of the posterior obtained when the Bayesian analysis was applied to a 5 5 image region surrounding that pixel.…”

“…1 We would like to thank D. Keren for suggesting the bootstrapping approach. 2 In addition, the Bayesian 2 algorithm assumed a circular symmetric rather than an x0y separable Gauss-Markov model, but this change had only a minor effect on the algorithm's performance. III.…”

“…Nonetheless, it is well known that the use of a mosaiced design can lead to aliasing artifacts in the final demosaiced image [2]. Often, these artifacts manifest themselves as chromatic mottle or splotches in the vicinity of high spatial frequency luminance variation in the image (see Fig.…”