Abstract
Usually, models are developed to describe visual performance at a specified background luminance level. As visual performance varies with not just stimulus contrast, but also absolute luminance, it has advantage to include the luminance as a variable in the model. Here I present a luminance-dependent visual image-processing based on the concept of implicit masking (Yang et al, 1995; Makous, 1997), implemented here with a front-end low-pass filter, a retinal local compressive nonlinearity described by a modified Naka-Rushton equation, a cortical representation in the Fourier domain, and a frequency dependent compressive nonlinearity. The model is used to fit CSFs over 7 mean luminance levels (Van Nes and Bouman, 1967), and to fit the Modelfest data (Carney et al, 1999).
Using Minkowski summation over every frequency components at a decision stage, the model results have a RMS error of 0.10 log unit with the CSF data and 0.11 log unit with the Modelfest data. For the Modelfest data, the error comes largely from stimuli #35 (a noise pattern) and #43 (a natural scene), where the model estimates are much lower than the experimental data.
After adding (A) a spatial aperture (Watson and Ahumada, 2005) at the front-end and (B) linear frequency summation windows prior to the Minkowski summation, the RMS error for fitting the Modelfest data reduced to 0.06. By just adding either (A) or (B) alone, the resulting RMS error was about 0.08. Nevertheless, adding these components did not improve the fit to the CSF data.