The principal focus of the image quality program is the development of rapidly computable, easy to use image quality metrics. Specifically, we are developing image discrimination metrics. Our models take two images as input and compute an estimate of the probability that an observer will see the difference between the two images. This estimate is represented by the signal-to-noise ratio, d'. Our simple model filters the images with a Difference of Gaussian contrast sensitivity filter, computes the RMS difference between the filtered images, and then divides by a contrast energy masking function. This simple model has been calibrated and can predict the presence or absence of targets in fixed, or identical, backgrounds (Ahumada, Watson, and Rohaly (1995); Rohaly, Ahumada, and Watson (1995); Ahumada (1996); Ahumada and Beard (1996); Ahumada and Beard (1997a,d); Rohaly, Ahumada, and Watson (1997)).
Both the contrast and the masking calculations in this model were based on global average luminance and average contrast energy measures. These global averages will be adequate if the images are fairly uniform in luminance and contrast. To allow the model to work when there are significant variations in luminance and contrast within the image, we are now generalizing the model by making both the contrast and contrast masking spatially local phenomena. Where previously we divided by the overall average luminance to get contrast, we now divide by a local average luminance image. And, rather than dividing the unmasked d' by a single masking value, a masking energy image is formed and used to compute masked images before the RMS difference is computed. We calibrated this model with four different masks in the fovea and the near periphery and found that the contrast pooling region (5 arc min) and the masking pooling regions (10 arc min) did not vary with eccentricity. The simple model with local contrast energy and local luminance is available in the Mathematica language, but is easily converted to other languages for doing mathematics by computer since the only complex operation is convolution by a Gaussian kernel. All other operations are simple pointwise image arithmetic. (Ahumada and Beard, accompanying report in PostScript ).