General Description
This model is similar to the single filter model for a homogeneous background except that contrast is computed using a local luminance image instead of a single value and no masking is computed. The masking may be not desired to save speed, because appropriate parameters are not known, or the images do not have much masking energy. The model takes two luminance images. One is the standard or background image; the other is different in some way. The images are blurred by an 'optical' blur function and then blurred again to generate local luminance images. These are used to convert to contrast images. The Minkowski distance between the two final images gives a d' estimate for the discriminability of the images.
Model applications
1) Detectability of image processing artifacts
2) Estimation target detection using simulated (exactly replicated)
backgrounds
Capabilities and limitations
1) Images need not be square or have square pixels
2) Luminance may vary across the image.
4) No provision for visual field inhomogeneity.
5) Model is not linear, so threshold estimation must be done by iteration.
System requirements: Mathematica version tested on Mathematica 2.2 for SPARC. Matlab version tested on Matlab 5.2 for MacIntosh.
Data requirements
Input data
1) Number of rows and columns in each image
2) Pixels per degree of visual angle in row and column directions
3) Luminance values for each image
Data format and units
1) Images are assumed to be rectangular arrays.
2) Luminance units do not matter, they are converted to contrast.
Model output data
d' estimate
Uses of model output
1/d' times the difference luminance image is approximate signal threshold
References
Ahumada (2005 ECVP)
A local contrast metric.
Ahumada and Beard
(1998 SID)
A simple vision model for inhomogeneous image quality assessment.
Ahumada, Beard, and Jones
(2005 VSS)
Modeling the detection of blurred visual targets in non-homogeneous
backgrounds.