Abstract
Aim. We measured psychophysical performance of human observers for discrimination of monochrome natural images that could be either blended one into the other (Tolhurst & Tadmor, 2000) or morphed (Parraga et al, 2000) . We propose a visual-cortex based model to predict such performance. The model consists of multiple narrowband spatial filters (Campbell & Robson, 1968) tuned to different spatial frequencies and orientations. Methods. Image discrimination was measured using a 2AFC procedure whereby observers binocularly identified which of two test images was different from a reference image. A staircase method was used to adjust the difference between the test and the reference images. We made related sets of images in which the changing target area could be a small or a large proportion of the overall size (3 deg square) of the image. Morphing the target area generated changes in texture, or shape, or both; blending the target area generated just changes in texture. Results and Conclusions. Observers' discrimination thresholds varied under different conditions, but were relatively unaffected by the size of the target area. To model the psychophysical data, we define an index of image discrimination threshold based on a weighted contrast difference (Minkowski sum with exponent 4) between pairs of images calculated over a range of spatial frequencies and orientations. At first, we treated the cues from different parts of the visual field equally; however, this predicted that the observers' thresholds should have been lower for the larger target sizes than for the smaller. We accounted for target area size, by modeling the regional variation in contrast sensitivity (Robson & Graham, 1981) and cortical magnification factor (Tolhurst & Ling, 1988) across the visual field. Now, the index of image discrimination remains constant across the different sizes of the changing target and appears to be a good predictor of observers' image discrimination thresholds.
EPSRC/Dstl GR/S56399/01 and GRS56405/01