Abstract
We compare several functional models of LGN population response in terms of their ability to predict human judgments of visual distortion. The model-derived Fisher Information matrix provides a bound on discrimination thresholds for the visibility of arbitrary distortions. In particular, the largest and smallest eigenvectors of this matrix represent the model-predicted most- and least-noticeable distortions that can be added to a given image. We used a two-alternative forced-choice task to measure the ability of human observers to detect distortions corresponding to addition of multiples of these eigenvectors. Results are quantified using the difference, D, of the log amplitude thresholds for detection of the two eigen-distortions. Two randomly-chosen distortion vectors would yield approximately D = 0, and larger values of D indicate that model sensitivity is better aligned with that of human perception. We tested four nested models: a linear model, a luminance normalized model (LN), a contrast normalized model (CN), and a contrast normalized model with separate On and Off pathways (On-Off). Parameters of each model were set by fitting to the TID 2008 database of human perceptual judgments. We found that each successive level of normalization increased our measure of human sensitivity alignment, (D: Linear, 4.36; LN, 6.85; CN, 8.16; On-Off, 8.62), suggesting that hierarchical normalization, even at the early stages of the visual system, plays a significant role in human perceptual sensitivity to complex stimuli. On the other hand, our crossvalidated correlation with the database(Pearson R: Linear, .66; LN, .74; CN, .83; On-Off, .82), revealed that the On-Off model did not increase over the CN model. We conclude that our method, based on synthesis of model-optimized stimuli, exposes representational limitations and capabilities that cross-validation on curated datasets does not.
Meeting abstract presented at VSS 2017