Abstract


= PDF Reprint,     = BibTeX entry,     = Online Abstract


Click to download PDF version Click to download BibTeX data Clik to view abstract L. Itti, J. Braun, C. Koch, Contrast Discrimination can Explain Orientation Discrimination, In: Investigative Ophthalmology and Visual Science (Proc. ARVO 1999), Vol. 40, No. 4, p. 3016, Mar 1999.

Abstract: Purpose: Many current early vision models consist of a population of noisy orientation-selective filters, followed by noiseless central decision. It has been argued (Bowne, Vis Res 1990;30:449-61) that all such models, with noise only at the sensory level (filters), cannot explain the differential dependence of contrast and orientation discrimination thresholds on stimulus contrast (Delta_c propto c^-0.4 versus Delta_theta propto c^-0.1, c>0.1). Most current models indeed predict improvement with contrast at the same rate for Delta_c/c and Delta_theta, both resulting from an overall improvement in signal-to-noise ratio with stimulus contrast. One way to reconcile both observations is to assume additional, task-dependent noise at the central decision stage (Bowne, 1990). Here, we argue that this apparent contradiction constitutes further evidence for a certain type of non-linear interactions among filters (``divisive inhibition'', also known as ``Heeger normalization''). [[[FIGURE]]] Results: A model with noiseless decision was able to simultaneously account for both observations (figure) by implementing strong non-linear excitatory and inhibitory interactions between filters tuned to similar orientations. Resulting from the interactions, the orientation tuning bandwidth of the filters broadened by 25% as c increased from 0.01 to 0.99, which partially canceled the improvement of Delta_theta with c, but did not affect Delta_c/c. Conclusion: ad hoc task-dependent central noise is unnecessary provided that filters interact. Far from constituting a weakness of current spatial vision models, the differential contrast dependence of different types of thresholds corroborates the current views as to the nature of interactions between filters.

Themes: Computational Modeling, Human Psychophysics

 

Copyright © 2000-2007 by the University of Southern California, iLab and Prof. Laurent Itti.
This page generated by bibTOhtml on Wed Feb 15 12:13:56 PST 2017