Abstract
Statistical features of images are crucial to discrimination of visual textures and image segmentation.
We compared the strength of different statistical cues and tested simple models for how they are processed.
Stimuli consisted of four arrays of black and white checks. In three of the arrays, checks were colored at random; in the fourth array (“the target”), a statistical bias was introduced in local first-order statistics (the number of white checks), local fourth-order statistics (the “even” texture), or in long-range correlations (bilateral symmetry). Each kind of bias was introduced in a graded fashion. Each array subtended 2.7 deg and was centered 4 deg from fixation. The number of checks in each array ranged from 6×6 to 16×16.
Subjects (N=7) were asked (100 ms presentation, 4-AFC without feedback) to identify the target.
For targets that were distinguished by their local statistics, 75% correct performance was achieved with sub-maximal levels of statistical structure. For targets that were distinguished by bilateral symmetry, performance never exceeded approximately 50% correct even with maximal statistical structure.
Some subjects showed a modest implicit priming effect when the target was in the same location on consecutive trials, suggesting a covert direction of attention by the statistically anomalous target.
Conditions with greater statistical structure (and greater fraction correct) were associated with shorter reaction times, but reaction times did not show a corresponding priming effect. We constructed a model (for fraction correct) consisting of detection, pooling, and decision stages. Discriminations based on local statistics were well fit by this model. There was a striking difference in the local detection stage between first-order and higher-order statistics, but similar pooling and decision stages. Performance for discriminations based on symmetry could not be fit in a satisfactory fashion by models of this simple structure.
Support: NIH Grant EY07977