Purchase this article with an account.
Susana Chung, Bosco Tjan; Response-triggered covariance analysis of letter features. Journal of Vision 2009;9(8):1000. doi: https://doi.org/10.1167/9.8.1000.
Download citation file:
© ARVO (1962-2015); The Authors (2016-present)
Abnormal feature integration has been suggested as a cause of letter crowding. Nandy and Tjan (2007) reported that the number of letter features utilized by human observers is similar for identifying crowded and non-crowded letters, but there are fewer valid features and more invalid features for crowded letters. Using a set of 26 lowercase letters constructed of Gaussian patches that could be individually turned on or off, last year we reported that the patch locations within a given letter that correlated with an observer's response are largely invariant between the crowded and non-crowded conditions. To ascertain that the result was not an artefact of the assumption that each patch location was independent, and to identify higher-order features used by human observers, this year we adapted a covariance-based reverse correlation technique to examine if the amount of first- and second-order features (formed by a conjunction of patches) utilized for identifying crowded and non-crowded letters remains similar. We considered only the false-alarm trials and only at the target-letter location. We used the principal component analysis to partially prewhiten the distribution of the stimuli that were presented, before computing the mean stimulus (first-order classification image) and the covariance matrix, for each of the 26 letter responses. From the covariance matrix, we computed the second-order classification image for each letter response in the form of a correlogram. The RMS of the pixel values in each classification image was used to quantify the amount of features present. The amount of first- and second-order letter features at the target location was significantly lower (p [[lt]]0.05) for crowded than for non-crowded letters. Considering that the way we perturbed the stimulus did not include any spurious letter features, our finding is consistent with previous report that crowding leads to a decrease in the number of valid features.
This PDF is available to Subscribers Only