Probing blur adaptation with reverse correlation

Abstract

Prolonged viewing of blurred or sharpened images makes subsequently viewed images look more sharp or blurred, respectively. A possible mechanism is provided by scale-space models of edge detection, which filter the image with Gaussian derivative operators of different scale (i.e. size), and detect edges by finding responses that are peaks across both space and filter scale. The position of the peak across space indicates edge location, and the position of the peak across filter scale indicates edge blur. To create the peak across filter scale (σ), the output of each filter is multiplied by σγ. γ is conventionally chosen so that the peak occurs in a filter matched in scale to the edge blur. If γ deviates from this default value, then the peak across scale occurs in a filter whose scale is a constant multiple of the edge blur (either larger or smaller). This provides a potential mechanism for blur adaptation: prolonged viewing of blurred or sharpened images could cause the value of γ to be adjusted, causing the peak across scale to occur in a filter with scale larger or smaller than the true edge blur, thereby leading to biases in perceived blur. We estimated the scale of the filter mediating blur discrimination using a task in which participants see two noisy edges and have to say which is the most blurred. By correlating the observer’s responses with the noise samples, we obtain a classification image, a linear approximation of the filter used to do the task. Participants adapted to a blurred or sharpened square wave grating, with Fourier components of frequency f multiplied by 1/f (for blurred gratings) or f (for sharpened gratings). Surprisingly, the classification images were unaffected by blur adaptation. This suggests that the mechanism for blur adaptation is located downstream of low-level blur analysis.