Abstract
Segmenting visual images into objects and identifying their surface properties require the analysis of local correlations, as these define the lines, edges, and texture. The relevant local correlations are captured by image statistics involving two or more nearby points. In natural images, these correlations occur together in a complex fashion, and across many spatial scales. This study asks how visual analysis of these correlations and their interactions depends on spatial scale. To analyze how the visual system processes multipoint correlations individually and in combination, we developed a space of artificial images in which these correlations can vary independently. The space focuses on specific types of correlations that are informative in natural images (Tkacik et al., PNAS 2010); this yields a 10-parameter domain of binary visual textures. Recently we showed that at a single spatial scale (14 min check size), visual sensitivity was concisely described by a Euclidean metric. Here, we extend the analysis to cover a 10-fold range of check sizes. In N=6 subjects, we measured texture segmentation thresholds (4-AFC paradigm) along all coordinate axes of the texture space and in coordinate planes covering combinations of image statistics from first- to fourth-order. Stimulus size (15 deg to 1.5 deg) varied in proportion to check size, to keep the number of checks constant. Over a fivefold range of check sizes (14 min to 2.8 min), sensitivities to all correlations remained in proportion to each other. For 1.4 min checks, sensitivity to two-point and higher-order correlations was selectively diminished, while sensitivity to first-order (luminance-driven) statistics was, as expected, preserved. We conclude that visual sensitivity to local statistics is approximately scale-invariant. Consequently, the tuning of visual sensitivity to the informativeness of image statistics in natural images (Briguglio et al., VSS 2013) holds across spatial scales as well.
Meeting abstract presented at VSS 2014