Abstract
Perceptual grouping and selection have been widely studied; however, their mechanisms remain poorly understood. Francis et al. (2017) and Kon and Francis (2021) introduced top-down control of connection and selection circuits that, together, implement a grouping strategy in a neural network model of visual cortex. Here we apply this model to two experiments that explore textural grouping. Beck (1966) used patterns consisting of left, middle, and right regions, each containing 36 nonoverlapping copies of a particular letter, e.g., a backwards letter L, a T, or a T slanted at 45 degrees. The task was to identify the boundary between regions that most naturally divided the pattern into two areas. Since participants tended to group regions if their elements had lines of the same orientations but not if they were of the same shape but slanted, Beck concluded that grouping was modulated by orientation rather than by shape similarity. Beck (1983) used a similar task except participants gave a rating to indicate the degree that a pattern segmented into top and bottom regions. Each image consisted of a matrix of U’s that varied in orientation, proximity, or alignment. Beck interpreted the results as supporting his hypothesis that elements’ proximity and alignment facilitate their linking, which is a key part of many models (e.g., Grossberg & Mingolla, 1985). According to our model, a good connection strategy connects elements within an area but not between areas, e.g., in two of three regions of the Beck (1966) stimuli or in half of the Beck (1983) stimuli. A good selection strategy places selection signals on connected elements in an area to segment them from elements in the neighboring area. Model results are generally consistent with empirical findings and thereby support Beck’s claim that orientation, alignment, and proximity, play fundamental roles in texture segmentation.