A comparison of density-based and feature-based texture boundary segmentation

Previous studies have demonstrated that density of texture elements is an important perceptual aspect of textural appearance, and can enable texture segmentation in the absence of other cues. We compared segmentation thresholds for two kinds of second-order boundaries comprised of two species of micropatterns (e.g., horizontal and vertical Gabors): (1) Feature boundaries, with the same number of total micropatterns on opposite sides but different numbers of each micropattern species within each side, and (2) Density boundaries, with different numbers of total micropatterns on opposite sides, but with the same number of both micropattern species within each side. Contrary to the predictions of a standard late-pooling Filter-Rectify-Filter (FRF) model in which different micropattern specific first-order channels are analyzed by different second stage filters before pooling, we observed lower segmentation thresholds for density boundaries than feature boundaries. This suggests that density boundaries may be detected by a different, early-pooling mechanism. In a second experiment, we considered how two species of micropatterns combine for boundary segmentation. When two single-micropattern-species density boundaries are superimposed in-phase to form a new density boundary, the boundaries formed by each micropattern species combine via probability summation. By contrast, when two single-micropattern-species density boundaries are superimposed in opposite-phase to form a feature boundary, segmentation performance is worse than for either single-micropattern-species boundary alone. We conclude that the mechanisms for density-based texture segmentation are not identical to the mechanisms for feature-based segmentation and that density-sensitive mechanisms most likely integrate across multiple first-order filters responsive to different micropattern species.