The classical receptive fields of V1 neurons contain only local visual input features. The visual system must group separate local elements into meaningful global features to infer the visual objects in the scene. Local features can group into regions, as in texture segmentation; or into contours which may represent boundaries of underlying objects. I propowse that the primary visual cortex (V1) contributes to both kinds of groupings with a single mechanism of cortical interactions/dynamics mediated by the horizontal connections, and that the dynamics enhance the saliencies of those features in the contours (compared with those in a noisy background) or near the region boundaries (compared with those away from the boundaries). Visual inputs specify the initial neural activity levels, and cortical dynamics modify the neural activities to achieve desired computations. Contours are thereby enhanced through dynamically integrating the mutual facilitation between contour segments, while region boundaries are manifested (and enhanced) in the dynamics because of the breakdown of translation invariance in image characteristics at the region boundaries. I will show analytically and empirically how global phenomena emerge from local features and finite range interactions, how saliency enhancement relates to the contour length and curvature, and how the neural interaction can be computationally designed for region segmentation and figure-ground segregation. The structure and behavior of the model are consistent with experimental observations.
展开▼
