Neural dynamics of perceptual grouping: Textures, boundaries, and emergent segmentations

Stephen Grossberg, Ennio MingollaPerception 1985 · DOI

Introduction

A fundamental challenge in visual perception is the transition from local feature detection to global scene understanding. Standard edge detection algorithms often fail in the absence of clear luminance gradients, such as in camouflage or illusory figures like the Kanizsa Triangle. This paper addresses the "uncertainty principle" of vision: the trade-off between orientation sensitivity and spatial localization. The authors seek to explain how neural dynamics resolve these ambiguities to create stable, coherent segmentations of textures and shaded regions in real-time.

Theoretical Framework: The BCS and FCS

The proposed model is a connectionist architecture consisting of two primary interacting systems, implemented via systems of nonlinear differential equations (shunting networks) to simulate real-time neural dynamics:

  1. Boundary Contour (BC) System: This system generates a structural "blueprint" of the scene. It utilizes a hierarchy of orientationally tuned interactions, starting with filters for local contrast detection, followed by a Cooperative-Competitive (CC) Loop. This feedback network employs short-range competition to sharpen edges and long-range cooperation to bridge gaps between aligned segments, synthesizing a global boundary.
  2. Feature Contour (FC) System: This system handles visible qualities such as color and brightness. It uses the boundaries established by the BC System as "fences" to contain the diffusive "filling-in" of featural signals.
Figure 1: Macrocircuit architecture of the visual system model.

The Boundary-Feature Tradeoff and Filling-In

A central insight of this theory is that boundaries themselves are "invisible"—they possess no intrinsic color or brightness. Perception of a boundary arises because it acts as a barrier to the diffusion of featural signals within the FC System. This process, known as "filling-in," occurs within compartments defined by the BC signals.

Figure 3: Schematic of the diffusive filling-in process within the Feature Contour System.

The model also introduces the concept of "end-cutting." Since local edge detectors are inherently weak at line endings, the BC system generates orthogonal boundaries at these points. This mechanism explains the perception of illusory contours and the completion of shapes from sparse inputs, effectively performing a real-time statistical decision to settle on the most likely global interpretation.

Experimental Stimuli and Results

The model was evaluated using classic psychophysical stimuli designed to "stress test" visual perception, including Beck’s texture displays, Glass patterns, and the Café Wall illusion. Computer simulations demonstrated that the BC System could successfully:

  • Generate illusory contours: Completing boundaries across gaps where no physical luminance difference exists.
  • Segment textures: Predicting "pop out" effects based on the statistical distribution of orientations.
  • Explain geometric-optical illusions: Showing how skewed neural responses to staggered tiles in the Café Wall illusion create the global perception of tilted lines.
  • Simulate Neon Color Spreading: Explaining why color appears to bleed out of certain patterns when gaps in BC "fences" allow featural signals to escape.

Discussion and Limitations

While the theory provides a robust framework for 2D monocular vision and unifies various Gestalt rules (proximity, similarity, good continuation) into a single mathematical framework, it faces limitations. In 1985, these simulations were computationally taxing. Furthermore, while the model is primarily bottom-up (preattentive), it acknowledges the role of top-down influences from an Object Recognition System, though the specifics of this interaction require further exploration. The proposed neural loops provide testable predictions for cortical areas V1 and V2, bridging the gap between psychology and neurophysiology.

Summary updated

BibTeX
@inproceedings{grossberg1985neural,
  author = {Stephen Grossberg and Ennio Mingolla},
  title = {{Neural dynamics of perceptual grouping: Textures, boundaries, and emergent segmentations}},
  booktitle = {Perception},
  year = {1985},
  doi = {10.3758/bf03198851},
}
Figure