A feedforward mechanism for human-like contour integration.

Area of Science:

• Computational neuroscience
• Computer vision
• Human visual perception

Background:

• Deep neural networks (DNNs) serve as experimental platforms for studying human visual perception.
• Contour integration, linking local edges into unified percepts, is a key aspect of visual processing.

Purpose of the Study:

• To investigate if feedforward convolutional neural networks (CNNs) can exhibit human-like contour integration and perceptual grouping.
• To identify the necessary properties within CNNs for achieving these human-like visual capacities.

Main Methods:

• Utilized ImageNet pre-trained feedforward CNNs fine-tuned for contour detection.
• Analyzed the impact of architectural properties, specifically progressively increasing receptive fields.
• Investigated the effect of biased fine-tuning on sensitivity to gradual curves.

Main Results:

• Feedforward CNNs, with specific architectural features and fine-tuning, demonstrated human-like contour integration.
• These networks showed human-like sensitivity to curvature, particularly for gradual curves (~20 degrees).
• Fine-tuning also revealed human-like uncrowding capabilities in feedforward networks.

Conclusions:

• Purely feedforward hierarchical computations in CNNs can implement Gestalt principles like 'good continuation' for contour integration and uncrowding.
• These findings suggest that later processing stages in human vision may play a more significant role in perceptual organization than previously thought.
• The study provides a computational existence proof for feedforward mechanisms underlying key aspects of human visual perception.