Proximity as a Ground-Truth Proxy for Training Texture Discrimination and Segmentation

Area of Science:

• Computational neuroscience
• Computer vision
• Perception psychology

Background:

• Perceptual systems segment scenes into meaningful regions.
• Accurate texture categorization is crucial for scene segmentation.
• Low-level mechanisms identify same/different texture patches.

Purpose of the Study:

• To investigate spatial proximity as a proxy for texture discrimination.
• To train decision variables and bounds directly from natural images.
• To improve scene segmentation using a proximity-based approach.

Main Methods:

• Utilized spatial proximity as ground truth for training.
• Applied decision variables and bounds to natural images without feedback.
• Integrated trained variables into a hierarchical Bayesian observer (HBO) model.

Main Results:

• Spatial proximity effectively trained accurate decision variables and bounds.
• Performance improved by using proximity for final decision adjustments.
• The HBO model achieved excellent image segmentation with arbitrary textures.

Conclusions:

• Proximity discrimination and texture discrimination share mathematically identical decision bounds under certain conditions.
• The proximity proxy is a plausible mechanism for natural selection in perceptual tasks.
• This approach offers a simple yet effective method for texture discrimination and scene segmentation.