A Texture Statistics Encoding Model Reveals Hierarchical Feature Selectivity across Human Visual Cortex

Area of Science:

• Neuroscience
• Computational Vision
• Cognitive Science

Background:

• Midlevel visual features like texture link low-level and high-level processing.
• The Portilla-Simoncelli (P-S) model is a candidate for predicting neural responses in early visual areas (V1-V4).
• The model's efficacy in higher visual cortex for natural scenes remains unexplored.

Purpose of the Study:

• To assess how well the P-S texture statistics model predicts neural responses in higher visual cortex to natural scenes.
• To investigate the hierarchical emergence of midlevel feature representations in the human visual system.

Main Methods:

• Constructed single-voxel encoding models using P-S texture statistics.
• Fit models to functional magnetic resonance imaging (fMRI) data from human subjects viewing natural scenes.
• Employed variance partitioning analyses to identify predictive features.

Main Results:

• The P-S model successfully predicted voxel responses in both early retinotopic and higher-level category-selective areas.
• Texture statistics representation is widespread across the visual cortex.
• Higher-order texture features increasingly predict responses in ventral and lateral visual cortex.
• Texture sensitivity patterns helped recover organizational axes, including semantic image content.

Conclusions:

• Texture statistics are broadly represented throughout the visual system.
• The complexity of texture representations increases hierarchically from early to higher visual areas.
• This work provides insights into the role of texture in visual processing and hierarchical feature emergence.