Invariant texture perception is harder with synthetic textures: Implications for models of texture processing

Area of Science:

• Computational vision
• Human visual perception
• Machine learning for image processing

Background:

• Texture synthesis models, like the Portilla-Simoncelli (P-S) model, are valuable tools for understanding human visual texture processing.
• The P-S model's summary statistics effectively synthesize natural textures and explain performance in tasks like crowding and search.
• These statistics may also correlate with the response properties of neurons in visual area V2.

Purpose of the Study:

• To determine if the summary statistics of the Portilla-Simoncelli (P-S) model are sufficient for texture discrimination under varying illumination conditions.
• To investigate the role of illumination invariance in human texture perception.
• To assess the limitations of current texture synthesis models in replicating human visual capabilities.

Main Methods:

• A match-to-sample task was employed using natural textures under different lighting conditions (diffuse overhead vs. side lighting).
• Participants performed the task with both original photographic textures and synthetic textures generated by the P-S model.
• Experiments included conditions with illumination changes between sample and test images, no illumination changes, matched power spectra, and unique texture sampling locations.

Main Results:

• A significant performance cost was observed when using P-S synthetic images compared to original images, even when illumination conditions were consistent.
• This performance decrease occurred in both illumination change and no change conditions, suggesting a general limitation of the P-S model's statistics.
• The negative effect persisted even when controlling for power spectra and minimizing image-based processing, confirming the model's inadequacy for invariant texture perception.

Conclusions:

• The summary statistics implemented in the Portilla-Simoncelli (P-S) model are insufficient for achieving illumination-invariant texture discrimination in human vision.
• Invariant texture perception relies on visual computations or representations not captured by the P-S algorithm.
• Further research is needed to develop texture models that better account for the complexities of human visual processing, particularly illumination invariance.