Efficient coding of natural scene statistics predicts discrimination thresholds for grayscale textures.

Area of Science:

• Computational neuroscience
• Visual perception
• Image statistics

Background:

• The efficient coding principle suggests neural systems optimize for predictable information.
• Previous work linked this principle to visual sensitivity for binary images via the "variance is salience" hypothesis.

Purpose of the Study:

• To test the "variance is salience" hypothesis for grayscale image statistics.
• To determine the limits of this hypothesis in predicting visual sensitivity.
• To explore the relationship between natural scene statistics and visual perception.

Main Methods:

• Defined a 66-dimensional space of local grayscale light-intensity correlations.
• Analyzed the relevance of these correlations in natural scenes.
• Conducted psychophysical experiments using a texture-segregation task with synthetic textures.

Main Results:

• The "variance is salience" hypothesis accurately predicted visual sensitivity to a wide range of second-order grayscale correlations.
• Correlations beyond second order were found to be non-salient.
• Predicted thresholds closely matched experimental results for over 300 second-order correlations (median fractional error <0.13).

Conclusions:

• The "variance is salience" hypothesis provides a detailed explanation for visual sensitivity to local grayscale image statistics.
• The hypothesis's predictive power extends to complex, high-dimensional image properties.
• This framework offers insights into how the visual system processes natural textures.