High frequency edges (but not contrast) predict where we fixate: A Bayesian system identification analysis.

Area of Science:

• Computational neuroscience
• Visual perception
• Machine learning

Background:

• Understanding visual attention is crucial for fields like artificial intelligence and human-computer interaction.
• Previous models have explored various image features to predict eye movements, but a definitive model remains elusive.

Purpose of the Study:

• To identify specific image characteristics that predict human fixation locations in natural scenes.
• To develop a computational model mapping image features to fixation probabilities using a Bayesian approach.

Main Methods:

• Employed a Bayesian system identification technique to model the relationship between image features and eye fixation probability.
• Utilized a large dataset of human eye-tracking data on natural images.
• Evaluated candidate feature maps including edges, contrast, and luminance at multiple spatial scales.

Main Results:

• High spatial frequency edge information was the dominant predictor of fixation probability.
• The optimal model incorporated a compressive non-linearity (square root) on edge detection filters.
• Low spatial frequency edges and contrast exhibited weaker, inhibitory effects on fixation probability.

Conclusions:

• Human visual attention is primarily driven by high-level edge information in natural images.
• The Bayesian framework effectively models the mapping between image features and visual fixations.
• Fixation probability is influenced by spatial averaging of feature weights over approximately 2 degrees, rather than center-surround inhibition.