A model of top-down attentional control during visual search in complex scenes

Area of Science:

• Computational vision
• Cognitive neuroscience
• Human visual attention

Background:

• Visual attention research focuses on predicting saccadic endpoints in naturalistic scenes.
• Stimulus-driven (bottom-up) attention guides viewing in task-free settings.
• Goal-driven (top-down) processes bias attention toward target-like features during search tasks.

Purpose of the Study:

• To develop a top-down computational model of visual attention during visual search.
• To predict saccadic endpoint distributions in complex scenes based on target similarity.

Main Methods:

• A top-down model was created using histogram matching to define feature similarity (orientation, spatial frequency).
• An informativeness measure predicted attentional guidance across feature dimensions.
• Model performance was evaluated using human eye-movement data during search tasks.

Main Results:

• The model successfully predicted saccadic endpoint distributions in search displays.
• Model accuracy was comparable to that of human observers' eye movements.

Conclusions:

• Top-down attentional guidance during visual search can be effectively modeled.
• Feature similarity to the search target is a key factor in predicting attention distribution.