A unified computational framework for visual attention dynamics.

Area of Science:

• Computational Neuroscience
• Computer Vision
• Physics

Background:

• Eye movements are crucial for visual attention, directing the fovea to areas of interest.
• Free visual exploration is a complex process influenced by image properties and individual cognitive states.

Purpose of the Study:

• To propose a novel theory of free visual exploration grounded in physics.
• To develop a model that integrates bottom-up visual principles with top-down semantic information.

Main Methods:

• Formulated a theory of eye movements based on the Principle of Least Action.
• Integrated semantic information using deep convolutional neural networks.
• Validated the model on a diverse image dataset for saliency prediction.

Main Results:

• Differential laws for eye movements emerged naturally from the physics-based framework.
• The model demonstrated strong performance in saliency prediction tasks.
• The theory successfully accounts for both low-level image features and high-level semantic content.

Conclusions:

• The proposed physics-based theory provides a unified framework for understanding eye movements in free visual exploration.
• Integrating semantic information enhances the model's predictive power.
• This approach offers new insights into the interplay between physics, cognition, and vision.