A dual foveal-peripheral visual processing model implements efficient saccade selection.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Computer vision

Background:

• Human visual search involves specialized pathways for object recognition ('What') and spatial localization ('Where').
• Current computer vision models often lack biomimetic efficiency in processing the full visual field.

Purpose of the Study:

• To develop a visuomotor model that mimics human visual search strategies.
• To improve the efficiency and range of visual search in artificial systems.

Main Methods:

• A dual-pathway model with a 'What' network (deep learning classifier) for foveal vision and a 'Where' network (log-polar encoding) for peripheral vision.
• An actor-critic framework where foveal accuracy trains the 'Where' network to generate an accuracy map for guiding eye movements (saccades).
• Testing on a digit-finding task in cluttered images.

Main Results:

• The model demonstrated a tenfold increase in the radius of detectable and recognizable targets.
• The log-polar encoding enabled sublinear processing of visual information, outperforming mainstream computer vision approaches.
• The 'Where' network successfully guided eye movements toward peripheral objects.

Conclusions:

• The proposed visuomotor model effectively integrates 'What' and 'Where' information for efficient visual search.
• Biomimetic approaches, particularly log-polar encoding, offer significant advantages for large-scale visual search tasks.
• This framework provides a novel computational approach to understanding and replicating human visual attention.