Foveated Retinotopy Improves Classification and Localization in Convolutional Neural Networks.

Area of Science:

• Computer Vision
• Computational Neuroscience
• Machine Learning

Background:

• Human and animal vision relies on foveated retinal organization for high-acuity central vision and low-resolution peripheral vision.
• This visual processing strategy is conserved in early visual pathways but is under-explored in artificial intelligence.

Purpose of the Study:

• To investigate the impact of incorporating a foveated retinotopic transformation into convolutional neural networks (CNNs) for image classification.
• To explore how this biologically-inspired preprocessing layer affects model accuracy, robustness, and object localization capabilities.

Main Methods:

• A log-polar mapping was applied to standard CNN models to simulate foveated vision.
• The modified models were retrained to evaluate performance changes.
• The sensitivity to fixation point shifts was analyzed to assess saliency mapping and object localization.

Main Results:

• The foveated retinotopic transformation achieved comparable classification accuracy to standard models.
• Significant improvements in robustness to scale and rotation were observed.
• The architecture demonstrated sensitivity to fixation point shifts, enabling effective saliency mapping for object localization.

Conclusions:

• Foveated retinotopy encodes prior geometric knowledge, offering a solution for visual search tasks.
• This approach presents a viable trade-off between classification robustness and localization accuracy.
• The findings provide a proof of concept for integrating biological vision principles into artificial networks for more robust and efficient computer vision systems.