Enhancing Robustness of Visual Object Localization by Introducing Retina-Inspired Mapping to Convolutional Neural Networks
Keywords: Foveated vision; Convolutional Neural Networks; Transfer learning; Visual categorisation; Saliency
TL;DR: We show that CNNs can fail when attacked by simple image rotation, and implement a biologically inspired retinotopic mapping with robust classification and localization performance.
Abstract: Foveated vision, a trait shared by many animals including humans, has yet to be fully exploited in machine learning applications despite its key contributions to biological visual function. In this study, we investigate whether retinotopic mapping, a critical component of foveated vision, can improve image categorization and localization performance when incorporated into deep convolutional neural networks (CNNs). In particular, we incorporated log-polar retinotopic mapping into the inputs of classic off-the-shelf CNNs and retrained these network on the ImageNet task. Surprisingly, the retinotopically mapped network performed equally well in classification but showed improved robustness to arbitrary image zooms and rotations, especially for isolated objects. In addition, this network showed improved classification localization when the foveated center of the transform was moved, mimicking a key capability of the human visual system that is lacking in standard CNNs. These results suggest that retinotopic mapping may underlie important preattentive visual processes.
Primary Area: applications to neuroscience & cognitive science
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 5651
Loading