Retinotopy Improves the Categorisation and Localisation of Visual Objects in CNNs

Abstract

Foveated vision is a trait shared by many animals, including humans, but its contribution to visual function compared to species lacking it is still under question. This study suggests that the retinotopic mapping which defines foveated vision may play a critical role in achieving efficient visual performance, notably for image categorisation and localisation. To test for this hypothesis, we transformed regular images by using a Log-polar mapping, and used this retinotopic images as the the input of convolutional neural networks (CNNs). We then applied transfer learning on pre-trained networks on the ImageNet challenge dataset. Our results show that surprisingly, the network re-trained on images which were compressed by the retinotopic mapping performs as well as the retrained network applied to regular images. Moreover, we observed that the retinotopic mapping improves the robustness and localisation of image classification, especially for isolated objects. This was specially acute on a custom version of the dataset which aimed to categorise images that contain or not an animal. In summary, these results suggest that such retinotopic mapping may be an important component of preattentive processes, a central cognitive characteristic of more advanced visual systems.