Skip to main content
. Author manuscript; available in PMC: 2023 May 16.
Published in final edited form as: Vision Res. 2022 Jul 10;200:108083. doi: 10.1016/j.visres.2022.108083

Figure 3. Human-like peripheral blur is optimal for object recognition.

Figure 3

(A) Schematic of the VGG-16 neural network architecture used to train images

(B) Top-5 accuracy of neural networks with varying peripheral blur. The accuracy of each network was calculated on test images after training it on foveated images with the corresponding blur profile. Each grey line corresponds to Top-5 test accuracies for individual instances of VGG-16 architecture trained from scratch with the random initialization, and the black line represents the average performance over the three instances.

(C) Example images for which the correct category was identified only by the foveated network (with human-like peripheral blur) but not the full-resolution (unfoveated) network. Below each image, the correct object label is shown (top), followed by its rank and posterior probability returned by the unfoveated network (black, second row) and by the foveated network (red, third row).