Corruption-Robust Deep Convolutional Networks Inspired by Primary Visual Cortex

While deep convolutional neural networks (DCNNs) make impressive progress in object classi fi cation, they often struggle to recognize objects in images corrupted with different common noise patterns. Recent studies show that simulating a primary visual cortex (V1) at the front of DCNNs leads to improvements in noise-robustness. We develop a corruption-robust DCNN model inspired by V1 orientation selection receptive fields. In this study, we use the ImageNet-C dataset to evaluate the classification accuracy under four categories of image corruption. Our model exhibits an average improvement of 5.1% over the baseline DCNN on the perturbation benchmarks of common image corruptions. Compared with the VOneResNet, the model shows enhanced robustness against all corruption categories except for the digital effect. The experimental results demonstrate that introducing a more accurate V1-inspired model will further improve the robustness of DCNNs against image perturbations.