Local Image Normalization For Data Augmentation

Several data augmentation techniques, such as Aug-Mix, CutMix, and MixUp, are widely recognized for enhancing the performance of deep learning (DL) models. In this work, we propose a novel and straightforward method to further improve the effectiveness of these augmentation techniques. Our approach leverages biologically inspired local image normalization, demonstrating significant enhancements in the generalization performance of DL models. We validate our method across a range of DL architectures-including ResNet-18, Res NeXt-29, DenseNet, and ResNet-50-and on various datasets such as CIFAR-10C, CIFAR-100C, and ImageNet-C. The robustness and effectiveness of our proposed method are thoroughly examined and showcased through comprehensive experiments. The proposed approach delivers a consistent improvement in terms of the mean Corruption Error (mCE) and RMS calibration error.