MoireDB: A Formula-driven Image Dataset for Robustness Enhancement

Yuto Matsuo, Yoshihiro Fukuhara, Yuki M Asano, Hirokatsu Kataoka, Akio Nakamura

Keywords: synthetic image, data augmentation, robustness in image classification

TL;DR: We propose MoireDB, a formula-generated image dataset, which allows us to enhance image classification robustness when used for data augmentation.

Abstract: Image recognition models have struggled to achieve robustness against real-world degradations and adversarial attacks. In this context, data augmentation methods like PixMix have been shown to enhance robustness. The PixMix framework utilizes generative Fractal arts and Feature Visualizations of CNNs (FVis) as mixing images, which are combined with images from the original training dataset. However, these mixing images suffer from copyright restrictions and high construction costs. To address these challenges, we propose Moire DataBase (MoireDB), a formula-driven Moiré image dataset. MoireDB eliminates copyright concerns, reduces dataset construction costs compared to previous mixing images, and effectively diversifies the perturbations applied to the original images during training. Since each Moiré image is generated from simple mathematical formulas, MoireDB is computationally efficient, eliminating the need for advanced image generation AI and minimizing resource consumption. Experiments on CIFAR-C and CIFAR-based adversarial robustness demonstrate that MoireDB-augmented images in the CIFAR training dataset partially outperform traditional augmentations based on Fractal arts and FVis.

Submission Number: 15