Domain Generalization with Vital Phase Augmentation

TL;DR

This work tackles domain generalization for image classification by addressing phase fluctuations in frequency-domain data augmentation. It introduces VIPAug, a vitality-aware augmentation framework that first detects vital phases via a $3\mathrm{D}$ DFT amplitude analysis and then applies differential phase variations: Gaussian-based variations with $\sigma_{\text{vital}}^2$ and $\sigma_{\text{nonvital}}^2$, or fractal-phase replacements for non-vital components, all while performing amplitude augmentation through APR-SP. By coupling these phase variations with amplitude augmentation and reconstructing via the inverse DFT, VIPAug biases the model to rely more on domain-invariant vital phases, yielding strong robustness on CIFAR-10/100 and competitive results on ImageNet-100/ImageNet. The method achieves state-of-the-art performance on CIFAR-10 and CIFAR-100 and near-state-of-the-art results on large-scale datasets, underscoring the practical value of phase-aware augmentation for real-world distribution shifts. The work also provides extensive ablations and analysis of robustness weights, fractal-phase replacement, and cross-DFT comparisons, offering a promising direction for phase-centric domain-generalization research.

Abstract

Deep neural networks have shown remarkable performance in image classification. However, their performance significantly deteriorates with corrupted input data. Domain generalization methods have been proposed to train robust models against out-of-distribution data. Data augmentation in the frequency domain is one of such approaches that enable a model to learn phase features to establish domain-invariant representations. This approach changes the amplitudes of the input data while preserving the phases. However, using fixed phases leads to susceptibility to phase fluctuations because amplitudes and phase fluctuations commonly occur in out-of-distribution. In this study, to address this problem, we introduce an approach using finite variation of the phases of input data rather than maintaining fixed phases. Based on the assumption that the degree of domain-invariant features varies for each phase, we propose a method to distinguish phases based on this degree. In addition, we propose a method called vital phase augmentation (VIPAug) that applies the variation to the phases differently according to the degree of domain-invariant features of given phases. The model depends more on the vital phases that contain more domain-invariant features for attaining robustness to amplitude and phase fluctuations. We present experimental evaluations of our proposed approach, which exhibited improved performance for both clean and corrupted data. VIPAug achieved SOTA performance on the benchmark CIFAR-10 and CIFAR-100 datasets, as well as near-SOTA performance on the ImageNet-100 and ImageNet datasets. Our code is available at https://github.com/excitedkid/vipaug.

Figures (6)

• Figure 1: Amplitude and phase of original and blur images, with reconstructed images. Amplitude and phase fluctuations can be observed in II compared with I. III-(a) retains vital phases of the original image and sets other phases to 0. In contrast, III-(b) retains the non-vital phases at the same ratio and sets other phases to 0. In III, the amplitude of the original image is kept unchanged. The amplitude and non-vital phases are shown to contain less domain-invariant features.
• Figure 2: The number of fluctuated phases from clean to corrupted domains. We extracted an arbitrary image from ImageNet-C and calculated the average value of all corruption types at corruption severity level 3. Phase fluctuations above the threshold value were counted. The range of phase is $[-\pi, \pi]$. Most phases were changed even when the threshold was small.
• Figure 3: Overall structure of VIPAug. VIPAug contains VIPAug-G and VIPAug-F. VIPAug-G introduces phase variations by using Gaussian distributions with different variances, $\sigma_{\text{vital }}^2$ and $\sigma_{\text{nonvital }}^2$. VIPAug-F employs the phases of the fractal images. We introduce finite phase variations by finding vital phases with a filter and applying variations with different strengths depending on the robustness weight. $A$, $P$, and $P^{\prime}$ denote the amplitude, phase, and varied phase spectrum.
• Figure 4: Vital phase detection method. The vital phase coordinates are found by applying the argmax function to the filter size of the amplitude spectrum region. The filter moves on the amplitude spectrums without overlaps. This figure is shown in the case of a gray scale image.
• Figure 5: The ablation analysis of modifications to the phase range of VIPAug-F on ImageNet-100 and ImageNet-100-C. The x-axis stands for the ratio of the modified phase range to the total phase spectrum.