Organizations: National Institute of Advanced Industrial Science and Technology (AIST) · Tokyo Denki University · Waseda University · Fundamental AI Lab, University of Technology Nuremberg · Visual Geometry Group, University of Oxford
Data augmentation is a key technique for improving the robustness of image classification models. However, many recent approaches rely on diffusion-based synthesis or complex feature mixing strategies, which introduce substantial computational overhead or require external datasets. In this work, we explore a different direction: procedural augmentation based on analytic interference patterns. Unlike conventional augmentation methods that rely on stochastic noise, feature mixing, or generative models, our approach exploits Moire interference to generate structured perturbations spanning a wide range of spatial frequencies. We propose a lightweight augmentation method that procedurally generates Moire textures on-the-fly using a closed-form mathematical formulation. The patterns are synthesized directly in memory with negligible computational cost (0.0026 seconds per image), mixed with training images during training, and immediately discarded, enabling a storage-free augmentation pipeline without external data. Extensive experiments with Vision Transformers demonstrate that the proposed method consistently improves robustness across multiple benchmarks, including ImageNet-C, ImageNet-R, and adversarial benchmarks, outperforming standard augmentation baselines and existing external-data-free augmentation approaches. These results suggest that analytic interference patterns provide a practical and efficient alternative to data-driven generative augmentation methods.
Figures & tables
Figure 1: Overview of our proposed MoireMix data augmentation framework. The method utilizes mathematical formulas to procedurally generate complex interference patterns on-the-fly during training.
Method
External Data-free
IN-C
IN-R
IN-PGD
IN-Moire
IN-val
Baseline
✓
34.4
15.3
1.3
20.3
69.2
Cutout
✓
36.6
16.4
1.8
21.6
70.5
AugMix
✓
39.1
17.2
1.1
27.7
70.7
RandAug
✓
41.8
19.6
1.9
27.5
73.1
MixUp
✓
45.1
19.8
1.3
35.7
74.1
CutMix
✓
41.0
19.3
1.2
27.9
76.2
Table 1: Performance comparison of various data augmentation methods on ImageNet-1k using a ViT-Base model. We report standard accuracy (IN-val) alongside robustness metrics including ImageNet-C, ImageNet-R, PGD adversarial attacks, and Moire-degraded images. Bold values indicate the best results among comparable external-data-free methods.
Figure 2: Learning dynamics of the ViT-Base model on ImageNet-1k across 100 training epochs. We compare Train/Validation Loss and Top-1 Accuracy between traditional spatial augmentations (dashed lines) and texture-blending approaches (solid lines). MoireMix (thick blue line) provides a strong regularization effect, characterized by slower initial convergence but leading to a stable and robust final model state.
Figure 3: Visual examples of procedurally generated texture images utilized as mixing sets in our on-the-fly augmentation pipeline. These specific examples illustrate the augmentation results obtained through a single addition operation. By applying the blending process once, the pipeline efficiently introduces diverse structural perturbations without external dependencies, storage overhead, or computational bottlenecks during training.
Method
Time [s]
IN-C
IN-R
IN-PGD
IN-Moire
IN-val
Baseline
-
34.4
15.3
1.3
20.3
69.2
PerlinNoiseMix
0.0592
54.9
20.6
1.2
32.1
72.6
DeadLeavesMix
0.0050
53.8
23.5
2.1
44.0
74.4
AFA-Mix
0.0172
48.6
24.0
4.7
45.9
74.2
StripeMix
0.0009
50.4
23.0
2.9
35.9
73.4
MoireMix (Ours)
0.0026
50.5
23.1
3.1
36.2
73.1
Table 2: Ablation study comparing different procedural texture types within the on-the-fly mixing pipeline. We report the average CPU generation time per image in seconds alongside standard and robustness benchmarks. MoireMix provides a superior balance between computational efficiency and robustness across diverse corruption types, whereas methods like Perlin Noise introduce significantly higher latency during the training process.
Figure 4: Fourier sensitivity heatmaps visualizing robustness across spatial frequencies and orientations. Each pixel represents the classification error under specific Fourier basis noise. The center indicates low frequencies, while the periphery represents high frequencies. MoireMix exhibits a more generalized reduction in error across the spectrum compared to AFA-Mix and StripeMix, which show localized robustness likely due to evaluation bias.
Number of Waves ( N )
Frequency Range ( fi )
Metric
1
[1,2]
[1,3]∗
[1,4]
[1,5]
[1,7]
[1,10]
[1,15]
[1,20]
[1,30]
Low ( 1≤fi≤33 )
Mid ( 34≤fi≤66 )
High ( 67≤fi≤100 )
Full ∗ ( 1≤fi≤100 )
IN-val
60.3
59.8
60.5
62.0
60.7
61.1
61.8
61.3
61.8
61.7
60.3
60.1
62.0
60.5
IN-R
12.1
11.7
11.8
13.0
12.2
12.4
12.8
12.7
12.7
12.7
11.0
12.5
12.8
11.8
IN-C
33.0
32.3
32.7
34.3
32.6
33.5
33.4
33.4
33.6
33.0
31.0
32.9
32.8
32.7
Table 3: Hyperparameter ablation for MoireMix generation, focusing on the number of superimposed waves and the sampled frequency range. Experiments were conducted using a ViT-Tiny model for 30 epochs. The results demonstrate that increasing structural complexity and utilizing high-frequency bands contribute significantly to generalization. Default settings used in main experiments are indicated by an asterisk (*).
CIFAR-10
CIFAR-100
Mixing Dataset
Clean
C
PGD
Clean
C
PGD
Baseline
95.6
73.6
8.7
78.7
50.0
3.2
Fractal arts
95.8
89.2
18.0
79.7
66.7
6.8
FVis
95.2
90.5
21.4
79.0
69.7
7.7
FractalDB
96.0
88.1
7.8
80.0
65.0
1.5
VisualAtom
95.6
89.2
6.1
78.6
66.6
1.5
Table 4: Robustness evaluation on CIFAR-10 and CIFAR-100 using WideResNet. Values represent Top-1 Accuracy (%); higher is better.
Department of Computer Science and Engineering, Southern University of Science and Technology · The University of Sydney · JD Explore Academy, JD.com Inc +1