Keuper Labs
← All publications

International Journal of Computer Vision (IJCV 2026), Volume 134, Article 461 · 2026

Fix Your Downsampling ASAP! Aliasing and Sinc Artifact Free Pooling in the Fourier Domain

Julia Grabinski, Steffen Jung, Janis Keuper, Margret Keuper

CNNsFrequency AnalysisAnti-AliasingAdversarial Robustness

Why this publication matters

Reducing the resolution of CNN feature maps can distort the information a network uses to make predictions. This paper extends alias-free FLC Pooling with ASAP, which also suppresses sinc interpolation artifacts. Experiments on ImageNet-1k, ImageNet-C, and CIFAR show more stable features and improved robustness against corruptions and adversarial attacks, with clean accuracy similar to the baseline models.

Abstract

Convolutional Neural Networks (CNNs) are successful in various computer vision tasks. From an image and signal processing point of view, this success is counter-intuitive, as the inherent spatial pyramid design of most CNNs is apparently violating basic signal processing laws, i.e. the Sampling Theorem in their downsampling operations. This issue has been broadly neglected until recent work in the context of adversarial attacks and distribution shifts showed that there is a strong correlation between the vulnerability of CNNs and aliasing artifacts induced by bandlimit-violating downsampling. As a remedy, we propose an alias-free downsampling operation in the frequency domain, denoted Frequency Low Cut Pooling (FLC Pooling) which we further extend to Aliasing and Sinc Artifact-free Pooling (ASAP). ASAP is alias-free and removes further artifacts from sinc-interpolation. Our experimental evaluation on ImageNet-1k, ImageNet-C and CIFAR datasets on various CNN architectures demonstrates that networks using FLC Pooling and ASAP as downsampling methods learn more stable features as measured by their robustness against common corruptions and adversarial attacks, while maintaining a clean accuracy similar to the respective baseline models.

Abstract source ↗

Figures

Comparison of MaxPooling, strided downsampling, FLC Pooling, and ASAP on a zebra image.
Figure 1. MaxPooling and strided downsampling distort image structure or introduce aliasing. FLC Pooling removes aliasing but retains ringing artifacts; ASAP also suppresses these sinc interpolation artifacts. Reproduced from Grabinski et al. (2026), CC BY 4.0; converted to WebP. View in source ↗
FLC Pooling pipeline using FFT, a low-frequency cut, and inverse FFT to downsample feature maps without aliasing.
Figure 3. FLC Pooling transforms the input into the Fourier domain, retains the central low-frequency components, and returns to the spatial domain with an inverse FFT. Reproduced from Grabinski et al. (2026), CC BY 4.0; cropped from the PDF and converted to WebP. View in source ↗

Cite this paper

Download .bib 
@article{grabinski2026asap,
  title = {Fix Your Downsampling ASAP! Aliasing and Sinc Artifact Free Pooling in the Fourier Domain},
  author = {Grabinski, Julia and Jung, Steffen and Keuper, Janis and Keuper, Margret},
  journal = {International Journal of Computer Vision},
  year = {2026},
  volume = {134},
  pages = {461},
  doi = {10.1007/s11263-026-03005-9},
  url = {https://link.springer.com/article/10.1007/s11263-026-03005-9}
}

Figures and abstract are reproduced from the linked research sources. Credit remains with the authors and publishers.