Reproduce “Noise & pattern: identity-anchored Tikhonov regularization for robust structural anomaly detection” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alexander Bauer, Klaus-Robert Müller
A Filtering Autoencoder (FAE) approach for structural anomaly detection incorporates identity-anchored Tikhonov regularization via symmetric Gaussian noise and an expressive corruption model. This method achieved state-of-the-art pixel-level anomaly segmentation on the MVTec AD benchmark with a 99.4 P-AUROC.
Key insights
- Symmetrically applying Gaussian noise to both input and target during autoencoder training asymptotically leads to an identity-anchored Tikhonov regularization, penalizing deviations of the reconstruction function's Jacobian from the identity.
- This regularization stabilizes reconstruction of normal regions, enhances fine-grained detail preservation, and improves the model's ability to correct anomalous regions.
- A diverse corruption model with varied shapes, textures, and opacities significantly improves the model's generalization capacity to unseen structural defects.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.