Reproduce “Wavelet-Enhanced PaDiM for Industrial 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
Cory Gardner, Byungseok Min, Tae-Hyuk Ahn
Wavelet-Enhanced PaDiM (WE-PaDiM) integrates Discrete Wavelet Transform into the PaDiM framework to provide a structured, frequency-domain approach for feature selection in unsupervised industrial anomaly detection. This method achieved an average Image-level ROC AUC of 99.32% and Pixel-level ROC AUC of 92.10% on the MVTec AD dataset, improving both detection and localization accuracy while maintaining computational efficiency.
Key insights
- Coarse, low-frequency information (LL subband) is most discriminative for overall image-level anomaly detection, while high-frequency detail subbands (LH, HL) are critical for precise pixel-level localization.
- The optimal wavelet configuration (type, level, subband selection) varies significantly based on the CNN backbone and the specific characteristics of anomalies (e.g., texture vs. object categories).
- Replacing random feature selection with a structured, frequency-domain approach provides a more interpretable mechanism, allowing targeted emphasis on features relevant to different anomaly types.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.