Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper addresses the vulnerability of reconstruction-based methods in unsupervised image anomaly detection to outlier leakage, which occurs when standard mean squared error (MSE) loss allows anomalous patterns to dominate the model's learning process. The authors introduce a Non-linear Reconstruction Loss that employs a sigmoid-based squashing function to suppress high-magnitude features, alongside a statistical calibration scheme for selecting a scaling factor from the confidence interval of normal feature distributions. Their approach significantly enhances anomaly detection performance, achieving up to 99.0% Image-AUROC on MVTec-AD, demonstrating its effectiveness in mitigating outlier leakage and improving anomaly localization in industrial inspection contexts.
Non-linear gradient suppression can drastically improve anomaly detection by preventing outlier leakage, achieving record performance on benchmark datasets.
Reconstruction-based methods are a cornerstone of unsupervised image anomaly detection, but they remain vulnerable to \emph{outlier leakage}, where standard mean squared error (MSE) loss drives the model to faithfully reconstruct anomalous patterns. We propose a Non-linear Reconstruction Loss that applies a sigmoid-based squashing function to suppress high-magnitude features, preventing outliers from dominating optimization while preserving sensitivity to normal patterns. In addition, we introduce a statistical calibration scheme that selects the scaling factor $k$ from the confidence interval (CI) of the normal feature distribution, enabling data-driven control of the suppression strength. Our approach achieves competitive or superior anomaly detection performance compared to state-of-the-art methods, reaching 99.0\% Image-AUROC and 97.3\% Pixel-AUROC on MVTec-AD, and 95.3\% Image-AUROC and 99.0\% Pixel-AUROC on VisA. These results indicate that non-linear gradient suppression is an effective mechanism for mitigating outlier leakage and improving anomaly localization in unified industrial inspection settings. The implementation is available at https://github.com/mintii13/Statistical-Non-linear-Reconstruction-Loss.git.