arxivcs.CV2026-07-14
Statistical Non-linear Reconstruction Loss for Image Anomaly Detection
Nguyen Minh Tri, Hoang Khuong Duy, Huynh Cong Viet Ngu
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 Recons…